1 /* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "oops/objArrayKlass.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/callGenerator.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/convertnode.hpp" 38 #include "opto/countbitsnode.hpp" 39 #include "opto/intrinsicnode.hpp" 40 #include "opto/idealKit.hpp" 41 #include "opto/mathexactnode.hpp" 42 #include "opto/movenode.hpp" 43 #include "opto/mulnode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/opaquenode.hpp" 46 #include "opto/parse.hpp" 47 #include "opto/runtime.hpp" 48 #include "opto/subnode.hpp" 49 #include "prims/nativeLookup.hpp" 50 #include "runtime/sharedRuntime.hpp" 51 #include "trace/traceMacros.hpp" 52 53 class LibraryIntrinsic : public InlineCallGenerator { 54 // Extend the set of intrinsics known to the runtime: 55 public: 56 private: 57 bool _is_virtual; 58 bool _does_virtual_dispatch; 59 int8_t _predicates_count; // Intrinsic is predicated by several conditions 60 int8_t _last_predicate; // Last generated predicate 61 vmIntrinsics::ID _intrinsic_id; 62 63 public: 64 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 65 : InlineCallGenerator(m), 66 _is_virtual(is_virtual), 67 _does_virtual_dispatch(does_virtual_dispatch), 68 _predicates_count((int8_t)predicates_count), 69 _last_predicate((int8_t)-1), 70 _intrinsic_id(id) 71 { 72 } 73 virtual bool is_intrinsic() const { return true; } 74 virtual bool is_virtual() const { return _is_virtual; } 75 virtual bool is_predicated() const { return _predicates_count > 0; } 76 virtual int predicates_count() const { return _predicates_count; } 77 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 78 virtual JVMState* generate(JVMState* jvms); 79 virtual Node* generate_predicate(JVMState* jvms, int predicate); 80 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 81 }; 82 83 84 // Local helper class for LibraryIntrinsic: 85 class LibraryCallKit : public GraphKit { 86 private: 87 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 88 Node* _result; // the result node, if any 89 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 90 91 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 92 93 public: 94 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 95 : GraphKit(jvms), 96 _intrinsic(intrinsic), 97 _result(NULL) 98 { 99 // Check if this is a root compile. In that case we don't have a caller. 100 if (!jvms->has_method()) { 101 _reexecute_sp = sp(); 102 } else { 103 // Find out how many arguments the interpreter needs when deoptimizing 104 // and save the stack pointer value so it can used by uncommon_trap. 105 // We find the argument count by looking at the declared signature. 106 bool ignored_will_link; 107 ciSignature* declared_signature = NULL; 108 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 109 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 110 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 111 } 112 } 113 114 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 115 116 ciMethod* caller() const { return jvms()->method(); } 117 int bci() const { return jvms()->bci(); } 118 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 119 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 120 ciMethod* callee() const { return _intrinsic->method(); } 121 122 bool try_to_inline(int predicate); 123 Node* try_to_predicate(int predicate); 124 125 void push_result() { 126 // Push the result onto the stack. 127 if (!stopped() && result() != NULL) { 128 BasicType bt = result()->bottom_type()->basic_type(); 129 push_node(bt, result()); 130 } 131 } 132 133 private: 134 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 135 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 136 } 137 138 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 139 void set_result(RegionNode* region, PhiNode* value); 140 Node* result() { return _result; } 141 142 virtual int reexecute_sp() { return _reexecute_sp; } 143 144 // Helper functions to inline natives 145 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 146 Node* generate_slow_guard(Node* test, RegionNode* region); 147 Node* generate_fair_guard(Node* test, RegionNode* region); 148 Node* generate_negative_guard(Node* index, RegionNode* region, 149 // resulting CastII of index: 150 Node* *pos_index = NULL); 151 Node* generate_limit_guard(Node* offset, Node* subseq_length, 152 Node* array_length, 153 RegionNode* region); 154 Node* generate_current_thread(Node* &tls_output); 155 Node* load_mirror_from_klass(Node* klass); 156 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 157 RegionNode* region, int null_path, 158 int offset); 159 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path) { 161 int offset = java_lang_Class::klass_offset_in_bytes(); 162 return load_klass_from_mirror_common(mirror, never_see_null, 163 region, null_path, 164 offset); 165 } 166 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 167 RegionNode* region, int null_path) { 168 int offset = java_lang_Class::array_klass_offset_in_bytes(); 169 return load_klass_from_mirror_common(mirror, never_see_null, 170 region, null_path, 171 offset); 172 } 173 Node* generate_access_flags_guard(Node* kls, 174 int modifier_mask, int modifier_bits, 175 RegionNode* region); 176 Node* generate_interface_guard(Node* kls, RegionNode* region); 177 Node* generate_array_guard(Node* kls, RegionNode* region) { 178 return generate_array_guard_common(kls, region, false, false); 179 } 180 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 181 return generate_array_guard_common(kls, region, false, true); 182 } 183 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 184 return generate_array_guard_common(kls, region, true, false); 185 } 186 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 187 return generate_array_guard_common(kls, region, true, true); 188 } 189 Node* generate_array_guard_common(Node* kls, RegionNode* region, 190 bool obj_array, bool not_array); 191 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 192 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 193 bool is_virtual = false, bool is_static = false); 194 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 195 return generate_method_call(method_id, false, true); 196 } 197 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 198 return generate_method_call(method_id, true, false); 199 } 200 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 201 202 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 203 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 204 bool inline_string_compareTo(); 205 bool inline_string_indexOf(); 206 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 207 bool inline_string_equals(); 208 Node* round_double_node(Node* n); 209 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 210 bool inline_math_native(vmIntrinsics::ID id); 211 bool inline_trig(vmIntrinsics::ID id); 212 bool inline_math(vmIntrinsics::ID id); 213 template <typename OverflowOp> 214 bool inline_math_overflow(Node* arg1, Node* arg2); 215 void inline_math_mathExact(Node* math, Node* test); 216 bool inline_math_addExactI(bool is_increment); 217 bool inline_math_addExactL(bool is_increment); 218 bool inline_math_multiplyExactI(); 219 bool inline_math_multiplyExactL(); 220 bool inline_math_negateExactI(); 221 bool inline_math_negateExactL(); 222 bool inline_math_subtractExactI(bool is_decrement); 223 bool inline_math_subtractExactL(bool is_decrement); 224 bool inline_exp(); 225 bool inline_pow(); 226 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 227 bool inline_min_max(vmIntrinsics::ID id); 228 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 229 // This returns Type::AnyPtr, RawPtr, or OopPtr. 230 int classify_unsafe_addr(Node* &base, Node* &offset); 231 Node* make_unsafe_address(Node* base, Node* offset); 232 // Helper for inline_unsafe_access. 233 // Generates the guards that check whether the result of 234 // Unsafe.getObject should be recorded in an SATB log buffer. 235 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 236 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 237 static bool klass_needs_init_guard(Node* kls); 238 bool inline_unsafe_allocate(); 239 bool inline_unsafe_copyMemory(); 240 bool inline_native_currentThread(); 241 #ifdef TRACE_HAVE_INTRINSICS 242 bool inline_native_classID(); 243 bool inline_native_threadID(); 244 #endif 245 bool inline_native_time_funcs(address method, const char* funcName); 246 bool inline_native_isInterrupted(); 247 bool inline_native_Class_query(vmIntrinsics::ID id); 248 bool inline_native_subtype_check(); 249 250 bool inline_native_newArray(); 251 bool inline_native_getLength(); 252 bool inline_array_copyOf(bool is_copyOfRange); 253 bool inline_array_equals(); 254 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 255 bool inline_native_clone(bool is_virtual); 256 bool inline_native_Reflection_getCallerClass(); 257 // Helper function for inlining native object hash method 258 bool inline_native_hashcode(bool is_virtual, bool is_static); 259 bool inline_native_getClass(); 260 261 // Helper functions for inlining arraycopy 262 bool inline_arraycopy(); 263 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 264 RegionNode* slow_region); 265 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 266 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 267 268 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 269 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 270 bool inline_unsafe_ordered_store(BasicType type); 271 bool inline_unsafe_fence(vmIntrinsics::ID id); 272 bool inline_fp_conversions(vmIntrinsics::ID id); 273 bool inline_number_methods(vmIntrinsics::ID id); 274 bool inline_reference_get(); 275 bool inline_Class_cast(); 276 bool inline_aescrypt_Block(vmIntrinsics::ID id); 277 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 278 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 279 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 280 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 281 bool inline_sha_implCompress(vmIntrinsics::ID id); 282 bool inline_digestBase_implCompressMB(int predicate); 283 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 284 bool long_state, address stubAddr, const char *stubName, 285 Node* src_start, Node* ofs, Node* limit); 286 Node* get_state_from_sha_object(Node *sha_object); 287 Node* get_state_from_sha5_object(Node *sha_object); 288 Node* inline_digestBase_implCompressMB_predicate(int predicate); 289 bool inline_encodeISOArray(); 290 bool inline_updateCRC32(); 291 bool inline_updateBytesCRC32(); 292 bool inline_updateByteBufferCRC32(); 293 bool inline_multiplyToLen(); 294 bool inline_squareToLen(); 295 bool inline_mulAdd(); 296 297 bool inline_profileBoolean(); 298 bool inline_isCompileConstant(); 299 }; 300 301 302 //---------------------------make_vm_intrinsic---------------------------- 303 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 304 vmIntrinsics::ID id = m->intrinsic_id(); 305 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 306 307 ccstr disable_intr = NULL; 308 309 if ((DisableIntrinsic[0] != '\0' 310 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 311 (method_has_option_value("DisableIntrinsic", disable_intr) 312 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 313 // disabled by a user request on the command line: 314 // example: -XX:DisableIntrinsic=_hashCode,_getClass 315 return NULL; 316 } 317 318 if (!m->is_loaded()) { 319 // do not attempt to inline unloaded methods 320 return NULL; 321 } 322 323 // Only a few intrinsics implement a virtual dispatch. 324 // They are expensive calls which are also frequently overridden. 325 if (is_virtual) { 326 switch (id) { 327 case vmIntrinsics::_hashCode: 328 case vmIntrinsics::_clone: 329 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 330 break; 331 default: 332 return NULL; 333 } 334 } 335 336 // -XX:-InlineNatives disables nearly all intrinsics: 337 if (!InlineNatives) { 338 switch (id) { 339 case vmIntrinsics::_indexOf: 340 case vmIntrinsics::_compareTo: 341 case vmIntrinsics::_equals: 342 case vmIntrinsics::_equalsC: 343 case vmIntrinsics::_getAndAddInt: 344 case vmIntrinsics::_getAndAddLong: 345 case vmIntrinsics::_getAndSetInt: 346 case vmIntrinsics::_getAndSetLong: 347 case vmIntrinsics::_getAndSetObject: 348 case vmIntrinsics::_loadFence: 349 case vmIntrinsics::_storeFence: 350 case vmIntrinsics::_fullFence: 351 break; // InlineNatives does not control String.compareTo 352 case vmIntrinsics::_Reference_get: 353 break; // InlineNatives does not control Reference.get 354 default: 355 return NULL; 356 } 357 } 358 359 int predicates = 0; 360 bool does_virtual_dispatch = false; 361 362 switch (id) { 363 case vmIntrinsics::_compareTo: 364 if (!SpecialStringCompareTo) return NULL; 365 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 366 break; 367 case vmIntrinsics::_indexOf: 368 if (!SpecialStringIndexOf) return NULL; 369 break; 370 case vmIntrinsics::_equals: 371 if (!SpecialStringEquals) return NULL; 372 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 373 break; 374 case vmIntrinsics::_equalsC: 375 if (!SpecialArraysEquals) return NULL; 376 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 377 break; 378 case vmIntrinsics::_arraycopy: 379 if (!InlineArrayCopy) return NULL; 380 break; 381 case vmIntrinsics::_copyMemory: 382 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 383 if (!InlineArrayCopy) return NULL; 384 break; 385 case vmIntrinsics::_hashCode: 386 if (!InlineObjectHash) return NULL; 387 does_virtual_dispatch = true; 388 break; 389 case vmIntrinsics::_clone: 390 does_virtual_dispatch = true; 391 case vmIntrinsics::_copyOf: 392 case vmIntrinsics::_copyOfRange: 393 if (!InlineObjectCopy) return NULL; 394 // These also use the arraycopy intrinsic mechanism: 395 if (!InlineArrayCopy) return NULL; 396 break; 397 case vmIntrinsics::_encodeISOArray: 398 if (!SpecialEncodeISOArray) return NULL; 399 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 400 break; 401 case vmIntrinsics::_checkIndex: 402 // We do not intrinsify this. The optimizer does fine with it. 403 return NULL; 404 405 case vmIntrinsics::_getCallerClass: 406 if (!InlineReflectionGetCallerClass) return NULL; 407 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 408 break; 409 410 case vmIntrinsics::_bitCount_i: 411 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 412 break; 413 414 case vmIntrinsics::_bitCount_l: 415 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 416 break; 417 418 case vmIntrinsics::_numberOfLeadingZeros_i: 419 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 420 break; 421 422 case vmIntrinsics::_numberOfLeadingZeros_l: 423 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 424 break; 425 426 case vmIntrinsics::_numberOfTrailingZeros_i: 427 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 428 break; 429 430 case vmIntrinsics::_numberOfTrailingZeros_l: 431 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 432 break; 433 434 case vmIntrinsics::_reverseBytes_c: 435 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 436 break; 437 case vmIntrinsics::_reverseBytes_s: 438 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 439 break; 440 case vmIntrinsics::_reverseBytes_i: 441 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 442 break; 443 case vmIntrinsics::_reverseBytes_l: 444 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 445 break; 446 447 case vmIntrinsics::_Reference_get: 448 // Use the intrinsic version of Reference.get() so that the value in 449 // the referent field can be registered by the G1 pre-barrier code. 450 // Also add memory barrier to prevent commoning reads from this field 451 // across safepoint since GC can change it value. 452 break; 453 454 case vmIntrinsics::_compareAndSwapObject: 455 #ifdef _LP64 456 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 457 #endif 458 break; 459 460 case vmIntrinsics::_compareAndSwapLong: 461 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 462 break; 463 464 case vmIntrinsics::_getAndAddInt: 465 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 466 break; 467 468 case vmIntrinsics::_getAndAddLong: 469 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 470 break; 471 472 case vmIntrinsics::_getAndSetInt: 473 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 474 break; 475 476 case vmIntrinsics::_getAndSetLong: 477 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 478 break; 479 480 case vmIntrinsics::_getAndSetObject: 481 #ifdef _LP64 482 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 483 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 484 break; 485 #else 486 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 487 break; 488 #endif 489 490 case vmIntrinsics::_aescrypt_encryptBlock: 491 case vmIntrinsics::_aescrypt_decryptBlock: 492 if (!UseAESIntrinsics) return NULL; 493 break; 494 495 case vmIntrinsics::_multiplyToLen: 496 if (!UseMultiplyToLenIntrinsic) return NULL; 497 break; 498 499 case vmIntrinsics::_squareToLen: 500 if (!UseSquareToLenIntrinsic) return NULL; 501 break; 502 503 case vmIntrinsics::_mulAdd: 504 if (!UseMulAddIntrinsic) return NULL; 505 break; 506 507 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 508 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 509 if (!UseAESIntrinsics) return NULL; 510 // these two require the predicated logic 511 predicates = 1; 512 break; 513 514 case vmIntrinsics::_sha_implCompress: 515 if (!UseSHA1Intrinsics) return NULL; 516 break; 517 518 case vmIntrinsics::_sha2_implCompress: 519 if (!UseSHA256Intrinsics) return NULL; 520 break; 521 522 case vmIntrinsics::_sha5_implCompress: 523 if (!UseSHA512Intrinsics) return NULL; 524 break; 525 526 case vmIntrinsics::_digestBase_implCompressMB: 527 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 528 predicates = 3; 529 break; 530 531 case vmIntrinsics::_updateCRC32: 532 case vmIntrinsics::_updateBytesCRC32: 533 case vmIntrinsics::_updateByteBufferCRC32: 534 if (!UseCRC32Intrinsics) return NULL; 535 break; 536 537 case vmIntrinsics::_incrementExactI: 538 case vmIntrinsics::_addExactI: 539 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 540 break; 541 case vmIntrinsics::_incrementExactL: 542 case vmIntrinsics::_addExactL: 543 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 544 break; 545 case vmIntrinsics::_decrementExactI: 546 case vmIntrinsics::_subtractExactI: 547 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 548 break; 549 case vmIntrinsics::_decrementExactL: 550 case vmIntrinsics::_subtractExactL: 551 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 552 break; 553 case vmIntrinsics::_negateExactI: 554 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 555 break; 556 case vmIntrinsics::_negateExactL: 557 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 558 break; 559 case vmIntrinsics::_multiplyExactI: 560 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 561 break; 562 case vmIntrinsics::_multiplyExactL: 563 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 564 break; 565 566 case vmIntrinsics::_getShortUnaligned: 567 case vmIntrinsics::_getCharUnaligned: 568 case vmIntrinsics::_getIntUnaligned: 569 case vmIntrinsics::_getLongUnaligned: 570 case vmIntrinsics::_putShortUnaligned: 571 case vmIntrinsics::_putCharUnaligned: 572 case vmIntrinsics::_putIntUnaligned: 573 case vmIntrinsics::_putLongUnaligned: 574 if (!UseUnalignedAccesses) return NULL; 575 break; 576 577 default: 578 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 579 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 580 break; 581 } 582 583 // -XX:-InlineClassNatives disables natives from the Class class. 584 // The flag applies to all reflective calls, notably Array.newArray 585 // (visible to Java programmers as Array.newInstance). 586 if (m->holder()->name() == ciSymbol::java_lang_Class() || 587 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 588 if (!InlineClassNatives) return NULL; 589 } 590 591 // -XX:-InlineThreadNatives disables natives from the Thread class. 592 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 593 if (!InlineThreadNatives) return NULL; 594 } 595 596 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 597 if (m->holder()->name() == ciSymbol::java_lang_Math() || 598 m->holder()->name() == ciSymbol::java_lang_Float() || 599 m->holder()->name() == ciSymbol::java_lang_Double()) { 600 if (!InlineMathNatives) return NULL; 601 } 602 603 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 604 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 605 if (!InlineUnsafeOps) return NULL; 606 } 607 608 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 609 } 610 611 //----------------------register_library_intrinsics----------------------- 612 // Initialize this file's data structures, for each Compile instance. 613 void Compile::register_library_intrinsics() { 614 // Nothing to do here. 615 } 616 617 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 618 LibraryCallKit kit(jvms, this); 619 Compile* C = kit.C; 620 int nodes = C->unique(); 621 #ifndef PRODUCT 622 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 623 char buf[1000]; 624 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 625 tty->print_cr("Intrinsic %s", str); 626 } 627 #endif 628 ciMethod* callee = kit.callee(); 629 const int bci = kit.bci(); 630 631 // Try to inline the intrinsic. 632 if (kit.try_to_inline(_last_predicate)) { 633 if (C->print_intrinsics() || C->print_inlining()) { 634 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 635 } 636 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 637 if (C->log()) { 638 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 639 vmIntrinsics::name_at(intrinsic_id()), 640 (is_virtual() ? " virtual='1'" : ""), 641 C->unique() - nodes); 642 } 643 // Push the result from the inlined method onto the stack. 644 kit.push_result(); 645 C->print_inlining_update(this); 646 return kit.transfer_exceptions_into_jvms(); 647 } 648 649 // The intrinsic bailed out 650 if (C->print_intrinsics() || C->print_inlining()) { 651 if (jvms->has_method()) { 652 // Not a root compile. 653 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 654 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 655 } else { 656 // Root compile 657 tty->print("Did not generate intrinsic %s%s at bci:%d in", 658 vmIntrinsics::name_at(intrinsic_id()), 659 (is_virtual() ? " (virtual)" : ""), bci); 660 } 661 } 662 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 663 C->print_inlining_update(this); 664 return NULL; 665 } 666 667 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 668 LibraryCallKit kit(jvms, this); 669 Compile* C = kit.C; 670 int nodes = C->unique(); 671 _last_predicate = predicate; 672 #ifndef PRODUCT 673 assert(is_predicated() && predicate < predicates_count(), "sanity"); 674 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 675 char buf[1000]; 676 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 677 tty->print_cr("Predicate for intrinsic %s", str); 678 } 679 #endif 680 ciMethod* callee = kit.callee(); 681 const int bci = kit.bci(); 682 683 Node* slow_ctl = kit.try_to_predicate(predicate); 684 if (!kit.failing()) { 685 if (C->print_intrinsics() || C->print_inlining()) { 686 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 687 } 688 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 689 if (C->log()) { 690 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 691 vmIntrinsics::name_at(intrinsic_id()), 692 (is_virtual() ? " virtual='1'" : ""), 693 C->unique() - nodes); 694 } 695 return slow_ctl; // Could be NULL if the check folds. 696 } 697 698 // The intrinsic bailed out 699 if (C->print_intrinsics() || C->print_inlining()) { 700 if (jvms->has_method()) { 701 // Not a root compile. 702 const char* msg = "failed to generate predicate for intrinsic"; 703 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 704 } else { 705 // Root compile 706 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 707 vmIntrinsics::name_at(intrinsic_id()), 708 (is_virtual() ? " (virtual)" : ""), bci); 709 } 710 } 711 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 712 return NULL; 713 } 714 715 bool LibraryCallKit::try_to_inline(int predicate) { 716 // Handle symbolic names for otherwise undistinguished boolean switches: 717 const bool is_store = true; 718 const bool is_native_ptr = true; 719 const bool is_static = true; 720 const bool is_volatile = true; 721 722 if (!jvms()->has_method()) { 723 // Root JVMState has a null method. 724 assert(map()->memory()->Opcode() == Op_Parm, ""); 725 // Insert the memory aliasing node 726 set_all_memory(reset_memory()); 727 } 728 assert(merged_memory(), ""); 729 730 731 switch (intrinsic_id()) { 732 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 733 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 734 case vmIntrinsics::_getClass: return inline_native_getClass(); 735 736 case vmIntrinsics::_dsin: 737 case vmIntrinsics::_dcos: 738 case vmIntrinsics::_dtan: 739 case vmIntrinsics::_dabs: 740 case vmIntrinsics::_datan2: 741 case vmIntrinsics::_dsqrt: 742 case vmIntrinsics::_dexp: 743 case vmIntrinsics::_dlog: 744 case vmIntrinsics::_dlog10: 745 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 746 747 case vmIntrinsics::_min: 748 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 749 750 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 751 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 752 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 753 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 754 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 755 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 756 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 757 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 758 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 759 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 760 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 761 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 762 763 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 764 765 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 766 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 767 case vmIntrinsics::_equals: return inline_string_equals(); 768 769 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 770 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 771 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 772 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 773 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 774 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 775 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 776 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 777 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 778 779 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 780 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 781 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 782 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 783 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 784 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 785 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 786 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 787 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 788 789 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 790 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 791 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 792 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 793 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 794 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 795 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 796 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 797 798 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 799 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 800 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 801 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 802 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 803 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 804 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 805 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 806 807 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 808 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 809 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 810 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 811 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 812 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 813 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 814 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 815 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 816 817 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 818 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 819 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 820 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 821 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 822 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 823 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 824 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 825 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 826 827 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 828 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 829 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 830 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 831 832 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 833 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 834 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 835 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 836 837 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 838 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 839 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 840 841 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 842 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 843 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 844 845 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 846 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 847 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 848 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 849 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 850 851 case vmIntrinsics::_loadFence: 852 case vmIntrinsics::_storeFence: 853 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 854 855 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 856 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 857 858 #ifdef TRACE_HAVE_INTRINSICS 859 case vmIntrinsics::_classID: return inline_native_classID(); 860 case vmIntrinsics::_threadID: return inline_native_threadID(); 861 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 862 #endif 863 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 864 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 865 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 866 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 867 case vmIntrinsics::_newArray: return inline_native_newArray(); 868 case vmIntrinsics::_getLength: return inline_native_getLength(); 869 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 870 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 871 case vmIntrinsics::_equalsC: return inline_array_equals(); 872 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 873 874 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 875 876 case vmIntrinsics::_isInstance: 877 case vmIntrinsics::_getModifiers: 878 case vmIntrinsics::_isInterface: 879 case vmIntrinsics::_isArray: 880 case vmIntrinsics::_isPrimitive: 881 case vmIntrinsics::_getSuperclass: 882 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 883 884 case vmIntrinsics::_floatToRawIntBits: 885 case vmIntrinsics::_floatToIntBits: 886 case vmIntrinsics::_intBitsToFloat: 887 case vmIntrinsics::_doubleToRawLongBits: 888 case vmIntrinsics::_doubleToLongBits: 889 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 890 891 case vmIntrinsics::_numberOfLeadingZeros_i: 892 case vmIntrinsics::_numberOfLeadingZeros_l: 893 case vmIntrinsics::_numberOfTrailingZeros_i: 894 case vmIntrinsics::_numberOfTrailingZeros_l: 895 case vmIntrinsics::_bitCount_i: 896 case vmIntrinsics::_bitCount_l: 897 case vmIntrinsics::_reverseBytes_i: 898 case vmIntrinsics::_reverseBytes_l: 899 case vmIntrinsics::_reverseBytes_s: 900 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 901 902 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 903 904 case vmIntrinsics::_Reference_get: return inline_reference_get(); 905 906 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 907 908 case vmIntrinsics::_aescrypt_encryptBlock: 909 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 910 911 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 912 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 913 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 914 915 case vmIntrinsics::_sha_implCompress: 916 case vmIntrinsics::_sha2_implCompress: 917 case vmIntrinsics::_sha5_implCompress: 918 return inline_sha_implCompress(intrinsic_id()); 919 920 case vmIntrinsics::_digestBase_implCompressMB: 921 return inline_digestBase_implCompressMB(predicate); 922 923 case vmIntrinsics::_multiplyToLen: 924 return inline_multiplyToLen(); 925 926 case vmIntrinsics::_squareToLen: 927 return inline_squareToLen(); 928 929 case vmIntrinsics::_mulAdd: 930 return inline_mulAdd(); 931 932 case vmIntrinsics::_encodeISOArray: 933 return inline_encodeISOArray(); 934 935 case vmIntrinsics::_updateCRC32: 936 return inline_updateCRC32(); 937 case vmIntrinsics::_updateBytesCRC32: 938 return inline_updateBytesCRC32(); 939 case vmIntrinsics::_updateByteBufferCRC32: 940 return inline_updateByteBufferCRC32(); 941 942 case vmIntrinsics::_profileBoolean: 943 return inline_profileBoolean(); 944 case vmIntrinsics::_isCompileConstant: 945 return inline_isCompileConstant(); 946 947 default: 948 // If you get here, it may be that someone has added a new intrinsic 949 // to the list in vmSymbols.hpp without implementing it here. 950 #ifndef PRODUCT 951 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 952 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 953 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 954 } 955 #endif 956 return false; 957 } 958 } 959 960 Node* LibraryCallKit::try_to_predicate(int predicate) { 961 if (!jvms()->has_method()) { 962 // Root JVMState has a null method. 963 assert(map()->memory()->Opcode() == Op_Parm, ""); 964 // Insert the memory aliasing node 965 set_all_memory(reset_memory()); 966 } 967 assert(merged_memory(), ""); 968 969 switch (intrinsic_id()) { 970 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 971 return inline_cipherBlockChaining_AESCrypt_predicate(false); 972 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 973 return inline_cipherBlockChaining_AESCrypt_predicate(true); 974 case vmIntrinsics::_digestBase_implCompressMB: 975 return inline_digestBase_implCompressMB_predicate(predicate); 976 977 default: 978 // If you get here, it may be that someone has added a new intrinsic 979 // to the list in vmSymbols.hpp without implementing it here. 980 #ifndef PRODUCT 981 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 982 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 983 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 984 } 985 #endif 986 Node* slow_ctl = control(); 987 set_control(top()); // No fast path instrinsic 988 return slow_ctl; 989 } 990 } 991 992 //------------------------------set_result------------------------------- 993 // Helper function for finishing intrinsics. 994 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 995 record_for_igvn(region); 996 set_control(_gvn.transform(region)); 997 set_result( _gvn.transform(value)); 998 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 999 } 1000 1001 //------------------------------generate_guard--------------------------- 1002 // Helper function for generating guarded fast-slow graph structures. 1003 // The given 'test', if true, guards a slow path. If the test fails 1004 // then a fast path can be taken. (We generally hope it fails.) 1005 // In all cases, GraphKit::control() is updated to the fast path. 1006 // The returned value represents the control for the slow path. 1007 // The return value is never 'top'; it is either a valid control 1008 // or NULL if it is obvious that the slow path can never be taken. 1009 // Also, if region and the slow control are not NULL, the slow edge 1010 // is appended to the region. 1011 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 1012 if (stopped()) { 1013 // Already short circuited. 1014 return NULL; 1015 } 1016 1017 // Build an if node and its projections. 1018 // If test is true we take the slow path, which we assume is uncommon. 1019 if (_gvn.type(test) == TypeInt::ZERO) { 1020 // The slow branch is never taken. No need to build this guard. 1021 return NULL; 1022 } 1023 1024 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1025 1026 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1027 if (if_slow == top()) { 1028 // The slow branch is never taken. No need to build this guard. 1029 return NULL; 1030 } 1031 1032 if (region != NULL) 1033 region->add_req(if_slow); 1034 1035 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1036 set_control(if_fast); 1037 1038 return if_slow; 1039 } 1040 1041 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1042 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1043 } 1044 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1045 return generate_guard(test, region, PROB_FAIR); 1046 } 1047 1048 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1049 Node* *pos_index) { 1050 if (stopped()) 1051 return NULL; // already stopped 1052 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1053 return NULL; // index is already adequately typed 1054 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1055 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1056 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1057 if (is_neg != NULL && pos_index != NULL) { 1058 // Emulate effect of Parse::adjust_map_after_if. 1059 Node* ccast = new CastIINode(index, TypeInt::POS); 1060 ccast->set_req(0, control()); 1061 (*pos_index) = _gvn.transform(ccast); 1062 } 1063 return is_neg; 1064 } 1065 1066 // Make sure that 'position' is a valid limit index, in [0..length]. 1067 // There are two equivalent plans for checking this: 1068 // A. (offset + copyLength) unsigned<= arrayLength 1069 // B. offset <= (arrayLength - copyLength) 1070 // We require that all of the values above, except for the sum and 1071 // difference, are already known to be non-negative. 1072 // Plan A is robust in the face of overflow, if offset and copyLength 1073 // are both hugely positive. 1074 // 1075 // Plan B is less direct and intuitive, but it does not overflow at 1076 // all, since the difference of two non-negatives is always 1077 // representable. Whenever Java methods must perform the equivalent 1078 // check they generally use Plan B instead of Plan A. 1079 // For the moment we use Plan A. 1080 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1081 Node* subseq_length, 1082 Node* array_length, 1083 RegionNode* region) { 1084 if (stopped()) 1085 return NULL; // already stopped 1086 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1087 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1088 return NULL; // common case of whole-array copy 1089 Node* last = subseq_length; 1090 if (!zero_offset) // last += offset 1091 last = _gvn.transform(new AddINode(last, offset)); 1092 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1093 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1094 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1095 return is_over; 1096 } 1097 1098 1099 //--------------------------generate_current_thread-------------------- 1100 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1101 ciKlass* thread_klass = env()->Thread_klass(); 1102 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1103 Node* thread = _gvn.transform(new ThreadLocalNode()); 1104 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1105 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1106 tls_output = thread; 1107 return threadObj; 1108 } 1109 1110 1111 //------------------------------make_string_method_node------------------------ 1112 // Helper method for String intrinsic functions. This version is called 1113 // with str1 and str2 pointing to String object nodes. 1114 // 1115 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1116 Node* no_ctrl = NULL; 1117 1118 // Get start addr of string 1119 Node* str1_value = load_String_value(no_ctrl, str1); 1120 Node* str1_offset = load_String_offset(no_ctrl, str1); 1121 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1122 1123 // Get length of string 1 1124 Node* str1_len = load_String_length(no_ctrl, str1); 1125 1126 Node* str2_value = load_String_value(no_ctrl, str2); 1127 Node* str2_offset = load_String_offset(no_ctrl, str2); 1128 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1129 1130 Node* str2_len = NULL; 1131 Node* result = NULL; 1132 1133 switch (opcode) { 1134 case Op_StrIndexOf: 1135 // Get length of string 2 1136 str2_len = load_String_length(no_ctrl, str2); 1137 1138 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1139 str1_start, str1_len, str2_start, str2_len); 1140 break; 1141 case Op_StrComp: 1142 // Get length of string 2 1143 str2_len = load_String_length(no_ctrl, str2); 1144 1145 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1146 str1_start, str1_len, str2_start, str2_len); 1147 break; 1148 case Op_StrEquals: 1149 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1150 str1_start, str2_start, str1_len); 1151 break; 1152 default: 1153 ShouldNotReachHere(); 1154 return NULL; 1155 } 1156 1157 // All these intrinsics have checks. 1158 C->set_has_split_ifs(true); // Has chance for split-if optimization 1159 1160 return _gvn.transform(result); 1161 } 1162 1163 // Helper method for String intrinsic functions. This version is called 1164 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1165 // to Int nodes containing the lenghts of str1 and str2. 1166 // 1167 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1168 Node* result = NULL; 1169 switch (opcode) { 1170 case Op_StrIndexOf: 1171 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1172 str1_start, cnt1, str2_start, cnt2); 1173 break; 1174 case Op_StrComp: 1175 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1176 str1_start, cnt1, str2_start, cnt2); 1177 break; 1178 case Op_StrEquals: 1179 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1180 str1_start, str2_start, cnt1); 1181 break; 1182 default: 1183 ShouldNotReachHere(); 1184 return NULL; 1185 } 1186 1187 // All these intrinsics have checks. 1188 C->set_has_split_ifs(true); // Has chance for split-if optimization 1189 1190 return _gvn.transform(result); 1191 } 1192 1193 //------------------------------inline_string_compareTo------------------------ 1194 // public int java.lang.String.compareTo(String anotherString); 1195 bool LibraryCallKit::inline_string_compareTo() { 1196 Node* receiver = null_check(argument(0)); 1197 Node* arg = null_check(argument(1)); 1198 if (stopped()) { 1199 return true; 1200 } 1201 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1202 return true; 1203 } 1204 1205 //------------------------------inline_string_equals------------------------ 1206 bool LibraryCallKit::inline_string_equals() { 1207 Node* receiver = null_check_receiver(); 1208 // NOTE: Do not null check argument for String.equals() because spec 1209 // allows to specify NULL as argument. 1210 Node* argument = this->argument(1); 1211 if (stopped()) { 1212 return true; 1213 } 1214 1215 // paths (plus control) merge 1216 RegionNode* region = new RegionNode(5); 1217 Node* phi = new PhiNode(region, TypeInt::BOOL); 1218 1219 // does source == target string? 1220 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1221 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1222 1223 Node* if_eq = generate_slow_guard(bol, NULL); 1224 if (if_eq != NULL) { 1225 // receiver == argument 1226 phi->init_req(2, intcon(1)); 1227 region->init_req(2, if_eq); 1228 } 1229 1230 // get String klass for instanceOf 1231 ciInstanceKlass* klass = env()->String_klass(); 1232 1233 if (!stopped()) { 1234 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1235 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1236 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1237 1238 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1239 //instanceOf == true, fallthrough 1240 1241 if (inst_false != NULL) { 1242 phi->init_req(3, intcon(0)); 1243 region->init_req(3, inst_false); 1244 } 1245 } 1246 1247 if (!stopped()) { 1248 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1249 1250 // Properly cast the argument to String 1251 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1252 // This path is taken only when argument's type is String:NotNull. 1253 argument = cast_not_null(argument, false); 1254 1255 Node* no_ctrl = NULL; 1256 1257 // Get start addr of receiver 1258 Node* receiver_val = load_String_value(no_ctrl, receiver); 1259 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1260 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1261 1262 // Get length of receiver 1263 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1264 1265 // Get start addr of argument 1266 Node* argument_val = load_String_value(no_ctrl, argument); 1267 Node* argument_offset = load_String_offset(no_ctrl, argument); 1268 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1269 1270 // Get length of argument 1271 Node* argument_cnt = load_String_length(no_ctrl, argument); 1272 1273 // Check for receiver count != argument count 1274 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1275 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1276 Node* if_ne = generate_slow_guard(bol, NULL); 1277 if (if_ne != NULL) { 1278 phi->init_req(4, intcon(0)); 1279 region->init_req(4, if_ne); 1280 } 1281 1282 // Check for count == 0 is done by assembler code for StrEquals. 1283 1284 if (!stopped()) { 1285 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1286 phi->init_req(1, equals); 1287 region->init_req(1, control()); 1288 } 1289 } 1290 1291 // post merge 1292 set_control(_gvn.transform(region)); 1293 record_for_igvn(region); 1294 1295 set_result(_gvn.transform(phi)); 1296 return true; 1297 } 1298 1299 //------------------------------inline_array_equals---------------------------- 1300 bool LibraryCallKit::inline_array_equals() { 1301 Node* arg1 = argument(0); 1302 Node* arg2 = argument(1); 1303 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1304 return true; 1305 } 1306 1307 // Java version of String.indexOf(constant string) 1308 // class StringDecl { 1309 // StringDecl(char[] ca) { 1310 // offset = 0; 1311 // count = ca.length; 1312 // value = ca; 1313 // } 1314 // int offset; 1315 // int count; 1316 // char[] value; 1317 // } 1318 // 1319 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1320 // int targetOffset, int cache_i, int md2) { 1321 // int cache = cache_i; 1322 // int sourceOffset = string_object.offset; 1323 // int sourceCount = string_object.count; 1324 // int targetCount = target_object.length; 1325 // 1326 // int targetCountLess1 = targetCount - 1; 1327 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1328 // 1329 // char[] source = string_object.value; 1330 // char[] target = target_object; 1331 // int lastChar = target[targetCountLess1]; 1332 // 1333 // outer_loop: 1334 // for (int i = sourceOffset; i < sourceEnd; ) { 1335 // int src = source[i + targetCountLess1]; 1336 // if (src == lastChar) { 1337 // // With random strings and a 4-character alphabet, 1338 // // reverse matching at this point sets up 0.8% fewer 1339 // // frames, but (paradoxically) makes 0.3% more probes. 1340 // // Since those probes are nearer the lastChar probe, 1341 // // there is may be a net D$ win with reverse matching. 1342 // // But, reversing loop inhibits unroll of inner loop 1343 // // for unknown reason. So, does running outer loop from 1344 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1345 // for (int j = 0; j < targetCountLess1; j++) { 1346 // if (target[targetOffset + j] != source[i+j]) { 1347 // if ((cache & (1 << source[i+j])) == 0) { 1348 // if (md2 < j+1) { 1349 // i += j+1; 1350 // continue outer_loop; 1351 // } 1352 // } 1353 // i += md2; 1354 // continue outer_loop; 1355 // } 1356 // } 1357 // return i - sourceOffset; 1358 // } 1359 // if ((cache & (1 << src)) == 0) { 1360 // i += targetCountLess1; 1361 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1362 // i++; 1363 // } 1364 // return -1; 1365 // } 1366 1367 //------------------------------string_indexOf------------------------ 1368 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1369 jint cache_i, jint md2_i) { 1370 1371 Node* no_ctrl = NULL; 1372 float likely = PROB_LIKELY(0.9); 1373 float unlikely = PROB_UNLIKELY(0.9); 1374 1375 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1376 1377 Node* source = load_String_value(no_ctrl, string_object); 1378 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1379 Node* sourceCount = load_String_length(no_ctrl, string_object); 1380 1381 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1382 jint target_length = target_array->length(); 1383 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1384 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1385 1386 // String.value field is known to be @Stable. 1387 if (UseImplicitStableValues) { 1388 target = cast_array_to_stable(target, target_type); 1389 } 1390 1391 IdealKit kit(this, false, true); 1392 #define __ kit. 1393 Node* zero = __ ConI(0); 1394 Node* one = __ ConI(1); 1395 Node* cache = __ ConI(cache_i); 1396 Node* md2 = __ ConI(md2_i); 1397 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1398 Node* targetCountLess1 = __ ConI(target_length - 1); 1399 Node* targetOffset = __ ConI(targetOffset_i); 1400 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1401 1402 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1403 Node* outer_loop = __ make_label(2 /* goto */); 1404 Node* return_ = __ make_label(1); 1405 1406 __ set(rtn,__ ConI(-1)); 1407 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1408 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1409 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1410 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1411 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1412 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1413 Node* tpj = __ AddI(targetOffset, __ value(j)); 1414 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1415 Node* ipj = __ AddI(__ value(i), __ value(j)); 1416 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1417 __ if_then(targ, BoolTest::ne, src2); { 1418 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1419 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1420 __ increment(i, __ AddI(__ value(j), one)); 1421 __ goto_(outer_loop); 1422 } __ end_if(); __ dead(j); 1423 }__ end_if(); __ dead(j); 1424 __ increment(i, md2); 1425 __ goto_(outer_loop); 1426 }__ end_if(); 1427 __ increment(j, one); 1428 }__ end_loop(); __ dead(j); 1429 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1430 __ goto_(return_); 1431 }__ end_if(); 1432 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1433 __ increment(i, targetCountLess1); 1434 }__ end_if(); 1435 __ increment(i, one); 1436 __ bind(outer_loop); 1437 }__ end_loop(); __ dead(i); 1438 __ bind(return_); 1439 1440 // Final sync IdealKit and GraphKit. 1441 final_sync(kit); 1442 Node* result = __ value(rtn); 1443 #undef __ 1444 C->set_has_loops(true); 1445 return result; 1446 } 1447 1448 //------------------------------inline_string_indexOf------------------------ 1449 bool LibraryCallKit::inline_string_indexOf() { 1450 Node* receiver = argument(0); 1451 Node* arg = argument(1); 1452 1453 Node* result; 1454 if (Matcher::has_match_rule(Op_StrIndexOf) && 1455 UseSSE42Intrinsics) { 1456 // Generate SSE4.2 version of indexOf 1457 // We currently only have match rules that use SSE4.2 1458 1459 receiver = null_check(receiver); 1460 arg = null_check(arg); 1461 if (stopped()) { 1462 return true; 1463 } 1464 1465 // Make the merge point 1466 RegionNode* result_rgn = new RegionNode(4); 1467 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1468 Node* no_ctrl = NULL; 1469 1470 // Get start addr of source string 1471 Node* source = load_String_value(no_ctrl, receiver); 1472 Node* source_offset = load_String_offset(no_ctrl, receiver); 1473 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1474 1475 // Get length of source string 1476 Node* source_cnt = load_String_length(no_ctrl, receiver); 1477 1478 // Get start addr of substring 1479 Node* substr = load_String_value(no_ctrl, arg); 1480 Node* substr_offset = load_String_offset(no_ctrl, arg); 1481 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1482 1483 // Get length of source string 1484 Node* substr_cnt = load_String_length(no_ctrl, arg); 1485 1486 // Check for substr count > string count 1487 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1488 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1489 Node* if_gt = generate_slow_guard(bol, NULL); 1490 if (if_gt != NULL) { 1491 result_phi->init_req(2, intcon(-1)); 1492 result_rgn->init_req(2, if_gt); 1493 } 1494 1495 if (!stopped()) { 1496 // Check for substr count == 0 1497 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1498 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1499 Node* if_zero = generate_slow_guard(bol, NULL); 1500 if (if_zero != NULL) { 1501 result_phi->init_req(3, intcon(0)); 1502 result_rgn->init_req(3, if_zero); 1503 } 1504 } 1505 1506 if (!stopped()) { 1507 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1508 result_phi->init_req(1, result); 1509 result_rgn->init_req(1, control()); 1510 } 1511 set_control(_gvn.transform(result_rgn)); 1512 record_for_igvn(result_rgn); 1513 result = _gvn.transform(result_phi); 1514 1515 } else { // Use LibraryCallKit::string_indexOf 1516 // don't intrinsify if argument isn't a constant string. 1517 if (!arg->is_Con()) { 1518 return false; 1519 } 1520 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1521 if (str_type == NULL) { 1522 return false; 1523 } 1524 ciInstanceKlass* klass = env()->String_klass(); 1525 ciObject* str_const = str_type->const_oop(); 1526 if (str_const == NULL || str_const->klass() != klass) { 1527 return false; 1528 } 1529 ciInstance* str = str_const->as_instance(); 1530 assert(str != NULL, "must be instance"); 1531 1532 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1533 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1534 1535 int o; 1536 int c; 1537 if (java_lang_String::has_offset_field()) { 1538 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1539 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1540 } else { 1541 o = 0; 1542 c = pat->length(); 1543 } 1544 1545 // constant strings have no offset and count == length which 1546 // simplifies the resulting code somewhat so lets optimize for that. 1547 if (o != 0 || c != pat->length()) { 1548 return false; 1549 } 1550 1551 receiver = null_check(receiver, T_OBJECT); 1552 // NOTE: No null check on the argument is needed since it's a constant String oop. 1553 if (stopped()) { 1554 return true; 1555 } 1556 1557 // The null string as a pattern always returns 0 (match at beginning of string) 1558 if (c == 0) { 1559 set_result(intcon(0)); 1560 return true; 1561 } 1562 1563 // Generate default indexOf 1564 jchar lastChar = pat->char_at(o + (c - 1)); 1565 int cache = 0; 1566 int i; 1567 for (i = 0; i < c - 1; i++) { 1568 assert(i < pat->length(), "out of range"); 1569 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1570 } 1571 1572 int md2 = c; 1573 for (i = 0; i < c - 1; i++) { 1574 assert(i < pat->length(), "out of range"); 1575 if (pat->char_at(o + i) == lastChar) { 1576 md2 = (c - 1) - i; 1577 } 1578 } 1579 1580 result = string_indexOf(receiver, pat, o, cache, md2); 1581 } 1582 set_result(result); 1583 return true; 1584 } 1585 1586 //--------------------------round_double_node-------------------------------- 1587 // Round a double node if necessary. 1588 Node* LibraryCallKit::round_double_node(Node* n) { 1589 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1590 n = _gvn.transform(new RoundDoubleNode(0, n)); 1591 return n; 1592 } 1593 1594 //------------------------------inline_math----------------------------------- 1595 // public static double Math.abs(double) 1596 // public static double Math.sqrt(double) 1597 // public static double Math.log(double) 1598 // public static double Math.log10(double) 1599 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1600 Node* arg = round_double_node(argument(0)); 1601 Node* n; 1602 switch (id) { 1603 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1604 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1605 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1606 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1607 default: fatal_unexpected_iid(id); break; 1608 } 1609 set_result(_gvn.transform(n)); 1610 return true; 1611 } 1612 1613 //------------------------------inline_trig---------------------------------- 1614 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1615 // argument reduction which will turn into a fast/slow diamond. 1616 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1617 Node* arg = round_double_node(argument(0)); 1618 Node* n = NULL; 1619 1620 switch (id) { 1621 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1622 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1623 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1624 default: fatal_unexpected_iid(id); break; 1625 } 1626 n = _gvn.transform(n); 1627 1628 // Rounding required? Check for argument reduction! 1629 if (Matcher::strict_fp_requires_explicit_rounding) { 1630 static const double pi_4 = 0.7853981633974483; 1631 static const double neg_pi_4 = -0.7853981633974483; 1632 // pi/2 in 80-bit extended precision 1633 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1634 // -pi/2 in 80-bit extended precision 1635 // 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}; 1636 // Cutoff value for using this argument reduction technique 1637 //static const double pi_2_minus_epsilon = 1.564660403643354; 1638 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1639 1640 // Pseudocode for sin: 1641 // if (x <= Math.PI / 4.0) { 1642 // if (x >= -Math.PI / 4.0) return fsin(x); 1643 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1644 // } else { 1645 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1646 // } 1647 // return StrictMath.sin(x); 1648 1649 // Pseudocode for cos: 1650 // if (x <= Math.PI / 4.0) { 1651 // if (x >= -Math.PI / 4.0) return fcos(x); 1652 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1653 // } else { 1654 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1655 // } 1656 // return StrictMath.cos(x); 1657 1658 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1659 // requires a special machine instruction to load it. Instead we'll try 1660 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1661 // probably do the math inside the SIN encoding. 1662 1663 // Make the merge point 1664 RegionNode* r = new RegionNode(3); 1665 Node* phi = new PhiNode(r, Type::DOUBLE); 1666 1667 // Flatten arg so we need only 1 test 1668 Node *abs = _gvn.transform(new AbsDNode(arg)); 1669 // Node for PI/4 constant 1670 Node *pi4 = makecon(TypeD::make(pi_4)); 1671 // Check PI/4 : abs(arg) 1672 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1673 // Check: If PI/4 < abs(arg) then go slow 1674 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1675 // Branch either way 1676 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1677 set_control(opt_iff(r,iff)); 1678 1679 // Set fast path result 1680 phi->init_req(2, n); 1681 1682 // Slow path - non-blocking leaf call 1683 Node* call = NULL; 1684 switch (id) { 1685 case vmIntrinsics::_dsin: 1686 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1687 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1688 "Sin", NULL, arg, top()); 1689 break; 1690 case vmIntrinsics::_dcos: 1691 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1692 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1693 "Cos", NULL, arg, top()); 1694 break; 1695 case vmIntrinsics::_dtan: 1696 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1697 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1698 "Tan", NULL, arg, top()); 1699 break; 1700 } 1701 assert(control()->in(0) == call, ""); 1702 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1703 r->init_req(1, control()); 1704 phi->init_req(1, slow_result); 1705 1706 // Post-merge 1707 set_control(_gvn.transform(r)); 1708 record_for_igvn(r); 1709 n = _gvn.transform(phi); 1710 1711 C->set_has_split_ifs(true); // Has chance for split-if optimization 1712 } 1713 set_result(n); 1714 return true; 1715 } 1716 1717 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1718 //------------------- 1719 //result=(result.isNaN())? funcAddr():result; 1720 // Check: If isNaN() by checking result!=result? then either trap 1721 // or go to runtime 1722 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1723 // Build the boolean node 1724 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1725 1726 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1727 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1728 // The pow or exp intrinsic returned a NaN, which requires a call 1729 // to the runtime. Recompile with the runtime call. 1730 uncommon_trap(Deoptimization::Reason_intrinsic, 1731 Deoptimization::Action_make_not_entrant); 1732 } 1733 return result; 1734 } else { 1735 // If this inlining ever returned NaN in the past, we compile a call 1736 // to the runtime to properly handle corner cases 1737 1738 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1739 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1740 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1741 1742 if (!if_slow->is_top()) { 1743 RegionNode* result_region = new RegionNode(3); 1744 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1745 1746 result_region->init_req(1, if_fast); 1747 result_val->init_req(1, result); 1748 1749 set_control(if_slow); 1750 1751 const TypePtr* no_memory_effects = NULL; 1752 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1753 no_memory_effects, 1754 x, top(), y, y ? top() : NULL); 1755 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1756 #ifdef ASSERT 1757 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1758 assert(value_top == top(), "second value must be top"); 1759 #endif 1760 1761 result_region->init_req(2, control()); 1762 result_val->init_req(2, value); 1763 set_control(_gvn.transform(result_region)); 1764 return _gvn.transform(result_val); 1765 } else { 1766 return result; 1767 } 1768 } 1769 } 1770 1771 //------------------------------inline_exp------------------------------------- 1772 // Inline exp instructions, if possible. The Intel hardware only misses 1773 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1774 bool LibraryCallKit::inline_exp() { 1775 Node* arg = round_double_node(argument(0)); 1776 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1777 1778 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1779 set_result(n); 1780 1781 C->set_has_split_ifs(true); // Has chance for split-if optimization 1782 return true; 1783 } 1784 1785 //------------------------------inline_pow------------------------------------- 1786 // Inline power instructions, if possible. 1787 bool LibraryCallKit::inline_pow() { 1788 // Pseudocode for pow 1789 // if (y == 2) { 1790 // return x * x; 1791 // } else { 1792 // if (x <= 0.0) { 1793 // long longy = (long)y; 1794 // if ((double)longy == y) { // if y is long 1795 // if (y + 1 == y) longy = 0; // huge number: even 1796 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1797 // } else { 1798 // result = NaN; 1799 // } 1800 // } else { 1801 // result = DPow(x,y); 1802 // } 1803 // if (result != result)? { 1804 // result = uncommon_trap() or runtime_call(); 1805 // } 1806 // return result; 1807 // } 1808 1809 Node* x = round_double_node(argument(0)); 1810 Node* y = round_double_node(argument(2)); 1811 1812 Node* result = NULL; 1813 1814 Node* const_two_node = makecon(TypeD::make(2.0)); 1815 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1816 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1817 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1818 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1819 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1820 1821 RegionNode* region_node = new RegionNode(3); 1822 region_node->init_req(1, if_true); 1823 1824 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1825 // special case for x^y where y == 2, we can convert it to x * x 1826 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1827 1828 // set control to if_false since we will now process the false branch 1829 set_control(if_false); 1830 1831 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1832 // Short form: skip the fancy tests and just check for NaN result. 1833 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1834 } else { 1835 // If this inlining ever returned NaN in the past, include all 1836 // checks + call to the runtime. 1837 1838 // Set the merge point for If node with condition of (x <= 0.0) 1839 // There are four possible paths to region node and phi node 1840 RegionNode *r = new RegionNode(4); 1841 Node *phi = new PhiNode(r, Type::DOUBLE); 1842 1843 // Build the first if node: if (x <= 0.0) 1844 // Node for 0 constant 1845 Node *zeronode = makecon(TypeD::ZERO); 1846 // Check x:0 1847 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1848 // Check: If (x<=0) then go complex path 1849 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1850 // Branch either way 1851 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1852 // Fast path taken; set region slot 3 1853 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1854 r->init_req(3,fast_taken); // Capture fast-control 1855 1856 // Fast path not-taken, i.e. slow path 1857 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1858 1859 // Set fast path result 1860 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1861 phi->init_req(3, fast_result); 1862 1863 // Complex path 1864 // Build the second if node (if y is long) 1865 // Node for (long)y 1866 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1867 // Node for (double)((long) y) 1868 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1869 // Check (double)((long) y) : y 1870 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1871 // Check if (y isn't long) then go to slow path 1872 1873 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1874 // Branch either way 1875 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1876 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1877 1878 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1879 1880 // Calculate DPow(abs(x), y)*(1 & (long)y) 1881 // Node for constant 1 1882 Node *conone = longcon(1); 1883 // 1& (long)y 1884 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1885 1886 // A huge number is always even. Detect a huge number by checking 1887 // if y + 1 == y and set integer to be tested for parity to 0. 1888 // Required for corner case: 1889 // (long)9.223372036854776E18 = max_jlong 1890 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1891 // max_jlong is odd but 9.223372036854776E18 is even 1892 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1893 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1894 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1895 Node* correctedsign = NULL; 1896 if (ConditionalMoveLimit != 0) { 1897 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1898 } else { 1899 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1900 RegionNode *r = new RegionNode(3); 1901 Node *phi = new PhiNode(r, TypeLong::LONG); 1902 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1903 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1904 phi->init_req(1, signnode); 1905 phi->init_req(2, longcon(0)); 1906 correctedsign = _gvn.transform(phi); 1907 ylong_path = _gvn.transform(r); 1908 record_for_igvn(r); 1909 } 1910 1911 // zero node 1912 Node *conzero = longcon(0); 1913 // Check (1&(long)y)==0? 1914 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1915 // Check if (1&(long)y)!=0?, if so the result is negative 1916 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1917 // abs(x) 1918 Node *absx=_gvn.transform(new AbsDNode(x)); 1919 // abs(x)^y 1920 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1921 // -abs(x)^y 1922 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1923 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1924 Node *signresult = NULL; 1925 if (ConditionalMoveLimit != 0) { 1926 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1927 } else { 1928 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1929 RegionNode *r = new RegionNode(3); 1930 Node *phi = new PhiNode(r, Type::DOUBLE); 1931 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1932 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1933 phi->init_req(1, absxpowy); 1934 phi->init_req(2, negabsxpowy); 1935 signresult = _gvn.transform(phi); 1936 ylong_path = _gvn.transform(r); 1937 record_for_igvn(r); 1938 } 1939 // Set complex path fast result 1940 r->init_req(2, ylong_path); 1941 phi->init_req(2, signresult); 1942 1943 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1944 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1945 r->init_req(1,slow_path); 1946 phi->init_req(1,slow_result); 1947 1948 // Post merge 1949 set_control(_gvn.transform(r)); 1950 record_for_igvn(r); 1951 result = _gvn.transform(phi); 1952 } 1953 1954 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1955 1956 // control from finish_pow_exp is now input to the region node 1957 region_node->set_req(2, control()); 1958 // the result from finish_pow_exp is now input to the phi node 1959 phi_node->init_req(2, result); 1960 set_control(_gvn.transform(region_node)); 1961 record_for_igvn(region_node); 1962 set_result(_gvn.transform(phi_node)); 1963 1964 C->set_has_split_ifs(true); // Has chance for split-if optimization 1965 return true; 1966 } 1967 1968 //------------------------------runtime_math----------------------------- 1969 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1970 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1971 "must be (DD)D or (D)D type"); 1972 1973 // Inputs 1974 Node* a = round_double_node(argument(0)); 1975 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1976 1977 const TypePtr* no_memory_effects = NULL; 1978 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1979 no_memory_effects, 1980 a, top(), b, b ? top() : NULL); 1981 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1982 #ifdef ASSERT 1983 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1984 assert(value_top == top(), "second value must be top"); 1985 #endif 1986 1987 set_result(value); 1988 return true; 1989 } 1990 1991 //------------------------------inline_math_native----------------------------- 1992 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1993 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1994 switch (id) { 1995 // These intrinsics are not properly supported on all hardware 1996 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1997 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1998 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1999 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2000 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2001 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2002 2003 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2004 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2005 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2006 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2007 2008 // These intrinsics are supported on all hardware 2009 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2010 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2011 2012 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2013 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2014 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2015 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2016 #undef FN_PTR 2017 2018 // These intrinsics are not yet correctly implemented 2019 case vmIntrinsics::_datan2: 2020 return false; 2021 2022 default: 2023 fatal_unexpected_iid(id); 2024 return false; 2025 } 2026 } 2027 2028 static bool is_simple_name(Node* n) { 2029 return (n->req() == 1 // constant 2030 || (n->is_Type() && n->as_Type()->type()->singleton()) 2031 || n->is_Proj() // parameter or return value 2032 || n->is_Phi() // local of some sort 2033 ); 2034 } 2035 2036 //----------------------------inline_min_max----------------------------------- 2037 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2038 set_result(generate_min_max(id, argument(0), argument(1))); 2039 return true; 2040 } 2041 2042 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2043 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2044 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2045 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2046 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2047 2048 { 2049 PreserveJVMState pjvms(this); 2050 PreserveReexecuteState preexecs(this); 2051 jvms()->set_should_reexecute(true); 2052 2053 set_control(slow_path); 2054 set_i_o(i_o()); 2055 2056 uncommon_trap(Deoptimization::Reason_intrinsic, 2057 Deoptimization::Action_none); 2058 } 2059 2060 set_control(fast_path); 2061 set_result(math); 2062 } 2063 2064 template <typename OverflowOp> 2065 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2066 typedef typename OverflowOp::MathOp MathOp; 2067 2068 MathOp* mathOp = new MathOp(arg1, arg2); 2069 Node* operation = _gvn.transform( mathOp ); 2070 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2071 inline_math_mathExact(operation, ofcheck); 2072 return true; 2073 } 2074 2075 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2076 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2077 } 2078 2079 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2080 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2081 } 2082 2083 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2084 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2085 } 2086 2087 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2088 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2089 } 2090 2091 bool LibraryCallKit::inline_math_negateExactI() { 2092 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2093 } 2094 2095 bool LibraryCallKit::inline_math_negateExactL() { 2096 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2097 } 2098 2099 bool LibraryCallKit::inline_math_multiplyExactI() { 2100 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2101 } 2102 2103 bool LibraryCallKit::inline_math_multiplyExactL() { 2104 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2105 } 2106 2107 Node* 2108 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2109 // These are the candidate return value: 2110 Node* xvalue = x0; 2111 Node* yvalue = y0; 2112 2113 if (xvalue == yvalue) { 2114 return xvalue; 2115 } 2116 2117 bool want_max = (id == vmIntrinsics::_max); 2118 2119 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2120 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2121 if (txvalue == NULL || tyvalue == NULL) return top(); 2122 // This is not really necessary, but it is consistent with a 2123 // hypothetical MaxINode::Value method: 2124 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2125 2126 // %%% This folding logic should (ideally) be in a different place. 2127 // Some should be inside IfNode, and there to be a more reliable 2128 // transformation of ?: style patterns into cmoves. We also want 2129 // more powerful optimizations around cmove and min/max. 2130 2131 // Try to find a dominating comparison of these guys. 2132 // It can simplify the index computation for Arrays.copyOf 2133 // and similar uses of System.arraycopy. 2134 // First, compute the normalized version of CmpI(x, y). 2135 int cmp_op = Op_CmpI; 2136 Node* xkey = xvalue; 2137 Node* ykey = yvalue; 2138 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2139 if (ideal_cmpxy->is_Cmp()) { 2140 // E.g., if we have CmpI(length - offset, count), 2141 // it might idealize to CmpI(length, count + offset) 2142 cmp_op = ideal_cmpxy->Opcode(); 2143 xkey = ideal_cmpxy->in(1); 2144 ykey = ideal_cmpxy->in(2); 2145 } 2146 2147 // Start by locating any relevant comparisons. 2148 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2149 Node* cmpxy = NULL; 2150 Node* cmpyx = NULL; 2151 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2152 Node* cmp = start_from->fast_out(k); 2153 if (cmp->outcnt() > 0 && // must have prior uses 2154 cmp->in(0) == NULL && // must be context-independent 2155 cmp->Opcode() == cmp_op) { // right kind of compare 2156 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2157 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2158 } 2159 } 2160 2161 const int NCMPS = 2; 2162 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2163 int cmpn; 2164 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2165 if (cmps[cmpn] != NULL) break; // find a result 2166 } 2167 if (cmpn < NCMPS) { 2168 // Look for a dominating test that tells us the min and max. 2169 int depth = 0; // Limit search depth for speed 2170 Node* dom = control(); 2171 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2172 if (++depth >= 100) break; 2173 Node* ifproj = dom; 2174 if (!ifproj->is_Proj()) continue; 2175 Node* iff = ifproj->in(0); 2176 if (!iff->is_If()) continue; 2177 Node* bol = iff->in(1); 2178 if (!bol->is_Bool()) continue; 2179 Node* cmp = bol->in(1); 2180 if (cmp == NULL) continue; 2181 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2182 if (cmps[cmpn] == cmp) break; 2183 if (cmpn == NCMPS) continue; 2184 BoolTest::mask btest = bol->as_Bool()->_test._test; 2185 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2186 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2187 // At this point, we know that 'x btest y' is true. 2188 switch (btest) { 2189 case BoolTest::eq: 2190 // They are proven equal, so we can collapse the min/max. 2191 // Either value is the answer. Choose the simpler. 2192 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2193 return yvalue; 2194 return xvalue; 2195 case BoolTest::lt: // x < y 2196 case BoolTest::le: // x <= y 2197 return (want_max ? yvalue : xvalue); 2198 case BoolTest::gt: // x > y 2199 case BoolTest::ge: // x >= y 2200 return (want_max ? xvalue : yvalue); 2201 } 2202 } 2203 } 2204 2205 // We failed to find a dominating test. 2206 // Let's pick a test that might GVN with prior tests. 2207 Node* best_bol = NULL; 2208 BoolTest::mask best_btest = BoolTest::illegal; 2209 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2210 Node* cmp = cmps[cmpn]; 2211 if (cmp == NULL) continue; 2212 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2213 Node* bol = cmp->fast_out(j); 2214 if (!bol->is_Bool()) continue; 2215 BoolTest::mask btest = bol->as_Bool()->_test._test; 2216 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2217 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2218 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2219 best_bol = bol->as_Bool(); 2220 best_btest = btest; 2221 } 2222 } 2223 } 2224 2225 Node* answer_if_true = NULL; 2226 Node* answer_if_false = NULL; 2227 switch (best_btest) { 2228 default: 2229 if (cmpxy == NULL) 2230 cmpxy = ideal_cmpxy; 2231 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2232 // and fall through: 2233 case BoolTest::lt: // x < y 2234 case BoolTest::le: // x <= y 2235 answer_if_true = (want_max ? yvalue : xvalue); 2236 answer_if_false = (want_max ? xvalue : yvalue); 2237 break; 2238 case BoolTest::gt: // x > y 2239 case BoolTest::ge: // x >= y 2240 answer_if_true = (want_max ? xvalue : yvalue); 2241 answer_if_false = (want_max ? yvalue : xvalue); 2242 break; 2243 } 2244 2245 jint hi, lo; 2246 if (want_max) { 2247 // We can sharpen the minimum. 2248 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2249 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2250 } else { 2251 // We can sharpen the maximum. 2252 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2253 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2254 } 2255 2256 // Use a flow-free graph structure, to avoid creating excess control edges 2257 // which could hinder other optimizations. 2258 // Since Math.min/max is often used with arraycopy, we want 2259 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2260 Node* cmov = CMoveNode::make(NULL, best_bol, 2261 answer_if_false, answer_if_true, 2262 TypeInt::make(lo, hi, widen)); 2263 2264 return _gvn.transform(cmov); 2265 2266 /* 2267 // This is not as desirable as it may seem, since Min and Max 2268 // nodes do not have a full set of optimizations. 2269 // And they would interfere, anyway, with 'if' optimizations 2270 // and with CMoveI canonical forms. 2271 switch (id) { 2272 case vmIntrinsics::_min: 2273 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2274 case vmIntrinsics::_max: 2275 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2276 default: 2277 ShouldNotReachHere(); 2278 } 2279 */ 2280 } 2281 2282 inline int 2283 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2284 const TypePtr* base_type = TypePtr::NULL_PTR; 2285 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2286 if (base_type == NULL) { 2287 // Unknown type. 2288 return Type::AnyPtr; 2289 } else if (base_type == TypePtr::NULL_PTR) { 2290 // Since this is a NULL+long form, we have to switch to a rawptr. 2291 base = _gvn.transform(new CastX2PNode(offset)); 2292 offset = MakeConX(0); 2293 return Type::RawPtr; 2294 } else if (base_type->base() == Type::RawPtr) { 2295 return Type::RawPtr; 2296 } else if (base_type->isa_oopptr()) { 2297 // Base is never null => always a heap address. 2298 if (base_type->ptr() == TypePtr::NotNull) { 2299 return Type::OopPtr; 2300 } 2301 // Offset is small => always a heap address. 2302 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2303 if (offset_type != NULL && 2304 base_type->offset() == 0 && // (should always be?) 2305 offset_type->_lo >= 0 && 2306 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2307 return Type::OopPtr; 2308 } 2309 // Otherwise, it might either be oop+off or NULL+addr. 2310 return Type::AnyPtr; 2311 } else { 2312 // No information: 2313 return Type::AnyPtr; 2314 } 2315 } 2316 2317 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2318 int kind = classify_unsafe_addr(base, offset); 2319 if (kind == Type::RawPtr) { 2320 return basic_plus_adr(top(), base, offset); 2321 } else { 2322 return basic_plus_adr(base, offset); 2323 } 2324 } 2325 2326 //--------------------------inline_number_methods----------------------------- 2327 // inline int Integer.numberOfLeadingZeros(int) 2328 // inline int Long.numberOfLeadingZeros(long) 2329 // 2330 // inline int Integer.numberOfTrailingZeros(int) 2331 // inline int Long.numberOfTrailingZeros(long) 2332 // 2333 // inline int Integer.bitCount(int) 2334 // inline int Long.bitCount(long) 2335 // 2336 // inline char Character.reverseBytes(char) 2337 // inline short Short.reverseBytes(short) 2338 // inline int Integer.reverseBytes(int) 2339 // inline long Long.reverseBytes(long) 2340 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2341 Node* arg = argument(0); 2342 Node* n; 2343 switch (id) { 2344 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2345 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2346 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2347 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2348 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2349 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2350 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2351 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2352 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2353 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2354 default: fatal_unexpected_iid(id); break; 2355 } 2356 set_result(_gvn.transform(n)); 2357 return true; 2358 } 2359 2360 //----------------------------inline_unsafe_access---------------------------- 2361 2362 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2363 2364 // Helper that guards and inserts a pre-barrier. 2365 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2366 Node* pre_val, bool need_mem_bar) { 2367 // We could be accessing the referent field of a reference object. If so, when G1 2368 // is enabled, we need to log the value in the referent field in an SATB buffer. 2369 // This routine performs some compile time filters and generates suitable 2370 // runtime filters that guard the pre-barrier code. 2371 // Also add memory barrier for non volatile load from the referent field 2372 // to prevent commoning of loads across safepoint. 2373 if (!UseG1GC && !need_mem_bar) 2374 return; 2375 2376 // Some compile time checks. 2377 2378 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2379 const TypeX* otype = offset->find_intptr_t_type(); 2380 if (otype != NULL && otype->is_con() && 2381 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2382 // Constant offset but not the reference_offset so just return 2383 return; 2384 } 2385 2386 // We only need to generate the runtime guards for instances. 2387 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2388 if (btype != NULL) { 2389 if (btype->isa_aryptr()) { 2390 // Array type so nothing to do 2391 return; 2392 } 2393 2394 const TypeInstPtr* itype = btype->isa_instptr(); 2395 if (itype != NULL) { 2396 // Can the klass of base_oop be statically determined to be 2397 // _not_ a sub-class of Reference and _not_ Object? 2398 ciKlass* klass = itype->klass(); 2399 if ( klass->is_loaded() && 2400 !klass->is_subtype_of(env()->Reference_klass()) && 2401 !env()->Object_klass()->is_subtype_of(klass)) { 2402 return; 2403 } 2404 } 2405 } 2406 2407 // The compile time filters did not reject base_oop/offset so 2408 // we need to generate the following runtime filters 2409 // 2410 // if (offset == java_lang_ref_Reference::_reference_offset) { 2411 // if (instance_of(base, java.lang.ref.Reference)) { 2412 // pre_barrier(_, pre_val, ...); 2413 // } 2414 // } 2415 2416 float likely = PROB_LIKELY( 0.999); 2417 float unlikely = PROB_UNLIKELY(0.999); 2418 2419 IdealKit ideal(this); 2420 #define __ ideal. 2421 2422 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2423 2424 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2425 // Update graphKit memory and control from IdealKit. 2426 sync_kit(ideal); 2427 2428 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2429 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2430 2431 // Update IdealKit memory and control from graphKit. 2432 __ sync_kit(this); 2433 2434 Node* one = __ ConI(1); 2435 // is_instof == 0 if base_oop == NULL 2436 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2437 2438 // Update graphKit from IdeakKit. 2439 sync_kit(ideal); 2440 2441 // Use the pre-barrier to record the value in the referent field 2442 pre_barrier(false /* do_load */, 2443 __ ctrl(), 2444 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2445 pre_val /* pre_val */, 2446 T_OBJECT); 2447 if (need_mem_bar) { 2448 // Add memory barrier to prevent commoning reads from this field 2449 // across safepoint since GC can change its value. 2450 insert_mem_bar(Op_MemBarCPUOrder); 2451 } 2452 // Update IdealKit from graphKit. 2453 __ sync_kit(this); 2454 2455 } __ end_if(); // _ref_type != ref_none 2456 } __ end_if(); // offset == referent_offset 2457 2458 // Final sync IdealKit and GraphKit. 2459 final_sync(ideal); 2460 #undef __ 2461 } 2462 2463 2464 // Interpret Unsafe.fieldOffset cookies correctly: 2465 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2466 2467 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2468 // Attempt to infer a sharper value type from the offset and base type. 2469 ciKlass* sharpened_klass = NULL; 2470 2471 // See if it is an instance field, with an object type. 2472 if (alias_type->field() != NULL) { 2473 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2474 if (alias_type->field()->type()->is_klass()) { 2475 sharpened_klass = alias_type->field()->type()->as_klass(); 2476 } 2477 } 2478 2479 // See if it is a narrow oop array. 2480 if (adr_type->isa_aryptr()) { 2481 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2482 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2483 if (elem_type != NULL) { 2484 sharpened_klass = elem_type->klass(); 2485 } 2486 } 2487 } 2488 2489 // The sharpened class might be unloaded if there is no class loader 2490 // contraint in place. 2491 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2492 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2493 2494 #ifndef PRODUCT 2495 if (C->print_intrinsics() || C->print_inlining()) { 2496 tty->print(" from base type: "); adr_type->dump(); 2497 tty->print(" sharpened value: "); tjp->dump(); 2498 } 2499 #endif 2500 // Sharpen the value type. 2501 return tjp; 2502 } 2503 return NULL; 2504 } 2505 2506 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2507 if (callee()->is_static()) return false; // caller must have the capability! 2508 2509 #ifndef PRODUCT 2510 { 2511 ResourceMark rm; 2512 // Check the signatures. 2513 ciSignature* sig = callee()->signature(); 2514 #ifdef ASSERT 2515 if (!is_store) { 2516 // Object getObject(Object base, int/long offset), etc. 2517 BasicType rtype = sig->return_type()->basic_type(); 2518 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2519 rtype = T_ADDRESS; // it is really a C void* 2520 assert(rtype == type, "getter must return the expected value"); 2521 if (!is_native_ptr) { 2522 assert(sig->count() == 2, "oop getter has 2 arguments"); 2523 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2524 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2525 } else { 2526 assert(sig->count() == 1, "native getter has 1 argument"); 2527 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2528 } 2529 } else { 2530 // void putObject(Object base, int/long offset, Object x), etc. 2531 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2532 if (!is_native_ptr) { 2533 assert(sig->count() == 3, "oop putter has 3 arguments"); 2534 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2535 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2536 } else { 2537 assert(sig->count() == 2, "native putter has 2 arguments"); 2538 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2539 } 2540 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2541 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2542 vtype = T_ADDRESS; // it is really a C void* 2543 assert(vtype == type, "putter must accept the expected value"); 2544 } 2545 #endif // ASSERT 2546 } 2547 #endif //PRODUCT 2548 2549 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2550 2551 Node* receiver = argument(0); // type: oop 2552 2553 // Build address expression. 2554 Node* adr; 2555 Node* heap_base_oop = top(); 2556 Node* offset = top(); 2557 Node* val; 2558 2559 if (!is_native_ptr) { 2560 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2561 Node* base = argument(1); // type: oop 2562 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2563 offset = argument(2); // type: long 2564 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2565 // to be plain byte offsets, which are also the same as those accepted 2566 // by oopDesc::field_base. 2567 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2568 "fieldOffset must be byte-scaled"); 2569 // 32-bit machines ignore the high half! 2570 offset = ConvL2X(offset); 2571 adr = make_unsafe_address(base, offset); 2572 heap_base_oop = base; 2573 val = is_store ? argument(4) : NULL; 2574 } else { 2575 Node* ptr = argument(1); // type: long 2576 ptr = ConvL2X(ptr); // adjust Java long to machine word 2577 adr = make_unsafe_address(NULL, ptr); 2578 val = is_store ? argument(3) : NULL; 2579 } 2580 2581 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2582 2583 // First guess at the value type. 2584 const Type *value_type = Type::get_const_basic_type(type); 2585 2586 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2587 // there was not enough information to nail it down. 2588 Compile::AliasType* alias_type = C->alias_type(adr_type); 2589 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2590 2591 // We will need memory barriers unless we can determine a unique 2592 // alias category for this reference. (Note: If for some reason 2593 // the barriers get omitted and the unsafe reference begins to "pollute" 2594 // the alias analysis of the rest of the graph, either Compile::can_alias 2595 // or Compile::must_alias will throw a diagnostic assert.) 2596 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2597 2598 // If we are reading the value of the referent field of a Reference 2599 // object (either by using Unsafe directly or through reflection) 2600 // then, if G1 is enabled, we need to record the referent in an 2601 // SATB log buffer using the pre-barrier mechanism. 2602 // Also we need to add memory barrier to prevent commoning reads 2603 // from this field across safepoint since GC can change its value. 2604 bool need_read_barrier = !is_native_ptr && !is_store && 2605 offset != top() && heap_base_oop != top(); 2606 2607 if (!is_store && type == T_OBJECT) { 2608 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2609 if (tjp != NULL) { 2610 value_type = tjp; 2611 } 2612 } 2613 2614 receiver = null_check(receiver); 2615 if (stopped()) { 2616 return true; 2617 } 2618 // Heap pointers get a null-check from the interpreter, 2619 // as a courtesy. However, this is not guaranteed by Unsafe, 2620 // and it is not possible to fully distinguish unintended nulls 2621 // from intended ones in this API. 2622 2623 if (is_volatile) { 2624 // We need to emit leading and trailing CPU membars (see below) in 2625 // addition to memory membars when is_volatile. This is a little 2626 // too strong, but avoids the need to insert per-alias-type 2627 // volatile membars (for stores; compare Parse::do_put_xxx), which 2628 // we cannot do effectively here because we probably only have a 2629 // rough approximation of type. 2630 need_mem_bar = true; 2631 // For Stores, place a memory ordering barrier now. 2632 if (is_store) { 2633 insert_mem_bar(Op_MemBarRelease); 2634 } else { 2635 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2636 insert_mem_bar(Op_MemBarVolatile); 2637 } 2638 } 2639 } 2640 2641 // Memory barrier to prevent normal and 'unsafe' accesses from 2642 // bypassing each other. Happens after null checks, so the 2643 // exception paths do not take memory state from the memory barrier, 2644 // so there's no problems making a strong assert about mixing users 2645 // of safe & unsafe memory. 2646 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2647 2648 if (!is_store) { 2649 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2650 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile); 2651 // load value 2652 switch (type) { 2653 case T_BOOLEAN: 2654 case T_CHAR: 2655 case T_BYTE: 2656 case T_SHORT: 2657 case T_INT: 2658 case T_LONG: 2659 case T_FLOAT: 2660 case T_DOUBLE: 2661 break; 2662 case T_OBJECT: 2663 if (need_read_barrier) { 2664 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2665 } 2666 break; 2667 case T_ADDRESS: 2668 // Cast to an int type. 2669 p = _gvn.transform(new CastP2XNode(NULL, p)); 2670 p = ConvX2UL(p); 2671 break; 2672 default: 2673 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2674 break; 2675 } 2676 // The load node has the control of the preceding MemBarCPUOrder. All 2677 // following nodes will have the control of the MemBarCPUOrder inserted at 2678 // the end of this method. So, pushing the load onto the stack at a later 2679 // point is fine. 2680 set_result(p); 2681 } else { 2682 // place effect of store into memory 2683 switch (type) { 2684 case T_DOUBLE: 2685 val = dstore_rounding(val); 2686 break; 2687 case T_ADDRESS: 2688 // Repackage the long as a pointer. 2689 val = ConvL2X(val); 2690 val = _gvn.transform(new CastX2PNode(val)); 2691 break; 2692 } 2693 2694 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2695 if (type != T_OBJECT ) { 2696 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2697 } else { 2698 // Possibly an oop being stored to Java heap or native memory 2699 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2700 // oop to Java heap. 2701 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2702 } else { 2703 // We can't tell at compile time if we are storing in the Java heap or outside 2704 // of it. So we need to emit code to conditionally do the proper type of 2705 // store. 2706 2707 IdealKit ideal(this); 2708 #define __ ideal. 2709 // QQQ who knows what probability is here?? 2710 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2711 // Sync IdealKit and graphKit. 2712 sync_kit(ideal); 2713 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2714 // Update IdealKit memory. 2715 __ sync_kit(this); 2716 } __ else_(); { 2717 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2718 } __ end_if(); 2719 // Final sync IdealKit and GraphKit. 2720 final_sync(ideal); 2721 #undef __ 2722 } 2723 } 2724 } 2725 2726 if (is_volatile) { 2727 if (!is_store) { 2728 insert_mem_bar(Op_MemBarAcquire); 2729 } else { 2730 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2731 insert_mem_bar(Op_MemBarVolatile); 2732 } 2733 } 2734 } 2735 2736 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2737 2738 return true; 2739 } 2740 2741 //----------------------------inline_unsafe_load_store---------------------------- 2742 // This method serves a couple of different customers (depending on LoadStoreKind): 2743 // 2744 // LS_cmpxchg: 2745 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2746 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2747 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2748 // 2749 // LS_xadd: 2750 // public int getAndAddInt( Object o, long offset, int delta) 2751 // public long getAndAddLong(Object o, long offset, long delta) 2752 // 2753 // LS_xchg: 2754 // int getAndSet(Object o, long offset, int newValue) 2755 // long getAndSet(Object o, long offset, long newValue) 2756 // Object getAndSet(Object o, long offset, Object newValue) 2757 // 2758 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2759 // This basic scheme here is the same as inline_unsafe_access, but 2760 // differs in enough details that combining them would make the code 2761 // overly confusing. (This is a true fact! I originally combined 2762 // them, but even I was confused by it!) As much code/comments as 2763 // possible are retained from inline_unsafe_access though to make 2764 // the correspondences clearer. - dl 2765 2766 if (callee()->is_static()) return false; // caller must have the capability! 2767 2768 #ifndef PRODUCT 2769 BasicType rtype; 2770 { 2771 ResourceMark rm; 2772 // Check the signatures. 2773 ciSignature* sig = callee()->signature(); 2774 rtype = sig->return_type()->basic_type(); 2775 if (kind == LS_xadd || kind == LS_xchg) { 2776 // Check the signatures. 2777 #ifdef ASSERT 2778 assert(rtype == type, "get and set must return the expected type"); 2779 assert(sig->count() == 3, "get and set has 3 arguments"); 2780 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2781 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2782 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2783 #endif // ASSERT 2784 } else if (kind == LS_cmpxchg) { 2785 // Check the signatures. 2786 #ifdef ASSERT 2787 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2788 assert(sig->count() == 4, "CAS has 4 arguments"); 2789 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2790 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2791 #endif // ASSERT 2792 } else { 2793 ShouldNotReachHere(); 2794 } 2795 } 2796 #endif //PRODUCT 2797 2798 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2799 2800 // Get arguments: 2801 Node* receiver = NULL; 2802 Node* base = NULL; 2803 Node* offset = NULL; 2804 Node* oldval = NULL; 2805 Node* newval = NULL; 2806 if (kind == LS_cmpxchg) { 2807 const bool two_slot_type = type2size[type] == 2; 2808 receiver = argument(0); // type: oop 2809 base = argument(1); // type: oop 2810 offset = argument(2); // type: long 2811 oldval = argument(4); // type: oop, int, or long 2812 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2813 } else if (kind == LS_xadd || kind == LS_xchg){ 2814 receiver = argument(0); // type: oop 2815 base = argument(1); // type: oop 2816 offset = argument(2); // type: long 2817 oldval = NULL; 2818 newval = argument(4); // type: oop, int, or long 2819 } 2820 2821 // Null check receiver. 2822 receiver = null_check(receiver); 2823 if (stopped()) { 2824 return true; 2825 } 2826 2827 // Build field offset expression. 2828 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2829 // to be plain byte offsets, which are also the same as those accepted 2830 // by oopDesc::field_base. 2831 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2832 // 32-bit machines ignore the high half of long offsets 2833 offset = ConvL2X(offset); 2834 Node* adr = make_unsafe_address(base, offset); 2835 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2836 2837 // For CAS, unlike inline_unsafe_access, there seems no point in 2838 // trying to refine types. Just use the coarse types here. 2839 const Type *value_type = Type::get_const_basic_type(type); 2840 Compile::AliasType* alias_type = C->alias_type(adr_type); 2841 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2842 2843 if (kind == LS_xchg && type == T_OBJECT) { 2844 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2845 if (tjp != NULL) { 2846 value_type = tjp; 2847 } 2848 } 2849 2850 int alias_idx = C->get_alias_index(adr_type); 2851 2852 // Memory-model-wise, a LoadStore acts like a little synchronized 2853 // block, so needs barriers on each side. These don't translate 2854 // into actual barriers on most machines, but we still need rest of 2855 // compiler to respect ordering. 2856 2857 insert_mem_bar(Op_MemBarRelease); 2858 insert_mem_bar(Op_MemBarCPUOrder); 2859 2860 // 4984716: MemBars must be inserted before this 2861 // memory node in order to avoid a false 2862 // dependency which will confuse the scheduler. 2863 Node *mem = memory(alias_idx); 2864 2865 // For now, we handle only those cases that actually exist: ints, 2866 // longs, and Object. Adding others should be straightforward. 2867 Node* load_store; 2868 switch(type) { 2869 case T_INT: 2870 if (kind == LS_xadd) { 2871 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2872 } else if (kind == LS_xchg) { 2873 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2874 } else if (kind == LS_cmpxchg) { 2875 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2876 } else { 2877 ShouldNotReachHere(); 2878 } 2879 break; 2880 case T_LONG: 2881 if (kind == LS_xadd) { 2882 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2883 } else if (kind == LS_xchg) { 2884 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2885 } else if (kind == LS_cmpxchg) { 2886 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2887 } else { 2888 ShouldNotReachHere(); 2889 } 2890 break; 2891 case T_OBJECT: 2892 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2893 // could be delayed during Parse (for example, in adjust_map_after_if()). 2894 // Execute transformation here to avoid barrier generation in such case. 2895 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2896 newval = _gvn.makecon(TypePtr::NULL_PTR); 2897 2898 // Reference stores need a store barrier. 2899 if (kind == LS_xchg) { 2900 // If pre-barrier must execute before the oop store, old value will require do_load here. 2901 if (!can_move_pre_barrier()) { 2902 pre_barrier(true /* do_load*/, 2903 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2904 NULL /* pre_val*/, 2905 T_OBJECT); 2906 } // Else move pre_barrier to use load_store value, see below. 2907 } else if (kind == LS_cmpxchg) { 2908 // Same as for newval above: 2909 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2910 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2911 } 2912 // The only known value which might get overwritten is oldval. 2913 pre_barrier(false /* do_load */, 2914 control(), NULL, NULL, max_juint, NULL, NULL, 2915 oldval /* pre_val */, 2916 T_OBJECT); 2917 } else { 2918 ShouldNotReachHere(); 2919 } 2920 2921 #ifdef _LP64 2922 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2923 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2924 if (kind == LS_xchg) { 2925 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2926 newval_enc, adr_type, value_type->make_narrowoop())); 2927 } else { 2928 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2929 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2930 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2931 newval_enc, oldval_enc)); 2932 } 2933 } else 2934 #endif 2935 { 2936 if (kind == LS_xchg) { 2937 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2938 } else { 2939 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2940 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2941 } 2942 } 2943 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2944 break; 2945 default: 2946 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2947 break; 2948 } 2949 2950 // SCMemProjNodes represent the memory state of a LoadStore. Their 2951 // main role is to prevent LoadStore nodes from being optimized away 2952 // when their results aren't used. 2953 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2954 set_memory(proj, alias_idx); 2955 2956 if (type == T_OBJECT && kind == LS_xchg) { 2957 #ifdef _LP64 2958 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2959 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2960 } 2961 #endif 2962 if (can_move_pre_barrier()) { 2963 // Don't need to load pre_val. The old value is returned by load_store. 2964 // The pre_barrier can execute after the xchg as long as no safepoint 2965 // gets inserted between them. 2966 pre_barrier(false /* do_load */, 2967 control(), NULL, NULL, max_juint, NULL, NULL, 2968 load_store /* pre_val */, 2969 T_OBJECT); 2970 } 2971 } 2972 2973 // Add the trailing membar surrounding the access 2974 insert_mem_bar(Op_MemBarCPUOrder); 2975 insert_mem_bar(Op_MemBarAcquire); 2976 2977 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2978 set_result(load_store); 2979 return true; 2980 } 2981 2982 //----------------------------inline_unsafe_ordered_store---------------------- 2983 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2984 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 2985 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 2986 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2987 // This is another variant of inline_unsafe_access, differing in 2988 // that it always issues store-store ("release") barrier and ensures 2989 // store-atomicity (which only matters for "long"). 2990 2991 if (callee()->is_static()) return false; // caller must have the capability! 2992 2993 #ifndef PRODUCT 2994 { 2995 ResourceMark rm; 2996 // Check the signatures. 2997 ciSignature* sig = callee()->signature(); 2998 #ifdef ASSERT 2999 BasicType rtype = sig->return_type()->basic_type(); 3000 assert(rtype == T_VOID, "must return void"); 3001 assert(sig->count() == 3, "has 3 arguments"); 3002 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3003 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3004 #endif // ASSERT 3005 } 3006 #endif //PRODUCT 3007 3008 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3009 3010 // Get arguments: 3011 Node* receiver = argument(0); // type: oop 3012 Node* base = argument(1); // type: oop 3013 Node* offset = argument(2); // type: long 3014 Node* val = argument(4); // type: oop, int, or long 3015 3016 // Null check receiver. 3017 receiver = null_check(receiver); 3018 if (stopped()) { 3019 return true; 3020 } 3021 3022 // Build field offset expression. 3023 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3024 // 32-bit machines ignore the high half of long offsets 3025 offset = ConvL2X(offset); 3026 Node* adr = make_unsafe_address(base, offset); 3027 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3028 const Type *value_type = Type::get_const_basic_type(type); 3029 Compile::AliasType* alias_type = C->alias_type(adr_type); 3030 3031 insert_mem_bar(Op_MemBarRelease); 3032 insert_mem_bar(Op_MemBarCPUOrder); 3033 // Ensure that the store is atomic for longs: 3034 const bool require_atomic_access = true; 3035 Node* store; 3036 if (type == T_OBJECT) // reference stores need a store barrier. 3037 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3038 else { 3039 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3040 } 3041 insert_mem_bar(Op_MemBarCPUOrder); 3042 return true; 3043 } 3044 3045 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3046 // Regardless of form, don't allow previous ld/st to move down, 3047 // then issue acquire, release, or volatile mem_bar. 3048 insert_mem_bar(Op_MemBarCPUOrder); 3049 switch(id) { 3050 case vmIntrinsics::_loadFence: 3051 insert_mem_bar(Op_LoadFence); 3052 return true; 3053 case vmIntrinsics::_storeFence: 3054 insert_mem_bar(Op_StoreFence); 3055 return true; 3056 case vmIntrinsics::_fullFence: 3057 insert_mem_bar(Op_MemBarVolatile); 3058 return true; 3059 default: 3060 fatal_unexpected_iid(id); 3061 return false; 3062 } 3063 } 3064 3065 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3066 if (!kls->is_Con()) { 3067 return true; 3068 } 3069 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3070 if (klsptr == NULL) { 3071 return true; 3072 } 3073 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3074 // don't need a guard for a klass that is already initialized 3075 return !ik->is_initialized(); 3076 } 3077 3078 //----------------------------inline_unsafe_allocate--------------------------- 3079 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3080 bool LibraryCallKit::inline_unsafe_allocate() { 3081 if (callee()->is_static()) return false; // caller must have the capability! 3082 3083 null_check_receiver(); // null-check, then ignore 3084 Node* cls = null_check(argument(1)); 3085 if (stopped()) return true; 3086 3087 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3088 kls = null_check(kls); 3089 if (stopped()) return true; // argument was like int.class 3090 3091 Node* test = NULL; 3092 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3093 // Note: The argument might still be an illegal value like 3094 // Serializable.class or Object[].class. The runtime will handle it. 3095 // But we must make an explicit check for initialization. 3096 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3097 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3098 // can generate code to load it as unsigned byte. 3099 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3100 Node* bits = intcon(InstanceKlass::fully_initialized); 3101 test = _gvn.transform(new SubINode(inst, bits)); 3102 // The 'test' is non-zero if we need to take a slow path. 3103 } 3104 3105 Node* obj = new_instance(kls, test); 3106 set_result(obj); 3107 return true; 3108 } 3109 3110 #ifdef TRACE_HAVE_INTRINSICS 3111 /* 3112 * oop -> myklass 3113 * myklass->trace_id |= USED 3114 * return myklass->trace_id & ~0x3 3115 */ 3116 bool LibraryCallKit::inline_native_classID() { 3117 null_check_receiver(); // null-check, then ignore 3118 Node* cls = null_check(argument(1), T_OBJECT); 3119 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3120 kls = null_check(kls, T_OBJECT); 3121 ByteSize offset = TRACE_ID_OFFSET; 3122 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3123 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3124 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3125 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3126 Node* clsused = longcon(0x01l); // set the class bit 3127 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3128 3129 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3130 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3131 set_result(andl); 3132 return true; 3133 } 3134 3135 bool LibraryCallKit::inline_native_threadID() { 3136 Node* tls_ptr = NULL; 3137 Node* cur_thr = generate_current_thread(tls_ptr); 3138 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3139 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3140 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3141 3142 Node* threadid = NULL; 3143 size_t thread_id_size = OSThread::thread_id_size(); 3144 if (thread_id_size == (size_t) BytesPerLong) { 3145 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3146 } else if (thread_id_size == (size_t) BytesPerInt) { 3147 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3148 } else { 3149 ShouldNotReachHere(); 3150 } 3151 set_result(threadid); 3152 return true; 3153 } 3154 #endif 3155 3156 //------------------------inline_native_time_funcs-------------- 3157 // inline code for System.currentTimeMillis() and System.nanoTime() 3158 // these have the same type and signature 3159 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3160 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3161 const TypePtr* no_memory_effects = NULL; 3162 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3163 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3164 #ifdef ASSERT 3165 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3166 assert(value_top == top(), "second value must be top"); 3167 #endif 3168 set_result(value); 3169 return true; 3170 } 3171 3172 //------------------------inline_native_currentThread------------------ 3173 bool LibraryCallKit::inline_native_currentThread() { 3174 Node* junk = NULL; 3175 set_result(generate_current_thread(junk)); 3176 return true; 3177 } 3178 3179 //------------------------inline_native_isInterrupted------------------ 3180 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3181 bool LibraryCallKit::inline_native_isInterrupted() { 3182 // Add a fast path to t.isInterrupted(clear_int): 3183 // (t == Thread.current() && 3184 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3185 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3186 // So, in the common case that the interrupt bit is false, 3187 // we avoid making a call into the VM. Even if the interrupt bit 3188 // is true, if the clear_int argument is false, we avoid the VM call. 3189 // However, if the receiver is not currentThread, we must call the VM, 3190 // because there must be some locking done around the operation. 3191 3192 // We only go to the fast case code if we pass two guards. 3193 // Paths which do not pass are accumulated in the slow_region. 3194 3195 enum { 3196 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3197 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3198 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3199 PATH_LIMIT 3200 }; 3201 3202 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3203 // out of the function. 3204 insert_mem_bar(Op_MemBarCPUOrder); 3205 3206 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3207 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3208 3209 RegionNode* slow_region = new RegionNode(1); 3210 record_for_igvn(slow_region); 3211 3212 // (a) Receiving thread must be the current thread. 3213 Node* rec_thr = argument(0); 3214 Node* tls_ptr = NULL; 3215 Node* cur_thr = generate_current_thread(tls_ptr); 3216 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3217 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3218 3219 generate_slow_guard(bol_thr, slow_region); 3220 3221 // (b) Interrupt bit on TLS must be false. 3222 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3223 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3224 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3225 3226 // Set the control input on the field _interrupted read to prevent it floating up. 3227 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3228 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3229 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3230 3231 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3232 3233 // First fast path: if (!TLS._interrupted) return false; 3234 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3235 result_rgn->init_req(no_int_result_path, false_bit); 3236 result_val->init_req(no_int_result_path, intcon(0)); 3237 3238 // drop through to next case 3239 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3240 3241 #ifndef TARGET_OS_FAMILY_windows 3242 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3243 Node* clr_arg = argument(1); 3244 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3245 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3246 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3247 3248 // Second fast path: ... else if (!clear_int) return true; 3249 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3250 result_rgn->init_req(no_clear_result_path, false_arg); 3251 result_val->init_req(no_clear_result_path, intcon(1)); 3252 3253 // drop through to next case 3254 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3255 #else 3256 // To return true on Windows you must read the _interrupted field 3257 // and check the the event state i.e. take the slow path. 3258 #endif // TARGET_OS_FAMILY_windows 3259 3260 // (d) Otherwise, go to the slow path. 3261 slow_region->add_req(control()); 3262 set_control( _gvn.transform(slow_region)); 3263 3264 if (stopped()) { 3265 // There is no slow path. 3266 result_rgn->init_req(slow_result_path, top()); 3267 result_val->init_req(slow_result_path, top()); 3268 } else { 3269 // non-virtual because it is a private non-static 3270 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3271 3272 Node* slow_val = set_results_for_java_call(slow_call); 3273 // this->control() comes from set_results_for_java_call 3274 3275 Node* fast_io = slow_call->in(TypeFunc::I_O); 3276 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3277 3278 // These two phis are pre-filled with copies of of the fast IO and Memory 3279 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3280 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3281 3282 result_rgn->init_req(slow_result_path, control()); 3283 result_io ->init_req(slow_result_path, i_o()); 3284 result_mem->init_req(slow_result_path, reset_memory()); 3285 result_val->init_req(slow_result_path, slow_val); 3286 3287 set_all_memory(_gvn.transform(result_mem)); 3288 set_i_o( _gvn.transform(result_io)); 3289 } 3290 3291 C->set_has_split_ifs(true); // Has chance for split-if optimization 3292 set_result(result_rgn, result_val); 3293 return true; 3294 } 3295 3296 //---------------------------load_mirror_from_klass---------------------------- 3297 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3298 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3299 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3300 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3301 } 3302 3303 //-----------------------load_klass_from_mirror_common------------------------- 3304 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3305 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3306 // and branch to the given path on the region. 3307 // If never_see_null, take an uncommon trap on null, so we can optimistically 3308 // compile for the non-null case. 3309 // If the region is NULL, force never_see_null = true. 3310 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3311 bool never_see_null, 3312 RegionNode* region, 3313 int null_path, 3314 int offset) { 3315 if (region == NULL) never_see_null = true; 3316 Node* p = basic_plus_adr(mirror, offset); 3317 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3318 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3319 Node* null_ctl = top(); 3320 kls = null_check_oop(kls, &null_ctl, never_see_null); 3321 if (region != NULL) { 3322 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3323 region->init_req(null_path, null_ctl); 3324 } else { 3325 assert(null_ctl == top(), "no loose ends"); 3326 } 3327 return kls; 3328 } 3329 3330 //--------------------(inline_native_Class_query helpers)--------------------- 3331 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3332 // Fall through if (mods & mask) == bits, take the guard otherwise. 3333 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3334 // Branch around if the given klass has the given modifier bit set. 3335 // Like generate_guard, adds a new path onto the region. 3336 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3337 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3338 Node* mask = intcon(modifier_mask); 3339 Node* bits = intcon(modifier_bits); 3340 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3341 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3342 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3343 return generate_fair_guard(bol, region); 3344 } 3345 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3346 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3347 } 3348 3349 //-------------------------inline_native_Class_query------------------- 3350 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3351 const Type* return_type = TypeInt::BOOL; 3352 Node* prim_return_value = top(); // what happens if it's a primitive class? 3353 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3354 bool expect_prim = false; // most of these guys expect to work on refs 3355 3356 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3357 3358 Node* mirror = argument(0); 3359 Node* obj = top(); 3360 3361 switch (id) { 3362 case vmIntrinsics::_isInstance: 3363 // nothing is an instance of a primitive type 3364 prim_return_value = intcon(0); 3365 obj = argument(1); 3366 break; 3367 case vmIntrinsics::_getModifiers: 3368 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3369 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3370 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3371 break; 3372 case vmIntrinsics::_isInterface: 3373 prim_return_value = intcon(0); 3374 break; 3375 case vmIntrinsics::_isArray: 3376 prim_return_value = intcon(0); 3377 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3378 break; 3379 case vmIntrinsics::_isPrimitive: 3380 prim_return_value = intcon(1); 3381 expect_prim = true; // obviously 3382 break; 3383 case vmIntrinsics::_getSuperclass: 3384 prim_return_value = null(); 3385 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3386 break; 3387 case vmIntrinsics::_getClassAccessFlags: 3388 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3389 return_type = TypeInt::INT; // not bool! 6297094 3390 break; 3391 default: 3392 fatal_unexpected_iid(id); 3393 break; 3394 } 3395 3396 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3397 if (mirror_con == NULL) return false; // cannot happen? 3398 3399 #ifndef PRODUCT 3400 if (C->print_intrinsics() || C->print_inlining()) { 3401 ciType* k = mirror_con->java_mirror_type(); 3402 if (k) { 3403 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3404 k->print_name(); 3405 tty->cr(); 3406 } 3407 } 3408 #endif 3409 3410 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3411 RegionNode* region = new RegionNode(PATH_LIMIT); 3412 record_for_igvn(region); 3413 PhiNode* phi = new PhiNode(region, return_type); 3414 3415 // The mirror will never be null of Reflection.getClassAccessFlags, however 3416 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3417 // if it is. See bug 4774291. 3418 3419 // For Reflection.getClassAccessFlags(), the null check occurs in 3420 // the wrong place; see inline_unsafe_access(), above, for a similar 3421 // situation. 3422 mirror = null_check(mirror); 3423 // If mirror or obj is dead, only null-path is taken. 3424 if (stopped()) return true; 3425 3426 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3427 3428 // Now load the mirror's klass metaobject, and null-check it. 3429 // Side-effects region with the control path if the klass is null. 3430 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3431 // If kls is null, we have a primitive mirror. 3432 phi->init_req(_prim_path, prim_return_value); 3433 if (stopped()) { set_result(region, phi); return true; } 3434 bool safe_for_replace = (region->in(_prim_path) == top()); 3435 3436 Node* p; // handy temp 3437 Node* null_ctl; 3438 3439 // Now that we have the non-null klass, we can perform the real query. 3440 // For constant classes, the query will constant-fold in LoadNode::Value. 3441 Node* query_value = top(); 3442 switch (id) { 3443 case vmIntrinsics::_isInstance: 3444 // nothing is an instance of a primitive type 3445 query_value = gen_instanceof(obj, kls, safe_for_replace); 3446 break; 3447 3448 case vmIntrinsics::_getModifiers: 3449 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3450 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3451 break; 3452 3453 case vmIntrinsics::_isInterface: 3454 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3455 if (generate_interface_guard(kls, region) != NULL) 3456 // A guard was added. If the guard is taken, it was an interface. 3457 phi->add_req(intcon(1)); 3458 // If we fall through, it's a plain class. 3459 query_value = intcon(0); 3460 break; 3461 3462 case vmIntrinsics::_isArray: 3463 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3464 if (generate_array_guard(kls, region) != NULL) 3465 // A guard was added. If the guard is taken, it was an array. 3466 phi->add_req(intcon(1)); 3467 // If we fall through, it's a plain class. 3468 query_value = intcon(0); 3469 break; 3470 3471 case vmIntrinsics::_isPrimitive: 3472 query_value = intcon(0); // "normal" path produces false 3473 break; 3474 3475 case vmIntrinsics::_getSuperclass: 3476 // The rules here are somewhat unfortunate, but we can still do better 3477 // with random logic than with a JNI call. 3478 // Interfaces store null or Object as _super, but must report null. 3479 // Arrays store an intermediate super as _super, but must report Object. 3480 // Other types can report the actual _super. 3481 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3482 if (generate_interface_guard(kls, region) != NULL) 3483 // A guard was added. If the guard is taken, it was an interface. 3484 phi->add_req(null()); 3485 if (generate_array_guard(kls, region) != NULL) 3486 // A guard was added. If the guard is taken, it was an array. 3487 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3488 // If we fall through, it's a plain class. Get its _super. 3489 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3490 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3491 null_ctl = top(); 3492 kls = null_check_oop(kls, &null_ctl); 3493 if (null_ctl != top()) { 3494 // If the guard is taken, Object.superClass is null (both klass and mirror). 3495 region->add_req(null_ctl); 3496 phi ->add_req(null()); 3497 } 3498 if (!stopped()) { 3499 query_value = load_mirror_from_klass(kls); 3500 } 3501 break; 3502 3503 case vmIntrinsics::_getClassAccessFlags: 3504 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3505 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3506 break; 3507 3508 default: 3509 fatal_unexpected_iid(id); 3510 break; 3511 } 3512 3513 // Fall-through is the normal case of a query to a real class. 3514 phi->init_req(1, query_value); 3515 region->init_req(1, control()); 3516 3517 C->set_has_split_ifs(true); // Has chance for split-if optimization 3518 set_result(region, phi); 3519 return true; 3520 } 3521 3522 //-------------------------inline_Class_cast------------------- 3523 bool LibraryCallKit::inline_Class_cast() { 3524 Node* mirror = argument(0); // Class 3525 Node* obj = argument(1); 3526 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3527 if (mirror_con == NULL) { 3528 return false; // dead path (mirror->is_top()). 3529 } 3530 if (obj == NULL || obj->is_top()) { 3531 return false; // dead path 3532 } 3533 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3534 3535 // First, see if Class.cast() can be folded statically. 3536 // java_mirror_type() returns non-null for compile-time Class constants. 3537 ciType* tm = mirror_con->java_mirror_type(); 3538 if (tm != NULL && tm->is_klass() && 3539 tp != NULL && tp->klass() != NULL) { 3540 if (!tp->klass()->is_loaded()) { 3541 // Don't use intrinsic when class is not loaded. 3542 return false; 3543 } else { 3544 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3545 if (static_res == Compile::SSC_always_true) { 3546 // isInstance() is true - fold the code. 3547 set_result(obj); 3548 return true; 3549 } else if (static_res == Compile::SSC_always_false) { 3550 // Don't use intrinsic, have to throw ClassCastException. 3551 // If the reference is null, the non-intrinsic bytecode will 3552 // be optimized appropriately. 3553 return false; 3554 } 3555 } 3556 } 3557 3558 // Bailout intrinsic and do normal inlining if exception path is frequent. 3559 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3560 return false; 3561 } 3562 3563 // Generate dynamic checks. 3564 // Class.cast() is java implementation of _checkcast bytecode. 3565 // Do checkcast (Parse::do_checkcast()) optimizations here. 3566 3567 mirror = null_check(mirror); 3568 // If mirror is dead, only null-path is taken. 3569 if (stopped()) { 3570 return true; 3571 } 3572 3573 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3574 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3575 RegionNode* region = new RegionNode(PATH_LIMIT); 3576 record_for_igvn(region); 3577 3578 // Now load the mirror's klass metaobject, and null-check it. 3579 // If kls is null, we have a primitive mirror and 3580 // nothing is an instance of a primitive type. 3581 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3582 3583 Node* res = top(); 3584 if (!stopped()) { 3585 Node* bad_type_ctrl = top(); 3586 // Do checkcast optimizations. 3587 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3588 region->init_req(_bad_type_path, bad_type_ctrl); 3589 } 3590 if (region->in(_prim_path) != top() || 3591 region->in(_bad_type_path) != top()) { 3592 // Let Interpreter throw ClassCastException. 3593 PreserveJVMState pjvms(this); 3594 set_control(_gvn.transform(region)); 3595 uncommon_trap(Deoptimization::Reason_intrinsic, 3596 Deoptimization::Action_maybe_recompile); 3597 } 3598 if (!stopped()) { 3599 set_result(res); 3600 } 3601 return true; 3602 } 3603 3604 3605 //--------------------------inline_native_subtype_check------------------------ 3606 // This intrinsic takes the JNI calls out of the heart of 3607 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3608 bool LibraryCallKit::inline_native_subtype_check() { 3609 // Pull both arguments off the stack. 3610 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3611 args[0] = argument(0); 3612 args[1] = argument(1); 3613 Node* klasses[2]; // corresponding Klasses: superk, subk 3614 klasses[0] = klasses[1] = top(); 3615 3616 enum { 3617 // A full decision tree on {superc is prim, subc is prim}: 3618 _prim_0_path = 1, // {P,N} => false 3619 // {P,P} & superc!=subc => false 3620 _prim_same_path, // {P,P} & superc==subc => true 3621 _prim_1_path, // {N,P} => false 3622 _ref_subtype_path, // {N,N} & subtype check wins => true 3623 _both_ref_path, // {N,N} & subtype check loses => false 3624 PATH_LIMIT 3625 }; 3626 3627 RegionNode* region = new RegionNode(PATH_LIMIT); 3628 Node* phi = new PhiNode(region, TypeInt::BOOL); 3629 record_for_igvn(region); 3630 3631 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3632 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3633 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3634 3635 // First null-check both mirrors and load each mirror's klass metaobject. 3636 int which_arg; 3637 for (which_arg = 0; which_arg <= 1; which_arg++) { 3638 Node* arg = args[which_arg]; 3639 arg = null_check(arg); 3640 if (stopped()) break; 3641 args[which_arg] = arg; 3642 3643 Node* p = basic_plus_adr(arg, class_klass_offset); 3644 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3645 klasses[which_arg] = _gvn.transform(kls); 3646 } 3647 3648 // Having loaded both klasses, test each for null. 3649 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3650 for (which_arg = 0; which_arg <= 1; which_arg++) { 3651 Node* kls = klasses[which_arg]; 3652 Node* null_ctl = top(); 3653 kls = null_check_oop(kls, &null_ctl, never_see_null); 3654 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3655 region->init_req(prim_path, null_ctl); 3656 if (stopped()) break; 3657 klasses[which_arg] = kls; 3658 } 3659 3660 if (!stopped()) { 3661 // now we have two reference types, in klasses[0..1] 3662 Node* subk = klasses[1]; // the argument to isAssignableFrom 3663 Node* superk = klasses[0]; // the receiver 3664 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3665 // now we have a successful reference subtype check 3666 region->set_req(_ref_subtype_path, control()); 3667 } 3668 3669 // If both operands are primitive (both klasses null), then 3670 // we must return true when they are identical primitives. 3671 // It is convenient to test this after the first null klass check. 3672 set_control(region->in(_prim_0_path)); // go back to first null check 3673 if (!stopped()) { 3674 // Since superc is primitive, make a guard for the superc==subc case. 3675 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3676 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3677 generate_guard(bol_eq, region, PROB_FAIR); 3678 if (region->req() == PATH_LIMIT+1) { 3679 // A guard was added. If the added guard is taken, superc==subc. 3680 region->swap_edges(PATH_LIMIT, _prim_same_path); 3681 region->del_req(PATH_LIMIT); 3682 } 3683 region->set_req(_prim_0_path, control()); // Not equal after all. 3684 } 3685 3686 // these are the only paths that produce 'true': 3687 phi->set_req(_prim_same_path, intcon(1)); 3688 phi->set_req(_ref_subtype_path, intcon(1)); 3689 3690 // pull together the cases: 3691 assert(region->req() == PATH_LIMIT, "sane region"); 3692 for (uint i = 1; i < region->req(); i++) { 3693 Node* ctl = region->in(i); 3694 if (ctl == NULL || ctl == top()) { 3695 region->set_req(i, top()); 3696 phi ->set_req(i, top()); 3697 } else if (phi->in(i) == NULL) { 3698 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3699 } 3700 } 3701 3702 set_control(_gvn.transform(region)); 3703 set_result(_gvn.transform(phi)); 3704 return true; 3705 } 3706 3707 //---------------------generate_array_guard_common------------------------ 3708 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3709 bool obj_array, bool not_array) { 3710 3711 if (stopped()) { 3712 return NULL; 3713 } 3714 3715 // If obj_array/non_array==false/false: 3716 // Branch around if the given klass is in fact an array (either obj or prim). 3717 // If obj_array/non_array==false/true: 3718 // Branch around if the given klass is not an array klass of any kind. 3719 // If obj_array/non_array==true/true: 3720 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3721 // If obj_array/non_array==true/false: 3722 // Branch around if the kls is an oop array (Object[] or subtype) 3723 // 3724 // Like generate_guard, adds a new path onto the region. 3725 jint layout_con = 0; 3726 Node* layout_val = get_layout_helper(kls, layout_con); 3727 if (layout_val == NULL) { 3728 bool query = (obj_array 3729 ? Klass::layout_helper_is_objArray(layout_con) 3730 : Klass::layout_helper_is_array(layout_con)); 3731 if (query == not_array) { 3732 return NULL; // never a branch 3733 } else { // always a branch 3734 Node* always_branch = control(); 3735 if (region != NULL) 3736 region->add_req(always_branch); 3737 set_control(top()); 3738 return always_branch; 3739 } 3740 } 3741 // Now test the correct condition. 3742 jint nval = (obj_array 3743 ? ((jint)Klass::_lh_array_tag_type_value 3744 << Klass::_lh_array_tag_shift) 3745 : Klass::_lh_neutral_value); 3746 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3747 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3748 // invert the test if we are looking for a non-array 3749 if (not_array) btest = BoolTest(btest).negate(); 3750 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3751 return generate_fair_guard(bol, region); 3752 } 3753 3754 3755 //-----------------------inline_native_newArray-------------------------- 3756 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3757 bool LibraryCallKit::inline_native_newArray() { 3758 Node* mirror = argument(0); 3759 Node* count_val = argument(1); 3760 3761 mirror = null_check(mirror); 3762 // If mirror or obj is dead, only null-path is taken. 3763 if (stopped()) return true; 3764 3765 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3766 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3767 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3768 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3769 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3770 3771 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3772 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3773 result_reg, _slow_path); 3774 Node* normal_ctl = control(); 3775 Node* no_array_ctl = result_reg->in(_slow_path); 3776 3777 // Generate code for the slow case. We make a call to newArray(). 3778 set_control(no_array_ctl); 3779 if (!stopped()) { 3780 // Either the input type is void.class, or else the 3781 // array klass has not yet been cached. Either the 3782 // ensuing call will throw an exception, or else it 3783 // will cache the array klass for next time. 3784 PreserveJVMState pjvms(this); 3785 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3786 Node* slow_result = set_results_for_java_call(slow_call); 3787 // this->control() comes from set_results_for_java_call 3788 result_reg->set_req(_slow_path, control()); 3789 result_val->set_req(_slow_path, slow_result); 3790 result_io ->set_req(_slow_path, i_o()); 3791 result_mem->set_req(_slow_path, reset_memory()); 3792 } 3793 3794 set_control(normal_ctl); 3795 if (!stopped()) { 3796 // Normal case: The array type has been cached in the java.lang.Class. 3797 // The following call works fine even if the array type is polymorphic. 3798 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3799 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3800 result_reg->init_req(_normal_path, control()); 3801 result_val->init_req(_normal_path, obj); 3802 result_io ->init_req(_normal_path, i_o()); 3803 result_mem->init_req(_normal_path, reset_memory()); 3804 } 3805 3806 // Return the combined state. 3807 set_i_o( _gvn.transform(result_io) ); 3808 set_all_memory( _gvn.transform(result_mem)); 3809 3810 C->set_has_split_ifs(true); // Has chance for split-if optimization 3811 set_result(result_reg, result_val); 3812 return true; 3813 } 3814 3815 //----------------------inline_native_getLength-------------------------- 3816 // public static native int java.lang.reflect.Array.getLength(Object array); 3817 bool LibraryCallKit::inline_native_getLength() { 3818 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3819 3820 Node* array = null_check(argument(0)); 3821 // If array is dead, only null-path is taken. 3822 if (stopped()) return true; 3823 3824 // Deoptimize if it is a non-array. 3825 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3826 3827 if (non_array != NULL) { 3828 PreserveJVMState pjvms(this); 3829 set_control(non_array); 3830 uncommon_trap(Deoptimization::Reason_intrinsic, 3831 Deoptimization::Action_maybe_recompile); 3832 } 3833 3834 // If control is dead, only non-array-path is taken. 3835 if (stopped()) return true; 3836 3837 // The works fine even if the array type is polymorphic. 3838 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3839 Node* result = load_array_length(array); 3840 3841 C->set_has_split_ifs(true); // Has chance for split-if optimization 3842 set_result(result); 3843 return true; 3844 } 3845 3846 //------------------------inline_array_copyOf---------------------------- 3847 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3848 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3849 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3850 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3851 3852 // Get the arguments. 3853 Node* original = argument(0); 3854 Node* start = is_copyOfRange? argument(1): intcon(0); 3855 Node* end = is_copyOfRange? argument(2): argument(1); 3856 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3857 3858 Node* newcopy; 3859 3860 // Set the original stack and the reexecute bit for the interpreter to reexecute 3861 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3862 { PreserveReexecuteState preexecs(this); 3863 jvms()->set_should_reexecute(true); 3864 3865 array_type_mirror = null_check(array_type_mirror); 3866 original = null_check(original); 3867 3868 // Check if a null path was taken unconditionally. 3869 if (stopped()) return true; 3870 3871 Node* orig_length = load_array_length(original); 3872 3873 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3874 klass_node = null_check(klass_node); 3875 3876 RegionNode* bailout = new RegionNode(1); 3877 record_for_igvn(bailout); 3878 3879 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3880 // Bail out if that is so. 3881 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3882 if (not_objArray != NULL) { 3883 // Improve the klass node's type from the new optimistic assumption: 3884 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3885 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3886 Node* cast = new CastPPNode(klass_node, akls); 3887 cast->init_req(0, control()); 3888 klass_node = _gvn.transform(cast); 3889 } 3890 3891 // Bail out if either start or end is negative. 3892 generate_negative_guard(start, bailout, &start); 3893 generate_negative_guard(end, bailout, &end); 3894 3895 Node* length = end; 3896 if (_gvn.type(start) != TypeInt::ZERO) { 3897 length = _gvn.transform(new SubINode(end, start)); 3898 } 3899 3900 // Bail out if length is negative. 3901 // Without this the new_array would throw 3902 // NegativeArraySizeException but IllegalArgumentException is what 3903 // should be thrown 3904 generate_negative_guard(length, bailout, &length); 3905 3906 if (bailout->req() > 1) { 3907 PreserveJVMState pjvms(this); 3908 set_control(_gvn.transform(bailout)); 3909 uncommon_trap(Deoptimization::Reason_intrinsic, 3910 Deoptimization::Action_maybe_recompile); 3911 } 3912 3913 if (!stopped()) { 3914 // How many elements will we copy from the original? 3915 // The answer is MinI(orig_length - start, length). 3916 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3917 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3918 3919 // Generate a direct call to the right arraycopy function(s). 3920 // We know the copy is disjoint but we might not know if the 3921 // oop stores need checking. 3922 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3923 // This will fail a store-check if x contains any non-nulls. 3924 3925 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3926 // loads/stores but it is legal only if we're sure the 3927 // Arrays.copyOf would succeed. So we need all input arguments 3928 // to the copyOf to be validated, including that the copy to the 3929 // new array won't trigger an ArrayStoreException. That subtype 3930 // check can be optimized if we know something on the type of 3931 // the input array from type speculation. 3932 if (_gvn.type(klass_node)->singleton()) { 3933 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3934 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3935 3936 int test = C->static_subtype_check(superk, subk); 3937 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3938 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3939 if (t_original->speculative_type() != NULL) { 3940 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3941 } 3942 } 3943 } 3944 3945 bool validated = false; 3946 // Reason_class_check rather than Reason_intrinsic because we 3947 // want to intrinsify even if this traps. 3948 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3949 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3950 klass_node); 3951 3952 if (not_subtype_ctrl != top()) { 3953 PreserveJVMState pjvms(this); 3954 set_control(not_subtype_ctrl); 3955 uncommon_trap(Deoptimization::Reason_class_check, 3956 Deoptimization::Action_make_not_entrant); 3957 assert(stopped(), "Should be stopped"); 3958 } 3959 validated = true; 3960 } 3961 3962 if (!stopped()) { 3963 newcopy = new_array(klass_node, length, 0); // no arguments to push 3964 3965 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3966 load_object_klass(original), klass_node); 3967 if (!is_copyOfRange) { 3968 ac->set_copyof(validated); 3969 } else { 3970 ac->set_copyofrange(validated); 3971 } 3972 Node* n = _gvn.transform(ac); 3973 if (n == ac) { 3974 ac->connect_outputs(this); 3975 } else { 3976 assert(validated, "shouldn't transform if all arguments not validated"); 3977 set_all_memory(n); 3978 } 3979 } 3980 } 3981 } // original reexecute is set back here 3982 3983 C->set_has_split_ifs(true); // Has chance for split-if optimization 3984 if (!stopped()) { 3985 set_result(newcopy); 3986 } 3987 return true; 3988 } 3989 3990 3991 //----------------------generate_virtual_guard--------------------------- 3992 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3993 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3994 RegionNode* slow_region) { 3995 ciMethod* method = callee(); 3996 int vtable_index = method->vtable_index(); 3997 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3998 err_msg_res("bad index %d", vtable_index)); 3999 // Get the Method* out of the appropriate vtable entry. 4000 int entry_offset = (InstanceKlass::vtable_start_offset() + 4001 vtable_index*vtableEntry::size()) * wordSize + 4002 vtableEntry::method_offset_in_bytes(); 4003 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4004 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4005 4006 // Compare the target method with the expected method (e.g., Object.hashCode). 4007 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4008 4009 Node* native_call = makecon(native_call_addr); 4010 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4011 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4012 4013 return generate_slow_guard(test_native, slow_region); 4014 } 4015 4016 //-----------------------generate_method_call---------------------------- 4017 // Use generate_method_call to make a slow-call to the real 4018 // method if the fast path fails. An alternative would be to 4019 // use a stub like OptoRuntime::slow_arraycopy_Java. 4020 // This only works for expanding the current library call, 4021 // not another intrinsic. (E.g., don't use this for making an 4022 // arraycopy call inside of the copyOf intrinsic.) 4023 CallJavaNode* 4024 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4025 // When compiling the intrinsic method itself, do not use this technique. 4026 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4027 4028 ciMethod* method = callee(); 4029 // ensure the JVMS we have will be correct for this call 4030 guarantee(method_id == method->intrinsic_id(), "must match"); 4031 4032 const TypeFunc* tf = TypeFunc::make(method); 4033 CallJavaNode* slow_call; 4034 if (is_static) { 4035 assert(!is_virtual, ""); 4036 slow_call = new CallStaticJavaNode(C, tf, 4037 SharedRuntime::get_resolve_static_call_stub(), 4038 method, bci()); 4039 } else if (is_virtual) { 4040 null_check_receiver(); 4041 int vtable_index = Method::invalid_vtable_index; 4042 if (UseInlineCaches) { 4043 // Suppress the vtable call 4044 } else { 4045 // hashCode and clone are not a miranda methods, 4046 // so the vtable index is fixed. 4047 // No need to use the linkResolver to get it. 4048 vtable_index = method->vtable_index(); 4049 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4050 err_msg_res("bad index %d", vtable_index)); 4051 } 4052 slow_call = new CallDynamicJavaNode(tf, 4053 SharedRuntime::get_resolve_virtual_call_stub(), 4054 method, vtable_index, bci()); 4055 } else { // neither virtual nor static: opt_virtual 4056 null_check_receiver(); 4057 slow_call = new CallStaticJavaNode(C, tf, 4058 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4059 method, bci()); 4060 slow_call->set_optimized_virtual(true); 4061 } 4062 set_arguments_for_java_call(slow_call); 4063 set_edges_for_java_call(slow_call); 4064 return slow_call; 4065 } 4066 4067 4068 /** 4069 * Build special case code for calls to hashCode on an object. This call may 4070 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4071 * slightly different code. 4072 */ 4073 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4074 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4075 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4076 4077 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4078 4079 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4080 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4081 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4082 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4083 Node* obj = NULL; 4084 if (!is_static) { 4085 // Check for hashing null object 4086 obj = null_check_receiver(); 4087 if (stopped()) return true; // unconditionally null 4088 result_reg->init_req(_null_path, top()); 4089 result_val->init_req(_null_path, top()); 4090 } else { 4091 // Do a null check, and return zero if null. 4092 // System.identityHashCode(null) == 0 4093 obj = argument(0); 4094 Node* null_ctl = top(); 4095 obj = null_check_oop(obj, &null_ctl); 4096 result_reg->init_req(_null_path, null_ctl); 4097 result_val->init_req(_null_path, _gvn.intcon(0)); 4098 } 4099 4100 // Unconditionally null? Then return right away. 4101 if (stopped()) { 4102 set_control( result_reg->in(_null_path)); 4103 if (!stopped()) 4104 set_result(result_val->in(_null_path)); 4105 return true; 4106 } 4107 4108 // We only go to the fast case code if we pass a number of guards. The 4109 // paths which do not pass are accumulated in the slow_region. 4110 RegionNode* slow_region = new RegionNode(1); 4111 record_for_igvn(slow_region); 4112 4113 // If this is a virtual call, we generate a funny guard. We pull out 4114 // the vtable entry corresponding to hashCode() from the target object. 4115 // If the target method which we are calling happens to be the native 4116 // Object hashCode() method, we pass the guard. We do not need this 4117 // guard for non-virtual calls -- the caller is known to be the native 4118 // Object hashCode(). 4119 if (is_virtual) { 4120 // After null check, get the object's klass. 4121 Node* obj_klass = load_object_klass(obj); 4122 generate_virtual_guard(obj_klass, slow_region); 4123 } 4124 4125 // Get the header out of the object, use LoadMarkNode when available 4126 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4127 // The control of the load must be NULL. Otherwise, the load can move before 4128 // the null check after castPP removal. 4129 Node* no_ctrl = NULL; 4130 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4131 4132 // Test the header to see if it is unlocked. 4133 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4134 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4135 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4136 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4137 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4138 4139 generate_slow_guard(test_unlocked, slow_region); 4140 4141 // Get the hash value and check to see that it has been properly assigned. 4142 // We depend on hash_mask being at most 32 bits and avoid the use of 4143 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4144 // vm: see markOop.hpp. 4145 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4146 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4147 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4148 // This hack lets the hash bits live anywhere in the mark object now, as long 4149 // as the shift drops the relevant bits into the low 32 bits. Note that 4150 // Java spec says that HashCode is an int so there's no point in capturing 4151 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4152 hshifted_header = ConvX2I(hshifted_header); 4153 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4154 4155 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4156 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4157 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4158 4159 generate_slow_guard(test_assigned, slow_region); 4160 4161 Node* init_mem = reset_memory(); 4162 // fill in the rest of the null path: 4163 result_io ->init_req(_null_path, i_o()); 4164 result_mem->init_req(_null_path, init_mem); 4165 4166 result_val->init_req(_fast_path, hash_val); 4167 result_reg->init_req(_fast_path, control()); 4168 result_io ->init_req(_fast_path, i_o()); 4169 result_mem->init_req(_fast_path, init_mem); 4170 4171 // Generate code for the slow case. We make a call to hashCode(). 4172 set_control(_gvn.transform(slow_region)); 4173 if (!stopped()) { 4174 // No need for PreserveJVMState, because we're using up the present state. 4175 set_all_memory(init_mem); 4176 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4177 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4178 Node* slow_result = set_results_for_java_call(slow_call); 4179 // this->control() comes from set_results_for_java_call 4180 result_reg->init_req(_slow_path, control()); 4181 result_val->init_req(_slow_path, slow_result); 4182 result_io ->set_req(_slow_path, i_o()); 4183 result_mem ->set_req(_slow_path, reset_memory()); 4184 } 4185 4186 // Return the combined state. 4187 set_i_o( _gvn.transform(result_io) ); 4188 set_all_memory( _gvn.transform(result_mem)); 4189 4190 set_result(result_reg, result_val); 4191 return true; 4192 } 4193 4194 //---------------------------inline_native_getClass---------------------------- 4195 // public final native Class<?> java.lang.Object.getClass(); 4196 // 4197 // Build special case code for calls to getClass on an object. 4198 bool LibraryCallKit::inline_native_getClass() { 4199 Node* obj = null_check_receiver(); 4200 if (stopped()) return true; 4201 set_result(load_mirror_from_klass(load_object_klass(obj))); 4202 return true; 4203 } 4204 4205 //-----------------inline_native_Reflection_getCallerClass--------------------- 4206 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4207 // 4208 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4209 // 4210 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4211 // in that it must skip particular security frames and checks for 4212 // caller sensitive methods. 4213 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4214 #ifndef PRODUCT 4215 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4216 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4217 } 4218 #endif 4219 4220 if (!jvms()->has_method()) { 4221 #ifndef PRODUCT 4222 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4223 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4224 } 4225 #endif 4226 return false; 4227 } 4228 4229 // Walk back up the JVM state to find the caller at the required 4230 // depth. 4231 JVMState* caller_jvms = jvms(); 4232 4233 // Cf. JVM_GetCallerClass 4234 // NOTE: Start the loop at depth 1 because the current JVM state does 4235 // not include the Reflection.getCallerClass() frame. 4236 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4237 ciMethod* m = caller_jvms->method(); 4238 switch (n) { 4239 case 0: 4240 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4241 break; 4242 case 1: 4243 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4244 if (!m->caller_sensitive()) { 4245 #ifndef PRODUCT 4246 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4247 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4248 } 4249 #endif 4250 return false; // bail-out; let JVM_GetCallerClass do the work 4251 } 4252 break; 4253 default: 4254 if (!m->is_ignored_by_security_stack_walk()) { 4255 // We have reached the desired frame; return the holder class. 4256 // Acquire method holder as java.lang.Class and push as constant. 4257 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4258 ciInstance* caller_mirror = caller_klass->java_mirror(); 4259 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4260 4261 #ifndef PRODUCT 4262 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4263 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()); 4264 tty->print_cr(" JVM state at this point:"); 4265 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4266 ciMethod* m = jvms()->of_depth(i)->method(); 4267 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4268 } 4269 } 4270 #endif 4271 return true; 4272 } 4273 break; 4274 } 4275 } 4276 4277 #ifndef PRODUCT 4278 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4279 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4280 tty->print_cr(" JVM state at this point:"); 4281 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4282 ciMethod* m = jvms()->of_depth(i)->method(); 4283 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4284 } 4285 } 4286 #endif 4287 4288 return false; // bail-out; let JVM_GetCallerClass do the work 4289 } 4290 4291 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4292 Node* arg = argument(0); 4293 Node* result; 4294 4295 switch (id) { 4296 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4297 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4298 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4299 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4300 4301 case vmIntrinsics::_doubleToLongBits: { 4302 // two paths (plus control) merge in a wood 4303 RegionNode *r = new RegionNode(3); 4304 Node *phi = new PhiNode(r, TypeLong::LONG); 4305 4306 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4307 // Build the boolean node 4308 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4309 4310 // Branch either way. 4311 // NaN case is less traveled, which makes all the difference. 4312 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4313 Node *opt_isnan = _gvn.transform(ifisnan); 4314 assert( opt_isnan->is_If(), "Expect an IfNode"); 4315 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4316 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4317 4318 set_control(iftrue); 4319 4320 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4321 Node *slow_result = longcon(nan_bits); // return NaN 4322 phi->init_req(1, _gvn.transform( slow_result )); 4323 r->init_req(1, iftrue); 4324 4325 // Else fall through 4326 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4327 set_control(iffalse); 4328 4329 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4330 r->init_req(2, iffalse); 4331 4332 // Post merge 4333 set_control(_gvn.transform(r)); 4334 record_for_igvn(r); 4335 4336 C->set_has_split_ifs(true); // Has chance for split-if optimization 4337 result = phi; 4338 assert(result->bottom_type()->isa_long(), "must be"); 4339 break; 4340 } 4341 4342 case vmIntrinsics::_floatToIntBits: { 4343 // two paths (plus control) merge in a wood 4344 RegionNode *r = new RegionNode(3); 4345 Node *phi = new PhiNode(r, TypeInt::INT); 4346 4347 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4348 // Build the boolean node 4349 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4350 4351 // Branch either way. 4352 // NaN case is less traveled, which makes all the difference. 4353 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4354 Node *opt_isnan = _gvn.transform(ifisnan); 4355 assert( opt_isnan->is_If(), "Expect an IfNode"); 4356 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4357 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4358 4359 set_control(iftrue); 4360 4361 static const jint nan_bits = 0x7fc00000; 4362 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4363 phi->init_req(1, _gvn.transform( slow_result )); 4364 r->init_req(1, iftrue); 4365 4366 // Else fall through 4367 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4368 set_control(iffalse); 4369 4370 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4371 r->init_req(2, iffalse); 4372 4373 // Post merge 4374 set_control(_gvn.transform(r)); 4375 record_for_igvn(r); 4376 4377 C->set_has_split_ifs(true); // Has chance for split-if optimization 4378 result = phi; 4379 assert(result->bottom_type()->isa_int(), "must be"); 4380 break; 4381 } 4382 4383 default: 4384 fatal_unexpected_iid(id); 4385 break; 4386 } 4387 set_result(_gvn.transform(result)); 4388 return true; 4389 } 4390 4391 #ifdef _LP64 4392 #define XTOP ,top() /*additional argument*/ 4393 #else //_LP64 4394 #define XTOP /*no additional argument*/ 4395 #endif //_LP64 4396 4397 //----------------------inline_unsafe_copyMemory------------------------- 4398 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4399 bool LibraryCallKit::inline_unsafe_copyMemory() { 4400 if (callee()->is_static()) return false; // caller must have the capability! 4401 null_check_receiver(); // null-check receiver 4402 if (stopped()) return true; 4403 4404 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4405 4406 Node* src_ptr = argument(1); // type: oop 4407 Node* src_off = ConvL2X(argument(2)); // type: long 4408 Node* dst_ptr = argument(4); // type: oop 4409 Node* dst_off = ConvL2X(argument(5)); // type: long 4410 Node* size = ConvL2X(argument(7)); // type: long 4411 4412 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4413 "fieldOffset must be byte-scaled"); 4414 4415 Node* src = make_unsafe_address(src_ptr, src_off); 4416 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4417 4418 // Conservatively insert a memory barrier on all memory slices. 4419 // Do not let writes of the copy source or destination float below the copy. 4420 insert_mem_bar(Op_MemBarCPUOrder); 4421 4422 // Call it. Note that the length argument is not scaled. 4423 make_runtime_call(RC_LEAF|RC_NO_FP, 4424 OptoRuntime::fast_arraycopy_Type(), 4425 StubRoutines::unsafe_arraycopy(), 4426 "unsafe_arraycopy", 4427 TypeRawPtr::BOTTOM, 4428 src, dst, size XTOP); 4429 4430 // Do not let reads of the copy destination float above the copy. 4431 insert_mem_bar(Op_MemBarCPUOrder); 4432 4433 return true; 4434 } 4435 4436 //------------------------clone_coping----------------------------------- 4437 // Helper function for inline_native_clone. 4438 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4439 assert(obj_size != NULL, ""); 4440 Node* raw_obj = alloc_obj->in(1); 4441 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4442 4443 AllocateNode* alloc = NULL; 4444 if (ReduceBulkZeroing) { 4445 // We will be completely responsible for initializing this object - 4446 // mark Initialize node as complete. 4447 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4448 // The object was just allocated - there should be no any stores! 4449 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4450 // Mark as complete_with_arraycopy so that on AllocateNode 4451 // expansion, we know this AllocateNode is initialized by an array 4452 // copy and a StoreStore barrier exists after the array copy. 4453 alloc->initialization()->set_complete_with_arraycopy(); 4454 } 4455 4456 // Copy the fastest available way. 4457 // TODO: generate fields copies for small objects instead. 4458 Node* src = obj; 4459 Node* dest = alloc_obj; 4460 Node* size = _gvn.transform(obj_size); 4461 4462 // Exclude the header but include array length to copy by 8 bytes words. 4463 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4464 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4465 instanceOopDesc::base_offset_in_bytes(); 4466 // base_off: 4467 // 8 - 32-bit VM 4468 // 12 - 64-bit VM, compressed klass 4469 // 16 - 64-bit VM, normal klass 4470 if (base_off % BytesPerLong != 0) { 4471 assert(UseCompressedClassPointers, ""); 4472 if (is_array) { 4473 // Exclude length to copy by 8 bytes words. 4474 base_off += sizeof(int); 4475 } else { 4476 // Include klass to copy by 8 bytes words. 4477 base_off = instanceOopDesc::klass_offset_in_bytes(); 4478 } 4479 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4480 } 4481 src = basic_plus_adr(src, base_off); 4482 dest = basic_plus_adr(dest, base_off); 4483 4484 // Compute the length also, if needed: 4485 Node* countx = size; 4486 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4487 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4488 4489 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4490 4491 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4492 ac->set_clonebasic(); 4493 Node* n = _gvn.transform(ac); 4494 if (n == ac) { 4495 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4496 } else { 4497 set_all_memory(n); 4498 } 4499 4500 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4501 if (card_mark) { 4502 assert(!is_array, ""); 4503 // Put in store barrier for any and all oops we are sticking 4504 // into this object. (We could avoid this if we could prove 4505 // that the object type contains no oop fields at all.) 4506 Node* no_particular_value = NULL; 4507 Node* no_particular_field = NULL; 4508 int raw_adr_idx = Compile::AliasIdxRaw; 4509 post_barrier(control(), 4510 memory(raw_adr_type), 4511 alloc_obj, 4512 no_particular_field, 4513 raw_adr_idx, 4514 no_particular_value, 4515 T_OBJECT, 4516 false); 4517 } 4518 4519 // Do not let reads from the cloned object float above the arraycopy. 4520 if (alloc != NULL) { 4521 // Do not let stores that initialize this object be reordered with 4522 // a subsequent store that would make this object accessible by 4523 // other threads. 4524 // Record what AllocateNode this StoreStore protects so that 4525 // escape analysis can go from the MemBarStoreStoreNode to the 4526 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4527 // based on the escape status of the AllocateNode. 4528 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4529 } else { 4530 insert_mem_bar(Op_MemBarCPUOrder); 4531 } 4532 } 4533 4534 //------------------------inline_native_clone---------------------------- 4535 // protected native Object java.lang.Object.clone(); 4536 // 4537 // Here are the simple edge cases: 4538 // null receiver => normal trap 4539 // virtual and clone was overridden => slow path to out-of-line clone 4540 // not cloneable or finalizer => slow path to out-of-line Object.clone 4541 // 4542 // The general case has two steps, allocation and copying. 4543 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4544 // 4545 // Copying also has two cases, oop arrays and everything else. 4546 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4547 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4548 // 4549 // These steps fold up nicely if and when the cloned object's klass 4550 // can be sharply typed as an object array, a type array, or an instance. 4551 // 4552 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4553 PhiNode* result_val; 4554 4555 // Set the reexecute bit for the interpreter to reexecute 4556 // the bytecode that invokes Object.clone if deoptimization happens. 4557 { PreserveReexecuteState preexecs(this); 4558 jvms()->set_should_reexecute(true); 4559 4560 Node* obj = null_check_receiver(); 4561 if (stopped()) return true; 4562 4563 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4564 4565 // If we are going to clone an instance, we need its exact type to 4566 // know the number and types of fields to convert the clone to 4567 // loads/stores. Maybe a speculative type can help us. 4568 if (!obj_type->klass_is_exact() && 4569 obj_type->speculative_type() != NULL && 4570 obj_type->speculative_type()->is_instance_klass()) { 4571 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4572 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4573 !spec_ik->has_injected_fields()) { 4574 ciKlass* k = obj_type->klass(); 4575 if (!k->is_instance_klass() || 4576 k->as_instance_klass()->is_interface() || 4577 k->as_instance_klass()->has_subklass()) { 4578 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4579 } 4580 } 4581 } 4582 4583 Node* obj_klass = load_object_klass(obj); 4584 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4585 const TypeOopPtr* toop = ((tklass != NULL) 4586 ? tklass->as_instance_type() 4587 : TypeInstPtr::NOTNULL); 4588 4589 // Conservatively insert a memory barrier on all memory slices. 4590 // Do not let writes into the original float below the clone. 4591 insert_mem_bar(Op_MemBarCPUOrder); 4592 4593 // paths into result_reg: 4594 enum { 4595 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4596 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4597 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4598 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4599 PATH_LIMIT 4600 }; 4601 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4602 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4603 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4604 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4605 record_for_igvn(result_reg); 4606 4607 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4608 int raw_adr_idx = Compile::AliasIdxRaw; 4609 4610 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4611 if (array_ctl != NULL) { 4612 // It's an array. 4613 PreserveJVMState pjvms(this); 4614 set_control(array_ctl); 4615 Node* obj_length = load_array_length(obj); 4616 Node* obj_size = NULL; 4617 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4618 4619 if (!use_ReduceInitialCardMarks()) { 4620 // If it is an oop array, it requires very special treatment, 4621 // because card marking is required on each card of the array. 4622 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4623 if (is_obja != NULL) { 4624 PreserveJVMState pjvms2(this); 4625 set_control(is_obja); 4626 // Generate a direct call to the right arraycopy function(s). 4627 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4628 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4629 ac->set_cloneoop(); 4630 Node* n = _gvn.transform(ac); 4631 assert(n == ac, "cannot disappear"); 4632 ac->connect_outputs(this); 4633 4634 result_reg->init_req(_objArray_path, control()); 4635 result_val->init_req(_objArray_path, alloc_obj); 4636 result_i_o ->set_req(_objArray_path, i_o()); 4637 result_mem ->set_req(_objArray_path, reset_memory()); 4638 } 4639 } 4640 // Otherwise, there are no card marks to worry about. 4641 // (We can dispense with card marks if we know the allocation 4642 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4643 // causes the non-eden paths to take compensating steps to 4644 // simulate a fresh allocation, so that no further 4645 // card marks are required in compiled code to initialize 4646 // the object.) 4647 4648 if (!stopped()) { 4649 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4650 4651 // Present the results of the copy. 4652 result_reg->init_req(_array_path, control()); 4653 result_val->init_req(_array_path, alloc_obj); 4654 result_i_o ->set_req(_array_path, i_o()); 4655 result_mem ->set_req(_array_path, reset_memory()); 4656 } 4657 } 4658 4659 // We only go to the instance fast case code if we pass a number of guards. 4660 // The paths which do not pass are accumulated in the slow_region. 4661 RegionNode* slow_region = new RegionNode(1); 4662 record_for_igvn(slow_region); 4663 if (!stopped()) { 4664 // It's an instance (we did array above). Make the slow-path tests. 4665 // If this is a virtual call, we generate a funny guard. We grab 4666 // the vtable entry corresponding to clone() from the target object. 4667 // If the target method which we are calling happens to be the 4668 // Object clone() method, we pass the guard. We do not need this 4669 // guard for non-virtual calls; the caller is known to be the native 4670 // Object clone(). 4671 if (is_virtual) { 4672 generate_virtual_guard(obj_klass, slow_region); 4673 } 4674 4675 // The object must be cloneable and must not have a finalizer. 4676 // Both of these conditions may be checked in a single test. 4677 // We could optimize the cloneable test further, but we don't care. 4678 generate_access_flags_guard(obj_klass, 4679 // Test both conditions: 4680 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4681 // Must be cloneable but not finalizer: 4682 JVM_ACC_IS_CLONEABLE, 4683 slow_region); 4684 } 4685 4686 if (!stopped()) { 4687 // It's an instance, and it passed the slow-path tests. 4688 PreserveJVMState pjvms(this); 4689 Node* obj_size = NULL; 4690 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4691 // is reexecuted if deoptimization occurs and there could be problems when merging 4692 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4693 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4694 4695 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4696 4697 // Present the results of the slow call. 4698 result_reg->init_req(_instance_path, control()); 4699 result_val->init_req(_instance_path, alloc_obj); 4700 result_i_o ->set_req(_instance_path, i_o()); 4701 result_mem ->set_req(_instance_path, reset_memory()); 4702 } 4703 4704 // Generate code for the slow case. We make a call to clone(). 4705 set_control(_gvn.transform(slow_region)); 4706 if (!stopped()) { 4707 PreserveJVMState pjvms(this); 4708 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4709 Node* slow_result = set_results_for_java_call(slow_call); 4710 // this->control() comes from set_results_for_java_call 4711 result_reg->init_req(_slow_path, control()); 4712 result_val->init_req(_slow_path, slow_result); 4713 result_i_o ->set_req(_slow_path, i_o()); 4714 result_mem ->set_req(_slow_path, reset_memory()); 4715 } 4716 4717 // Return the combined state. 4718 set_control( _gvn.transform(result_reg)); 4719 set_i_o( _gvn.transform(result_i_o)); 4720 set_all_memory( _gvn.transform(result_mem)); 4721 } // original reexecute is set back here 4722 4723 set_result(_gvn.transform(result_val)); 4724 return true; 4725 } 4726 4727 // If we have a tighly coupled allocation, the arraycopy may take care 4728 // of the array initialization. If one of the guards we insert between 4729 // the allocation and the arraycopy causes a deoptimization, an 4730 // unitialized array will escape the compiled method. To prevent that 4731 // we set the JVM state for uncommon traps between the allocation and 4732 // the arraycopy to the state before the allocation so, in case of 4733 // deoptimization, we'll reexecute the allocation and the 4734 // initialization. 4735 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4736 if (alloc != NULL) { 4737 ciMethod* trap_method = alloc->jvms()->method(); 4738 int trap_bci = alloc->jvms()->bci(); 4739 4740 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4741 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4742 // Make sure there's no store between the allocation and the 4743 // arraycopy otherwise visible side effects could be rexecuted 4744 // in case of deoptimization and cause incorrect execution. 4745 bool no_interfering_store = true; 4746 Node* mem = alloc->in(TypeFunc::Memory); 4747 if (mem->is_MergeMem()) { 4748 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4749 Node* n = mms.memory(); 4750 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4751 assert(n->is_Store(), "what else?"); 4752 no_interfering_store = false; 4753 break; 4754 } 4755 } 4756 } else { 4757 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4758 Node* n = mms.memory(); 4759 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4760 assert(n->is_Store(), "what else?"); 4761 no_interfering_store = false; 4762 break; 4763 } 4764 } 4765 } 4766 4767 if (no_interfering_store) { 4768 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4769 uint size = alloc->req(); 4770 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4771 old_jvms->set_map(sfpt); 4772 for (uint i = 0; i < size; i++) { 4773 sfpt->init_req(i, alloc->in(i)); 4774 } 4775 // re-push array length for deoptimization 4776 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4777 old_jvms->set_sp(old_jvms->sp()+1); 4778 old_jvms->set_monoff(old_jvms->monoff()+1); 4779 old_jvms->set_scloff(old_jvms->scloff()+1); 4780 old_jvms->set_endoff(old_jvms->endoff()+1); 4781 old_jvms->set_should_reexecute(true); 4782 4783 sfpt->set_i_o(map()->i_o()); 4784 sfpt->set_memory(map()->memory()); 4785 sfpt->set_control(map()->control()); 4786 4787 JVMState* saved_jvms = jvms(); 4788 saved_reexecute_sp = _reexecute_sp; 4789 4790 set_jvms(sfpt->jvms()); 4791 _reexecute_sp = jvms()->sp(); 4792 4793 return saved_jvms; 4794 } 4795 } 4796 } 4797 return NULL; 4798 } 4799 4800 // In case of a deoptimization, we restart execution at the 4801 // allocation, allocating a new array. We would leave an uninitialized 4802 // array in the heap that GCs wouldn't expect. Move the allocation 4803 // after the traps so we don't allocate the array if we 4804 // deoptimize. This is possible because tightly_coupled_allocation() 4805 // guarantees there's no observer of the allocated array at this point 4806 // and the control flow is simple enough. 4807 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4808 if (saved_jvms != NULL && !stopped()) { 4809 assert(alloc != NULL, "only with a tightly coupled allocation"); 4810 // restore JVM state to the state at the arraycopy 4811 saved_jvms->map()->set_control(map()->control()); 4812 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4813 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4814 // If we've improved the types of some nodes (null check) while 4815 // emitting the guards, propagate them to the current state 4816 map()->replaced_nodes().apply(saved_jvms->map()); 4817 set_jvms(saved_jvms); 4818 _reexecute_sp = saved_reexecute_sp; 4819 4820 // Remove the allocation from above the guards 4821 CallProjections callprojs; 4822 alloc->extract_projections(&callprojs, true); 4823 InitializeNode* init = alloc->initialization(); 4824 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4825 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4826 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4827 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4828 4829 // move the allocation here (after the guards) 4830 _gvn.hash_delete(alloc); 4831 alloc->set_req(TypeFunc::Control, control()); 4832 alloc->set_req(TypeFunc::I_O, i_o()); 4833 Node *mem = reset_memory(); 4834 set_all_memory(mem); 4835 alloc->set_req(TypeFunc::Memory, mem); 4836 set_control(init->proj_out(TypeFunc::Control)); 4837 set_i_o(callprojs.fallthrough_ioproj); 4838 4839 // Update memory as done in GraphKit::set_output_for_allocation() 4840 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4841 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4842 if (ary_type->isa_aryptr() && length_type != NULL) { 4843 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4844 } 4845 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4846 int elemidx = C->get_alias_index(telemref); 4847 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4848 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4849 4850 Node* allocx = _gvn.transform(alloc); 4851 assert(allocx == alloc, "where has the allocation gone?"); 4852 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4853 4854 _gvn.hash_delete(dest); 4855 dest->set_req(0, control()); 4856 Node* destx = _gvn.transform(dest); 4857 assert(destx == dest, "where has the allocation result gone?"); 4858 } 4859 } 4860 4861 4862 //------------------------------inline_arraycopy----------------------- 4863 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4864 // Object dest, int destPos, 4865 // int length); 4866 bool LibraryCallKit::inline_arraycopy() { 4867 // Get the arguments. 4868 Node* src = argument(0); // type: oop 4869 Node* src_offset = argument(1); // type: int 4870 Node* dest = argument(2); // type: oop 4871 Node* dest_offset = argument(3); // type: int 4872 Node* length = argument(4); // type: int 4873 4874 4875 // Check for allocation before we add nodes that would confuse 4876 // tightly_coupled_allocation() 4877 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4878 4879 int saved_reexecute_sp = -1; 4880 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4881 // See arraycopy_restore_alloc_state() comment 4882 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4883 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4884 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4885 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4886 4887 // The following tests must be performed 4888 // (1) src and dest are arrays. 4889 // (2) src and dest arrays must have elements of the same BasicType 4890 // (3) src and dest must not be null. 4891 // (4) src_offset must not be negative. 4892 // (5) dest_offset must not be negative. 4893 // (6) length must not be negative. 4894 // (7) src_offset + length must not exceed length of src. 4895 // (8) dest_offset + length must not exceed length of dest. 4896 // (9) each element of an oop array must be assignable 4897 4898 // (3) src and dest must not be null. 4899 // always do this here because we need the JVM state for uncommon traps 4900 Node* null_ctl = top(); 4901 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4902 assert(null_ctl->is_top(), "no null control here"); 4903 dest = null_check(dest, T_ARRAY); 4904 4905 if (!can_emit_guards) { 4906 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4907 // guards but the arraycopy node could still take advantage of a 4908 // tightly allocated allocation. tightly_coupled_allocation() is 4909 // called again to make sure it takes the null check above into 4910 // account: the null check is mandatory and if it caused an 4911 // uncommon trap to be emitted then the allocation can't be 4912 // considered tightly coupled in this context. 4913 alloc = tightly_coupled_allocation(dest, NULL); 4914 } 4915 4916 bool validated = false; 4917 4918 const Type* src_type = _gvn.type(src); 4919 const Type* dest_type = _gvn.type(dest); 4920 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4921 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4922 4923 // Do we have the type of src? 4924 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4925 // Do we have the type of dest? 4926 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4927 // Is the type for src from speculation? 4928 bool src_spec = false; 4929 // Is the type for dest from speculation? 4930 bool dest_spec = false; 4931 4932 if ((!has_src || !has_dest) && can_emit_guards) { 4933 // We don't have sufficient type information, let's see if 4934 // speculative types can help. We need to have types for both src 4935 // and dest so that it pays off. 4936 4937 // Do we already have or could we have type information for src 4938 bool could_have_src = has_src; 4939 // Do we already have or could we have type information for dest 4940 bool could_have_dest = has_dest; 4941 4942 ciKlass* src_k = NULL; 4943 if (!has_src) { 4944 src_k = src_type->speculative_type_not_null(); 4945 if (src_k != NULL && src_k->is_array_klass()) { 4946 could_have_src = true; 4947 } 4948 } 4949 4950 ciKlass* dest_k = NULL; 4951 if (!has_dest) { 4952 dest_k = dest_type->speculative_type_not_null(); 4953 if (dest_k != NULL && dest_k->is_array_klass()) { 4954 could_have_dest = true; 4955 } 4956 } 4957 4958 if (could_have_src && could_have_dest) { 4959 // This is going to pay off so emit the required guards 4960 if (!has_src) { 4961 src = maybe_cast_profiled_obj(src, src_k, true); 4962 src_type = _gvn.type(src); 4963 top_src = src_type->isa_aryptr(); 4964 has_src = (top_src != NULL && top_src->klass() != NULL); 4965 src_spec = true; 4966 } 4967 if (!has_dest) { 4968 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4969 dest_type = _gvn.type(dest); 4970 top_dest = dest_type->isa_aryptr(); 4971 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4972 dest_spec = true; 4973 } 4974 } 4975 } 4976 4977 if (has_src && has_dest && can_emit_guards) { 4978 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4979 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4980 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4981 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4982 4983 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4984 // If both arrays are object arrays then having the exact types 4985 // for both will remove the need for a subtype check at runtime 4986 // before the call and may make it possible to pick a faster copy 4987 // routine (without a subtype check on every element) 4988 // Do we have the exact type of src? 4989 bool could_have_src = src_spec; 4990 // Do we have the exact type of dest? 4991 bool could_have_dest = dest_spec; 4992 ciKlass* src_k = top_src->klass(); 4993 ciKlass* dest_k = top_dest->klass(); 4994 if (!src_spec) { 4995 src_k = src_type->speculative_type_not_null(); 4996 if (src_k != NULL && src_k->is_array_klass()) { 4997 could_have_src = true; 4998 } 4999 } 5000 if (!dest_spec) { 5001 dest_k = dest_type->speculative_type_not_null(); 5002 if (dest_k != NULL && dest_k->is_array_klass()) { 5003 could_have_dest = true; 5004 } 5005 } 5006 if (could_have_src && could_have_dest) { 5007 // If we can have both exact types, emit the missing guards 5008 if (could_have_src && !src_spec) { 5009 src = maybe_cast_profiled_obj(src, src_k, true); 5010 } 5011 if (could_have_dest && !dest_spec) { 5012 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5013 } 5014 } 5015 } 5016 } 5017 5018 ciMethod* trap_method = method(); 5019 int trap_bci = bci(); 5020 if (saved_jvms != NULL) { 5021 trap_method = alloc->jvms()->method(); 5022 trap_bci = alloc->jvms()->bci(); 5023 } 5024 5025 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5026 can_emit_guards && 5027 !src->is_top() && !dest->is_top()) { 5028 // validate arguments: enables transformation the ArrayCopyNode 5029 validated = true; 5030 5031 RegionNode* slow_region = new RegionNode(1); 5032 record_for_igvn(slow_region); 5033 5034 // (1) src and dest are arrays. 5035 generate_non_array_guard(load_object_klass(src), slow_region); 5036 generate_non_array_guard(load_object_klass(dest), slow_region); 5037 5038 // (2) src and dest arrays must have elements of the same BasicType 5039 // done at macro expansion or at Ideal transformation time 5040 5041 // (4) src_offset must not be negative. 5042 generate_negative_guard(src_offset, slow_region); 5043 5044 // (5) dest_offset must not be negative. 5045 generate_negative_guard(dest_offset, slow_region); 5046 5047 // (7) src_offset + length must not exceed length of src. 5048 generate_limit_guard(src_offset, length, 5049 load_array_length(src), 5050 slow_region); 5051 5052 // (8) dest_offset + length must not exceed length of dest. 5053 generate_limit_guard(dest_offset, length, 5054 load_array_length(dest), 5055 slow_region); 5056 5057 // (9) each element of an oop array must be assignable 5058 Node* src_klass = load_object_klass(src); 5059 Node* dest_klass = load_object_klass(dest); 5060 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5061 5062 if (not_subtype_ctrl != top()) { 5063 PreserveJVMState pjvms(this); 5064 set_control(not_subtype_ctrl); 5065 uncommon_trap(Deoptimization::Reason_intrinsic, 5066 Deoptimization::Action_make_not_entrant); 5067 assert(stopped(), "Should be stopped"); 5068 } 5069 { 5070 PreserveJVMState pjvms(this); 5071 set_control(_gvn.transform(slow_region)); 5072 uncommon_trap(Deoptimization::Reason_intrinsic, 5073 Deoptimization::Action_make_not_entrant); 5074 assert(stopped(), "Should be stopped"); 5075 } 5076 } 5077 5078 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5079 5080 if (stopped()) { 5081 return true; 5082 } 5083 5084 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5085 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5086 // so the compiler has a chance to eliminate them: during macro expansion, 5087 // we have to set their control (CastPP nodes are eliminated). 5088 load_object_klass(src), load_object_klass(dest), 5089 load_array_length(src), load_array_length(dest)); 5090 5091 ac->set_arraycopy(validated); 5092 5093 Node* n = _gvn.transform(ac); 5094 if (n == ac) { 5095 ac->connect_outputs(this); 5096 } else { 5097 assert(validated, "shouldn't transform if all arguments not validated"); 5098 set_all_memory(n); 5099 } 5100 5101 return true; 5102 } 5103 5104 5105 // Helper function which determines if an arraycopy immediately follows 5106 // an allocation, with no intervening tests or other escapes for the object. 5107 AllocateArrayNode* 5108 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5109 RegionNode* slow_region) { 5110 if (stopped()) return NULL; // no fast path 5111 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5112 5113 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5114 if (alloc == NULL) return NULL; 5115 5116 Node* rawmem = memory(Compile::AliasIdxRaw); 5117 // Is the allocation's memory state untouched? 5118 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5119 // Bail out if there have been raw-memory effects since the allocation. 5120 // (Example: There might have been a call or safepoint.) 5121 return NULL; 5122 } 5123 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5124 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5125 return NULL; 5126 } 5127 5128 // There must be no unexpected observers of this allocation. 5129 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5130 Node* obs = ptr->fast_out(i); 5131 if (obs != this->map()) { 5132 return NULL; 5133 } 5134 } 5135 5136 // This arraycopy must unconditionally follow the allocation of the ptr. 5137 Node* alloc_ctl = ptr->in(0); 5138 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5139 5140 Node* ctl = control(); 5141 while (ctl != alloc_ctl) { 5142 // There may be guards which feed into the slow_region. 5143 // Any other control flow means that we might not get a chance 5144 // to finish initializing the allocated object. 5145 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5146 IfNode* iff = ctl->in(0)->as_If(); 5147 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5148 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5149 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5150 ctl = iff->in(0); // This test feeds the known slow_region. 5151 continue; 5152 } 5153 // One more try: Various low-level checks bottom out in 5154 // uncommon traps. If the debug-info of the trap omits 5155 // any reference to the allocation, as we've already 5156 // observed, then there can be no objection to the trap. 5157 bool found_trap = false; 5158 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5159 Node* obs = not_ctl->fast_out(j); 5160 if (obs->in(0) == not_ctl && obs->is_Call() && 5161 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5162 found_trap = true; break; 5163 } 5164 } 5165 if (found_trap) { 5166 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5167 continue; 5168 } 5169 } 5170 return NULL; 5171 } 5172 5173 // If we get this far, we have an allocation which immediately 5174 // precedes the arraycopy, and we can take over zeroing the new object. 5175 // The arraycopy will finish the initialization, and provide 5176 // a new control state to which we will anchor the destination pointer. 5177 5178 return alloc; 5179 } 5180 5181 //-------------inline_encodeISOArray----------------------------------- 5182 // encode char[] to byte[] in ISO_8859_1 5183 bool LibraryCallKit::inline_encodeISOArray() { 5184 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5185 // no receiver since it is static method 5186 Node *src = argument(0); 5187 Node *src_offset = argument(1); 5188 Node *dst = argument(2); 5189 Node *dst_offset = argument(3); 5190 Node *length = argument(4); 5191 5192 const Type* src_type = src->Value(&_gvn); 5193 const Type* dst_type = dst->Value(&_gvn); 5194 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5195 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5196 if (top_src == NULL || top_src->klass() == NULL || 5197 top_dest == NULL || top_dest->klass() == NULL) { 5198 // failed array check 5199 return false; 5200 } 5201 5202 // Figure out the size and type of the elements we will be copying. 5203 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5204 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5205 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5206 return false; 5207 } 5208 Node* src_start = array_element_address(src, src_offset, src_elem); 5209 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5210 // 'src_start' points to src array + scaled offset 5211 // 'dst_start' points to dst array + scaled offset 5212 5213 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5214 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5215 enc = _gvn.transform(enc); 5216 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5217 set_memory(res_mem, mtype); 5218 set_result(enc); 5219 return true; 5220 } 5221 5222 //-------------inline_multiplyToLen----------------------------------- 5223 bool LibraryCallKit::inline_multiplyToLen() { 5224 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5225 5226 address stubAddr = StubRoutines::multiplyToLen(); 5227 if (stubAddr == NULL) { 5228 return false; // Intrinsic's stub is not implemented on this platform 5229 } 5230 const char* stubName = "multiplyToLen"; 5231 5232 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5233 5234 Node* x = argument(1); 5235 Node* xlen = argument(2); 5236 Node* y = argument(3); 5237 Node* ylen = argument(4); 5238 Node* z = argument(5); 5239 5240 const Type* x_type = x->Value(&_gvn); 5241 const Type* y_type = y->Value(&_gvn); 5242 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5243 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5244 if (top_x == NULL || top_x->klass() == NULL || 5245 top_y == NULL || top_y->klass() == NULL) { 5246 // failed array check 5247 return false; 5248 } 5249 5250 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5251 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5252 if (x_elem != T_INT || y_elem != T_INT) { 5253 return false; 5254 } 5255 5256 // Set the original stack and the reexecute bit for the interpreter to reexecute 5257 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5258 // on the return from z array allocation in runtime. 5259 { PreserveReexecuteState preexecs(this); 5260 jvms()->set_should_reexecute(true); 5261 5262 Node* x_start = array_element_address(x, intcon(0), x_elem); 5263 Node* y_start = array_element_address(y, intcon(0), y_elem); 5264 // 'x_start' points to x array + scaled xlen 5265 // 'y_start' points to y array + scaled ylen 5266 5267 // Allocate the result array 5268 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5269 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5270 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5271 5272 IdealKit ideal(this); 5273 5274 #define __ ideal. 5275 Node* one = __ ConI(1); 5276 Node* zero = __ ConI(0); 5277 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5278 __ set(need_alloc, zero); 5279 __ set(z_alloc, z); 5280 __ if_then(z, BoolTest::eq, null()); { 5281 __ increment (need_alloc, one); 5282 } __ else_(); { 5283 // Update graphKit memory and control from IdealKit. 5284 sync_kit(ideal); 5285 Node* zlen_arg = load_array_length(z); 5286 // Update IdealKit memory and control from graphKit. 5287 __ sync_kit(this); 5288 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5289 __ increment (need_alloc, one); 5290 } __ end_if(); 5291 } __ end_if(); 5292 5293 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5294 // Update graphKit memory and control from IdealKit. 5295 sync_kit(ideal); 5296 Node * narr = new_array(klass_node, zlen, 1); 5297 // Update IdealKit memory and control from graphKit. 5298 __ sync_kit(this); 5299 __ set(z_alloc, narr); 5300 } __ end_if(); 5301 5302 sync_kit(ideal); 5303 z = __ value(z_alloc); 5304 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5305 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5306 // Final sync IdealKit and GraphKit. 5307 final_sync(ideal); 5308 #undef __ 5309 5310 Node* z_start = array_element_address(z, intcon(0), T_INT); 5311 5312 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5313 OptoRuntime::multiplyToLen_Type(), 5314 stubAddr, stubName, TypePtr::BOTTOM, 5315 x_start, xlen, y_start, ylen, z_start, zlen); 5316 } // original reexecute is set back here 5317 5318 C->set_has_split_ifs(true); // Has chance for split-if optimization 5319 set_result(z); 5320 return true; 5321 } 5322 5323 //-------------inline_squareToLen------------------------------------ 5324 bool LibraryCallKit::inline_squareToLen() { 5325 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5326 5327 address stubAddr = StubRoutines::squareToLen(); 5328 if (stubAddr == NULL) { 5329 return false; // Intrinsic's stub is not implemented on this platform 5330 } 5331 const char* stubName = "squareToLen"; 5332 5333 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5334 5335 Node* x = argument(0); 5336 Node* len = argument(1); 5337 Node* z = argument(2); 5338 Node* zlen = argument(3); 5339 5340 const Type* x_type = x->Value(&_gvn); 5341 const Type* z_type = z->Value(&_gvn); 5342 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5343 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5344 if (top_x == NULL || top_x->klass() == NULL || 5345 top_z == NULL || top_z->klass() == NULL) { 5346 // failed array check 5347 return false; 5348 } 5349 5350 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5351 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5352 if (x_elem != T_INT || z_elem != T_INT) { 5353 return false; 5354 } 5355 5356 5357 Node* x_start = array_element_address(x, intcon(0), x_elem); 5358 Node* z_start = array_element_address(z, intcon(0), z_elem); 5359 5360 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5361 OptoRuntime::squareToLen_Type(), 5362 stubAddr, stubName, TypePtr::BOTTOM, 5363 x_start, len, z_start, zlen); 5364 5365 set_result(z); 5366 return true; 5367 } 5368 5369 //-------------inline_mulAdd------------------------------------------ 5370 bool LibraryCallKit::inline_mulAdd() { 5371 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5372 5373 address stubAddr = StubRoutines::mulAdd(); 5374 if (stubAddr == NULL) { 5375 return false; // Intrinsic's stub is not implemented on this platform 5376 } 5377 const char* stubName = "mulAdd"; 5378 5379 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5380 5381 Node* out = argument(0); 5382 Node* in = argument(1); 5383 Node* offset = argument(2); 5384 Node* len = argument(3); 5385 Node* k = argument(4); 5386 5387 const Type* out_type = out->Value(&_gvn); 5388 const Type* in_type = in->Value(&_gvn); 5389 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5390 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5391 if (top_out == NULL || top_out->klass() == NULL || 5392 top_in == NULL || top_in->klass() == NULL) { 5393 // failed array check 5394 return false; 5395 } 5396 5397 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5398 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5399 if (out_elem != T_INT || in_elem != T_INT) { 5400 return false; 5401 } 5402 5403 Node* outlen = load_array_length(out); 5404 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5405 Node* out_start = array_element_address(out, intcon(0), out_elem); 5406 Node* in_start = array_element_address(in, intcon(0), in_elem); 5407 5408 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5409 OptoRuntime::mulAdd_Type(), 5410 stubAddr, stubName, TypePtr::BOTTOM, 5411 out_start,in_start, new_offset, len, k); 5412 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5413 set_result(result); 5414 return true; 5415 } 5416 5417 5418 /** 5419 * Calculate CRC32 for byte. 5420 * int java.util.zip.CRC32.update(int crc, int b) 5421 */ 5422 bool LibraryCallKit::inline_updateCRC32() { 5423 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5424 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5425 // no receiver since it is static method 5426 Node* crc = argument(0); // type: int 5427 Node* b = argument(1); // type: int 5428 5429 /* 5430 * int c = ~ crc; 5431 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5432 * b = b ^ (c >>> 8); 5433 * crc = ~b; 5434 */ 5435 5436 Node* M1 = intcon(-1); 5437 crc = _gvn.transform(new XorINode(crc, M1)); 5438 Node* result = _gvn.transform(new XorINode(crc, b)); 5439 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5440 5441 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5442 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5443 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5444 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5445 5446 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5447 result = _gvn.transform(new XorINode(crc, result)); 5448 result = _gvn.transform(new XorINode(result, M1)); 5449 set_result(result); 5450 return true; 5451 } 5452 5453 /** 5454 * Calculate CRC32 for byte[] array. 5455 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5456 */ 5457 bool LibraryCallKit::inline_updateBytesCRC32() { 5458 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5459 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5460 // no receiver since it is static method 5461 Node* crc = argument(0); // type: int 5462 Node* src = argument(1); // type: oop 5463 Node* offset = argument(2); // type: int 5464 Node* length = argument(3); // type: int 5465 5466 const Type* src_type = src->Value(&_gvn); 5467 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5468 if (top_src == NULL || top_src->klass() == NULL) { 5469 // failed array check 5470 return false; 5471 } 5472 5473 // Figure out the size and type of the elements we will be copying. 5474 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5475 if (src_elem != T_BYTE) { 5476 return false; 5477 } 5478 5479 // 'src_start' points to src array + scaled offset 5480 Node* src_start = array_element_address(src, offset, src_elem); 5481 5482 // We assume that range check is done by caller. 5483 // TODO: generate range check (offset+length < src.length) in debug VM. 5484 5485 // Call the stub. 5486 address stubAddr = StubRoutines::updateBytesCRC32(); 5487 const char *stubName = "updateBytesCRC32"; 5488 5489 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5490 stubAddr, stubName, TypePtr::BOTTOM, 5491 crc, src_start, length); 5492 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5493 set_result(result); 5494 return true; 5495 } 5496 5497 /** 5498 * Calculate CRC32 for ByteBuffer. 5499 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5500 */ 5501 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5502 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5503 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5504 // no receiver since it is static method 5505 Node* crc = argument(0); // type: int 5506 Node* src = argument(1); // type: long 5507 Node* offset = argument(3); // type: int 5508 Node* length = argument(4); // type: int 5509 5510 src = ConvL2X(src); // adjust Java long to machine word 5511 Node* base = _gvn.transform(new CastX2PNode(src)); 5512 offset = ConvI2X(offset); 5513 5514 // 'src_start' points to src array + scaled offset 5515 Node* src_start = basic_plus_adr(top(), base, offset); 5516 5517 // Call the stub. 5518 address stubAddr = StubRoutines::updateBytesCRC32(); 5519 const char *stubName = "updateBytesCRC32"; 5520 5521 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5522 stubAddr, stubName, TypePtr::BOTTOM, 5523 crc, src_start, length); 5524 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5525 set_result(result); 5526 return true; 5527 } 5528 5529 //----------------------------inline_reference_get---------------------------- 5530 // public T java.lang.ref.Reference.get(); 5531 bool LibraryCallKit::inline_reference_get() { 5532 const int referent_offset = java_lang_ref_Reference::referent_offset; 5533 guarantee(referent_offset > 0, "should have already been set"); 5534 5535 // Get the argument: 5536 Node* reference_obj = null_check_receiver(); 5537 if (stopped()) return true; 5538 5539 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5540 5541 ciInstanceKlass* klass = env()->Object_klass(); 5542 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5543 5544 Node* no_ctrl = NULL; 5545 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5546 5547 // Use the pre-barrier to record the value in the referent field 5548 pre_barrier(false /* do_load */, 5549 control(), 5550 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5551 result /* pre_val */, 5552 T_OBJECT); 5553 5554 // Add memory barrier to prevent commoning reads from this field 5555 // across safepoint since GC can change its value. 5556 insert_mem_bar(Op_MemBarCPUOrder); 5557 5558 set_result(result); 5559 return true; 5560 } 5561 5562 5563 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5564 bool is_exact=true, bool is_static=false) { 5565 5566 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5567 assert(tinst != NULL, "obj is null"); 5568 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5569 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5570 5571 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5572 ciSymbol::make(fieldTypeString), 5573 is_static); 5574 if (field == NULL) return (Node *) NULL; 5575 assert (field != NULL, "undefined field"); 5576 5577 // Next code copied from Parse::do_get_xxx(): 5578 5579 // Compute address and memory type. 5580 int offset = field->offset_in_bytes(); 5581 bool is_vol = field->is_volatile(); 5582 ciType* field_klass = field->type(); 5583 assert(field_klass->is_loaded(), "should be loaded"); 5584 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5585 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5586 BasicType bt = field->layout_type(); 5587 5588 // Build the resultant type of the load 5589 const Type *type; 5590 if (bt == T_OBJECT) { 5591 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5592 } else { 5593 type = Type::get_const_basic_type(bt); 5594 } 5595 5596 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5597 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5598 } 5599 // Build the load. 5600 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5601 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol); 5602 // If reference is volatile, prevent following memory ops from 5603 // floating up past the volatile read. Also prevents commoning 5604 // another volatile read. 5605 if (is_vol) { 5606 // Memory barrier includes bogus read of value to force load BEFORE membar 5607 insert_mem_bar(Op_MemBarAcquire, loadedField); 5608 } 5609 return loadedField; 5610 } 5611 5612 5613 //------------------------------inline_aescrypt_Block----------------------- 5614 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5615 address stubAddr; 5616 const char *stubName; 5617 assert(UseAES, "need AES instruction support"); 5618 5619 switch(id) { 5620 case vmIntrinsics::_aescrypt_encryptBlock: 5621 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5622 stubName = "aescrypt_encryptBlock"; 5623 break; 5624 case vmIntrinsics::_aescrypt_decryptBlock: 5625 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5626 stubName = "aescrypt_decryptBlock"; 5627 break; 5628 } 5629 if (stubAddr == NULL) return false; 5630 5631 Node* aescrypt_object = argument(0); 5632 Node* src = argument(1); 5633 Node* src_offset = argument(2); 5634 Node* dest = argument(3); 5635 Node* dest_offset = argument(4); 5636 5637 // (1) src and dest are arrays. 5638 const Type* src_type = src->Value(&_gvn); 5639 const Type* dest_type = dest->Value(&_gvn); 5640 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5641 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5642 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5643 5644 // for the quick and dirty code we will skip all the checks. 5645 // we are just trying to get the call to be generated. 5646 Node* src_start = src; 5647 Node* dest_start = dest; 5648 if (src_offset != NULL || dest_offset != NULL) { 5649 assert(src_offset != NULL && dest_offset != NULL, ""); 5650 src_start = array_element_address(src, src_offset, T_BYTE); 5651 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5652 } 5653 5654 // now need to get the start of its expanded key array 5655 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5656 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5657 if (k_start == NULL) return false; 5658 5659 if (Matcher::pass_original_key_for_aes()) { 5660 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5661 // compatibility issues between Java key expansion and SPARC crypto instructions 5662 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5663 if (original_k_start == NULL) return false; 5664 5665 // Call the stub. 5666 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5667 stubAddr, stubName, TypePtr::BOTTOM, 5668 src_start, dest_start, k_start, original_k_start); 5669 } else { 5670 // Call the stub. 5671 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5672 stubAddr, stubName, TypePtr::BOTTOM, 5673 src_start, dest_start, k_start); 5674 } 5675 5676 return true; 5677 } 5678 5679 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5680 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5681 address stubAddr; 5682 const char *stubName; 5683 5684 assert(UseAES, "need AES instruction support"); 5685 5686 switch(id) { 5687 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5688 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5689 stubName = "cipherBlockChaining_encryptAESCrypt"; 5690 break; 5691 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5692 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5693 stubName = "cipherBlockChaining_decryptAESCrypt"; 5694 break; 5695 } 5696 if (stubAddr == NULL) return false; 5697 5698 Node* cipherBlockChaining_object = argument(0); 5699 Node* src = argument(1); 5700 Node* src_offset = argument(2); 5701 Node* len = argument(3); 5702 Node* dest = argument(4); 5703 Node* dest_offset = argument(5); 5704 5705 // (1) src and dest are arrays. 5706 const Type* src_type = src->Value(&_gvn); 5707 const Type* dest_type = dest->Value(&_gvn); 5708 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5709 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5710 assert (top_src != NULL && top_src->klass() != NULL 5711 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5712 5713 // checks are the responsibility of the caller 5714 Node* src_start = src; 5715 Node* dest_start = dest; 5716 if (src_offset != NULL || dest_offset != NULL) { 5717 assert(src_offset != NULL && dest_offset != NULL, ""); 5718 src_start = array_element_address(src, src_offset, T_BYTE); 5719 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5720 } 5721 5722 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5723 // (because of the predicated logic executed earlier). 5724 // so we cast it here safely. 5725 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5726 5727 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5728 if (embeddedCipherObj == NULL) return false; 5729 5730 // cast it to what we know it will be at runtime 5731 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5732 assert(tinst != NULL, "CBC obj is null"); 5733 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5734 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5735 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5736 5737 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5738 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5739 const TypeOopPtr* xtype = aklass->as_instance_type(); 5740 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5741 aescrypt_object = _gvn.transform(aescrypt_object); 5742 5743 // we need to get the start of the aescrypt_object's expanded key array 5744 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5745 if (k_start == NULL) return false; 5746 5747 // similarly, get the start address of the r vector 5748 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5749 if (objRvec == NULL) return false; 5750 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5751 5752 Node* cbcCrypt; 5753 if (Matcher::pass_original_key_for_aes()) { 5754 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5755 // compatibility issues between Java key expansion and SPARC crypto instructions 5756 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5757 if (original_k_start == NULL) return false; 5758 5759 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5760 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5761 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5762 stubAddr, stubName, TypePtr::BOTTOM, 5763 src_start, dest_start, k_start, r_start, len, original_k_start); 5764 } else { 5765 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5766 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5767 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5768 stubAddr, stubName, TypePtr::BOTTOM, 5769 src_start, dest_start, k_start, r_start, len); 5770 } 5771 5772 // return cipher length (int) 5773 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5774 set_result(retvalue); 5775 return true; 5776 } 5777 5778 //------------------------------get_key_start_from_aescrypt_object----------------------- 5779 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5780 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5781 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5782 if (objAESCryptKey == NULL) return (Node *) NULL; 5783 5784 // now have the array, need to get the start address of the K array 5785 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5786 return k_start; 5787 } 5788 5789 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5790 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5791 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5792 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5793 if (objAESCryptKey == NULL) return (Node *) NULL; 5794 5795 // now have the array, need to get the start address of the lastKey array 5796 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5797 return original_k_start; 5798 } 5799 5800 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5801 // Return node representing slow path of predicate check. 5802 // the pseudo code we want to emulate with this predicate is: 5803 // for encryption: 5804 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5805 // for decryption: 5806 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5807 // note cipher==plain is more conservative than the original java code but that's OK 5808 // 5809 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5810 // The receiver was checked for NULL already. 5811 Node* objCBC = argument(0); 5812 5813 // Load embeddedCipher field of CipherBlockChaining object. 5814 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5815 5816 // get AESCrypt klass for instanceOf check 5817 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5818 // will have same classloader as CipherBlockChaining object 5819 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5820 assert(tinst != NULL, "CBCobj is null"); 5821 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5822 5823 // we want to do an instanceof comparison against the AESCrypt class 5824 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5825 if (!klass_AESCrypt->is_loaded()) { 5826 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5827 Node* ctrl = control(); 5828 set_control(top()); // no regular fast path 5829 return ctrl; 5830 } 5831 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5832 5833 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5834 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5835 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5836 5837 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5838 5839 // for encryption, we are done 5840 if (!decrypting) 5841 return instof_false; // even if it is NULL 5842 5843 // for decryption, we need to add a further check to avoid 5844 // taking the intrinsic path when cipher and plain are the same 5845 // see the original java code for why. 5846 RegionNode* region = new RegionNode(3); 5847 region->init_req(1, instof_false); 5848 Node* src = argument(1); 5849 Node* dest = argument(4); 5850 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5851 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5852 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5853 region->init_req(2, src_dest_conjoint); 5854 5855 record_for_igvn(region); 5856 return _gvn.transform(region); 5857 } 5858 5859 //------------------------------inline_sha_implCompress----------------------- 5860 // 5861 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 5862 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 5863 // 5864 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 5865 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 5866 // 5867 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 5868 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 5869 // 5870 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 5871 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 5872 5873 Node* sha_obj = argument(0); 5874 Node* src = argument(1); // type oop 5875 Node* ofs = argument(2); // type int 5876 5877 const Type* src_type = src->Value(&_gvn); 5878 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5879 if (top_src == NULL || top_src->klass() == NULL) { 5880 // failed array check 5881 return false; 5882 } 5883 // Figure out the size and type of the elements we will be copying. 5884 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5885 if (src_elem != T_BYTE) { 5886 return false; 5887 } 5888 // 'src_start' points to src array + offset 5889 Node* src_start = array_element_address(src, ofs, src_elem); 5890 Node* state = NULL; 5891 address stubAddr; 5892 const char *stubName; 5893 5894 switch(id) { 5895 case vmIntrinsics::_sha_implCompress: 5896 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 5897 state = get_state_from_sha_object(sha_obj); 5898 stubAddr = StubRoutines::sha1_implCompress(); 5899 stubName = "sha1_implCompress"; 5900 break; 5901 case vmIntrinsics::_sha2_implCompress: 5902 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 5903 state = get_state_from_sha_object(sha_obj); 5904 stubAddr = StubRoutines::sha256_implCompress(); 5905 stubName = "sha256_implCompress"; 5906 break; 5907 case vmIntrinsics::_sha5_implCompress: 5908 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 5909 state = get_state_from_sha5_object(sha_obj); 5910 stubAddr = StubRoutines::sha512_implCompress(); 5911 stubName = "sha512_implCompress"; 5912 break; 5913 default: 5914 fatal_unexpected_iid(id); 5915 return false; 5916 } 5917 if (state == NULL) return false; 5918 5919 // Call the stub. 5920 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 5921 stubAddr, stubName, TypePtr::BOTTOM, 5922 src_start, state); 5923 5924 return true; 5925 } 5926 5927 //------------------------------inline_digestBase_implCompressMB----------------------- 5928 // 5929 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 5930 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 5931 // 5932 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 5933 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5934 "need SHA1/SHA256/SHA512 instruction support"); 5935 assert((uint)predicate < 3, "sanity"); 5936 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 5937 5938 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 5939 Node* src = argument(1); // byte[] array 5940 Node* ofs = argument(2); // type int 5941 Node* limit = argument(3); // type int 5942 5943 const Type* src_type = src->Value(&_gvn); 5944 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5945 if (top_src == NULL || top_src->klass() == NULL) { 5946 // failed array check 5947 return false; 5948 } 5949 // Figure out the size and type of the elements we will be copying. 5950 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5951 if (src_elem != T_BYTE) { 5952 return false; 5953 } 5954 // 'src_start' points to src array + offset 5955 Node* src_start = array_element_address(src, ofs, src_elem); 5956 5957 const char* klass_SHA_name = NULL; 5958 const char* stub_name = NULL; 5959 address stub_addr = NULL; 5960 bool long_state = false; 5961 5962 switch (predicate) { 5963 case 0: 5964 if (UseSHA1Intrinsics) { 5965 klass_SHA_name = "sun/security/provider/SHA"; 5966 stub_name = "sha1_implCompressMB"; 5967 stub_addr = StubRoutines::sha1_implCompressMB(); 5968 } 5969 break; 5970 case 1: 5971 if (UseSHA256Intrinsics) { 5972 klass_SHA_name = "sun/security/provider/SHA2"; 5973 stub_name = "sha256_implCompressMB"; 5974 stub_addr = StubRoutines::sha256_implCompressMB(); 5975 } 5976 break; 5977 case 2: 5978 if (UseSHA512Intrinsics) { 5979 klass_SHA_name = "sun/security/provider/SHA5"; 5980 stub_name = "sha512_implCompressMB"; 5981 stub_addr = StubRoutines::sha512_implCompressMB(); 5982 long_state = true; 5983 } 5984 break; 5985 default: 5986 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5987 } 5988 if (klass_SHA_name != NULL) { 5989 // get DigestBase klass to lookup for SHA klass 5990 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 5991 assert(tinst != NULL, "digestBase_obj is not instance???"); 5992 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5993 5994 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5995 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 5996 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5997 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 5998 } 5999 return false; 6000 } 6001 //------------------------------inline_sha_implCompressMB----------------------- 6002 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6003 bool long_state, address stubAddr, const char *stubName, 6004 Node* src_start, Node* ofs, Node* limit) { 6005 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6006 const TypeOopPtr* xtype = aklass->as_instance_type(); 6007 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6008 sha_obj = _gvn.transform(sha_obj); 6009 6010 Node* state; 6011 if (long_state) { 6012 state = get_state_from_sha5_object(sha_obj); 6013 } else { 6014 state = get_state_from_sha_object(sha_obj); 6015 } 6016 if (state == NULL) return false; 6017 6018 // Call the stub. 6019 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6020 OptoRuntime::digestBase_implCompressMB_Type(), 6021 stubAddr, stubName, TypePtr::BOTTOM, 6022 src_start, state, ofs, limit); 6023 // return ofs (int) 6024 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6025 set_result(result); 6026 6027 return true; 6028 } 6029 6030 //------------------------------get_state_from_sha_object----------------------- 6031 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6032 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6033 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6034 if (sha_state == NULL) return (Node *) NULL; 6035 6036 // now have the array, need to get the start address of the state array 6037 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6038 return state; 6039 } 6040 6041 //------------------------------get_state_from_sha5_object----------------------- 6042 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6043 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6044 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6045 if (sha_state == NULL) return (Node *) NULL; 6046 6047 // now have the array, need to get the start address of the state array 6048 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6049 return state; 6050 } 6051 6052 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6053 // Return node representing slow path of predicate check. 6054 // the pseudo code we want to emulate with this predicate is: 6055 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6056 // 6057 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6058 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6059 "need SHA1/SHA256/SHA512 instruction support"); 6060 assert((uint)predicate < 3, "sanity"); 6061 6062 // The receiver was checked for NULL already. 6063 Node* digestBaseObj = argument(0); 6064 6065 // get DigestBase klass for instanceOf check 6066 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6067 assert(tinst != NULL, "digestBaseObj is null"); 6068 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6069 6070 const char* klass_SHA_name = NULL; 6071 switch (predicate) { 6072 case 0: 6073 if (UseSHA1Intrinsics) { 6074 // we want to do an instanceof comparison against the SHA class 6075 klass_SHA_name = "sun/security/provider/SHA"; 6076 } 6077 break; 6078 case 1: 6079 if (UseSHA256Intrinsics) { 6080 // we want to do an instanceof comparison against the SHA2 class 6081 klass_SHA_name = "sun/security/provider/SHA2"; 6082 } 6083 break; 6084 case 2: 6085 if (UseSHA512Intrinsics) { 6086 // we want to do an instanceof comparison against the SHA5 class 6087 klass_SHA_name = "sun/security/provider/SHA5"; 6088 } 6089 break; 6090 default: 6091 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6092 } 6093 6094 ciKlass* klass_SHA = NULL; 6095 if (klass_SHA_name != NULL) { 6096 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6097 } 6098 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6099 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6100 Node* ctrl = control(); 6101 set_control(top()); // no intrinsic path 6102 return ctrl; 6103 } 6104 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6105 6106 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6107 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6108 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6109 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6110 6111 return instof_false; // even if it is NULL 6112 } 6113 6114 bool LibraryCallKit::inline_profileBoolean() { 6115 Node* counts = argument(1); 6116 const TypeAryPtr* ary = NULL; 6117 ciArray* aobj = NULL; 6118 if (counts->is_Con() 6119 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6120 && (aobj = ary->const_oop()->as_array()) != NULL 6121 && (aobj->length() == 2)) { 6122 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6123 jint false_cnt = aobj->element_value(0).as_int(); 6124 jint true_cnt = aobj->element_value(1).as_int(); 6125 6126 method()->set_injected_profile(true); 6127 6128 if (C->log() != NULL) { 6129 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6130 false_cnt, true_cnt); 6131 } 6132 6133 if (false_cnt + true_cnt == 0) { 6134 // According to profile, never executed. 6135 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6136 Deoptimization::Action_reinterpret); 6137 return true; 6138 } 6139 6140 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6141 // is a number of each value occurrences. 6142 Node* result = argument(0); 6143 if (false_cnt == 0 || true_cnt == 0) { 6144 // According to profile, one value has been never seen. 6145 int expected_val = (false_cnt == 0) ? 1 : 0; 6146 6147 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6148 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6149 6150 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6151 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6152 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6153 6154 { // Slow path: uncommon trap for never seen value and then reexecute 6155 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6156 // the value has been seen at least once. 6157 PreserveJVMState pjvms(this); 6158 PreserveReexecuteState preexecs(this); 6159 jvms()->set_should_reexecute(true); 6160 6161 set_control(slow_path); 6162 set_i_o(i_o()); 6163 6164 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6165 Deoptimization::Action_reinterpret); 6166 } 6167 // The guard for never seen value enables sharpening of the result and 6168 // returning a constant. It allows to eliminate branches on the same value 6169 // later on. 6170 set_control(fast_path); 6171 result = intcon(expected_val); 6172 } 6173 // Stop profiling. 6174 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6175 // By replacing method body with profile data (represented as ProfileBooleanNode 6176 // on IR level) we effectively disable profiling. 6177 // It enables full speed execution once optimized code is generated. 6178 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6179 C->record_for_igvn(profile); 6180 set_result(profile); 6181 return true; 6182 } else { 6183 // Continue profiling. 6184 // Profile data isn't available at the moment. So, execute method's bytecode version. 6185 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6186 // is compiled and counters aren't available since corresponding MethodHandle 6187 // isn't a compile-time constant. 6188 return false; 6189 } 6190 } 6191 6192 bool LibraryCallKit::inline_isCompileConstant() { 6193 Node* n = argument(0); 6194 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6195 return true; 6196 }