1 /* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "oops/objArrayKlass.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/callGenerator.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/cfgnode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/countbitsnode.hpp" 38 #include "opto/intrinsicnode.hpp" 39 #include "opto/idealKit.hpp" 40 #include "opto/mathexactnode.hpp" 41 #include "opto/movenode.hpp" 42 #include "opto/mulnode.hpp" 43 #include "opto/narrowptrnode.hpp" 44 #include "opto/parse.hpp" 45 #include "opto/runtime.hpp" 46 #include "opto/subnode.hpp" 47 #include "prims/nativeLookup.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 #include "trace/traceMacros.hpp" 50 51 class LibraryIntrinsic : public InlineCallGenerator { 52 // Extend the set of intrinsics known to the runtime: 53 public: 54 private: 55 bool _is_virtual; 56 bool _does_virtual_dispatch; 57 int8_t _predicates_count; // Intrinsic is predicated by several conditions 58 int8_t _last_predicate; // Last generated predicate 59 vmIntrinsics::ID _intrinsic_id; 60 61 public: 62 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 63 : InlineCallGenerator(m), 64 _is_virtual(is_virtual), 65 _does_virtual_dispatch(does_virtual_dispatch), 66 _predicates_count((int8_t)predicates_count), 67 _last_predicate((int8_t)-1), 68 _intrinsic_id(id) 69 { 70 } 71 virtual bool is_intrinsic() const { return true; } 72 virtual bool is_virtual() const { return _is_virtual; } 73 virtual bool is_predicated() const { return _predicates_count > 0; } 74 virtual int predicates_count() const { return _predicates_count; } 75 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 76 virtual JVMState* generate(JVMState* jvms); 77 virtual Node* generate_predicate(JVMState* jvms, int predicate); 78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 79 }; 80 81 82 // Local helper class for LibraryIntrinsic: 83 class LibraryCallKit : public GraphKit { 84 private: 85 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 86 Node* _result; // the result node, if any 87 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 88 89 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 90 91 public: 92 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 93 : GraphKit(jvms), 94 _intrinsic(intrinsic), 95 _result(NULL) 96 { 97 // Check if this is a root compile. In that case we don't have a caller. 98 if (!jvms->has_method()) { 99 _reexecute_sp = sp(); 100 } else { 101 // Find out how many arguments the interpreter needs when deoptimizing 102 // and save the stack pointer value so it can used by uncommon_trap. 103 // We find the argument count by looking at the declared signature. 104 bool ignored_will_link; 105 ciSignature* declared_signature = NULL; 106 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 107 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 108 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 109 } 110 } 111 112 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 113 114 ciMethod* caller() const { return jvms()->method(); } 115 int bci() const { return jvms()->bci(); } 116 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 117 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 118 ciMethod* callee() const { return _intrinsic->method(); } 119 120 bool try_to_inline(int predicate); 121 Node* try_to_predicate(int predicate); 122 123 void push_result() { 124 // Push the result onto the stack. 125 if (!stopped() && result() != NULL) { 126 BasicType bt = result()->bottom_type()->basic_type(); 127 push_node(bt, result()); 128 } 129 } 130 131 private: 132 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 133 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 134 } 135 136 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 137 void set_result(RegionNode* region, PhiNode* value); 138 Node* result() { return _result; } 139 140 virtual int reexecute_sp() { return _reexecute_sp; } 141 142 // Helper functions to inline natives 143 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 144 Node* generate_slow_guard(Node* test, RegionNode* region); 145 Node* generate_fair_guard(Node* test, RegionNode* region); 146 Node* generate_negative_guard(Node* index, RegionNode* region, 147 // resulting CastII of index: 148 Node* *pos_index = NULL); 149 Node* generate_limit_guard(Node* offset, Node* subseq_length, 150 Node* array_length, 151 RegionNode* region); 152 Node* generate_current_thread(Node* &tls_output); 153 Node* load_mirror_from_klass(Node* klass); 154 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 155 RegionNode* region, int null_path, 156 int offset); 157 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 158 RegionNode* region, int null_path) { 159 int offset = java_lang_Class::klass_offset_in_bytes(); 160 return load_klass_from_mirror_common(mirror, never_see_null, 161 region, null_path, 162 offset); 163 } 164 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 165 RegionNode* region, int null_path) { 166 int offset = java_lang_Class::array_klass_offset_in_bytes(); 167 return load_klass_from_mirror_common(mirror, never_see_null, 168 region, null_path, 169 offset); 170 } 171 Node* generate_access_flags_guard(Node* kls, 172 int modifier_mask, int modifier_bits, 173 RegionNode* region); 174 Node* generate_interface_guard(Node* kls, RegionNode* region); 175 Node* generate_array_guard(Node* kls, RegionNode* region) { 176 return generate_array_guard_common(kls, region, false, false); 177 } 178 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 179 return generate_array_guard_common(kls, region, false, true); 180 } 181 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 182 return generate_array_guard_common(kls, region, true, false); 183 } 184 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 185 return generate_array_guard_common(kls, region, true, true); 186 } 187 Node* generate_array_guard_common(Node* kls, RegionNode* region, 188 bool obj_array, bool not_array); 189 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 190 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 191 bool is_virtual = false, bool is_static = false); 192 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 193 return generate_method_call(method_id, false, true); 194 } 195 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 196 return generate_method_call(method_id, true, false); 197 } 198 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 199 200 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 201 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 202 bool inline_string_compareTo(); 203 bool inline_string_indexOf(); 204 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 205 bool inline_string_equals(); 206 Node* round_double_node(Node* n); 207 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 208 bool inline_math_native(vmIntrinsics::ID id); 209 bool inline_trig(vmIntrinsics::ID id); 210 bool inline_math(vmIntrinsics::ID id); 211 template <typename OverflowOp> 212 bool inline_math_overflow(Node* arg1, Node* arg2); 213 void inline_math_mathExact(Node* math, Node* test); 214 bool inline_math_addExactI(bool is_increment); 215 bool inline_math_addExactL(bool is_increment); 216 bool inline_math_multiplyExactI(); 217 bool inline_math_multiplyExactL(); 218 bool inline_math_negateExactI(); 219 bool inline_math_negateExactL(); 220 bool inline_math_subtractExactI(bool is_decrement); 221 bool inline_math_subtractExactL(bool is_decrement); 222 bool inline_exp(); 223 bool inline_pow(); 224 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 225 bool inline_min_max(vmIntrinsics::ID id); 226 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 227 // This returns Type::AnyPtr, RawPtr, or OopPtr. 228 int classify_unsafe_addr(Node* &base, Node* &offset); 229 Node* make_unsafe_address(Node* base, Node* offset); 230 // Helper for inline_unsafe_access. 231 // Generates the guards that check whether the result of 232 // Unsafe.getObject should be recorded in an SATB log buffer. 233 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 234 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 235 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 236 static bool klass_needs_init_guard(Node* kls); 237 bool inline_unsafe_allocate(); 238 bool inline_unsafe_copyMemory(); 239 bool inline_native_currentThread(); 240 #ifdef TRACE_HAVE_INTRINSICS 241 bool inline_native_classID(); 242 bool inline_native_threadID(); 243 #endif 244 bool inline_native_time_funcs(address method, const char* funcName); 245 bool inline_native_isInterrupted(); 246 bool inline_native_Class_query(vmIntrinsics::ID id); 247 bool inline_native_subtype_check(); 248 249 bool inline_native_newArray(); 250 bool inline_native_getLength(); 251 bool inline_array_copyOf(bool is_copyOfRange); 252 bool inline_array_equals(); 253 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 254 bool inline_native_clone(bool is_virtual); 255 bool inline_native_Reflection_getCallerClass(); 256 // Helper function for inlining native object hash method 257 bool inline_native_hashcode(bool is_virtual, bool is_static); 258 bool inline_native_getClass(); 259 260 // Helper functions for inlining arraycopy 261 bool inline_arraycopy(); 262 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 263 RegionNode* slow_region); 264 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 265 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 266 bool inline_unsafe_ordered_store(BasicType type); 267 bool inline_unsafe_fence(vmIntrinsics::ID id); 268 bool inline_fp_conversions(vmIntrinsics::ID id); 269 bool inline_number_methods(vmIntrinsics::ID id); 270 bool inline_reference_get(); 271 bool inline_Class_cast(); 272 bool inline_aescrypt_Block(vmIntrinsics::ID id); 273 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 274 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 275 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 276 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 277 bool inline_sha_implCompress(vmIntrinsics::ID id); 278 bool inline_digestBase_implCompressMB(int predicate); 279 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 280 bool long_state, address stubAddr, const char *stubName, 281 Node* src_start, Node* ofs, Node* limit); 282 Node* get_state_from_sha_object(Node *sha_object); 283 Node* get_state_from_sha5_object(Node *sha_object); 284 Node* inline_digestBase_implCompressMB_predicate(int predicate); 285 bool inline_encodeISOArray(); 286 bool inline_updateCRC32(); 287 bool inline_updateBytesCRC32(); 288 bool inline_updateByteBufferCRC32(); 289 bool inline_multiplyToLen(); 290 }; 291 292 293 //---------------------------make_vm_intrinsic---------------------------- 294 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 295 vmIntrinsics::ID id = m->intrinsic_id(); 296 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 297 298 ccstr disable_intr = NULL; 299 300 if ((DisableIntrinsic[0] != '\0' 301 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 302 (method_has_option_value("DisableIntrinsic", disable_intr) 303 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 304 // disabled by a user request on the command line: 305 // example: -XX:DisableIntrinsic=_hashCode,_getClass 306 return NULL; 307 } 308 309 if (!m->is_loaded()) { 310 // do not attempt to inline unloaded methods 311 return NULL; 312 } 313 314 // Only a few intrinsics implement a virtual dispatch. 315 // They are expensive calls which are also frequently overridden. 316 if (is_virtual) { 317 switch (id) { 318 case vmIntrinsics::_hashCode: 319 case vmIntrinsics::_clone: 320 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 321 break; 322 default: 323 return NULL; 324 } 325 } 326 327 // -XX:-InlineNatives disables nearly all intrinsics: 328 if (!InlineNatives) { 329 switch (id) { 330 case vmIntrinsics::_indexOf: 331 case vmIntrinsics::_compareTo: 332 case vmIntrinsics::_equals: 333 case vmIntrinsics::_equalsC: 334 case vmIntrinsics::_getAndAddInt: 335 case vmIntrinsics::_getAndAddLong: 336 case vmIntrinsics::_getAndSetInt: 337 case vmIntrinsics::_getAndSetLong: 338 case vmIntrinsics::_getAndSetObject: 339 case vmIntrinsics::_loadFence: 340 case vmIntrinsics::_storeFence: 341 case vmIntrinsics::_fullFence: 342 break; // InlineNatives does not control String.compareTo 343 case vmIntrinsics::_Reference_get: 344 break; // InlineNatives does not control Reference.get 345 default: 346 return NULL; 347 } 348 } 349 350 int predicates = 0; 351 bool does_virtual_dispatch = false; 352 353 switch (id) { 354 case vmIntrinsics::_compareTo: 355 if (!SpecialStringCompareTo) return NULL; 356 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 357 break; 358 case vmIntrinsics::_indexOf: 359 if (!SpecialStringIndexOf) return NULL; 360 break; 361 case vmIntrinsics::_equals: 362 if (!SpecialStringEquals) return NULL; 363 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 364 break; 365 case vmIntrinsics::_equalsC: 366 if (!SpecialArraysEquals) return NULL; 367 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 368 break; 369 case vmIntrinsics::_arraycopy: 370 if (!InlineArrayCopy) return NULL; 371 break; 372 case vmIntrinsics::_copyMemory: 373 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 374 if (!InlineArrayCopy) return NULL; 375 break; 376 case vmIntrinsics::_hashCode: 377 if (!InlineObjectHash) return NULL; 378 does_virtual_dispatch = true; 379 break; 380 case vmIntrinsics::_clone: 381 does_virtual_dispatch = true; 382 case vmIntrinsics::_copyOf: 383 case vmIntrinsics::_copyOfRange: 384 if (!InlineObjectCopy) return NULL; 385 // These also use the arraycopy intrinsic mechanism: 386 if (!InlineArrayCopy) return NULL; 387 break; 388 case vmIntrinsics::_encodeISOArray: 389 if (!SpecialEncodeISOArray) return NULL; 390 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 391 break; 392 case vmIntrinsics::_checkIndex: 393 // We do not intrinsify this. The optimizer does fine with it. 394 return NULL; 395 396 case vmIntrinsics::_getCallerClass: 397 if (!InlineReflectionGetCallerClass) return NULL; 398 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 399 break; 400 401 case vmIntrinsics::_bitCount_i: 402 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 403 break; 404 405 case vmIntrinsics::_bitCount_l: 406 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 407 break; 408 409 case vmIntrinsics::_numberOfLeadingZeros_i: 410 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 411 break; 412 413 case vmIntrinsics::_numberOfLeadingZeros_l: 414 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 415 break; 416 417 case vmIntrinsics::_numberOfTrailingZeros_i: 418 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 419 break; 420 421 case vmIntrinsics::_numberOfTrailingZeros_l: 422 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 423 break; 424 425 case vmIntrinsics::_reverseBytes_c: 426 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 427 break; 428 case vmIntrinsics::_reverseBytes_s: 429 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 430 break; 431 case vmIntrinsics::_reverseBytes_i: 432 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 433 break; 434 case vmIntrinsics::_reverseBytes_l: 435 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 436 break; 437 438 case vmIntrinsics::_Reference_get: 439 // Use the intrinsic version of Reference.get() so that the value in 440 // the referent field can be registered by the G1 pre-barrier code. 441 // Also add memory barrier to prevent commoning reads from this field 442 // across safepoint since GC can change it value. 443 break; 444 445 case vmIntrinsics::_compareAndSwapObject: 446 #ifdef _LP64 447 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 448 #endif 449 break; 450 451 case vmIntrinsics::_compareAndSwapLong: 452 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 453 break; 454 455 case vmIntrinsics::_getAndAddInt: 456 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 457 break; 458 459 case vmIntrinsics::_getAndAddLong: 460 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 461 break; 462 463 case vmIntrinsics::_getAndSetInt: 464 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 465 break; 466 467 case vmIntrinsics::_getAndSetLong: 468 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 469 break; 470 471 case vmIntrinsics::_getAndSetObject: 472 #ifdef _LP64 473 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 474 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 475 break; 476 #else 477 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 478 break; 479 #endif 480 481 case vmIntrinsics::_aescrypt_encryptBlock: 482 case vmIntrinsics::_aescrypt_decryptBlock: 483 if (!UseAESIntrinsics) return NULL; 484 break; 485 486 case vmIntrinsics::_multiplyToLen: 487 if (!UseMultiplyToLenIntrinsic) return NULL; 488 break; 489 490 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 491 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 492 if (!UseAESIntrinsics) return NULL; 493 // these two require the predicated logic 494 predicates = 1; 495 break; 496 497 case vmIntrinsics::_sha_implCompress: 498 if (!UseSHA1Intrinsics) return NULL; 499 break; 500 501 case vmIntrinsics::_sha2_implCompress: 502 if (!UseSHA256Intrinsics) return NULL; 503 break; 504 505 case vmIntrinsics::_sha5_implCompress: 506 if (!UseSHA512Intrinsics) return NULL; 507 break; 508 509 case vmIntrinsics::_digestBase_implCompressMB: 510 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 511 predicates = 3; 512 break; 513 514 case vmIntrinsics::_updateCRC32: 515 case vmIntrinsics::_updateBytesCRC32: 516 case vmIntrinsics::_updateByteBufferCRC32: 517 if (!UseCRC32Intrinsics) return NULL; 518 break; 519 520 case vmIntrinsics::_incrementExactI: 521 case vmIntrinsics::_addExactI: 522 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 523 break; 524 case vmIntrinsics::_incrementExactL: 525 case vmIntrinsics::_addExactL: 526 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 527 break; 528 case vmIntrinsics::_decrementExactI: 529 case vmIntrinsics::_subtractExactI: 530 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 531 break; 532 case vmIntrinsics::_decrementExactL: 533 case vmIntrinsics::_subtractExactL: 534 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 535 break; 536 case vmIntrinsics::_negateExactI: 537 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 538 break; 539 case vmIntrinsics::_negateExactL: 540 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 541 break; 542 case vmIntrinsics::_multiplyExactI: 543 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 544 break; 545 case vmIntrinsics::_multiplyExactL: 546 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 547 break; 548 549 default: 550 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 551 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 552 break; 553 } 554 555 // -XX:-InlineClassNatives disables natives from the Class class. 556 // The flag applies to all reflective calls, notably Array.newArray 557 // (visible to Java programmers as Array.newInstance). 558 if (m->holder()->name() == ciSymbol::java_lang_Class() || 559 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 560 if (!InlineClassNatives) return NULL; 561 } 562 563 // -XX:-InlineThreadNatives disables natives from the Thread class. 564 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 565 if (!InlineThreadNatives) return NULL; 566 } 567 568 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 569 if (m->holder()->name() == ciSymbol::java_lang_Math() || 570 m->holder()->name() == ciSymbol::java_lang_Float() || 571 m->holder()->name() == ciSymbol::java_lang_Double()) { 572 if (!InlineMathNatives) return NULL; 573 } 574 575 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 576 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 577 if (!InlineUnsafeOps) return NULL; 578 } 579 580 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 581 } 582 583 //----------------------register_library_intrinsics----------------------- 584 // Initialize this file's data structures, for each Compile instance. 585 void Compile::register_library_intrinsics() { 586 // Nothing to do here. 587 } 588 589 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 590 LibraryCallKit kit(jvms, this); 591 Compile* C = kit.C; 592 int nodes = C->unique(); 593 #ifndef PRODUCT 594 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 595 char buf[1000]; 596 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 597 tty->print_cr("Intrinsic %s", str); 598 } 599 #endif 600 ciMethod* callee = kit.callee(); 601 const int bci = kit.bci(); 602 603 // Try to inline the intrinsic. 604 if (kit.try_to_inline(_last_predicate)) { 605 if (C->print_intrinsics() || C->print_inlining()) { 606 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 607 } 608 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 609 if (C->log()) { 610 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 611 vmIntrinsics::name_at(intrinsic_id()), 612 (is_virtual() ? " virtual='1'" : ""), 613 C->unique() - nodes); 614 } 615 // Push the result from the inlined method onto the stack. 616 kit.push_result(); 617 C->print_inlining_update(this); 618 return kit.transfer_exceptions_into_jvms(); 619 } 620 621 // The intrinsic bailed out 622 if (C->print_intrinsics() || C->print_inlining()) { 623 if (jvms->has_method()) { 624 // Not a root compile. 625 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 626 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 627 } else { 628 // Root compile 629 tty->print("Did not generate intrinsic %s%s at bci:%d in", 630 vmIntrinsics::name_at(intrinsic_id()), 631 (is_virtual() ? " (virtual)" : ""), bci); 632 } 633 } 634 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 635 C->print_inlining_update(this); 636 return NULL; 637 } 638 639 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 640 LibraryCallKit kit(jvms, this); 641 Compile* C = kit.C; 642 int nodes = C->unique(); 643 _last_predicate = predicate; 644 #ifndef PRODUCT 645 assert(is_predicated() && predicate < predicates_count(), "sanity"); 646 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 647 char buf[1000]; 648 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 649 tty->print_cr("Predicate for intrinsic %s", str); 650 } 651 #endif 652 ciMethod* callee = kit.callee(); 653 const int bci = kit.bci(); 654 655 Node* slow_ctl = kit.try_to_predicate(predicate); 656 if (!kit.failing()) { 657 if (C->print_intrinsics() || C->print_inlining()) { 658 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 659 } 660 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 661 if (C->log()) { 662 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 663 vmIntrinsics::name_at(intrinsic_id()), 664 (is_virtual() ? " virtual='1'" : ""), 665 C->unique() - nodes); 666 } 667 return slow_ctl; // Could be NULL if the check folds. 668 } 669 670 // The intrinsic bailed out 671 if (C->print_intrinsics() || C->print_inlining()) { 672 if (jvms->has_method()) { 673 // Not a root compile. 674 const char* msg = "failed to generate predicate for intrinsic"; 675 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 676 } else { 677 // Root compile 678 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 679 vmIntrinsics::name_at(intrinsic_id()), 680 (is_virtual() ? " (virtual)" : ""), bci); 681 } 682 } 683 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 684 return NULL; 685 } 686 687 bool LibraryCallKit::try_to_inline(int predicate) { 688 // Handle symbolic names for otherwise undistinguished boolean switches: 689 const bool is_store = true; 690 const bool is_native_ptr = true; 691 const bool is_static = true; 692 const bool is_volatile = true; 693 694 if (!jvms()->has_method()) { 695 // Root JVMState has a null method. 696 assert(map()->memory()->Opcode() == Op_Parm, ""); 697 // Insert the memory aliasing node 698 set_all_memory(reset_memory()); 699 } 700 assert(merged_memory(), ""); 701 702 703 switch (intrinsic_id()) { 704 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 705 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 706 case vmIntrinsics::_getClass: return inline_native_getClass(); 707 708 case vmIntrinsics::_dsin: 709 case vmIntrinsics::_dcos: 710 case vmIntrinsics::_dtan: 711 case vmIntrinsics::_dabs: 712 case vmIntrinsics::_datan2: 713 case vmIntrinsics::_dsqrt: 714 case vmIntrinsics::_dexp: 715 case vmIntrinsics::_dlog: 716 case vmIntrinsics::_dlog10: 717 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 718 719 case vmIntrinsics::_min: 720 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 721 722 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 723 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 724 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 725 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 726 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 727 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 728 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 729 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 730 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 731 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 732 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 733 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 734 735 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 736 737 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 738 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 739 case vmIntrinsics::_equals: return inline_string_equals(); 740 741 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 742 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 743 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 744 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 745 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 746 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 747 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 748 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 749 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 750 751 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 752 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 753 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 754 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 755 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 756 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 757 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 758 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 759 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 760 761 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 762 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 763 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 764 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 765 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 766 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 767 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 768 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 769 770 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 771 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 772 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 773 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 774 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 775 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 776 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 777 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 778 779 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 780 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 781 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 782 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 783 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 784 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 785 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 786 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 787 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 788 789 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 790 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 791 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 792 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 793 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 794 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 795 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 796 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 797 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 798 799 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 800 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 801 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 802 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 803 804 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 805 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 806 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 807 808 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 809 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 810 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 811 812 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 813 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 814 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 815 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 816 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 817 818 case vmIntrinsics::_loadFence: 819 case vmIntrinsics::_storeFence: 820 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 821 822 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 823 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 824 825 #ifdef TRACE_HAVE_INTRINSICS 826 case vmIntrinsics::_classID: return inline_native_classID(); 827 case vmIntrinsics::_threadID: return inline_native_threadID(); 828 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 829 #endif 830 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 831 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 832 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 833 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 834 case vmIntrinsics::_newArray: return inline_native_newArray(); 835 case vmIntrinsics::_getLength: return inline_native_getLength(); 836 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 837 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 838 case vmIntrinsics::_equalsC: return inline_array_equals(); 839 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 840 841 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 842 843 case vmIntrinsics::_isInstance: 844 case vmIntrinsics::_getModifiers: 845 case vmIntrinsics::_isInterface: 846 case vmIntrinsics::_isArray: 847 case vmIntrinsics::_isPrimitive: 848 case vmIntrinsics::_getSuperclass: 849 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 850 851 case vmIntrinsics::_floatToRawIntBits: 852 case vmIntrinsics::_floatToIntBits: 853 case vmIntrinsics::_intBitsToFloat: 854 case vmIntrinsics::_doubleToRawLongBits: 855 case vmIntrinsics::_doubleToLongBits: 856 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 857 858 case vmIntrinsics::_numberOfLeadingZeros_i: 859 case vmIntrinsics::_numberOfLeadingZeros_l: 860 case vmIntrinsics::_numberOfTrailingZeros_i: 861 case vmIntrinsics::_numberOfTrailingZeros_l: 862 case vmIntrinsics::_bitCount_i: 863 case vmIntrinsics::_bitCount_l: 864 case vmIntrinsics::_reverseBytes_i: 865 case vmIntrinsics::_reverseBytes_l: 866 case vmIntrinsics::_reverseBytes_s: 867 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 868 869 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 870 871 case vmIntrinsics::_Reference_get: return inline_reference_get(); 872 873 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 874 875 case vmIntrinsics::_aescrypt_encryptBlock: 876 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 877 878 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 879 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 880 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 881 882 case vmIntrinsics::_sha_implCompress: 883 case vmIntrinsics::_sha2_implCompress: 884 case vmIntrinsics::_sha5_implCompress: 885 return inline_sha_implCompress(intrinsic_id()); 886 887 case vmIntrinsics::_digestBase_implCompressMB: 888 return inline_digestBase_implCompressMB(predicate); 889 890 case vmIntrinsics::_multiplyToLen: 891 return inline_multiplyToLen(); 892 893 case vmIntrinsics::_encodeISOArray: 894 return inline_encodeISOArray(); 895 896 case vmIntrinsics::_updateCRC32: 897 return inline_updateCRC32(); 898 case vmIntrinsics::_updateBytesCRC32: 899 return inline_updateBytesCRC32(); 900 case vmIntrinsics::_updateByteBufferCRC32: 901 return inline_updateByteBufferCRC32(); 902 903 default: 904 // If you get here, it may be that someone has added a new intrinsic 905 // to the list in vmSymbols.hpp without implementing it here. 906 #ifndef PRODUCT 907 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 908 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 909 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 910 } 911 #endif 912 return false; 913 } 914 } 915 916 Node* LibraryCallKit::try_to_predicate(int predicate) { 917 if (!jvms()->has_method()) { 918 // Root JVMState has a null method. 919 assert(map()->memory()->Opcode() == Op_Parm, ""); 920 // Insert the memory aliasing node 921 set_all_memory(reset_memory()); 922 } 923 assert(merged_memory(), ""); 924 925 switch (intrinsic_id()) { 926 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 927 return inline_cipherBlockChaining_AESCrypt_predicate(false); 928 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 929 return inline_cipherBlockChaining_AESCrypt_predicate(true); 930 case vmIntrinsics::_digestBase_implCompressMB: 931 return inline_digestBase_implCompressMB_predicate(predicate); 932 933 default: 934 // If you get here, it may be that someone has added a new intrinsic 935 // to the list in vmSymbols.hpp without implementing it here. 936 #ifndef PRODUCT 937 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 938 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 939 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 940 } 941 #endif 942 Node* slow_ctl = control(); 943 set_control(top()); // No fast path instrinsic 944 return slow_ctl; 945 } 946 } 947 948 //------------------------------set_result------------------------------- 949 // Helper function for finishing intrinsics. 950 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 951 record_for_igvn(region); 952 set_control(_gvn.transform(region)); 953 set_result( _gvn.transform(value)); 954 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 955 } 956 957 //------------------------------generate_guard--------------------------- 958 // Helper function for generating guarded fast-slow graph structures. 959 // The given 'test', if true, guards a slow path. If the test fails 960 // then a fast path can be taken. (We generally hope it fails.) 961 // In all cases, GraphKit::control() is updated to the fast path. 962 // The returned value represents the control for the slow path. 963 // The return value is never 'top'; it is either a valid control 964 // or NULL if it is obvious that the slow path can never be taken. 965 // Also, if region and the slow control are not NULL, the slow edge 966 // is appended to the region. 967 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 968 if (stopped()) { 969 // Already short circuited. 970 return NULL; 971 } 972 973 // Build an if node and its projections. 974 // If test is true we take the slow path, which we assume is uncommon. 975 if (_gvn.type(test) == TypeInt::ZERO) { 976 // The slow branch is never taken. No need to build this guard. 977 return NULL; 978 } 979 980 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 981 982 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 983 if (if_slow == top()) { 984 // The slow branch is never taken. No need to build this guard. 985 return NULL; 986 } 987 988 if (region != NULL) 989 region->add_req(if_slow); 990 991 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 992 set_control(if_fast); 993 994 return if_slow; 995 } 996 997 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 998 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 999 } 1000 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1001 return generate_guard(test, region, PROB_FAIR); 1002 } 1003 1004 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1005 Node* *pos_index) { 1006 if (stopped()) 1007 return NULL; // already stopped 1008 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1009 return NULL; // index is already adequately typed 1010 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1011 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1012 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1013 if (is_neg != NULL && pos_index != NULL) { 1014 // Emulate effect of Parse::adjust_map_after_if. 1015 Node* ccast = new CastIINode(index, TypeInt::POS); 1016 ccast->set_req(0, control()); 1017 (*pos_index) = _gvn.transform(ccast); 1018 } 1019 return is_neg; 1020 } 1021 1022 // Make sure that 'position' is a valid limit index, in [0..length]. 1023 // There are two equivalent plans for checking this: 1024 // A. (offset + copyLength) unsigned<= arrayLength 1025 // B. offset <= (arrayLength - copyLength) 1026 // We require that all of the values above, except for the sum and 1027 // difference, are already known to be non-negative. 1028 // Plan A is robust in the face of overflow, if offset and copyLength 1029 // are both hugely positive. 1030 // 1031 // Plan B is less direct and intuitive, but it does not overflow at 1032 // all, since the difference of two non-negatives is always 1033 // representable. Whenever Java methods must perform the equivalent 1034 // check they generally use Plan B instead of Plan A. 1035 // For the moment we use Plan A. 1036 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1037 Node* subseq_length, 1038 Node* array_length, 1039 RegionNode* region) { 1040 if (stopped()) 1041 return NULL; // already stopped 1042 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1043 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1044 return NULL; // common case of whole-array copy 1045 Node* last = subseq_length; 1046 if (!zero_offset) // last += offset 1047 last = _gvn.transform(new AddINode(last, offset)); 1048 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1049 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1050 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1051 return is_over; 1052 } 1053 1054 1055 //--------------------------generate_current_thread-------------------- 1056 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1057 ciKlass* thread_klass = env()->Thread_klass(); 1058 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1059 Node* thread = _gvn.transform(new ThreadLocalNode()); 1060 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1061 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1062 tls_output = thread; 1063 return threadObj; 1064 } 1065 1066 1067 //------------------------------make_string_method_node------------------------ 1068 // Helper method for String intrinsic functions. This version is called 1069 // with str1 and str2 pointing to String object nodes. 1070 // 1071 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1072 Node* no_ctrl = NULL; 1073 1074 // Get start addr of string 1075 Node* str1_value = load_String_value(no_ctrl, str1); 1076 Node* str1_offset = load_String_offset(no_ctrl, str1); 1077 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1078 1079 // Get length of string 1 1080 Node* str1_len = load_String_length(no_ctrl, str1); 1081 1082 Node* str2_value = load_String_value(no_ctrl, str2); 1083 Node* str2_offset = load_String_offset(no_ctrl, str2); 1084 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1085 1086 Node* str2_len = NULL; 1087 Node* result = NULL; 1088 1089 switch (opcode) { 1090 case Op_StrIndexOf: 1091 // Get length of string 2 1092 str2_len = load_String_length(no_ctrl, str2); 1093 1094 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1095 str1_start, str1_len, str2_start, str2_len); 1096 break; 1097 case Op_StrComp: 1098 // Get length of string 2 1099 str2_len = load_String_length(no_ctrl, str2); 1100 1101 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1102 str1_start, str1_len, str2_start, str2_len); 1103 break; 1104 case Op_StrEquals: 1105 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1106 str1_start, str2_start, str1_len); 1107 break; 1108 default: 1109 ShouldNotReachHere(); 1110 return NULL; 1111 } 1112 1113 // All these intrinsics have checks. 1114 C->set_has_split_ifs(true); // Has chance for split-if optimization 1115 1116 return _gvn.transform(result); 1117 } 1118 1119 // Helper method for String intrinsic functions. This version is called 1120 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1121 // to Int nodes containing the lenghts of str1 and str2. 1122 // 1123 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1124 Node* result = NULL; 1125 switch (opcode) { 1126 case Op_StrIndexOf: 1127 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1128 str1_start, cnt1, str2_start, cnt2); 1129 break; 1130 case Op_StrComp: 1131 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1132 str1_start, cnt1, str2_start, cnt2); 1133 break; 1134 case Op_StrEquals: 1135 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1136 str1_start, str2_start, cnt1); 1137 break; 1138 default: 1139 ShouldNotReachHere(); 1140 return NULL; 1141 } 1142 1143 // All these intrinsics have checks. 1144 C->set_has_split_ifs(true); // Has chance for split-if optimization 1145 1146 return _gvn.transform(result); 1147 } 1148 1149 //------------------------------inline_string_compareTo------------------------ 1150 // public int java.lang.String.compareTo(String anotherString); 1151 bool LibraryCallKit::inline_string_compareTo() { 1152 Node* receiver = null_check(argument(0)); 1153 Node* arg = null_check(argument(1)); 1154 if (stopped()) { 1155 return true; 1156 } 1157 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1158 return true; 1159 } 1160 1161 //------------------------------inline_string_equals------------------------ 1162 bool LibraryCallKit::inline_string_equals() { 1163 Node* receiver = null_check_receiver(); 1164 // NOTE: Do not null check argument for String.equals() because spec 1165 // allows to specify NULL as argument. 1166 Node* argument = this->argument(1); 1167 if (stopped()) { 1168 return true; 1169 } 1170 1171 // paths (plus control) merge 1172 RegionNode* region = new RegionNode(5); 1173 Node* phi = new PhiNode(region, TypeInt::BOOL); 1174 1175 // does source == target string? 1176 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1177 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1178 1179 Node* if_eq = generate_slow_guard(bol, NULL); 1180 if (if_eq != NULL) { 1181 // receiver == argument 1182 phi->init_req(2, intcon(1)); 1183 region->init_req(2, if_eq); 1184 } 1185 1186 // get String klass for instanceOf 1187 ciInstanceKlass* klass = env()->String_klass(); 1188 1189 if (!stopped()) { 1190 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1191 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1192 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1193 1194 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1195 //instanceOf == true, fallthrough 1196 1197 if (inst_false != NULL) { 1198 phi->init_req(3, intcon(0)); 1199 region->init_req(3, inst_false); 1200 } 1201 } 1202 1203 if (!stopped()) { 1204 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1205 1206 // Properly cast the argument to String 1207 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1208 // This path is taken only when argument's type is String:NotNull. 1209 argument = cast_not_null(argument, false); 1210 1211 Node* no_ctrl = NULL; 1212 1213 // Get start addr of receiver 1214 Node* receiver_val = load_String_value(no_ctrl, receiver); 1215 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1216 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1217 1218 // Get length of receiver 1219 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1220 1221 // Get start addr of argument 1222 Node* argument_val = load_String_value(no_ctrl, argument); 1223 Node* argument_offset = load_String_offset(no_ctrl, argument); 1224 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1225 1226 // Get length of argument 1227 Node* argument_cnt = load_String_length(no_ctrl, argument); 1228 1229 // Check for receiver count != argument count 1230 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1231 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1232 Node* if_ne = generate_slow_guard(bol, NULL); 1233 if (if_ne != NULL) { 1234 phi->init_req(4, intcon(0)); 1235 region->init_req(4, if_ne); 1236 } 1237 1238 // Check for count == 0 is done by assembler code for StrEquals. 1239 1240 if (!stopped()) { 1241 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1242 phi->init_req(1, equals); 1243 region->init_req(1, control()); 1244 } 1245 } 1246 1247 // post merge 1248 set_control(_gvn.transform(region)); 1249 record_for_igvn(region); 1250 1251 set_result(_gvn.transform(phi)); 1252 return true; 1253 } 1254 1255 //------------------------------inline_array_equals---------------------------- 1256 bool LibraryCallKit::inline_array_equals() { 1257 Node* arg1 = argument(0); 1258 Node* arg2 = argument(1); 1259 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1260 return true; 1261 } 1262 1263 // Java version of String.indexOf(constant string) 1264 // class StringDecl { 1265 // StringDecl(char[] ca) { 1266 // offset = 0; 1267 // count = ca.length; 1268 // value = ca; 1269 // } 1270 // int offset; 1271 // int count; 1272 // char[] value; 1273 // } 1274 // 1275 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1276 // int targetOffset, int cache_i, int md2) { 1277 // int cache = cache_i; 1278 // int sourceOffset = string_object.offset; 1279 // int sourceCount = string_object.count; 1280 // int targetCount = target_object.length; 1281 // 1282 // int targetCountLess1 = targetCount - 1; 1283 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1284 // 1285 // char[] source = string_object.value; 1286 // char[] target = target_object; 1287 // int lastChar = target[targetCountLess1]; 1288 // 1289 // outer_loop: 1290 // for (int i = sourceOffset; i < sourceEnd; ) { 1291 // int src = source[i + targetCountLess1]; 1292 // if (src == lastChar) { 1293 // // With random strings and a 4-character alphabet, 1294 // // reverse matching at this point sets up 0.8% fewer 1295 // // frames, but (paradoxically) makes 0.3% more probes. 1296 // // Since those probes are nearer the lastChar probe, 1297 // // there is may be a net D$ win with reverse matching. 1298 // // But, reversing loop inhibits unroll of inner loop 1299 // // for unknown reason. So, does running outer loop from 1300 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1301 // for (int j = 0; j < targetCountLess1; j++) { 1302 // if (target[targetOffset + j] != source[i+j]) { 1303 // if ((cache & (1 << source[i+j])) == 0) { 1304 // if (md2 < j+1) { 1305 // i += j+1; 1306 // continue outer_loop; 1307 // } 1308 // } 1309 // i += md2; 1310 // continue outer_loop; 1311 // } 1312 // } 1313 // return i - sourceOffset; 1314 // } 1315 // if ((cache & (1 << src)) == 0) { 1316 // i += targetCountLess1; 1317 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1318 // i++; 1319 // } 1320 // return -1; 1321 // } 1322 1323 //------------------------------string_indexOf------------------------ 1324 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1325 jint cache_i, jint md2_i) { 1326 1327 Node* no_ctrl = NULL; 1328 float likely = PROB_LIKELY(0.9); 1329 float unlikely = PROB_UNLIKELY(0.9); 1330 1331 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1332 1333 Node* source = load_String_value(no_ctrl, string_object); 1334 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1335 Node* sourceCount = load_String_length(no_ctrl, string_object); 1336 1337 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1338 jint target_length = target_array->length(); 1339 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1340 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1341 1342 // String.value field is known to be @Stable. 1343 if (UseImplicitStableValues) { 1344 target = cast_array_to_stable(target, target_type); 1345 } 1346 1347 IdealKit kit(this, false, true); 1348 #define __ kit. 1349 Node* zero = __ ConI(0); 1350 Node* one = __ ConI(1); 1351 Node* cache = __ ConI(cache_i); 1352 Node* md2 = __ ConI(md2_i); 1353 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1354 Node* targetCountLess1 = __ ConI(target_length - 1); 1355 Node* targetOffset = __ ConI(targetOffset_i); 1356 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1357 1358 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1359 Node* outer_loop = __ make_label(2 /* goto */); 1360 Node* return_ = __ make_label(1); 1361 1362 __ set(rtn,__ ConI(-1)); 1363 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1364 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1365 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1366 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1367 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1368 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1369 Node* tpj = __ AddI(targetOffset, __ value(j)); 1370 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1371 Node* ipj = __ AddI(__ value(i), __ value(j)); 1372 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1373 __ if_then(targ, BoolTest::ne, src2); { 1374 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1375 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1376 __ increment(i, __ AddI(__ value(j), one)); 1377 __ goto_(outer_loop); 1378 } __ end_if(); __ dead(j); 1379 }__ end_if(); __ dead(j); 1380 __ increment(i, md2); 1381 __ goto_(outer_loop); 1382 }__ end_if(); 1383 __ increment(j, one); 1384 }__ end_loop(); __ dead(j); 1385 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1386 __ goto_(return_); 1387 }__ end_if(); 1388 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1389 __ increment(i, targetCountLess1); 1390 }__ end_if(); 1391 __ increment(i, one); 1392 __ bind(outer_loop); 1393 }__ end_loop(); __ dead(i); 1394 __ bind(return_); 1395 1396 // Final sync IdealKit and GraphKit. 1397 final_sync(kit); 1398 Node* result = __ value(rtn); 1399 #undef __ 1400 C->set_has_loops(true); 1401 return result; 1402 } 1403 1404 //------------------------------inline_string_indexOf------------------------ 1405 bool LibraryCallKit::inline_string_indexOf() { 1406 Node* receiver = argument(0); 1407 Node* arg = argument(1); 1408 1409 Node* result; 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 // Make the merge point 1422 RegionNode* result_rgn = new RegionNode(4); 1423 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1424 Node* no_ctrl = NULL; 1425 1426 // Get start addr of source string 1427 Node* source = load_String_value(no_ctrl, receiver); 1428 Node* source_offset = load_String_offset(no_ctrl, receiver); 1429 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1430 1431 // Get length of source string 1432 Node* source_cnt = load_String_length(no_ctrl, receiver); 1433 1434 // Get start addr of substring 1435 Node* substr = load_String_value(no_ctrl, arg); 1436 Node* substr_offset = load_String_offset(no_ctrl, arg); 1437 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1438 1439 // Get length of source string 1440 Node* substr_cnt = load_String_length(no_ctrl, arg); 1441 1442 // Check for substr count > string count 1443 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1444 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1445 Node* if_gt = generate_slow_guard(bol, NULL); 1446 if (if_gt != NULL) { 1447 result_phi->init_req(2, intcon(-1)); 1448 result_rgn->init_req(2, if_gt); 1449 } 1450 1451 if (!stopped()) { 1452 // Check for substr count == 0 1453 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1454 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1455 Node* if_zero = generate_slow_guard(bol, NULL); 1456 if (if_zero != NULL) { 1457 result_phi->init_req(3, intcon(0)); 1458 result_rgn->init_req(3, if_zero); 1459 } 1460 } 1461 1462 if (!stopped()) { 1463 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1464 result_phi->init_req(1, result); 1465 result_rgn->init_req(1, control()); 1466 } 1467 set_control(_gvn.transform(result_rgn)); 1468 record_for_igvn(result_rgn); 1469 result = _gvn.transform(result_phi); 1470 1471 } else { // Use LibraryCallKit::string_indexOf 1472 // don't intrinsify if argument isn't a constant string. 1473 if (!arg->is_Con()) { 1474 return false; 1475 } 1476 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1477 if (str_type == NULL) { 1478 return false; 1479 } 1480 ciInstanceKlass* klass = env()->String_klass(); 1481 ciObject* str_const = str_type->const_oop(); 1482 if (str_const == NULL || str_const->klass() != klass) { 1483 return false; 1484 } 1485 ciInstance* str = str_const->as_instance(); 1486 assert(str != NULL, "must be instance"); 1487 1488 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1489 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1490 1491 int o; 1492 int c; 1493 if (java_lang_String::has_offset_field()) { 1494 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1495 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1496 } else { 1497 o = 0; 1498 c = pat->length(); 1499 } 1500 1501 // constant strings have no offset and count == length which 1502 // simplifies the resulting code somewhat so lets optimize for that. 1503 if (o != 0 || c != pat->length()) { 1504 return false; 1505 } 1506 1507 receiver = null_check(receiver, T_OBJECT); 1508 // NOTE: No null check on the argument is needed since it's a constant String oop. 1509 if (stopped()) { 1510 return true; 1511 } 1512 1513 // The null string as a pattern always returns 0 (match at beginning of string) 1514 if (c == 0) { 1515 set_result(intcon(0)); 1516 return true; 1517 } 1518 1519 // Generate default indexOf 1520 jchar lastChar = pat->char_at(o + (c - 1)); 1521 int cache = 0; 1522 int i; 1523 for (i = 0; i < c - 1; i++) { 1524 assert(i < pat->length(), "out of range"); 1525 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1526 } 1527 1528 int md2 = c; 1529 for (i = 0; i < c - 1; i++) { 1530 assert(i < pat->length(), "out of range"); 1531 if (pat->char_at(o + i) == lastChar) { 1532 md2 = (c - 1) - i; 1533 } 1534 } 1535 1536 result = string_indexOf(receiver, pat, o, cache, md2); 1537 } 1538 set_result(result); 1539 return true; 1540 } 1541 1542 //--------------------------round_double_node-------------------------------- 1543 // Round a double node if necessary. 1544 Node* LibraryCallKit::round_double_node(Node* n) { 1545 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1546 n = _gvn.transform(new RoundDoubleNode(0, n)); 1547 return n; 1548 } 1549 1550 //------------------------------inline_math----------------------------------- 1551 // public static double Math.abs(double) 1552 // public static double Math.sqrt(double) 1553 // public static double Math.log(double) 1554 // public static double Math.log10(double) 1555 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1556 Node* arg = round_double_node(argument(0)); 1557 Node* n; 1558 switch (id) { 1559 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1560 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1561 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1562 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1563 default: fatal_unexpected_iid(id); break; 1564 } 1565 set_result(_gvn.transform(n)); 1566 return true; 1567 } 1568 1569 //------------------------------inline_trig---------------------------------- 1570 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1571 // argument reduction which will turn into a fast/slow diamond. 1572 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1573 Node* arg = round_double_node(argument(0)); 1574 Node* n = NULL; 1575 1576 switch (id) { 1577 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1578 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1579 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1580 default: fatal_unexpected_iid(id); break; 1581 } 1582 n = _gvn.transform(n); 1583 1584 // Rounding required? Check for argument reduction! 1585 if (Matcher::strict_fp_requires_explicit_rounding) { 1586 static const double pi_4 = 0.7853981633974483; 1587 static const double neg_pi_4 = -0.7853981633974483; 1588 // pi/2 in 80-bit extended precision 1589 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1590 // -pi/2 in 80-bit extended precision 1591 // 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}; 1592 // Cutoff value for using this argument reduction technique 1593 //static const double pi_2_minus_epsilon = 1.564660403643354; 1594 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1595 1596 // Pseudocode for sin: 1597 // if (x <= Math.PI / 4.0) { 1598 // if (x >= -Math.PI / 4.0) return fsin(x); 1599 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1600 // } else { 1601 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1602 // } 1603 // return StrictMath.sin(x); 1604 1605 // Pseudocode for cos: 1606 // if (x <= Math.PI / 4.0) { 1607 // if (x >= -Math.PI / 4.0) return fcos(x); 1608 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1609 // } else { 1610 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1611 // } 1612 // return StrictMath.cos(x); 1613 1614 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1615 // requires a special machine instruction to load it. Instead we'll try 1616 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1617 // probably do the math inside the SIN encoding. 1618 1619 // Make the merge point 1620 RegionNode* r = new RegionNode(3); 1621 Node* phi = new PhiNode(r, Type::DOUBLE); 1622 1623 // Flatten arg so we need only 1 test 1624 Node *abs = _gvn.transform(new AbsDNode(arg)); 1625 // Node for PI/4 constant 1626 Node *pi4 = makecon(TypeD::make(pi_4)); 1627 // Check PI/4 : abs(arg) 1628 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1629 // Check: If PI/4 < abs(arg) then go slow 1630 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1631 // Branch either way 1632 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1633 set_control(opt_iff(r,iff)); 1634 1635 // Set fast path result 1636 phi->init_req(2, n); 1637 1638 // Slow path - non-blocking leaf call 1639 Node* call = NULL; 1640 switch (id) { 1641 case vmIntrinsics::_dsin: 1642 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1643 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1644 "Sin", NULL, arg, top()); 1645 break; 1646 case vmIntrinsics::_dcos: 1647 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1648 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1649 "Cos", NULL, arg, top()); 1650 break; 1651 case vmIntrinsics::_dtan: 1652 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1653 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1654 "Tan", NULL, arg, top()); 1655 break; 1656 } 1657 assert(control()->in(0) == call, ""); 1658 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1659 r->init_req(1, control()); 1660 phi->init_req(1, slow_result); 1661 1662 // Post-merge 1663 set_control(_gvn.transform(r)); 1664 record_for_igvn(r); 1665 n = _gvn.transform(phi); 1666 1667 C->set_has_split_ifs(true); // Has chance for split-if optimization 1668 } 1669 set_result(n); 1670 return true; 1671 } 1672 1673 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1674 //------------------- 1675 //result=(result.isNaN())? funcAddr():result; 1676 // Check: If isNaN() by checking result!=result? then either trap 1677 // or go to runtime 1678 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1679 // Build the boolean node 1680 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1681 1682 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1683 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1684 // The pow or exp intrinsic returned a NaN, which requires a call 1685 // to the runtime. Recompile with the runtime call. 1686 uncommon_trap(Deoptimization::Reason_intrinsic, 1687 Deoptimization::Action_make_not_entrant); 1688 } 1689 return result; 1690 } else { 1691 // If this inlining ever returned NaN in the past, we compile a call 1692 // to the runtime to properly handle corner cases 1693 1694 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1695 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1696 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1697 1698 if (!if_slow->is_top()) { 1699 RegionNode* result_region = new RegionNode(3); 1700 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1701 1702 result_region->init_req(1, if_fast); 1703 result_val->init_req(1, result); 1704 1705 set_control(if_slow); 1706 1707 const TypePtr* no_memory_effects = NULL; 1708 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1709 no_memory_effects, 1710 x, top(), y, y ? top() : NULL); 1711 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1712 #ifdef ASSERT 1713 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1714 assert(value_top == top(), "second value must be top"); 1715 #endif 1716 1717 result_region->init_req(2, control()); 1718 result_val->init_req(2, value); 1719 set_control(_gvn.transform(result_region)); 1720 return _gvn.transform(result_val); 1721 } else { 1722 return result; 1723 } 1724 } 1725 } 1726 1727 //------------------------------inline_exp------------------------------------- 1728 // Inline exp instructions, if possible. The Intel hardware only misses 1729 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1730 bool LibraryCallKit::inline_exp() { 1731 Node* arg = round_double_node(argument(0)); 1732 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1733 1734 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1735 set_result(n); 1736 1737 C->set_has_split_ifs(true); // Has chance for split-if optimization 1738 return true; 1739 } 1740 1741 //------------------------------inline_pow------------------------------------- 1742 // Inline power instructions, if possible. 1743 bool LibraryCallKit::inline_pow() { 1744 // Pseudocode for pow 1745 // if (y == 2) { 1746 // return x * x; 1747 // } else { 1748 // if (x <= 0.0) { 1749 // long longy = (long)y; 1750 // if ((double)longy == y) { // if y is long 1751 // if (y + 1 == y) longy = 0; // huge number: even 1752 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1753 // } else { 1754 // result = NaN; 1755 // } 1756 // } else { 1757 // result = DPow(x,y); 1758 // } 1759 // if (result != result)? { 1760 // result = uncommon_trap() or runtime_call(); 1761 // } 1762 // return result; 1763 // } 1764 1765 Node* x = round_double_node(argument(0)); 1766 Node* y = round_double_node(argument(2)); 1767 1768 Node* result = NULL; 1769 1770 Node* const_two_node = makecon(TypeD::make(2.0)); 1771 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1772 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1773 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1774 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1775 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1776 1777 RegionNode* region_node = new RegionNode(3); 1778 region_node->init_req(1, if_true); 1779 1780 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1781 // special case for x^y where y == 2, we can convert it to x * x 1782 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1783 1784 // set control to if_false since we will now process the false branch 1785 set_control(if_false); 1786 1787 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1788 // Short form: skip the fancy tests and just check for NaN result. 1789 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1790 } else { 1791 // If this inlining ever returned NaN in the past, include all 1792 // checks + call to the runtime. 1793 1794 // Set the merge point for If node with condition of (x <= 0.0) 1795 // There are four possible paths to region node and phi node 1796 RegionNode *r = new RegionNode(4); 1797 Node *phi = new PhiNode(r, Type::DOUBLE); 1798 1799 // Build the first if node: if (x <= 0.0) 1800 // Node for 0 constant 1801 Node *zeronode = makecon(TypeD::ZERO); 1802 // Check x:0 1803 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1804 // Check: If (x<=0) then go complex path 1805 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1806 // Branch either way 1807 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1808 // Fast path taken; set region slot 3 1809 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1810 r->init_req(3,fast_taken); // Capture fast-control 1811 1812 // Fast path not-taken, i.e. slow path 1813 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1814 1815 // Set fast path result 1816 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1817 phi->init_req(3, fast_result); 1818 1819 // Complex path 1820 // Build the second if node (if y is long) 1821 // Node for (long)y 1822 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1823 // Node for (double)((long) y) 1824 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1825 // Check (double)((long) y) : y 1826 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1827 // Check if (y isn't long) then go to slow path 1828 1829 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1830 // Branch either way 1831 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1832 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1833 1834 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1835 1836 // Calculate DPow(abs(x), y)*(1 & (long)y) 1837 // Node for constant 1 1838 Node *conone = longcon(1); 1839 // 1& (long)y 1840 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1841 1842 // A huge number is always even. Detect a huge number by checking 1843 // if y + 1 == y and set integer to be tested for parity to 0. 1844 // Required for corner case: 1845 // (long)9.223372036854776E18 = max_jlong 1846 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1847 // max_jlong is odd but 9.223372036854776E18 is even 1848 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1849 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1850 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1851 Node* correctedsign = NULL; 1852 if (ConditionalMoveLimit != 0) { 1853 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1854 } else { 1855 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1856 RegionNode *r = new RegionNode(3); 1857 Node *phi = new PhiNode(r, TypeLong::LONG); 1858 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1859 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1860 phi->init_req(1, signnode); 1861 phi->init_req(2, longcon(0)); 1862 correctedsign = _gvn.transform(phi); 1863 ylong_path = _gvn.transform(r); 1864 record_for_igvn(r); 1865 } 1866 1867 // zero node 1868 Node *conzero = longcon(0); 1869 // Check (1&(long)y)==0? 1870 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1871 // Check if (1&(long)y)!=0?, if so the result is negative 1872 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1873 // abs(x) 1874 Node *absx=_gvn.transform(new AbsDNode(x)); 1875 // abs(x)^y 1876 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1877 // -abs(x)^y 1878 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1879 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1880 Node *signresult = NULL; 1881 if (ConditionalMoveLimit != 0) { 1882 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1883 } else { 1884 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1885 RegionNode *r = new RegionNode(3); 1886 Node *phi = new PhiNode(r, Type::DOUBLE); 1887 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1888 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1889 phi->init_req(1, absxpowy); 1890 phi->init_req(2, negabsxpowy); 1891 signresult = _gvn.transform(phi); 1892 ylong_path = _gvn.transform(r); 1893 record_for_igvn(r); 1894 } 1895 // Set complex path fast result 1896 r->init_req(2, ylong_path); 1897 phi->init_req(2, signresult); 1898 1899 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1900 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1901 r->init_req(1,slow_path); 1902 phi->init_req(1,slow_result); 1903 1904 // Post merge 1905 set_control(_gvn.transform(r)); 1906 record_for_igvn(r); 1907 result = _gvn.transform(phi); 1908 } 1909 1910 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1911 1912 // control from finish_pow_exp is now input to the region node 1913 region_node->set_req(2, control()); 1914 // the result from finish_pow_exp is now input to the phi node 1915 phi_node->init_req(2, result); 1916 set_control(_gvn.transform(region_node)); 1917 record_for_igvn(region_node); 1918 set_result(_gvn.transform(phi_node)); 1919 1920 C->set_has_split_ifs(true); // Has chance for split-if optimization 1921 return true; 1922 } 1923 1924 //------------------------------runtime_math----------------------------- 1925 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1926 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1927 "must be (DD)D or (D)D type"); 1928 1929 // Inputs 1930 Node* a = round_double_node(argument(0)); 1931 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1932 1933 const TypePtr* no_memory_effects = NULL; 1934 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1935 no_memory_effects, 1936 a, top(), b, b ? top() : NULL); 1937 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1938 #ifdef ASSERT 1939 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1940 assert(value_top == top(), "second value must be top"); 1941 #endif 1942 1943 set_result(value); 1944 return true; 1945 } 1946 1947 //------------------------------inline_math_native----------------------------- 1948 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1949 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1950 switch (id) { 1951 // These intrinsics are not properly supported on all hardware 1952 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1953 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1954 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1955 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1956 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1957 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1958 1959 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1960 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1961 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1962 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1963 1964 // These intrinsics are supported on all hardware 1965 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1966 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1967 1968 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1969 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1970 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1971 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1972 #undef FN_PTR 1973 1974 // These intrinsics are not yet correctly implemented 1975 case vmIntrinsics::_datan2: 1976 return false; 1977 1978 default: 1979 fatal_unexpected_iid(id); 1980 return false; 1981 } 1982 } 1983 1984 static bool is_simple_name(Node* n) { 1985 return (n->req() == 1 // constant 1986 || (n->is_Type() && n->as_Type()->type()->singleton()) 1987 || n->is_Proj() // parameter or return value 1988 || n->is_Phi() // local of some sort 1989 ); 1990 } 1991 1992 //----------------------------inline_min_max----------------------------------- 1993 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1994 set_result(generate_min_max(id, argument(0), argument(1))); 1995 return true; 1996 } 1997 1998 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1999 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2000 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2001 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2002 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2003 2004 { 2005 PreserveJVMState pjvms(this); 2006 PreserveReexecuteState preexecs(this); 2007 jvms()->set_should_reexecute(true); 2008 2009 set_control(slow_path); 2010 set_i_o(i_o()); 2011 2012 uncommon_trap(Deoptimization::Reason_intrinsic, 2013 Deoptimization::Action_none); 2014 } 2015 2016 set_control(fast_path); 2017 set_result(math); 2018 } 2019 2020 template <typename OverflowOp> 2021 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2022 typedef typename OverflowOp::MathOp MathOp; 2023 2024 MathOp* mathOp = new MathOp(arg1, arg2); 2025 Node* operation = _gvn.transform( mathOp ); 2026 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2027 inline_math_mathExact(operation, ofcheck); 2028 return true; 2029 } 2030 2031 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2032 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2033 } 2034 2035 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2036 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2037 } 2038 2039 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2040 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2041 } 2042 2043 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2044 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2045 } 2046 2047 bool LibraryCallKit::inline_math_negateExactI() { 2048 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2049 } 2050 2051 bool LibraryCallKit::inline_math_negateExactL() { 2052 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2053 } 2054 2055 bool LibraryCallKit::inline_math_multiplyExactI() { 2056 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2057 } 2058 2059 bool LibraryCallKit::inline_math_multiplyExactL() { 2060 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2061 } 2062 2063 Node* 2064 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2065 // These are the candidate return value: 2066 Node* xvalue = x0; 2067 Node* yvalue = y0; 2068 2069 if (xvalue == yvalue) { 2070 return xvalue; 2071 } 2072 2073 bool want_max = (id == vmIntrinsics::_max); 2074 2075 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2076 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2077 if (txvalue == NULL || tyvalue == NULL) return top(); 2078 // This is not really necessary, but it is consistent with a 2079 // hypothetical MaxINode::Value method: 2080 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2081 2082 // %%% This folding logic should (ideally) be in a different place. 2083 // Some should be inside IfNode, and there to be a more reliable 2084 // transformation of ?: style patterns into cmoves. We also want 2085 // more powerful optimizations around cmove and min/max. 2086 2087 // Try to find a dominating comparison of these guys. 2088 // It can simplify the index computation for Arrays.copyOf 2089 // and similar uses of System.arraycopy. 2090 // First, compute the normalized version of CmpI(x, y). 2091 int cmp_op = Op_CmpI; 2092 Node* xkey = xvalue; 2093 Node* ykey = yvalue; 2094 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2095 if (ideal_cmpxy->is_Cmp()) { 2096 // E.g., if we have CmpI(length - offset, count), 2097 // it might idealize to CmpI(length, count + offset) 2098 cmp_op = ideal_cmpxy->Opcode(); 2099 xkey = ideal_cmpxy->in(1); 2100 ykey = ideal_cmpxy->in(2); 2101 } 2102 2103 // Start by locating any relevant comparisons. 2104 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2105 Node* cmpxy = NULL; 2106 Node* cmpyx = NULL; 2107 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2108 Node* cmp = start_from->fast_out(k); 2109 if (cmp->outcnt() > 0 && // must have prior uses 2110 cmp->in(0) == NULL && // must be context-independent 2111 cmp->Opcode() == cmp_op) { // right kind of compare 2112 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2113 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2114 } 2115 } 2116 2117 const int NCMPS = 2; 2118 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2119 int cmpn; 2120 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2121 if (cmps[cmpn] != NULL) break; // find a result 2122 } 2123 if (cmpn < NCMPS) { 2124 // Look for a dominating test that tells us the min and max. 2125 int depth = 0; // Limit search depth for speed 2126 Node* dom = control(); 2127 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2128 if (++depth >= 100) break; 2129 Node* ifproj = dom; 2130 if (!ifproj->is_Proj()) continue; 2131 Node* iff = ifproj->in(0); 2132 if (!iff->is_If()) continue; 2133 Node* bol = iff->in(1); 2134 if (!bol->is_Bool()) continue; 2135 Node* cmp = bol->in(1); 2136 if (cmp == NULL) continue; 2137 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2138 if (cmps[cmpn] == cmp) break; 2139 if (cmpn == NCMPS) continue; 2140 BoolTest::mask btest = bol->as_Bool()->_test._test; 2141 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2142 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2143 // At this point, we know that 'x btest y' is true. 2144 switch (btest) { 2145 case BoolTest::eq: 2146 // They are proven equal, so we can collapse the min/max. 2147 // Either value is the answer. Choose the simpler. 2148 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2149 return yvalue; 2150 return xvalue; 2151 case BoolTest::lt: // x < y 2152 case BoolTest::le: // x <= y 2153 return (want_max ? yvalue : xvalue); 2154 case BoolTest::gt: // x > y 2155 case BoolTest::ge: // x >= y 2156 return (want_max ? xvalue : yvalue); 2157 } 2158 } 2159 } 2160 2161 // We failed to find a dominating test. 2162 // Let's pick a test that might GVN with prior tests. 2163 Node* best_bol = NULL; 2164 BoolTest::mask best_btest = BoolTest::illegal; 2165 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2166 Node* cmp = cmps[cmpn]; 2167 if (cmp == NULL) continue; 2168 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2169 Node* bol = cmp->fast_out(j); 2170 if (!bol->is_Bool()) continue; 2171 BoolTest::mask btest = bol->as_Bool()->_test._test; 2172 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2173 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2174 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2175 best_bol = bol->as_Bool(); 2176 best_btest = btest; 2177 } 2178 } 2179 } 2180 2181 Node* answer_if_true = NULL; 2182 Node* answer_if_false = NULL; 2183 switch (best_btest) { 2184 default: 2185 if (cmpxy == NULL) 2186 cmpxy = ideal_cmpxy; 2187 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2188 // and fall through: 2189 case BoolTest::lt: // x < y 2190 case BoolTest::le: // x <= y 2191 answer_if_true = (want_max ? yvalue : xvalue); 2192 answer_if_false = (want_max ? xvalue : yvalue); 2193 break; 2194 case BoolTest::gt: // x > y 2195 case BoolTest::ge: // x >= y 2196 answer_if_true = (want_max ? xvalue : yvalue); 2197 answer_if_false = (want_max ? yvalue : xvalue); 2198 break; 2199 } 2200 2201 jint hi, lo; 2202 if (want_max) { 2203 // We can sharpen the minimum. 2204 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2205 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2206 } else { 2207 // We can sharpen the maximum. 2208 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2209 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2210 } 2211 2212 // Use a flow-free graph structure, to avoid creating excess control edges 2213 // which could hinder other optimizations. 2214 // Since Math.min/max is often used with arraycopy, we want 2215 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2216 Node* cmov = CMoveNode::make(NULL, best_bol, 2217 answer_if_false, answer_if_true, 2218 TypeInt::make(lo, hi, widen)); 2219 2220 return _gvn.transform(cmov); 2221 2222 /* 2223 // This is not as desirable as it may seem, since Min and Max 2224 // nodes do not have a full set of optimizations. 2225 // And they would interfere, anyway, with 'if' optimizations 2226 // and with CMoveI canonical forms. 2227 switch (id) { 2228 case vmIntrinsics::_min: 2229 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2230 case vmIntrinsics::_max: 2231 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2232 default: 2233 ShouldNotReachHere(); 2234 } 2235 */ 2236 } 2237 2238 inline int 2239 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2240 const TypePtr* base_type = TypePtr::NULL_PTR; 2241 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2242 if (base_type == NULL) { 2243 // Unknown type. 2244 return Type::AnyPtr; 2245 } else if (base_type == TypePtr::NULL_PTR) { 2246 // Since this is a NULL+long form, we have to switch to a rawptr. 2247 base = _gvn.transform(new CastX2PNode(offset)); 2248 offset = MakeConX(0); 2249 return Type::RawPtr; 2250 } else if (base_type->base() == Type::RawPtr) { 2251 return Type::RawPtr; 2252 } else if (base_type->isa_oopptr()) { 2253 // Base is never null => always a heap address. 2254 if (base_type->ptr() == TypePtr::NotNull) { 2255 return Type::OopPtr; 2256 } 2257 // Offset is small => always a heap address. 2258 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2259 if (offset_type != NULL && 2260 base_type->offset() == 0 && // (should always be?) 2261 offset_type->_lo >= 0 && 2262 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2263 return Type::OopPtr; 2264 } 2265 // Otherwise, it might either be oop+off or NULL+addr. 2266 return Type::AnyPtr; 2267 } else { 2268 // No information: 2269 return Type::AnyPtr; 2270 } 2271 } 2272 2273 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2274 int kind = classify_unsafe_addr(base, offset); 2275 if (kind == Type::RawPtr) { 2276 return basic_plus_adr(top(), base, offset); 2277 } else { 2278 return basic_plus_adr(base, offset); 2279 } 2280 } 2281 2282 //--------------------------inline_number_methods----------------------------- 2283 // inline int Integer.numberOfLeadingZeros(int) 2284 // inline int Long.numberOfLeadingZeros(long) 2285 // 2286 // inline int Integer.numberOfTrailingZeros(int) 2287 // inline int Long.numberOfTrailingZeros(long) 2288 // 2289 // inline int Integer.bitCount(int) 2290 // inline int Long.bitCount(long) 2291 // 2292 // inline char Character.reverseBytes(char) 2293 // inline short Short.reverseBytes(short) 2294 // inline int Integer.reverseBytes(int) 2295 // inline long Long.reverseBytes(long) 2296 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2297 Node* arg = argument(0); 2298 Node* n; 2299 switch (id) { 2300 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2301 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2302 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2303 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2304 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2305 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2306 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2307 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2308 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2309 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2310 default: fatal_unexpected_iid(id); break; 2311 } 2312 set_result(_gvn.transform(n)); 2313 return true; 2314 } 2315 2316 //----------------------------inline_unsafe_access---------------------------- 2317 2318 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2319 2320 // Helper that guards and inserts a pre-barrier. 2321 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2322 Node* pre_val, bool need_mem_bar) { 2323 // We could be accessing the referent field of a reference object. If so, when G1 2324 // is enabled, we need to log the value in the referent field in an SATB buffer. 2325 // This routine performs some compile time filters and generates suitable 2326 // runtime filters that guard the pre-barrier code. 2327 // Also add memory barrier for non volatile load from the referent field 2328 // to prevent commoning of loads across safepoint. 2329 if (!UseG1GC && !need_mem_bar) 2330 return; 2331 2332 // Some compile time checks. 2333 2334 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2335 const TypeX* otype = offset->find_intptr_t_type(); 2336 if (otype != NULL && otype->is_con() && 2337 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2338 // Constant offset but not the reference_offset so just return 2339 return; 2340 } 2341 2342 // We only need to generate the runtime guards for instances. 2343 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2344 if (btype != NULL) { 2345 if (btype->isa_aryptr()) { 2346 // Array type so nothing to do 2347 return; 2348 } 2349 2350 const TypeInstPtr* itype = btype->isa_instptr(); 2351 if (itype != NULL) { 2352 // Can the klass of base_oop be statically determined to be 2353 // _not_ a sub-class of Reference and _not_ Object? 2354 ciKlass* klass = itype->klass(); 2355 if ( klass->is_loaded() && 2356 !klass->is_subtype_of(env()->Reference_klass()) && 2357 !env()->Object_klass()->is_subtype_of(klass)) { 2358 return; 2359 } 2360 } 2361 } 2362 2363 // The compile time filters did not reject base_oop/offset so 2364 // we need to generate the following runtime filters 2365 // 2366 // if (offset == java_lang_ref_Reference::_reference_offset) { 2367 // if (instance_of(base, java.lang.ref.Reference)) { 2368 // pre_barrier(_, pre_val, ...); 2369 // } 2370 // } 2371 2372 float likely = PROB_LIKELY( 0.999); 2373 float unlikely = PROB_UNLIKELY(0.999); 2374 2375 IdealKit ideal(this); 2376 #define __ ideal. 2377 2378 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2379 2380 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2381 // Update graphKit memory and control from IdealKit. 2382 sync_kit(ideal); 2383 2384 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2385 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2386 2387 // Update IdealKit memory and control from graphKit. 2388 __ sync_kit(this); 2389 2390 Node* one = __ ConI(1); 2391 // is_instof == 0 if base_oop == NULL 2392 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2393 2394 // Update graphKit from IdeakKit. 2395 sync_kit(ideal); 2396 2397 // Use the pre-barrier to record the value in the referent field 2398 pre_barrier(false /* do_load */, 2399 __ ctrl(), 2400 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2401 pre_val /* pre_val */, 2402 T_OBJECT); 2403 if (need_mem_bar) { 2404 // Add memory barrier to prevent commoning reads from this field 2405 // across safepoint since GC can change its value. 2406 insert_mem_bar(Op_MemBarCPUOrder); 2407 } 2408 // Update IdealKit from graphKit. 2409 __ sync_kit(this); 2410 2411 } __ end_if(); // _ref_type != ref_none 2412 } __ end_if(); // offset == referent_offset 2413 2414 // Final sync IdealKit and GraphKit. 2415 final_sync(ideal); 2416 #undef __ 2417 } 2418 2419 2420 // Interpret Unsafe.fieldOffset cookies correctly: 2421 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2422 2423 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2424 // Attempt to infer a sharper value type from the offset and base type. 2425 ciKlass* sharpened_klass = NULL; 2426 2427 // See if it is an instance field, with an object type. 2428 if (alias_type->field() != NULL) { 2429 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2430 if (alias_type->field()->type()->is_klass()) { 2431 sharpened_klass = alias_type->field()->type()->as_klass(); 2432 } 2433 } 2434 2435 // See if it is a narrow oop array. 2436 if (adr_type->isa_aryptr()) { 2437 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2438 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2439 if (elem_type != NULL) { 2440 sharpened_klass = elem_type->klass(); 2441 } 2442 } 2443 } 2444 2445 // The sharpened class might be unloaded if there is no class loader 2446 // contraint in place. 2447 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2448 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2449 2450 #ifndef PRODUCT 2451 if (C->print_intrinsics() || C->print_inlining()) { 2452 tty->print(" from base type: "); adr_type->dump(); 2453 tty->print(" sharpened value: "); tjp->dump(); 2454 } 2455 #endif 2456 // Sharpen the value type. 2457 return tjp; 2458 } 2459 return NULL; 2460 } 2461 2462 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2463 if (callee()->is_static()) return false; // caller must have the capability! 2464 2465 #ifndef PRODUCT 2466 { 2467 ResourceMark rm; 2468 // Check the signatures. 2469 ciSignature* sig = callee()->signature(); 2470 #ifdef ASSERT 2471 if (!is_store) { 2472 // Object getObject(Object base, int/long offset), etc. 2473 BasicType rtype = sig->return_type()->basic_type(); 2474 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2475 rtype = T_ADDRESS; // it is really a C void* 2476 assert(rtype == type, "getter must return the expected value"); 2477 if (!is_native_ptr) { 2478 assert(sig->count() == 2, "oop getter has 2 arguments"); 2479 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2480 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2481 } else { 2482 assert(sig->count() == 1, "native getter has 1 argument"); 2483 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2484 } 2485 } else { 2486 // void putObject(Object base, int/long offset, Object x), etc. 2487 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2488 if (!is_native_ptr) { 2489 assert(sig->count() == 3, "oop putter has 3 arguments"); 2490 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2491 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2492 } else { 2493 assert(sig->count() == 2, "native putter has 2 arguments"); 2494 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2495 } 2496 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2497 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2498 vtype = T_ADDRESS; // it is really a C void* 2499 assert(vtype == type, "putter must accept the expected value"); 2500 } 2501 #endif // ASSERT 2502 } 2503 #endif //PRODUCT 2504 2505 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2506 2507 Node* receiver = argument(0); // type: oop 2508 2509 // Build address expression. See the code in inline_unsafe_prefetch. 2510 Node* adr; 2511 Node* heap_base_oop = top(); 2512 Node* offset = top(); 2513 Node* val; 2514 2515 if (!is_native_ptr) { 2516 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2517 Node* base = argument(1); // type: oop 2518 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2519 offset = argument(2); // type: long 2520 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2521 // to be plain byte offsets, which are also the same as those accepted 2522 // by oopDesc::field_base. 2523 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2524 "fieldOffset must be byte-scaled"); 2525 // 32-bit machines ignore the high half! 2526 offset = ConvL2X(offset); 2527 adr = make_unsafe_address(base, offset); 2528 heap_base_oop = base; 2529 val = is_store ? argument(4) : NULL; 2530 } else { 2531 Node* ptr = argument(1); // type: long 2532 ptr = ConvL2X(ptr); // adjust Java long to machine word 2533 adr = make_unsafe_address(NULL, ptr); 2534 val = is_store ? argument(3) : NULL; 2535 } 2536 2537 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2538 2539 // First guess at the value type. 2540 const Type *value_type = Type::get_const_basic_type(type); 2541 2542 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2543 // there was not enough information to nail it down. 2544 Compile::AliasType* alias_type = C->alias_type(adr_type); 2545 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2546 2547 // We will need memory barriers unless we can determine a unique 2548 // alias category for this reference. (Note: If for some reason 2549 // the barriers get omitted and the unsafe reference begins to "pollute" 2550 // the alias analysis of the rest of the graph, either Compile::can_alias 2551 // or Compile::must_alias will throw a diagnostic assert.) 2552 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2553 2554 // If we are reading the value of the referent field of a Reference 2555 // object (either by using Unsafe directly or through reflection) 2556 // then, if G1 is enabled, we need to record the referent in an 2557 // SATB log buffer using the pre-barrier mechanism. 2558 // Also we need to add memory barrier to prevent commoning reads 2559 // from this field across safepoint since GC can change its value. 2560 bool need_read_barrier = !is_native_ptr && !is_store && 2561 offset != top() && heap_base_oop != top(); 2562 2563 if (!is_store && type == T_OBJECT) { 2564 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2565 if (tjp != NULL) { 2566 value_type = tjp; 2567 } 2568 } 2569 2570 receiver = null_check(receiver); 2571 if (stopped()) { 2572 return true; 2573 } 2574 // Heap pointers get a null-check from the interpreter, 2575 // as a courtesy. However, this is not guaranteed by Unsafe, 2576 // and it is not possible to fully distinguish unintended nulls 2577 // from intended ones in this API. 2578 2579 if (is_volatile) { 2580 // We need to emit leading and trailing CPU membars (see below) in 2581 // addition to memory membars when is_volatile. This is a little 2582 // too strong, but avoids the need to insert per-alias-type 2583 // volatile membars (for stores; compare Parse::do_put_xxx), which 2584 // we cannot do effectively here because we probably only have a 2585 // rough approximation of type. 2586 need_mem_bar = true; 2587 // For Stores, place a memory ordering barrier now. 2588 if (is_store) { 2589 insert_mem_bar(Op_MemBarRelease); 2590 } else { 2591 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2592 insert_mem_bar(Op_MemBarVolatile); 2593 } 2594 } 2595 } 2596 2597 // Memory barrier to prevent normal and 'unsafe' accesses from 2598 // bypassing each other. Happens after null checks, so the 2599 // exception paths do not take memory state from the memory barrier, 2600 // so there's no problems making a strong assert about mixing users 2601 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2602 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2603 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2604 2605 if (!is_store) { 2606 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2607 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile); 2608 // load value 2609 switch (type) { 2610 case T_BOOLEAN: 2611 case T_CHAR: 2612 case T_BYTE: 2613 case T_SHORT: 2614 case T_INT: 2615 case T_LONG: 2616 case T_FLOAT: 2617 case T_DOUBLE: 2618 break; 2619 case T_OBJECT: 2620 if (need_read_barrier) { 2621 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2622 } 2623 break; 2624 case T_ADDRESS: 2625 // Cast to an int type. 2626 p = _gvn.transform(new CastP2XNode(NULL, p)); 2627 p = ConvX2UL(p); 2628 break; 2629 default: 2630 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2631 break; 2632 } 2633 // The load node has the control of the preceding MemBarCPUOrder. All 2634 // following nodes will have the control of the MemBarCPUOrder inserted at 2635 // the end of this method. So, pushing the load onto the stack at a later 2636 // point is fine. 2637 set_result(p); 2638 } else { 2639 // place effect of store into memory 2640 switch (type) { 2641 case T_DOUBLE: 2642 val = dstore_rounding(val); 2643 break; 2644 case T_ADDRESS: 2645 // Repackage the long as a pointer. 2646 val = ConvL2X(val); 2647 val = _gvn.transform(new CastX2PNode(val)); 2648 break; 2649 } 2650 2651 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2652 if (type != T_OBJECT ) { 2653 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2654 } else { 2655 // Possibly an oop being stored to Java heap or native memory 2656 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2657 // oop to Java heap. 2658 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2659 } else { 2660 // We can't tell at compile time if we are storing in the Java heap or outside 2661 // of it. So we need to emit code to conditionally do the proper type of 2662 // store. 2663 2664 IdealKit ideal(this); 2665 #define __ ideal. 2666 // QQQ who knows what probability is here?? 2667 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2668 // Sync IdealKit and graphKit. 2669 sync_kit(ideal); 2670 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2671 // Update IdealKit memory. 2672 __ sync_kit(this); 2673 } __ else_(); { 2674 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2675 } __ end_if(); 2676 // Final sync IdealKit and GraphKit. 2677 final_sync(ideal); 2678 #undef __ 2679 } 2680 } 2681 } 2682 2683 if (is_volatile) { 2684 if (!is_store) { 2685 insert_mem_bar(Op_MemBarAcquire); 2686 } else { 2687 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2688 insert_mem_bar(Op_MemBarVolatile); 2689 } 2690 } 2691 } 2692 2693 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2694 2695 return true; 2696 } 2697 2698 //----------------------------inline_unsafe_prefetch---------------------------- 2699 2700 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2701 #ifndef PRODUCT 2702 { 2703 ResourceMark rm; 2704 // Check the signatures. 2705 ciSignature* sig = callee()->signature(); 2706 #ifdef ASSERT 2707 // Object getObject(Object base, int/long offset), etc. 2708 BasicType rtype = sig->return_type()->basic_type(); 2709 if (!is_native_ptr) { 2710 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2711 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2712 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2713 } else { 2714 assert(sig->count() == 1, "native prefetch has 1 argument"); 2715 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2716 } 2717 #endif // ASSERT 2718 } 2719 #endif // !PRODUCT 2720 2721 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2722 2723 const int idx = is_static ? 0 : 1; 2724 if (!is_static) { 2725 null_check_receiver(); 2726 if (stopped()) { 2727 return true; 2728 } 2729 } 2730 2731 // Build address expression. See the code in inline_unsafe_access. 2732 Node *adr; 2733 if (!is_native_ptr) { 2734 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2735 Node* base = argument(idx + 0); // type: oop 2736 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2737 Node* offset = argument(idx + 1); // type: long 2738 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2739 // to be plain byte offsets, which are also the same as those accepted 2740 // by oopDesc::field_base. 2741 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2742 "fieldOffset must be byte-scaled"); 2743 // 32-bit machines ignore the high half! 2744 offset = ConvL2X(offset); 2745 adr = make_unsafe_address(base, offset); 2746 } else { 2747 Node* ptr = argument(idx + 0); // type: long 2748 ptr = ConvL2X(ptr); // adjust Java long to machine word 2749 adr = make_unsafe_address(NULL, ptr); 2750 } 2751 2752 // Generate the read or write prefetch 2753 Node *prefetch; 2754 if (is_store) { 2755 prefetch = new PrefetchWriteNode(i_o(), adr); 2756 } else { 2757 prefetch = new PrefetchReadNode(i_o(), adr); 2758 } 2759 prefetch->init_req(0, control()); 2760 set_i_o(_gvn.transform(prefetch)); 2761 2762 return true; 2763 } 2764 2765 //----------------------------inline_unsafe_load_store---------------------------- 2766 // This method serves a couple of different customers (depending on LoadStoreKind): 2767 // 2768 // LS_cmpxchg: 2769 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2770 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2771 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2772 // 2773 // LS_xadd: 2774 // public int getAndAddInt( Object o, long offset, int delta) 2775 // public long getAndAddLong(Object o, long offset, long delta) 2776 // 2777 // LS_xchg: 2778 // int getAndSet(Object o, long offset, int newValue) 2779 // long getAndSet(Object o, long offset, long newValue) 2780 // Object getAndSet(Object o, long offset, Object newValue) 2781 // 2782 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2783 // This basic scheme here is the same as inline_unsafe_access, but 2784 // differs in enough details that combining them would make the code 2785 // overly confusing. (This is a true fact! I originally combined 2786 // them, but even I was confused by it!) As much code/comments as 2787 // possible are retained from inline_unsafe_access though to make 2788 // the correspondences clearer. - dl 2789 2790 if (callee()->is_static()) return false; // caller must have the capability! 2791 2792 #ifndef PRODUCT 2793 BasicType rtype; 2794 { 2795 ResourceMark rm; 2796 // Check the signatures. 2797 ciSignature* sig = callee()->signature(); 2798 rtype = sig->return_type()->basic_type(); 2799 if (kind == LS_xadd || kind == LS_xchg) { 2800 // Check the signatures. 2801 #ifdef ASSERT 2802 assert(rtype == type, "get and set must return the expected type"); 2803 assert(sig->count() == 3, "get and set has 3 arguments"); 2804 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2805 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2806 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2807 #endif // ASSERT 2808 } else if (kind == LS_cmpxchg) { 2809 // Check the signatures. 2810 #ifdef ASSERT 2811 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2812 assert(sig->count() == 4, "CAS has 4 arguments"); 2813 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2814 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2815 #endif // ASSERT 2816 } else { 2817 ShouldNotReachHere(); 2818 } 2819 } 2820 #endif //PRODUCT 2821 2822 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2823 2824 // Get arguments: 2825 Node* receiver = NULL; 2826 Node* base = NULL; 2827 Node* offset = NULL; 2828 Node* oldval = NULL; 2829 Node* newval = NULL; 2830 if (kind == LS_cmpxchg) { 2831 const bool two_slot_type = type2size[type] == 2; 2832 receiver = argument(0); // type: oop 2833 base = argument(1); // type: oop 2834 offset = argument(2); // type: long 2835 oldval = argument(4); // type: oop, int, or long 2836 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2837 } else if (kind == LS_xadd || kind == LS_xchg){ 2838 receiver = argument(0); // type: oop 2839 base = argument(1); // type: oop 2840 offset = argument(2); // type: long 2841 oldval = NULL; 2842 newval = argument(4); // type: oop, int, or long 2843 } 2844 2845 // Null check receiver. 2846 receiver = null_check(receiver); 2847 if (stopped()) { 2848 return true; 2849 } 2850 2851 // Build field offset expression. 2852 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2853 // to be plain byte offsets, which are also the same as those accepted 2854 // by oopDesc::field_base. 2855 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2856 // 32-bit machines ignore the high half of long offsets 2857 offset = ConvL2X(offset); 2858 Node* adr = make_unsafe_address(base, offset); 2859 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2860 2861 // For CAS, unlike inline_unsafe_access, there seems no point in 2862 // trying to refine types. Just use the coarse types here. 2863 const Type *value_type = Type::get_const_basic_type(type); 2864 Compile::AliasType* alias_type = C->alias_type(adr_type); 2865 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2866 2867 if (kind == LS_xchg && type == T_OBJECT) { 2868 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2869 if (tjp != NULL) { 2870 value_type = tjp; 2871 } 2872 } 2873 2874 int alias_idx = C->get_alias_index(adr_type); 2875 2876 // Memory-model-wise, a LoadStore acts like a little synchronized 2877 // block, so needs barriers on each side. These don't translate 2878 // into actual barriers on most machines, but we still need rest of 2879 // compiler to respect ordering. 2880 2881 insert_mem_bar(Op_MemBarRelease); 2882 insert_mem_bar(Op_MemBarCPUOrder); 2883 2884 // 4984716: MemBars must be inserted before this 2885 // memory node in order to avoid a false 2886 // dependency which will confuse the scheduler. 2887 Node *mem = memory(alias_idx); 2888 2889 // For now, we handle only those cases that actually exist: ints, 2890 // longs, and Object. Adding others should be straightforward. 2891 Node* load_store; 2892 switch(type) { 2893 case T_INT: 2894 if (kind == LS_xadd) { 2895 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2896 } else if (kind == LS_xchg) { 2897 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2898 } else if (kind == LS_cmpxchg) { 2899 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2900 } else { 2901 ShouldNotReachHere(); 2902 } 2903 break; 2904 case T_LONG: 2905 if (kind == LS_xadd) { 2906 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2907 } else if (kind == LS_xchg) { 2908 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2909 } else if (kind == LS_cmpxchg) { 2910 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2911 } else { 2912 ShouldNotReachHere(); 2913 } 2914 break; 2915 case T_OBJECT: 2916 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2917 // could be delayed during Parse (for example, in adjust_map_after_if()). 2918 // Execute transformation here to avoid barrier generation in such case. 2919 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2920 newval = _gvn.makecon(TypePtr::NULL_PTR); 2921 2922 // Reference stores need a store barrier. 2923 if (kind == LS_xchg) { 2924 // If pre-barrier must execute before the oop store, old value will require do_load here. 2925 if (!can_move_pre_barrier()) { 2926 pre_barrier(true /* do_load*/, 2927 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2928 NULL /* pre_val*/, 2929 T_OBJECT); 2930 } // Else move pre_barrier to use load_store value, see below. 2931 } else if (kind == LS_cmpxchg) { 2932 // Same as for newval above: 2933 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2934 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2935 } 2936 // The only known value which might get overwritten is oldval. 2937 pre_barrier(false /* do_load */, 2938 control(), NULL, NULL, max_juint, NULL, NULL, 2939 oldval /* pre_val */, 2940 T_OBJECT); 2941 } else { 2942 ShouldNotReachHere(); 2943 } 2944 2945 #ifdef _LP64 2946 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2947 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2948 if (kind == LS_xchg) { 2949 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2950 newval_enc, adr_type, value_type->make_narrowoop())); 2951 } else { 2952 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2953 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2954 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2955 newval_enc, oldval_enc)); 2956 } 2957 } else 2958 #endif 2959 { 2960 if (kind == LS_xchg) { 2961 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2962 } else { 2963 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2964 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2965 } 2966 } 2967 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2968 break; 2969 default: 2970 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2971 break; 2972 } 2973 2974 // SCMemProjNodes represent the memory state of a LoadStore. Their 2975 // main role is to prevent LoadStore nodes from being optimized away 2976 // when their results aren't used. 2977 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2978 set_memory(proj, alias_idx); 2979 2980 if (type == T_OBJECT && kind == LS_xchg) { 2981 #ifdef _LP64 2982 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2983 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2984 } 2985 #endif 2986 if (can_move_pre_barrier()) { 2987 // Don't need to load pre_val. The old value is returned by load_store. 2988 // The pre_barrier can execute after the xchg as long as no safepoint 2989 // gets inserted between them. 2990 pre_barrier(false /* do_load */, 2991 control(), NULL, NULL, max_juint, NULL, NULL, 2992 load_store /* pre_val */, 2993 T_OBJECT); 2994 } 2995 } 2996 2997 // Add the trailing membar surrounding the access 2998 insert_mem_bar(Op_MemBarCPUOrder); 2999 insert_mem_bar(Op_MemBarAcquire); 3000 3001 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3002 set_result(load_store); 3003 return true; 3004 } 3005 3006 //----------------------------inline_unsafe_ordered_store---------------------- 3007 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3008 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3009 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3010 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3011 // This is another variant of inline_unsafe_access, differing in 3012 // that it always issues store-store ("release") barrier and ensures 3013 // store-atomicity (which only matters for "long"). 3014 3015 if (callee()->is_static()) return false; // caller must have the capability! 3016 3017 #ifndef PRODUCT 3018 { 3019 ResourceMark rm; 3020 // Check the signatures. 3021 ciSignature* sig = callee()->signature(); 3022 #ifdef ASSERT 3023 BasicType rtype = sig->return_type()->basic_type(); 3024 assert(rtype == T_VOID, "must return void"); 3025 assert(sig->count() == 3, "has 3 arguments"); 3026 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3027 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3028 #endif // ASSERT 3029 } 3030 #endif //PRODUCT 3031 3032 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3033 3034 // Get arguments: 3035 Node* receiver = argument(0); // type: oop 3036 Node* base = argument(1); // type: oop 3037 Node* offset = argument(2); // type: long 3038 Node* val = argument(4); // type: oop, int, or long 3039 3040 // Null check receiver. 3041 receiver = null_check(receiver); 3042 if (stopped()) { 3043 return true; 3044 } 3045 3046 // Build field offset expression. 3047 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3048 // 32-bit machines ignore the high half of long offsets 3049 offset = ConvL2X(offset); 3050 Node* adr = make_unsafe_address(base, offset); 3051 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3052 const Type *value_type = Type::get_const_basic_type(type); 3053 Compile::AliasType* alias_type = C->alias_type(adr_type); 3054 3055 insert_mem_bar(Op_MemBarRelease); 3056 insert_mem_bar(Op_MemBarCPUOrder); 3057 // Ensure that the store is atomic for longs: 3058 const bool require_atomic_access = true; 3059 Node* store; 3060 if (type == T_OBJECT) // reference stores need a store barrier. 3061 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3062 else { 3063 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3064 } 3065 insert_mem_bar(Op_MemBarCPUOrder); 3066 return true; 3067 } 3068 3069 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3070 // Regardless of form, don't allow previous ld/st to move down, 3071 // then issue acquire, release, or volatile mem_bar. 3072 insert_mem_bar(Op_MemBarCPUOrder); 3073 switch(id) { 3074 case vmIntrinsics::_loadFence: 3075 insert_mem_bar(Op_LoadFence); 3076 return true; 3077 case vmIntrinsics::_storeFence: 3078 insert_mem_bar(Op_StoreFence); 3079 return true; 3080 case vmIntrinsics::_fullFence: 3081 insert_mem_bar(Op_MemBarVolatile); 3082 return true; 3083 default: 3084 fatal_unexpected_iid(id); 3085 return false; 3086 } 3087 } 3088 3089 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3090 if (!kls->is_Con()) { 3091 return true; 3092 } 3093 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3094 if (klsptr == NULL) { 3095 return true; 3096 } 3097 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3098 // don't need a guard for a klass that is already initialized 3099 return !ik->is_initialized(); 3100 } 3101 3102 //----------------------------inline_unsafe_allocate--------------------------- 3103 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3104 bool LibraryCallKit::inline_unsafe_allocate() { 3105 if (callee()->is_static()) return false; // caller must have the capability! 3106 3107 null_check_receiver(); // null-check, then ignore 3108 Node* cls = null_check(argument(1)); 3109 if (stopped()) return true; 3110 3111 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3112 kls = null_check(kls); 3113 if (stopped()) return true; // argument was like int.class 3114 3115 Node* test = NULL; 3116 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3117 // Note: The argument might still be an illegal value like 3118 // Serializable.class or Object[].class. The runtime will handle it. 3119 // But we must make an explicit check for initialization. 3120 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3121 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3122 // can generate code to load it as unsigned byte. 3123 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3124 Node* bits = intcon(InstanceKlass::fully_initialized); 3125 test = _gvn.transform(new SubINode(inst, bits)); 3126 // The 'test' is non-zero if we need to take a slow path. 3127 } 3128 3129 Node* obj = new_instance(kls, test); 3130 set_result(obj); 3131 return true; 3132 } 3133 3134 #ifdef TRACE_HAVE_INTRINSICS 3135 /* 3136 * oop -> myklass 3137 * myklass->trace_id |= USED 3138 * return myklass->trace_id & ~0x3 3139 */ 3140 bool LibraryCallKit::inline_native_classID() { 3141 null_check_receiver(); // null-check, then ignore 3142 Node* cls = null_check(argument(1), T_OBJECT); 3143 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3144 kls = null_check(kls, T_OBJECT); 3145 ByteSize offset = TRACE_ID_OFFSET; 3146 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3147 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3148 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3149 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3150 Node* clsused = longcon(0x01l); // set the class bit 3151 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3152 3153 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3154 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3155 set_result(andl); 3156 return true; 3157 } 3158 3159 bool LibraryCallKit::inline_native_threadID() { 3160 Node* tls_ptr = NULL; 3161 Node* cur_thr = generate_current_thread(tls_ptr); 3162 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3163 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3164 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3165 3166 Node* threadid = NULL; 3167 size_t thread_id_size = OSThread::thread_id_size(); 3168 if (thread_id_size == (size_t) BytesPerLong) { 3169 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3170 } else if (thread_id_size == (size_t) BytesPerInt) { 3171 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3172 } else { 3173 ShouldNotReachHere(); 3174 } 3175 set_result(threadid); 3176 return true; 3177 } 3178 #endif 3179 3180 //------------------------inline_native_time_funcs-------------- 3181 // inline code for System.currentTimeMillis() and System.nanoTime() 3182 // these have the same type and signature 3183 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3184 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3185 const TypePtr* no_memory_effects = NULL; 3186 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3187 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3188 #ifdef ASSERT 3189 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3190 assert(value_top == top(), "second value must be top"); 3191 #endif 3192 set_result(value); 3193 return true; 3194 } 3195 3196 //------------------------inline_native_currentThread------------------ 3197 bool LibraryCallKit::inline_native_currentThread() { 3198 Node* junk = NULL; 3199 set_result(generate_current_thread(junk)); 3200 return true; 3201 } 3202 3203 //------------------------inline_native_isInterrupted------------------ 3204 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3205 bool LibraryCallKit::inline_native_isInterrupted() { 3206 // Add a fast path to t.isInterrupted(clear_int): 3207 // (t == Thread.current() && 3208 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3209 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3210 // So, in the common case that the interrupt bit is false, 3211 // we avoid making a call into the VM. Even if the interrupt bit 3212 // is true, if the clear_int argument is false, we avoid the VM call. 3213 // However, if the receiver is not currentThread, we must call the VM, 3214 // because there must be some locking done around the operation. 3215 3216 // We only go to the fast case code if we pass two guards. 3217 // Paths which do not pass are accumulated in the slow_region. 3218 3219 enum { 3220 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3221 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3222 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3223 PATH_LIMIT 3224 }; 3225 3226 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3227 // out of the function. 3228 insert_mem_bar(Op_MemBarCPUOrder); 3229 3230 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3231 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3232 3233 RegionNode* slow_region = new RegionNode(1); 3234 record_for_igvn(slow_region); 3235 3236 // (a) Receiving thread must be the current thread. 3237 Node* rec_thr = argument(0); 3238 Node* tls_ptr = NULL; 3239 Node* cur_thr = generate_current_thread(tls_ptr); 3240 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3241 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3242 3243 generate_slow_guard(bol_thr, slow_region); 3244 3245 // (b) Interrupt bit on TLS must be false. 3246 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3247 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3248 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3249 3250 // Set the control input on the field _interrupted read to prevent it floating up. 3251 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3252 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3253 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3254 3255 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3256 3257 // First fast path: if (!TLS._interrupted) return false; 3258 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3259 result_rgn->init_req(no_int_result_path, false_bit); 3260 result_val->init_req(no_int_result_path, intcon(0)); 3261 3262 // drop through to next case 3263 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3264 3265 #ifndef TARGET_OS_FAMILY_windows 3266 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3267 Node* clr_arg = argument(1); 3268 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3269 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3270 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3271 3272 // Second fast path: ... else if (!clear_int) return true; 3273 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3274 result_rgn->init_req(no_clear_result_path, false_arg); 3275 result_val->init_req(no_clear_result_path, intcon(1)); 3276 3277 // drop through to next case 3278 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3279 #else 3280 // To return true on Windows you must read the _interrupted field 3281 // and check the the event state i.e. take the slow path. 3282 #endif // TARGET_OS_FAMILY_windows 3283 3284 // (d) Otherwise, go to the slow path. 3285 slow_region->add_req(control()); 3286 set_control( _gvn.transform(slow_region)); 3287 3288 if (stopped()) { 3289 // There is no slow path. 3290 result_rgn->init_req(slow_result_path, top()); 3291 result_val->init_req(slow_result_path, top()); 3292 } else { 3293 // non-virtual because it is a private non-static 3294 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3295 3296 Node* slow_val = set_results_for_java_call(slow_call); 3297 // this->control() comes from set_results_for_java_call 3298 3299 Node* fast_io = slow_call->in(TypeFunc::I_O); 3300 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3301 3302 // These two phis are pre-filled with copies of of the fast IO and Memory 3303 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3304 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3305 3306 result_rgn->init_req(slow_result_path, control()); 3307 result_io ->init_req(slow_result_path, i_o()); 3308 result_mem->init_req(slow_result_path, reset_memory()); 3309 result_val->init_req(slow_result_path, slow_val); 3310 3311 set_all_memory(_gvn.transform(result_mem)); 3312 set_i_o( _gvn.transform(result_io)); 3313 } 3314 3315 C->set_has_split_ifs(true); // Has chance for split-if optimization 3316 set_result(result_rgn, result_val); 3317 return true; 3318 } 3319 3320 //---------------------------load_mirror_from_klass---------------------------- 3321 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3322 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3323 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3324 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3325 } 3326 3327 //-----------------------load_klass_from_mirror_common------------------------- 3328 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3329 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3330 // and branch to the given path on the region. 3331 // If never_see_null, take an uncommon trap on null, so we can optimistically 3332 // compile for the non-null case. 3333 // If the region is NULL, force never_see_null = true. 3334 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3335 bool never_see_null, 3336 RegionNode* region, 3337 int null_path, 3338 int offset) { 3339 if (region == NULL) never_see_null = true; 3340 Node* p = basic_plus_adr(mirror, offset); 3341 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3342 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3343 Node* null_ctl = top(); 3344 kls = null_check_oop(kls, &null_ctl, never_see_null); 3345 if (region != NULL) { 3346 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3347 region->init_req(null_path, null_ctl); 3348 } else { 3349 assert(null_ctl == top(), "no loose ends"); 3350 } 3351 return kls; 3352 } 3353 3354 //--------------------(inline_native_Class_query helpers)--------------------- 3355 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3356 // Fall through if (mods & mask) == bits, take the guard otherwise. 3357 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3358 // Branch around if the given klass has the given modifier bit set. 3359 // Like generate_guard, adds a new path onto the region. 3360 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3361 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3362 Node* mask = intcon(modifier_mask); 3363 Node* bits = intcon(modifier_bits); 3364 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3365 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3366 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3367 return generate_fair_guard(bol, region); 3368 } 3369 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3370 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3371 } 3372 3373 //-------------------------inline_native_Class_query------------------- 3374 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3375 const Type* return_type = TypeInt::BOOL; 3376 Node* prim_return_value = top(); // what happens if it's a primitive class? 3377 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3378 bool expect_prim = false; // most of these guys expect to work on refs 3379 3380 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3381 3382 Node* mirror = argument(0); 3383 Node* obj = top(); 3384 3385 switch (id) { 3386 case vmIntrinsics::_isInstance: 3387 // nothing is an instance of a primitive type 3388 prim_return_value = intcon(0); 3389 obj = argument(1); 3390 break; 3391 case vmIntrinsics::_getModifiers: 3392 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3393 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3394 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3395 break; 3396 case vmIntrinsics::_isInterface: 3397 prim_return_value = intcon(0); 3398 break; 3399 case vmIntrinsics::_isArray: 3400 prim_return_value = intcon(0); 3401 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3402 break; 3403 case vmIntrinsics::_isPrimitive: 3404 prim_return_value = intcon(1); 3405 expect_prim = true; // obviously 3406 break; 3407 case vmIntrinsics::_getSuperclass: 3408 prim_return_value = null(); 3409 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3410 break; 3411 case vmIntrinsics::_getClassAccessFlags: 3412 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3413 return_type = TypeInt::INT; // not bool! 6297094 3414 break; 3415 default: 3416 fatal_unexpected_iid(id); 3417 break; 3418 } 3419 3420 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3421 if (mirror_con == NULL) return false; // cannot happen? 3422 3423 #ifndef PRODUCT 3424 if (C->print_intrinsics() || C->print_inlining()) { 3425 ciType* k = mirror_con->java_mirror_type(); 3426 if (k) { 3427 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3428 k->print_name(); 3429 tty->cr(); 3430 } 3431 } 3432 #endif 3433 3434 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3435 RegionNode* region = new RegionNode(PATH_LIMIT); 3436 record_for_igvn(region); 3437 PhiNode* phi = new PhiNode(region, return_type); 3438 3439 // The mirror will never be null of Reflection.getClassAccessFlags, however 3440 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3441 // if it is. See bug 4774291. 3442 3443 // For Reflection.getClassAccessFlags(), the null check occurs in 3444 // the wrong place; see inline_unsafe_access(), above, for a similar 3445 // situation. 3446 mirror = null_check(mirror); 3447 // If mirror or obj is dead, only null-path is taken. 3448 if (stopped()) return true; 3449 3450 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3451 3452 // Now load the mirror's klass metaobject, and null-check it. 3453 // Side-effects region with the control path if the klass is null. 3454 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3455 // If kls is null, we have a primitive mirror. 3456 phi->init_req(_prim_path, prim_return_value); 3457 if (stopped()) { set_result(region, phi); return true; } 3458 bool safe_for_replace = (region->in(_prim_path) == top()); 3459 3460 Node* p; // handy temp 3461 Node* null_ctl; 3462 3463 // Now that we have the non-null klass, we can perform the real query. 3464 // For constant classes, the query will constant-fold in LoadNode::Value. 3465 Node* query_value = top(); 3466 switch (id) { 3467 case vmIntrinsics::_isInstance: 3468 // nothing is an instance of a primitive type 3469 query_value = gen_instanceof(obj, kls, safe_for_replace); 3470 break; 3471 3472 case vmIntrinsics::_getModifiers: 3473 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3474 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3475 break; 3476 3477 case vmIntrinsics::_isInterface: 3478 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3479 if (generate_interface_guard(kls, region) != NULL) 3480 // A guard was added. If the guard is taken, it was an interface. 3481 phi->add_req(intcon(1)); 3482 // If we fall through, it's a plain class. 3483 query_value = intcon(0); 3484 break; 3485 3486 case vmIntrinsics::_isArray: 3487 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3488 if (generate_array_guard(kls, region) != NULL) 3489 // A guard was added. If the guard is taken, it was an array. 3490 phi->add_req(intcon(1)); 3491 // If we fall through, it's a plain class. 3492 query_value = intcon(0); 3493 break; 3494 3495 case vmIntrinsics::_isPrimitive: 3496 query_value = intcon(0); // "normal" path produces false 3497 break; 3498 3499 case vmIntrinsics::_getSuperclass: 3500 // The rules here are somewhat unfortunate, but we can still do better 3501 // with random logic than with a JNI call. 3502 // Interfaces store null or Object as _super, but must report null. 3503 // Arrays store an intermediate super as _super, but must report Object. 3504 // Other types can report the actual _super. 3505 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3506 if (generate_interface_guard(kls, region) != NULL) 3507 // A guard was added. If the guard is taken, it was an interface. 3508 phi->add_req(null()); 3509 if (generate_array_guard(kls, region) != NULL) 3510 // A guard was added. If the guard is taken, it was an array. 3511 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3512 // If we fall through, it's a plain class. Get its _super. 3513 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3514 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3515 null_ctl = top(); 3516 kls = null_check_oop(kls, &null_ctl); 3517 if (null_ctl != top()) { 3518 // If the guard is taken, Object.superClass is null (both klass and mirror). 3519 region->add_req(null_ctl); 3520 phi ->add_req(null()); 3521 } 3522 if (!stopped()) { 3523 query_value = load_mirror_from_klass(kls); 3524 } 3525 break; 3526 3527 case vmIntrinsics::_getClassAccessFlags: 3528 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3529 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3530 break; 3531 3532 default: 3533 fatal_unexpected_iid(id); 3534 break; 3535 } 3536 3537 // Fall-through is the normal case of a query to a real class. 3538 phi->init_req(1, query_value); 3539 region->init_req(1, control()); 3540 3541 C->set_has_split_ifs(true); // Has chance for split-if optimization 3542 set_result(region, phi); 3543 return true; 3544 } 3545 3546 //-------------------------inline_Class_cast------------------- 3547 bool LibraryCallKit::inline_Class_cast() { 3548 Node* mirror = argument(0); // Class 3549 Node* obj = argument(1); 3550 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3551 if (mirror_con == NULL) { 3552 return false; // dead path (mirror->is_top()). 3553 } 3554 if (obj == NULL || obj->is_top()) { 3555 return false; // dead path 3556 } 3557 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3558 3559 // First, see if Class.cast() can be folded statically. 3560 // java_mirror_type() returns non-null for compile-time Class constants. 3561 ciType* tm = mirror_con->java_mirror_type(); 3562 if (tm != NULL && tm->is_klass() && 3563 tp != NULL && tp->klass() != NULL) { 3564 if (!tp->klass()->is_loaded()) { 3565 // Don't use intrinsic when class is not loaded. 3566 return false; 3567 } else { 3568 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3569 if (static_res == Compile::SSC_always_true) { 3570 // isInstance() is true - fold the code. 3571 set_result(obj); 3572 return true; 3573 } else if (static_res == Compile::SSC_always_false) { 3574 // Don't use intrinsic, have to throw ClassCastException. 3575 // If the reference is null, the non-intrinsic bytecode will 3576 // be optimized appropriately. 3577 return false; 3578 } 3579 } 3580 } 3581 3582 // Bailout intrinsic and do normal inlining if exception path is frequent. 3583 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3584 return false; 3585 } 3586 3587 // Generate dynamic checks. 3588 // Class.cast() is java implementation of _checkcast bytecode. 3589 // Do checkcast (Parse::do_checkcast()) optimizations here. 3590 3591 mirror = null_check(mirror); 3592 // If mirror is dead, only null-path is taken. 3593 if (stopped()) { 3594 return true; 3595 } 3596 3597 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3598 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3599 RegionNode* region = new RegionNode(PATH_LIMIT); 3600 record_for_igvn(region); 3601 3602 // Now load the mirror's klass metaobject, and null-check it. 3603 // If kls is null, we have a primitive mirror and 3604 // nothing is an instance of a primitive type. 3605 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3606 3607 Node* res = top(); 3608 if (!stopped()) { 3609 Node* bad_type_ctrl = top(); 3610 // Do checkcast optimizations. 3611 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3612 region->init_req(_bad_type_path, bad_type_ctrl); 3613 } 3614 if (region->in(_prim_path) != top() || 3615 region->in(_bad_type_path) != top()) { 3616 // Let Interpreter throw ClassCastException. 3617 PreserveJVMState pjvms(this); 3618 set_control(_gvn.transform(region)); 3619 uncommon_trap(Deoptimization::Reason_intrinsic, 3620 Deoptimization::Action_maybe_recompile); 3621 } 3622 if (!stopped()) { 3623 set_result(res); 3624 } 3625 return true; 3626 } 3627 3628 3629 //--------------------------inline_native_subtype_check------------------------ 3630 // This intrinsic takes the JNI calls out of the heart of 3631 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3632 bool LibraryCallKit::inline_native_subtype_check() { 3633 // Pull both arguments off the stack. 3634 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3635 args[0] = argument(0); 3636 args[1] = argument(1); 3637 Node* klasses[2]; // corresponding Klasses: superk, subk 3638 klasses[0] = klasses[1] = top(); 3639 3640 enum { 3641 // A full decision tree on {superc is prim, subc is prim}: 3642 _prim_0_path = 1, // {P,N} => false 3643 // {P,P} & superc!=subc => false 3644 _prim_same_path, // {P,P} & superc==subc => true 3645 _prim_1_path, // {N,P} => false 3646 _ref_subtype_path, // {N,N} & subtype check wins => true 3647 _both_ref_path, // {N,N} & subtype check loses => false 3648 PATH_LIMIT 3649 }; 3650 3651 RegionNode* region = new RegionNode(PATH_LIMIT); 3652 Node* phi = new PhiNode(region, TypeInt::BOOL); 3653 record_for_igvn(region); 3654 3655 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3656 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3657 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3658 3659 // First null-check both mirrors and load each mirror's klass metaobject. 3660 int which_arg; 3661 for (which_arg = 0; which_arg <= 1; which_arg++) { 3662 Node* arg = args[which_arg]; 3663 arg = null_check(arg); 3664 if (stopped()) break; 3665 args[which_arg] = arg; 3666 3667 Node* p = basic_plus_adr(arg, class_klass_offset); 3668 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3669 klasses[which_arg] = _gvn.transform(kls); 3670 } 3671 3672 // Having loaded both klasses, test each for null. 3673 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3674 for (which_arg = 0; which_arg <= 1; which_arg++) { 3675 Node* kls = klasses[which_arg]; 3676 Node* null_ctl = top(); 3677 kls = null_check_oop(kls, &null_ctl, never_see_null); 3678 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3679 region->init_req(prim_path, null_ctl); 3680 if (stopped()) break; 3681 klasses[which_arg] = kls; 3682 } 3683 3684 if (!stopped()) { 3685 // now we have two reference types, in klasses[0..1] 3686 Node* subk = klasses[1]; // the argument to isAssignableFrom 3687 Node* superk = klasses[0]; // the receiver 3688 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3689 // now we have a successful reference subtype check 3690 region->set_req(_ref_subtype_path, control()); 3691 } 3692 3693 // If both operands are primitive (both klasses null), then 3694 // we must return true when they are identical primitives. 3695 // It is convenient to test this after the first null klass check. 3696 set_control(region->in(_prim_0_path)); // go back to first null check 3697 if (!stopped()) { 3698 // Since superc is primitive, make a guard for the superc==subc case. 3699 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3700 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3701 generate_guard(bol_eq, region, PROB_FAIR); 3702 if (region->req() == PATH_LIMIT+1) { 3703 // A guard was added. If the added guard is taken, superc==subc. 3704 region->swap_edges(PATH_LIMIT, _prim_same_path); 3705 region->del_req(PATH_LIMIT); 3706 } 3707 region->set_req(_prim_0_path, control()); // Not equal after all. 3708 } 3709 3710 // these are the only paths that produce 'true': 3711 phi->set_req(_prim_same_path, intcon(1)); 3712 phi->set_req(_ref_subtype_path, intcon(1)); 3713 3714 // pull together the cases: 3715 assert(region->req() == PATH_LIMIT, "sane region"); 3716 for (uint i = 1; i < region->req(); i++) { 3717 Node* ctl = region->in(i); 3718 if (ctl == NULL || ctl == top()) { 3719 region->set_req(i, top()); 3720 phi ->set_req(i, top()); 3721 } else if (phi->in(i) == NULL) { 3722 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3723 } 3724 } 3725 3726 set_control(_gvn.transform(region)); 3727 set_result(_gvn.transform(phi)); 3728 return true; 3729 } 3730 3731 //---------------------generate_array_guard_common------------------------ 3732 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3733 bool obj_array, bool not_array) { 3734 // If obj_array/non_array==false/false: 3735 // Branch around if the given klass is in fact an array (either obj or prim). 3736 // If obj_array/non_array==false/true: 3737 // Branch around if the given klass is not an array klass of any kind. 3738 // If obj_array/non_array==true/true: 3739 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3740 // If obj_array/non_array==true/false: 3741 // Branch around if the kls is an oop array (Object[] or subtype) 3742 // 3743 // Like generate_guard, adds a new path onto the region. 3744 jint layout_con = 0; 3745 Node* layout_val = get_layout_helper(kls, layout_con); 3746 if (layout_val == NULL) { 3747 bool query = (obj_array 3748 ? Klass::layout_helper_is_objArray(layout_con) 3749 : Klass::layout_helper_is_array(layout_con)); 3750 if (query == not_array) { 3751 return NULL; // never a branch 3752 } else { // always a branch 3753 Node* always_branch = control(); 3754 if (region != NULL) 3755 region->add_req(always_branch); 3756 set_control(top()); 3757 return always_branch; 3758 } 3759 } 3760 // Now test the correct condition. 3761 jint nval = (obj_array 3762 ? ((jint)Klass::_lh_array_tag_type_value 3763 << Klass::_lh_array_tag_shift) 3764 : Klass::_lh_neutral_value); 3765 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3766 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3767 // invert the test if we are looking for a non-array 3768 if (not_array) btest = BoolTest(btest).negate(); 3769 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3770 return generate_fair_guard(bol, region); 3771 } 3772 3773 3774 //-----------------------inline_native_newArray-------------------------- 3775 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3776 bool LibraryCallKit::inline_native_newArray() { 3777 Node* mirror = argument(0); 3778 Node* count_val = argument(1); 3779 3780 mirror = null_check(mirror); 3781 // If mirror or obj is dead, only null-path is taken. 3782 if (stopped()) return true; 3783 3784 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3785 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3786 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3787 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3788 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3789 3790 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3791 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3792 result_reg, _slow_path); 3793 Node* normal_ctl = control(); 3794 Node* no_array_ctl = result_reg->in(_slow_path); 3795 3796 // Generate code for the slow case. We make a call to newArray(). 3797 set_control(no_array_ctl); 3798 if (!stopped()) { 3799 // Either the input type is void.class, or else the 3800 // array klass has not yet been cached. Either the 3801 // ensuing call will throw an exception, or else it 3802 // will cache the array klass for next time. 3803 PreserveJVMState pjvms(this); 3804 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3805 Node* slow_result = set_results_for_java_call(slow_call); 3806 // this->control() comes from set_results_for_java_call 3807 result_reg->set_req(_slow_path, control()); 3808 result_val->set_req(_slow_path, slow_result); 3809 result_io ->set_req(_slow_path, i_o()); 3810 result_mem->set_req(_slow_path, reset_memory()); 3811 } 3812 3813 set_control(normal_ctl); 3814 if (!stopped()) { 3815 // Normal case: The array type has been cached in the java.lang.Class. 3816 // The following call works fine even if the array type is polymorphic. 3817 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3818 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3819 result_reg->init_req(_normal_path, control()); 3820 result_val->init_req(_normal_path, obj); 3821 result_io ->init_req(_normal_path, i_o()); 3822 result_mem->init_req(_normal_path, reset_memory()); 3823 } 3824 3825 // Return the combined state. 3826 set_i_o( _gvn.transform(result_io) ); 3827 set_all_memory( _gvn.transform(result_mem)); 3828 3829 C->set_has_split_ifs(true); // Has chance for split-if optimization 3830 set_result(result_reg, result_val); 3831 return true; 3832 } 3833 3834 //----------------------inline_native_getLength-------------------------- 3835 // public static native int java.lang.reflect.Array.getLength(Object array); 3836 bool LibraryCallKit::inline_native_getLength() { 3837 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3838 3839 Node* array = null_check(argument(0)); 3840 // If array is dead, only null-path is taken. 3841 if (stopped()) return true; 3842 3843 // Deoptimize if it is a non-array. 3844 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3845 3846 if (non_array != NULL) { 3847 PreserveJVMState pjvms(this); 3848 set_control(non_array); 3849 uncommon_trap(Deoptimization::Reason_intrinsic, 3850 Deoptimization::Action_maybe_recompile); 3851 } 3852 3853 // If control is dead, only non-array-path is taken. 3854 if (stopped()) return true; 3855 3856 // The works fine even if the array type is polymorphic. 3857 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3858 Node* result = load_array_length(array); 3859 3860 C->set_has_split_ifs(true); // Has chance for split-if optimization 3861 set_result(result); 3862 return true; 3863 } 3864 3865 //------------------------inline_array_copyOf---------------------------- 3866 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3867 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3868 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3869 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3870 3871 // Get the arguments. 3872 Node* original = argument(0); 3873 Node* start = is_copyOfRange? argument(1): intcon(0); 3874 Node* end = is_copyOfRange? argument(2): argument(1); 3875 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3876 3877 Node* newcopy; 3878 3879 // Set the original stack and the reexecute bit for the interpreter to reexecute 3880 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3881 { PreserveReexecuteState preexecs(this); 3882 jvms()->set_should_reexecute(true); 3883 3884 array_type_mirror = null_check(array_type_mirror); 3885 original = null_check(original); 3886 3887 // Check if a null path was taken unconditionally. 3888 if (stopped()) return true; 3889 3890 Node* orig_length = load_array_length(original); 3891 3892 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3893 klass_node = null_check(klass_node); 3894 3895 RegionNode* bailout = new RegionNode(1); 3896 record_for_igvn(bailout); 3897 3898 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3899 // Bail out if that is so. 3900 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3901 if (not_objArray != NULL) { 3902 // Improve the klass node's type from the new optimistic assumption: 3903 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3904 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3905 Node* cast = new CastPPNode(klass_node, akls); 3906 cast->init_req(0, control()); 3907 klass_node = _gvn.transform(cast); 3908 } 3909 3910 // Bail out if either start or end is negative. 3911 generate_negative_guard(start, bailout, &start); 3912 generate_negative_guard(end, bailout, &end); 3913 3914 Node* length = end; 3915 if (_gvn.type(start) != TypeInt::ZERO) { 3916 length = _gvn.transform(new SubINode(end, start)); 3917 } 3918 3919 // Bail out if length is negative. 3920 // Without this the new_array would throw 3921 // NegativeArraySizeException but IllegalArgumentException is what 3922 // should be thrown 3923 generate_negative_guard(length, bailout, &length); 3924 3925 if (bailout->req() > 1) { 3926 PreserveJVMState pjvms(this); 3927 set_control(_gvn.transform(bailout)); 3928 uncommon_trap(Deoptimization::Reason_intrinsic, 3929 Deoptimization::Action_maybe_recompile); 3930 } 3931 3932 if (!stopped()) { 3933 // How many elements will we copy from the original? 3934 // The answer is MinI(orig_length - start, length). 3935 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3936 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3937 3938 newcopy = new_array(klass_node, length, 0); // no arguments to push 3939 3940 // Generate a direct call to the right arraycopy function(s). 3941 // We know the copy is disjoint but we might not know if the 3942 // oop stores need checking. 3943 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3944 // This will fail a store-check if x contains any non-nulls. 3945 3946 Node* alloc = tightly_coupled_allocation(newcopy, NULL); 3947 3948 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, alloc != NULL, 3949 load_object_klass(original), klass_node); 3950 if (!is_copyOfRange) { 3951 ac->set_copyof(); 3952 } else { 3953 ac->set_copyofrange(); 3954 } 3955 Node* n = _gvn.transform(ac); 3956 assert(n == ac, "cannot disappear"); 3957 ac->connect_outputs(this); 3958 } 3959 } // original reexecute is set back here 3960 3961 C->set_has_split_ifs(true); // Has chance for split-if optimization 3962 if (!stopped()) { 3963 set_result(newcopy); 3964 } 3965 return true; 3966 } 3967 3968 3969 //----------------------generate_virtual_guard--------------------------- 3970 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3971 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3972 RegionNode* slow_region) { 3973 ciMethod* method = callee(); 3974 int vtable_index = method->vtable_index(); 3975 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3976 err_msg_res("bad index %d", vtable_index)); 3977 // Get the Method* out of the appropriate vtable entry. 3978 int entry_offset = (InstanceKlass::vtable_start_offset() + 3979 vtable_index*vtableEntry::size()) * wordSize + 3980 vtableEntry::method_offset_in_bytes(); 3981 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3982 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3983 3984 // Compare the target method with the expected method (e.g., Object.hashCode). 3985 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3986 3987 Node* native_call = makecon(native_call_addr); 3988 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3989 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3990 3991 return generate_slow_guard(test_native, slow_region); 3992 } 3993 3994 //-----------------------generate_method_call---------------------------- 3995 // Use generate_method_call to make a slow-call to the real 3996 // method if the fast path fails. An alternative would be to 3997 // use a stub like OptoRuntime::slow_arraycopy_Java. 3998 // This only works for expanding the current library call, 3999 // not another intrinsic. (E.g., don't use this for making an 4000 // arraycopy call inside of the copyOf intrinsic.) 4001 CallJavaNode* 4002 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4003 // When compiling the intrinsic method itself, do not use this technique. 4004 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4005 4006 ciMethod* method = callee(); 4007 // ensure the JVMS we have will be correct for this call 4008 guarantee(method_id == method->intrinsic_id(), "must match"); 4009 4010 const TypeFunc* tf = TypeFunc::make(method); 4011 CallJavaNode* slow_call; 4012 if (is_static) { 4013 assert(!is_virtual, ""); 4014 slow_call = new CallStaticJavaNode(C, tf, 4015 SharedRuntime::get_resolve_static_call_stub(), 4016 method, bci()); 4017 } else if (is_virtual) { 4018 null_check_receiver(); 4019 int vtable_index = Method::invalid_vtable_index; 4020 if (UseInlineCaches) { 4021 // Suppress the vtable call 4022 } else { 4023 // hashCode and clone are not a miranda methods, 4024 // so the vtable index is fixed. 4025 // No need to use the linkResolver to get it. 4026 vtable_index = method->vtable_index(); 4027 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4028 err_msg_res("bad index %d", vtable_index)); 4029 } 4030 slow_call = new CallDynamicJavaNode(tf, 4031 SharedRuntime::get_resolve_virtual_call_stub(), 4032 method, vtable_index, bci()); 4033 } else { // neither virtual nor static: opt_virtual 4034 null_check_receiver(); 4035 slow_call = new CallStaticJavaNode(C, tf, 4036 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4037 method, bci()); 4038 slow_call->set_optimized_virtual(true); 4039 } 4040 set_arguments_for_java_call(slow_call); 4041 set_edges_for_java_call(slow_call); 4042 return slow_call; 4043 } 4044 4045 4046 /** 4047 * Build special case code for calls to hashCode on an object. This call may 4048 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4049 * slightly different code. 4050 */ 4051 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4052 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4053 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4054 4055 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4056 4057 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4058 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4059 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4060 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4061 Node* obj = NULL; 4062 if (!is_static) { 4063 // Check for hashing null object 4064 obj = null_check_receiver(); 4065 if (stopped()) return true; // unconditionally null 4066 result_reg->init_req(_null_path, top()); 4067 result_val->init_req(_null_path, top()); 4068 } else { 4069 // Do a null check, and return zero if null. 4070 // System.identityHashCode(null) == 0 4071 obj = argument(0); 4072 Node* null_ctl = top(); 4073 obj = null_check_oop(obj, &null_ctl); 4074 result_reg->init_req(_null_path, null_ctl); 4075 result_val->init_req(_null_path, _gvn.intcon(0)); 4076 } 4077 4078 // Unconditionally null? Then return right away. 4079 if (stopped()) { 4080 set_control( result_reg->in(_null_path)); 4081 if (!stopped()) 4082 set_result(result_val->in(_null_path)); 4083 return true; 4084 } 4085 4086 // We only go to the fast case code if we pass a number of guards. The 4087 // paths which do not pass are accumulated in the slow_region. 4088 RegionNode* slow_region = new RegionNode(1); 4089 record_for_igvn(slow_region); 4090 4091 // If this is a virtual call, we generate a funny guard. We pull out 4092 // the vtable entry corresponding to hashCode() from the target object. 4093 // If the target method which we are calling happens to be the native 4094 // Object hashCode() method, we pass the guard. We do not need this 4095 // guard for non-virtual calls -- the caller is known to be the native 4096 // Object hashCode(). 4097 if (is_virtual) { 4098 // After null check, get the object's klass. 4099 Node* obj_klass = load_object_klass(obj); 4100 generate_virtual_guard(obj_klass, slow_region); 4101 } 4102 4103 // Get the header out of the object, use LoadMarkNode when available 4104 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4105 // The control of the load must be NULL. Otherwise, the load can move before 4106 // the null check after castPP removal. 4107 Node* no_ctrl = NULL; 4108 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4109 4110 // Test the header to see if it is unlocked. 4111 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4112 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4113 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4114 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4115 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4116 4117 generate_slow_guard(test_unlocked, slow_region); 4118 4119 // Get the hash value and check to see that it has been properly assigned. 4120 // We depend on hash_mask being at most 32 bits and avoid the use of 4121 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4122 // vm: see markOop.hpp. 4123 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4124 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4125 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4126 // This hack lets the hash bits live anywhere in the mark object now, as long 4127 // as the shift drops the relevant bits into the low 32 bits. Note that 4128 // Java spec says that HashCode is an int so there's no point in capturing 4129 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4130 hshifted_header = ConvX2I(hshifted_header); 4131 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4132 4133 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4134 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4135 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4136 4137 generate_slow_guard(test_assigned, slow_region); 4138 4139 Node* init_mem = reset_memory(); 4140 // fill in the rest of the null path: 4141 result_io ->init_req(_null_path, i_o()); 4142 result_mem->init_req(_null_path, init_mem); 4143 4144 result_val->init_req(_fast_path, hash_val); 4145 result_reg->init_req(_fast_path, control()); 4146 result_io ->init_req(_fast_path, i_o()); 4147 result_mem->init_req(_fast_path, init_mem); 4148 4149 // Generate code for the slow case. We make a call to hashCode(). 4150 set_control(_gvn.transform(slow_region)); 4151 if (!stopped()) { 4152 // No need for PreserveJVMState, because we're using up the present state. 4153 set_all_memory(init_mem); 4154 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4155 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4156 Node* slow_result = set_results_for_java_call(slow_call); 4157 // this->control() comes from set_results_for_java_call 4158 result_reg->init_req(_slow_path, control()); 4159 result_val->init_req(_slow_path, slow_result); 4160 result_io ->set_req(_slow_path, i_o()); 4161 result_mem ->set_req(_slow_path, reset_memory()); 4162 } 4163 4164 // Return the combined state. 4165 set_i_o( _gvn.transform(result_io) ); 4166 set_all_memory( _gvn.transform(result_mem)); 4167 4168 set_result(result_reg, result_val); 4169 return true; 4170 } 4171 4172 //---------------------------inline_native_getClass---------------------------- 4173 // public final native Class<?> java.lang.Object.getClass(); 4174 // 4175 // Build special case code for calls to getClass on an object. 4176 bool LibraryCallKit::inline_native_getClass() { 4177 Node* obj = null_check_receiver(); 4178 if (stopped()) return true; 4179 set_result(load_mirror_from_klass(load_object_klass(obj))); 4180 return true; 4181 } 4182 4183 //-----------------inline_native_Reflection_getCallerClass--------------------- 4184 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4185 // 4186 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4187 // 4188 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4189 // in that it must skip particular security frames and checks for 4190 // caller sensitive methods. 4191 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4192 #ifndef PRODUCT 4193 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4194 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4195 } 4196 #endif 4197 4198 if (!jvms()->has_method()) { 4199 #ifndef PRODUCT 4200 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4201 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4202 } 4203 #endif 4204 return false; 4205 } 4206 4207 // Walk back up the JVM state to find the caller at the required 4208 // depth. 4209 JVMState* caller_jvms = jvms(); 4210 4211 // Cf. JVM_GetCallerClass 4212 // NOTE: Start the loop at depth 1 because the current JVM state does 4213 // not include the Reflection.getCallerClass() frame. 4214 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4215 ciMethod* m = caller_jvms->method(); 4216 switch (n) { 4217 case 0: 4218 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4219 break; 4220 case 1: 4221 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4222 if (!m->caller_sensitive()) { 4223 #ifndef PRODUCT 4224 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4225 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4226 } 4227 #endif 4228 return false; // bail-out; let JVM_GetCallerClass do the work 4229 } 4230 break; 4231 default: 4232 if (!m->is_ignored_by_security_stack_walk()) { 4233 // We have reached the desired frame; return the holder class. 4234 // Acquire method holder as java.lang.Class and push as constant. 4235 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4236 ciInstance* caller_mirror = caller_klass->java_mirror(); 4237 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4238 4239 #ifndef PRODUCT 4240 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4241 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()); 4242 tty->print_cr(" JVM state at this point:"); 4243 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4244 ciMethod* m = jvms()->of_depth(i)->method(); 4245 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4246 } 4247 } 4248 #endif 4249 return true; 4250 } 4251 break; 4252 } 4253 } 4254 4255 #ifndef PRODUCT 4256 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4257 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4258 tty->print_cr(" JVM state at this point:"); 4259 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4260 ciMethod* m = jvms()->of_depth(i)->method(); 4261 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4262 } 4263 } 4264 #endif 4265 4266 return false; // bail-out; let JVM_GetCallerClass do the work 4267 } 4268 4269 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4270 Node* arg = argument(0); 4271 Node* result; 4272 4273 switch (id) { 4274 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4275 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4276 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4277 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4278 4279 case vmIntrinsics::_doubleToLongBits: { 4280 // two paths (plus control) merge in a wood 4281 RegionNode *r = new RegionNode(3); 4282 Node *phi = new PhiNode(r, TypeLong::LONG); 4283 4284 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4285 // Build the boolean node 4286 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4287 4288 // Branch either way. 4289 // NaN case is less traveled, which makes all the difference. 4290 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4291 Node *opt_isnan = _gvn.transform(ifisnan); 4292 assert( opt_isnan->is_If(), "Expect an IfNode"); 4293 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4294 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4295 4296 set_control(iftrue); 4297 4298 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4299 Node *slow_result = longcon(nan_bits); // return NaN 4300 phi->init_req(1, _gvn.transform( slow_result )); 4301 r->init_req(1, iftrue); 4302 4303 // Else fall through 4304 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4305 set_control(iffalse); 4306 4307 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4308 r->init_req(2, iffalse); 4309 4310 // Post merge 4311 set_control(_gvn.transform(r)); 4312 record_for_igvn(r); 4313 4314 C->set_has_split_ifs(true); // Has chance for split-if optimization 4315 result = phi; 4316 assert(result->bottom_type()->isa_long(), "must be"); 4317 break; 4318 } 4319 4320 case vmIntrinsics::_floatToIntBits: { 4321 // two paths (plus control) merge in a wood 4322 RegionNode *r = new RegionNode(3); 4323 Node *phi = new PhiNode(r, TypeInt::INT); 4324 4325 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4326 // Build the boolean node 4327 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4328 4329 // Branch either way. 4330 // NaN case is less traveled, which makes all the difference. 4331 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4332 Node *opt_isnan = _gvn.transform(ifisnan); 4333 assert( opt_isnan->is_If(), "Expect an IfNode"); 4334 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4335 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4336 4337 set_control(iftrue); 4338 4339 static const jint nan_bits = 0x7fc00000; 4340 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4341 phi->init_req(1, _gvn.transform( slow_result )); 4342 r->init_req(1, iftrue); 4343 4344 // Else fall through 4345 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4346 set_control(iffalse); 4347 4348 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4349 r->init_req(2, iffalse); 4350 4351 // Post merge 4352 set_control(_gvn.transform(r)); 4353 record_for_igvn(r); 4354 4355 C->set_has_split_ifs(true); // Has chance for split-if optimization 4356 result = phi; 4357 assert(result->bottom_type()->isa_int(), "must be"); 4358 break; 4359 } 4360 4361 default: 4362 fatal_unexpected_iid(id); 4363 break; 4364 } 4365 set_result(_gvn.transform(result)); 4366 return true; 4367 } 4368 4369 #ifdef _LP64 4370 #define XTOP ,top() /*additional argument*/ 4371 #else //_LP64 4372 #define XTOP /*no additional argument*/ 4373 #endif //_LP64 4374 4375 //----------------------inline_unsafe_copyMemory------------------------- 4376 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4377 bool LibraryCallKit::inline_unsafe_copyMemory() { 4378 if (callee()->is_static()) return false; // caller must have the capability! 4379 null_check_receiver(); // null-check receiver 4380 if (stopped()) return true; 4381 4382 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4383 4384 Node* src_ptr = argument(1); // type: oop 4385 Node* src_off = ConvL2X(argument(2)); // type: long 4386 Node* dst_ptr = argument(4); // type: oop 4387 Node* dst_off = ConvL2X(argument(5)); // type: long 4388 Node* size = ConvL2X(argument(7)); // type: long 4389 4390 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4391 "fieldOffset must be byte-scaled"); 4392 4393 Node* src = make_unsafe_address(src_ptr, src_off); 4394 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4395 4396 // Conservatively insert a memory barrier on all memory slices. 4397 // Do not let writes of the copy source or destination float below the copy. 4398 insert_mem_bar(Op_MemBarCPUOrder); 4399 4400 // Call it. Note that the length argument is not scaled. 4401 make_runtime_call(RC_LEAF|RC_NO_FP, 4402 OptoRuntime::fast_arraycopy_Type(), 4403 StubRoutines::unsafe_arraycopy(), 4404 "unsafe_arraycopy", 4405 TypeRawPtr::BOTTOM, 4406 src, dst, size XTOP); 4407 4408 // Do not let reads of the copy destination float above the copy. 4409 insert_mem_bar(Op_MemBarCPUOrder); 4410 4411 return true; 4412 } 4413 4414 //------------------------clone_coping----------------------------------- 4415 // Helper function for inline_native_clone. 4416 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4417 assert(obj_size != NULL, ""); 4418 Node* raw_obj = alloc_obj->in(1); 4419 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4420 4421 AllocateNode* alloc = NULL; 4422 if (ReduceBulkZeroing) { 4423 // We will be completely responsible for initializing this object - 4424 // mark Initialize node as complete. 4425 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4426 // The object was just allocated - there should be no any stores! 4427 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4428 // Mark as complete_with_arraycopy so that on AllocateNode 4429 // expansion, we know this AllocateNode is initialized by an array 4430 // copy and a StoreStore barrier exists after the array copy. 4431 alloc->initialization()->set_complete_with_arraycopy(); 4432 } 4433 4434 // Copy the fastest available way. 4435 // TODO: generate fields copies for small objects instead. 4436 Node* src = obj; 4437 Node* dest = alloc_obj; 4438 Node* size = _gvn.transform(obj_size); 4439 4440 // Exclude the header but include array length to copy by 8 bytes words. 4441 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4442 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4443 instanceOopDesc::base_offset_in_bytes(); 4444 // base_off: 4445 // 8 - 32-bit VM 4446 // 12 - 64-bit VM, compressed klass 4447 // 16 - 64-bit VM, normal klass 4448 if (base_off % BytesPerLong != 0) { 4449 assert(UseCompressedClassPointers, ""); 4450 if (is_array) { 4451 // Exclude length to copy by 8 bytes words. 4452 base_off += sizeof(int); 4453 } else { 4454 // Include klass to copy by 8 bytes words. 4455 base_off = instanceOopDesc::klass_offset_in_bytes(); 4456 } 4457 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4458 } 4459 src = basic_plus_adr(src, base_off); 4460 dest = basic_plus_adr(dest, base_off); 4461 4462 // Compute the length also, if needed: 4463 Node* countx = size; 4464 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4465 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4466 4467 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4468 4469 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4470 ac->set_clonebasic(); 4471 Node* n = _gvn.transform(ac); 4472 if (n == ac) { 4473 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4474 } else { 4475 set_all_memory(n); 4476 } 4477 4478 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4479 if (card_mark) { 4480 assert(!is_array, ""); 4481 // Put in store barrier for any and all oops we are sticking 4482 // into this object. (We could avoid this if we could prove 4483 // that the object type contains no oop fields at all.) 4484 Node* no_particular_value = NULL; 4485 Node* no_particular_field = NULL; 4486 int raw_adr_idx = Compile::AliasIdxRaw; 4487 post_barrier(control(), 4488 memory(raw_adr_type), 4489 alloc_obj, 4490 no_particular_field, 4491 raw_adr_idx, 4492 no_particular_value, 4493 T_OBJECT, 4494 false); 4495 } 4496 4497 // Do not let reads from the cloned object float above the arraycopy. 4498 if (alloc != NULL) { 4499 // Do not let stores that initialize this object be reordered with 4500 // a subsequent store that would make this object accessible by 4501 // other threads. 4502 // Record what AllocateNode this StoreStore protects so that 4503 // escape analysis can go from the MemBarStoreStoreNode to the 4504 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4505 // based on the escape status of the AllocateNode. 4506 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4507 } else { 4508 insert_mem_bar(Op_MemBarCPUOrder); 4509 } 4510 } 4511 4512 //------------------------inline_native_clone---------------------------- 4513 // protected native Object java.lang.Object.clone(); 4514 // 4515 // Here are the simple edge cases: 4516 // null receiver => normal trap 4517 // virtual and clone was overridden => slow path to out-of-line clone 4518 // not cloneable or finalizer => slow path to out-of-line Object.clone 4519 // 4520 // The general case has two steps, allocation and copying. 4521 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4522 // 4523 // Copying also has two cases, oop arrays and everything else. 4524 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4525 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4526 // 4527 // These steps fold up nicely if and when the cloned object's klass 4528 // can be sharply typed as an object array, a type array, or an instance. 4529 // 4530 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4531 PhiNode* result_val; 4532 4533 // Set the reexecute bit for the interpreter to reexecute 4534 // the bytecode that invokes Object.clone if deoptimization happens. 4535 { PreserveReexecuteState preexecs(this); 4536 jvms()->set_should_reexecute(true); 4537 4538 Node* obj = null_check_receiver(); 4539 if (stopped()) return true; 4540 4541 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4542 4543 // If we are going to clone an instance, we need its exact type to 4544 // know the number and types of fields to convert the clone to 4545 // loads/stores. Maybe a speculative type can help us. 4546 if (!obj_type->klass_is_exact() && 4547 obj_type->speculative_type() != NULL && 4548 obj_type->speculative_type()->is_instance_klass()) { 4549 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4550 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4551 !spec_ik->has_injected_fields()) { 4552 ciKlass* k = obj_type->klass(); 4553 if (!k->is_instance_klass() || 4554 k->as_instance_klass()->is_interface() || 4555 k->as_instance_klass()->has_subklass()) { 4556 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4557 } 4558 } 4559 } 4560 4561 Node* obj_klass = load_object_klass(obj); 4562 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4563 const TypeOopPtr* toop = ((tklass != NULL) 4564 ? tklass->as_instance_type() 4565 : TypeInstPtr::NOTNULL); 4566 4567 // Conservatively insert a memory barrier on all memory slices. 4568 // Do not let writes into the original float below the clone. 4569 insert_mem_bar(Op_MemBarCPUOrder); 4570 4571 // paths into result_reg: 4572 enum { 4573 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4574 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4575 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4576 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4577 PATH_LIMIT 4578 }; 4579 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4580 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4581 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4582 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4583 record_for_igvn(result_reg); 4584 4585 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4586 int raw_adr_idx = Compile::AliasIdxRaw; 4587 4588 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4589 if (array_ctl != NULL) { 4590 // It's an array. 4591 PreserveJVMState pjvms(this); 4592 set_control(array_ctl); 4593 Node* obj_length = load_array_length(obj); 4594 Node* obj_size = NULL; 4595 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4596 4597 if (!use_ReduceInitialCardMarks()) { 4598 // If it is an oop array, it requires very special treatment, 4599 // because card marking is required on each card of the array. 4600 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4601 if (is_obja != NULL) { 4602 PreserveJVMState pjvms2(this); 4603 set_control(is_obja); 4604 // Generate a direct call to the right arraycopy function(s). 4605 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4606 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4607 ac->set_cloneoop(); 4608 Node* n = _gvn.transform(ac); 4609 assert(n == ac, "cannot disappear"); 4610 ac->connect_outputs(this); 4611 4612 result_reg->init_req(_objArray_path, control()); 4613 result_val->init_req(_objArray_path, alloc_obj); 4614 result_i_o ->set_req(_objArray_path, i_o()); 4615 result_mem ->set_req(_objArray_path, reset_memory()); 4616 } 4617 } 4618 // Otherwise, there are no card marks to worry about. 4619 // (We can dispense with card marks if we know the allocation 4620 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4621 // causes the non-eden paths to take compensating steps to 4622 // simulate a fresh allocation, so that no further 4623 // card marks are required in compiled code to initialize 4624 // the object.) 4625 4626 if (!stopped()) { 4627 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4628 4629 // Present the results of the copy. 4630 result_reg->init_req(_array_path, control()); 4631 result_val->init_req(_array_path, alloc_obj); 4632 result_i_o ->set_req(_array_path, i_o()); 4633 result_mem ->set_req(_array_path, reset_memory()); 4634 } 4635 } 4636 4637 // We only go to the instance fast case code if we pass a number of guards. 4638 // The paths which do not pass are accumulated in the slow_region. 4639 RegionNode* slow_region = new RegionNode(1); 4640 record_for_igvn(slow_region); 4641 if (!stopped()) { 4642 // It's an instance (we did array above). Make the slow-path tests. 4643 // If this is a virtual call, we generate a funny guard. We grab 4644 // the vtable entry corresponding to clone() from the target object. 4645 // If the target method which we are calling happens to be the 4646 // Object clone() method, we pass the guard. We do not need this 4647 // guard for non-virtual calls; the caller is known to be the native 4648 // Object clone(). 4649 if (is_virtual) { 4650 generate_virtual_guard(obj_klass, slow_region); 4651 } 4652 4653 // The object must be cloneable and must not have a finalizer. 4654 // Both of these conditions may be checked in a single test. 4655 // We could optimize the cloneable test further, but we don't care. 4656 generate_access_flags_guard(obj_klass, 4657 // Test both conditions: 4658 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4659 // Must be cloneable but not finalizer: 4660 JVM_ACC_IS_CLONEABLE, 4661 slow_region); 4662 } 4663 4664 if (!stopped()) { 4665 // It's an instance, and it passed the slow-path tests. 4666 PreserveJVMState pjvms(this); 4667 Node* obj_size = NULL; 4668 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4669 // is reexecuted if deoptimization occurs and there could be problems when merging 4670 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4671 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4672 4673 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4674 4675 // Present the results of the slow call. 4676 result_reg->init_req(_instance_path, control()); 4677 result_val->init_req(_instance_path, alloc_obj); 4678 result_i_o ->set_req(_instance_path, i_o()); 4679 result_mem ->set_req(_instance_path, reset_memory()); 4680 } 4681 4682 // Generate code for the slow case. We make a call to clone(). 4683 set_control(_gvn.transform(slow_region)); 4684 if (!stopped()) { 4685 PreserveJVMState pjvms(this); 4686 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4687 Node* slow_result = set_results_for_java_call(slow_call); 4688 // this->control() comes from set_results_for_java_call 4689 result_reg->init_req(_slow_path, control()); 4690 result_val->init_req(_slow_path, slow_result); 4691 result_i_o ->set_req(_slow_path, i_o()); 4692 result_mem ->set_req(_slow_path, reset_memory()); 4693 } 4694 4695 // Return the combined state. 4696 set_control( _gvn.transform(result_reg)); 4697 set_i_o( _gvn.transform(result_i_o)); 4698 set_all_memory( _gvn.transform(result_mem)); 4699 } // original reexecute is set back here 4700 4701 set_result(_gvn.transform(result_val)); 4702 return true; 4703 } 4704 4705 //------------------------------inline_arraycopy----------------------- 4706 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4707 // Object dest, int destPos, 4708 // int length); 4709 bool LibraryCallKit::inline_arraycopy() { 4710 // Get the arguments. 4711 Node* src = argument(0); // type: oop 4712 Node* src_offset = argument(1); // type: int 4713 Node* dest = argument(2); // type: oop 4714 Node* dest_offset = argument(3); // type: int 4715 Node* length = argument(4); // type: int 4716 4717 // The following tests must be performed 4718 // (1) src and dest are arrays. 4719 // (2) src and dest arrays must have elements of the same BasicType 4720 // (3) src and dest must not be null. 4721 // (4) src_offset must not be negative. 4722 // (5) dest_offset must not be negative. 4723 // (6) length must not be negative. 4724 // (7) src_offset + length must not exceed length of src. 4725 // (8) dest_offset + length must not exceed length of dest. 4726 // (9) each element of an oop array must be assignable 4727 4728 // (3) src and dest must not be null. 4729 // always do this here because we need the JVM state for uncommon traps 4730 src = null_check(src, T_ARRAY); 4731 dest = null_check(dest, T_ARRAY); 4732 4733 // Check for allocation before we add nodes that would confuse 4734 // tightly_coupled_allocation() 4735 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4736 4737 SafePointNode* sfpt = NULL; 4738 if (alloc != NULL) { 4739 // The JVM state for uncommon traps between the allocation and 4740 // arraycopy is set to the state before the allocation: if the 4741 // initialization is performed by the array copy, we don't want to 4742 // go back to the interpreter with an unitialized array. 4743 JVMState* old_jvms = alloc->jvms(); 4744 JVMState* jvms = old_jvms->clone_shallow(C); 4745 uint size = alloc->req(); 4746 sfpt = new SafePointNode(size, jvms); 4747 jvms->set_map(sfpt); 4748 for (uint i = 0; i < size; i++) { 4749 sfpt->init_req(i, alloc->in(i)); 4750 } 4751 // re-push array length for deoptimization 4752 sfpt->ins_req(jvms->stkoff() + jvms->sp(), alloc->in(AllocateNode::ALength)); 4753 jvms->set_sp(jvms->sp()+1); 4754 jvms->set_monoff(jvms->monoff()+1); 4755 jvms->set_scloff(jvms->scloff()+1); 4756 jvms->set_endoff(jvms->endoff()+1); 4757 jvms->set_should_reexecute(true); 4758 4759 sfpt->set_i_o(map()->i_o()); 4760 sfpt->set_memory(map()->memory()); 4761 } 4762 4763 bool validated = false; 4764 4765 const Type* src_type = _gvn.type(src); 4766 const Type* dest_type = _gvn.type(dest); 4767 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4768 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4769 4770 // Do we have the type of src? 4771 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4772 // Do we have the type of dest? 4773 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4774 // Is the type for src from speculation? 4775 bool src_spec = false; 4776 // Is the type for dest from speculation? 4777 bool dest_spec = false; 4778 4779 if (!has_src || !has_dest) { 4780 // We don't have sufficient type information, let's see if 4781 // speculative types can help. We need to have types for both src 4782 // and dest so that it pays off. 4783 4784 // Do we already have or could we have type information for src 4785 bool could_have_src = has_src; 4786 // Do we already have or could we have type information for dest 4787 bool could_have_dest = has_dest; 4788 4789 ciKlass* src_k = NULL; 4790 if (!has_src) { 4791 src_k = src_type->speculative_type_not_null(); 4792 if (src_k != NULL && src_k->is_array_klass()) { 4793 could_have_src = true; 4794 } 4795 } 4796 4797 ciKlass* dest_k = NULL; 4798 if (!has_dest) { 4799 dest_k = dest_type->speculative_type_not_null(); 4800 if (dest_k != NULL && dest_k->is_array_klass()) { 4801 could_have_dest = true; 4802 } 4803 } 4804 4805 if (could_have_src && could_have_dest) { 4806 // This is going to pay off so emit the required guards 4807 if (!has_src) { 4808 src = maybe_cast_profiled_obj(src, src_k, true, sfpt); 4809 src_type = _gvn.type(src); 4810 top_src = src_type->isa_aryptr(); 4811 has_src = (top_src != NULL && top_src->klass() != NULL); 4812 src_spec = true; 4813 } 4814 if (!has_dest) { 4815 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4816 dest_type = _gvn.type(dest); 4817 top_dest = dest_type->isa_aryptr(); 4818 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4819 dest_spec = true; 4820 } 4821 } 4822 } 4823 4824 if (has_src && has_dest) { 4825 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4826 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4827 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4828 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4829 4830 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4831 // If both arrays are object arrays then having the exact types 4832 // for both will remove the need for a subtype check at runtime 4833 // before the call and may make it possible to pick a faster copy 4834 // routine (without a subtype check on every element) 4835 // Do we have the exact type of src? 4836 bool could_have_src = src_spec; 4837 // Do we have the exact type of dest? 4838 bool could_have_dest = dest_spec; 4839 ciKlass* src_k = top_src->klass(); 4840 ciKlass* dest_k = top_dest->klass(); 4841 if (!src_spec) { 4842 src_k = src_type->speculative_type_not_null(); 4843 if (src_k != NULL && src_k->is_array_klass()) { 4844 could_have_src = true; 4845 } 4846 } 4847 if (!dest_spec) { 4848 dest_k = dest_type->speculative_type_not_null(); 4849 if (dest_k != NULL && dest_k->is_array_klass()) { 4850 could_have_dest = true; 4851 } 4852 } 4853 if (could_have_src && could_have_dest) { 4854 // If we can have both exact types, emit the missing guards 4855 if (could_have_src && !src_spec) { 4856 src = maybe_cast_profiled_obj(src, src_k, true, sfpt); 4857 } 4858 if (could_have_dest && !dest_spec) { 4859 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4860 } 4861 } 4862 } 4863 } 4864 4865 if (!too_many_traps(Deoptimization::Reason_intrinsic) && !src->is_top() && !dest->is_top()) { 4866 // validate arguments: enables transformation the ArrayCopyNode 4867 validated = true; 4868 4869 RegionNode* slow_region = new RegionNode(1); 4870 record_for_igvn(slow_region); 4871 4872 // (1) src and dest are arrays. 4873 generate_non_array_guard(load_object_klass(src), slow_region); 4874 generate_non_array_guard(load_object_klass(dest), slow_region); 4875 4876 // (2) src and dest arrays must have elements of the same BasicType 4877 // done at macro expansion or at Ideal transformation time 4878 4879 // (4) src_offset must not be negative. 4880 generate_negative_guard(src_offset, slow_region); 4881 4882 // (5) dest_offset must not be negative. 4883 generate_negative_guard(dest_offset, slow_region); 4884 4885 // (7) src_offset + length must not exceed length of src. 4886 generate_limit_guard(src_offset, length, 4887 load_array_length(src), 4888 slow_region); 4889 4890 // (8) dest_offset + length must not exceed length of dest. 4891 generate_limit_guard(dest_offset, length, 4892 load_array_length(dest), 4893 slow_region); 4894 4895 // (9) each element of an oop array must be assignable 4896 Node* src_klass = load_object_klass(src); 4897 Node* dest_klass = load_object_klass(dest); 4898 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4899 4900 if (not_subtype_ctrl != top()) { 4901 if (sfpt != NULL) { 4902 GraphKit kit(sfpt->jvms()); 4903 PreserveJVMState pjvms(&kit); 4904 kit.set_control(not_subtype_ctrl); 4905 kit.uncommon_trap(Deoptimization::Reason_intrinsic, 4906 Deoptimization::Action_make_not_entrant); 4907 assert(kit.stopped(), "Should be stopped"); 4908 } else { 4909 PreserveJVMState pjvms(this); 4910 set_control(not_subtype_ctrl); 4911 uncommon_trap(Deoptimization::Reason_intrinsic, 4912 Deoptimization::Action_make_not_entrant); 4913 assert(stopped(), "Should be stopped"); 4914 } 4915 } 4916 if (sfpt != NULL) { 4917 GraphKit kit(sfpt->jvms()); 4918 kit.set_control(_gvn.transform(slow_region)); 4919 kit.uncommon_trap(Deoptimization::Reason_intrinsic, 4920 Deoptimization::Action_make_not_entrant); 4921 assert(kit.stopped(), "Should be stopped"); 4922 } else { 4923 PreserveJVMState pjvms(this); 4924 set_control(_gvn.transform(slow_region)); 4925 uncommon_trap(Deoptimization::Reason_intrinsic, 4926 Deoptimization::Action_make_not_entrant); 4927 assert(stopped(), "Should be stopped"); 4928 } 4929 } 4930 4931 if (stopped()) { 4932 return true; 4933 } 4934 4935 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 4936 // Create LoadRange and LoadKlass nodes for use during macro expansion here 4937 // so the compiler has a chance to eliminate them: during macro expansion, 4938 // we have to set their control (CastPP nodes are eliminated). 4939 load_object_klass(src), load_object_klass(dest), 4940 load_array_length(src), load_array_length(dest)); 4941 4942 ac->set_arraycopy(validated); 4943 4944 Node* n = _gvn.transform(ac); 4945 if (n == ac) { 4946 ac->connect_outputs(this); 4947 } else { 4948 assert(validated, "shouldn't transform if all arguments not validated"); 4949 set_all_memory(n); 4950 } 4951 4952 return true; 4953 } 4954 4955 4956 // Helper function which determines if an arraycopy immediately follows 4957 // an allocation, with no intervening tests or other escapes for the object. 4958 AllocateArrayNode* 4959 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4960 RegionNode* slow_region) { 4961 if (stopped()) return NULL; // no fast path 4962 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4963 4964 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4965 if (alloc == NULL) return NULL; 4966 4967 Node* rawmem = memory(Compile::AliasIdxRaw); 4968 // Is the allocation's memory state untouched? 4969 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4970 // Bail out if there have been raw-memory effects since the allocation. 4971 // (Example: There might have been a call or safepoint.) 4972 return NULL; 4973 } 4974 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4975 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4976 return NULL; 4977 } 4978 4979 // There must be no unexpected observers of this allocation. 4980 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4981 Node* obs = ptr->fast_out(i); 4982 if (obs != this->map()) { 4983 return NULL; 4984 } 4985 } 4986 4987 // This arraycopy must unconditionally follow the allocation of the ptr. 4988 Node* alloc_ctl = ptr->in(0); 4989 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4990 4991 Node* ctl = control(); 4992 while (ctl != alloc_ctl) { 4993 // There may be guards which feed into the slow_region. 4994 // Any other control flow means that we might not get a chance 4995 // to finish initializing the allocated object. 4996 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4997 IfNode* iff = ctl->in(0)->as_If(); 4998 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 4999 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5000 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5001 ctl = iff->in(0); // This test feeds the known slow_region. 5002 continue; 5003 } 5004 // One more try: Various low-level checks bottom out in 5005 // uncommon traps. If the debug-info of the trap omits 5006 // any reference to the allocation, as we've already 5007 // observed, then there can be no objection to the trap. 5008 bool found_trap = false; 5009 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5010 Node* obs = not_ctl->fast_out(j); 5011 if (obs->in(0) == not_ctl && obs->is_Call() && 5012 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5013 found_trap = true; break; 5014 } 5015 } 5016 if (found_trap) { 5017 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5018 continue; 5019 } 5020 } 5021 return NULL; 5022 } 5023 5024 // If we get this far, we have an allocation which immediately 5025 // precedes the arraycopy, and we can take over zeroing the new object. 5026 // The arraycopy will finish the initialization, and provide 5027 // a new control state to which we will anchor the destination pointer. 5028 5029 return alloc; 5030 } 5031 5032 //-------------inline_encodeISOArray----------------------------------- 5033 // encode char[] to byte[] in ISO_8859_1 5034 bool LibraryCallKit::inline_encodeISOArray() { 5035 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5036 // no receiver since it is static method 5037 Node *src = argument(0); 5038 Node *src_offset = argument(1); 5039 Node *dst = argument(2); 5040 Node *dst_offset = argument(3); 5041 Node *length = argument(4); 5042 5043 const Type* src_type = src->Value(&_gvn); 5044 const Type* dst_type = dst->Value(&_gvn); 5045 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5046 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5047 if (top_src == NULL || top_src->klass() == NULL || 5048 top_dest == NULL || top_dest->klass() == NULL) { 5049 // failed array check 5050 return false; 5051 } 5052 5053 // Figure out the size and type of the elements we will be copying. 5054 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5055 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5056 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5057 return false; 5058 } 5059 Node* src_start = array_element_address(src, src_offset, src_elem); 5060 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5061 // 'src_start' points to src array + scaled offset 5062 // 'dst_start' points to dst array + scaled offset 5063 5064 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5065 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5066 enc = _gvn.transform(enc); 5067 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5068 set_memory(res_mem, mtype); 5069 set_result(enc); 5070 return true; 5071 } 5072 5073 //-------------inline_multiplyToLen----------------------------------- 5074 bool LibraryCallKit::inline_multiplyToLen() { 5075 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5076 5077 address stubAddr = StubRoutines::multiplyToLen(); 5078 if (stubAddr == NULL) { 5079 return false; // Intrinsic's stub is not implemented on this platform 5080 } 5081 const char* stubName = "multiplyToLen"; 5082 5083 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5084 5085 Node* x = argument(1); 5086 Node* xlen = argument(2); 5087 Node* y = argument(3); 5088 Node* ylen = argument(4); 5089 Node* z = argument(5); 5090 5091 const Type* x_type = x->Value(&_gvn); 5092 const Type* y_type = y->Value(&_gvn); 5093 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5094 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5095 if (top_x == NULL || top_x->klass() == NULL || 5096 top_y == NULL || top_y->klass() == NULL) { 5097 // failed array check 5098 return false; 5099 } 5100 5101 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5102 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5103 if (x_elem != T_INT || y_elem != T_INT) { 5104 return false; 5105 } 5106 5107 // Set the original stack and the reexecute bit for the interpreter to reexecute 5108 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5109 // on the return from z array allocation in runtime. 5110 { PreserveReexecuteState preexecs(this); 5111 jvms()->set_should_reexecute(true); 5112 5113 Node* x_start = array_element_address(x, intcon(0), x_elem); 5114 Node* y_start = array_element_address(y, intcon(0), y_elem); 5115 // 'x_start' points to x array + scaled xlen 5116 // 'y_start' points to y array + scaled ylen 5117 5118 // Allocate the result array 5119 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5120 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5121 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5122 5123 IdealKit ideal(this); 5124 5125 #define __ ideal. 5126 Node* one = __ ConI(1); 5127 Node* zero = __ ConI(0); 5128 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5129 __ set(need_alloc, zero); 5130 __ set(z_alloc, z); 5131 __ if_then(z, BoolTest::eq, null()); { 5132 __ increment (need_alloc, one); 5133 } __ else_(); { 5134 // Update graphKit memory and control from IdealKit. 5135 sync_kit(ideal); 5136 Node* zlen_arg = load_array_length(z); 5137 // Update IdealKit memory and control from graphKit. 5138 __ sync_kit(this); 5139 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5140 __ increment (need_alloc, one); 5141 } __ end_if(); 5142 } __ end_if(); 5143 5144 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5145 // Update graphKit memory and control from IdealKit. 5146 sync_kit(ideal); 5147 Node * narr = new_array(klass_node, zlen, 1); 5148 // Update IdealKit memory and control from graphKit. 5149 __ sync_kit(this); 5150 __ set(z_alloc, narr); 5151 } __ end_if(); 5152 5153 sync_kit(ideal); 5154 z = __ value(z_alloc); 5155 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5156 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5157 // Final sync IdealKit and GraphKit. 5158 final_sync(ideal); 5159 #undef __ 5160 5161 Node* z_start = array_element_address(z, intcon(0), T_INT); 5162 5163 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5164 OptoRuntime::multiplyToLen_Type(), 5165 stubAddr, stubName, TypePtr::BOTTOM, 5166 x_start, xlen, y_start, ylen, z_start, zlen); 5167 } // original reexecute is set back here 5168 5169 C->set_has_split_ifs(true); // Has chance for split-if optimization 5170 set_result(z); 5171 return true; 5172 } 5173 5174 5175 /** 5176 * Calculate CRC32 for byte. 5177 * int java.util.zip.CRC32.update(int crc, int b) 5178 */ 5179 bool LibraryCallKit::inline_updateCRC32() { 5180 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5181 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5182 // no receiver since it is static method 5183 Node* crc = argument(0); // type: int 5184 Node* b = argument(1); // type: int 5185 5186 /* 5187 * int c = ~ crc; 5188 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5189 * b = b ^ (c >>> 8); 5190 * crc = ~b; 5191 */ 5192 5193 Node* M1 = intcon(-1); 5194 crc = _gvn.transform(new XorINode(crc, M1)); 5195 Node* result = _gvn.transform(new XorINode(crc, b)); 5196 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5197 5198 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5199 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5200 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5201 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5202 5203 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5204 result = _gvn.transform(new XorINode(crc, result)); 5205 result = _gvn.transform(new XorINode(result, M1)); 5206 set_result(result); 5207 return true; 5208 } 5209 5210 /** 5211 * Calculate CRC32 for byte[] array. 5212 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5213 */ 5214 bool LibraryCallKit::inline_updateBytesCRC32() { 5215 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5216 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5217 // no receiver since it is static method 5218 Node* crc = argument(0); // type: int 5219 Node* src = argument(1); // type: oop 5220 Node* offset = argument(2); // type: int 5221 Node* length = argument(3); // type: int 5222 5223 const Type* src_type = src->Value(&_gvn); 5224 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5225 if (top_src == NULL || top_src->klass() == NULL) { 5226 // failed array check 5227 return false; 5228 } 5229 5230 // Figure out the size and type of the elements we will be copying. 5231 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5232 if (src_elem != T_BYTE) { 5233 return false; 5234 } 5235 5236 // 'src_start' points to src array + scaled offset 5237 Node* src_start = array_element_address(src, offset, src_elem); 5238 5239 // We assume that range check is done by caller. 5240 // TODO: generate range check (offset+length < src.length) in debug VM. 5241 5242 // Call the stub. 5243 address stubAddr = StubRoutines::updateBytesCRC32(); 5244 const char *stubName = "updateBytesCRC32"; 5245 5246 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5247 stubAddr, stubName, TypePtr::BOTTOM, 5248 crc, src_start, length); 5249 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5250 set_result(result); 5251 return true; 5252 } 5253 5254 /** 5255 * Calculate CRC32 for ByteBuffer. 5256 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5257 */ 5258 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5259 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5260 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5261 // no receiver since it is static method 5262 Node* crc = argument(0); // type: int 5263 Node* src = argument(1); // type: long 5264 Node* offset = argument(3); // type: int 5265 Node* length = argument(4); // type: int 5266 5267 src = ConvL2X(src); // adjust Java long to machine word 5268 Node* base = _gvn.transform(new CastX2PNode(src)); 5269 offset = ConvI2X(offset); 5270 5271 // 'src_start' points to src array + scaled offset 5272 Node* src_start = basic_plus_adr(top(), base, offset); 5273 5274 // Call the stub. 5275 address stubAddr = StubRoutines::updateBytesCRC32(); 5276 const char *stubName = "updateBytesCRC32"; 5277 5278 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5279 stubAddr, stubName, TypePtr::BOTTOM, 5280 crc, src_start, length); 5281 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5282 set_result(result); 5283 return true; 5284 } 5285 5286 //----------------------------inline_reference_get---------------------------- 5287 // public T java.lang.ref.Reference.get(); 5288 bool LibraryCallKit::inline_reference_get() { 5289 const int referent_offset = java_lang_ref_Reference::referent_offset; 5290 guarantee(referent_offset > 0, "should have already been set"); 5291 5292 // Get the argument: 5293 Node* reference_obj = null_check_receiver(); 5294 if (stopped()) return true; 5295 5296 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5297 5298 ciInstanceKlass* klass = env()->Object_klass(); 5299 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5300 5301 Node* no_ctrl = NULL; 5302 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5303 5304 // Use the pre-barrier to record the value in the referent field 5305 pre_barrier(false /* do_load */, 5306 control(), 5307 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5308 result /* pre_val */, 5309 T_OBJECT); 5310 5311 // Add memory barrier to prevent commoning reads from this field 5312 // across safepoint since GC can change its value. 5313 insert_mem_bar(Op_MemBarCPUOrder); 5314 5315 set_result(result); 5316 return true; 5317 } 5318 5319 5320 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5321 bool is_exact=true, bool is_static=false) { 5322 5323 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5324 assert(tinst != NULL, "obj is null"); 5325 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5326 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5327 5328 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5329 ciSymbol::make(fieldTypeString), 5330 is_static); 5331 if (field == NULL) return (Node *) NULL; 5332 assert (field != NULL, "undefined field"); 5333 5334 // Next code copied from Parse::do_get_xxx(): 5335 5336 // Compute address and memory type. 5337 int offset = field->offset_in_bytes(); 5338 bool is_vol = field->is_volatile(); 5339 ciType* field_klass = field->type(); 5340 assert(field_klass->is_loaded(), "should be loaded"); 5341 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5342 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5343 BasicType bt = field->layout_type(); 5344 5345 // Build the resultant type of the load 5346 const Type *type; 5347 if (bt == T_OBJECT) { 5348 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5349 } else { 5350 type = Type::get_const_basic_type(bt); 5351 } 5352 5353 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5354 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5355 } 5356 // Build the load. 5357 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5358 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol); 5359 // If reference is volatile, prevent following memory ops from 5360 // floating up past the volatile read. Also prevents commoning 5361 // another volatile read. 5362 if (is_vol) { 5363 // Memory barrier includes bogus read of value to force load BEFORE membar 5364 insert_mem_bar(Op_MemBarAcquire, loadedField); 5365 } 5366 return loadedField; 5367 } 5368 5369 5370 //------------------------------inline_aescrypt_Block----------------------- 5371 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5372 address stubAddr; 5373 const char *stubName; 5374 assert(UseAES, "need AES instruction support"); 5375 5376 switch(id) { 5377 case vmIntrinsics::_aescrypt_encryptBlock: 5378 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5379 stubName = "aescrypt_encryptBlock"; 5380 break; 5381 case vmIntrinsics::_aescrypt_decryptBlock: 5382 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5383 stubName = "aescrypt_decryptBlock"; 5384 break; 5385 } 5386 if (stubAddr == NULL) return false; 5387 5388 Node* aescrypt_object = argument(0); 5389 Node* src = argument(1); 5390 Node* src_offset = argument(2); 5391 Node* dest = argument(3); 5392 Node* dest_offset = argument(4); 5393 5394 // (1) src and dest are arrays. 5395 const Type* src_type = src->Value(&_gvn); 5396 const Type* dest_type = dest->Value(&_gvn); 5397 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5398 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5399 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5400 5401 // for the quick and dirty code we will skip all the checks. 5402 // we are just trying to get the call to be generated. 5403 Node* src_start = src; 5404 Node* dest_start = dest; 5405 if (src_offset != NULL || dest_offset != NULL) { 5406 assert(src_offset != NULL && dest_offset != NULL, ""); 5407 src_start = array_element_address(src, src_offset, T_BYTE); 5408 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5409 } 5410 5411 // now need to get the start of its expanded key array 5412 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5413 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5414 if (k_start == NULL) return false; 5415 5416 if (Matcher::pass_original_key_for_aes()) { 5417 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5418 // compatibility issues between Java key expansion and SPARC crypto instructions 5419 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5420 if (original_k_start == NULL) return false; 5421 5422 // Call the stub. 5423 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5424 stubAddr, stubName, TypePtr::BOTTOM, 5425 src_start, dest_start, k_start, original_k_start); 5426 } else { 5427 // Call the stub. 5428 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5429 stubAddr, stubName, TypePtr::BOTTOM, 5430 src_start, dest_start, k_start); 5431 } 5432 5433 return true; 5434 } 5435 5436 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5437 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5438 address stubAddr; 5439 const char *stubName; 5440 5441 assert(UseAES, "need AES instruction support"); 5442 5443 switch(id) { 5444 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5445 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5446 stubName = "cipherBlockChaining_encryptAESCrypt"; 5447 break; 5448 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5449 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5450 stubName = "cipherBlockChaining_decryptAESCrypt"; 5451 break; 5452 } 5453 if (stubAddr == NULL) return false; 5454 5455 Node* cipherBlockChaining_object = argument(0); 5456 Node* src = argument(1); 5457 Node* src_offset = argument(2); 5458 Node* len = argument(3); 5459 Node* dest = argument(4); 5460 Node* dest_offset = argument(5); 5461 5462 // (1) src and dest are arrays. 5463 const Type* src_type = src->Value(&_gvn); 5464 const Type* dest_type = dest->Value(&_gvn); 5465 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5466 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5467 assert (top_src != NULL && top_src->klass() != NULL 5468 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5469 5470 // checks are the responsibility of the caller 5471 Node* src_start = src; 5472 Node* dest_start = dest; 5473 if (src_offset != NULL || dest_offset != NULL) { 5474 assert(src_offset != NULL && dest_offset != NULL, ""); 5475 src_start = array_element_address(src, src_offset, T_BYTE); 5476 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5477 } 5478 5479 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5480 // (because of the predicated logic executed earlier). 5481 // so we cast it here safely. 5482 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5483 5484 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5485 if (embeddedCipherObj == NULL) return false; 5486 5487 // cast it to what we know it will be at runtime 5488 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5489 assert(tinst != NULL, "CBC obj is null"); 5490 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5491 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5492 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5493 5494 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5495 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5496 const TypeOopPtr* xtype = aklass->as_instance_type(); 5497 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5498 aescrypt_object = _gvn.transform(aescrypt_object); 5499 5500 // we need to get the start of the aescrypt_object's expanded key array 5501 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5502 if (k_start == NULL) return false; 5503 5504 // similarly, get the start address of the r vector 5505 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5506 if (objRvec == NULL) return false; 5507 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5508 5509 Node* cbcCrypt; 5510 if (Matcher::pass_original_key_for_aes()) { 5511 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5512 // compatibility issues between Java key expansion and SPARC crypto instructions 5513 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5514 if (original_k_start == NULL) return false; 5515 5516 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5517 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5518 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5519 stubAddr, stubName, TypePtr::BOTTOM, 5520 src_start, dest_start, k_start, r_start, len, original_k_start); 5521 } else { 5522 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5523 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5524 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5525 stubAddr, stubName, TypePtr::BOTTOM, 5526 src_start, dest_start, k_start, r_start, len); 5527 } 5528 5529 // return cipher length (int) 5530 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5531 set_result(retvalue); 5532 return true; 5533 } 5534 5535 //------------------------------get_key_start_from_aescrypt_object----------------------- 5536 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5537 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5538 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5539 if (objAESCryptKey == NULL) return (Node *) NULL; 5540 5541 // now have the array, need to get the start address of the K array 5542 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5543 return k_start; 5544 } 5545 5546 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5547 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5548 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5549 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5550 if (objAESCryptKey == NULL) return (Node *) NULL; 5551 5552 // now have the array, need to get the start address of the lastKey array 5553 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5554 return original_k_start; 5555 } 5556 5557 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5558 // Return node representing slow path of predicate check. 5559 // the pseudo code we want to emulate with this predicate is: 5560 // for encryption: 5561 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5562 // for decryption: 5563 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5564 // note cipher==plain is more conservative than the original java code but that's OK 5565 // 5566 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5567 // The receiver was checked for NULL already. 5568 Node* objCBC = argument(0); 5569 5570 // Load embeddedCipher field of CipherBlockChaining object. 5571 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5572 5573 // get AESCrypt klass for instanceOf check 5574 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5575 // will have same classloader as CipherBlockChaining object 5576 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5577 assert(tinst != NULL, "CBCobj is null"); 5578 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5579 5580 // we want to do an instanceof comparison against the AESCrypt class 5581 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5582 if (!klass_AESCrypt->is_loaded()) { 5583 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5584 Node* ctrl = control(); 5585 set_control(top()); // no regular fast path 5586 return ctrl; 5587 } 5588 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5589 5590 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5591 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5592 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5593 5594 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5595 5596 // for encryption, we are done 5597 if (!decrypting) 5598 return instof_false; // even if it is NULL 5599 5600 // for decryption, we need to add a further check to avoid 5601 // taking the intrinsic path when cipher and plain are the same 5602 // see the original java code for why. 5603 RegionNode* region = new RegionNode(3); 5604 region->init_req(1, instof_false); 5605 Node* src = argument(1); 5606 Node* dest = argument(4); 5607 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5608 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5609 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5610 region->init_req(2, src_dest_conjoint); 5611 5612 record_for_igvn(region); 5613 return _gvn.transform(region); 5614 } 5615 5616 //------------------------------inline_sha_implCompress----------------------- 5617 // 5618 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 5619 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 5620 // 5621 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 5622 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 5623 // 5624 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 5625 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 5626 // 5627 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 5628 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 5629 5630 Node* sha_obj = argument(0); 5631 Node* src = argument(1); // type oop 5632 Node* ofs = argument(2); // type int 5633 5634 const Type* src_type = src->Value(&_gvn); 5635 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5636 if (top_src == NULL || top_src->klass() == NULL) { 5637 // failed array check 5638 return false; 5639 } 5640 // Figure out the size and type of the elements we will be copying. 5641 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5642 if (src_elem != T_BYTE) { 5643 return false; 5644 } 5645 // 'src_start' points to src array + offset 5646 Node* src_start = array_element_address(src, ofs, src_elem); 5647 Node* state = NULL; 5648 address stubAddr; 5649 const char *stubName; 5650 5651 switch(id) { 5652 case vmIntrinsics::_sha_implCompress: 5653 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 5654 state = get_state_from_sha_object(sha_obj); 5655 stubAddr = StubRoutines::sha1_implCompress(); 5656 stubName = "sha1_implCompress"; 5657 break; 5658 case vmIntrinsics::_sha2_implCompress: 5659 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 5660 state = get_state_from_sha_object(sha_obj); 5661 stubAddr = StubRoutines::sha256_implCompress(); 5662 stubName = "sha256_implCompress"; 5663 break; 5664 case vmIntrinsics::_sha5_implCompress: 5665 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 5666 state = get_state_from_sha5_object(sha_obj); 5667 stubAddr = StubRoutines::sha512_implCompress(); 5668 stubName = "sha512_implCompress"; 5669 break; 5670 default: 5671 fatal_unexpected_iid(id); 5672 return false; 5673 } 5674 if (state == NULL) return false; 5675 5676 // Call the stub. 5677 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 5678 stubAddr, stubName, TypePtr::BOTTOM, 5679 src_start, state); 5680 5681 return true; 5682 } 5683 5684 //------------------------------inline_digestBase_implCompressMB----------------------- 5685 // 5686 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 5687 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 5688 // 5689 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 5690 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5691 "need SHA1/SHA256/SHA512 instruction support"); 5692 assert((uint)predicate < 3, "sanity"); 5693 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 5694 5695 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 5696 Node* src = argument(1); // byte[] array 5697 Node* ofs = argument(2); // type int 5698 Node* limit = argument(3); // type int 5699 5700 const Type* src_type = src->Value(&_gvn); 5701 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5702 if (top_src == NULL || top_src->klass() == NULL) { 5703 // failed array check 5704 return false; 5705 } 5706 // Figure out the size and type of the elements we will be copying. 5707 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5708 if (src_elem != T_BYTE) { 5709 return false; 5710 } 5711 // 'src_start' points to src array + offset 5712 Node* src_start = array_element_address(src, ofs, src_elem); 5713 5714 const char* klass_SHA_name = NULL; 5715 const char* stub_name = NULL; 5716 address stub_addr = NULL; 5717 bool long_state = false; 5718 5719 switch (predicate) { 5720 case 0: 5721 if (UseSHA1Intrinsics) { 5722 klass_SHA_name = "sun/security/provider/SHA"; 5723 stub_name = "sha1_implCompressMB"; 5724 stub_addr = StubRoutines::sha1_implCompressMB(); 5725 } 5726 break; 5727 case 1: 5728 if (UseSHA256Intrinsics) { 5729 klass_SHA_name = "sun/security/provider/SHA2"; 5730 stub_name = "sha256_implCompressMB"; 5731 stub_addr = StubRoutines::sha256_implCompressMB(); 5732 } 5733 break; 5734 case 2: 5735 if (UseSHA512Intrinsics) { 5736 klass_SHA_name = "sun/security/provider/SHA5"; 5737 stub_name = "sha512_implCompressMB"; 5738 stub_addr = StubRoutines::sha512_implCompressMB(); 5739 long_state = true; 5740 } 5741 break; 5742 default: 5743 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5744 } 5745 if (klass_SHA_name != NULL) { 5746 // get DigestBase klass to lookup for SHA klass 5747 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 5748 assert(tinst != NULL, "digestBase_obj is not instance???"); 5749 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5750 5751 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5752 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 5753 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5754 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 5755 } 5756 return false; 5757 } 5758 //------------------------------inline_sha_implCompressMB----------------------- 5759 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 5760 bool long_state, address stubAddr, const char *stubName, 5761 Node* src_start, Node* ofs, Node* limit) { 5762 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 5763 const TypeOopPtr* xtype = aklass->as_instance_type(); 5764 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 5765 sha_obj = _gvn.transform(sha_obj); 5766 5767 Node* state; 5768 if (long_state) { 5769 state = get_state_from_sha5_object(sha_obj); 5770 } else { 5771 state = get_state_from_sha_object(sha_obj); 5772 } 5773 if (state == NULL) return false; 5774 5775 // Call the stub. 5776 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5777 OptoRuntime::digestBase_implCompressMB_Type(), 5778 stubAddr, stubName, TypePtr::BOTTOM, 5779 src_start, state, ofs, limit); 5780 // return ofs (int) 5781 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5782 set_result(result); 5783 5784 return true; 5785 } 5786 5787 //------------------------------get_state_from_sha_object----------------------- 5788 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 5789 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 5790 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 5791 if (sha_state == NULL) return (Node *) NULL; 5792 5793 // now have the array, need to get the start address of the state array 5794 Node* state = array_element_address(sha_state, intcon(0), T_INT); 5795 return state; 5796 } 5797 5798 //------------------------------get_state_from_sha5_object----------------------- 5799 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 5800 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 5801 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 5802 if (sha_state == NULL) return (Node *) NULL; 5803 5804 // now have the array, need to get the start address of the state array 5805 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 5806 return state; 5807 } 5808 5809 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 5810 // Return node representing slow path of predicate check. 5811 // the pseudo code we want to emulate with this predicate is: 5812 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 5813 // 5814 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 5815 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5816 "need SHA1/SHA256/SHA512 instruction support"); 5817 assert((uint)predicate < 3, "sanity"); 5818 5819 // The receiver was checked for NULL already. 5820 Node* digestBaseObj = argument(0); 5821 5822 // get DigestBase klass for instanceOf check 5823 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 5824 assert(tinst != NULL, "digestBaseObj is null"); 5825 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5826 5827 const char* klass_SHA_name = NULL; 5828 switch (predicate) { 5829 case 0: 5830 if (UseSHA1Intrinsics) { 5831 // we want to do an instanceof comparison against the SHA class 5832 klass_SHA_name = "sun/security/provider/SHA"; 5833 } 5834 break; 5835 case 1: 5836 if (UseSHA256Intrinsics) { 5837 // we want to do an instanceof comparison against the SHA2 class 5838 klass_SHA_name = "sun/security/provider/SHA2"; 5839 } 5840 break; 5841 case 2: 5842 if (UseSHA512Intrinsics) { 5843 // we want to do an instanceof comparison against the SHA5 class 5844 klass_SHA_name = "sun/security/provider/SHA5"; 5845 } 5846 break; 5847 default: 5848 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5849 } 5850 5851 ciKlass* klass_SHA = NULL; 5852 if (klass_SHA_name != NULL) { 5853 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5854 } 5855 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 5856 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 5857 Node* ctrl = control(); 5858 set_control(top()); // no intrinsic path 5859 return ctrl; 5860 } 5861 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5862 5863 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 5864 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 5865 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5866 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5867 5868 return instof_false; // even if it is NULL 5869 }