1 /* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "oops/objArrayKlass.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/callGenerator.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/convertnode.hpp" 38 #include "opto/countbitsnode.hpp" 39 #include "opto/intrinsicnode.hpp" 40 #include "opto/idealKit.hpp" 41 #include "opto/mathexactnode.hpp" 42 #include "opto/movenode.hpp" 43 #include "opto/mulnode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/opaquenode.hpp" 46 #include "opto/parse.hpp" 47 #include "opto/runtime.hpp" 48 #include "opto/subnode.hpp" 49 #include "prims/nativeLookup.hpp" 50 #include "runtime/sharedRuntime.hpp" 51 #include "trace/traceMacros.hpp" 52 53 class LibraryIntrinsic : public InlineCallGenerator { 54 // Extend the set of intrinsics known to the runtime: 55 public: 56 private: 57 bool _is_virtual; 58 bool _does_virtual_dispatch; 59 int8_t _predicates_count; // Intrinsic is predicated by several conditions 60 int8_t _last_predicate; // Last generated predicate 61 vmIntrinsics::ID _intrinsic_id; 62 63 public: 64 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 65 : InlineCallGenerator(m), 66 _is_virtual(is_virtual), 67 _does_virtual_dispatch(does_virtual_dispatch), 68 _predicates_count((int8_t)predicates_count), 69 _last_predicate((int8_t)-1), 70 _intrinsic_id(id) 71 { 72 } 73 virtual bool is_intrinsic() const { return true; } 74 virtual bool is_virtual() const { return _is_virtual; } 75 virtual bool is_predicated() const { return _predicates_count > 0; } 76 virtual int predicates_count() const { return _predicates_count; } 77 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 78 virtual JVMState* generate(JVMState* jvms); 79 virtual Node* generate_predicate(JVMState* jvms, int predicate); 80 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 81 }; 82 83 84 // Local helper class for LibraryIntrinsic: 85 class LibraryCallKit : public GraphKit { 86 private: 87 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 88 Node* _result; // the result node, if any 89 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 90 91 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 92 93 public: 94 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 95 : GraphKit(jvms), 96 _intrinsic(intrinsic), 97 _result(NULL) 98 { 99 // Check if this is a root compile. In that case we don't have a caller. 100 if (!jvms->has_method()) { 101 _reexecute_sp = sp(); 102 } else { 103 // Find out how many arguments the interpreter needs when deoptimizing 104 // and save the stack pointer value so it can used by uncommon_trap. 105 // We find the argument count by looking at the declared signature. 106 bool ignored_will_link; 107 ciSignature* declared_signature = NULL; 108 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 109 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 110 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 111 } 112 } 113 114 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 115 116 ciMethod* caller() const { return jvms()->method(); } 117 int bci() const { return jvms()->bci(); } 118 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 119 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 120 ciMethod* callee() const { return _intrinsic->method(); } 121 122 bool try_to_inline(int predicate); 123 Node* try_to_predicate(int predicate); 124 125 void push_result() { 126 // Push the result onto the stack. 127 if (!stopped() && result() != NULL) { 128 BasicType bt = result()->bottom_type()->basic_type(); 129 push_node(bt, result()); 130 } 131 } 132 133 private: 134 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 135 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 136 } 137 138 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 139 void set_result(RegionNode* region, PhiNode* value); 140 Node* result() { return _result; } 141 142 virtual int reexecute_sp() { return _reexecute_sp; } 143 144 // Helper functions to inline natives 145 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 146 Node* generate_slow_guard(Node* test, RegionNode* region); 147 Node* generate_fair_guard(Node* test, RegionNode* region); 148 Node* generate_negative_guard(Node* index, RegionNode* region, 149 // resulting CastII of index: 150 Node* *pos_index = NULL); 151 Node* generate_limit_guard(Node* offset, Node* subseq_length, 152 Node* array_length, 153 RegionNode* region); 154 Node* generate_current_thread(Node* &tls_output); 155 Node* load_mirror_from_klass(Node* klass); 156 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 157 RegionNode* region, int null_path, 158 int offset); 159 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path) { 161 int offset = java_lang_Class::klass_offset_in_bytes(); 162 return load_klass_from_mirror_common(mirror, never_see_null, 163 region, null_path, 164 offset); 165 } 166 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 167 RegionNode* region, int null_path) { 168 int offset = java_lang_Class::array_klass_offset_in_bytes(); 169 return load_klass_from_mirror_common(mirror, never_see_null, 170 region, null_path, 171 offset); 172 } 173 Node* generate_access_flags_guard(Node* kls, 174 int modifier_mask, int modifier_bits, 175 RegionNode* region); 176 Node* generate_interface_guard(Node* kls, RegionNode* region); 177 Node* generate_array_guard(Node* kls, RegionNode* region) { 178 return generate_array_guard_common(kls, region, false, false); 179 } 180 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 181 return generate_array_guard_common(kls, region, false, true); 182 } 183 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 184 return generate_array_guard_common(kls, region, true, false); 185 } 186 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 187 return generate_array_guard_common(kls, region, true, true); 188 } 189 Node* generate_array_guard_common(Node* kls, RegionNode* region, 190 bool obj_array, bool not_array); 191 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 192 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 193 bool is_virtual = false, bool is_static = false); 194 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 195 return generate_method_call(method_id, false, true); 196 } 197 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 198 return generate_method_call(method_id, true, false); 199 } 200 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 201 202 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 203 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 204 bool inline_string_compareTo(); 205 bool inline_string_indexOf(); 206 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 207 bool inline_string_equals(); 208 Node* round_double_node(Node* n); 209 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 210 bool inline_math_native(vmIntrinsics::ID id); 211 bool inline_trig(vmIntrinsics::ID id); 212 bool inline_math(vmIntrinsics::ID id); 213 template <typename OverflowOp> 214 bool inline_math_overflow(Node* arg1, Node* arg2); 215 void inline_math_mathExact(Node* math, Node* test); 216 bool inline_math_addExactI(bool is_increment); 217 bool inline_math_addExactL(bool is_increment); 218 bool inline_math_multiplyExactI(); 219 bool inline_math_multiplyExactL(); 220 bool inline_math_negateExactI(); 221 bool inline_math_negateExactL(); 222 bool inline_math_subtractExactI(bool is_decrement); 223 bool inline_math_subtractExactL(bool is_decrement); 224 bool inline_exp(); 225 bool inline_pow(); 226 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 227 bool inline_min_max(vmIntrinsics::ID id); 228 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 229 // This returns Type::AnyPtr, RawPtr, or OopPtr. 230 int classify_unsafe_addr(Node* &base, Node* &offset); 231 Node* make_unsafe_address(Node* base, Node* offset); 232 // Helper for inline_unsafe_access. 233 // Generates the guards that check whether the result of 234 // Unsafe.getObject should be recorded in an SATB log buffer. 235 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 236 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 237 static bool klass_needs_init_guard(Node* kls); 238 bool inline_unsafe_allocate(); 239 bool inline_unsafe_copyMemory(); 240 bool inline_native_currentThread(); 241 #ifdef TRACE_HAVE_INTRINSICS 242 bool inline_native_classID(); 243 bool inline_native_threadID(); 244 #endif 245 bool inline_native_time_funcs(address method, const char* funcName); 246 bool inline_native_isInterrupted(); 247 bool inline_native_Class_query(vmIntrinsics::ID id); 248 bool inline_native_subtype_check(); 249 250 bool inline_native_newArray(); 251 bool inline_native_getLength(); 252 bool inline_array_copyOf(bool is_copyOfRange); 253 bool inline_array_equals(); 254 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 255 bool inline_native_clone(bool is_virtual); 256 bool inline_native_Reflection_getCallerClass(); 257 // Helper function for inlining native object hash method 258 bool inline_native_hashcode(bool is_virtual, bool is_static); 259 bool inline_native_getClass(); 260 261 // Helper functions for inlining arraycopy 262 bool inline_arraycopy(); 263 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 264 RegionNode* slow_region); 265 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 266 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 267 268 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 269 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 270 bool inline_unsafe_ordered_store(BasicType type); 271 bool inline_unsafe_fence(vmIntrinsics::ID id); 272 bool inline_fp_conversions(vmIntrinsics::ID id); 273 bool inline_number_methods(vmIntrinsics::ID id); 274 bool inline_reference_get(); 275 bool inline_Class_cast(); 276 bool inline_aescrypt_Block(vmIntrinsics::ID id); 277 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 278 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 279 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 280 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 281 bool inline_sha_implCompress(vmIntrinsics::ID id); 282 bool inline_digestBase_implCompressMB(int predicate); 283 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 284 bool long_state, address stubAddr, const char *stubName, 285 Node* src_start, Node* ofs, Node* limit); 286 Node* get_state_from_sha_object(Node *sha_object); 287 Node* get_state_from_sha5_object(Node *sha_object); 288 Node* inline_digestBase_implCompressMB_predicate(int predicate); 289 bool inline_encodeISOArray(); 290 bool inline_updateCRC32(); 291 bool inline_updateBytesCRC32(); 292 bool inline_updateByteBufferCRC32(); 293 bool inline_multiplyToLen(); 294 295 bool inline_profileBoolean(); 296 }; 297 298 299 //---------------------------make_vm_intrinsic---------------------------- 300 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 301 vmIntrinsics::ID id = m->intrinsic_id(); 302 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 303 304 ccstr disable_intr = NULL; 305 306 if ((DisableIntrinsic[0] != '\0' 307 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 308 (method_has_option_value("DisableIntrinsic", disable_intr) 309 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 310 // disabled by a user request on the command line: 311 // example: -XX:DisableIntrinsic=_hashCode,_getClass 312 return NULL; 313 } 314 315 if (!m->is_loaded()) { 316 // do not attempt to inline unloaded methods 317 return NULL; 318 } 319 320 // Only a few intrinsics implement a virtual dispatch. 321 // They are expensive calls which are also frequently overridden. 322 if (is_virtual) { 323 switch (id) { 324 case vmIntrinsics::_hashCode: 325 case vmIntrinsics::_clone: 326 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 327 break; 328 default: 329 return NULL; 330 } 331 } 332 333 // -XX:-InlineNatives disables nearly all intrinsics: 334 if (!InlineNatives) { 335 switch (id) { 336 case vmIntrinsics::_indexOf: 337 case vmIntrinsics::_compareTo: 338 case vmIntrinsics::_equals: 339 case vmIntrinsics::_equalsC: 340 case vmIntrinsics::_getAndAddInt: 341 case vmIntrinsics::_getAndAddLong: 342 case vmIntrinsics::_getAndSetInt: 343 case vmIntrinsics::_getAndSetLong: 344 case vmIntrinsics::_getAndSetObject: 345 case vmIntrinsics::_loadFence: 346 case vmIntrinsics::_storeFence: 347 case vmIntrinsics::_fullFence: 348 break; // InlineNatives does not control String.compareTo 349 case vmIntrinsics::_Reference_get: 350 break; // InlineNatives does not control Reference.get 351 default: 352 return NULL; 353 } 354 } 355 356 int predicates = 0; 357 bool does_virtual_dispatch = false; 358 359 switch (id) { 360 case vmIntrinsics::_compareTo: 361 if (!SpecialStringCompareTo) return NULL; 362 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 363 break; 364 case vmIntrinsics::_indexOf: 365 if (!SpecialStringIndexOf) return NULL; 366 break; 367 case vmIntrinsics::_equals: 368 if (!SpecialStringEquals) return NULL; 369 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 370 break; 371 case vmIntrinsics::_equalsC: 372 if (!SpecialArraysEquals) return NULL; 373 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 374 break; 375 case vmIntrinsics::_arraycopy: 376 if (!InlineArrayCopy) return NULL; 377 break; 378 case vmIntrinsics::_copyMemory: 379 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 380 if (!InlineArrayCopy) return NULL; 381 break; 382 case vmIntrinsics::_hashCode: 383 if (!InlineObjectHash) return NULL; 384 does_virtual_dispatch = true; 385 break; 386 case vmIntrinsics::_clone: 387 does_virtual_dispatch = true; 388 case vmIntrinsics::_copyOf: 389 case vmIntrinsics::_copyOfRange: 390 if (!InlineObjectCopy) return NULL; 391 // These also use the arraycopy intrinsic mechanism: 392 if (!InlineArrayCopy) return NULL; 393 break; 394 case vmIntrinsics::_encodeISOArray: 395 if (!SpecialEncodeISOArray) return NULL; 396 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 397 break; 398 case vmIntrinsics::_checkIndex: 399 // We do not intrinsify this. The optimizer does fine with it. 400 return NULL; 401 402 case vmIntrinsics::_getCallerClass: 403 if (!InlineReflectionGetCallerClass) return NULL; 404 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 405 break; 406 407 case vmIntrinsics::_bitCount_i: 408 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 409 break; 410 411 case vmIntrinsics::_bitCount_l: 412 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 413 break; 414 415 case vmIntrinsics::_numberOfLeadingZeros_i: 416 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 417 break; 418 419 case vmIntrinsics::_numberOfLeadingZeros_l: 420 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 421 break; 422 423 case vmIntrinsics::_numberOfTrailingZeros_i: 424 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 425 break; 426 427 case vmIntrinsics::_numberOfTrailingZeros_l: 428 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 429 break; 430 431 case vmIntrinsics::_reverseBytes_c: 432 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 433 break; 434 case vmIntrinsics::_reverseBytes_s: 435 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 436 break; 437 case vmIntrinsics::_reverseBytes_i: 438 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 439 break; 440 case vmIntrinsics::_reverseBytes_l: 441 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 442 break; 443 444 case vmIntrinsics::_Reference_get: 445 // Use the intrinsic version of Reference.get() so that the value in 446 // the referent field can be registered by the G1 pre-barrier code. 447 // Also add memory barrier to prevent commoning reads from this field 448 // across safepoint since GC can change it value. 449 break; 450 451 case vmIntrinsics::_compareAndSwapObject: 452 #ifdef _LP64 453 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 454 #endif 455 break; 456 457 case vmIntrinsics::_compareAndSwapLong: 458 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 459 break; 460 461 case vmIntrinsics::_getAndAddInt: 462 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 463 break; 464 465 case vmIntrinsics::_getAndAddLong: 466 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 467 break; 468 469 case vmIntrinsics::_getAndSetInt: 470 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 471 break; 472 473 case vmIntrinsics::_getAndSetLong: 474 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 475 break; 476 477 case vmIntrinsics::_getAndSetObject: 478 #ifdef _LP64 479 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 480 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 481 break; 482 #else 483 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 484 break; 485 #endif 486 487 case vmIntrinsics::_aescrypt_encryptBlock: 488 case vmIntrinsics::_aescrypt_decryptBlock: 489 if (!UseAESIntrinsics) return NULL; 490 break; 491 492 case vmIntrinsics::_multiplyToLen: 493 if (!UseMultiplyToLenIntrinsic) return NULL; 494 break; 495 496 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 497 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 498 if (!UseAESIntrinsics) return NULL; 499 // these two require the predicated logic 500 predicates = 1; 501 break; 502 503 case vmIntrinsics::_sha_implCompress: 504 if (!UseSHA1Intrinsics) return NULL; 505 break; 506 507 case vmIntrinsics::_sha2_implCompress: 508 if (!UseSHA256Intrinsics) return NULL; 509 break; 510 511 case vmIntrinsics::_sha5_implCompress: 512 if (!UseSHA512Intrinsics) return NULL; 513 break; 514 515 case vmIntrinsics::_digestBase_implCompressMB: 516 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 517 predicates = 3; 518 break; 519 520 case vmIntrinsics::_updateCRC32: 521 case vmIntrinsics::_updateBytesCRC32: 522 case vmIntrinsics::_updateByteBufferCRC32: 523 if (!UseCRC32Intrinsics) return NULL; 524 break; 525 526 case vmIntrinsics::_incrementExactI: 527 case vmIntrinsics::_addExactI: 528 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 529 break; 530 case vmIntrinsics::_incrementExactL: 531 case vmIntrinsics::_addExactL: 532 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 533 break; 534 case vmIntrinsics::_decrementExactI: 535 case vmIntrinsics::_subtractExactI: 536 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 537 break; 538 case vmIntrinsics::_decrementExactL: 539 case vmIntrinsics::_subtractExactL: 540 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 541 break; 542 case vmIntrinsics::_negateExactI: 543 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 544 break; 545 case vmIntrinsics::_negateExactL: 546 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 547 break; 548 case vmIntrinsics::_multiplyExactI: 549 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 550 break; 551 case vmIntrinsics::_multiplyExactL: 552 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 553 break; 554 555 default: 556 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 557 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 558 break; 559 } 560 561 // -XX:-InlineClassNatives disables natives from the Class class. 562 // The flag applies to all reflective calls, notably Array.newArray 563 // (visible to Java programmers as Array.newInstance). 564 if (m->holder()->name() == ciSymbol::java_lang_Class() || 565 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 566 if (!InlineClassNatives) return NULL; 567 } 568 569 // -XX:-InlineThreadNatives disables natives from the Thread class. 570 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 571 if (!InlineThreadNatives) return NULL; 572 } 573 574 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 575 if (m->holder()->name() == ciSymbol::java_lang_Math() || 576 m->holder()->name() == ciSymbol::java_lang_Float() || 577 m->holder()->name() == ciSymbol::java_lang_Double()) { 578 if (!InlineMathNatives) return NULL; 579 } 580 581 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 582 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 583 if (!InlineUnsafeOps) return NULL; 584 } 585 586 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 587 } 588 589 //----------------------register_library_intrinsics----------------------- 590 // Initialize this file's data structures, for each Compile instance. 591 void Compile::register_library_intrinsics() { 592 // Nothing to do here. 593 } 594 595 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 596 LibraryCallKit kit(jvms, this); 597 Compile* C = kit.C; 598 int nodes = C->unique(); 599 #ifndef PRODUCT 600 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 601 char buf[1000]; 602 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 603 tty->print_cr("Intrinsic %s", str); 604 } 605 #endif 606 ciMethod* callee = kit.callee(); 607 const int bci = kit.bci(); 608 609 // Try to inline the intrinsic. 610 if (kit.try_to_inline(_last_predicate)) { 611 if (C->print_intrinsics() || C->print_inlining()) { 612 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 613 } 614 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 615 if (C->log()) { 616 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 617 vmIntrinsics::name_at(intrinsic_id()), 618 (is_virtual() ? " virtual='1'" : ""), 619 C->unique() - nodes); 620 } 621 // Push the result from the inlined method onto the stack. 622 kit.push_result(); 623 C->print_inlining_update(this); 624 return kit.transfer_exceptions_into_jvms(); 625 } 626 627 // The intrinsic bailed out 628 if (C->print_intrinsics() || C->print_inlining()) { 629 if (jvms->has_method()) { 630 // Not a root compile. 631 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 632 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 633 } else { 634 // Root compile 635 tty->print("Did not generate intrinsic %s%s at bci:%d in", 636 vmIntrinsics::name_at(intrinsic_id()), 637 (is_virtual() ? " (virtual)" : ""), bci); 638 } 639 } 640 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 641 C->print_inlining_update(this); 642 return NULL; 643 } 644 645 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 646 LibraryCallKit kit(jvms, this); 647 Compile* C = kit.C; 648 int nodes = C->unique(); 649 _last_predicate = predicate; 650 #ifndef PRODUCT 651 assert(is_predicated() && predicate < predicates_count(), "sanity"); 652 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 653 char buf[1000]; 654 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 655 tty->print_cr("Predicate for intrinsic %s", str); 656 } 657 #endif 658 ciMethod* callee = kit.callee(); 659 const int bci = kit.bci(); 660 661 Node* slow_ctl = kit.try_to_predicate(predicate); 662 if (!kit.failing()) { 663 if (C->print_intrinsics() || C->print_inlining()) { 664 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 665 } 666 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 667 if (C->log()) { 668 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 669 vmIntrinsics::name_at(intrinsic_id()), 670 (is_virtual() ? " virtual='1'" : ""), 671 C->unique() - nodes); 672 } 673 return slow_ctl; // Could be NULL if the check folds. 674 } 675 676 // The intrinsic bailed out 677 if (C->print_intrinsics() || C->print_inlining()) { 678 if (jvms->has_method()) { 679 // Not a root compile. 680 const char* msg = "failed to generate predicate for intrinsic"; 681 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 682 } else { 683 // Root compile 684 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 685 vmIntrinsics::name_at(intrinsic_id()), 686 (is_virtual() ? " (virtual)" : ""), bci); 687 } 688 } 689 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 690 return NULL; 691 } 692 693 bool LibraryCallKit::try_to_inline(int predicate) { 694 // Handle symbolic names for otherwise undistinguished boolean switches: 695 const bool is_store = true; 696 const bool is_native_ptr = true; 697 const bool is_static = true; 698 const bool is_volatile = true; 699 700 if (!jvms()->has_method()) { 701 // Root JVMState has a null method. 702 assert(map()->memory()->Opcode() == Op_Parm, ""); 703 // Insert the memory aliasing node 704 set_all_memory(reset_memory()); 705 } 706 assert(merged_memory(), ""); 707 708 709 switch (intrinsic_id()) { 710 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 711 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 712 case vmIntrinsics::_getClass: return inline_native_getClass(); 713 714 case vmIntrinsics::_dsin: 715 case vmIntrinsics::_dcos: 716 case vmIntrinsics::_dtan: 717 case vmIntrinsics::_dabs: 718 case vmIntrinsics::_datan2: 719 case vmIntrinsics::_dsqrt: 720 case vmIntrinsics::_dexp: 721 case vmIntrinsics::_dlog: 722 case vmIntrinsics::_dlog10: 723 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 724 725 case vmIntrinsics::_min: 726 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 727 728 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 729 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 730 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 731 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 732 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 733 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 734 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 735 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 736 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 737 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 738 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 739 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 740 741 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 742 743 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 744 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 745 case vmIntrinsics::_equals: return inline_string_equals(); 746 747 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 748 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 749 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 750 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 751 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 752 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 753 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 754 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 755 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 756 757 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 758 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 759 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 760 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 761 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 762 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 763 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 764 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 765 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 766 767 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 768 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 769 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 770 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 771 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 772 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 773 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 774 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 775 776 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 777 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 778 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 779 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 780 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 781 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 782 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 783 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 784 785 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 786 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 787 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 788 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 789 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 790 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 791 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 792 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 793 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 794 795 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 796 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 797 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 798 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 799 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 800 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 801 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 802 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 803 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 804 805 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 806 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 807 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 808 809 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 810 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 811 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 812 813 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 814 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 815 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 816 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 817 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 818 819 case vmIntrinsics::_loadFence: 820 case vmIntrinsics::_storeFence: 821 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 822 823 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 824 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 825 826 #ifdef TRACE_HAVE_INTRINSICS 827 case vmIntrinsics::_classID: return inline_native_classID(); 828 case vmIntrinsics::_threadID: return inline_native_threadID(); 829 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 830 #endif 831 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 832 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 833 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 834 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 835 case vmIntrinsics::_newArray: return inline_native_newArray(); 836 case vmIntrinsics::_getLength: return inline_native_getLength(); 837 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 838 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 839 case vmIntrinsics::_equalsC: return inline_array_equals(); 840 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 841 842 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 843 844 case vmIntrinsics::_isInstance: 845 case vmIntrinsics::_getModifiers: 846 case vmIntrinsics::_isInterface: 847 case vmIntrinsics::_isArray: 848 case vmIntrinsics::_isPrimitive: 849 case vmIntrinsics::_getSuperclass: 850 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 851 852 case vmIntrinsics::_floatToRawIntBits: 853 case vmIntrinsics::_floatToIntBits: 854 case vmIntrinsics::_intBitsToFloat: 855 case vmIntrinsics::_doubleToRawLongBits: 856 case vmIntrinsics::_doubleToLongBits: 857 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 858 859 case vmIntrinsics::_numberOfLeadingZeros_i: 860 case vmIntrinsics::_numberOfLeadingZeros_l: 861 case vmIntrinsics::_numberOfTrailingZeros_i: 862 case vmIntrinsics::_numberOfTrailingZeros_l: 863 case vmIntrinsics::_bitCount_i: 864 case vmIntrinsics::_bitCount_l: 865 case vmIntrinsics::_reverseBytes_i: 866 case vmIntrinsics::_reverseBytes_l: 867 case vmIntrinsics::_reverseBytes_s: 868 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 869 870 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 871 872 case vmIntrinsics::_Reference_get: return inline_reference_get(); 873 874 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 875 876 case vmIntrinsics::_aescrypt_encryptBlock: 877 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 878 879 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 880 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 881 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 882 883 case vmIntrinsics::_sha_implCompress: 884 case vmIntrinsics::_sha2_implCompress: 885 case vmIntrinsics::_sha5_implCompress: 886 return inline_sha_implCompress(intrinsic_id()); 887 888 case vmIntrinsics::_digestBase_implCompressMB: 889 return inline_digestBase_implCompressMB(predicate); 890 891 case vmIntrinsics::_multiplyToLen: 892 return inline_multiplyToLen(); 893 894 case vmIntrinsics::_encodeISOArray: 895 return inline_encodeISOArray(); 896 897 case vmIntrinsics::_updateCRC32: 898 return inline_updateCRC32(); 899 case vmIntrinsics::_updateBytesCRC32: 900 return inline_updateBytesCRC32(); 901 case vmIntrinsics::_updateByteBufferCRC32: 902 return inline_updateByteBufferCRC32(); 903 904 case vmIntrinsics::_profileBoolean: 905 return inline_profileBoolean(); 906 907 default: 908 // If you get here, it may be that someone has added a new intrinsic 909 // to the list in vmSymbols.hpp without implementing it here. 910 #ifndef PRODUCT 911 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 912 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 913 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 914 } 915 #endif 916 return false; 917 } 918 } 919 920 Node* LibraryCallKit::try_to_predicate(int predicate) { 921 if (!jvms()->has_method()) { 922 // Root JVMState has a null method. 923 assert(map()->memory()->Opcode() == Op_Parm, ""); 924 // Insert the memory aliasing node 925 set_all_memory(reset_memory()); 926 } 927 assert(merged_memory(), ""); 928 929 switch (intrinsic_id()) { 930 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 931 return inline_cipherBlockChaining_AESCrypt_predicate(false); 932 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 933 return inline_cipherBlockChaining_AESCrypt_predicate(true); 934 case vmIntrinsics::_digestBase_implCompressMB: 935 return inline_digestBase_implCompressMB_predicate(predicate); 936 937 default: 938 // If you get here, it may be that someone has added a new intrinsic 939 // to the list in vmSymbols.hpp without implementing it here. 940 #ifndef PRODUCT 941 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 942 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 943 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 944 } 945 #endif 946 Node* slow_ctl = control(); 947 set_control(top()); // No fast path instrinsic 948 return slow_ctl; 949 } 950 } 951 952 //------------------------------set_result------------------------------- 953 // Helper function for finishing intrinsics. 954 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 955 record_for_igvn(region); 956 set_control(_gvn.transform(region)); 957 set_result( _gvn.transform(value)); 958 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 959 } 960 961 //------------------------------generate_guard--------------------------- 962 // Helper function for generating guarded fast-slow graph structures. 963 // The given 'test', if true, guards a slow path. If the test fails 964 // then a fast path can be taken. (We generally hope it fails.) 965 // In all cases, GraphKit::control() is updated to the fast path. 966 // The returned value represents the control for the slow path. 967 // The return value is never 'top'; it is either a valid control 968 // or NULL if it is obvious that the slow path can never be taken. 969 // Also, if region and the slow control are not NULL, the slow edge 970 // is appended to the region. 971 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 972 if (stopped()) { 973 // Already short circuited. 974 return NULL; 975 } 976 977 // Build an if node and its projections. 978 // If test is true we take the slow path, which we assume is uncommon. 979 if (_gvn.type(test) == TypeInt::ZERO) { 980 // The slow branch is never taken. No need to build this guard. 981 return NULL; 982 } 983 984 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 985 986 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 987 if (if_slow == top()) { 988 // The slow branch is never taken. No need to build this guard. 989 return NULL; 990 } 991 992 if (region != NULL) 993 region->add_req(if_slow); 994 995 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 996 set_control(if_fast); 997 998 return if_slow; 999 } 1000 1001 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1002 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1003 } 1004 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1005 return generate_guard(test, region, PROB_FAIR); 1006 } 1007 1008 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1009 Node* *pos_index) { 1010 if (stopped()) 1011 return NULL; // already stopped 1012 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1013 return NULL; // index is already adequately typed 1014 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1015 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1016 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1017 if (is_neg != NULL && pos_index != NULL) { 1018 // Emulate effect of Parse::adjust_map_after_if. 1019 Node* ccast = new CastIINode(index, TypeInt::POS); 1020 ccast->set_req(0, control()); 1021 (*pos_index) = _gvn.transform(ccast); 1022 } 1023 return is_neg; 1024 } 1025 1026 // Make sure that 'position' is a valid limit index, in [0..length]. 1027 // There are two equivalent plans for checking this: 1028 // A. (offset + copyLength) unsigned<= arrayLength 1029 // B. offset <= (arrayLength - copyLength) 1030 // We require that all of the values above, except for the sum and 1031 // difference, are already known to be non-negative. 1032 // Plan A is robust in the face of overflow, if offset and copyLength 1033 // are both hugely positive. 1034 // 1035 // Plan B is less direct and intuitive, but it does not overflow at 1036 // all, since the difference of two non-negatives is always 1037 // representable. Whenever Java methods must perform the equivalent 1038 // check they generally use Plan B instead of Plan A. 1039 // For the moment we use Plan A. 1040 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1041 Node* subseq_length, 1042 Node* array_length, 1043 RegionNode* region) { 1044 if (stopped()) 1045 return NULL; // already stopped 1046 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1047 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1048 return NULL; // common case of whole-array copy 1049 Node* last = subseq_length; 1050 if (!zero_offset) // last += offset 1051 last = _gvn.transform(new AddINode(last, offset)); 1052 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1053 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1054 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1055 return is_over; 1056 } 1057 1058 1059 //--------------------------generate_current_thread-------------------- 1060 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1061 ciKlass* thread_klass = env()->Thread_klass(); 1062 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1063 Node* thread = _gvn.transform(new ThreadLocalNode()); 1064 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1065 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1066 tls_output = thread; 1067 return threadObj; 1068 } 1069 1070 1071 //------------------------------make_string_method_node------------------------ 1072 // Helper method for String intrinsic functions. This version is called 1073 // with str1 and str2 pointing to String object nodes. 1074 // 1075 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1076 Node* no_ctrl = NULL; 1077 1078 // Get start addr of string 1079 Node* str1_value = load_String_value(no_ctrl, str1); 1080 Node* str1_offset = load_String_offset(no_ctrl, str1); 1081 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1082 1083 // Get length of string 1 1084 Node* str1_len = load_String_length(no_ctrl, str1); 1085 1086 Node* str2_value = load_String_value(no_ctrl, str2); 1087 Node* str2_offset = load_String_offset(no_ctrl, str2); 1088 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1089 1090 Node* str2_len = NULL; 1091 Node* result = NULL; 1092 1093 switch (opcode) { 1094 case Op_StrIndexOf: 1095 // Get length of string 2 1096 str2_len = load_String_length(no_ctrl, str2); 1097 1098 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1099 str1_start, str1_len, str2_start, str2_len); 1100 break; 1101 case Op_StrComp: 1102 // Get length of string 2 1103 str2_len = load_String_length(no_ctrl, str2); 1104 1105 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1106 str1_start, str1_len, str2_start, str2_len); 1107 break; 1108 case Op_StrEquals: 1109 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1110 str1_start, str2_start, str1_len); 1111 break; 1112 default: 1113 ShouldNotReachHere(); 1114 return NULL; 1115 } 1116 1117 // All these intrinsics have checks. 1118 C->set_has_split_ifs(true); // Has chance for split-if optimization 1119 1120 return _gvn.transform(result); 1121 } 1122 1123 // Helper method for String intrinsic functions. This version is called 1124 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1125 // to Int nodes containing the lenghts of str1 and str2. 1126 // 1127 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1128 Node* result = NULL; 1129 switch (opcode) { 1130 case Op_StrIndexOf: 1131 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1132 str1_start, cnt1, str2_start, cnt2); 1133 break; 1134 case Op_StrComp: 1135 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1136 str1_start, cnt1, str2_start, cnt2); 1137 break; 1138 case Op_StrEquals: 1139 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1140 str1_start, str2_start, cnt1); 1141 break; 1142 default: 1143 ShouldNotReachHere(); 1144 return NULL; 1145 } 1146 1147 // All these intrinsics have checks. 1148 C->set_has_split_ifs(true); // Has chance for split-if optimization 1149 1150 return _gvn.transform(result); 1151 } 1152 1153 //------------------------------inline_string_compareTo------------------------ 1154 // public int java.lang.String.compareTo(String anotherString); 1155 bool LibraryCallKit::inline_string_compareTo() { 1156 Node* receiver = null_check(argument(0)); 1157 Node* arg = null_check(argument(1)); 1158 if (stopped()) { 1159 return true; 1160 } 1161 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1162 return true; 1163 } 1164 1165 //------------------------------inline_string_equals------------------------ 1166 bool LibraryCallKit::inline_string_equals() { 1167 Node* receiver = null_check_receiver(); 1168 // NOTE: Do not null check argument for String.equals() because spec 1169 // allows to specify NULL as argument. 1170 Node* argument = this->argument(1); 1171 if (stopped()) { 1172 return true; 1173 } 1174 1175 // paths (plus control) merge 1176 RegionNode* region = new RegionNode(5); 1177 Node* phi = new PhiNode(region, TypeInt::BOOL); 1178 1179 // does source == target string? 1180 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1181 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1182 1183 Node* if_eq = generate_slow_guard(bol, NULL); 1184 if (if_eq != NULL) { 1185 // receiver == argument 1186 phi->init_req(2, intcon(1)); 1187 region->init_req(2, if_eq); 1188 } 1189 1190 // get String klass for instanceOf 1191 ciInstanceKlass* klass = env()->String_klass(); 1192 1193 if (!stopped()) { 1194 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1195 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1196 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1197 1198 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1199 //instanceOf == true, fallthrough 1200 1201 if (inst_false != NULL) { 1202 phi->init_req(3, intcon(0)); 1203 region->init_req(3, inst_false); 1204 } 1205 } 1206 1207 if (!stopped()) { 1208 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1209 1210 // Properly cast the argument to String 1211 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1212 // This path is taken only when argument's type is String:NotNull. 1213 argument = cast_not_null(argument, false); 1214 1215 Node* no_ctrl = NULL; 1216 1217 // Get start addr of receiver 1218 Node* receiver_val = load_String_value(no_ctrl, receiver); 1219 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1220 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1221 1222 // Get length of receiver 1223 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1224 1225 // Get start addr of argument 1226 Node* argument_val = load_String_value(no_ctrl, argument); 1227 Node* argument_offset = load_String_offset(no_ctrl, argument); 1228 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1229 1230 // Get length of argument 1231 Node* argument_cnt = load_String_length(no_ctrl, argument); 1232 1233 // Check for receiver count != argument count 1234 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1235 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1236 Node* if_ne = generate_slow_guard(bol, NULL); 1237 if (if_ne != NULL) { 1238 phi->init_req(4, intcon(0)); 1239 region->init_req(4, if_ne); 1240 } 1241 1242 // Check for count == 0 is done by assembler code for StrEquals. 1243 1244 if (!stopped()) { 1245 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1246 phi->init_req(1, equals); 1247 region->init_req(1, control()); 1248 } 1249 } 1250 1251 // post merge 1252 set_control(_gvn.transform(region)); 1253 record_for_igvn(region); 1254 1255 set_result(_gvn.transform(phi)); 1256 return true; 1257 } 1258 1259 //------------------------------inline_array_equals---------------------------- 1260 bool LibraryCallKit::inline_array_equals() { 1261 Node* arg1 = argument(0); 1262 Node* arg2 = argument(1); 1263 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1264 return true; 1265 } 1266 1267 // Java version of String.indexOf(constant string) 1268 // class StringDecl { 1269 // StringDecl(char[] ca) { 1270 // offset = 0; 1271 // count = ca.length; 1272 // value = ca; 1273 // } 1274 // int offset; 1275 // int count; 1276 // char[] value; 1277 // } 1278 // 1279 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1280 // int targetOffset, int cache_i, int md2) { 1281 // int cache = cache_i; 1282 // int sourceOffset = string_object.offset; 1283 // int sourceCount = string_object.count; 1284 // int targetCount = target_object.length; 1285 // 1286 // int targetCountLess1 = targetCount - 1; 1287 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1288 // 1289 // char[] source = string_object.value; 1290 // char[] target = target_object; 1291 // int lastChar = target[targetCountLess1]; 1292 // 1293 // outer_loop: 1294 // for (int i = sourceOffset; i < sourceEnd; ) { 1295 // int src = source[i + targetCountLess1]; 1296 // if (src == lastChar) { 1297 // // With random strings and a 4-character alphabet, 1298 // // reverse matching at this point sets up 0.8% fewer 1299 // // frames, but (paradoxically) makes 0.3% more probes. 1300 // // Since those probes are nearer the lastChar probe, 1301 // // there is may be a net D$ win with reverse matching. 1302 // // But, reversing loop inhibits unroll of inner loop 1303 // // for unknown reason. So, does running outer loop from 1304 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1305 // for (int j = 0; j < targetCountLess1; j++) { 1306 // if (target[targetOffset + j] != source[i+j]) { 1307 // if ((cache & (1 << source[i+j])) == 0) { 1308 // if (md2 < j+1) { 1309 // i += j+1; 1310 // continue outer_loop; 1311 // } 1312 // } 1313 // i += md2; 1314 // continue outer_loop; 1315 // } 1316 // } 1317 // return i - sourceOffset; 1318 // } 1319 // if ((cache & (1 << src)) == 0) { 1320 // i += targetCountLess1; 1321 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1322 // i++; 1323 // } 1324 // return -1; 1325 // } 1326 1327 //------------------------------string_indexOf------------------------ 1328 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1329 jint cache_i, jint md2_i) { 1330 1331 Node* no_ctrl = NULL; 1332 float likely = PROB_LIKELY(0.9); 1333 float unlikely = PROB_UNLIKELY(0.9); 1334 1335 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1336 1337 Node* source = load_String_value(no_ctrl, string_object); 1338 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1339 Node* sourceCount = load_String_length(no_ctrl, string_object); 1340 1341 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1342 jint target_length = target_array->length(); 1343 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1344 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1345 1346 // String.value field is known to be @Stable. 1347 if (UseImplicitStableValues) { 1348 target = cast_array_to_stable(target, target_type); 1349 } 1350 1351 IdealKit kit(this, false, true); 1352 #define __ kit. 1353 Node* zero = __ ConI(0); 1354 Node* one = __ ConI(1); 1355 Node* cache = __ ConI(cache_i); 1356 Node* md2 = __ ConI(md2_i); 1357 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1358 Node* targetCountLess1 = __ ConI(target_length - 1); 1359 Node* targetOffset = __ ConI(targetOffset_i); 1360 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1361 1362 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1363 Node* outer_loop = __ make_label(2 /* goto */); 1364 Node* return_ = __ make_label(1); 1365 1366 __ set(rtn,__ ConI(-1)); 1367 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1368 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1369 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1370 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1371 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1372 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1373 Node* tpj = __ AddI(targetOffset, __ value(j)); 1374 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1375 Node* ipj = __ AddI(__ value(i), __ value(j)); 1376 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1377 __ if_then(targ, BoolTest::ne, src2); { 1378 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1379 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1380 __ increment(i, __ AddI(__ value(j), one)); 1381 __ goto_(outer_loop); 1382 } __ end_if(); __ dead(j); 1383 }__ end_if(); __ dead(j); 1384 __ increment(i, md2); 1385 __ goto_(outer_loop); 1386 }__ end_if(); 1387 __ increment(j, one); 1388 }__ end_loop(); __ dead(j); 1389 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1390 __ goto_(return_); 1391 }__ end_if(); 1392 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1393 __ increment(i, targetCountLess1); 1394 }__ end_if(); 1395 __ increment(i, one); 1396 __ bind(outer_loop); 1397 }__ end_loop(); __ dead(i); 1398 __ bind(return_); 1399 1400 // Final sync IdealKit and GraphKit. 1401 final_sync(kit); 1402 Node* result = __ value(rtn); 1403 #undef __ 1404 C->set_has_loops(true); 1405 return result; 1406 } 1407 1408 //------------------------------inline_string_indexOf------------------------ 1409 bool LibraryCallKit::inline_string_indexOf() { 1410 Node* receiver = argument(0); 1411 Node* arg = argument(1); 1412 1413 Node* result; 1414 if (Matcher::has_match_rule(Op_StrIndexOf) && 1415 UseSSE42Intrinsics) { 1416 // Generate SSE4.2 version of indexOf 1417 // We currently only have match rules that use SSE4.2 1418 1419 receiver = null_check(receiver); 1420 arg = null_check(arg); 1421 if (stopped()) { 1422 return true; 1423 } 1424 1425 // Make the merge point 1426 RegionNode* result_rgn = new RegionNode(4); 1427 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1428 Node* no_ctrl = NULL; 1429 1430 // Get start addr of source string 1431 Node* source = load_String_value(no_ctrl, receiver); 1432 Node* source_offset = load_String_offset(no_ctrl, receiver); 1433 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1434 1435 // Get length of source string 1436 Node* source_cnt = load_String_length(no_ctrl, receiver); 1437 1438 // Get start addr of substring 1439 Node* substr = load_String_value(no_ctrl, arg); 1440 Node* substr_offset = load_String_offset(no_ctrl, arg); 1441 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1442 1443 // Get length of source string 1444 Node* substr_cnt = load_String_length(no_ctrl, arg); 1445 1446 // Check for substr count > string count 1447 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1448 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1449 Node* if_gt = generate_slow_guard(bol, NULL); 1450 if (if_gt != NULL) { 1451 result_phi->init_req(2, intcon(-1)); 1452 result_rgn->init_req(2, if_gt); 1453 } 1454 1455 if (!stopped()) { 1456 // Check for substr count == 0 1457 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1458 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1459 Node* if_zero = generate_slow_guard(bol, NULL); 1460 if (if_zero != NULL) { 1461 result_phi->init_req(3, intcon(0)); 1462 result_rgn->init_req(3, if_zero); 1463 } 1464 } 1465 1466 if (!stopped()) { 1467 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1468 result_phi->init_req(1, result); 1469 result_rgn->init_req(1, control()); 1470 } 1471 set_control(_gvn.transform(result_rgn)); 1472 record_for_igvn(result_rgn); 1473 result = _gvn.transform(result_phi); 1474 1475 } else { // Use LibraryCallKit::string_indexOf 1476 // don't intrinsify if argument isn't a constant string. 1477 if (!arg->is_Con()) { 1478 return false; 1479 } 1480 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1481 if (str_type == NULL) { 1482 return false; 1483 } 1484 ciInstanceKlass* klass = env()->String_klass(); 1485 ciObject* str_const = str_type->const_oop(); 1486 if (str_const == NULL || str_const->klass() != klass) { 1487 return false; 1488 } 1489 ciInstance* str = str_const->as_instance(); 1490 assert(str != NULL, "must be instance"); 1491 1492 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1493 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1494 1495 int o; 1496 int c; 1497 if (java_lang_String::has_offset_field()) { 1498 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1499 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1500 } else { 1501 o = 0; 1502 c = pat->length(); 1503 } 1504 1505 // constant strings have no offset and count == length which 1506 // simplifies the resulting code somewhat so lets optimize for that. 1507 if (o != 0 || c != pat->length()) { 1508 return false; 1509 } 1510 1511 receiver = null_check(receiver, T_OBJECT); 1512 // NOTE: No null check on the argument is needed since it's a constant String oop. 1513 if (stopped()) { 1514 return true; 1515 } 1516 1517 // The null string as a pattern always returns 0 (match at beginning of string) 1518 if (c == 0) { 1519 set_result(intcon(0)); 1520 return true; 1521 } 1522 1523 // Generate default indexOf 1524 jchar lastChar = pat->char_at(o + (c - 1)); 1525 int cache = 0; 1526 int i; 1527 for (i = 0; i < c - 1; i++) { 1528 assert(i < pat->length(), "out of range"); 1529 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1530 } 1531 1532 int md2 = c; 1533 for (i = 0; i < c - 1; i++) { 1534 assert(i < pat->length(), "out of range"); 1535 if (pat->char_at(o + i) == lastChar) { 1536 md2 = (c - 1) - i; 1537 } 1538 } 1539 1540 result = string_indexOf(receiver, pat, o, cache, md2); 1541 } 1542 set_result(result); 1543 return true; 1544 } 1545 1546 //--------------------------round_double_node-------------------------------- 1547 // Round a double node if necessary. 1548 Node* LibraryCallKit::round_double_node(Node* n) { 1549 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1550 n = _gvn.transform(new RoundDoubleNode(0, n)); 1551 return n; 1552 } 1553 1554 //------------------------------inline_math----------------------------------- 1555 // public static double Math.abs(double) 1556 // public static double Math.sqrt(double) 1557 // public static double Math.log(double) 1558 // public static double Math.log10(double) 1559 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1560 Node* arg = round_double_node(argument(0)); 1561 Node* n; 1562 switch (id) { 1563 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1564 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1565 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1566 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1567 default: fatal_unexpected_iid(id); break; 1568 } 1569 set_result(_gvn.transform(n)); 1570 return true; 1571 } 1572 1573 //------------------------------inline_trig---------------------------------- 1574 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1575 // argument reduction which will turn into a fast/slow diamond. 1576 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1577 Node* arg = round_double_node(argument(0)); 1578 Node* n = NULL; 1579 1580 switch (id) { 1581 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1582 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1583 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1584 default: fatal_unexpected_iid(id); break; 1585 } 1586 n = _gvn.transform(n); 1587 1588 // Rounding required? Check for argument reduction! 1589 if (Matcher::strict_fp_requires_explicit_rounding) { 1590 static const double pi_4 = 0.7853981633974483; 1591 static const double neg_pi_4 = -0.7853981633974483; 1592 // pi/2 in 80-bit extended precision 1593 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1594 // -pi/2 in 80-bit extended precision 1595 // 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}; 1596 // Cutoff value for using this argument reduction technique 1597 //static const double pi_2_minus_epsilon = 1.564660403643354; 1598 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1599 1600 // Pseudocode for sin: 1601 // if (x <= Math.PI / 4.0) { 1602 // if (x >= -Math.PI / 4.0) return fsin(x); 1603 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1604 // } else { 1605 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1606 // } 1607 // return StrictMath.sin(x); 1608 1609 // Pseudocode for cos: 1610 // if (x <= Math.PI / 4.0) { 1611 // if (x >= -Math.PI / 4.0) return fcos(x); 1612 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1613 // } else { 1614 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1615 // } 1616 // return StrictMath.cos(x); 1617 1618 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1619 // requires a special machine instruction to load it. Instead we'll try 1620 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1621 // probably do the math inside the SIN encoding. 1622 1623 // Make the merge point 1624 RegionNode* r = new RegionNode(3); 1625 Node* phi = new PhiNode(r, Type::DOUBLE); 1626 1627 // Flatten arg so we need only 1 test 1628 Node *abs = _gvn.transform(new AbsDNode(arg)); 1629 // Node for PI/4 constant 1630 Node *pi4 = makecon(TypeD::make(pi_4)); 1631 // Check PI/4 : abs(arg) 1632 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1633 // Check: If PI/4 < abs(arg) then go slow 1634 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1635 // Branch either way 1636 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1637 set_control(opt_iff(r,iff)); 1638 1639 // Set fast path result 1640 phi->init_req(2, n); 1641 1642 // Slow path - non-blocking leaf call 1643 Node* call = NULL; 1644 switch (id) { 1645 case vmIntrinsics::_dsin: 1646 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1647 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1648 "Sin", NULL, arg, top()); 1649 break; 1650 case vmIntrinsics::_dcos: 1651 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1652 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1653 "Cos", NULL, arg, top()); 1654 break; 1655 case vmIntrinsics::_dtan: 1656 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1657 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1658 "Tan", NULL, arg, top()); 1659 break; 1660 } 1661 assert(control()->in(0) == call, ""); 1662 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1663 r->init_req(1, control()); 1664 phi->init_req(1, slow_result); 1665 1666 // Post-merge 1667 set_control(_gvn.transform(r)); 1668 record_for_igvn(r); 1669 n = _gvn.transform(phi); 1670 1671 C->set_has_split_ifs(true); // Has chance for split-if optimization 1672 } 1673 set_result(n); 1674 return true; 1675 } 1676 1677 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1678 //------------------- 1679 //result=(result.isNaN())? funcAddr():result; 1680 // Check: If isNaN() by checking result!=result? then either trap 1681 // or go to runtime 1682 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1683 // Build the boolean node 1684 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1685 1686 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1687 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1688 // The pow or exp intrinsic returned a NaN, which requires a call 1689 // to the runtime. Recompile with the runtime call. 1690 uncommon_trap(Deoptimization::Reason_intrinsic, 1691 Deoptimization::Action_make_not_entrant); 1692 } 1693 return result; 1694 } else { 1695 // If this inlining ever returned NaN in the past, we compile a call 1696 // to the runtime to properly handle corner cases 1697 1698 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1699 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1700 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1701 1702 if (!if_slow->is_top()) { 1703 RegionNode* result_region = new RegionNode(3); 1704 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1705 1706 result_region->init_req(1, if_fast); 1707 result_val->init_req(1, result); 1708 1709 set_control(if_slow); 1710 1711 const TypePtr* no_memory_effects = NULL; 1712 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1713 no_memory_effects, 1714 x, top(), y, y ? top() : NULL); 1715 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1716 #ifdef ASSERT 1717 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1718 assert(value_top == top(), "second value must be top"); 1719 #endif 1720 1721 result_region->init_req(2, control()); 1722 result_val->init_req(2, value); 1723 set_control(_gvn.transform(result_region)); 1724 return _gvn.transform(result_val); 1725 } else { 1726 return result; 1727 } 1728 } 1729 } 1730 1731 //------------------------------inline_exp------------------------------------- 1732 // Inline exp instructions, if possible. The Intel hardware only misses 1733 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1734 bool LibraryCallKit::inline_exp() { 1735 Node* arg = round_double_node(argument(0)); 1736 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1737 1738 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1739 set_result(n); 1740 1741 C->set_has_split_ifs(true); // Has chance for split-if optimization 1742 return true; 1743 } 1744 1745 //------------------------------inline_pow------------------------------------- 1746 // Inline power instructions, if possible. 1747 bool LibraryCallKit::inline_pow() { 1748 // Pseudocode for pow 1749 // if (y == 2) { 1750 // return x * x; 1751 // } else { 1752 // if (x <= 0.0) { 1753 // long longy = (long)y; 1754 // if ((double)longy == y) { // if y is long 1755 // if (y + 1 == y) longy = 0; // huge number: even 1756 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1757 // } else { 1758 // result = NaN; 1759 // } 1760 // } else { 1761 // result = DPow(x,y); 1762 // } 1763 // if (result != result)? { 1764 // result = uncommon_trap() or runtime_call(); 1765 // } 1766 // return result; 1767 // } 1768 1769 Node* x = round_double_node(argument(0)); 1770 Node* y = round_double_node(argument(2)); 1771 1772 Node* result = NULL; 1773 1774 Node* const_two_node = makecon(TypeD::make(2.0)); 1775 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1776 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1777 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1778 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1779 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1780 1781 RegionNode* region_node = new RegionNode(3); 1782 region_node->init_req(1, if_true); 1783 1784 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1785 // special case for x^y where y == 2, we can convert it to x * x 1786 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1787 1788 // set control to if_false since we will now process the false branch 1789 set_control(if_false); 1790 1791 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1792 // Short form: skip the fancy tests and just check for NaN result. 1793 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1794 } else { 1795 // If this inlining ever returned NaN in the past, include all 1796 // checks + call to the runtime. 1797 1798 // Set the merge point for If node with condition of (x <= 0.0) 1799 // There are four possible paths to region node and phi node 1800 RegionNode *r = new RegionNode(4); 1801 Node *phi = new PhiNode(r, Type::DOUBLE); 1802 1803 // Build the first if node: if (x <= 0.0) 1804 // Node for 0 constant 1805 Node *zeronode = makecon(TypeD::ZERO); 1806 // Check x:0 1807 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1808 // Check: If (x<=0) then go complex path 1809 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1810 // Branch either way 1811 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1812 // Fast path taken; set region slot 3 1813 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1814 r->init_req(3,fast_taken); // Capture fast-control 1815 1816 // Fast path not-taken, i.e. slow path 1817 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1818 1819 // Set fast path result 1820 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1821 phi->init_req(3, fast_result); 1822 1823 // Complex path 1824 // Build the second if node (if y is long) 1825 // Node for (long)y 1826 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1827 // Node for (double)((long) y) 1828 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1829 // Check (double)((long) y) : y 1830 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1831 // Check if (y isn't long) then go to slow path 1832 1833 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1834 // Branch either way 1835 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1836 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1837 1838 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1839 1840 // Calculate DPow(abs(x), y)*(1 & (long)y) 1841 // Node for constant 1 1842 Node *conone = longcon(1); 1843 // 1& (long)y 1844 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1845 1846 // A huge number is always even. Detect a huge number by checking 1847 // if y + 1 == y and set integer to be tested for parity to 0. 1848 // Required for corner case: 1849 // (long)9.223372036854776E18 = max_jlong 1850 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1851 // max_jlong is odd but 9.223372036854776E18 is even 1852 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1853 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1854 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1855 Node* correctedsign = NULL; 1856 if (ConditionalMoveLimit != 0) { 1857 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1858 } else { 1859 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1860 RegionNode *r = new RegionNode(3); 1861 Node *phi = new PhiNode(r, TypeLong::LONG); 1862 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1863 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1864 phi->init_req(1, signnode); 1865 phi->init_req(2, longcon(0)); 1866 correctedsign = _gvn.transform(phi); 1867 ylong_path = _gvn.transform(r); 1868 record_for_igvn(r); 1869 } 1870 1871 // zero node 1872 Node *conzero = longcon(0); 1873 // Check (1&(long)y)==0? 1874 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1875 // Check if (1&(long)y)!=0?, if so the result is negative 1876 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1877 // abs(x) 1878 Node *absx=_gvn.transform(new AbsDNode(x)); 1879 // abs(x)^y 1880 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1881 // -abs(x)^y 1882 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1883 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1884 Node *signresult = NULL; 1885 if (ConditionalMoveLimit != 0) { 1886 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1887 } else { 1888 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1889 RegionNode *r = new RegionNode(3); 1890 Node *phi = new PhiNode(r, Type::DOUBLE); 1891 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1892 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1893 phi->init_req(1, absxpowy); 1894 phi->init_req(2, negabsxpowy); 1895 signresult = _gvn.transform(phi); 1896 ylong_path = _gvn.transform(r); 1897 record_for_igvn(r); 1898 } 1899 // Set complex path fast result 1900 r->init_req(2, ylong_path); 1901 phi->init_req(2, signresult); 1902 1903 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1904 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1905 r->init_req(1,slow_path); 1906 phi->init_req(1,slow_result); 1907 1908 // Post merge 1909 set_control(_gvn.transform(r)); 1910 record_for_igvn(r); 1911 result = _gvn.transform(phi); 1912 } 1913 1914 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1915 1916 // control from finish_pow_exp is now input to the region node 1917 region_node->set_req(2, control()); 1918 // the result from finish_pow_exp is now input to the phi node 1919 phi_node->init_req(2, result); 1920 set_control(_gvn.transform(region_node)); 1921 record_for_igvn(region_node); 1922 set_result(_gvn.transform(phi_node)); 1923 1924 C->set_has_split_ifs(true); // Has chance for split-if optimization 1925 return true; 1926 } 1927 1928 //------------------------------runtime_math----------------------------- 1929 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1930 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1931 "must be (DD)D or (D)D type"); 1932 1933 // Inputs 1934 Node* a = round_double_node(argument(0)); 1935 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1936 1937 const TypePtr* no_memory_effects = NULL; 1938 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1939 no_memory_effects, 1940 a, top(), b, b ? top() : NULL); 1941 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1942 #ifdef ASSERT 1943 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1944 assert(value_top == top(), "second value must be top"); 1945 #endif 1946 1947 set_result(value); 1948 return true; 1949 } 1950 1951 //------------------------------inline_math_native----------------------------- 1952 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1953 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1954 switch (id) { 1955 // These intrinsics are not properly supported on all hardware 1956 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1957 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1958 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1959 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1960 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1961 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1962 1963 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1964 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1965 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1966 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1967 1968 // These intrinsics are supported on all hardware 1969 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1970 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1971 1972 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1973 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1974 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1975 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1976 #undef FN_PTR 1977 1978 // These intrinsics are not yet correctly implemented 1979 case vmIntrinsics::_datan2: 1980 return false; 1981 1982 default: 1983 fatal_unexpected_iid(id); 1984 return false; 1985 } 1986 } 1987 1988 static bool is_simple_name(Node* n) { 1989 return (n->req() == 1 // constant 1990 || (n->is_Type() && n->as_Type()->type()->singleton()) 1991 || n->is_Proj() // parameter or return value 1992 || n->is_Phi() // local of some sort 1993 ); 1994 } 1995 1996 //----------------------------inline_min_max----------------------------------- 1997 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1998 set_result(generate_min_max(id, argument(0), argument(1))); 1999 return true; 2000 } 2001 2002 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2003 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2004 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2005 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2006 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2007 2008 { 2009 PreserveJVMState pjvms(this); 2010 PreserveReexecuteState preexecs(this); 2011 jvms()->set_should_reexecute(true); 2012 2013 set_control(slow_path); 2014 set_i_o(i_o()); 2015 2016 uncommon_trap(Deoptimization::Reason_intrinsic, 2017 Deoptimization::Action_none); 2018 } 2019 2020 set_control(fast_path); 2021 set_result(math); 2022 } 2023 2024 template <typename OverflowOp> 2025 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2026 typedef typename OverflowOp::MathOp MathOp; 2027 2028 MathOp* mathOp = new MathOp(arg1, arg2); 2029 Node* operation = _gvn.transform( mathOp ); 2030 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2031 inline_math_mathExact(operation, ofcheck); 2032 return true; 2033 } 2034 2035 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2036 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2037 } 2038 2039 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2040 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2041 } 2042 2043 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2044 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2045 } 2046 2047 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2048 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2049 } 2050 2051 bool LibraryCallKit::inline_math_negateExactI() { 2052 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2053 } 2054 2055 bool LibraryCallKit::inline_math_negateExactL() { 2056 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2057 } 2058 2059 bool LibraryCallKit::inline_math_multiplyExactI() { 2060 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2061 } 2062 2063 bool LibraryCallKit::inline_math_multiplyExactL() { 2064 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2065 } 2066 2067 Node* 2068 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2069 // These are the candidate return value: 2070 Node* xvalue = x0; 2071 Node* yvalue = y0; 2072 2073 if (xvalue == yvalue) { 2074 return xvalue; 2075 } 2076 2077 bool want_max = (id == vmIntrinsics::_max); 2078 2079 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2080 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2081 if (txvalue == NULL || tyvalue == NULL) return top(); 2082 // This is not really necessary, but it is consistent with a 2083 // hypothetical MaxINode::Value method: 2084 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2085 2086 // %%% This folding logic should (ideally) be in a different place. 2087 // Some should be inside IfNode, and there to be a more reliable 2088 // transformation of ?: style patterns into cmoves. We also want 2089 // more powerful optimizations around cmove and min/max. 2090 2091 // Try to find a dominating comparison of these guys. 2092 // It can simplify the index computation for Arrays.copyOf 2093 // and similar uses of System.arraycopy. 2094 // First, compute the normalized version of CmpI(x, y). 2095 int cmp_op = Op_CmpI; 2096 Node* xkey = xvalue; 2097 Node* ykey = yvalue; 2098 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2099 if (ideal_cmpxy->is_Cmp()) { 2100 // E.g., if we have CmpI(length - offset, count), 2101 // it might idealize to CmpI(length, count + offset) 2102 cmp_op = ideal_cmpxy->Opcode(); 2103 xkey = ideal_cmpxy->in(1); 2104 ykey = ideal_cmpxy->in(2); 2105 } 2106 2107 // Start by locating any relevant comparisons. 2108 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2109 Node* cmpxy = NULL; 2110 Node* cmpyx = NULL; 2111 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2112 Node* cmp = start_from->fast_out(k); 2113 if (cmp->outcnt() > 0 && // must have prior uses 2114 cmp->in(0) == NULL && // must be context-independent 2115 cmp->Opcode() == cmp_op) { // right kind of compare 2116 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2117 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2118 } 2119 } 2120 2121 const int NCMPS = 2; 2122 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2123 int cmpn; 2124 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2125 if (cmps[cmpn] != NULL) break; // find a result 2126 } 2127 if (cmpn < NCMPS) { 2128 // Look for a dominating test that tells us the min and max. 2129 int depth = 0; // Limit search depth for speed 2130 Node* dom = control(); 2131 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2132 if (++depth >= 100) break; 2133 Node* ifproj = dom; 2134 if (!ifproj->is_Proj()) continue; 2135 Node* iff = ifproj->in(0); 2136 if (!iff->is_If()) continue; 2137 Node* bol = iff->in(1); 2138 if (!bol->is_Bool()) continue; 2139 Node* cmp = bol->in(1); 2140 if (cmp == NULL) continue; 2141 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2142 if (cmps[cmpn] == cmp) break; 2143 if (cmpn == NCMPS) continue; 2144 BoolTest::mask btest = bol->as_Bool()->_test._test; 2145 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2146 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2147 // At this point, we know that 'x btest y' is true. 2148 switch (btest) { 2149 case BoolTest::eq: 2150 // They are proven equal, so we can collapse the min/max. 2151 // Either value is the answer. Choose the simpler. 2152 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2153 return yvalue; 2154 return xvalue; 2155 case BoolTest::lt: // x < y 2156 case BoolTest::le: // x <= y 2157 return (want_max ? yvalue : xvalue); 2158 case BoolTest::gt: // x > y 2159 case BoolTest::ge: // x >= y 2160 return (want_max ? xvalue : yvalue); 2161 } 2162 } 2163 } 2164 2165 // We failed to find a dominating test. 2166 // Let's pick a test that might GVN with prior tests. 2167 Node* best_bol = NULL; 2168 BoolTest::mask best_btest = BoolTest::illegal; 2169 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2170 Node* cmp = cmps[cmpn]; 2171 if (cmp == NULL) continue; 2172 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2173 Node* bol = cmp->fast_out(j); 2174 if (!bol->is_Bool()) continue; 2175 BoolTest::mask btest = bol->as_Bool()->_test._test; 2176 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2177 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2178 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2179 best_bol = bol->as_Bool(); 2180 best_btest = btest; 2181 } 2182 } 2183 } 2184 2185 Node* answer_if_true = NULL; 2186 Node* answer_if_false = NULL; 2187 switch (best_btest) { 2188 default: 2189 if (cmpxy == NULL) 2190 cmpxy = ideal_cmpxy; 2191 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2192 // and fall through: 2193 case BoolTest::lt: // x < y 2194 case BoolTest::le: // x <= y 2195 answer_if_true = (want_max ? yvalue : xvalue); 2196 answer_if_false = (want_max ? xvalue : yvalue); 2197 break; 2198 case BoolTest::gt: // x > y 2199 case BoolTest::ge: // x >= y 2200 answer_if_true = (want_max ? xvalue : yvalue); 2201 answer_if_false = (want_max ? yvalue : xvalue); 2202 break; 2203 } 2204 2205 jint hi, lo; 2206 if (want_max) { 2207 // We can sharpen the minimum. 2208 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2209 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2210 } else { 2211 // We can sharpen the maximum. 2212 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2213 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2214 } 2215 2216 // Use a flow-free graph structure, to avoid creating excess control edges 2217 // which could hinder other optimizations. 2218 // Since Math.min/max is often used with arraycopy, we want 2219 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2220 Node* cmov = CMoveNode::make(NULL, best_bol, 2221 answer_if_false, answer_if_true, 2222 TypeInt::make(lo, hi, widen)); 2223 2224 return _gvn.transform(cmov); 2225 2226 /* 2227 // This is not as desirable as it may seem, since Min and Max 2228 // nodes do not have a full set of optimizations. 2229 // And they would interfere, anyway, with 'if' optimizations 2230 // and with CMoveI canonical forms. 2231 switch (id) { 2232 case vmIntrinsics::_min: 2233 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2234 case vmIntrinsics::_max: 2235 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2236 default: 2237 ShouldNotReachHere(); 2238 } 2239 */ 2240 } 2241 2242 inline int 2243 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2244 const TypePtr* base_type = TypePtr::NULL_PTR; 2245 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2246 if (base_type == NULL) { 2247 // Unknown type. 2248 return Type::AnyPtr; 2249 } else if (base_type == TypePtr::NULL_PTR) { 2250 // Since this is a NULL+long form, we have to switch to a rawptr. 2251 base = _gvn.transform(new CastX2PNode(offset)); 2252 offset = MakeConX(0); 2253 return Type::RawPtr; 2254 } else if (base_type->base() == Type::RawPtr) { 2255 return Type::RawPtr; 2256 } else if (base_type->isa_oopptr()) { 2257 // Base is never null => always a heap address. 2258 if (base_type->ptr() == TypePtr::NotNull) { 2259 return Type::OopPtr; 2260 } 2261 // Offset is small => always a heap address. 2262 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2263 if (offset_type != NULL && 2264 base_type->offset() == 0 && // (should always be?) 2265 offset_type->_lo >= 0 && 2266 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2267 return Type::OopPtr; 2268 } 2269 // Otherwise, it might either be oop+off or NULL+addr. 2270 return Type::AnyPtr; 2271 } else { 2272 // No information: 2273 return Type::AnyPtr; 2274 } 2275 } 2276 2277 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2278 int kind = classify_unsafe_addr(base, offset); 2279 if (kind == Type::RawPtr) { 2280 return basic_plus_adr(top(), base, offset); 2281 } else { 2282 return basic_plus_adr(base, offset); 2283 } 2284 } 2285 2286 //--------------------------inline_number_methods----------------------------- 2287 // inline int Integer.numberOfLeadingZeros(int) 2288 // inline int Long.numberOfLeadingZeros(long) 2289 // 2290 // inline int Integer.numberOfTrailingZeros(int) 2291 // inline int Long.numberOfTrailingZeros(long) 2292 // 2293 // inline int Integer.bitCount(int) 2294 // inline int Long.bitCount(long) 2295 // 2296 // inline char Character.reverseBytes(char) 2297 // inline short Short.reverseBytes(short) 2298 // inline int Integer.reverseBytes(int) 2299 // inline long Long.reverseBytes(long) 2300 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2301 Node* arg = argument(0); 2302 Node* n; 2303 switch (id) { 2304 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2305 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2306 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2307 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2308 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2309 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2310 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2311 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2312 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2313 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2314 default: fatal_unexpected_iid(id); break; 2315 } 2316 set_result(_gvn.transform(n)); 2317 return true; 2318 } 2319 2320 //----------------------------inline_unsafe_access---------------------------- 2321 2322 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2323 2324 // Helper that guards and inserts a pre-barrier. 2325 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2326 Node* pre_val, bool need_mem_bar) { 2327 // We could be accessing the referent field of a reference object. If so, when G1 2328 // is enabled, we need to log the value in the referent field in an SATB buffer. 2329 // This routine performs some compile time filters and generates suitable 2330 // runtime filters that guard the pre-barrier code. 2331 // Also add memory barrier for non volatile load from the referent field 2332 // to prevent commoning of loads across safepoint. 2333 if (!UseG1GC && !need_mem_bar) 2334 return; 2335 2336 // Some compile time checks. 2337 2338 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2339 const TypeX* otype = offset->find_intptr_t_type(); 2340 if (otype != NULL && otype->is_con() && 2341 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2342 // Constant offset but not the reference_offset so just return 2343 return; 2344 } 2345 2346 // We only need to generate the runtime guards for instances. 2347 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2348 if (btype != NULL) { 2349 if (btype->isa_aryptr()) { 2350 // Array type so nothing to do 2351 return; 2352 } 2353 2354 const TypeInstPtr* itype = btype->isa_instptr(); 2355 if (itype != NULL) { 2356 // Can the klass of base_oop be statically determined to be 2357 // _not_ a sub-class of Reference and _not_ Object? 2358 ciKlass* klass = itype->klass(); 2359 if ( klass->is_loaded() && 2360 !klass->is_subtype_of(env()->Reference_klass()) && 2361 !env()->Object_klass()->is_subtype_of(klass)) { 2362 return; 2363 } 2364 } 2365 } 2366 2367 // The compile time filters did not reject base_oop/offset so 2368 // we need to generate the following runtime filters 2369 // 2370 // if (offset == java_lang_ref_Reference::_reference_offset) { 2371 // if (instance_of(base, java.lang.ref.Reference)) { 2372 // pre_barrier(_, pre_val, ...); 2373 // } 2374 // } 2375 2376 float likely = PROB_LIKELY( 0.999); 2377 float unlikely = PROB_UNLIKELY(0.999); 2378 2379 IdealKit ideal(this); 2380 #define __ ideal. 2381 2382 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2383 2384 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2385 // Update graphKit memory and control from IdealKit. 2386 sync_kit(ideal); 2387 2388 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2389 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2390 2391 // Update IdealKit memory and control from graphKit. 2392 __ sync_kit(this); 2393 2394 Node* one = __ ConI(1); 2395 // is_instof == 0 if base_oop == NULL 2396 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2397 2398 // Update graphKit from IdeakKit. 2399 sync_kit(ideal); 2400 2401 // Use the pre-barrier to record the value in the referent field 2402 pre_barrier(false /* do_load */, 2403 __ ctrl(), 2404 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2405 pre_val /* pre_val */, 2406 T_OBJECT); 2407 if (need_mem_bar) { 2408 // Add memory barrier to prevent commoning reads from this field 2409 // across safepoint since GC can change its value. 2410 insert_mem_bar(Op_MemBarCPUOrder); 2411 } 2412 // Update IdealKit from graphKit. 2413 __ sync_kit(this); 2414 2415 } __ end_if(); // _ref_type != ref_none 2416 } __ end_if(); // offset == referent_offset 2417 2418 // Final sync IdealKit and GraphKit. 2419 final_sync(ideal); 2420 #undef __ 2421 } 2422 2423 2424 // Interpret Unsafe.fieldOffset cookies correctly: 2425 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2426 2427 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2428 // Attempt to infer a sharper value type from the offset and base type. 2429 ciKlass* sharpened_klass = NULL; 2430 2431 // See if it is an instance field, with an object type. 2432 if (alias_type->field() != NULL) { 2433 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2434 if (alias_type->field()->type()->is_klass()) { 2435 sharpened_klass = alias_type->field()->type()->as_klass(); 2436 } 2437 } 2438 2439 // See if it is a narrow oop array. 2440 if (adr_type->isa_aryptr()) { 2441 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2442 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2443 if (elem_type != NULL) { 2444 sharpened_klass = elem_type->klass(); 2445 } 2446 } 2447 } 2448 2449 // The sharpened class might be unloaded if there is no class loader 2450 // contraint in place. 2451 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2452 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2453 2454 #ifndef PRODUCT 2455 if (C->print_intrinsics() || C->print_inlining()) { 2456 tty->print(" from base type: "); adr_type->dump(); 2457 tty->print(" sharpened value: "); tjp->dump(); 2458 } 2459 #endif 2460 // Sharpen the value type. 2461 return tjp; 2462 } 2463 return NULL; 2464 } 2465 2466 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2467 if (callee()->is_static()) return false; // caller must have the capability! 2468 2469 #ifndef PRODUCT 2470 { 2471 ResourceMark rm; 2472 // Check the signatures. 2473 ciSignature* sig = callee()->signature(); 2474 #ifdef ASSERT 2475 if (!is_store) { 2476 // Object getObject(Object base, int/long offset), etc. 2477 BasicType rtype = sig->return_type()->basic_type(); 2478 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2479 rtype = T_ADDRESS; // it is really a C void* 2480 assert(rtype == type, "getter must return the expected value"); 2481 if (!is_native_ptr) { 2482 assert(sig->count() == 2, "oop getter has 2 arguments"); 2483 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2484 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2485 } else { 2486 assert(sig->count() == 1, "native getter has 1 argument"); 2487 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2488 } 2489 } else { 2490 // void putObject(Object base, int/long offset, Object x), etc. 2491 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2492 if (!is_native_ptr) { 2493 assert(sig->count() == 3, "oop putter has 3 arguments"); 2494 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2495 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2496 } else { 2497 assert(sig->count() == 2, "native putter has 2 arguments"); 2498 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2499 } 2500 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2501 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2502 vtype = T_ADDRESS; // it is really a C void* 2503 assert(vtype == type, "putter must accept the expected value"); 2504 } 2505 #endif // ASSERT 2506 } 2507 #endif //PRODUCT 2508 2509 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2510 2511 Node* receiver = argument(0); // type: oop 2512 2513 // Build address expression. 2514 Node* adr; 2515 Node* heap_base_oop = top(); 2516 Node* offset = top(); 2517 Node* val; 2518 2519 if (!is_native_ptr) { 2520 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2521 Node* base = argument(1); // type: oop 2522 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2523 offset = argument(2); // type: long 2524 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2525 // to be plain byte offsets, which are also the same as those accepted 2526 // by oopDesc::field_base. 2527 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2528 "fieldOffset must be byte-scaled"); 2529 // 32-bit machines ignore the high half! 2530 offset = ConvL2X(offset); 2531 adr = make_unsafe_address(base, offset); 2532 heap_base_oop = base; 2533 val = is_store ? argument(4) : NULL; 2534 } else { 2535 Node* ptr = argument(1); // type: long 2536 ptr = ConvL2X(ptr); // adjust Java long to machine word 2537 adr = make_unsafe_address(NULL, ptr); 2538 val = is_store ? argument(3) : NULL; 2539 } 2540 2541 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2542 2543 // First guess at the value type. 2544 const Type *value_type = Type::get_const_basic_type(type); 2545 2546 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2547 // there was not enough information to nail it down. 2548 Compile::AliasType* alias_type = C->alias_type(adr_type); 2549 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2550 2551 // We will need memory barriers unless we can determine a unique 2552 // alias category for this reference. (Note: If for some reason 2553 // the barriers get omitted and the unsafe reference begins to "pollute" 2554 // the alias analysis of the rest of the graph, either Compile::can_alias 2555 // or Compile::must_alias will throw a diagnostic assert.) 2556 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2557 2558 // If we are reading the value of the referent field of a Reference 2559 // object (either by using Unsafe directly or through reflection) 2560 // then, if G1 is enabled, we need to record the referent in an 2561 // SATB log buffer using the pre-barrier mechanism. 2562 // Also we need to add memory barrier to prevent commoning reads 2563 // from this field across safepoint since GC can change its value. 2564 bool need_read_barrier = !is_native_ptr && !is_store && 2565 offset != top() && heap_base_oop != top(); 2566 2567 if (!is_store && type == T_OBJECT) { 2568 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2569 if (tjp != NULL) { 2570 value_type = tjp; 2571 } 2572 } 2573 2574 receiver = null_check(receiver); 2575 if (stopped()) { 2576 return true; 2577 } 2578 // Heap pointers get a null-check from the interpreter, 2579 // as a courtesy. However, this is not guaranteed by Unsafe, 2580 // and it is not possible to fully distinguish unintended nulls 2581 // from intended ones in this API. 2582 2583 if (is_volatile) { 2584 // We need to emit leading and trailing CPU membars (see below) in 2585 // addition to memory membars when is_volatile. This is a little 2586 // too strong, but avoids the need to insert per-alias-type 2587 // volatile membars (for stores; compare Parse::do_put_xxx), which 2588 // we cannot do effectively here because we probably only have a 2589 // rough approximation of type. 2590 need_mem_bar = true; 2591 // For Stores, place a memory ordering barrier now. 2592 if (is_store) { 2593 insert_mem_bar(Op_MemBarRelease); 2594 } else { 2595 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2596 insert_mem_bar(Op_MemBarVolatile); 2597 } 2598 } 2599 } 2600 2601 // Memory barrier to prevent normal and 'unsafe' accesses from 2602 // bypassing each other. Happens after null checks, so the 2603 // exception paths do not take memory state from the memory barrier, 2604 // so there's no problems making a strong assert about mixing users 2605 // of safe & unsafe memory. 2606 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2607 2608 if (!is_store) { 2609 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2610 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile); 2611 // load value 2612 switch (type) { 2613 case T_BOOLEAN: 2614 case T_CHAR: 2615 case T_BYTE: 2616 case T_SHORT: 2617 case T_INT: 2618 case T_LONG: 2619 case T_FLOAT: 2620 case T_DOUBLE: 2621 break; 2622 case T_OBJECT: 2623 if (need_read_barrier) { 2624 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2625 } 2626 break; 2627 case T_ADDRESS: 2628 // Cast to an int type. 2629 p = _gvn.transform(new CastP2XNode(NULL, p)); 2630 p = ConvX2UL(p); 2631 break; 2632 default: 2633 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2634 break; 2635 } 2636 // The load node has the control of the preceding MemBarCPUOrder. All 2637 // following nodes will have the control of the MemBarCPUOrder inserted at 2638 // the end of this method. So, pushing the load onto the stack at a later 2639 // point is fine. 2640 set_result(p); 2641 } else { 2642 // place effect of store into memory 2643 switch (type) { 2644 case T_DOUBLE: 2645 val = dstore_rounding(val); 2646 break; 2647 case T_ADDRESS: 2648 // Repackage the long as a pointer. 2649 val = ConvL2X(val); 2650 val = _gvn.transform(new CastX2PNode(val)); 2651 break; 2652 } 2653 2654 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2655 if (type != T_OBJECT ) { 2656 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2657 } else { 2658 // Possibly an oop being stored to Java heap or native memory 2659 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2660 // oop to Java heap. 2661 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2662 } else { 2663 // We can't tell at compile time if we are storing in the Java heap or outside 2664 // of it. So we need to emit code to conditionally do the proper type of 2665 // store. 2666 2667 IdealKit ideal(this); 2668 #define __ ideal. 2669 // QQQ who knows what probability is here?? 2670 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2671 // Sync IdealKit and graphKit. 2672 sync_kit(ideal); 2673 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2674 // Update IdealKit memory. 2675 __ sync_kit(this); 2676 } __ else_(); { 2677 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2678 } __ end_if(); 2679 // Final sync IdealKit and GraphKit. 2680 final_sync(ideal); 2681 #undef __ 2682 } 2683 } 2684 } 2685 2686 if (is_volatile) { 2687 if (!is_store) { 2688 insert_mem_bar(Op_MemBarAcquire); 2689 } else { 2690 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2691 insert_mem_bar(Op_MemBarVolatile); 2692 } 2693 } 2694 } 2695 2696 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2697 2698 return true; 2699 } 2700 2701 //----------------------------inline_unsafe_load_store---------------------------- 2702 // This method serves a couple of different customers (depending on LoadStoreKind): 2703 // 2704 // LS_cmpxchg: 2705 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2706 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2707 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2708 // 2709 // LS_xadd: 2710 // public int getAndAddInt( Object o, long offset, int delta) 2711 // public long getAndAddLong(Object o, long offset, long delta) 2712 // 2713 // LS_xchg: 2714 // int getAndSet(Object o, long offset, int newValue) 2715 // long getAndSet(Object o, long offset, long newValue) 2716 // Object getAndSet(Object o, long offset, Object newValue) 2717 // 2718 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2719 // This basic scheme here is the same as inline_unsafe_access, but 2720 // differs in enough details that combining them would make the code 2721 // overly confusing. (This is a true fact! I originally combined 2722 // them, but even I was confused by it!) As much code/comments as 2723 // possible are retained from inline_unsafe_access though to make 2724 // the correspondences clearer. - dl 2725 2726 if (callee()->is_static()) return false; // caller must have the capability! 2727 2728 #ifndef PRODUCT 2729 BasicType rtype; 2730 { 2731 ResourceMark rm; 2732 // Check the signatures. 2733 ciSignature* sig = callee()->signature(); 2734 rtype = sig->return_type()->basic_type(); 2735 if (kind == LS_xadd || kind == LS_xchg) { 2736 // Check the signatures. 2737 #ifdef ASSERT 2738 assert(rtype == type, "get and set must return the expected type"); 2739 assert(sig->count() == 3, "get and set has 3 arguments"); 2740 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2741 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2742 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2743 #endif // ASSERT 2744 } else if (kind == LS_cmpxchg) { 2745 // Check the signatures. 2746 #ifdef ASSERT 2747 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2748 assert(sig->count() == 4, "CAS has 4 arguments"); 2749 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2750 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2751 #endif // ASSERT 2752 } else { 2753 ShouldNotReachHere(); 2754 } 2755 } 2756 #endif //PRODUCT 2757 2758 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2759 2760 // Get arguments: 2761 Node* receiver = NULL; 2762 Node* base = NULL; 2763 Node* offset = NULL; 2764 Node* oldval = NULL; 2765 Node* newval = NULL; 2766 if (kind == LS_cmpxchg) { 2767 const bool two_slot_type = type2size[type] == 2; 2768 receiver = argument(0); // type: oop 2769 base = argument(1); // type: oop 2770 offset = argument(2); // type: long 2771 oldval = argument(4); // type: oop, int, or long 2772 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2773 } else if (kind == LS_xadd || kind == LS_xchg){ 2774 receiver = argument(0); // type: oop 2775 base = argument(1); // type: oop 2776 offset = argument(2); // type: long 2777 oldval = NULL; 2778 newval = argument(4); // type: oop, int, or long 2779 } 2780 2781 // Null check receiver. 2782 receiver = null_check(receiver); 2783 if (stopped()) { 2784 return true; 2785 } 2786 2787 // Build field offset expression. 2788 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2789 // to be plain byte offsets, which are also the same as those accepted 2790 // by oopDesc::field_base. 2791 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2792 // 32-bit machines ignore the high half of long offsets 2793 offset = ConvL2X(offset); 2794 Node* adr = make_unsafe_address(base, offset); 2795 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2796 2797 // For CAS, unlike inline_unsafe_access, there seems no point in 2798 // trying to refine types. Just use the coarse types here. 2799 const Type *value_type = Type::get_const_basic_type(type); 2800 Compile::AliasType* alias_type = C->alias_type(adr_type); 2801 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2802 2803 if (kind == LS_xchg && type == T_OBJECT) { 2804 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2805 if (tjp != NULL) { 2806 value_type = tjp; 2807 } 2808 } 2809 2810 int alias_idx = C->get_alias_index(adr_type); 2811 2812 // Memory-model-wise, a LoadStore acts like a little synchronized 2813 // block, so needs barriers on each side. These don't translate 2814 // into actual barriers on most machines, but we still need rest of 2815 // compiler to respect ordering. 2816 2817 insert_mem_bar(Op_MemBarRelease); 2818 insert_mem_bar(Op_MemBarCPUOrder); 2819 2820 // 4984716: MemBars must be inserted before this 2821 // memory node in order to avoid a false 2822 // dependency which will confuse the scheduler. 2823 Node *mem = memory(alias_idx); 2824 2825 // For now, we handle only those cases that actually exist: ints, 2826 // longs, and Object. Adding others should be straightforward. 2827 Node* load_store; 2828 switch(type) { 2829 case T_INT: 2830 if (kind == LS_xadd) { 2831 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2832 } else if (kind == LS_xchg) { 2833 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2834 } else if (kind == LS_cmpxchg) { 2835 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2836 } else { 2837 ShouldNotReachHere(); 2838 } 2839 break; 2840 case T_LONG: 2841 if (kind == LS_xadd) { 2842 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2843 } else if (kind == LS_xchg) { 2844 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2845 } else if (kind == LS_cmpxchg) { 2846 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2847 } else { 2848 ShouldNotReachHere(); 2849 } 2850 break; 2851 case T_OBJECT: 2852 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2853 // could be delayed during Parse (for example, in adjust_map_after_if()). 2854 // Execute transformation here to avoid barrier generation in such case. 2855 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2856 newval = _gvn.makecon(TypePtr::NULL_PTR); 2857 2858 // Reference stores need a store barrier. 2859 if (kind == LS_xchg) { 2860 // If pre-barrier must execute before the oop store, old value will require do_load here. 2861 if (!can_move_pre_barrier()) { 2862 pre_barrier(true /* do_load*/, 2863 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2864 NULL /* pre_val*/, 2865 T_OBJECT); 2866 } // Else move pre_barrier to use load_store value, see below. 2867 } else if (kind == LS_cmpxchg) { 2868 // Same as for newval above: 2869 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2870 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2871 } 2872 // The only known value which might get overwritten is oldval. 2873 pre_barrier(false /* do_load */, 2874 control(), NULL, NULL, max_juint, NULL, NULL, 2875 oldval /* pre_val */, 2876 T_OBJECT); 2877 } else { 2878 ShouldNotReachHere(); 2879 } 2880 2881 #ifdef _LP64 2882 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2883 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2884 if (kind == LS_xchg) { 2885 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2886 newval_enc, adr_type, value_type->make_narrowoop())); 2887 } else { 2888 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2889 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2890 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2891 newval_enc, oldval_enc)); 2892 } 2893 } else 2894 #endif 2895 { 2896 if (kind == LS_xchg) { 2897 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2898 } else { 2899 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2900 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2901 } 2902 } 2903 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2904 break; 2905 default: 2906 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2907 break; 2908 } 2909 2910 // SCMemProjNodes represent the memory state of a LoadStore. Their 2911 // main role is to prevent LoadStore nodes from being optimized away 2912 // when their results aren't used. 2913 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2914 set_memory(proj, alias_idx); 2915 2916 if (type == T_OBJECT && kind == LS_xchg) { 2917 #ifdef _LP64 2918 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2919 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2920 } 2921 #endif 2922 if (can_move_pre_barrier()) { 2923 // Don't need to load pre_val. The old value is returned by load_store. 2924 // The pre_barrier can execute after the xchg as long as no safepoint 2925 // gets inserted between them. 2926 pre_barrier(false /* do_load */, 2927 control(), NULL, NULL, max_juint, NULL, NULL, 2928 load_store /* pre_val */, 2929 T_OBJECT); 2930 } 2931 } 2932 2933 // Add the trailing membar surrounding the access 2934 insert_mem_bar(Op_MemBarCPUOrder); 2935 insert_mem_bar(Op_MemBarAcquire); 2936 2937 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2938 set_result(load_store); 2939 return true; 2940 } 2941 2942 //----------------------------inline_unsafe_ordered_store---------------------- 2943 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2944 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 2945 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 2946 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2947 // This is another variant of inline_unsafe_access, differing in 2948 // that it always issues store-store ("release") barrier and ensures 2949 // store-atomicity (which only matters for "long"). 2950 2951 if (callee()->is_static()) return false; // caller must have the capability! 2952 2953 #ifndef PRODUCT 2954 { 2955 ResourceMark rm; 2956 // Check the signatures. 2957 ciSignature* sig = callee()->signature(); 2958 #ifdef ASSERT 2959 BasicType rtype = sig->return_type()->basic_type(); 2960 assert(rtype == T_VOID, "must return void"); 2961 assert(sig->count() == 3, "has 3 arguments"); 2962 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2963 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2964 #endif // ASSERT 2965 } 2966 #endif //PRODUCT 2967 2968 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2969 2970 // Get arguments: 2971 Node* receiver = argument(0); // type: oop 2972 Node* base = argument(1); // type: oop 2973 Node* offset = argument(2); // type: long 2974 Node* val = argument(4); // type: oop, int, or long 2975 2976 // Null check receiver. 2977 receiver = null_check(receiver); 2978 if (stopped()) { 2979 return true; 2980 } 2981 2982 // Build field offset expression. 2983 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2984 // 32-bit machines ignore the high half of long offsets 2985 offset = ConvL2X(offset); 2986 Node* adr = make_unsafe_address(base, offset); 2987 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2988 const Type *value_type = Type::get_const_basic_type(type); 2989 Compile::AliasType* alias_type = C->alias_type(adr_type); 2990 2991 insert_mem_bar(Op_MemBarRelease); 2992 insert_mem_bar(Op_MemBarCPUOrder); 2993 // Ensure that the store is atomic for longs: 2994 const bool require_atomic_access = true; 2995 Node* store; 2996 if (type == T_OBJECT) // reference stores need a store barrier. 2997 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 2998 else { 2999 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3000 } 3001 insert_mem_bar(Op_MemBarCPUOrder); 3002 return true; 3003 } 3004 3005 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3006 // Regardless of form, don't allow previous ld/st to move down, 3007 // then issue acquire, release, or volatile mem_bar. 3008 insert_mem_bar(Op_MemBarCPUOrder); 3009 switch(id) { 3010 case vmIntrinsics::_loadFence: 3011 insert_mem_bar(Op_LoadFence); 3012 return true; 3013 case vmIntrinsics::_storeFence: 3014 insert_mem_bar(Op_StoreFence); 3015 return true; 3016 case vmIntrinsics::_fullFence: 3017 insert_mem_bar(Op_MemBarVolatile); 3018 return true; 3019 default: 3020 fatal_unexpected_iid(id); 3021 return false; 3022 } 3023 } 3024 3025 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3026 if (!kls->is_Con()) { 3027 return true; 3028 } 3029 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3030 if (klsptr == NULL) { 3031 return true; 3032 } 3033 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3034 // don't need a guard for a klass that is already initialized 3035 return !ik->is_initialized(); 3036 } 3037 3038 //----------------------------inline_unsafe_allocate--------------------------- 3039 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3040 bool LibraryCallKit::inline_unsafe_allocate() { 3041 if (callee()->is_static()) return false; // caller must have the capability! 3042 3043 null_check_receiver(); // null-check, then ignore 3044 Node* cls = null_check(argument(1)); 3045 if (stopped()) return true; 3046 3047 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3048 kls = null_check(kls); 3049 if (stopped()) return true; // argument was like int.class 3050 3051 Node* test = NULL; 3052 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3053 // Note: The argument might still be an illegal value like 3054 // Serializable.class or Object[].class. The runtime will handle it. 3055 // But we must make an explicit check for initialization. 3056 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3057 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3058 // can generate code to load it as unsigned byte. 3059 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3060 Node* bits = intcon(InstanceKlass::fully_initialized); 3061 test = _gvn.transform(new SubINode(inst, bits)); 3062 // The 'test' is non-zero if we need to take a slow path. 3063 } 3064 3065 Node* obj = new_instance(kls, test); 3066 set_result(obj); 3067 return true; 3068 } 3069 3070 #ifdef TRACE_HAVE_INTRINSICS 3071 /* 3072 * oop -> myklass 3073 * myklass->trace_id |= USED 3074 * return myklass->trace_id & ~0x3 3075 */ 3076 bool LibraryCallKit::inline_native_classID() { 3077 null_check_receiver(); // null-check, then ignore 3078 Node* cls = null_check(argument(1), T_OBJECT); 3079 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3080 kls = null_check(kls, T_OBJECT); 3081 ByteSize offset = TRACE_ID_OFFSET; 3082 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3083 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3084 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3085 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3086 Node* clsused = longcon(0x01l); // set the class bit 3087 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3088 3089 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3090 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3091 set_result(andl); 3092 return true; 3093 } 3094 3095 bool LibraryCallKit::inline_native_threadID() { 3096 Node* tls_ptr = NULL; 3097 Node* cur_thr = generate_current_thread(tls_ptr); 3098 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3099 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3100 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3101 3102 Node* threadid = NULL; 3103 size_t thread_id_size = OSThread::thread_id_size(); 3104 if (thread_id_size == (size_t) BytesPerLong) { 3105 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3106 } else if (thread_id_size == (size_t) BytesPerInt) { 3107 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3108 } else { 3109 ShouldNotReachHere(); 3110 } 3111 set_result(threadid); 3112 return true; 3113 } 3114 #endif 3115 3116 //------------------------inline_native_time_funcs-------------- 3117 // inline code for System.currentTimeMillis() and System.nanoTime() 3118 // these have the same type and signature 3119 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3120 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3121 const TypePtr* no_memory_effects = NULL; 3122 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3123 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3124 #ifdef ASSERT 3125 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3126 assert(value_top == top(), "second value must be top"); 3127 #endif 3128 set_result(value); 3129 return true; 3130 } 3131 3132 //------------------------inline_native_currentThread------------------ 3133 bool LibraryCallKit::inline_native_currentThread() { 3134 Node* junk = NULL; 3135 set_result(generate_current_thread(junk)); 3136 return true; 3137 } 3138 3139 //------------------------inline_native_isInterrupted------------------ 3140 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3141 bool LibraryCallKit::inline_native_isInterrupted() { 3142 // Add a fast path to t.isInterrupted(clear_int): 3143 // (t == Thread.current() && 3144 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3145 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3146 // So, in the common case that the interrupt bit is false, 3147 // we avoid making a call into the VM. Even if the interrupt bit 3148 // is true, if the clear_int argument is false, we avoid the VM call. 3149 // However, if the receiver is not currentThread, we must call the VM, 3150 // because there must be some locking done around the operation. 3151 3152 // We only go to the fast case code if we pass two guards. 3153 // Paths which do not pass are accumulated in the slow_region. 3154 3155 enum { 3156 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3157 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3158 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3159 PATH_LIMIT 3160 }; 3161 3162 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3163 // out of the function. 3164 insert_mem_bar(Op_MemBarCPUOrder); 3165 3166 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3167 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3168 3169 RegionNode* slow_region = new RegionNode(1); 3170 record_for_igvn(slow_region); 3171 3172 // (a) Receiving thread must be the current thread. 3173 Node* rec_thr = argument(0); 3174 Node* tls_ptr = NULL; 3175 Node* cur_thr = generate_current_thread(tls_ptr); 3176 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3177 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3178 3179 generate_slow_guard(bol_thr, slow_region); 3180 3181 // (b) Interrupt bit on TLS must be false. 3182 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3183 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3184 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3185 3186 // Set the control input on the field _interrupted read to prevent it floating up. 3187 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3188 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3189 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3190 3191 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3192 3193 // First fast path: if (!TLS._interrupted) return false; 3194 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3195 result_rgn->init_req(no_int_result_path, false_bit); 3196 result_val->init_req(no_int_result_path, intcon(0)); 3197 3198 // drop through to next case 3199 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3200 3201 #ifndef TARGET_OS_FAMILY_windows 3202 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3203 Node* clr_arg = argument(1); 3204 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3205 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3206 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3207 3208 // Second fast path: ... else if (!clear_int) return true; 3209 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3210 result_rgn->init_req(no_clear_result_path, false_arg); 3211 result_val->init_req(no_clear_result_path, intcon(1)); 3212 3213 // drop through to next case 3214 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3215 #else 3216 // To return true on Windows you must read the _interrupted field 3217 // and check the the event state i.e. take the slow path. 3218 #endif // TARGET_OS_FAMILY_windows 3219 3220 // (d) Otherwise, go to the slow path. 3221 slow_region->add_req(control()); 3222 set_control( _gvn.transform(slow_region)); 3223 3224 if (stopped()) { 3225 // There is no slow path. 3226 result_rgn->init_req(slow_result_path, top()); 3227 result_val->init_req(slow_result_path, top()); 3228 } else { 3229 // non-virtual because it is a private non-static 3230 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3231 3232 Node* slow_val = set_results_for_java_call(slow_call); 3233 // this->control() comes from set_results_for_java_call 3234 3235 Node* fast_io = slow_call->in(TypeFunc::I_O); 3236 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3237 3238 // These two phis are pre-filled with copies of of the fast IO and Memory 3239 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3240 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3241 3242 result_rgn->init_req(slow_result_path, control()); 3243 result_io ->init_req(slow_result_path, i_o()); 3244 result_mem->init_req(slow_result_path, reset_memory()); 3245 result_val->init_req(slow_result_path, slow_val); 3246 3247 set_all_memory(_gvn.transform(result_mem)); 3248 set_i_o( _gvn.transform(result_io)); 3249 } 3250 3251 C->set_has_split_ifs(true); // Has chance for split-if optimization 3252 set_result(result_rgn, result_val); 3253 return true; 3254 } 3255 3256 //---------------------------load_mirror_from_klass---------------------------- 3257 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3258 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3259 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3260 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3261 } 3262 3263 //-----------------------load_klass_from_mirror_common------------------------- 3264 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3265 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3266 // and branch to the given path on the region. 3267 // If never_see_null, take an uncommon trap on null, so we can optimistically 3268 // compile for the non-null case. 3269 // If the region is NULL, force never_see_null = true. 3270 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3271 bool never_see_null, 3272 RegionNode* region, 3273 int null_path, 3274 int offset) { 3275 if (region == NULL) never_see_null = true; 3276 Node* p = basic_plus_adr(mirror, offset); 3277 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3278 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3279 Node* null_ctl = top(); 3280 kls = null_check_oop(kls, &null_ctl, never_see_null); 3281 if (region != NULL) { 3282 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3283 region->init_req(null_path, null_ctl); 3284 } else { 3285 assert(null_ctl == top(), "no loose ends"); 3286 } 3287 return kls; 3288 } 3289 3290 //--------------------(inline_native_Class_query helpers)--------------------- 3291 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3292 // Fall through if (mods & mask) == bits, take the guard otherwise. 3293 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3294 // Branch around if the given klass has the given modifier bit set. 3295 // Like generate_guard, adds a new path onto the region. 3296 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3297 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3298 Node* mask = intcon(modifier_mask); 3299 Node* bits = intcon(modifier_bits); 3300 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3301 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3302 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3303 return generate_fair_guard(bol, region); 3304 } 3305 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3306 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3307 } 3308 3309 //-------------------------inline_native_Class_query------------------- 3310 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3311 const Type* return_type = TypeInt::BOOL; 3312 Node* prim_return_value = top(); // what happens if it's a primitive class? 3313 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3314 bool expect_prim = false; // most of these guys expect to work on refs 3315 3316 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3317 3318 Node* mirror = argument(0); 3319 Node* obj = top(); 3320 3321 switch (id) { 3322 case vmIntrinsics::_isInstance: 3323 // nothing is an instance of a primitive type 3324 prim_return_value = intcon(0); 3325 obj = argument(1); 3326 break; 3327 case vmIntrinsics::_getModifiers: 3328 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3329 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3330 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3331 break; 3332 case vmIntrinsics::_isInterface: 3333 prim_return_value = intcon(0); 3334 break; 3335 case vmIntrinsics::_isArray: 3336 prim_return_value = intcon(0); 3337 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3338 break; 3339 case vmIntrinsics::_isPrimitive: 3340 prim_return_value = intcon(1); 3341 expect_prim = true; // obviously 3342 break; 3343 case vmIntrinsics::_getSuperclass: 3344 prim_return_value = null(); 3345 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3346 break; 3347 case vmIntrinsics::_getClassAccessFlags: 3348 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3349 return_type = TypeInt::INT; // not bool! 6297094 3350 break; 3351 default: 3352 fatal_unexpected_iid(id); 3353 break; 3354 } 3355 3356 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3357 if (mirror_con == NULL) return false; // cannot happen? 3358 3359 #ifndef PRODUCT 3360 if (C->print_intrinsics() || C->print_inlining()) { 3361 ciType* k = mirror_con->java_mirror_type(); 3362 if (k) { 3363 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3364 k->print_name(); 3365 tty->cr(); 3366 } 3367 } 3368 #endif 3369 3370 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3371 RegionNode* region = new RegionNode(PATH_LIMIT); 3372 record_for_igvn(region); 3373 PhiNode* phi = new PhiNode(region, return_type); 3374 3375 // The mirror will never be null of Reflection.getClassAccessFlags, however 3376 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3377 // if it is. See bug 4774291. 3378 3379 // For Reflection.getClassAccessFlags(), the null check occurs in 3380 // the wrong place; see inline_unsafe_access(), above, for a similar 3381 // situation. 3382 mirror = null_check(mirror); 3383 // If mirror or obj is dead, only null-path is taken. 3384 if (stopped()) return true; 3385 3386 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3387 3388 // Now load the mirror's klass metaobject, and null-check it. 3389 // Side-effects region with the control path if the klass is null. 3390 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3391 // If kls is null, we have a primitive mirror. 3392 phi->init_req(_prim_path, prim_return_value); 3393 if (stopped()) { set_result(region, phi); return true; } 3394 bool safe_for_replace = (region->in(_prim_path) == top()); 3395 3396 Node* p; // handy temp 3397 Node* null_ctl; 3398 3399 // Now that we have the non-null klass, we can perform the real query. 3400 // For constant classes, the query will constant-fold in LoadNode::Value. 3401 Node* query_value = top(); 3402 switch (id) { 3403 case vmIntrinsics::_isInstance: 3404 // nothing is an instance of a primitive type 3405 query_value = gen_instanceof(obj, kls, safe_for_replace); 3406 break; 3407 3408 case vmIntrinsics::_getModifiers: 3409 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3410 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3411 break; 3412 3413 case vmIntrinsics::_isInterface: 3414 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3415 if (generate_interface_guard(kls, region) != NULL) 3416 // A guard was added. If the guard is taken, it was an interface. 3417 phi->add_req(intcon(1)); 3418 // If we fall through, it's a plain class. 3419 query_value = intcon(0); 3420 break; 3421 3422 case vmIntrinsics::_isArray: 3423 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3424 if (generate_array_guard(kls, region) != NULL) 3425 // A guard was added. If the guard is taken, it was an array. 3426 phi->add_req(intcon(1)); 3427 // If we fall through, it's a plain class. 3428 query_value = intcon(0); 3429 break; 3430 3431 case vmIntrinsics::_isPrimitive: 3432 query_value = intcon(0); // "normal" path produces false 3433 break; 3434 3435 case vmIntrinsics::_getSuperclass: 3436 // The rules here are somewhat unfortunate, but we can still do better 3437 // with random logic than with a JNI call. 3438 // Interfaces store null or Object as _super, but must report null. 3439 // Arrays store an intermediate super as _super, but must report Object. 3440 // Other types can report the actual _super. 3441 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3442 if (generate_interface_guard(kls, region) != NULL) 3443 // A guard was added. If the guard is taken, it was an interface. 3444 phi->add_req(null()); 3445 if (generate_array_guard(kls, region) != NULL) 3446 // A guard was added. If the guard is taken, it was an array. 3447 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3448 // If we fall through, it's a plain class. Get its _super. 3449 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3450 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3451 null_ctl = top(); 3452 kls = null_check_oop(kls, &null_ctl); 3453 if (null_ctl != top()) { 3454 // If the guard is taken, Object.superClass is null (both klass and mirror). 3455 region->add_req(null_ctl); 3456 phi ->add_req(null()); 3457 } 3458 if (!stopped()) { 3459 query_value = load_mirror_from_klass(kls); 3460 } 3461 break; 3462 3463 case vmIntrinsics::_getClassAccessFlags: 3464 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3465 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3466 break; 3467 3468 default: 3469 fatal_unexpected_iid(id); 3470 break; 3471 } 3472 3473 // Fall-through is the normal case of a query to a real class. 3474 phi->init_req(1, query_value); 3475 region->init_req(1, control()); 3476 3477 C->set_has_split_ifs(true); // Has chance for split-if optimization 3478 set_result(region, phi); 3479 return true; 3480 } 3481 3482 //-------------------------inline_Class_cast------------------- 3483 bool LibraryCallKit::inline_Class_cast() { 3484 Node* mirror = argument(0); // Class 3485 Node* obj = argument(1); 3486 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3487 if (mirror_con == NULL) { 3488 return false; // dead path (mirror->is_top()). 3489 } 3490 if (obj == NULL || obj->is_top()) { 3491 return false; // dead path 3492 } 3493 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3494 3495 // First, see if Class.cast() can be folded statically. 3496 // java_mirror_type() returns non-null for compile-time Class constants. 3497 ciType* tm = mirror_con->java_mirror_type(); 3498 if (tm != NULL && tm->is_klass() && 3499 tp != NULL && tp->klass() != NULL) { 3500 if (!tp->klass()->is_loaded()) { 3501 // Don't use intrinsic when class is not loaded. 3502 return false; 3503 } else { 3504 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3505 if (static_res == Compile::SSC_always_true) { 3506 // isInstance() is true - fold the code. 3507 set_result(obj); 3508 return true; 3509 } else if (static_res == Compile::SSC_always_false) { 3510 // Don't use intrinsic, have to throw ClassCastException. 3511 // If the reference is null, the non-intrinsic bytecode will 3512 // be optimized appropriately. 3513 return false; 3514 } 3515 } 3516 } 3517 3518 // Bailout intrinsic and do normal inlining if exception path is frequent. 3519 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3520 return false; 3521 } 3522 3523 // Generate dynamic checks. 3524 // Class.cast() is java implementation of _checkcast bytecode. 3525 // Do checkcast (Parse::do_checkcast()) optimizations here. 3526 3527 mirror = null_check(mirror); 3528 // If mirror is dead, only null-path is taken. 3529 if (stopped()) { 3530 return true; 3531 } 3532 3533 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3534 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3535 RegionNode* region = new RegionNode(PATH_LIMIT); 3536 record_for_igvn(region); 3537 3538 // Now load the mirror's klass metaobject, and null-check it. 3539 // If kls is null, we have a primitive mirror and 3540 // nothing is an instance of a primitive type. 3541 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3542 3543 Node* res = top(); 3544 if (!stopped()) { 3545 Node* bad_type_ctrl = top(); 3546 // Do checkcast optimizations. 3547 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3548 region->init_req(_bad_type_path, bad_type_ctrl); 3549 } 3550 if (region->in(_prim_path) != top() || 3551 region->in(_bad_type_path) != top()) { 3552 // Let Interpreter throw ClassCastException. 3553 PreserveJVMState pjvms(this); 3554 set_control(_gvn.transform(region)); 3555 uncommon_trap(Deoptimization::Reason_intrinsic, 3556 Deoptimization::Action_maybe_recompile); 3557 } 3558 if (!stopped()) { 3559 set_result(res); 3560 } 3561 return true; 3562 } 3563 3564 3565 //--------------------------inline_native_subtype_check------------------------ 3566 // This intrinsic takes the JNI calls out of the heart of 3567 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3568 bool LibraryCallKit::inline_native_subtype_check() { 3569 // Pull both arguments off the stack. 3570 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3571 args[0] = argument(0); 3572 args[1] = argument(1); 3573 Node* klasses[2]; // corresponding Klasses: superk, subk 3574 klasses[0] = klasses[1] = top(); 3575 3576 enum { 3577 // A full decision tree on {superc is prim, subc is prim}: 3578 _prim_0_path = 1, // {P,N} => false 3579 // {P,P} & superc!=subc => false 3580 _prim_same_path, // {P,P} & superc==subc => true 3581 _prim_1_path, // {N,P} => false 3582 _ref_subtype_path, // {N,N} & subtype check wins => true 3583 _both_ref_path, // {N,N} & subtype check loses => false 3584 PATH_LIMIT 3585 }; 3586 3587 RegionNode* region = new RegionNode(PATH_LIMIT); 3588 Node* phi = new PhiNode(region, TypeInt::BOOL); 3589 record_for_igvn(region); 3590 3591 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3592 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3593 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3594 3595 // First null-check both mirrors and load each mirror's klass metaobject. 3596 int which_arg; 3597 for (which_arg = 0; which_arg <= 1; which_arg++) { 3598 Node* arg = args[which_arg]; 3599 arg = null_check(arg); 3600 if (stopped()) break; 3601 args[which_arg] = arg; 3602 3603 Node* p = basic_plus_adr(arg, class_klass_offset); 3604 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3605 klasses[which_arg] = _gvn.transform(kls); 3606 } 3607 3608 // Having loaded both klasses, test each for null. 3609 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3610 for (which_arg = 0; which_arg <= 1; which_arg++) { 3611 Node* kls = klasses[which_arg]; 3612 Node* null_ctl = top(); 3613 kls = null_check_oop(kls, &null_ctl, never_see_null); 3614 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3615 region->init_req(prim_path, null_ctl); 3616 if (stopped()) break; 3617 klasses[which_arg] = kls; 3618 } 3619 3620 if (!stopped()) { 3621 // now we have two reference types, in klasses[0..1] 3622 Node* subk = klasses[1]; // the argument to isAssignableFrom 3623 Node* superk = klasses[0]; // the receiver 3624 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3625 // now we have a successful reference subtype check 3626 region->set_req(_ref_subtype_path, control()); 3627 } 3628 3629 // If both operands are primitive (both klasses null), then 3630 // we must return true when they are identical primitives. 3631 // It is convenient to test this after the first null klass check. 3632 set_control(region->in(_prim_0_path)); // go back to first null check 3633 if (!stopped()) { 3634 // Since superc is primitive, make a guard for the superc==subc case. 3635 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3636 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3637 generate_guard(bol_eq, region, PROB_FAIR); 3638 if (region->req() == PATH_LIMIT+1) { 3639 // A guard was added. If the added guard is taken, superc==subc. 3640 region->swap_edges(PATH_LIMIT, _prim_same_path); 3641 region->del_req(PATH_LIMIT); 3642 } 3643 region->set_req(_prim_0_path, control()); // Not equal after all. 3644 } 3645 3646 // these are the only paths that produce 'true': 3647 phi->set_req(_prim_same_path, intcon(1)); 3648 phi->set_req(_ref_subtype_path, intcon(1)); 3649 3650 // pull together the cases: 3651 assert(region->req() == PATH_LIMIT, "sane region"); 3652 for (uint i = 1; i < region->req(); i++) { 3653 Node* ctl = region->in(i); 3654 if (ctl == NULL || ctl == top()) { 3655 region->set_req(i, top()); 3656 phi ->set_req(i, top()); 3657 } else if (phi->in(i) == NULL) { 3658 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3659 } 3660 } 3661 3662 set_control(_gvn.transform(region)); 3663 set_result(_gvn.transform(phi)); 3664 return true; 3665 } 3666 3667 //---------------------generate_array_guard_common------------------------ 3668 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3669 bool obj_array, bool not_array) { 3670 // If obj_array/non_array==false/false: 3671 // Branch around if the given klass is in fact an array (either obj or prim). 3672 // If obj_array/non_array==false/true: 3673 // Branch around if the given klass is not an array klass of any kind. 3674 // If obj_array/non_array==true/true: 3675 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3676 // If obj_array/non_array==true/false: 3677 // Branch around if the kls is an oop array (Object[] or subtype) 3678 // 3679 // Like generate_guard, adds a new path onto the region. 3680 jint layout_con = 0; 3681 Node* layout_val = get_layout_helper(kls, layout_con); 3682 if (layout_val == NULL) { 3683 bool query = (obj_array 3684 ? Klass::layout_helper_is_objArray(layout_con) 3685 : Klass::layout_helper_is_array(layout_con)); 3686 if (query == not_array) { 3687 return NULL; // never a branch 3688 } else { // always a branch 3689 Node* always_branch = control(); 3690 if (region != NULL) 3691 region->add_req(always_branch); 3692 set_control(top()); 3693 return always_branch; 3694 } 3695 } 3696 // Now test the correct condition. 3697 jint nval = (obj_array 3698 ? ((jint)Klass::_lh_array_tag_type_value 3699 << Klass::_lh_array_tag_shift) 3700 : Klass::_lh_neutral_value); 3701 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3702 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3703 // invert the test if we are looking for a non-array 3704 if (not_array) btest = BoolTest(btest).negate(); 3705 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3706 return generate_fair_guard(bol, region); 3707 } 3708 3709 3710 //-----------------------inline_native_newArray-------------------------- 3711 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3712 bool LibraryCallKit::inline_native_newArray() { 3713 Node* mirror = argument(0); 3714 Node* count_val = argument(1); 3715 3716 mirror = null_check(mirror); 3717 // If mirror or obj is dead, only null-path is taken. 3718 if (stopped()) return true; 3719 3720 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3721 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3722 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3723 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3724 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3725 3726 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3727 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3728 result_reg, _slow_path); 3729 Node* normal_ctl = control(); 3730 Node* no_array_ctl = result_reg->in(_slow_path); 3731 3732 // Generate code for the slow case. We make a call to newArray(). 3733 set_control(no_array_ctl); 3734 if (!stopped()) { 3735 // Either the input type is void.class, or else the 3736 // array klass has not yet been cached. Either the 3737 // ensuing call will throw an exception, or else it 3738 // will cache the array klass for next time. 3739 PreserveJVMState pjvms(this); 3740 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3741 Node* slow_result = set_results_for_java_call(slow_call); 3742 // this->control() comes from set_results_for_java_call 3743 result_reg->set_req(_slow_path, control()); 3744 result_val->set_req(_slow_path, slow_result); 3745 result_io ->set_req(_slow_path, i_o()); 3746 result_mem->set_req(_slow_path, reset_memory()); 3747 } 3748 3749 set_control(normal_ctl); 3750 if (!stopped()) { 3751 // Normal case: The array type has been cached in the java.lang.Class. 3752 // The following call works fine even if the array type is polymorphic. 3753 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3754 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3755 result_reg->init_req(_normal_path, control()); 3756 result_val->init_req(_normal_path, obj); 3757 result_io ->init_req(_normal_path, i_o()); 3758 result_mem->init_req(_normal_path, reset_memory()); 3759 } 3760 3761 // Return the combined state. 3762 set_i_o( _gvn.transform(result_io) ); 3763 set_all_memory( _gvn.transform(result_mem)); 3764 3765 C->set_has_split_ifs(true); // Has chance for split-if optimization 3766 set_result(result_reg, result_val); 3767 return true; 3768 } 3769 3770 //----------------------inline_native_getLength-------------------------- 3771 // public static native int java.lang.reflect.Array.getLength(Object array); 3772 bool LibraryCallKit::inline_native_getLength() { 3773 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3774 3775 Node* array = null_check(argument(0)); 3776 // If array is dead, only null-path is taken. 3777 if (stopped()) return true; 3778 3779 // Deoptimize if it is a non-array. 3780 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3781 3782 if (non_array != NULL) { 3783 PreserveJVMState pjvms(this); 3784 set_control(non_array); 3785 uncommon_trap(Deoptimization::Reason_intrinsic, 3786 Deoptimization::Action_maybe_recompile); 3787 } 3788 3789 // If control is dead, only non-array-path is taken. 3790 if (stopped()) return true; 3791 3792 // The works fine even if the array type is polymorphic. 3793 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3794 Node* result = load_array_length(array); 3795 3796 C->set_has_split_ifs(true); // Has chance for split-if optimization 3797 set_result(result); 3798 return true; 3799 } 3800 3801 //------------------------inline_array_copyOf---------------------------- 3802 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3803 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3804 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3805 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3806 3807 // Get the arguments. 3808 Node* original = argument(0); 3809 Node* start = is_copyOfRange? argument(1): intcon(0); 3810 Node* end = is_copyOfRange? argument(2): argument(1); 3811 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3812 3813 Node* newcopy; 3814 3815 // Set the original stack and the reexecute bit for the interpreter to reexecute 3816 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3817 { PreserveReexecuteState preexecs(this); 3818 jvms()->set_should_reexecute(true); 3819 3820 array_type_mirror = null_check(array_type_mirror); 3821 original = null_check(original); 3822 3823 // Check if a null path was taken unconditionally. 3824 if (stopped()) return true; 3825 3826 Node* orig_length = load_array_length(original); 3827 3828 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3829 klass_node = null_check(klass_node); 3830 3831 RegionNode* bailout = new RegionNode(1); 3832 record_for_igvn(bailout); 3833 3834 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3835 // Bail out if that is so. 3836 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3837 if (not_objArray != NULL) { 3838 // Improve the klass node's type from the new optimistic assumption: 3839 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3840 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3841 Node* cast = new CastPPNode(klass_node, akls); 3842 cast->init_req(0, control()); 3843 klass_node = _gvn.transform(cast); 3844 } 3845 3846 // Bail out if either start or end is negative. 3847 generate_negative_guard(start, bailout, &start); 3848 generate_negative_guard(end, bailout, &end); 3849 3850 Node* length = end; 3851 if (_gvn.type(start) != TypeInt::ZERO) { 3852 length = _gvn.transform(new SubINode(end, start)); 3853 } 3854 3855 // Bail out if length is negative. 3856 // Without this the new_array would throw 3857 // NegativeArraySizeException but IllegalArgumentException is what 3858 // should be thrown 3859 generate_negative_guard(length, bailout, &length); 3860 3861 if (bailout->req() > 1) { 3862 PreserveJVMState pjvms(this); 3863 set_control(_gvn.transform(bailout)); 3864 uncommon_trap(Deoptimization::Reason_intrinsic, 3865 Deoptimization::Action_maybe_recompile); 3866 } 3867 3868 if (!stopped()) { 3869 // How many elements will we copy from the original? 3870 // The answer is MinI(orig_length - start, length). 3871 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3872 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3873 3874 // Generate a direct call to the right arraycopy function(s). 3875 // We know the copy is disjoint but we might not know if the 3876 // oop stores need checking. 3877 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3878 // This will fail a store-check if x contains any non-nulls. 3879 3880 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3881 // loads/stores but it is legal only if we're sure the 3882 // Arrays.copyOf would succeed. So we need all input arguments 3883 // to the copyOf to be validated, including that the copy to the 3884 // new array won't trigger an ArrayStoreException. That subtype 3885 // check can be optimized if we know something on the type of 3886 // the input array from type speculation. 3887 if (_gvn.type(klass_node)->singleton()) { 3888 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3889 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3890 3891 int test = C->static_subtype_check(superk, subk); 3892 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3893 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3894 if (t_original->speculative_type() != NULL) { 3895 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3896 } 3897 } 3898 } 3899 3900 bool validated = false; 3901 // Reason_class_check rather than Reason_intrinsic because we 3902 // want to intrinsify even if this traps. 3903 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3904 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3905 klass_node); 3906 3907 if (not_subtype_ctrl != top()) { 3908 PreserveJVMState pjvms(this); 3909 set_control(not_subtype_ctrl); 3910 uncommon_trap(Deoptimization::Reason_class_check, 3911 Deoptimization::Action_make_not_entrant); 3912 assert(stopped(), "Should be stopped"); 3913 } 3914 validated = true; 3915 } 3916 3917 newcopy = new_array(klass_node, length, 0); // no arguments to push 3918 3919 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3920 load_object_klass(original), klass_node); 3921 if (!is_copyOfRange) { 3922 ac->set_copyof(validated); 3923 } else { 3924 ac->set_copyofrange(validated); 3925 } 3926 Node* n = _gvn.transform(ac); 3927 if (n == ac) { 3928 ac->connect_outputs(this); 3929 } else { 3930 assert(validated, "shouldn't transform if all arguments not validated"); 3931 set_all_memory(n); 3932 } 3933 } 3934 } // original reexecute is set back here 3935 3936 C->set_has_split_ifs(true); // Has chance for split-if optimization 3937 if (!stopped()) { 3938 set_result(newcopy); 3939 } 3940 return true; 3941 } 3942 3943 3944 //----------------------generate_virtual_guard--------------------------- 3945 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3946 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3947 RegionNode* slow_region) { 3948 ciMethod* method = callee(); 3949 int vtable_index = method->vtable_index(); 3950 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3951 err_msg_res("bad index %d", vtable_index)); 3952 // Get the Method* out of the appropriate vtable entry. 3953 int entry_offset = (InstanceKlass::vtable_start_offset() + 3954 vtable_index*vtableEntry::size()) * wordSize + 3955 vtableEntry::method_offset_in_bytes(); 3956 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3957 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3958 3959 // Compare the target method with the expected method (e.g., Object.hashCode). 3960 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3961 3962 Node* native_call = makecon(native_call_addr); 3963 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3964 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3965 3966 return generate_slow_guard(test_native, slow_region); 3967 } 3968 3969 //-----------------------generate_method_call---------------------------- 3970 // Use generate_method_call to make a slow-call to the real 3971 // method if the fast path fails. An alternative would be to 3972 // use a stub like OptoRuntime::slow_arraycopy_Java. 3973 // This only works for expanding the current library call, 3974 // not another intrinsic. (E.g., don't use this for making an 3975 // arraycopy call inside of the copyOf intrinsic.) 3976 CallJavaNode* 3977 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3978 // When compiling the intrinsic method itself, do not use this technique. 3979 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3980 3981 ciMethod* method = callee(); 3982 // ensure the JVMS we have will be correct for this call 3983 guarantee(method_id == method->intrinsic_id(), "must match"); 3984 3985 const TypeFunc* tf = TypeFunc::make(method); 3986 CallJavaNode* slow_call; 3987 if (is_static) { 3988 assert(!is_virtual, ""); 3989 slow_call = new CallStaticJavaNode(C, tf, 3990 SharedRuntime::get_resolve_static_call_stub(), 3991 method, bci()); 3992 } else if (is_virtual) { 3993 null_check_receiver(); 3994 int vtable_index = Method::invalid_vtable_index; 3995 if (UseInlineCaches) { 3996 // Suppress the vtable call 3997 } else { 3998 // hashCode and clone are not a miranda methods, 3999 // so the vtable index is fixed. 4000 // No need to use the linkResolver to get it. 4001 vtable_index = method->vtable_index(); 4002 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4003 err_msg_res("bad index %d", vtable_index)); 4004 } 4005 slow_call = new CallDynamicJavaNode(tf, 4006 SharedRuntime::get_resolve_virtual_call_stub(), 4007 method, vtable_index, bci()); 4008 } else { // neither virtual nor static: opt_virtual 4009 null_check_receiver(); 4010 slow_call = new CallStaticJavaNode(C, tf, 4011 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4012 method, bci()); 4013 slow_call->set_optimized_virtual(true); 4014 } 4015 set_arguments_for_java_call(slow_call); 4016 set_edges_for_java_call(slow_call); 4017 return slow_call; 4018 } 4019 4020 4021 /** 4022 * Build special case code for calls to hashCode on an object. This call may 4023 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4024 * slightly different code. 4025 */ 4026 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4027 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4028 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4029 4030 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4031 4032 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4033 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4034 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4035 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4036 Node* obj = NULL; 4037 if (!is_static) { 4038 // Check for hashing null object 4039 obj = null_check_receiver(); 4040 if (stopped()) return true; // unconditionally null 4041 result_reg->init_req(_null_path, top()); 4042 result_val->init_req(_null_path, top()); 4043 } else { 4044 // Do a null check, and return zero if null. 4045 // System.identityHashCode(null) == 0 4046 obj = argument(0); 4047 Node* null_ctl = top(); 4048 obj = null_check_oop(obj, &null_ctl); 4049 result_reg->init_req(_null_path, null_ctl); 4050 result_val->init_req(_null_path, _gvn.intcon(0)); 4051 } 4052 4053 // Unconditionally null? Then return right away. 4054 if (stopped()) { 4055 set_control( result_reg->in(_null_path)); 4056 if (!stopped()) 4057 set_result(result_val->in(_null_path)); 4058 return true; 4059 } 4060 4061 // We only go to the fast case code if we pass a number of guards. The 4062 // paths which do not pass are accumulated in the slow_region. 4063 RegionNode* slow_region = new RegionNode(1); 4064 record_for_igvn(slow_region); 4065 4066 // If this is a virtual call, we generate a funny guard. We pull out 4067 // the vtable entry corresponding to hashCode() from the target object. 4068 // If the target method which we are calling happens to be the native 4069 // Object hashCode() method, we pass the guard. We do not need this 4070 // guard for non-virtual calls -- the caller is known to be the native 4071 // Object hashCode(). 4072 if (is_virtual) { 4073 // After null check, get the object's klass. 4074 Node* obj_klass = load_object_klass(obj); 4075 generate_virtual_guard(obj_klass, slow_region); 4076 } 4077 4078 // Get the header out of the object, use LoadMarkNode when available 4079 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4080 // The control of the load must be NULL. Otherwise, the load can move before 4081 // the null check after castPP removal. 4082 Node* no_ctrl = NULL; 4083 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4084 4085 // Test the header to see if it is unlocked. 4086 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4087 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4088 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4089 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4090 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4091 4092 generate_slow_guard(test_unlocked, slow_region); 4093 4094 // Get the hash value and check to see that it has been properly assigned. 4095 // We depend on hash_mask being at most 32 bits and avoid the use of 4096 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4097 // vm: see markOop.hpp. 4098 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4099 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4100 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4101 // This hack lets the hash bits live anywhere in the mark object now, as long 4102 // as the shift drops the relevant bits into the low 32 bits. Note that 4103 // Java spec says that HashCode is an int so there's no point in capturing 4104 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4105 hshifted_header = ConvX2I(hshifted_header); 4106 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4107 4108 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4109 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4110 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4111 4112 generate_slow_guard(test_assigned, slow_region); 4113 4114 Node* init_mem = reset_memory(); 4115 // fill in the rest of the null path: 4116 result_io ->init_req(_null_path, i_o()); 4117 result_mem->init_req(_null_path, init_mem); 4118 4119 result_val->init_req(_fast_path, hash_val); 4120 result_reg->init_req(_fast_path, control()); 4121 result_io ->init_req(_fast_path, i_o()); 4122 result_mem->init_req(_fast_path, init_mem); 4123 4124 // Generate code for the slow case. We make a call to hashCode(). 4125 set_control(_gvn.transform(slow_region)); 4126 if (!stopped()) { 4127 // No need for PreserveJVMState, because we're using up the present state. 4128 set_all_memory(init_mem); 4129 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4130 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4131 Node* slow_result = set_results_for_java_call(slow_call); 4132 // this->control() comes from set_results_for_java_call 4133 result_reg->init_req(_slow_path, control()); 4134 result_val->init_req(_slow_path, slow_result); 4135 result_io ->set_req(_slow_path, i_o()); 4136 result_mem ->set_req(_slow_path, reset_memory()); 4137 } 4138 4139 // Return the combined state. 4140 set_i_o( _gvn.transform(result_io) ); 4141 set_all_memory( _gvn.transform(result_mem)); 4142 4143 set_result(result_reg, result_val); 4144 return true; 4145 } 4146 4147 //---------------------------inline_native_getClass---------------------------- 4148 // public final native Class<?> java.lang.Object.getClass(); 4149 // 4150 // Build special case code for calls to getClass on an object. 4151 bool LibraryCallKit::inline_native_getClass() { 4152 Node* obj = null_check_receiver(); 4153 if (stopped()) return true; 4154 set_result(load_mirror_from_klass(load_object_klass(obj))); 4155 return true; 4156 } 4157 4158 //-----------------inline_native_Reflection_getCallerClass--------------------- 4159 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4160 // 4161 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4162 // 4163 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4164 // in that it must skip particular security frames and checks for 4165 // caller sensitive methods. 4166 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4167 #ifndef PRODUCT 4168 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4169 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4170 } 4171 #endif 4172 4173 if (!jvms()->has_method()) { 4174 #ifndef PRODUCT 4175 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4176 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4177 } 4178 #endif 4179 return false; 4180 } 4181 4182 // Walk back up the JVM state to find the caller at the required 4183 // depth. 4184 JVMState* caller_jvms = jvms(); 4185 4186 // Cf. JVM_GetCallerClass 4187 // NOTE: Start the loop at depth 1 because the current JVM state does 4188 // not include the Reflection.getCallerClass() frame. 4189 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4190 ciMethod* m = caller_jvms->method(); 4191 switch (n) { 4192 case 0: 4193 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4194 break; 4195 case 1: 4196 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4197 if (!m->caller_sensitive()) { 4198 #ifndef PRODUCT 4199 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4200 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4201 } 4202 #endif 4203 return false; // bail-out; let JVM_GetCallerClass do the work 4204 } 4205 break; 4206 default: 4207 if (!m->is_ignored_by_security_stack_walk()) { 4208 // We have reached the desired frame; return the holder class. 4209 // Acquire method holder as java.lang.Class and push as constant. 4210 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4211 ciInstance* caller_mirror = caller_klass->java_mirror(); 4212 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4213 4214 #ifndef PRODUCT 4215 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4216 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()); 4217 tty->print_cr(" JVM state at this point:"); 4218 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4219 ciMethod* m = jvms()->of_depth(i)->method(); 4220 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4221 } 4222 } 4223 #endif 4224 return true; 4225 } 4226 break; 4227 } 4228 } 4229 4230 #ifndef PRODUCT 4231 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4232 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4233 tty->print_cr(" JVM state at this point:"); 4234 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4235 ciMethod* m = jvms()->of_depth(i)->method(); 4236 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4237 } 4238 } 4239 #endif 4240 4241 return false; // bail-out; let JVM_GetCallerClass do the work 4242 } 4243 4244 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4245 Node* arg = argument(0); 4246 Node* result; 4247 4248 switch (id) { 4249 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4250 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4251 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4252 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4253 4254 case vmIntrinsics::_doubleToLongBits: { 4255 // two paths (plus control) merge in a wood 4256 RegionNode *r = new RegionNode(3); 4257 Node *phi = new PhiNode(r, TypeLong::LONG); 4258 4259 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4260 // Build the boolean node 4261 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4262 4263 // Branch either way. 4264 // NaN case is less traveled, which makes all the difference. 4265 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4266 Node *opt_isnan = _gvn.transform(ifisnan); 4267 assert( opt_isnan->is_If(), "Expect an IfNode"); 4268 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4269 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4270 4271 set_control(iftrue); 4272 4273 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4274 Node *slow_result = longcon(nan_bits); // return NaN 4275 phi->init_req(1, _gvn.transform( slow_result )); 4276 r->init_req(1, iftrue); 4277 4278 // Else fall through 4279 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4280 set_control(iffalse); 4281 4282 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4283 r->init_req(2, iffalse); 4284 4285 // Post merge 4286 set_control(_gvn.transform(r)); 4287 record_for_igvn(r); 4288 4289 C->set_has_split_ifs(true); // Has chance for split-if optimization 4290 result = phi; 4291 assert(result->bottom_type()->isa_long(), "must be"); 4292 break; 4293 } 4294 4295 case vmIntrinsics::_floatToIntBits: { 4296 // two paths (plus control) merge in a wood 4297 RegionNode *r = new RegionNode(3); 4298 Node *phi = new PhiNode(r, TypeInt::INT); 4299 4300 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4301 // Build the boolean node 4302 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4303 4304 // Branch either way. 4305 // NaN case is less traveled, which makes all the difference. 4306 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4307 Node *opt_isnan = _gvn.transform(ifisnan); 4308 assert( opt_isnan->is_If(), "Expect an IfNode"); 4309 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4310 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4311 4312 set_control(iftrue); 4313 4314 static const jint nan_bits = 0x7fc00000; 4315 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4316 phi->init_req(1, _gvn.transform( slow_result )); 4317 r->init_req(1, iftrue); 4318 4319 // Else fall through 4320 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4321 set_control(iffalse); 4322 4323 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4324 r->init_req(2, iffalse); 4325 4326 // Post merge 4327 set_control(_gvn.transform(r)); 4328 record_for_igvn(r); 4329 4330 C->set_has_split_ifs(true); // Has chance for split-if optimization 4331 result = phi; 4332 assert(result->bottom_type()->isa_int(), "must be"); 4333 break; 4334 } 4335 4336 default: 4337 fatal_unexpected_iid(id); 4338 break; 4339 } 4340 set_result(_gvn.transform(result)); 4341 return true; 4342 } 4343 4344 #ifdef _LP64 4345 #define XTOP ,top() /*additional argument*/ 4346 #else //_LP64 4347 #define XTOP /*no additional argument*/ 4348 #endif //_LP64 4349 4350 //----------------------inline_unsafe_copyMemory------------------------- 4351 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4352 bool LibraryCallKit::inline_unsafe_copyMemory() { 4353 if (callee()->is_static()) return false; // caller must have the capability! 4354 null_check_receiver(); // null-check receiver 4355 if (stopped()) return true; 4356 4357 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4358 4359 Node* src_ptr = argument(1); // type: oop 4360 Node* src_off = ConvL2X(argument(2)); // type: long 4361 Node* dst_ptr = argument(4); // type: oop 4362 Node* dst_off = ConvL2X(argument(5)); // type: long 4363 Node* size = ConvL2X(argument(7)); // type: long 4364 4365 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4366 "fieldOffset must be byte-scaled"); 4367 4368 Node* src = make_unsafe_address(src_ptr, src_off); 4369 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4370 4371 // Conservatively insert a memory barrier on all memory slices. 4372 // Do not let writes of the copy source or destination float below the copy. 4373 insert_mem_bar(Op_MemBarCPUOrder); 4374 4375 // Call it. Note that the length argument is not scaled. 4376 make_runtime_call(RC_LEAF|RC_NO_FP, 4377 OptoRuntime::fast_arraycopy_Type(), 4378 StubRoutines::unsafe_arraycopy(), 4379 "unsafe_arraycopy", 4380 TypeRawPtr::BOTTOM, 4381 src, dst, size XTOP); 4382 4383 // Do not let reads of the copy destination float above the copy. 4384 insert_mem_bar(Op_MemBarCPUOrder); 4385 4386 return true; 4387 } 4388 4389 //------------------------clone_coping----------------------------------- 4390 // Helper function for inline_native_clone. 4391 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4392 assert(obj_size != NULL, ""); 4393 Node* raw_obj = alloc_obj->in(1); 4394 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4395 4396 AllocateNode* alloc = NULL; 4397 if (ReduceBulkZeroing) { 4398 // We will be completely responsible for initializing this object - 4399 // mark Initialize node as complete. 4400 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4401 // The object was just allocated - there should be no any stores! 4402 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4403 // Mark as complete_with_arraycopy so that on AllocateNode 4404 // expansion, we know this AllocateNode is initialized by an array 4405 // copy and a StoreStore barrier exists after the array copy. 4406 alloc->initialization()->set_complete_with_arraycopy(); 4407 } 4408 4409 // Copy the fastest available way. 4410 // TODO: generate fields copies for small objects instead. 4411 Node* src = obj; 4412 Node* dest = alloc_obj; 4413 Node* size = _gvn.transform(obj_size); 4414 4415 // Exclude the header but include array length to copy by 8 bytes words. 4416 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4417 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4418 instanceOopDesc::base_offset_in_bytes(); 4419 // base_off: 4420 // 8 - 32-bit VM 4421 // 12 - 64-bit VM, compressed klass 4422 // 16 - 64-bit VM, normal klass 4423 if (base_off % BytesPerLong != 0) { 4424 assert(UseCompressedClassPointers, ""); 4425 if (is_array) { 4426 // Exclude length to copy by 8 bytes words. 4427 base_off += sizeof(int); 4428 } else { 4429 // Include klass to copy by 8 bytes words. 4430 base_off = instanceOopDesc::klass_offset_in_bytes(); 4431 } 4432 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4433 } 4434 src = basic_plus_adr(src, base_off); 4435 dest = basic_plus_adr(dest, base_off); 4436 4437 // Compute the length also, if needed: 4438 Node* countx = size; 4439 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4440 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4441 4442 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4443 4444 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4445 ac->set_clonebasic(); 4446 Node* n = _gvn.transform(ac); 4447 if (n == ac) { 4448 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4449 } else { 4450 set_all_memory(n); 4451 } 4452 4453 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4454 if (card_mark) { 4455 assert(!is_array, ""); 4456 // Put in store barrier for any and all oops we are sticking 4457 // into this object. (We could avoid this if we could prove 4458 // that the object type contains no oop fields at all.) 4459 Node* no_particular_value = NULL; 4460 Node* no_particular_field = NULL; 4461 int raw_adr_idx = Compile::AliasIdxRaw; 4462 post_barrier(control(), 4463 memory(raw_adr_type), 4464 alloc_obj, 4465 no_particular_field, 4466 raw_adr_idx, 4467 no_particular_value, 4468 T_OBJECT, 4469 false); 4470 } 4471 4472 // Do not let reads from the cloned object float above the arraycopy. 4473 if (alloc != NULL) { 4474 // Do not let stores that initialize this object be reordered with 4475 // a subsequent store that would make this object accessible by 4476 // other threads. 4477 // Record what AllocateNode this StoreStore protects so that 4478 // escape analysis can go from the MemBarStoreStoreNode to the 4479 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4480 // based on the escape status of the AllocateNode. 4481 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4482 } else { 4483 insert_mem_bar(Op_MemBarCPUOrder); 4484 } 4485 } 4486 4487 //------------------------inline_native_clone---------------------------- 4488 // protected native Object java.lang.Object.clone(); 4489 // 4490 // Here are the simple edge cases: 4491 // null receiver => normal trap 4492 // virtual and clone was overridden => slow path to out-of-line clone 4493 // not cloneable or finalizer => slow path to out-of-line Object.clone 4494 // 4495 // The general case has two steps, allocation and copying. 4496 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4497 // 4498 // Copying also has two cases, oop arrays and everything else. 4499 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4500 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4501 // 4502 // These steps fold up nicely if and when the cloned object's klass 4503 // can be sharply typed as an object array, a type array, or an instance. 4504 // 4505 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4506 PhiNode* result_val; 4507 4508 // Set the reexecute bit for the interpreter to reexecute 4509 // the bytecode that invokes Object.clone if deoptimization happens. 4510 { PreserveReexecuteState preexecs(this); 4511 jvms()->set_should_reexecute(true); 4512 4513 Node* obj = null_check_receiver(); 4514 if (stopped()) return true; 4515 4516 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4517 4518 // If we are going to clone an instance, we need its exact type to 4519 // know the number and types of fields to convert the clone to 4520 // loads/stores. Maybe a speculative type can help us. 4521 if (!obj_type->klass_is_exact() && 4522 obj_type->speculative_type() != NULL && 4523 obj_type->speculative_type()->is_instance_klass()) { 4524 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4525 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4526 !spec_ik->has_injected_fields()) { 4527 ciKlass* k = obj_type->klass(); 4528 if (!k->is_instance_klass() || 4529 k->as_instance_klass()->is_interface() || 4530 k->as_instance_klass()->has_subklass()) { 4531 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4532 } 4533 } 4534 } 4535 4536 Node* obj_klass = load_object_klass(obj); 4537 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4538 const TypeOopPtr* toop = ((tklass != NULL) 4539 ? tklass->as_instance_type() 4540 : TypeInstPtr::NOTNULL); 4541 4542 // Conservatively insert a memory barrier on all memory slices. 4543 // Do not let writes into the original float below the clone. 4544 insert_mem_bar(Op_MemBarCPUOrder); 4545 4546 // paths into result_reg: 4547 enum { 4548 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4549 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4550 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4551 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4552 PATH_LIMIT 4553 }; 4554 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4555 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4556 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4557 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4558 record_for_igvn(result_reg); 4559 4560 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4561 int raw_adr_idx = Compile::AliasIdxRaw; 4562 4563 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4564 if (array_ctl != NULL) { 4565 // It's an array. 4566 PreserveJVMState pjvms(this); 4567 set_control(array_ctl); 4568 Node* obj_length = load_array_length(obj); 4569 Node* obj_size = NULL; 4570 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4571 4572 if (!use_ReduceInitialCardMarks()) { 4573 // If it is an oop array, it requires very special treatment, 4574 // because card marking is required on each card of the array. 4575 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4576 if (is_obja != NULL) { 4577 PreserveJVMState pjvms2(this); 4578 set_control(is_obja); 4579 // Generate a direct call to the right arraycopy function(s). 4580 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4581 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4582 ac->set_cloneoop(); 4583 Node* n = _gvn.transform(ac); 4584 assert(n == ac, "cannot disappear"); 4585 ac->connect_outputs(this); 4586 4587 result_reg->init_req(_objArray_path, control()); 4588 result_val->init_req(_objArray_path, alloc_obj); 4589 result_i_o ->set_req(_objArray_path, i_o()); 4590 result_mem ->set_req(_objArray_path, reset_memory()); 4591 } 4592 } 4593 // Otherwise, there are no card marks to worry about. 4594 // (We can dispense with card marks if we know the allocation 4595 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4596 // causes the non-eden paths to take compensating steps to 4597 // simulate a fresh allocation, so that no further 4598 // card marks are required in compiled code to initialize 4599 // the object.) 4600 4601 if (!stopped()) { 4602 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4603 4604 // Present the results of the copy. 4605 result_reg->init_req(_array_path, control()); 4606 result_val->init_req(_array_path, alloc_obj); 4607 result_i_o ->set_req(_array_path, i_o()); 4608 result_mem ->set_req(_array_path, reset_memory()); 4609 } 4610 } 4611 4612 // We only go to the instance fast case code if we pass a number of guards. 4613 // The paths which do not pass are accumulated in the slow_region. 4614 RegionNode* slow_region = new RegionNode(1); 4615 record_for_igvn(slow_region); 4616 if (!stopped()) { 4617 // It's an instance (we did array above). Make the slow-path tests. 4618 // If this is a virtual call, we generate a funny guard. We grab 4619 // the vtable entry corresponding to clone() from the target object. 4620 // If the target method which we are calling happens to be the 4621 // Object clone() method, we pass the guard. We do not need this 4622 // guard for non-virtual calls; the caller is known to be the native 4623 // Object clone(). 4624 if (is_virtual) { 4625 generate_virtual_guard(obj_klass, slow_region); 4626 } 4627 4628 // The object must be cloneable and must not have a finalizer. 4629 // Both of these conditions may be checked in a single test. 4630 // We could optimize the cloneable test further, but we don't care. 4631 generate_access_flags_guard(obj_klass, 4632 // Test both conditions: 4633 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4634 // Must be cloneable but not finalizer: 4635 JVM_ACC_IS_CLONEABLE, 4636 slow_region); 4637 } 4638 4639 if (!stopped()) { 4640 // It's an instance, and it passed the slow-path tests. 4641 PreserveJVMState pjvms(this); 4642 Node* obj_size = NULL; 4643 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4644 // is reexecuted if deoptimization occurs and there could be problems when merging 4645 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4646 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4647 4648 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4649 4650 // Present the results of the slow call. 4651 result_reg->init_req(_instance_path, control()); 4652 result_val->init_req(_instance_path, alloc_obj); 4653 result_i_o ->set_req(_instance_path, i_o()); 4654 result_mem ->set_req(_instance_path, reset_memory()); 4655 } 4656 4657 // Generate code for the slow case. We make a call to clone(). 4658 set_control(_gvn.transform(slow_region)); 4659 if (!stopped()) { 4660 PreserveJVMState pjvms(this); 4661 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4662 Node* slow_result = set_results_for_java_call(slow_call); 4663 // this->control() comes from set_results_for_java_call 4664 result_reg->init_req(_slow_path, control()); 4665 result_val->init_req(_slow_path, slow_result); 4666 result_i_o ->set_req(_slow_path, i_o()); 4667 result_mem ->set_req(_slow_path, reset_memory()); 4668 } 4669 4670 // Return the combined state. 4671 set_control( _gvn.transform(result_reg)); 4672 set_i_o( _gvn.transform(result_i_o)); 4673 set_all_memory( _gvn.transform(result_mem)); 4674 } // original reexecute is set back here 4675 4676 set_result(_gvn.transform(result_val)); 4677 return true; 4678 } 4679 4680 // If we have a tighly coupled allocation, the arraycopy may take care 4681 // of the array initialization. If one of the guards we insert between 4682 // the allocation and the arraycopy causes a deoptimization, an 4683 // unitialized array will escape the compiled method. To prevent that 4684 // we set the JVM state for uncommon traps between the allocation and 4685 // the arraycopy to the state before the allocation so, in case of 4686 // deoptimization, we'll reexecute the allocation and the 4687 // initialization. 4688 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4689 if (alloc != NULL) { 4690 ciMethod* trap_method = alloc->jvms()->method(); 4691 int trap_bci = alloc->jvms()->bci(); 4692 4693 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4694 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4695 // Make sure there's no store between the allocation and the 4696 // arraycopy otherwise visible side effects could be rexecuted 4697 // in case of deoptimization and cause incorrect execution. 4698 bool no_interfering_store = true; 4699 Node* mem = alloc->in(TypeFunc::Memory); 4700 if (mem->is_MergeMem()) { 4701 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4702 Node* n = mms.memory(); 4703 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4704 assert(n->is_Store(), "what else?"); 4705 no_interfering_store = false; 4706 break; 4707 } 4708 } 4709 } else { 4710 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4711 Node* n = mms.memory(); 4712 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4713 assert(n->is_Store(), "what else?"); 4714 no_interfering_store = false; 4715 break; 4716 } 4717 } 4718 } 4719 4720 if (no_interfering_store) { 4721 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4722 uint size = alloc->req(); 4723 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4724 old_jvms->set_map(sfpt); 4725 for (uint i = 0; i < size; i++) { 4726 sfpt->init_req(i, alloc->in(i)); 4727 } 4728 // re-push array length for deoptimization 4729 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4730 old_jvms->set_sp(old_jvms->sp()+1); 4731 old_jvms->set_monoff(old_jvms->monoff()+1); 4732 old_jvms->set_scloff(old_jvms->scloff()+1); 4733 old_jvms->set_endoff(old_jvms->endoff()+1); 4734 old_jvms->set_should_reexecute(true); 4735 4736 sfpt->set_i_o(map()->i_o()); 4737 sfpt->set_memory(map()->memory()); 4738 sfpt->set_control(map()->control()); 4739 4740 JVMState* saved_jvms = jvms(); 4741 saved_reexecute_sp = _reexecute_sp; 4742 4743 set_jvms(sfpt->jvms()); 4744 _reexecute_sp = jvms()->sp(); 4745 4746 return saved_jvms; 4747 } 4748 } 4749 } 4750 return NULL; 4751 } 4752 4753 // In case of a deoptimization, we restart execution at the 4754 // allocation, allocating a new array. We would leave an uninitialized 4755 // array in the heap that GCs wouldn't expect. Move the allocation 4756 // after the traps so we don't allocate the array if we 4757 // deoptimize. This is possible because tightly_coupled_allocation() 4758 // guarantees there's no observer of the allocated array at this point 4759 // and the control flow is simple enough. 4760 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4761 if (saved_jvms != NULL) { 4762 // restore JVM state to the state at the arraycopy 4763 saved_jvms->map()->set_control(map()->control()); 4764 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4765 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4766 // If we've improved the types of some nodes (null check) while 4767 // emitting the guards, propagate them to the current state 4768 map()->replaced_nodes().apply(saved_jvms->map()); 4769 set_jvms(saved_jvms); 4770 _reexecute_sp = saved_reexecute_sp; 4771 4772 // Remove the allocation from above the guards 4773 CallProjections callprojs; 4774 alloc->extract_projections(&callprojs, true); 4775 InitializeNode* init = alloc->initialization(); 4776 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4777 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4778 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4779 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4780 4781 // move the allocation here (after the guards) 4782 _gvn.hash_delete(alloc); 4783 alloc->set_req(TypeFunc::Control, control()); 4784 alloc->set_req(TypeFunc::I_O, i_o()); 4785 Node *mem = reset_memory(); 4786 set_all_memory(mem); 4787 alloc->set_req(TypeFunc::Memory, mem); 4788 set_control(init->proj_out(TypeFunc::Control)); 4789 set_i_o(callprojs.fallthrough_ioproj); 4790 4791 // Update memory as done in GraphKit::set_output_for_allocation() 4792 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4793 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4794 if (ary_type->isa_aryptr() && length_type != NULL) { 4795 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4796 } 4797 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4798 int elemidx = C->get_alias_index(telemref); 4799 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4800 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4801 4802 Node* allocx = _gvn.transform(alloc); 4803 assert(allocx == alloc, "where has the allocation gone?"); 4804 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4805 4806 _gvn.hash_delete(dest); 4807 dest->set_req(0, control()); 4808 Node* destx = _gvn.transform(dest); 4809 assert(destx == dest, "where has the allocation result gone?"); 4810 } 4811 } 4812 4813 4814 //------------------------------inline_arraycopy----------------------- 4815 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4816 // Object dest, int destPos, 4817 // int length); 4818 bool LibraryCallKit::inline_arraycopy() { 4819 // Get the arguments. 4820 Node* src = argument(0); // type: oop 4821 Node* src_offset = argument(1); // type: int 4822 Node* dest = argument(2); // type: oop 4823 Node* dest_offset = argument(3); // type: int 4824 Node* length = argument(4); // type: int 4825 4826 4827 // Check for allocation before we add nodes that would confuse 4828 // tightly_coupled_allocation() 4829 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4830 4831 int saved_reexecute_sp = -1; 4832 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4833 4834 // The following tests must be performed 4835 // (1) src and dest are arrays. 4836 // (2) src and dest arrays must have elements of the same BasicType 4837 // (3) src and dest must not be null. 4838 // (4) src_offset must not be negative. 4839 // (5) dest_offset must not be negative. 4840 // (6) length must not be negative. 4841 // (7) src_offset + length must not exceed length of src. 4842 // (8) dest_offset + length must not exceed length of dest. 4843 // (9) each element of an oop array must be assignable 4844 4845 // (3) src and dest must not be null. 4846 // always do this here because we need the JVM state for uncommon traps 4847 Node* null_ctl = top(); 4848 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4849 assert(null_ctl->is_top(), "no null control here"); 4850 dest = null_check(dest, T_ARRAY); 4851 4852 if (saved_jvms == NULL && alloc != NULL) { 4853 // We're not emitting the guards, see if we have a tightly 4854 // allocation now that we've done the null check 4855 alloc = tightly_coupled_allocation(dest, NULL); 4856 } 4857 4858 bool validated = false; 4859 4860 const Type* src_type = _gvn.type(src); 4861 const Type* dest_type = _gvn.type(dest); 4862 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4863 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4864 4865 // Do we have the type of src? 4866 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4867 // Do we have the type of dest? 4868 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4869 // Is the type for src from speculation? 4870 bool src_spec = false; 4871 // Is the type for dest from speculation? 4872 bool dest_spec = false; 4873 4874 if ((!has_src || !has_dest) && (alloc == NULL || saved_jvms != NULL)) { 4875 // We don't have sufficient type information, let's see if 4876 // speculative types can help. We need to have types for both src 4877 // and dest so that it pays off. 4878 4879 // Do we already have or could we have type information for src 4880 bool could_have_src = has_src; 4881 // Do we already have or could we have type information for dest 4882 bool could_have_dest = has_dest; 4883 4884 ciKlass* src_k = NULL; 4885 if (!has_src) { 4886 src_k = src_type->speculative_type_not_null(); 4887 if (src_k != NULL && src_k->is_array_klass()) { 4888 could_have_src = true; 4889 } 4890 } 4891 4892 ciKlass* dest_k = NULL; 4893 if (!has_dest) { 4894 dest_k = dest_type->speculative_type_not_null(); 4895 if (dest_k != NULL && dest_k->is_array_klass()) { 4896 could_have_dest = true; 4897 } 4898 } 4899 4900 if (could_have_src && could_have_dest) { 4901 // This is going to pay off so emit the required guards 4902 if (!has_src) { 4903 src = maybe_cast_profiled_obj(src, src_k, true); 4904 src_type = _gvn.type(src); 4905 top_src = src_type->isa_aryptr(); 4906 has_src = (top_src != NULL && top_src->klass() != NULL); 4907 src_spec = true; 4908 } 4909 if (!has_dest) { 4910 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4911 dest_type = _gvn.type(dest); 4912 top_dest = dest_type->isa_aryptr(); 4913 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4914 dest_spec = true; 4915 } 4916 } 4917 } 4918 4919 if (has_src && has_dest && (alloc == NULL || saved_jvms != NULL)) { 4920 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4921 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4922 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4923 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4924 4925 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4926 // If both arrays are object arrays then having the exact types 4927 // for both will remove the need for a subtype check at runtime 4928 // before the call and may make it possible to pick a faster copy 4929 // routine (without a subtype check on every element) 4930 // Do we have the exact type of src? 4931 bool could_have_src = src_spec; 4932 // Do we have the exact type of dest? 4933 bool could_have_dest = dest_spec; 4934 ciKlass* src_k = top_src->klass(); 4935 ciKlass* dest_k = top_dest->klass(); 4936 if (!src_spec) { 4937 src_k = src_type->speculative_type_not_null(); 4938 if (src_k != NULL && src_k->is_array_klass()) { 4939 could_have_src = true; 4940 } 4941 } 4942 if (!dest_spec) { 4943 dest_k = dest_type->speculative_type_not_null(); 4944 if (dest_k != NULL && dest_k->is_array_klass()) { 4945 could_have_dest = true; 4946 } 4947 } 4948 if (could_have_src && could_have_dest) { 4949 // If we can have both exact types, emit the missing guards 4950 if (could_have_src && !src_spec) { 4951 src = maybe_cast_profiled_obj(src, src_k, true); 4952 } 4953 if (could_have_dest && !dest_spec) { 4954 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4955 } 4956 } 4957 } 4958 } 4959 4960 ciMethod* trap_method = method(); 4961 int trap_bci = bci(); 4962 if (saved_jvms != NULL) { 4963 trap_method = alloc->jvms()->method(); 4964 trap_bci = alloc->jvms()->bci(); 4965 } 4966 4967 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 4968 (alloc == NULL || saved_jvms != NULL) && 4969 !src->is_top() && !dest->is_top()) { 4970 // validate arguments: enables transformation the ArrayCopyNode 4971 validated = true; 4972 4973 RegionNode* slow_region = new RegionNode(1); 4974 record_for_igvn(slow_region); 4975 4976 // (1) src and dest are arrays. 4977 generate_non_array_guard(load_object_klass(src), slow_region); 4978 generate_non_array_guard(load_object_klass(dest), slow_region); 4979 4980 // (2) src and dest arrays must have elements of the same BasicType 4981 // done at macro expansion or at Ideal transformation time 4982 4983 // (4) src_offset must not be negative. 4984 generate_negative_guard(src_offset, slow_region); 4985 4986 // (5) dest_offset must not be negative. 4987 generate_negative_guard(dest_offset, slow_region); 4988 4989 // (7) src_offset + length must not exceed length of src. 4990 generate_limit_guard(src_offset, length, 4991 load_array_length(src), 4992 slow_region); 4993 4994 // (8) dest_offset + length must not exceed length of dest. 4995 generate_limit_guard(dest_offset, length, 4996 load_array_length(dest), 4997 slow_region); 4998 4999 // (9) each element of an oop array must be assignable 5000 Node* src_klass = load_object_klass(src); 5001 Node* dest_klass = load_object_klass(dest); 5002 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5003 5004 if (not_subtype_ctrl != top()) { 5005 PreserveJVMState pjvms(this); 5006 set_control(not_subtype_ctrl); 5007 uncommon_trap(Deoptimization::Reason_intrinsic, 5008 Deoptimization::Action_make_not_entrant); 5009 assert(stopped(), "Should be stopped"); 5010 } 5011 { 5012 PreserveJVMState pjvms(this); 5013 set_control(_gvn.transform(slow_region)); 5014 uncommon_trap(Deoptimization::Reason_intrinsic, 5015 Deoptimization::Action_make_not_entrant); 5016 assert(stopped(), "Should be stopped"); 5017 } 5018 } 5019 5020 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5021 5022 if (stopped()) { 5023 return true; 5024 } 5025 5026 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5027 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5028 // so the compiler has a chance to eliminate them: during macro expansion, 5029 // we have to set their control (CastPP nodes are eliminated). 5030 load_object_klass(src), load_object_klass(dest), 5031 load_array_length(src), load_array_length(dest)); 5032 5033 ac->set_arraycopy(validated); 5034 5035 Node* n = _gvn.transform(ac); 5036 if (n == ac) { 5037 ac->connect_outputs(this); 5038 } else { 5039 assert(validated, "shouldn't transform if all arguments not validated"); 5040 set_all_memory(n); 5041 } 5042 5043 return true; 5044 } 5045 5046 5047 // Helper function which determines if an arraycopy immediately follows 5048 // an allocation, with no intervening tests or other escapes for the object. 5049 AllocateArrayNode* 5050 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5051 RegionNode* slow_region) { 5052 if (stopped()) return NULL; // no fast path 5053 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5054 5055 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5056 if (alloc == NULL) return NULL; 5057 5058 Node* rawmem = memory(Compile::AliasIdxRaw); 5059 // Is the allocation's memory state untouched? 5060 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5061 // Bail out if there have been raw-memory effects since the allocation. 5062 // (Example: There might have been a call or safepoint.) 5063 return NULL; 5064 } 5065 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5066 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5067 return NULL; 5068 } 5069 5070 // There must be no unexpected observers of this allocation. 5071 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5072 Node* obs = ptr->fast_out(i); 5073 if (obs != this->map()) { 5074 return NULL; 5075 } 5076 } 5077 5078 // This arraycopy must unconditionally follow the allocation of the ptr. 5079 Node* alloc_ctl = ptr->in(0); 5080 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5081 5082 Node* ctl = control(); 5083 while (ctl != alloc_ctl) { 5084 // There may be guards which feed into the slow_region. 5085 // Any other control flow means that we might not get a chance 5086 // to finish initializing the allocated object. 5087 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5088 IfNode* iff = ctl->in(0)->as_If(); 5089 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5090 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5091 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5092 ctl = iff->in(0); // This test feeds the known slow_region. 5093 continue; 5094 } 5095 // One more try: Various low-level checks bottom out in 5096 // uncommon traps. If the debug-info of the trap omits 5097 // any reference to the allocation, as we've already 5098 // observed, then there can be no objection to the trap. 5099 bool found_trap = false; 5100 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5101 Node* obs = not_ctl->fast_out(j); 5102 if (obs->in(0) == not_ctl && obs->is_Call() && 5103 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5104 found_trap = true; break; 5105 } 5106 } 5107 if (found_trap) { 5108 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5109 continue; 5110 } 5111 } 5112 return NULL; 5113 } 5114 5115 // If we get this far, we have an allocation which immediately 5116 // precedes the arraycopy, and we can take over zeroing the new object. 5117 // The arraycopy will finish the initialization, and provide 5118 // a new control state to which we will anchor the destination pointer. 5119 5120 return alloc; 5121 } 5122 5123 //-------------inline_encodeISOArray----------------------------------- 5124 // encode char[] to byte[] in ISO_8859_1 5125 bool LibraryCallKit::inline_encodeISOArray() { 5126 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5127 // no receiver since it is static method 5128 Node *src = argument(0); 5129 Node *src_offset = argument(1); 5130 Node *dst = argument(2); 5131 Node *dst_offset = argument(3); 5132 Node *length = argument(4); 5133 5134 const Type* src_type = src->Value(&_gvn); 5135 const Type* dst_type = dst->Value(&_gvn); 5136 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5137 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5138 if (top_src == NULL || top_src->klass() == NULL || 5139 top_dest == NULL || top_dest->klass() == NULL) { 5140 // failed array check 5141 return false; 5142 } 5143 5144 // Figure out the size and type of the elements we will be copying. 5145 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5146 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5147 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5148 return false; 5149 } 5150 Node* src_start = array_element_address(src, src_offset, src_elem); 5151 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5152 // 'src_start' points to src array + scaled offset 5153 // 'dst_start' points to dst array + scaled offset 5154 5155 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5156 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5157 enc = _gvn.transform(enc); 5158 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5159 set_memory(res_mem, mtype); 5160 set_result(enc); 5161 return true; 5162 } 5163 5164 //-------------inline_multiplyToLen----------------------------------- 5165 bool LibraryCallKit::inline_multiplyToLen() { 5166 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5167 5168 address stubAddr = StubRoutines::multiplyToLen(); 5169 if (stubAddr == NULL) { 5170 return false; // Intrinsic's stub is not implemented on this platform 5171 } 5172 const char* stubName = "multiplyToLen"; 5173 5174 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5175 5176 Node* x = argument(1); 5177 Node* xlen = argument(2); 5178 Node* y = argument(3); 5179 Node* ylen = argument(4); 5180 Node* z = argument(5); 5181 5182 const Type* x_type = x->Value(&_gvn); 5183 const Type* y_type = y->Value(&_gvn); 5184 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5185 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5186 if (top_x == NULL || top_x->klass() == NULL || 5187 top_y == NULL || top_y->klass() == NULL) { 5188 // failed array check 5189 return false; 5190 } 5191 5192 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5193 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5194 if (x_elem != T_INT || y_elem != T_INT) { 5195 return false; 5196 } 5197 5198 // Set the original stack and the reexecute bit for the interpreter to reexecute 5199 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5200 // on the return from z array allocation in runtime. 5201 { PreserveReexecuteState preexecs(this); 5202 jvms()->set_should_reexecute(true); 5203 5204 Node* x_start = array_element_address(x, intcon(0), x_elem); 5205 Node* y_start = array_element_address(y, intcon(0), y_elem); 5206 // 'x_start' points to x array + scaled xlen 5207 // 'y_start' points to y array + scaled ylen 5208 5209 // Allocate the result array 5210 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5211 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5212 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5213 5214 IdealKit ideal(this); 5215 5216 #define __ ideal. 5217 Node* one = __ ConI(1); 5218 Node* zero = __ ConI(0); 5219 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5220 __ set(need_alloc, zero); 5221 __ set(z_alloc, z); 5222 __ if_then(z, BoolTest::eq, null()); { 5223 __ increment (need_alloc, one); 5224 } __ else_(); { 5225 // Update graphKit memory and control from IdealKit. 5226 sync_kit(ideal); 5227 Node* zlen_arg = load_array_length(z); 5228 // Update IdealKit memory and control from graphKit. 5229 __ sync_kit(this); 5230 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5231 __ increment (need_alloc, one); 5232 } __ end_if(); 5233 } __ end_if(); 5234 5235 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5236 // Update graphKit memory and control from IdealKit. 5237 sync_kit(ideal); 5238 Node * narr = new_array(klass_node, zlen, 1); 5239 // Update IdealKit memory and control from graphKit. 5240 __ sync_kit(this); 5241 __ set(z_alloc, narr); 5242 } __ end_if(); 5243 5244 sync_kit(ideal); 5245 z = __ value(z_alloc); 5246 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5247 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5248 // Final sync IdealKit and GraphKit. 5249 final_sync(ideal); 5250 #undef __ 5251 5252 Node* z_start = array_element_address(z, intcon(0), T_INT); 5253 5254 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5255 OptoRuntime::multiplyToLen_Type(), 5256 stubAddr, stubName, TypePtr::BOTTOM, 5257 x_start, xlen, y_start, ylen, z_start, zlen); 5258 } // original reexecute is set back here 5259 5260 C->set_has_split_ifs(true); // Has chance for split-if optimization 5261 set_result(z); 5262 return true; 5263 } 5264 5265 5266 /** 5267 * Calculate CRC32 for byte. 5268 * int java.util.zip.CRC32.update(int crc, int b) 5269 */ 5270 bool LibraryCallKit::inline_updateCRC32() { 5271 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5272 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5273 // no receiver since it is static method 5274 Node* crc = argument(0); // type: int 5275 Node* b = argument(1); // type: int 5276 5277 /* 5278 * int c = ~ crc; 5279 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5280 * b = b ^ (c >>> 8); 5281 * crc = ~b; 5282 */ 5283 5284 Node* M1 = intcon(-1); 5285 crc = _gvn.transform(new XorINode(crc, M1)); 5286 Node* result = _gvn.transform(new XorINode(crc, b)); 5287 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5288 5289 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5290 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5291 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5292 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5293 5294 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5295 result = _gvn.transform(new XorINode(crc, result)); 5296 result = _gvn.transform(new XorINode(result, M1)); 5297 set_result(result); 5298 return true; 5299 } 5300 5301 /** 5302 * Calculate CRC32 for byte[] array. 5303 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5304 */ 5305 bool LibraryCallKit::inline_updateBytesCRC32() { 5306 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5307 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5308 // no receiver since it is static method 5309 Node* crc = argument(0); // type: int 5310 Node* src = argument(1); // type: oop 5311 Node* offset = argument(2); // type: int 5312 Node* length = argument(3); // type: int 5313 5314 const Type* src_type = src->Value(&_gvn); 5315 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5316 if (top_src == NULL || top_src->klass() == NULL) { 5317 // failed array check 5318 return false; 5319 } 5320 5321 // Figure out the size and type of the elements we will be copying. 5322 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5323 if (src_elem != T_BYTE) { 5324 return false; 5325 } 5326 5327 // 'src_start' points to src array + scaled offset 5328 Node* src_start = array_element_address(src, offset, src_elem); 5329 5330 // We assume that range check is done by caller. 5331 // TODO: generate range check (offset+length < src.length) in debug VM. 5332 5333 // Call the stub. 5334 address stubAddr = StubRoutines::updateBytesCRC32(); 5335 const char *stubName = "updateBytesCRC32"; 5336 5337 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5338 stubAddr, stubName, TypePtr::BOTTOM, 5339 crc, src_start, length); 5340 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5341 set_result(result); 5342 return true; 5343 } 5344 5345 /** 5346 * Calculate CRC32 for ByteBuffer. 5347 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5348 */ 5349 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5350 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5351 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5352 // no receiver since it is static method 5353 Node* crc = argument(0); // type: int 5354 Node* src = argument(1); // type: long 5355 Node* offset = argument(3); // type: int 5356 Node* length = argument(4); // type: int 5357 5358 src = ConvL2X(src); // adjust Java long to machine word 5359 Node* base = _gvn.transform(new CastX2PNode(src)); 5360 offset = ConvI2X(offset); 5361 5362 // 'src_start' points to src array + scaled offset 5363 Node* src_start = basic_plus_adr(top(), base, offset); 5364 5365 // Call the stub. 5366 address stubAddr = StubRoutines::updateBytesCRC32(); 5367 const char *stubName = "updateBytesCRC32"; 5368 5369 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5370 stubAddr, stubName, TypePtr::BOTTOM, 5371 crc, src_start, length); 5372 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5373 set_result(result); 5374 return true; 5375 } 5376 5377 //----------------------------inline_reference_get---------------------------- 5378 // public T java.lang.ref.Reference.get(); 5379 bool LibraryCallKit::inline_reference_get() { 5380 const int referent_offset = java_lang_ref_Reference::referent_offset; 5381 guarantee(referent_offset > 0, "should have already been set"); 5382 5383 // Get the argument: 5384 Node* reference_obj = null_check_receiver(); 5385 if (stopped()) return true; 5386 5387 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5388 5389 ciInstanceKlass* klass = env()->Object_klass(); 5390 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5391 5392 Node* no_ctrl = NULL; 5393 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5394 5395 // Use the pre-barrier to record the value in the referent field 5396 pre_barrier(false /* do_load */, 5397 control(), 5398 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5399 result /* pre_val */, 5400 T_OBJECT); 5401 5402 // Add memory barrier to prevent commoning reads from this field 5403 // across safepoint since GC can change its value. 5404 insert_mem_bar(Op_MemBarCPUOrder); 5405 5406 set_result(result); 5407 return true; 5408 } 5409 5410 5411 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5412 bool is_exact=true, bool is_static=false) { 5413 5414 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5415 assert(tinst != NULL, "obj is null"); 5416 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5417 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5418 5419 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5420 ciSymbol::make(fieldTypeString), 5421 is_static); 5422 if (field == NULL) return (Node *) NULL; 5423 assert (field != NULL, "undefined field"); 5424 5425 // Next code copied from Parse::do_get_xxx(): 5426 5427 // Compute address and memory type. 5428 int offset = field->offset_in_bytes(); 5429 bool is_vol = field->is_volatile(); 5430 ciType* field_klass = field->type(); 5431 assert(field_klass->is_loaded(), "should be loaded"); 5432 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5433 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5434 BasicType bt = field->layout_type(); 5435 5436 // Build the resultant type of the load 5437 const Type *type; 5438 if (bt == T_OBJECT) { 5439 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5440 } else { 5441 type = Type::get_const_basic_type(bt); 5442 } 5443 5444 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5445 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5446 } 5447 // Build the load. 5448 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5449 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol); 5450 // If reference is volatile, prevent following memory ops from 5451 // floating up past the volatile read. Also prevents commoning 5452 // another volatile read. 5453 if (is_vol) { 5454 // Memory barrier includes bogus read of value to force load BEFORE membar 5455 insert_mem_bar(Op_MemBarAcquire, loadedField); 5456 } 5457 return loadedField; 5458 } 5459 5460 5461 //------------------------------inline_aescrypt_Block----------------------- 5462 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5463 address stubAddr; 5464 const char *stubName; 5465 assert(UseAES, "need AES instruction support"); 5466 5467 switch(id) { 5468 case vmIntrinsics::_aescrypt_encryptBlock: 5469 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5470 stubName = "aescrypt_encryptBlock"; 5471 break; 5472 case vmIntrinsics::_aescrypt_decryptBlock: 5473 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5474 stubName = "aescrypt_decryptBlock"; 5475 break; 5476 } 5477 if (stubAddr == NULL) return false; 5478 5479 Node* aescrypt_object = argument(0); 5480 Node* src = argument(1); 5481 Node* src_offset = argument(2); 5482 Node* dest = argument(3); 5483 Node* dest_offset = argument(4); 5484 5485 // (1) src and dest are arrays. 5486 const Type* src_type = src->Value(&_gvn); 5487 const Type* dest_type = dest->Value(&_gvn); 5488 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5489 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5490 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5491 5492 // for the quick and dirty code we will skip all the checks. 5493 // we are just trying to get the call to be generated. 5494 Node* src_start = src; 5495 Node* dest_start = dest; 5496 if (src_offset != NULL || dest_offset != NULL) { 5497 assert(src_offset != NULL && dest_offset != NULL, ""); 5498 src_start = array_element_address(src, src_offset, T_BYTE); 5499 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5500 } 5501 5502 // now need to get the start of its expanded key array 5503 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5504 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5505 if (k_start == NULL) return false; 5506 5507 if (Matcher::pass_original_key_for_aes()) { 5508 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5509 // compatibility issues between Java key expansion and SPARC crypto instructions 5510 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5511 if (original_k_start == NULL) return false; 5512 5513 // Call the stub. 5514 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5515 stubAddr, stubName, TypePtr::BOTTOM, 5516 src_start, dest_start, k_start, original_k_start); 5517 } else { 5518 // Call the stub. 5519 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5520 stubAddr, stubName, TypePtr::BOTTOM, 5521 src_start, dest_start, k_start); 5522 } 5523 5524 return true; 5525 } 5526 5527 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5528 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5529 address stubAddr; 5530 const char *stubName; 5531 5532 assert(UseAES, "need AES instruction support"); 5533 5534 switch(id) { 5535 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5536 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5537 stubName = "cipherBlockChaining_encryptAESCrypt"; 5538 break; 5539 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5540 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5541 stubName = "cipherBlockChaining_decryptAESCrypt"; 5542 break; 5543 } 5544 if (stubAddr == NULL) return false; 5545 5546 Node* cipherBlockChaining_object = argument(0); 5547 Node* src = argument(1); 5548 Node* src_offset = argument(2); 5549 Node* len = argument(3); 5550 Node* dest = argument(4); 5551 Node* dest_offset = argument(5); 5552 5553 // (1) src and dest are arrays. 5554 const Type* src_type = src->Value(&_gvn); 5555 const Type* dest_type = dest->Value(&_gvn); 5556 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5557 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5558 assert (top_src != NULL && top_src->klass() != NULL 5559 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5560 5561 // checks are the responsibility of the caller 5562 Node* src_start = src; 5563 Node* dest_start = dest; 5564 if (src_offset != NULL || dest_offset != NULL) { 5565 assert(src_offset != NULL && dest_offset != NULL, ""); 5566 src_start = array_element_address(src, src_offset, T_BYTE); 5567 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5568 } 5569 5570 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5571 // (because of the predicated logic executed earlier). 5572 // so we cast it here safely. 5573 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5574 5575 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5576 if (embeddedCipherObj == NULL) return false; 5577 5578 // cast it to what we know it will be at runtime 5579 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5580 assert(tinst != NULL, "CBC obj is null"); 5581 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5582 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5583 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5584 5585 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5586 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5587 const TypeOopPtr* xtype = aklass->as_instance_type(); 5588 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5589 aescrypt_object = _gvn.transform(aescrypt_object); 5590 5591 // we need to get the start of the aescrypt_object's expanded key array 5592 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5593 if (k_start == NULL) return false; 5594 5595 // similarly, get the start address of the r vector 5596 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5597 if (objRvec == NULL) return false; 5598 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5599 5600 Node* cbcCrypt; 5601 if (Matcher::pass_original_key_for_aes()) { 5602 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5603 // compatibility issues between Java key expansion and SPARC crypto instructions 5604 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5605 if (original_k_start == NULL) return false; 5606 5607 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5608 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5609 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5610 stubAddr, stubName, TypePtr::BOTTOM, 5611 src_start, dest_start, k_start, r_start, len, original_k_start); 5612 } else { 5613 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5614 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5615 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5616 stubAddr, stubName, TypePtr::BOTTOM, 5617 src_start, dest_start, k_start, r_start, len); 5618 } 5619 5620 // return cipher length (int) 5621 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5622 set_result(retvalue); 5623 return true; 5624 } 5625 5626 //------------------------------get_key_start_from_aescrypt_object----------------------- 5627 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5628 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5629 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5630 if (objAESCryptKey == NULL) return (Node *) NULL; 5631 5632 // now have the array, need to get the start address of the K array 5633 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5634 return k_start; 5635 } 5636 5637 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5638 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5639 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5640 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5641 if (objAESCryptKey == NULL) return (Node *) NULL; 5642 5643 // now have the array, need to get the start address of the lastKey array 5644 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5645 return original_k_start; 5646 } 5647 5648 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5649 // Return node representing slow path of predicate check. 5650 // the pseudo code we want to emulate with this predicate is: 5651 // for encryption: 5652 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5653 // for decryption: 5654 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5655 // note cipher==plain is more conservative than the original java code but that's OK 5656 // 5657 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5658 // The receiver was checked for NULL already. 5659 Node* objCBC = argument(0); 5660 5661 // Load embeddedCipher field of CipherBlockChaining object. 5662 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5663 5664 // get AESCrypt klass for instanceOf check 5665 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5666 // will have same classloader as CipherBlockChaining object 5667 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5668 assert(tinst != NULL, "CBCobj is null"); 5669 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5670 5671 // we want to do an instanceof comparison against the AESCrypt class 5672 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5673 if (!klass_AESCrypt->is_loaded()) { 5674 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5675 Node* ctrl = control(); 5676 set_control(top()); // no regular fast path 5677 return ctrl; 5678 } 5679 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5680 5681 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5682 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5683 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5684 5685 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5686 5687 // for encryption, we are done 5688 if (!decrypting) 5689 return instof_false; // even if it is NULL 5690 5691 // for decryption, we need to add a further check to avoid 5692 // taking the intrinsic path when cipher and plain are the same 5693 // see the original java code for why. 5694 RegionNode* region = new RegionNode(3); 5695 region->init_req(1, instof_false); 5696 Node* src = argument(1); 5697 Node* dest = argument(4); 5698 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5699 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5700 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5701 region->init_req(2, src_dest_conjoint); 5702 5703 record_for_igvn(region); 5704 return _gvn.transform(region); 5705 } 5706 5707 //------------------------------inline_sha_implCompress----------------------- 5708 // 5709 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 5710 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 5711 // 5712 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 5713 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 5714 // 5715 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 5716 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 5717 // 5718 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 5719 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 5720 5721 Node* sha_obj = argument(0); 5722 Node* src = argument(1); // type oop 5723 Node* ofs = argument(2); // type int 5724 5725 const Type* src_type = src->Value(&_gvn); 5726 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5727 if (top_src == NULL || top_src->klass() == NULL) { 5728 // failed array check 5729 return false; 5730 } 5731 // Figure out the size and type of the elements we will be copying. 5732 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5733 if (src_elem != T_BYTE) { 5734 return false; 5735 } 5736 // 'src_start' points to src array + offset 5737 Node* src_start = array_element_address(src, ofs, src_elem); 5738 Node* state = NULL; 5739 address stubAddr; 5740 const char *stubName; 5741 5742 switch(id) { 5743 case vmIntrinsics::_sha_implCompress: 5744 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 5745 state = get_state_from_sha_object(sha_obj); 5746 stubAddr = StubRoutines::sha1_implCompress(); 5747 stubName = "sha1_implCompress"; 5748 break; 5749 case vmIntrinsics::_sha2_implCompress: 5750 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 5751 state = get_state_from_sha_object(sha_obj); 5752 stubAddr = StubRoutines::sha256_implCompress(); 5753 stubName = "sha256_implCompress"; 5754 break; 5755 case vmIntrinsics::_sha5_implCompress: 5756 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 5757 state = get_state_from_sha5_object(sha_obj); 5758 stubAddr = StubRoutines::sha512_implCompress(); 5759 stubName = "sha512_implCompress"; 5760 break; 5761 default: 5762 fatal_unexpected_iid(id); 5763 return false; 5764 } 5765 if (state == NULL) return false; 5766 5767 // Call the stub. 5768 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 5769 stubAddr, stubName, TypePtr::BOTTOM, 5770 src_start, state); 5771 5772 return true; 5773 } 5774 5775 //------------------------------inline_digestBase_implCompressMB----------------------- 5776 // 5777 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 5778 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 5779 // 5780 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 5781 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5782 "need SHA1/SHA256/SHA512 instruction support"); 5783 assert((uint)predicate < 3, "sanity"); 5784 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 5785 5786 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 5787 Node* src = argument(1); // byte[] array 5788 Node* ofs = argument(2); // type int 5789 Node* limit = argument(3); // type int 5790 5791 const Type* src_type = src->Value(&_gvn); 5792 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5793 if (top_src == NULL || top_src->klass() == NULL) { 5794 // failed array check 5795 return false; 5796 } 5797 // Figure out the size and type of the elements we will be copying. 5798 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5799 if (src_elem != T_BYTE) { 5800 return false; 5801 } 5802 // 'src_start' points to src array + offset 5803 Node* src_start = array_element_address(src, ofs, src_elem); 5804 5805 const char* klass_SHA_name = NULL; 5806 const char* stub_name = NULL; 5807 address stub_addr = NULL; 5808 bool long_state = false; 5809 5810 switch (predicate) { 5811 case 0: 5812 if (UseSHA1Intrinsics) { 5813 klass_SHA_name = "sun/security/provider/SHA"; 5814 stub_name = "sha1_implCompressMB"; 5815 stub_addr = StubRoutines::sha1_implCompressMB(); 5816 } 5817 break; 5818 case 1: 5819 if (UseSHA256Intrinsics) { 5820 klass_SHA_name = "sun/security/provider/SHA2"; 5821 stub_name = "sha256_implCompressMB"; 5822 stub_addr = StubRoutines::sha256_implCompressMB(); 5823 } 5824 break; 5825 case 2: 5826 if (UseSHA512Intrinsics) { 5827 klass_SHA_name = "sun/security/provider/SHA5"; 5828 stub_name = "sha512_implCompressMB"; 5829 stub_addr = StubRoutines::sha512_implCompressMB(); 5830 long_state = true; 5831 } 5832 break; 5833 default: 5834 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5835 } 5836 if (klass_SHA_name != NULL) { 5837 // get DigestBase klass to lookup for SHA klass 5838 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 5839 assert(tinst != NULL, "digestBase_obj is not instance???"); 5840 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5841 5842 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5843 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 5844 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5845 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 5846 } 5847 return false; 5848 } 5849 //------------------------------inline_sha_implCompressMB----------------------- 5850 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 5851 bool long_state, address stubAddr, const char *stubName, 5852 Node* src_start, Node* ofs, Node* limit) { 5853 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 5854 const TypeOopPtr* xtype = aklass->as_instance_type(); 5855 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 5856 sha_obj = _gvn.transform(sha_obj); 5857 5858 Node* state; 5859 if (long_state) { 5860 state = get_state_from_sha5_object(sha_obj); 5861 } else { 5862 state = get_state_from_sha_object(sha_obj); 5863 } 5864 if (state == NULL) return false; 5865 5866 // Call the stub. 5867 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5868 OptoRuntime::digestBase_implCompressMB_Type(), 5869 stubAddr, stubName, TypePtr::BOTTOM, 5870 src_start, state, ofs, limit); 5871 // return ofs (int) 5872 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5873 set_result(result); 5874 5875 return true; 5876 } 5877 5878 //------------------------------get_state_from_sha_object----------------------- 5879 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 5880 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 5881 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 5882 if (sha_state == NULL) return (Node *) NULL; 5883 5884 // now have the array, need to get the start address of the state array 5885 Node* state = array_element_address(sha_state, intcon(0), T_INT); 5886 return state; 5887 } 5888 5889 //------------------------------get_state_from_sha5_object----------------------- 5890 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 5891 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 5892 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 5893 if (sha_state == NULL) return (Node *) NULL; 5894 5895 // now have the array, need to get the start address of the state array 5896 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 5897 return state; 5898 } 5899 5900 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 5901 // Return node representing slow path of predicate check. 5902 // the pseudo code we want to emulate with this predicate is: 5903 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 5904 // 5905 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 5906 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5907 "need SHA1/SHA256/SHA512 instruction support"); 5908 assert((uint)predicate < 3, "sanity"); 5909 5910 // The receiver was checked for NULL already. 5911 Node* digestBaseObj = argument(0); 5912 5913 // get DigestBase klass for instanceOf check 5914 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 5915 assert(tinst != NULL, "digestBaseObj is null"); 5916 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5917 5918 const char* klass_SHA_name = NULL; 5919 switch (predicate) { 5920 case 0: 5921 if (UseSHA1Intrinsics) { 5922 // we want to do an instanceof comparison against the SHA class 5923 klass_SHA_name = "sun/security/provider/SHA"; 5924 } 5925 break; 5926 case 1: 5927 if (UseSHA256Intrinsics) { 5928 // we want to do an instanceof comparison against the SHA2 class 5929 klass_SHA_name = "sun/security/provider/SHA2"; 5930 } 5931 break; 5932 case 2: 5933 if (UseSHA512Intrinsics) { 5934 // we want to do an instanceof comparison against the SHA5 class 5935 klass_SHA_name = "sun/security/provider/SHA5"; 5936 } 5937 break; 5938 default: 5939 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5940 } 5941 5942 ciKlass* klass_SHA = NULL; 5943 if (klass_SHA_name != NULL) { 5944 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5945 } 5946 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 5947 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 5948 Node* ctrl = control(); 5949 set_control(top()); // no intrinsic path 5950 return ctrl; 5951 } 5952 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5953 5954 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 5955 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 5956 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5957 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5958 5959 return instof_false; // even if it is NULL 5960 } 5961 5962 bool LibraryCallKit::inline_profileBoolean() { 5963 Node* counts = argument(1); 5964 const TypeAryPtr* ary = NULL; 5965 ciArray* aobj = NULL; 5966 if (counts->is_Con() 5967 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 5968 && (aobj = ary->const_oop()->as_array()) != NULL 5969 && (aobj->length() == 2)) { 5970 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 5971 jint false_cnt = aobj->element_value(0).as_int(); 5972 jint true_cnt = aobj->element_value(1).as_int(); 5973 5974 method()->set_injected_profile(true); 5975 5976 if (C->log() != NULL) { 5977 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 5978 false_cnt, true_cnt); 5979 } 5980 5981 if (false_cnt + true_cnt == 0) { 5982 // According to profile, never executed. 5983 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 5984 Deoptimization::Action_reinterpret); 5985 return true; 5986 } 5987 // Stop profiling. 5988 // MethodHandleImpl::profileBoolean() has profiling logic in it's bytecode. 5989 // By replacing method's body with profile data (represented as ProfileBooleanNode 5990 // on IR level) we effectively disable profiling. 5991 // It enables full speed execution once optimized code is generated. 5992 Node* profile = _gvn.transform(new ProfileBooleanNode(argument(0), false_cnt, true_cnt)); 5993 C->record_for_igvn(profile); 5994 set_result(profile); 5995 return true; 5996 } else { 5997 // Continue profiling. 5998 // Profile data isn't available at the moment. So, execute method's bytecode version. 5999 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6000 // is compiled and counters aren't available since corresponding MethodHandle 6001 // isn't a compile-time constant. 6002 return false; 6003 } 6004 }