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