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_ghash_processBlocks(); 282 bool inline_sha_implCompress(vmIntrinsics::ID id); 283 bool inline_digestBase_implCompressMB(int predicate); 284 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 285 bool long_state, address stubAddr, const char *stubName, 286 Node* src_start, Node* ofs, Node* limit); 287 Node* get_state_from_sha_object(Node *sha_object); 288 Node* get_state_from_sha5_object(Node *sha_object); 289 Node* inline_digestBase_implCompressMB_predicate(int predicate); 290 bool inline_encodeISOArray(); 291 bool inline_updateCRC32(); 292 bool inline_updateBytesCRC32(); 293 bool inline_updateByteBufferCRC32(); 294 bool inline_multiplyToLen(); 295 bool inline_squareToLen(); 296 bool inline_mulAdd(); 297 298 bool inline_profileBoolean(); 299 bool inline_isCompileConstant(); 300 }; 301 302 303 //---------------------------make_vm_intrinsic---------------------------- 304 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 305 vmIntrinsics::ID id = m->intrinsic_id(); 306 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 307 308 ccstr disable_intr = NULL; 309 310 if ((DisableIntrinsic[0] != '\0' 311 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 312 (method_has_option_value("DisableIntrinsic", disable_intr) 313 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 314 // disabled by a user request on the command line: 315 // example: -XX:DisableIntrinsic=_hashCode,_getClass 316 return NULL; 317 } 318 319 if (!m->is_loaded()) { 320 // do not attempt to inline unloaded methods 321 return NULL; 322 } 323 324 // Only a few intrinsics implement a virtual dispatch. 325 // They are expensive calls which are also frequently overridden. 326 if (is_virtual) { 327 switch (id) { 328 case vmIntrinsics::_hashCode: 329 case vmIntrinsics::_clone: 330 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 331 break; 332 default: 333 return NULL; 334 } 335 } 336 337 // -XX:-InlineNatives disables nearly all intrinsics: 338 if (!InlineNatives) { 339 switch (id) { 340 case vmIntrinsics::_indexOf: 341 case vmIntrinsics::_compareTo: 342 case vmIntrinsics::_equals: 343 case vmIntrinsics::_equalsC: 344 case vmIntrinsics::_getAndAddInt: 345 case vmIntrinsics::_getAndAddLong: 346 case vmIntrinsics::_getAndSetInt: 347 case vmIntrinsics::_getAndSetLong: 348 case vmIntrinsics::_getAndSetObject: 349 case vmIntrinsics::_loadFence: 350 case vmIntrinsics::_storeFence: 351 case vmIntrinsics::_fullFence: 352 break; // InlineNatives does not control String.compareTo 353 case vmIntrinsics::_Reference_get: 354 break; // InlineNatives does not control Reference.get 355 default: 356 return NULL; 357 } 358 } 359 360 int predicates = 0; 361 bool does_virtual_dispatch = false; 362 363 switch (id) { 364 case vmIntrinsics::_compareTo: 365 if (!SpecialStringCompareTo) return NULL; 366 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 367 break; 368 case vmIntrinsics::_indexOf: 369 if (!SpecialStringIndexOf) return NULL; 370 break; 371 case vmIntrinsics::_equals: 372 if (!SpecialStringEquals) return NULL; 373 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 374 break; 375 case vmIntrinsics::_equalsC: 376 if (!SpecialArraysEquals) return NULL; 377 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 378 break; 379 case vmIntrinsics::_arraycopy: 380 if (!InlineArrayCopy) return NULL; 381 break; 382 case vmIntrinsics::_copyMemory: 383 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 384 if (!InlineArrayCopy) return NULL; 385 break; 386 case vmIntrinsics::_hashCode: 387 if (!InlineObjectHash) return NULL; 388 does_virtual_dispatch = true; 389 break; 390 case vmIntrinsics::_clone: 391 does_virtual_dispatch = true; 392 case vmIntrinsics::_copyOf: 393 case vmIntrinsics::_copyOfRange: 394 if (!InlineObjectCopy) return NULL; 395 // These also use the arraycopy intrinsic mechanism: 396 if (!InlineArrayCopy) return NULL; 397 break; 398 case vmIntrinsics::_encodeISOArray: 399 if (!SpecialEncodeISOArray) return NULL; 400 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 401 break; 402 case vmIntrinsics::_checkIndex: 403 // We do not intrinsify this. The optimizer does fine with it. 404 return NULL; 405 406 case vmIntrinsics::_getCallerClass: 407 if (!InlineReflectionGetCallerClass) return NULL; 408 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 409 break; 410 411 case vmIntrinsics::_bitCount_i: 412 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 413 break; 414 415 case vmIntrinsics::_bitCount_l: 416 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 417 break; 418 419 case vmIntrinsics::_numberOfLeadingZeros_i: 420 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 421 break; 422 423 case vmIntrinsics::_numberOfLeadingZeros_l: 424 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 425 break; 426 427 case vmIntrinsics::_numberOfTrailingZeros_i: 428 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 429 break; 430 431 case vmIntrinsics::_numberOfTrailingZeros_l: 432 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 433 break; 434 435 case vmIntrinsics::_reverseBytes_c: 436 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 437 break; 438 case vmIntrinsics::_reverseBytes_s: 439 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 440 break; 441 case vmIntrinsics::_reverseBytes_i: 442 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 443 break; 444 case vmIntrinsics::_reverseBytes_l: 445 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 446 break; 447 448 case vmIntrinsics::_Reference_get: 449 // Use the intrinsic version of Reference.get() so that the value in 450 // the referent field can be registered by the G1 pre-barrier code. 451 // Also add memory barrier to prevent commoning reads from this field 452 // across safepoint since GC can change it value. 453 break; 454 455 case vmIntrinsics::_compareAndSwapObject: 456 #ifdef _LP64 457 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 458 #endif 459 break; 460 461 case vmIntrinsics::_compareAndSwapLong: 462 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 463 break; 464 465 case vmIntrinsics::_getAndAddInt: 466 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 467 break; 468 469 case vmIntrinsics::_getAndAddLong: 470 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 471 break; 472 473 case vmIntrinsics::_getAndSetInt: 474 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 475 break; 476 477 case vmIntrinsics::_getAndSetLong: 478 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 479 break; 480 481 case vmIntrinsics::_getAndSetObject: 482 #ifdef _LP64 483 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 484 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 485 break; 486 #else 487 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 488 break; 489 #endif 490 491 case vmIntrinsics::_aescrypt_encryptBlock: 492 case vmIntrinsics::_aescrypt_decryptBlock: 493 if (!UseAESIntrinsics) return NULL; 494 break; 495 496 case vmIntrinsics::_multiplyToLen: 497 if (!UseMultiplyToLenIntrinsic) return NULL; 498 break; 499 500 case vmIntrinsics::_squareToLen: 501 if (!UseSquareToLenIntrinsic) return NULL; 502 break; 503 504 case vmIntrinsics::_mulAdd: 505 if (!UseMulAddIntrinsic) return NULL; 506 break; 507 508 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 509 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 510 if (!UseAESIntrinsics) return NULL; 511 // these two require the predicated logic 512 predicates = 1; 513 break; 514 515 case vmIntrinsics::_sha_implCompress: 516 if (!UseSHA1Intrinsics) return NULL; 517 break; 518 519 case vmIntrinsics::_sha2_implCompress: 520 if (!UseSHA256Intrinsics) return NULL; 521 break; 522 523 case vmIntrinsics::_sha5_implCompress: 524 if (!UseSHA512Intrinsics) return NULL; 525 break; 526 527 case vmIntrinsics::_digestBase_implCompressMB: 528 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 529 predicates = 3; 530 break; 531 532 case vmIntrinsics::_ghash_processBlocks: 533 if (!UseGHASHIntrinsics) return NULL; 534 break; 535 536 case vmIntrinsics::_updateCRC32: 537 case vmIntrinsics::_updateBytesCRC32: 538 case vmIntrinsics::_updateByteBufferCRC32: 539 if (!UseCRC32Intrinsics) return NULL; 540 break; 541 542 case vmIntrinsics::_incrementExactI: 543 case vmIntrinsics::_addExactI: 544 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 545 break; 546 case vmIntrinsics::_incrementExactL: 547 case vmIntrinsics::_addExactL: 548 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 549 break; 550 case vmIntrinsics::_decrementExactI: 551 case vmIntrinsics::_subtractExactI: 552 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 553 break; 554 case vmIntrinsics::_decrementExactL: 555 case vmIntrinsics::_subtractExactL: 556 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 557 break; 558 case vmIntrinsics::_negateExactI: 559 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 560 break; 561 case vmIntrinsics::_negateExactL: 562 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 563 break; 564 case vmIntrinsics::_multiplyExactI: 565 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 566 break; 567 case vmIntrinsics::_multiplyExactL: 568 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 569 break; 570 571 case vmIntrinsics::_getShortUnaligned: 572 case vmIntrinsics::_getCharUnaligned: 573 case vmIntrinsics::_getIntUnaligned: 574 case vmIntrinsics::_getLongUnaligned: 575 case vmIntrinsics::_putShortUnaligned: 576 case vmIntrinsics::_putCharUnaligned: 577 case vmIntrinsics::_putIntUnaligned: 578 case vmIntrinsics::_putLongUnaligned: 579 if (!UseUnalignedAccesses) return NULL; 580 break; 581 582 default: 583 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 584 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 585 break; 586 } 587 588 // -XX:-InlineClassNatives disables natives from the Class class. 589 // The flag applies to all reflective calls, notably Array.newArray 590 // (visible to Java programmers as Array.newInstance). 591 if (m->holder()->name() == ciSymbol::java_lang_Class() || 592 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 593 if (!InlineClassNatives) return NULL; 594 } 595 596 // -XX:-InlineThreadNatives disables natives from the Thread class. 597 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 598 if (!InlineThreadNatives) return NULL; 599 } 600 601 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 602 if (m->holder()->name() == ciSymbol::java_lang_Math() || 603 m->holder()->name() == ciSymbol::java_lang_Float() || 604 m->holder()->name() == ciSymbol::java_lang_Double()) { 605 if (!InlineMathNatives) return NULL; 606 } 607 608 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 609 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 610 if (!InlineUnsafeOps) return NULL; 611 } 612 613 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 614 } 615 616 //----------------------register_library_intrinsics----------------------- 617 // Initialize this file's data structures, for each Compile instance. 618 void Compile::register_library_intrinsics() { 619 // Nothing to do here. 620 } 621 622 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 623 LibraryCallKit kit(jvms, this); 624 Compile* C = kit.C; 625 int nodes = C->unique(); 626 #ifndef PRODUCT 627 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 628 char buf[1000]; 629 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 630 tty->print_cr("Intrinsic %s", str); 631 } 632 #endif 633 ciMethod* callee = kit.callee(); 634 const int bci = kit.bci(); 635 636 // Try to inline the intrinsic. 637 if (kit.try_to_inline(_last_predicate)) { 638 if (C->print_intrinsics() || C->print_inlining()) { 639 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 640 } 641 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 642 if (C->log()) { 643 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 644 vmIntrinsics::name_at(intrinsic_id()), 645 (is_virtual() ? " virtual='1'" : ""), 646 C->unique() - nodes); 647 } 648 // Push the result from the inlined method onto the stack. 649 kit.push_result(); 650 C->print_inlining_update(this); 651 return kit.transfer_exceptions_into_jvms(); 652 } 653 654 // The intrinsic bailed out 655 if (C->print_intrinsics() || C->print_inlining()) { 656 if (jvms->has_method()) { 657 // Not a root compile. 658 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 659 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 660 } else { 661 // Root compile 662 tty->print("Did not generate intrinsic %s%s at bci:%d in", 663 vmIntrinsics::name_at(intrinsic_id()), 664 (is_virtual() ? " (virtual)" : ""), bci); 665 } 666 } 667 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 668 C->print_inlining_update(this); 669 return NULL; 670 } 671 672 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 673 LibraryCallKit kit(jvms, this); 674 Compile* C = kit.C; 675 int nodes = C->unique(); 676 _last_predicate = predicate; 677 #ifndef PRODUCT 678 assert(is_predicated() && predicate < predicates_count(), "sanity"); 679 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 680 char buf[1000]; 681 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 682 tty->print_cr("Predicate for intrinsic %s", str); 683 } 684 #endif 685 ciMethod* callee = kit.callee(); 686 const int bci = kit.bci(); 687 688 Node* slow_ctl = kit.try_to_predicate(predicate); 689 if (!kit.failing()) { 690 if (C->print_intrinsics() || C->print_inlining()) { 691 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 692 } 693 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 694 if (C->log()) { 695 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 696 vmIntrinsics::name_at(intrinsic_id()), 697 (is_virtual() ? " virtual='1'" : ""), 698 C->unique() - nodes); 699 } 700 return slow_ctl; // Could be NULL if the check folds. 701 } 702 703 // The intrinsic bailed out 704 if (C->print_intrinsics() || C->print_inlining()) { 705 if (jvms->has_method()) { 706 // Not a root compile. 707 const char* msg = "failed to generate predicate for intrinsic"; 708 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 709 } else { 710 // Root compile 711 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 712 vmIntrinsics::name_at(intrinsic_id()), 713 (is_virtual() ? " (virtual)" : ""), bci); 714 } 715 } 716 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 717 return NULL; 718 } 719 720 bool LibraryCallKit::try_to_inline(int predicate) { 721 // Handle symbolic names for otherwise undistinguished boolean switches: 722 const bool is_store = true; 723 const bool is_native_ptr = true; 724 const bool is_static = true; 725 const bool is_volatile = true; 726 727 if (!jvms()->has_method()) { 728 // Root JVMState has a null method. 729 assert(map()->memory()->Opcode() == Op_Parm, ""); 730 // Insert the memory aliasing node 731 set_all_memory(reset_memory()); 732 } 733 assert(merged_memory(), ""); 734 735 736 switch (intrinsic_id()) { 737 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 738 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 739 case vmIntrinsics::_getClass: return inline_native_getClass(); 740 741 case vmIntrinsics::_dsin: 742 case vmIntrinsics::_dcos: 743 case vmIntrinsics::_dtan: 744 case vmIntrinsics::_dabs: 745 case vmIntrinsics::_datan2: 746 case vmIntrinsics::_dsqrt: 747 case vmIntrinsics::_dexp: 748 case vmIntrinsics::_dlog: 749 case vmIntrinsics::_dlog10: 750 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 751 752 case vmIntrinsics::_min: 753 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 754 755 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 756 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 757 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 758 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 759 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 760 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 761 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 762 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 763 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 764 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 765 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 766 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 767 768 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 769 770 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 771 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 772 case vmIntrinsics::_equals: return inline_string_equals(); 773 774 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 775 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 776 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 777 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 778 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 779 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 780 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 781 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 782 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 783 784 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 785 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 786 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 787 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 788 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 789 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 790 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 791 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 792 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 793 794 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 795 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 796 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 797 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 798 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 799 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 800 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 801 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 802 803 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 804 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 805 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 806 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 807 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 808 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 809 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 810 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 811 812 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 813 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 814 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 815 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 816 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 817 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 818 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 819 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 820 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 821 822 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 823 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 824 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 825 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 826 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 827 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 828 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 829 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 830 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 831 832 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 833 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 834 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 835 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 836 837 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 838 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 839 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 840 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 841 842 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 843 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 844 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 845 846 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 847 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 848 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 849 850 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 851 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 852 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 853 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 854 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 855 856 case vmIntrinsics::_loadFence: 857 case vmIntrinsics::_storeFence: 858 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 859 860 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 861 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 862 863 #ifdef TRACE_HAVE_INTRINSICS 864 case vmIntrinsics::_classID: return inline_native_classID(); 865 case vmIntrinsics::_threadID: return inline_native_threadID(); 866 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 867 #endif 868 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 869 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 870 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 871 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 872 case vmIntrinsics::_newArray: return inline_native_newArray(); 873 case vmIntrinsics::_getLength: return inline_native_getLength(); 874 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 875 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 876 case vmIntrinsics::_equalsC: return inline_array_equals(); 877 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 878 879 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 880 881 case vmIntrinsics::_isInstance: 882 case vmIntrinsics::_getModifiers: 883 case vmIntrinsics::_isInterface: 884 case vmIntrinsics::_isArray: 885 case vmIntrinsics::_isPrimitive: 886 case vmIntrinsics::_getSuperclass: 887 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 888 889 case vmIntrinsics::_floatToRawIntBits: 890 case vmIntrinsics::_floatToIntBits: 891 case vmIntrinsics::_intBitsToFloat: 892 case vmIntrinsics::_doubleToRawLongBits: 893 case vmIntrinsics::_doubleToLongBits: 894 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 895 896 case vmIntrinsics::_numberOfLeadingZeros_i: 897 case vmIntrinsics::_numberOfLeadingZeros_l: 898 case vmIntrinsics::_numberOfTrailingZeros_i: 899 case vmIntrinsics::_numberOfTrailingZeros_l: 900 case vmIntrinsics::_bitCount_i: 901 case vmIntrinsics::_bitCount_l: 902 case vmIntrinsics::_reverseBytes_i: 903 case vmIntrinsics::_reverseBytes_l: 904 case vmIntrinsics::_reverseBytes_s: 905 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 906 907 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 908 909 case vmIntrinsics::_Reference_get: return inline_reference_get(); 910 911 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 912 913 case vmIntrinsics::_aescrypt_encryptBlock: 914 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 915 916 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 917 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 918 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 919 920 case vmIntrinsics::_sha_implCompress: 921 case vmIntrinsics::_sha2_implCompress: 922 case vmIntrinsics::_sha5_implCompress: 923 return inline_sha_implCompress(intrinsic_id()); 924 925 case vmIntrinsics::_digestBase_implCompressMB: 926 return inline_digestBase_implCompressMB(predicate); 927 928 case vmIntrinsics::_multiplyToLen: 929 return inline_multiplyToLen(); 930 931 case vmIntrinsics::_squareToLen: 932 return inline_squareToLen(); 933 934 case vmIntrinsics::_mulAdd: 935 return inline_mulAdd(); 936 937 case vmIntrinsics::_ghash_processBlocks: 938 return inline_ghash_processBlocks(); 939 940 case vmIntrinsics::_encodeISOArray: 941 return inline_encodeISOArray(); 942 943 case vmIntrinsics::_updateCRC32: 944 return inline_updateCRC32(); 945 case vmIntrinsics::_updateBytesCRC32: 946 return inline_updateBytesCRC32(); 947 case vmIntrinsics::_updateByteBufferCRC32: 948 return inline_updateByteBufferCRC32(); 949 950 case vmIntrinsics::_profileBoolean: 951 return inline_profileBoolean(); 952 case vmIntrinsics::_isCompileConstant: 953 return inline_isCompileConstant(); 954 955 default: 956 // If you get here, it may be that someone has added a new intrinsic 957 // to the list in vmSymbols.hpp without implementing it here. 958 #ifndef PRODUCT 959 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 960 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 961 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 962 } 963 #endif 964 return false; 965 } 966 } 967 968 Node* LibraryCallKit::try_to_predicate(int predicate) { 969 if (!jvms()->has_method()) { 970 // Root JVMState has a null method. 971 assert(map()->memory()->Opcode() == Op_Parm, ""); 972 // Insert the memory aliasing node 973 set_all_memory(reset_memory()); 974 } 975 assert(merged_memory(), ""); 976 977 switch (intrinsic_id()) { 978 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 979 return inline_cipherBlockChaining_AESCrypt_predicate(false); 980 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 981 return inline_cipherBlockChaining_AESCrypt_predicate(true); 982 case vmIntrinsics::_digestBase_implCompressMB: 983 return inline_digestBase_implCompressMB_predicate(predicate); 984 985 default: 986 // If you get here, it may be that someone has added a new intrinsic 987 // to the list in vmSymbols.hpp without implementing it here. 988 #ifndef PRODUCT 989 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 990 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 991 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 992 } 993 #endif 994 Node* slow_ctl = control(); 995 set_control(top()); // No fast path instrinsic 996 return slow_ctl; 997 } 998 } 999 1000 //------------------------------set_result------------------------------- 1001 // Helper function for finishing intrinsics. 1002 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 1003 record_for_igvn(region); 1004 set_control(_gvn.transform(region)); 1005 set_result( _gvn.transform(value)); 1006 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 1007 } 1008 1009 //------------------------------generate_guard--------------------------- 1010 // Helper function for generating guarded fast-slow graph structures. 1011 // The given 'test', if true, guards a slow path. If the test fails 1012 // then a fast path can be taken. (We generally hope it fails.) 1013 // In all cases, GraphKit::control() is updated to the fast path. 1014 // The returned value represents the control for the slow path. 1015 // The return value is never 'top'; it is either a valid control 1016 // or NULL if it is obvious that the slow path can never be taken. 1017 // Also, if region and the slow control are not NULL, the slow edge 1018 // is appended to the region. 1019 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 1020 if (stopped()) { 1021 // Already short circuited. 1022 return NULL; 1023 } 1024 1025 // Build an if node and its projections. 1026 // If test is true we take the slow path, which we assume is uncommon. 1027 if (_gvn.type(test) == TypeInt::ZERO) { 1028 // The slow branch is never taken. No need to build this guard. 1029 return NULL; 1030 } 1031 1032 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1033 1034 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1035 if (if_slow == top()) { 1036 // The slow branch is never taken. No need to build this guard. 1037 return NULL; 1038 } 1039 1040 if (region != NULL) 1041 region->add_req(if_slow); 1042 1043 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1044 set_control(if_fast); 1045 1046 return if_slow; 1047 } 1048 1049 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1050 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1051 } 1052 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1053 return generate_guard(test, region, PROB_FAIR); 1054 } 1055 1056 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1057 Node* *pos_index) { 1058 if (stopped()) 1059 return NULL; // already stopped 1060 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1061 return NULL; // index is already adequately typed 1062 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1063 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1064 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1065 if (is_neg != NULL && pos_index != NULL) { 1066 // Emulate effect of Parse::adjust_map_after_if. 1067 Node* ccast = new CastIINode(index, TypeInt::POS); 1068 ccast->set_req(0, control()); 1069 (*pos_index) = _gvn.transform(ccast); 1070 } 1071 return is_neg; 1072 } 1073 1074 // Make sure that 'position' is a valid limit index, in [0..length]. 1075 // There are two equivalent plans for checking this: 1076 // A. (offset + copyLength) unsigned<= arrayLength 1077 // B. offset <= (arrayLength - copyLength) 1078 // We require that all of the values above, except for the sum and 1079 // difference, are already known to be non-negative. 1080 // Plan A is robust in the face of overflow, if offset and copyLength 1081 // are both hugely positive. 1082 // 1083 // Plan B is less direct and intuitive, but it does not overflow at 1084 // all, since the difference of two non-negatives is always 1085 // representable. Whenever Java methods must perform the equivalent 1086 // check they generally use Plan B instead of Plan A. 1087 // For the moment we use Plan A. 1088 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1089 Node* subseq_length, 1090 Node* array_length, 1091 RegionNode* region) { 1092 if (stopped()) 1093 return NULL; // already stopped 1094 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1095 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1096 return NULL; // common case of whole-array copy 1097 Node* last = subseq_length; 1098 if (!zero_offset) // last += offset 1099 last = _gvn.transform(new AddINode(last, offset)); 1100 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1101 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1102 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1103 return is_over; 1104 } 1105 1106 1107 //--------------------------generate_current_thread-------------------- 1108 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1109 ciKlass* thread_klass = env()->Thread_klass(); 1110 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1111 Node* thread = _gvn.transform(new ThreadLocalNode()); 1112 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1113 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1114 tls_output = thread; 1115 return threadObj; 1116 } 1117 1118 1119 //------------------------------make_string_method_node------------------------ 1120 // Helper method for String intrinsic functions. This version is called 1121 // with str1 and str2 pointing to String object nodes. 1122 // 1123 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1124 Node* no_ctrl = NULL; 1125 1126 // Get start addr of string 1127 Node* str1_value = load_String_value(no_ctrl, str1); 1128 Node* str1_offset = load_String_offset(no_ctrl, str1); 1129 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1130 1131 // Get length of string 1 1132 Node* str1_len = load_String_length(no_ctrl, str1); 1133 1134 Node* str2_value = load_String_value(no_ctrl, str2); 1135 Node* str2_offset = load_String_offset(no_ctrl, str2); 1136 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1137 1138 Node* str2_len = NULL; 1139 Node* result = NULL; 1140 1141 switch (opcode) { 1142 case Op_StrIndexOf: 1143 // Get length of string 2 1144 str2_len = load_String_length(no_ctrl, str2); 1145 1146 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1147 str1_start, str1_len, str2_start, str2_len); 1148 break; 1149 case Op_StrComp: 1150 // Get length of string 2 1151 str2_len = load_String_length(no_ctrl, str2); 1152 1153 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1154 str1_start, str1_len, str2_start, str2_len); 1155 break; 1156 case Op_StrEquals: 1157 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1158 str1_start, str2_start, str1_len); 1159 break; 1160 default: 1161 ShouldNotReachHere(); 1162 return NULL; 1163 } 1164 1165 // All these intrinsics have checks. 1166 C->set_has_split_ifs(true); // Has chance for split-if optimization 1167 1168 return _gvn.transform(result); 1169 } 1170 1171 // Helper method for String intrinsic functions. This version is called 1172 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1173 // to Int nodes containing the lenghts of str1 and str2. 1174 // 1175 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1176 Node* result = NULL; 1177 switch (opcode) { 1178 case Op_StrIndexOf: 1179 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1180 str1_start, cnt1, str2_start, cnt2); 1181 break; 1182 case Op_StrComp: 1183 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1184 str1_start, cnt1, str2_start, cnt2); 1185 break; 1186 case Op_StrEquals: 1187 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1188 str1_start, str2_start, cnt1); 1189 break; 1190 default: 1191 ShouldNotReachHere(); 1192 return NULL; 1193 } 1194 1195 // All these intrinsics have checks. 1196 C->set_has_split_ifs(true); // Has chance for split-if optimization 1197 1198 return _gvn.transform(result); 1199 } 1200 1201 //------------------------------inline_string_compareTo------------------------ 1202 // public int java.lang.String.compareTo(String anotherString); 1203 bool LibraryCallKit::inline_string_compareTo() { 1204 Node* receiver = null_check(argument(0)); 1205 Node* arg = null_check(argument(1)); 1206 if (stopped()) { 1207 return true; 1208 } 1209 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1210 return true; 1211 } 1212 1213 //------------------------------inline_string_equals------------------------ 1214 bool LibraryCallKit::inline_string_equals() { 1215 Node* receiver = null_check_receiver(); 1216 // NOTE: Do not null check argument for String.equals() because spec 1217 // allows to specify NULL as argument. 1218 Node* argument = this->argument(1); 1219 if (stopped()) { 1220 return true; 1221 } 1222 1223 // paths (plus control) merge 1224 RegionNode* region = new RegionNode(5); 1225 Node* phi = new PhiNode(region, TypeInt::BOOL); 1226 1227 // does source == target string? 1228 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1229 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1230 1231 Node* if_eq = generate_slow_guard(bol, NULL); 1232 if (if_eq != NULL) { 1233 // receiver == argument 1234 phi->init_req(2, intcon(1)); 1235 region->init_req(2, if_eq); 1236 } 1237 1238 // get String klass for instanceOf 1239 ciInstanceKlass* klass = env()->String_klass(); 1240 1241 if (!stopped()) { 1242 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1243 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1244 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1245 1246 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1247 //instanceOf == true, fallthrough 1248 1249 if (inst_false != NULL) { 1250 phi->init_req(3, intcon(0)); 1251 region->init_req(3, inst_false); 1252 } 1253 } 1254 1255 if (!stopped()) { 1256 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1257 1258 // Properly cast the argument to String 1259 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1260 // This path is taken only when argument's type is String:NotNull. 1261 argument = cast_not_null(argument, false); 1262 1263 Node* no_ctrl = NULL; 1264 1265 // Get start addr of receiver 1266 Node* receiver_val = load_String_value(no_ctrl, receiver); 1267 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1268 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1269 1270 // Get length of receiver 1271 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1272 1273 // Get start addr of argument 1274 Node* argument_val = load_String_value(no_ctrl, argument); 1275 Node* argument_offset = load_String_offset(no_ctrl, argument); 1276 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1277 1278 // Get length of argument 1279 Node* argument_cnt = load_String_length(no_ctrl, argument); 1280 1281 // Check for receiver count != argument count 1282 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1283 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1284 Node* if_ne = generate_slow_guard(bol, NULL); 1285 if (if_ne != NULL) { 1286 phi->init_req(4, intcon(0)); 1287 region->init_req(4, if_ne); 1288 } 1289 1290 // Check for count == 0 is done by assembler code for StrEquals. 1291 1292 if (!stopped()) { 1293 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1294 phi->init_req(1, equals); 1295 region->init_req(1, control()); 1296 } 1297 } 1298 1299 // post merge 1300 set_control(_gvn.transform(region)); 1301 record_for_igvn(region); 1302 1303 set_result(_gvn.transform(phi)); 1304 return true; 1305 } 1306 1307 //------------------------------inline_array_equals---------------------------- 1308 bool LibraryCallKit::inline_array_equals() { 1309 Node* arg1 = argument(0); 1310 Node* arg2 = argument(1); 1311 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1312 return true; 1313 } 1314 1315 // Java version of String.indexOf(constant string) 1316 // class StringDecl { 1317 // StringDecl(char[] ca) { 1318 // offset = 0; 1319 // count = ca.length; 1320 // value = ca; 1321 // } 1322 // int offset; 1323 // int count; 1324 // char[] value; 1325 // } 1326 // 1327 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1328 // int targetOffset, int cache_i, int md2) { 1329 // int cache = cache_i; 1330 // int sourceOffset = string_object.offset; 1331 // int sourceCount = string_object.count; 1332 // int targetCount = target_object.length; 1333 // 1334 // int targetCountLess1 = targetCount - 1; 1335 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1336 // 1337 // char[] source = string_object.value; 1338 // char[] target = target_object; 1339 // int lastChar = target[targetCountLess1]; 1340 // 1341 // outer_loop: 1342 // for (int i = sourceOffset; i < sourceEnd; ) { 1343 // int src = source[i + targetCountLess1]; 1344 // if (src == lastChar) { 1345 // // With random strings and a 4-character alphabet, 1346 // // reverse matching at this point sets up 0.8% fewer 1347 // // frames, but (paradoxically) makes 0.3% more probes. 1348 // // Since those probes are nearer the lastChar probe, 1349 // // there is may be a net D$ win with reverse matching. 1350 // // But, reversing loop inhibits unroll of inner loop 1351 // // for unknown reason. So, does running outer loop from 1352 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1353 // for (int j = 0; j < targetCountLess1; j++) { 1354 // if (target[targetOffset + j] != source[i+j]) { 1355 // if ((cache & (1 << source[i+j])) == 0) { 1356 // if (md2 < j+1) { 1357 // i += j+1; 1358 // continue outer_loop; 1359 // } 1360 // } 1361 // i += md2; 1362 // continue outer_loop; 1363 // } 1364 // } 1365 // return i - sourceOffset; 1366 // } 1367 // if ((cache & (1 << src)) == 0) { 1368 // i += targetCountLess1; 1369 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1370 // i++; 1371 // } 1372 // return -1; 1373 // } 1374 1375 //------------------------------string_indexOf------------------------ 1376 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1377 jint cache_i, jint md2_i) { 1378 1379 Node* no_ctrl = NULL; 1380 float likely = PROB_LIKELY(0.9); 1381 float unlikely = PROB_UNLIKELY(0.9); 1382 1383 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1384 1385 Node* source = load_String_value(no_ctrl, string_object); 1386 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1387 Node* sourceCount = load_String_length(no_ctrl, string_object); 1388 1389 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1390 jint target_length = target_array->length(); 1391 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1392 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1393 1394 // String.value field is known to be @Stable. 1395 if (UseImplicitStableValues) { 1396 target = cast_array_to_stable(target, target_type); 1397 } 1398 1399 IdealKit kit(this, false, true); 1400 #define __ kit. 1401 Node* zero = __ ConI(0); 1402 Node* one = __ ConI(1); 1403 Node* cache = __ ConI(cache_i); 1404 Node* md2 = __ ConI(md2_i); 1405 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1406 Node* targetCountLess1 = __ ConI(target_length - 1); 1407 Node* targetOffset = __ ConI(targetOffset_i); 1408 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1409 1410 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1411 Node* outer_loop = __ make_label(2 /* goto */); 1412 Node* return_ = __ make_label(1); 1413 1414 __ set(rtn,__ ConI(-1)); 1415 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1416 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1417 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1418 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1419 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1420 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1421 Node* tpj = __ AddI(targetOffset, __ value(j)); 1422 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1423 Node* ipj = __ AddI(__ value(i), __ value(j)); 1424 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1425 __ if_then(targ, BoolTest::ne, src2); { 1426 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1427 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1428 __ increment(i, __ AddI(__ value(j), one)); 1429 __ goto_(outer_loop); 1430 } __ end_if(); __ dead(j); 1431 }__ end_if(); __ dead(j); 1432 __ increment(i, md2); 1433 __ goto_(outer_loop); 1434 }__ end_if(); 1435 __ increment(j, one); 1436 }__ end_loop(); __ dead(j); 1437 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1438 __ goto_(return_); 1439 }__ end_if(); 1440 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1441 __ increment(i, targetCountLess1); 1442 }__ end_if(); 1443 __ increment(i, one); 1444 __ bind(outer_loop); 1445 }__ end_loop(); __ dead(i); 1446 __ bind(return_); 1447 1448 // Final sync IdealKit and GraphKit. 1449 final_sync(kit); 1450 Node* result = __ value(rtn); 1451 #undef __ 1452 C->set_has_loops(true); 1453 return result; 1454 } 1455 1456 //------------------------------inline_string_indexOf------------------------ 1457 bool LibraryCallKit::inline_string_indexOf() { 1458 Node* receiver = argument(0); 1459 Node* arg = argument(1); 1460 1461 Node* result; 1462 if (Matcher::has_match_rule(Op_StrIndexOf) && 1463 UseSSE42Intrinsics) { 1464 // Generate SSE4.2 version of indexOf 1465 // We currently only have match rules that use SSE4.2 1466 1467 receiver = null_check(receiver); 1468 arg = null_check(arg); 1469 if (stopped()) { 1470 return true; 1471 } 1472 1473 // Make the merge point 1474 RegionNode* result_rgn = new RegionNode(4); 1475 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1476 Node* no_ctrl = NULL; 1477 1478 // Get start addr of source string 1479 Node* source = load_String_value(no_ctrl, receiver); 1480 Node* source_offset = load_String_offset(no_ctrl, receiver); 1481 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1482 1483 // Get length of source string 1484 Node* source_cnt = load_String_length(no_ctrl, receiver); 1485 1486 // Get start addr of substring 1487 Node* substr = load_String_value(no_ctrl, arg); 1488 Node* substr_offset = load_String_offset(no_ctrl, arg); 1489 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1490 1491 // Get length of source string 1492 Node* substr_cnt = load_String_length(no_ctrl, arg); 1493 1494 // Check for substr count > string count 1495 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1496 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1497 Node* if_gt = generate_slow_guard(bol, NULL); 1498 if (if_gt != NULL) { 1499 result_phi->init_req(2, intcon(-1)); 1500 result_rgn->init_req(2, if_gt); 1501 } 1502 1503 if (!stopped()) { 1504 // Check for substr count == 0 1505 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1506 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1507 Node* if_zero = generate_slow_guard(bol, NULL); 1508 if (if_zero != NULL) { 1509 result_phi->init_req(3, intcon(0)); 1510 result_rgn->init_req(3, if_zero); 1511 } 1512 } 1513 1514 if (!stopped()) { 1515 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1516 result_phi->init_req(1, result); 1517 result_rgn->init_req(1, control()); 1518 } 1519 set_control(_gvn.transform(result_rgn)); 1520 record_for_igvn(result_rgn); 1521 result = _gvn.transform(result_phi); 1522 1523 } else { // Use LibraryCallKit::string_indexOf 1524 // don't intrinsify if argument isn't a constant string. 1525 if (!arg->is_Con()) { 1526 return false; 1527 } 1528 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1529 if (str_type == NULL) { 1530 return false; 1531 } 1532 ciInstanceKlass* klass = env()->String_klass(); 1533 ciObject* str_const = str_type->const_oop(); 1534 if (str_const == NULL || str_const->klass() != klass) { 1535 return false; 1536 } 1537 ciInstance* str = str_const->as_instance(); 1538 assert(str != NULL, "must be instance"); 1539 1540 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1541 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1542 1543 int o; 1544 int c; 1545 if (java_lang_String::has_offset_field()) { 1546 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1547 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1548 } else { 1549 o = 0; 1550 c = pat->length(); 1551 } 1552 1553 // constant strings have no offset and count == length which 1554 // simplifies the resulting code somewhat so lets optimize for that. 1555 if (o != 0 || c != pat->length()) { 1556 return false; 1557 } 1558 1559 receiver = null_check(receiver, T_OBJECT); 1560 // NOTE: No null check on the argument is needed since it's a constant String oop. 1561 if (stopped()) { 1562 return true; 1563 } 1564 1565 // The null string as a pattern always returns 0 (match at beginning of string) 1566 if (c == 0) { 1567 set_result(intcon(0)); 1568 return true; 1569 } 1570 1571 // Generate default indexOf 1572 jchar lastChar = pat->char_at(o + (c - 1)); 1573 int cache = 0; 1574 int i; 1575 for (i = 0; i < c - 1; i++) { 1576 assert(i < pat->length(), "out of range"); 1577 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1578 } 1579 1580 int md2 = c; 1581 for (i = 0; i < c - 1; i++) { 1582 assert(i < pat->length(), "out of range"); 1583 if (pat->char_at(o + i) == lastChar) { 1584 md2 = (c - 1) - i; 1585 } 1586 } 1587 1588 result = string_indexOf(receiver, pat, o, cache, md2); 1589 } 1590 set_result(result); 1591 return true; 1592 } 1593 1594 //--------------------------round_double_node-------------------------------- 1595 // Round a double node if necessary. 1596 Node* LibraryCallKit::round_double_node(Node* n) { 1597 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1598 n = _gvn.transform(new RoundDoubleNode(0, n)); 1599 return n; 1600 } 1601 1602 //------------------------------inline_math----------------------------------- 1603 // public static double Math.abs(double) 1604 // public static double Math.sqrt(double) 1605 // public static double Math.log(double) 1606 // public static double Math.log10(double) 1607 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1608 Node* arg = round_double_node(argument(0)); 1609 Node* n; 1610 switch (id) { 1611 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1612 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1613 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1614 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1615 default: fatal_unexpected_iid(id); break; 1616 } 1617 set_result(_gvn.transform(n)); 1618 return true; 1619 } 1620 1621 //------------------------------inline_trig---------------------------------- 1622 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1623 // argument reduction which will turn into a fast/slow diamond. 1624 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1625 Node* arg = round_double_node(argument(0)); 1626 Node* n = NULL; 1627 1628 switch (id) { 1629 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1630 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1631 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1632 default: fatal_unexpected_iid(id); break; 1633 } 1634 n = _gvn.transform(n); 1635 1636 // Rounding required? Check for argument reduction! 1637 if (Matcher::strict_fp_requires_explicit_rounding) { 1638 static const double pi_4 = 0.7853981633974483; 1639 static const double neg_pi_4 = -0.7853981633974483; 1640 // pi/2 in 80-bit extended precision 1641 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1642 // -pi/2 in 80-bit extended precision 1643 // 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}; 1644 // Cutoff value for using this argument reduction technique 1645 //static const double pi_2_minus_epsilon = 1.564660403643354; 1646 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1647 1648 // Pseudocode for sin: 1649 // if (x <= Math.PI / 4.0) { 1650 // if (x >= -Math.PI / 4.0) return fsin(x); 1651 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1652 // } else { 1653 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1654 // } 1655 // return StrictMath.sin(x); 1656 1657 // Pseudocode for cos: 1658 // if (x <= Math.PI / 4.0) { 1659 // if (x >= -Math.PI / 4.0) return fcos(x); 1660 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1661 // } else { 1662 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1663 // } 1664 // return StrictMath.cos(x); 1665 1666 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1667 // requires a special machine instruction to load it. Instead we'll try 1668 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1669 // probably do the math inside the SIN encoding. 1670 1671 // Make the merge point 1672 RegionNode* r = new RegionNode(3); 1673 Node* phi = new PhiNode(r, Type::DOUBLE); 1674 1675 // Flatten arg so we need only 1 test 1676 Node *abs = _gvn.transform(new AbsDNode(arg)); 1677 // Node for PI/4 constant 1678 Node *pi4 = makecon(TypeD::make(pi_4)); 1679 // Check PI/4 : abs(arg) 1680 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1681 // Check: If PI/4 < abs(arg) then go slow 1682 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1683 // Branch either way 1684 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1685 set_control(opt_iff(r,iff)); 1686 1687 // Set fast path result 1688 phi->init_req(2, n); 1689 1690 // Slow path - non-blocking leaf call 1691 Node* call = NULL; 1692 switch (id) { 1693 case vmIntrinsics::_dsin: 1694 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1695 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1696 "Sin", NULL, arg, top()); 1697 break; 1698 case vmIntrinsics::_dcos: 1699 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1700 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1701 "Cos", NULL, arg, top()); 1702 break; 1703 case vmIntrinsics::_dtan: 1704 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1705 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1706 "Tan", NULL, arg, top()); 1707 break; 1708 } 1709 assert(control()->in(0) == call, ""); 1710 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1711 r->init_req(1, control()); 1712 phi->init_req(1, slow_result); 1713 1714 // Post-merge 1715 set_control(_gvn.transform(r)); 1716 record_for_igvn(r); 1717 n = _gvn.transform(phi); 1718 1719 C->set_has_split_ifs(true); // Has chance for split-if optimization 1720 } 1721 set_result(n); 1722 return true; 1723 } 1724 1725 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1726 //------------------- 1727 //result=(result.isNaN())? funcAddr():result; 1728 // Check: If isNaN() by checking result!=result? then either trap 1729 // or go to runtime 1730 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1731 // Build the boolean node 1732 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1733 1734 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1735 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1736 // The pow or exp intrinsic returned a NaN, which requires a call 1737 // to the runtime. Recompile with the runtime call. 1738 uncommon_trap(Deoptimization::Reason_intrinsic, 1739 Deoptimization::Action_make_not_entrant); 1740 } 1741 return result; 1742 } else { 1743 // If this inlining ever returned NaN in the past, we compile a call 1744 // to the runtime to properly handle corner cases 1745 1746 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1747 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1748 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1749 1750 if (!if_slow->is_top()) { 1751 RegionNode* result_region = new RegionNode(3); 1752 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1753 1754 result_region->init_req(1, if_fast); 1755 result_val->init_req(1, result); 1756 1757 set_control(if_slow); 1758 1759 const TypePtr* no_memory_effects = NULL; 1760 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1761 no_memory_effects, 1762 x, top(), y, y ? top() : NULL); 1763 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1764 #ifdef ASSERT 1765 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1766 assert(value_top == top(), "second value must be top"); 1767 #endif 1768 1769 result_region->init_req(2, control()); 1770 result_val->init_req(2, value); 1771 set_control(_gvn.transform(result_region)); 1772 return _gvn.transform(result_val); 1773 } else { 1774 return result; 1775 } 1776 } 1777 } 1778 1779 //------------------------------inline_exp------------------------------------- 1780 // Inline exp instructions, if possible. The Intel hardware only misses 1781 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1782 bool LibraryCallKit::inline_exp() { 1783 Node* arg = round_double_node(argument(0)); 1784 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1785 1786 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1787 set_result(n); 1788 1789 C->set_has_split_ifs(true); // Has chance for split-if optimization 1790 return true; 1791 } 1792 1793 //------------------------------inline_pow------------------------------------- 1794 // Inline power instructions, if possible. 1795 bool LibraryCallKit::inline_pow() { 1796 // Pseudocode for pow 1797 // if (y == 2) { 1798 // return x * x; 1799 // } else { 1800 // if (x <= 0.0) { 1801 // long longy = (long)y; 1802 // if ((double)longy == y) { // if y is long 1803 // if (y + 1 == y) longy = 0; // huge number: even 1804 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1805 // } else { 1806 // result = NaN; 1807 // } 1808 // } else { 1809 // result = DPow(x,y); 1810 // } 1811 // if (result != result)? { 1812 // result = uncommon_trap() or runtime_call(); 1813 // } 1814 // return result; 1815 // } 1816 1817 Node* x = round_double_node(argument(0)); 1818 Node* y = round_double_node(argument(2)); 1819 1820 Node* result = NULL; 1821 1822 Node* const_two_node = makecon(TypeD::make(2.0)); 1823 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1824 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1825 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1826 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1827 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1828 1829 RegionNode* region_node = new RegionNode(3); 1830 region_node->init_req(1, if_true); 1831 1832 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1833 // special case for x^y where y == 2, we can convert it to x * x 1834 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1835 1836 // set control to if_false since we will now process the false branch 1837 set_control(if_false); 1838 1839 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1840 // Short form: skip the fancy tests and just check for NaN result. 1841 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1842 } else { 1843 // If this inlining ever returned NaN in the past, include all 1844 // checks + call to the runtime. 1845 1846 // Set the merge point for If node with condition of (x <= 0.0) 1847 // There are four possible paths to region node and phi node 1848 RegionNode *r = new RegionNode(4); 1849 Node *phi = new PhiNode(r, Type::DOUBLE); 1850 1851 // Build the first if node: if (x <= 0.0) 1852 // Node for 0 constant 1853 Node *zeronode = makecon(TypeD::ZERO); 1854 // Check x:0 1855 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1856 // Check: If (x<=0) then go complex path 1857 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1858 // Branch either way 1859 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1860 // Fast path taken; set region slot 3 1861 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1862 r->init_req(3,fast_taken); // Capture fast-control 1863 1864 // Fast path not-taken, i.e. slow path 1865 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1866 1867 // Set fast path result 1868 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1869 phi->init_req(3, fast_result); 1870 1871 // Complex path 1872 // Build the second if node (if y is long) 1873 // Node for (long)y 1874 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1875 // Node for (double)((long) y) 1876 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1877 // Check (double)((long) y) : y 1878 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1879 // Check if (y isn't long) then go to slow path 1880 1881 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1882 // Branch either way 1883 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1884 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1885 1886 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1887 1888 // Calculate DPow(abs(x), y)*(1 & (long)y) 1889 // Node for constant 1 1890 Node *conone = longcon(1); 1891 // 1& (long)y 1892 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1893 1894 // A huge number is always even. Detect a huge number by checking 1895 // if y + 1 == y and set integer to be tested for parity to 0. 1896 // Required for corner case: 1897 // (long)9.223372036854776E18 = max_jlong 1898 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1899 // max_jlong is odd but 9.223372036854776E18 is even 1900 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1901 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1902 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1903 Node* correctedsign = NULL; 1904 if (ConditionalMoveLimit != 0) { 1905 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1906 } else { 1907 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1908 RegionNode *r = new RegionNode(3); 1909 Node *phi = new PhiNode(r, TypeLong::LONG); 1910 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1911 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1912 phi->init_req(1, signnode); 1913 phi->init_req(2, longcon(0)); 1914 correctedsign = _gvn.transform(phi); 1915 ylong_path = _gvn.transform(r); 1916 record_for_igvn(r); 1917 } 1918 1919 // zero node 1920 Node *conzero = longcon(0); 1921 // Check (1&(long)y)==0? 1922 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1923 // Check if (1&(long)y)!=0?, if so the result is negative 1924 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1925 // abs(x) 1926 Node *absx=_gvn.transform(new AbsDNode(x)); 1927 // abs(x)^y 1928 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1929 // -abs(x)^y 1930 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1931 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1932 Node *signresult = NULL; 1933 if (ConditionalMoveLimit != 0) { 1934 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1935 } else { 1936 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1937 RegionNode *r = new RegionNode(3); 1938 Node *phi = new PhiNode(r, Type::DOUBLE); 1939 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1940 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1941 phi->init_req(1, absxpowy); 1942 phi->init_req(2, negabsxpowy); 1943 signresult = _gvn.transform(phi); 1944 ylong_path = _gvn.transform(r); 1945 record_for_igvn(r); 1946 } 1947 // Set complex path fast result 1948 r->init_req(2, ylong_path); 1949 phi->init_req(2, signresult); 1950 1951 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1952 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1953 r->init_req(1,slow_path); 1954 phi->init_req(1,slow_result); 1955 1956 // Post merge 1957 set_control(_gvn.transform(r)); 1958 record_for_igvn(r); 1959 result = _gvn.transform(phi); 1960 } 1961 1962 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1963 1964 // control from finish_pow_exp is now input to the region node 1965 region_node->set_req(2, control()); 1966 // the result from finish_pow_exp is now input to the phi node 1967 phi_node->init_req(2, result); 1968 set_control(_gvn.transform(region_node)); 1969 record_for_igvn(region_node); 1970 set_result(_gvn.transform(phi_node)); 1971 1972 C->set_has_split_ifs(true); // Has chance for split-if optimization 1973 return true; 1974 } 1975 1976 //------------------------------runtime_math----------------------------- 1977 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1978 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1979 "must be (DD)D or (D)D type"); 1980 1981 // Inputs 1982 Node* a = round_double_node(argument(0)); 1983 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1984 1985 const TypePtr* no_memory_effects = NULL; 1986 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1987 no_memory_effects, 1988 a, top(), b, b ? top() : NULL); 1989 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1990 #ifdef ASSERT 1991 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1992 assert(value_top == top(), "second value must be top"); 1993 #endif 1994 1995 set_result(value); 1996 return true; 1997 } 1998 1999 //------------------------------inline_math_native----------------------------- 2000 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 2001 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 2002 switch (id) { 2003 // These intrinsics are not properly supported on all hardware 2004 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 2005 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 2006 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 2007 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2008 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2009 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2010 2011 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2012 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2013 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2014 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2015 2016 // These intrinsics are supported on all hardware 2017 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2018 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2019 2020 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2021 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2022 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2023 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2024 #undef FN_PTR 2025 2026 // These intrinsics are not yet correctly implemented 2027 case vmIntrinsics::_datan2: 2028 return false; 2029 2030 default: 2031 fatal_unexpected_iid(id); 2032 return false; 2033 } 2034 } 2035 2036 static bool is_simple_name(Node* n) { 2037 return (n->req() == 1 // constant 2038 || (n->is_Type() && n->as_Type()->type()->singleton()) 2039 || n->is_Proj() // parameter or return value 2040 || n->is_Phi() // local of some sort 2041 ); 2042 } 2043 2044 //----------------------------inline_min_max----------------------------------- 2045 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2046 set_result(generate_min_max(id, argument(0), argument(1))); 2047 return true; 2048 } 2049 2050 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2051 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2052 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2053 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2054 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2055 2056 { 2057 PreserveJVMState pjvms(this); 2058 PreserveReexecuteState preexecs(this); 2059 jvms()->set_should_reexecute(true); 2060 2061 set_control(slow_path); 2062 set_i_o(i_o()); 2063 2064 uncommon_trap(Deoptimization::Reason_intrinsic, 2065 Deoptimization::Action_none); 2066 } 2067 2068 set_control(fast_path); 2069 set_result(math); 2070 } 2071 2072 template <typename OverflowOp> 2073 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2074 typedef typename OverflowOp::MathOp MathOp; 2075 2076 MathOp* mathOp = new MathOp(arg1, arg2); 2077 Node* operation = _gvn.transform( mathOp ); 2078 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2079 inline_math_mathExact(operation, ofcheck); 2080 return true; 2081 } 2082 2083 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2084 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2085 } 2086 2087 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2088 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2089 } 2090 2091 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2092 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2093 } 2094 2095 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2096 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2097 } 2098 2099 bool LibraryCallKit::inline_math_negateExactI() { 2100 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2101 } 2102 2103 bool LibraryCallKit::inline_math_negateExactL() { 2104 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2105 } 2106 2107 bool LibraryCallKit::inline_math_multiplyExactI() { 2108 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2109 } 2110 2111 bool LibraryCallKit::inline_math_multiplyExactL() { 2112 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2113 } 2114 2115 Node* 2116 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2117 // These are the candidate return value: 2118 Node* xvalue = x0; 2119 Node* yvalue = y0; 2120 2121 if (xvalue == yvalue) { 2122 return xvalue; 2123 } 2124 2125 bool want_max = (id == vmIntrinsics::_max); 2126 2127 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2128 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2129 if (txvalue == NULL || tyvalue == NULL) return top(); 2130 // This is not really necessary, but it is consistent with a 2131 // hypothetical MaxINode::Value method: 2132 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2133 2134 // %%% This folding logic should (ideally) be in a different place. 2135 // Some should be inside IfNode, and there to be a more reliable 2136 // transformation of ?: style patterns into cmoves. We also want 2137 // more powerful optimizations around cmove and min/max. 2138 2139 // Try to find a dominating comparison of these guys. 2140 // It can simplify the index computation for Arrays.copyOf 2141 // and similar uses of System.arraycopy. 2142 // First, compute the normalized version of CmpI(x, y). 2143 int cmp_op = Op_CmpI; 2144 Node* xkey = xvalue; 2145 Node* ykey = yvalue; 2146 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2147 if (ideal_cmpxy->is_Cmp()) { 2148 // E.g., if we have CmpI(length - offset, count), 2149 // it might idealize to CmpI(length, count + offset) 2150 cmp_op = ideal_cmpxy->Opcode(); 2151 xkey = ideal_cmpxy->in(1); 2152 ykey = ideal_cmpxy->in(2); 2153 } 2154 2155 // Start by locating any relevant comparisons. 2156 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2157 Node* cmpxy = NULL; 2158 Node* cmpyx = NULL; 2159 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2160 Node* cmp = start_from->fast_out(k); 2161 if (cmp->outcnt() > 0 && // must have prior uses 2162 cmp->in(0) == NULL && // must be context-independent 2163 cmp->Opcode() == cmp_op) { // right kind of compare 2164 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2165 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2166 } 2167 } 2168 2169 const int NCMPS = 2; 2170 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2171 int cmpn; 2172 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2173 if (cmps[cmpn] != NULL) break; // find a result 2174 } 2175 if (cmpn < NCMPS) { 2176 // Look for a dominating test that tells us the min and max. 2177 int depth = 0; // Limit search depth for speed 2178 Node* dom = control(); 2179 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2180 if (++depth >= 100) break; 2181 Node* ifproj = dom; 2182 if (!ifproj->is_Proj()) continue; 2183 Node* iff = ifproj->in(0); 2184 if (!iff->is_If()) continue; 2185 Node* bol = iff->in(1); 2186 if (!bol->is_Bool()) continue; 2187 Node* cmp = bol->in(1); 2188 if (cmp == NULL) continue; 2189 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2190 if (cmps[cmpn] == cmp) break; 2191 if (cmpn == NCMPS) continue; 2192 BoolTest::mask btest = bol->as_Bool()->_test._test; 2193 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2194 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2195 // At this point, we know that 'x btest y' is true. 2196 switch (btest) { 2197 case BoolTest::eq: 2198 // They are proven equal, so we can collapse the min/max. 2199 // Either value is the answer. Choose the simpler. 2200 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2201 return yvalue; 2202 return xvalue; 2203 case BoolTest::lt: // x < y 2204 case BoolTest::le: // x <= y 2205 return (want_max ? yvalue : xvalue); 2206 case BoolTest::gt: // x > y 2207 case BoolTest::ge: // x >= y 2208 return (want_max ? xvalue : yvalue); 2209 } 2210 } 2211 } 2212 2213 // We failed to find a dominating test. 2214 // Let's pick a test that might GVN with prior tests. 2215 Node* best_bol = NULL; 2216 BoolTest::mask best_btest = BoolTest::illegal; 2217 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2218 Node* cmp = cmps[cmpn]; 2219 if (cmp == NULL) continue; 2220 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2221 Node* bol = cmp->fast_out(j); 2222 if (!bol->is_Bool()) continue; 2223 BoolTest::mask btest = bol->as_Bool()->_test._test; 2224 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2225 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2226 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2227 best_bol = bol->as_Bool(); 2228 best_btest = btest; 2229 } 2230 } 2231 } 2232 2233 Node* answer_if_true = NULL; 2234 Node* answer_if_false = NULL; 2235 switch (best_btest) { 2236 default: 2237 if (cmpxy == NULL) 2238 cmpxy = ideal_cmpxy; 2239 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2240 // and fall through: 2241 case BoolTest::lt: // x < y 2242 case BoolTest::le: // x <= y 2243 answer_if_true = (want_max ? yvalue : xvalue); 2244 answer_if_false = (want_max ? xvalue : yvalue); 2245 break; 2246 case BoolTest::gt: // x > y 2247 case BoolTest::ge: // x >= y 2248 answer_if_true = (want_max ? xvalue : yvalue); 2249 answer_if_false = (want_max ? yvalue : xvalue); 2250 break; 2251 } 2252 2253 jint hi, lo; 2254 if (want_max) { 2255 // We can sharpen the minimum. 2256 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2257 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2258 } else { 2259 // We can sharpen the maximum. 2260 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2261 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2262 } 2263 2264 // Use a flow-free graph structure, to avoid creating excess control edges 2265 // which could hinder other optimizations. 2266 // Since Math.min/max is often used with arraycopy, we want 2267 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2268 Node* cmov = CMoveNode::make(NULL, best_bol, 2269 answer_if_false, answer_if_true, 2270 TypeInt::make(lo, hi, widen)); 2271 2272 return _gvn.transform(cmov); 2273 2274 /* 2275 // This is not as desirable as it may seem, since Min and Max 2276 // nodes do not have a full set of optimizations. 2277 // And they would interfere, anyway, with 'if' optimizations 2278 // and with CMoveI canonical forms. 2279 switch (id) { 2280 case vmIntrinsics::_min: 2281 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2282 case vmIntrinsics::_max: 2283 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2284 default: 2285 ShouldNotReachHere(); 2286 } 2287 */ 2288 } 2289 2290 inline int 2291 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2292 const TypePtr* base_type = TypePtr::NULL_PTR; 2293 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2294 if (base_type == NULL) { 2295 // Unknown type. 2296 return Type::AnyPtr; 2297 } else if (base_type == TypePtr::NULL_PTR) { 2298 // Since this is a NULL+long form, we have to switch to a rawptr. 2299 base = _gvn.transform(new CastX2PNode(offset)); 2300 offset = MakeConX(0); 2301 return Type::RawPtr; 2302 } else if (base_type->base() == Type::RawPtr) { 2303 return Type::RawPtr; 2304 } else if (base_type->isa_oopptr()) { 2305 // Base is never null => always a heap address. 2306 if (base_type->ptr() == TypePtr::NotNull) { 2307 return Type::OopPtr; 2308 } 2309 // Offset is small => always a heap address. 2310 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2311 if (offset_type != NULL && 2312 base_type->offset() == 0 && // (should always be?) 2313 offset_type->_lo >= 0 && 2314 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2315 return Type::OopPtr; 2316 } 2317 // Otherwise, it might either be oop+off or NULL+addr. 2318 return Type::AnyPtr; 2319 } else { 2320 // No information: 2321 return Type::AnyPtr; 2322 } 2323 } 2324 2325 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2326 int kind = classify_unsafe_addr(base, offset); 2327 if (kind == Type::RawPtr) { 2328 return basic_plus_adr(top(), base, offset); 2329 } else { 2330 return basic_plus_adr(base, offset); 2331 } 2332 } 2333 2334 //--------------------------inline_number_methods----------------------------- 2335 // inline int Integer.numberOfLeadingZeros(int) 2336 // inline int Long.numberOfLeadingZeros(long) 2337 // 2338 // inline int Integer.numberOfTrailingZeros(int) 2339 // inline int Long.numberOfTrailingZeros(long) 2340 // 2341 // inline int Integer.bitCount(int) 2342 // inline int Long.bitCount(long) 2343 // 2344 // inline char Character.reverseBytes(char) 2345 // inline short Short.reverseBytes(short) 2346 // inline int Integer.reverseBytes(int) 2347 // inline long Long.reverseBytes(long) 2348 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2349 Node* arg = argument(0); 2350 Node* n; 2351 switch (id) { 2352 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2353 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2354 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2355 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2356 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2357 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2358 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2359 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2360 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2361 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2362 default: fatal_unexpected_iid(id); break; 2363 } 2364 set_result(_gvn.transform(n)); 2365 return true; 2366 } 2367 2368 //----------------------------inline_unsafe_access---------------------------- 2369 2370 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2371 2372 // Helper that guards and inserts a pre-barrier. 2373 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2374 Node* pre_val, bool need_mem_bar) { 2375 // We could be accessing the referent field of a reference object. If so, when G1 2376 // is enabled, we need to log the value in the referent field in an SATB buffer. 2377 // This routine performs some compile time filters and generates suitable 2378 // runtime filters that guard the pre-barrier code. 2379 // Also add memory barrier for non volatile load from the referent field 2380 // to prevent commoning of loads across safepoint. 2381 if (!UseG1GC && !need_mem_bar) 2382 return; 2383 2384 // Some compile time checks. 2385 2386 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2387 const TypeX* otype = offset->find_intptr_t_type(); 2388 if (otype != NULL && otype->is_con() && 2389 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2390 // Constant offset but not the reference_offset so just return 2391 return; 2392 } 2393 2394 // We only need to generate the runtime guards for instances. 2395 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2396 if (btype != NULL) { 2397 if (btype->isa_aryptr()) { 2398 // Array type so nothing to do 2399 return; 2400 } 2401 2402 const TypeInstPtr* itype = btype->isa_instptr(); 2403 if (itype != NULL) { 2404 // Can the klass of base_oop be statically determined to be 2405 // _not_ a sub-class of Reference and _not_ Object? 2406 ciKlass* klass = itype->klass(); 2407 if ( klass->is_loaded() && 2408 !klass->is_subtype_of(env()->Reference_klass()) && 2409 !env()->Object_klass()->is_subtype_of(klass)) { 2410 return; 2411 } 2412 } 2413 } 2414 2415 // The compile time filters did not reject base_oop/offset so 2416 // we need to generate the following runtime filters 2417 // 2418 // if (offset == java_lang_ref_Reference::_reference_offset) { 2419 // if (instance_of(base, java.lang.ref.Reference)) { 2420 // pre_barrier(_, pre_val, ...); 2421 // } 2422 // } 2423 2424 float likely = PROB_LIKELY( 0.999); 2425 float unlikely = PROB_UNLIKELY(0.999); 2426 2427 IdealKit ideal(this); 2428 #define __ ideal. 2429 2430 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2431 2432 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2433 // Update graphKit memory and control from IdealKit. 2434 sync_kit(ideal); 2435 2436 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2437 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2438 2439 // Update IdealKit memory and control from graphKit. 2440 __ sync_kit(this); 2441 2442 Node* one = __ ConI(1); 2443 // is_instof == 0 if base_oop == NULL 2444 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2445 2446 // Update graphKit from IdeakKit. 2447 sync_kit(ideal); 2448 2449 // Use the pre-barrier to record the value in the referent field 2450 pre_barrier(false /* do_load */, 2451 __ ctrl(), 2452 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2453 pre_val /* pre_val */, 2454 T_OBJECT); 2455 if (need_mem_bar) { 2456 // Add memory barrier to prevent commoning reads from this field 2457 // across safepoint since GC can change its value. 2458 insert_mem_bar(Op_MemBarCPUOrder); 2459 } 2460 // Update IdealKit from graphKit. 2461 __ sync_kit(this); 2462 2463 } __ end_if(); // _ref_type != ref_none 2464 } __ end_if(); // offset == referent_offset 2465 2466 // Final sync IdealKit and GraphKit. 2467 final_sync(ideal); 2468 #undef __ 2469 } 2470 2471 2472 // Interpret Unsafe.fieldOffset cookies correctly: 2473 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2474 2475 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2476 // Attempt to infer a sharper value type from the offset and base type. 2477 ciKlass* sharpened_klass = NULL; 2478 2479 // See if it is an instance field, with an object type. 2480 if (alias_type->field() != NULL) { 2481 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2482 if (alias_type->field()->type()->is_klass()) { 2483 sharpened_klass = alias_type->field()->type()->as_klass(); 2484 } 2485 } 2486 2487 // See if it is a narrow oop array. 2488 if (adr_type->isa_aryptr()) { 2489 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2490 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2491 if (elem_type != NULL) { 2492 sharpened_klass = elem_type->klass(); 2493 } 2494 } 2495 } 2496 2497 // The sharpened class might be unloaded if there is no class loader 2498 // contraint in place. 2499 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2500 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2501 2502 #ifndef PRODUCT 2503 if (C->print_intrinsics() || C->print_inlining()) { 2504 tty->print(" from base type: "); adr_type->dump(); 2505 tty->print(" sharpened value: "); tjp->dump(); 2506 } 2507 #endif 2508 // Sharpen the value type. 2509 return tjp; 2510 } 2511 return NULL; 2512 } 2513 2514 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2515 if (callee()->is_static()) return false; // caller must have the capability! 2516 2517 #ifndef PRODUCT 2518 { 2519 ResourceMark rm; 2520 // Check the signatures. 2521 ciSignature* sig = callee()->signature(); 2522 #ifdef ASSERT 2523 if (!is_store) { 2524 // Object getObject(Object base, int/long offset), etc. 2525 BasicType rtype = sig->return_type()->basic_type(); 2526 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2527 rtype = T_ADDRESS; // it is really a C void* 2528 assert(rtype == type, "getter must return the expected value"); 2529 if (!is_native_ptr) { 2530 assert(sig->count() == 2, "oop getter has 2 arguments"); 2531 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2532 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2533 } else { 2534 assert(sig->count() == 1, "native getter has 1 argument"); 2535 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2536 } 2537 } else { 2538 // void putObject(Object base, int/long offset, Object x), etc. 2539 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2540 if (!is_native_ptr) { 2541 assert(sig->count() == 3, "oop putter has 3 arguments"); 2542 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2543 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2544 } else { 2545 assert(sig->count() == 2, "native putter has 2 arguments"); 2546 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2547 } 2548 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2549 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2550 vtype = T_ADDRESS; // it is really a C void* 2551 assert(vtype == type, "putter must accept the expected value"); 2552 } 2553 #endif // ASSERT 2554 } 2555 #endif //PRODUCT 2556 2557 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2558 2559 Node* receiver = argument(0); // type: oop 2560 2561 // Build address expression. 2562 Node* adr; 2563 Node* heap_base_oop = top(); 2564 Node* offset = top(); 2565 Node* val; 2566 2567 if (!is_native_ptr) { 2568 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2569 Node* base = argument(1); // type: oop 2570 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2571 offset = argument(2); // type: long 2572 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2573 // to be plain byte offsets, which are also the same as those accepted 2574 // by oopDesc::field_base. 2575 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2576 "fieldOffset must be byte-scaled"); 2577 // 32-bit machines ignore the high half! 2578 offset = ConvL2X(offset); 2579 adr = make_unsafe_address(base, offset); 2580 heap_base_oop = base; 2581 val = is_store ? argument(4) : NULL; 2582 } else { 2583 Node* ptr = argument(1); // type: long 2584 ptr = ConvL2X(ptr); // adjust Java long to machine word 2585 adr = make_unsafe_address(NULL, ptr); 2586 val = is_store ? argument(3) : NULL; 2587 } 2588 2589 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2590 2591 // First guess at the value type. 2592 const Type *value_type = Type::get_const_basic_type(type); 2593 2594 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2595 // there was not enough information to nail it down. 2596 Compile::AliasType* alias_type = C->alias_type(adr_type); 2597 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2598 2599 // We will need memory barriers unless we can determine a unique 2600 // alias category for this reference. (Note: If for some reason 2601 // the barriers get omitted and the unsafe reference begins to "pollute" 2602 // the alias analysis of the rest of the graph, either Compile::can_alias 2603 // or Compile::must_alias will throw a diagnostic assert.) 2604 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2605 2606 // If we are reading the value of the referent field of a Reference 2607 // object (either by using Unsafe directly or through reflection) 2608 // then, if G1 is enabled, we need to record the referent in an 2609 // SATB log buffer using the pre-barrier mechanism. 2610 // Also we need to add memory barrier to prevent commoning reads 2611 // from this field across safepoint since GC can change its value. 2612 bool need_read_barrier = !is_native_ptr && !is_store && 2613 offset != top() && heap_base_oop != top(); 2614 2615 if (!is_store && type == T_OBJECT) { 2616 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2617 if (tjp != NULL) { 2618 value_type = tjp; 2619 } 2620 } 2621 2622 receiver = null_check(receiver); 2623 if (stopped()) { 2624 return true; 2625 } 2626 // Heap pointers get a null-check from the interpreter, 2627 // as a courtesy. However, this is not guaranteed by Unsafe, 2628 // and it is not possible to fully distinguish unintended nulls 2629 // from intended ones in this API. 2630 2631 if (is_volatile) { 2632 // We need to emit leading and trailing CPU membars (see below) in 2633 // addition to memory membars when is_volatile. This is a little 2634 // too strong, but avoids the need to insert per-alias-type 2635 // volatile membars (for stores; compare Parse::do_put_xxx), which 2636 // we cannot do effectively here because we probably only have a 2637 // rough approximation of type. 2638 need_mem_bar = true; 2639 // For Stores, place a memory ordering barrier now. 2640 if (is_store) { 2641 insert_mem_bar(Op_MemBarRelease); 2642 } else { 2643 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2644 insert_mem_bar(Op_MemBarVolatile); 2645 } 2646 } 2647 } 2648 2649 // Memory barrier to prevent normal and 'unsafe' accesses from 2650 // bypassing each other. Happens after null checks, so the 2651 // exception paths do not take memory state from the memory barrier, 2652 // so there's no problems making a strong assert about mixing users 2653 // of safe & unsafe memory. 2654 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2655 2656 if (!is_store) { 2657 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2658 // To be valid, unsafe loads may depend on other conditions than 2659 // the one that guards them: pin the Load node 2660 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile); 2661 // load value 2662 switch (type) { 2663 case T_BOOLEAN: 2664 case T_CHAR: 2665 case T_BYTE: 2666 case T_SHORT: 2667 case T_INT: 2668 case T_LONG: 2669 case T_FLOAT: 2670 case T_DOUBLE: 2671 break; 2672 case T_OBJECT: 2673 if (need_read_barrier) { 2674 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2675 } 2676 break; 2677 case T_ADDRESS: 2678 // Cast to an int type. 2679 p = _gvn.transform(new CastP2XNode(NULL, p)); 2680 p = ConvX2UL(p); 2681 break; 2682 default: 2683 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2684 break; 2685 } 2686 // The load node has the control of the preceding MemBarCPUOrder. All 2687 // following nodes will have the control of the MemBarCPUOrder inserted at 2688 // the end of this method. So, pushing the load onto the stack at a later 2689 // point is fine. 2690 set_result(p); 2691 } else { 2692 // place effect of store into memory 2693 switch (type) { 2694 case T_DOUBLE: 2695 val = dstore_rounding(val); 2696 break; 2697 case T_ADDRESS: 2698 // Repackage the long as a pointer. 2699 val = ConvL2X(val); 2700 val = _gvn.transform(new CastX2PNode(val)); 2701 break; 2702 } 2703 2704 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2705 if (type != T_OBJECT ) { 2706 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2707 } else { 2708 // Possibly an oop being stored to Java heap or native memory 2709 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2710 // oop to Java heap. 2711 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2712 } else { 2713 // We can't tell at compile time if we are storing in the Java heap or outside 2714 // of it. So we need to emit code to conditionally do the proper type of 2715 // store. 2716 2717 IdealKit ideal(this); 2718 #define __ ideal. 2719 // QQQ who knows what probability is here?? 2720 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2721 // Sync IdealKit and graphKit. 2722 sync_kit(ideal); 2723 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2724 // Update IdealKit memory. 2725 __ sync_kit(this); 2726 } __ else_(); { 2727 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2728 } __ end_if(); 2729 // Final sync IdealKit and GraphKit. 2730 final_sync(ideal); 2731 #undef __ 2732 } 2733 } 2734 } 2735 2736 if (is_volatile) { 2737 if (!is_store) { 2738 insert_mem_bar(Op_MemBarAcquire); 2739 } else { 2740 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2741 insert_mem_bar(Op_MemBarVolatile); 2742 } 2743 } 2744 } 2745 2746 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2747 2748 return true; 2749 } 2750 2751 //----------------------------inline_unsafe_load_store---------------------------- 2752 // This method serves a couple of different customers (depending on LoadStoreKind): 2753 // 2754 // LS_cmpxchg: 2755 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2756 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2757 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2758 // 2759 // LS_xadd: 2760 // public int getAndAddInt( Object o, long offset, int delta) 2761 // public long getAndAddLong(Object o, long offset, long delta) 2762 // 2763 // LS_xchg: 2764 // int getAndSet(Object o, long offset, int newValue) 2765 // long getAndSet(Object o, long offset, long newValue) 2766 // Object getAndSet(Object o, long offset, Object newValue) 2767 // 2768 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2769 // This basic scheme here is the same as inline_unsafe_access, but 2770 // differs in enough details that combining them would make the code 2771 // overly confusing. (This is a true fact! I originally combined 2772 // them, but even I was confused by it!) As much code/comments as 2773 // possible are retained from inline_unsafe_access though to make 2774 // the correspondences clearer. - dl 2775 2776 if (callee()->is_static()) return false; // caller must have the capability! 2777 2778 #ifndef PRODUCT 2779 BasicType rtype; 2780 { 2781 ResourceMark rm; 2782 // Check the signatures. 2783 ciSignature* sig = callee()->signature(); 2784 rtype = sig->return_type()->basic_type(); 2785 if (kind == LS_xadd || kind == LS_xchg) { 2786 // Check the signatures. 2787 #ifdef ASSERT 2788 assert(rtype == type, "get and set must return the expected type"); 2789 assert(sig->count() == 3, "get and set has 3 arguments"); 2790 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2791 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2792 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2793 #endif // ASSERT 2794 } else if (kind == LS_cmpxchg) { 2795 // Check the signatures. 2796 #ifdef ASSERT 2797 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2798 assert(sig->count() == 4, "CAS has 4 arguments"); 2799 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2800 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2801 #endif // ASSERT 2802 } else { 2803 ShouldNotReachHere(); 2804 } 2805 } 2806 #endif //PRODUCT 2807 2808 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2809 2810 // Get arguments: 2811 Node* receiver = NULL; 2812 Node* base = NULL; 2813 Node* offset = NULL; 2814 Node* oldval = NULL; 2815 Node* newval = NULL; 2816 if (kind == LS_cmpxchg) { 2817 const bool two_slot_type = type2size[type] == 2; 2818 receiver = argument(0); // type: oop 2819 base = argument(1); // type: oop 2820 offset = argument(2); // type: long 2821 oldval = argument(4); // type: oop, int, or long 2822 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2823 } else if (kind == LS_xadd || kind == LS_xchg){ 2824 receiver = argument(0); // type: oop 2825 base = argument(1); // type: oop 2826 offset = argument(2); // type: long 2827 oldval = NULL; 2828 newval = argument(4); // type: oop, int, or long 2829 } 2830 2831 // Null check receiver. 2832 receiver = null_check(receiver); 2833 if (stopped()) { 2834 return true; 2835 } 2836 2837 // Build field offset expression. 2838 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2839 // to be plain byte offsets, which are also the same as those accepted 2840 // by oopDesc::field_base. 2841 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2842 // 32-bit machines ignore the high half of long offsets 2843 offset = ConvL2X(offset); 2844 Node* adr = make_unsafe_address(base, offset); 2845 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2846 2847 // For CAS, unlike inline_unsafe_access, there seems no point in 2848 // trying to refine types. Just use the coarse types here. 2849 const Type *value_type = Type::get_const_basic_type(type); 2850 Compile::AliasType* alias_type = C->alias_type(adr_type); 2851 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2852 2853 if (kind == LS_xchg && type == T_OBJECT) { 2854 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2855 if (tjp != NULL) { 2856 value_type = tjp; 2857 } 2858 } 2859 2860 int alias_idx = C->get_alias_index(adr_type); 2861 2862 // Memory-model-wise, a LoadStore acts like a little synchronized 2863 // block, so needs barriers on each side. These don't translate 2864 // into actual barriers on most machines, but we still need rest of 2865 // compiler to respect ordering. 2866 2867 insert_mem_bar(Op_MemBarRelease); 2868 insert_mem_bar(Op_MemBarCPUOrder); 2869 2870 // 4984716: MemBars must be inserted before this 2871 // memory node in order to avoid a false 2872 // dependency which will confuse the scheduler. 2873 Node *mem = memory(alias_idx); 2874 2875 // For now, we handle only those cases that actually exist: ints, 2876 // longs, and Object. Adding others should be straightforward. 2877 Node* load_store; 2878 switch(type) { 2879 case T_INT: 2880 if (kind == LS_xadd) { 2881 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2882 } else if (kind == LS_xchg) { 2883 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2884 } else if (kind == LS_cmpxchg) { 2885 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2886 } else { 2887 ShouldNotReachHere(); 2888 } 2889 break; 2890 case T_LONG: 2891 if (kind == LS_xadd) { 2892 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2893 } else if (kind == LS_xchg) { 2894 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2895 } else if (kind == LS_cmpxchg) { 2896 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2897 } else { 2898 ShouldNotReachHere(); 2899 } 2900 break; 2901 case T_OBJECT: 2902 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2903 // could be delayed during Parse (for example, in adjust_map_after_if()). 2904 // Execute transformation here to avoid barrier generation in such case. 2905 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2906 newval = _gvn.makecon(TypePtr::NULL_PTR); 2907 2908 // Reference stores need a store barrier. 2909 if (kind == LS_xchg) { 2910 // If pre-barrier must execute before the oop store, old value will require do_load here. 2911 if (!can_move_pre_barrier()) { 2912 pre_barrier(true /* do_load*/, 2913 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2914 NULL /* pre_val*/, 2915 T_OBJECT); 2916 } // Else move pre_barrier to use load_store value, see below. 2917 } else if (kind == LS_cmpxchg) { 2918 // Same as for newval above: 2919 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2920 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2921 } 2922 // The only known value which might get overwritten is oldval. 2923 pre_barrier(false /* do_load */, 2924 control(), NULL, NULL, max_juint, NULL, NULL, 2925 oldval /* pre_val */, 2926 T_OBJECT); 2927 } else { 2928 ShouldNotReachHere(); 2929 } 2930 2931 #ifdef _LP64 2932 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2933 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2934 if (kind == LS_xchg) { 2935 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2936 newval_enc, adr_type, value_type->make_narrowoop())); 2937 } else { 2938 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2939 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2940 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2941 newval_enc, oldval_enc)); 2942 } 2943 } else 2944 #endif 2945 { 2946 if (kind == LS_xchg) { 2947 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2948 } else { 2949 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2950 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2951 } 2952 } 2953 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2954 break; 2955 default: 2956 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2957 break; 2958 } 2959 2960 // SCMemProjNodes represent the memory state of a LoadStore. Their 2961 // main role is to prevent LoadStore nodes from being optimized away 2962 // when their results aren't used. 2963 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2964 set_memory(proj, alias_idx); 2965 2966 if (type == T_OBJECT && kind == LS_xchg) { 2967 #ifdef _LP64 2968 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2969 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2970 } 2971 #endif 2972 if (can_move_pre_barrier()) { 2973 // Don't need to load pre_val. The old value is returned by load_store. 2974 // The pre_barrier can execute after the xchg as long as no safepoint 2975 // gets inserted between them. 2976 pre_barrier(false /* do_load */, 2977 control(), NULL, NULL, max_juint, NULL, NULL, 2978 load_store /* pre_val */, 2979 T_OBJECT); 2980 } 2981 } 2982 2983 // Add the trailing membar surrounding the access 2984 insert_mem_bar(Op_MemBarCPUOrder); 2985 insert_mem_bar(Op_MemBarAcquire); 2986 2987 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2988 set_result(load_store); 2989 return true; 2990 } 2991 2992 //----------------------------inline_unsafe_ordered_store---------------------- 2993 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2994 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 2995 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 2996 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2997 // This is another variant of inline_unsafe_access, differing in 2998 // that it always issues store-store ("release") barrier and ensures 2999 // store-atomicity (which only matters for "long"). 3000 3001 if (callee()->is_static()) return false; // caller must have the capability! 3002 3003 #ifndef PRODUCT 3004 { 3005 ResourceMark rm; 3006 // Check the signatures. 3007 ciSignature* sig = callee()->signature(); 3008 #ifdef ASSERT 3009 BasicType rtype = sig->return_type()->basic_type(); 3010 assert(rtype == T_VOID, "must return void"); 3011 assert(sig->count() == 3, "has 3 arguments"); 3012 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3013 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3014 #endif // ASSERT 3015 } 3016 #endif //PRODUCT 3017 3018 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3019 3020 // Get arguments: 3021 Node* receiver = argument(0); // type: oop 3022 Node* base = argument(1); // type: oop 3023 Node* offset = argument(2); // type: long 3024 Node* val = argument(4); // type: oop, int, or long 3025 3026 // Null check receiver. 3027 receiver = null_check(receiver); 3028 if (stopped()) { 3029 return true; 3030 } 3031 3032 // Build field offset expression. 3033 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3034 // 32-bit machines ignore the high half of long offsets 3035 offset = ConvL2X(offset); 3036 Node* adr = make_unsafe_address(base, offset); 3037 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3038 const Type *value_type = Type::get_const_basic_type(type); 3039 Compile::AliasType* alias_type = C->alias_type(adr_type); 3040 3041 insert_mem_bar(Op_MemBarRelease); 3042 insert_mem_bar(Op_MemBarCPUOrder); 3043 // Ensure that the store is atomic for longs: 3044 const bool require_atomic_access = true; 3045 Node* store; 3046 if (type == T_OBJECT) // reference stores need a store barrier. 3047 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3048 else { 3049 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3050 } 3051 insert_mem_bar(Op_MemBarCPUOrder); 3052 return true; 3053 } 3054 3055 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3056 // Regardless of form, don't allow previous ld/st to move down, 3057 // then issue acquire, release, or volatile mem_bar. 3058 insert_mem_bar(Op_MemBarCPUOrder); 3059 switch(id) { 3060 case vmIntrinsics::_loadFence: 3061 insert_mem_bar(Op_LoadFence); 3062 return true; 3063 case vmIntrinsics::_storeFence: 3064 insert_mem_bar(Op_StoreFence); 3065 return true; 3066 case vmIntrinsics::_fullFence: 3067 insert_mem_bar(Op_MemBarVolatile); 3068 return true; 3069 default: 3070 fatal_unexpected_iid(id); 3071 return false; 3072 } 3073 } 3074 3075 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3076 if (!kls->is_Con()) { 3077 return true; 3078 } 3079 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3080 if (klsptr == NULL) { 3081 return true; 3082 } 3083 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3084 // don't need a guard for a klass that is already initialized 3085 return !ik->is_initialized(); 3086 } 3087 3088 //----------------------------inline_unsafe_allocate--------------------------- 3089 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3090 bool LibraryCallKit::inline_unsafe_allocate() { 3091 if (callee()->is_static()) return false; // caller must have the capability! 3092 3093 null_check_receiver(); // null-check, then ignore 3094 Node* cls = null_check(argument(1)); 3095 if (stopped()) return true; 3096 3097 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3098 kls = null_check(kls); 3099 if (stopped()) return true; // argument was like int.class 3100 3101 Node* test = NULL; 3102 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3103 // Note: The argument might still be an illegal value like 3104 // Serializable.class or Object[].class. The runtime will handle it. 3105 // But we must make an explicit check for initialization. 3106 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3107 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3108 // can generate code to load it as unsigned byte. 3109 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3110 Node* bits = intcon(InstanceKlass::fully_initialized); 3111 test = _gvn.transform(new SubINode(inst, bits)); 3112 // The 'test' is non-zero if we need to take a slow path. 3113 } 3114 3115 Node* obj = new_instance(kls, test); 3116 set_result(obj); 3117 return true; 3118 } 3119 3120 #ifdef TRACE_HAVE_INTRINSICS 3121 /* 3122 * oop -> myklass 3123 * myklass->trace_id |= USED 3124 * return myklass->trace_id & ~0x3 3125 */ 3126 bool LibraryCallKit::inline_native_classID() { 3127 null_check_receiver(); // null-check, then ignore 3128 Node* cls = null_check(argument(1), T_OBJECT); 3129 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3130 kls = null_check(kls, T_OBJECT); 3131 ByteSize offset = TRACE_ID_OFFSET; 3132 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3133 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3134 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3135 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3136 Node* clsused = longcon(0x01l); // set the class bit 3137 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3138 3139 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3140 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3141 set_result(andl); 3142 return true; 3143 } 3144 3145 bool LibraryCallKit::inline_native_threadID() { 3146 Node* tls_ptr = NULL; 3147 Node* cur_thr = generate_current_thread(tls_ptr); 3148 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3149 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3150 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3151 3152 Node* threadid = NULL; 3153 size_t thread_id_size = OSThread::thread_id_size(); 3154 if (thread_id_size == (size_t) BytesPerLong) { 3155 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3156 } else if (thread_id_size == (size_t) BytesPerInt) { 3157 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3158 } else { 3159 ShouldNotReachHere(); 3160 } 3161 set_result(threadid); 3162 return true; 3163 } 3164 #endif 3165 3166 //------------------------inline_native_time_funcs-------------- 3167 // inline code for System.currentTimeMillis() and System.nanoTime() 3168 // these have the same type and signature 3169 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3170 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3171 const TypePtr* no_memory_effects = NULL; 3172 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3173 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3174 #ifdef ASSERT 3175 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3176 assert(value_top == top(), "second value must be top"); 3177 #endif 3178 set_result(value); 3179 return true; 3180 } 3181 3182 //------------------------inline_native_currentThread------------------ 3183 bool LibraryCallKit::inline_native_currentThread() { 3184 Node* junk = NULL; 3185 set_result(generate_current_thread(junk)); 3186 return true; 3187 } 3188 3189 //------------------------inline_native_isInterrupted------------------ 3190 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3191 bool LibraryCallKit::inline_native_isInterrupted() { 3192 // Add a fast path to t.isInterrupted(clear_int): 3193 // (t == Thread.current() && 3194 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3195 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3196 // So, in the common case that the interrupt bit is false, 3197 // we avoid making a call into the VM. Even if the interrupt bit 3198 // is true, if the clear_int argument is false, we avoid the VM call. 3199 // However, if the receiver is not currentThread, we must call the VM, 3200 // because there must be some locking done around the operation. 3201 3202 // We only go to the fast case code if we pass two guards. 3203 // Paths which do not pass are accumulated in the slow_region. 3204 3205 enum { 3206 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3207 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3208 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3209 PATH_LIMIT 3210 }; 3211 3212 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3213 // out of the function. 3214 insert_mem_bar(Op_MemBarCPUOrder); 3215 3216 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3217 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3218 3219 RegionNode* slow_region = new RegionNode(1); 3220 record_for_igvn(slow_region); 3221 3222 // (a) Receiving thread must be the current thread. 3223 Node* rec_thr = argument(0); 3224 Node* tls_ptr = NULL; 3225 Node* cur_thr = generate_current_thread(tls_ptr); 3226 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3227 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3228 3229 generate_slow_guard(bol_thr, slow_region); 3230 3231 // (b) Interrupt bit on TLS must be false. 3232 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3233 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3234 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3235 3236 // Set the control input on the field _interrupted read to prevent it floating up. 3237 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3238 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3239 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3240 3241 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3242 3243 // First fast path: if (!TLS._interrupted) return false; 3244 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3245 result_rgn->init_req(no_int_result_path, false_bit); 3246 result_val->init_req(no_int_result_path, intcon(0)); 3247 3248 // drop through to next case 3249 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3250 3251 #ifndef TARGET_OS_FAMILY_windows 3252 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3253 Node* clr_arg = argument(1); 3254 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3255 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3256 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3257 3258 // Second fast path: ... else if (!clear_int) return true; 3259 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3260 result_rgn->init_req(no_clear_result_path, false_arg); 3261 result_val->init_req(no_clear_result_path, intcon(1)); 3262 3263 // drop through to next case 3264 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3265 #else 3266 // To return true on Windows you must read the _interrupted field 3267 // and check the the event state i.e. take the slow path. 3268 #endif // TARGET_OS_FAMILY_windows 3269 3270 // (d) Otherwise, go to the slow path. 3271 slow_region->add_req(control()); 3272 set_control( _gvn.transform(slow_region)); 3273 3274 if (stopped()) { 3275 // There is no slow path. 3276 result_rgn->init_req(slow_result_path, top()); 3277 result_val->init_req(slow_result_path, top()); 3278 } else { 3279 // non-virtual because it is a private non-static 3280 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3281 3282 Node* slow_val = set_results_for_java_call(slow_call); 3283 // this->control() comes from set_results_for_java_call 3284 3285 Node* fast_io = slow_call->in(TypeFunc::I_O); 3286 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3287 3288 // These two phis are pre-filled with copies of of the fast IO and Memory 3289 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3290 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3291 3292 result_rgn->init_req(slow_result_path, control()); 3293 result_io ->init_req(slow_result_path, i_o()); 3294 result_mem->init_req(slow_result_path, reset_memory()); 3295 result_val->init_req(slow_result_path, slow_val); 3296 3297 set_all_memory(_gvn.transform(result_mem)); 3298 set_i_o( _gvn.transform(result_io)); 3299 } 3300 3301 C->set_has_split_ifs(true); // Has chance for split-if optimization 3302 set_result(result_rgn, result_val); 3303 return true; 3304 } 3305 3306 //---------------------------load_mirror_from_klass---------------------------- 3307 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3308 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3309 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3310 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3311 } 3312 3313 //-----------------------load_klass_from_mirror_common------------------------- 3314 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3315 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3316 // and branch to the given path on the region. 3317 // If never_see_null, take an uncommon trap on null, so we can optimistically 3318 // compile for the non-null case. 3319 // If the region is NULL, force never_see_null = true. 3320 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3321 bool never_see_null, 3322 RegionNode* region, 3323 int null_path, 3324 int offset) { 3325 if (region == NULL) never_see_null = true; 3326 Node* p = basic_plus_adr(mirror, offset); 3327 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3328 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3329 Node* null_ctl = top(); 3330 kls = null_check_oop(kls, &null_ctl, never_see_null); 3331 if (region != NULL) { 3332 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3333 region->init_req(null_path, null_ctl); 3334 } else { 3335 assert(null_ctl == top(), "no loose ends"); 3336 } 3337 return kls; 3338 } 3339 3340 //--------------------(inline_native_Class_query helpers)--------------------- 3341 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3342 // Fall through if (mods & mask) == bits, take the guard otherwise. 3343 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3344 // Branch around if the given klass has the given modifier bit set. 3345 // Like generate_guard, adds a new path onto the region. 3346 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3347 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3348 Node* mask = intcon(modifier_mask); 3349 Node* bits = intcon(modifier_bits); 3350 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3351 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3352 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3353 return generate_fair_guard(bol, region); 3354 } 3355 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3356 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3357 } 3358 3359 //-------------------------inline_native_Class_query------------------- 3360 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3361 const Type* return_type = TypeInt::BOOL; 3362 Node* prim_return_value = top(); // what happens if it's a primitive class? 3363 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3364 bool expect_prim = false; // most of these guys expect to work on refs 3365 3366 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3367 3368 Node* mirror = argument(0); 3369 Node* obj = top(); 3370 3371 switch (id) { 3372 case vmIntrinsics::_isInstance: 3373 // nothing is an instance of a primitive type 3374 prim_return_value = intcon(0); 3375 obj = argument(1); 3376 break; 3377 case vmIntrinsics::_getModifiers: 3378 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3379 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3380 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3381 break; 3382 case vmIntrinsics::_isInterface: 3383 prim_return_value = intcon(0); 3384 break; 3385 case vmIntrinsics::_isArray: 3386 prim_return_value = intcon(0); 3387 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3388 break; 3389 case vmIntrinsics::_isPrimitive: 3390 prim_return_value = intcon(1); 3391 expect_prim = true; // obviously 3392 break; 3393 case vmIntrinsics::_getSuperclass: 3394 prim_return_value = null(); 3395 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3396 break; 3397 case vmIntrinsics::_getClassAccessFlags: 3398 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3399 return_type = TypeInt::INT; // not bool! 6297094 3400 break; 3401 default: 3402 fatal_unexpected_iid(id); 3403 break; 3404 } 3405 3406 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3407 if (mirror_con == NULL) return false; // cannot happen? 3408 3409 #ifndef PRODUCT 3410 if (C->print_intrinsics() || C->print_inlining()) { 3411 ciType* k = mirror_con->java_mirror_type(); 3412 if (k) { 3413 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3414 k->print_name(); 3415 tty->cr(); 3416 } 3417 } 3418 #endif 3419 3420 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3421 RegionNode* region = new RegionNode(PATH_LIMIT); 3422 record_for_igvn(region); 3423 PhiNode* phi = new PhiNode(region, return_type); 3424 3425 // The mirror will never be null of Reflection.getClassAccessFlags, however 3426 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3427 // if it is. See bug 4774291. 3428 3429 // For Reflection.getClassAccessFlags(), the null check occurs in 3430 // the wrong place; see inline_unsafe_access(), above, for a similar 3431 // situation. 3432 mirror = null_check(mirror); 3433 // If mirror or obj is dead, only null-path is taken. 3434 if (stopped()) return true; 3435 3436 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3437 3438 // Now load the mirror's klass metaobject, and null-check it. 3439 // Side-effects region with the control path if the klass is null. 3440 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3441 // If kls is null, we have a primitive mirror. 3442 phi->init_req(_prim_path, prim_return_value); 3443 if (stopped()) { set_result(region, phi); return true; } 3444 bool safe_for_replace = (region->in(_prim_path) == top()); 3445 3446 Node* p; // handy temp 3447 Node* null_ctl; 3448 3449 // Now that we have the non-null klass, we can perform the real query. 3450 // For constant classes, the query will constant-fold in LoadNode::Value. 3451 Node* query_value = top(); 3452 switch (id) { 3453 case vmIntrinsics::_isInstance: 3454 // nothing is an instance of a primitive type 3455 query_value = gen_instanceof(obj, kls, safe_for_replace); 3456 break; 3457 3458 case vmIntrinsics::_getModifiers: 3459 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3460 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3461 break; 3462 3463 case vmIntrinsics::_isInterface: 3464 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3465 if (generate_interface_guard(kls, region) != NULL) 3466 // A guard was added. If the guard is taken, it was an interface. 3467 phi->add_req(intcon(1)); 3468 // If we fall through, it's a plain class. 3469 query_value = intcon(0); 3470 break; 3471 3472 case vmIntrinsics::_isArray: 3473 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3474 if (generate_array_guard(kls, region) != NULL) 3475 // A guard was added. If the guard is taken, it was an array. 3476 phi->add_req(intcon(1)); 3477 // If we fall through, it's a plain class. 3478 query_value = intcon(0); 3479 break; 3480 3481 case vmIntrinsics::_isPrimitive: 3482 query_value = intcon(0); // "normal" path produces false 3483 break; 3484 3485 case vmIntrinsics::_getSuperclass: 3486 // The rules here are somewhat unfortunate, but we can still do better 3487 // with random logic than with a JNI call. 3488 // Interfaces store null or Object as _super, but must report null. 3489 // Arrays store an intermediate super as _super, but must report Object. 3490 // Other types can report the actual _super. 3491 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3492 if (generate_interface_guard(kls, region) != NULL) 3493 // A guard was added. If the guard is taken, it was an interface. 3494 phi->add_req(null()); 3495 if (generate_array_guard(kls, region) != NULL) 3496 // A guard was added. If the guard is taken, it was an array. 3497 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3498 // If we fall through, it's a plain class. Get its _super. 3499 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3500 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3501 null_ctl = top(); 3502 kls = null_check_oop(kls, &null_ctl); 3503 if (null_ctl != top()) { 3504 // If the guard is taken, Object.superClass is null (both klass and mirror). 3505 region->add_req(null_ctl); 3506 phi ->add_req(null()); 3507 } 3508 if (!stopped()) { 3509 query_value = load_mirror_from_klass(kls); 3510 } 3511 break; 3512 3513 case vmIntrinsics::_getClassAccessFlags: 3514 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3515 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3516 break; 3517 3518 default: 3519 fatal_unexpected_iid(id); 3520 break; 3521 } 3522 3523 // Fall-through is the normal case of a query to a real class. 3524 phi->init_req(1, query_value); 3525 region->init_req(1, control()); 3526 3527 C->set_has_split_ifs(true); // Has chance for split-if optimization 3528 set_result(region, phi); 3529 return true; 3530 } 3531 3532 //-------------------------inline_Class_cast------------------- 3533 bool LibraryCallKit::inline_Class_cast() { 3534 Node* mirror = argument(0); // Class 3535 Node* obj = argument(1); 3536 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3537 if (mirror_con == NULL) { 3538 return false; // dead path (mirror->is_top()). 3539 } 3540 if (obj == NULL || obj->is_top()) { 3541 return false; // dead path 3542 } 3543 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3544 3545 // First, see if Class.cast() can be folded statically. 3546 // java_mirror_type() returns non-null for compile-time Class constants. 3547 ciType* tm = mirror_con->java_mirror_type(); 3548 if (tm != NULL && tm->is_klass() && 3549 tp != NULL && tp->klass() != NULL) { 3550 if (!tp->klass()->is_loaded()) { 3551 // Don't use intrinsic when class is not loaded. 3552 return false; 3553 } else { 3554 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3555 if (static_res == Compile::SSC_always_true) { 3556 // isInstance() is true - fold the code. 3557 set_result(obj); 3558 return true; 3559 } else if (static_res == Compile::SSC_always_false) { 3560 // Don't use intrinsic, have to throw ClassCastException. 3561 // If the reference is null, the non-intrinsic bytecode will 3562 // be optimized appropriately. 3563 return false; 3564 } 3565 } 3566 } 3567 3568 // Bailout intrinsic and do normal inlining if exception path is frequent. 3569 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3570 return false; 3571 } 3572 3573 // Generate dynamic checks. 3574 // Class.cast() is java implementation of _checkcast bytecode. 3575 // Do checkcast (Parse::do_checkcast()) optimizations here. 3576 3577 mirror = null_check(mirror); 3578 // If mirror is dead, only null-path is taken. 3579 if (stopped()) { 3580 return true; 3581 } 3582 3583 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3584 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3585 RegionNode* region = new RegionNode(PATH_LIMIT); 3586 record_for_igvn(region); 3587 3588 // Now load the mirror's klass metaobject, and null-check it. 3589 // If kls is null, we have a primitive mirror and 3590 // nothing is an instance of a primitive type. 3591 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3592 3593 Node* res = top(); 3594 if (!stopped()) { 3595 Node* bad_type_ctrl = top(); 3596 // Do checkcast optimizations. 3597 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3598 region->init_req(_bad_type_path, bad_type_ctrl); 3599 } 3600 if (region->in(_prim_path) != top() || 3601 region->in(_bad_type_path) != top()) { 3602 // Let Interpreter throw ClassCastException. 3603 PreserveJVMState pjvms(this); 3604 set_control(_gvn.transform(region)); 3605 uncommon_trap(Deoptimization::Reason_intrinsic, 3606 Deoptimization::Action_maybe_recompile); 3607 } 3608 if (!stopped()) { 3609 set_result(res); 3610 } 3611 return true; 3612 } 3613 3614 3615 //--------------------------inline_native_subtype_check------------------------ 3616 // This intrinsic takes the JNI calls out of the heart of 3617 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3618 bool LibraryCallKit::inline_native_subtype_check() { 3619 // Pull both arguments off the stack. 3620 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3621 args[0] = argument(0); 3622 args[1] = argument(1); 3623 Node* klasses[2]; // corresponding Klasses: superk, subk 3624 klasses[0] = klasses[1] = top(); 3625 3626 enum { 3627 // A full decision tree on {superc is prim, subc is prim}: 3628 _prim_0_path = 1, // {P,N} => false 3629 // {P,P} & superc!=subc => false 3630 _prim_same_path, // {P,P} & superc==subc => true 3631 _prim_1_path, // {N,P} => false 3632 _ref_subtype_path, // {N,N} & subtype check wins => true 3633 _both_ref_path, // {N,N} & subtype check loses => false 3634 PATH_LIMIT 3635 }; 3636 3637 RegionNode* region = new RegionNode(PATH_LIMIT); 3638 Node* phi = new PhiNode(region, TypeInt::BOOL); 3639 record_for_igvn(region); 3640 3641 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3642 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3643 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3644 3645 // First null-check both mirrors and load each mirror's klass metaobject. 3646 int which_arg; 3647 for (which_arg = 0; which_arg <= 1; which_arg++) { 3648 Node* arg = args[which_arg]; 3649 arg = null_check(arg); 3650 if (stopped()) break; 3651 args[which_arg] = arg; 3652 3653 Node* p = basic_plus_adr(arg, class_klass_offset); 3654 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3655 klasses[which_arg] = _gvn.transform(kls); 3656 } 3657 3658 // Having loaded both klasses, test each for null. 3659 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3660 for (which_arg = 0; which_arg <= 1; which_arg++) { 3661 Node* kls = klasses[which_arg]; 3662 Node* null_ctl = top(); 3663 kls = null_check_oop(kls, &null_ctl, never_see_null); 3664 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3665 region->init_req(prim_path, null_ctl); 3666 if (stopped()) break; 3667 klasses[which_arg] = kls; 3668 } 3669 3670 if (!stopped()) { 3671 // now we have two reference types, in klasses[0..1] 3672 Node* subk = klasses[1]; // the argument to isAssignableFrom 3673 Node* superk = klasses[0]; // the receiver 3674 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3675 // now we have a successful reference subtype check 3676 region->set_req(_ref_subtype_path, control()); 3677 } 3678 3679 // If both operands are primitive (both klasses null), then 3680 // we must return true when they are identical primitives. 3681 // It is convenient to test this after the first null klass check. 3682 set_control(region->in(_prim_0_path)); // go back to first null check 3683 if (!stopped()) { 3684 // Since superc is primitive, make a guard for the superc==subc case. 3685 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3686 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3687 generate_guard(bol_eq, region, PROB_FAIR); 3688 if (region->req() == PATH_LIMIT+1) { 3689 // A guard was added. If the added guard is taken, superc==subc. 3690 region->swap_edges(PATH_LIMIT, _prim_same_path); 3691 region->del_req(PATH_LIMIT); 3692 } 3693 region->set_req(_prim_0_path, control()); // Not equal after all. 3694 } 3695 3696 // these are the only paths that produce 'true': 3697 phi->set_req(_prim_same_path, intcon(1)); 3698 phi->set_req(_ref_subtype_path, intcon(1)); 3699 3700 // pull together the cases: 3701 assert(region->req() == PATH_LIMIT, "sane region"); 3702 for (uint i = 1; i < region->req(); i++) { 3703 Node* ctl = region->in(i); 3704 if (ctl == NULL || ctl == top()) { 3705 region->set_req(i, top()); 3706 phi ->set_req(i, top()); 3707 } else if (phi->in(i) == NULL) { 3708 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3709 } 3710 } 3711 3712 set_control(_gvn.transform(region)); 3713 set_result(_gvn.transform(phi)); 3714 return true; 3715 } 3716 3717 //---------------------generate_array_guard_common------------------------ 3718 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3719 bool obj_array, bool not_array) { 3720 3721 if (stopped()) { 3722 return NULL; 3723 } 3724 3725 // If obj_array/non_array==false/false: 3726 // Branch around if the given klass is in fact an array (either obj or prim). 3727 // If obj_array/non_array==false/true: 3728 // Branch around if the given klass is not an array klass of any kind. 3729 // If obj_array/non_array==true/true: 3730 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3731 // If obj_array/non_array==true/false: 3732 // Branch around if the kls is an oop array (Object[] or subtype) 3733 // 3734 // Like generate_guard, adds a new path onto the region. 3735 jint layout_con = 0; 3736 Node* layout_val = get_layout_helper(kls, layout_con); 3737 if (layout_val == NULL) { 3738 bool query = (obj_array 3739 ? Klass::layout_helper_is_objArray(layout_con) 3740 : Klass::layout_helper_is_array(layout_con)); 3741 if (query == not_array) { 3742 return NULL; // never a branch 3743 } else { // always a branch 3744 Node* always_branch = control(); 3745 if (region != NULL) 3746 region->add_req(always_branch); 3747 set_control(top()); 3748 return always_branch; 3749 } 3750 } 3751 // Now test the correct condition. 3752 jint nval = (obj_array 3753 ? ((jint)Klass::_lh_array_tag_type_value 3754 << Klass::_lh_array_tag_shift) 3755 : Klass::_lh_neutral_value); 3756 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3757 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3758 // invert the test if we are looking for a non-array 3759 if (not_array) btest = BoolTest(btest).negate(); 3760 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3761 return generate_fair_guard(bol, region); 3762 } 3763 3764 3765 //-----------------------inline_native_newArray-------------------------- 3766 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3767 bool LibraryCallKit::inline_native_newArray() { 3768 Node* mirror = argument(0); 3769 Node* count_val = argument(1); 3770 3771 mirror = null_check(mirror); 3772 // If mirror or obj is dead, only null-path is taken. 3773 if (stopped()) return true; 3774 3775 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3776 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3777 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3778 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3779 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3780 3781 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3782 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3783 result_reg, _slow_path); 3784 Node* normal_ctl = control(); 3785 Node* no_array_ctl = result_reg->in(_slow_path); 3786 3787 // Generate code for the slow case. We make a call to newArray(). 3788 set_control(no_array_ctl); 3789 if (!stopped()) { 3790 // Either the input type is void.class, or else the 3791 // array klass has not yet been cached. Either the 3792 // ensuing call will throw an exception, or else it 3793 // will cache the array klass for next time. 3794 PreserveJVMState pjvms(this); 3795 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3796 Node* slow_result = set_results_for_java_call(slow_call); 3797 // this->control() comes from set_results_for_java_call 3798 result_reg->set_req(_slow_path, control()); 3799 result_val->set_req(_slow_path, slow_result); 3800 result_io ->set_req(_slow_path, i_o()); 3801 result_mem->set_req(_slow_path, reset_memory()); 3802 } 3803 3804 set_control(normal_ctl); 3805 if (!stopped()) { 3806 // Normal case: The array type has been cached in the java.lang.Class. 3807 // The following call works fine even if the array type is polymorphic. 3808 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3809 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3810 result_reg->init_req(_normal_path, control()); 3811 result_val->init_req(_normal_path, obj); 3812 result_io ->init_req(_normal_path, i_o()); 3813 result_mem->init_req(_normal_path, reset_memory()); 3814 } 3815 3816 // Return the combined state. 3817 set_i_o( _gvn.transform(result_io) ); 3818 set_all_memory( _gvn.transform(result_mem)); 3819 3820 C->set_has_split_ifs(true); // Has chance for split-if optimization 3821 set_result(result_reg, result_val); 3822 return true; 3823 } 3824 3825 //----------------------inline_native_getLength-------------------------- 3826 // public static native int java.lang.reflect.Array.getLength(Object array); 3827 bool LibraryCallKit::inline_native_getLength() { 3828 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3829 3830 Node* array = null_check(argument(0)); 3831 // If array is dead, only null-path is taken. 3832 if (stopped()) return true; 3833 3834 // Deoptimize if it is a non-array. 3835 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3836 3837 if (non_array != NULL) { 3838 PreserveJVMState pjvms(this); 3839 set_control(non_array); 3840 uncommon_trap(Deoptimization::Reason_intrinsic, 3841 Deoptimization::Action_maybe_recompile); 3842 } 3843 3844 // If control is dead, only non-array-path is taken. 3845 if (stopped()) return true; 3846 3847 // The works fine even if the array type is polymorphic. 3848 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3849 Node* result = load_array_length(array); 3850 3851 C->set_has_split_ifs(true); // Has chance for split-if optimization 3852 set_result(result); 3853 return true; 3854 } 3855 3856 //------------------------inline_array_copyOf---------------------------- 3857 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3858 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3859 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3860 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3861 3862 // Get the arguments. 3863 Node* original = argument(0); 3864 Node* start = is_copyOfRange? argument(1): intcon(0); 3865 Node* end = is_copyOfRange? argument(2): argument(1); 3866 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3867 3868 Node* newcopy; 3869 3870 // Set the original stack and the reexecute bit for the interpreter to reexecute 3871 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3872 { PreserveReexecuteState preexecs(this); 3873 jvms()->set_should_reexecute(true); 3874 3875 array_type_mirror = null_check(array_type_mirror); 3876 original = null_check(original); 3877 3878 // Check if a null path was taken unconditionally. 3879 if (stopped()) return true; 3880 3881 Node* orig_length = load_array_length(original); 3882 3883 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3884 klass_node = null_check(klass_node); 3885 3886 RegionNode* bailout = new RegionNode(1); 3887 record_for_igvn(bailout); 3888 3889 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3890 // Bail out if that is so. 3891 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3892 if (not_objArray != NULL) { 3893 // Improve the klass node's type from the new optimistic assumption: 3894 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3895 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3896 Node* cast = new CastPPNode(klass_node, akls); 3897 cast->init_req(0, control()); 3898 klass_node = _gvn.transform(cast); 3899 } 3900 3901 // Bail out if either start or end is negative. 3902 generate_negative_guard(start, bailout, &start); 3903 generate_negative_guard(end, bailout, &end); 3904 3905 Node* length = end; 3906 if (_gvn.type(start) != TypeInt::ZERO) { 3907 length = _gvn.transform(new SubINode(end, start)); 3908 } 3909 3910 // Bail out if length is negative. 3911 // Without this the new_array would throw 3912 // NegativeArraySizeException but IllegalArgumentException is what 3913 // should be thrown 3914 generate_negative_guard(length, bailout, &length); 3915 3916 if (bailout->req() > 1) { 3917 PreserveJVMState pjvms(this); 3918 set_control(_gvn.transform(bailout)); 3919 uncommon_trap(Deoptimization::Reason_intrinsic, 3920 Deoptimization::Action_maybe_recompile); 3921 } 3922 3923 if (!stopped()) { 3924 // How many elements will we copy from the original? 3925 // The answer is MinI(orig_length - start, length). 3926 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3927 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3928 3929 // Generate a direct call to the right arraycopy function(s). 3930 // We know the copy is disjoint but we might not know if the 3931 // oop stores need checking. 3932 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3933 // This will fail a store-check if x contains any non-nulls. 3934 3935 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3936 // loads/stores but it is legal only if we're sure the 3937 // Arrays.copyOf would succeed. So we need all input arguments 3938 // to the copyOf to be validated, including that the copy to the 3939 // new array won't trigger an ArrayStoreException. That subtype 3940 // check can be optimized if we know something on the type of 3941 // the input array from type speculation. 3942 if (_gvn.type(klass_node)->singleton()) { 3943 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3944 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3945 3946 int test = C->static_subtype_check(superk, subk); 3947 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3948 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3949 if (t_original->speculative_type() != NULL) { 3950 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3951 } 3952 } 3953 } 3954 3955 bool validated = false; 3956 // Reason_class_check rather than Reason_intrinsic because we 3957 // want to intrinsify even if this traps. 3958 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3959 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3960 klass_node); 3961 3962 if (not_subtype_ctrl != top()) { 3963 PreserveJVMState pjvms(this); 3964 set_control(not_subtype_ctrl); 3965 uncommon_trap(Deoptimization::Reason_class_check, 3966 Deoptimization::Action_make_not_entrant); 3967 assert(stopped(), "Should be stopped"); 3968 } 3969 validated = true; 3970 } 3971 3972 if (!stopped()) { 3973 newcopy = new_array(klass_node, length, 0); // no arguments to push 3974 3975 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3976 load_object_klass(original), klass_node); 3977 if (!is_copyOfRange) { 3978 ac->set_copyof(validated); 3979 } else { 3980 ac->set_copyofrange(validated); 3981 } 3982 Node* n = _gvn.transform(ac); 3983 if (n == ac) { 3984 ac->connect_outputs(this); 3985 } else { 3986 assert(validated, "shouldn't transform if all arguments not validated"); 3987 set_all_memory(n); 3988 } 3989 } 3990 } 3991 } // original reexecute is set back here 3992 3993 C->set_has_split_ifs(true); // Has chance for split-if optimization 3994 if (!stopped()) { 3995 set_result(newcopy); 3996 } 3997 return true; 3998 } 3999 4000 4001 //----------------------generate_virtual_guard--------------------------- 4002 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4003 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4004 RegionNode* slow_region) { 4005 ciMethod* method = callee(); 4006 int vtable_index = method->vtable_index(); 4007 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4008 err_msg_res("bad index %d", vtable_index)); 4009 // Get the Method* out of the appropriate vtable entry. 4010 int entry_offset = (InstanceKlass::vtable_start_offset() + 4011 vtable_index*vtableEntry::size()) * wordSize + 4012 vtableEntry::method_offset_in_bytes(); 4013 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4014 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4015 4016 // Compare the target method with the expected method (e.g., Object.hashCode). 4017 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4018 4019 Node* native_call = makecon(native_call_addr); 4020 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4021 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4022 4023 return generate_slow_guard(test_native, slow_region); 4024 } 4025 4026 //-----------------------generate_method_call---------------------------- 4027 // Use generate_method_call to make a slow-call to the real 4028 // method if the fast path fails. An alternative would be to 4029 // use a stub like OptoRuntime::slow_arraycopy_Java. 4030 // This only works for expanding the current library call, 4031 // not another intrinsic. (E.g., don't use this for making an 4032 // arraycopy call inside of the copyOf intrinsic.) 4033 CallJavaNode* 4034 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4035 // When compiling the intrinsic method itself, do not use this technique. 4036 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4037 4038 ciMethod* method = callee(); 4039 // ensure the JVMS we have will be correct for this call 4040 guarantee(method_id == method->intrinsic_id(), "must match"); 4041 4042 const TypeFunc* tf = TypeFunc::make(method); 4043 CallJavaNode* slow_call; 4044 if (is_static) { 4045 assert(!is_virtual, ""); 4046 slow_call = new CallStaticJavaNode(C, tf, 4047 SharedRuntime::get_resolve_static_call_stub(), 4048 method, bci()); 4049 } else if (is_virtual) { 4050 null_check_receiver(); 4051 int vtable_index = Method::invalid_vtable_index; 4052 if (UseInlineCaches) { 4053 // Suppress the vtable call 4054 } else { 4055 // hashCode and clone are not a miranda methods, 4056 // so the vtable index is fixed. 4057 // No need to use the linkResolver to get it. 4058 vtable_index = method->vtable_index(); 4059 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4060 err_msg_res("bad index %d", vtable_index)); 4061 } 4062 slow_call = new CallDynamicJavaNode(tf, 4063 SharedRuntime::get_resolve_virtual_call_stub(), 4064 method, vtable_index, bci()); 4065 } else { // neither virtual nor static: opt_virtual 4066 null_check_receiver(); 4067 slow_call = new CallStaticJavaNode(C, tf, 4068 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4069 method, bci()); 4070 slow_call->set_optimized_virtual(true); 4071 } 4072 set_arguments_for_java_call(slow_call); 4073 set_edges_for_java_call(slow_call); 4074 return slow_call; 4075 } 4076 4077 4078 /** 4079 * Build special case code for calls to hashCode on an object. This call may 4080 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4081 * slightly different code. 4082 */ 4083 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4084 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4085 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4086 4087 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4088 4089 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4090 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4091 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4092 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4093 Node* obj = NULL; 4094 if (!is_static) { 4095 // Check for hashing null object 4096 obj = null_check_receiver(); 4097 if (stopped()) return true; // unconditionally null 4098 result_reg->init_req(_null_path, top()); 4099 result_val->init_req(_null_path, top()); 4100 } else { 4101 // Do a null check, and return zero if null. 4102 // System.identityHashCode(null) == 0 4103 obj = argument(0); 4104 Node* null_ctl = top(); 4105 obj = null_check_oop(obj, &null_ctl); 4106 result_reg->init_req(_null_path, null_ctl); 4107 result_val->init_req(_null_path, _gvn.intcon(0)); 4108 } 4109 4110 // Unconditionally null? Then return right away. 4111 if (stopped()) { 4112 set_control( result_reg->in(_null_path)); 4113 if (!stopped()) 4114 set_result(result_val->in(_null_path)); 4115 return true; 4116 } 4117 4118 // We only go to the fast case code if we pass a number of guards. The 4119 // paths which do not pass are accumulated in the slow_region. 4120 RegionNode* slow_region = new RegionNode(1); 4121 record_for_igvn(slow_region); 4122 4123 // If this is a virtual call, we generate a funny guard. We pull out 4124 // the vtable entry corresponding to hashCode() from the target object. 4125 // If the target method which we are calling happens to be the native 4126 // Object hashCode() method, we pass the guard. We do not need this 4127 // guard for non-virtual calls -- the caller is known to be the native 4128 // Object hashCode(). 4129 if (is_virtual) { 4130 // After null check, get the object's klass. 4131 Node* obj_klass = load_object_klass(obj); 4132 generate_virtual_guard(obj_klass, slow_region); 4133 } 4134 4135 // Get the header out of the object, use LoadMarkNode when available 4136 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4137 // The control of the load must be NULL. Otherwise, the load can move before 4138 // the null check after castPP removal. 4139 Node* no_ctrl = NULL; 4140 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4141 4142 // Test the header to see if it is unlocked. 4143 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4144 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4145 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4146 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4147 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4148 4149 generate_slow_guard(test_unlocked, slow_region); 4150 4151 // Get the hash value and check to see that it has been properly assigned. 4152 // We depend on hash_mask being at most 32 bits and avoid the use of 4153 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4154 // vm: see markOop.hpp. 4155 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4156 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4157 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4158 // This hack lets the hash bits live anywhere in the mark object now, as long 4159 // as the shift drops the relevant bits into the low 32 bits. Note that 4160 // Java spec says that HashCode is an int so there's no point in capturing 4161 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4162 hshifted_header = ConvX2I(hshifted_header); 4163 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4164 4165 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4166 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4167 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4168 4169 generate_slow_guard(test_assigned, slow_region); 4170 4171 Node* init_mem = reset_memory(); 4172 // fill in the rest of the null path: 4173 result_io ->init_req(_null_path, i_o()); 4174 result_mem->init_req(_null_path, init_mem); 4175 4176 result_val->init_req(_fast_path, hash_val); 4177 result_reg->init_req(_fast_path, control()); 4178 result_io ->init_req(_fast_path, i_o()); 4179 result_mem->init_req(_fast_path, init_mem); 4180 4181 // Generate code for the slow case. We make a call to hashCode(). 4182 set_control(_gvn.transform(slow_region)); 4183 if (!stopped()) { 4184 // No need for PreserveJVMState, because we're using up the present state. 4185 set_all_memory(init_mem); 4186 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4187 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4188 Node* slow_result = set_results_for_java_call(slow_call); 4189 // this->control() comes from set_results_for_java_call 4190 result_reg->init_req(_slow_path, control()); 4191 result_val->init_req(_slow_path, slow_result); 4192 result_io ->set_req(_slow_path, i_o()); 4193 result_mem ->set_req(_slow_path, reset_memory()); 4194 } 4195 4196 // Return the combined state. 4197 set_i_o( _gvn.transform(result_io) ); 4198 set_all_memory( _gvn.transform(result_mem)); 4199 4200 set_result(result_reg, result_val); 4201 return true; 4202 } 4203 4204 //---------------------------inline_native_getClass---------------------------- 4205 // public final native Class<?> java.lang.Object.getClass(); 4206 // 4207 // Build special case code for calls to getClass on an object. 4208 bool LibraryCallKit::inline_native_getClass() { 4209 Node* obj = null_check_receiver(); 4210 if (stopped()) return true; 4211 set_result(load_mirror_from_klass(load_object_klass(obj))); 4212 return true; 4213 } 4214 4215 //-----------------inline_native_Reflection_getCallerClass--------------------- 4216 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4217 // 4218 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4219 // 4220 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4221 // in that it must skip particular security frames and checks for 4222 // caller sensitive methods. 4223 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4224 #ifndef PRODUCT 4225 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4226 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4227 } 4228 #endif 4229 4230 if (!jvms()->has_method()) { 4231 #ifndef PRODUCT 4232 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4233 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4234 } 4235 #endif 4236 return false; 4237 } 4238 4239 // Walk back up the JVM state to find the caller at the required 4240 // depth. 4241 JVMState* caller_jvms = jvms(); 4242 4243 // Cf. JVM_GetCallerClass 4244 // NOTE: Start the loop at depth 1 because the current JVM state does 4245 // not include the Reflection.getCallerClass() frame. 4246 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4247 ciMethod* m = caller_jvms->method(); 4248 switch (n) { 4249 case 0: 4250 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4251 break; 4252 case 1: 4253 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4254 if (!m->caller_sensitive()) { 4255 #ifndef PRODUCT 4256 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4257 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4258 } 4259 #endif 4260 return false; // bail-out; let JVM_GetCallerClass do the work 4261 } 4262 break; 4263 default: 4264 if (!m->is_ignored_by_security_stack_walk()) { 4265 // We have reached the desired frame; return the holder class. 4266 // Acquire method holder as java.lang.Class and push as constant. 4267 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4268 ciInstance* caller_mirror = caller_klass->java_mirror(); 4269 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4270 4271 #ifndef PRODUCT 4272 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4273 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()); 4274 tty->print_cr(" JVM state at this point:"); 4275 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4276 ciMethod* m = jvms()->of_depth(i)->method(); 4277 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4278 } 4279 } 4280 #endif 4281 return true; 4282 } 4283 break; 4284 } 4285 } 4286 4287 #ifndef PRODUCT 4288 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4289 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4290 tty->print_cr(" JVM state at this point:"); 4291 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4292 ciMethod* m = jvms()->of_depth(i)->method(); 4293 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4294 } 4295 } 4296 #endif 4297 4298 return false; // bail-out; let JVM_GetCallerClass do the work 4299 } 4300 4301 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4302 Node* arg = argument(0); 4303 Node* result; 4304 4305 switch (id) { 4306 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4307 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4308 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4309 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4310 4311 case vmIntrinsics::_doubleToLongBits: { 4312 // two paths (plus control) merge in a wood 4313 RegionNode *r = new RegionNode(3); 4314 Node *phi = new PhiNode(r, TypeLong::LONG); 4315 4316 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4317 // Build the boolean node 4318 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4319 4320 // Branch either way. 4321 // NaN case is less traveled, which makes all the difference. 4322 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4323 Node *opt_isnan = _gvn.transform(ifisnan); 4324 assert( opt_isnan->is_If(), "Expect an IfNode"); 4325 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4326 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4327 4328 set_control(iftrue); 4329 4330 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4331 Node *slow_result = longcon(nan_bits); // return NaN 4332 phi->init_req(1, _gvn.transform( slow_result )); 4333 r->init_req(1, iftrue); 4334 4335 // Else fall through 4336 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4337 set_control(iffalse); 4338 4339 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4340 r->init_req(2, iffalse); 4341 4342 // Post merge 4343 set_control(_gvn.transform(r)); 4344 record_for_igvn(r); 4345 4346 C->set_has_split_ifs(true); // Has chance for split-if optimization 4347 result = phi; 4348 assert(result->bottom_type()->isa_long(), "must be"); 4349 break; 4350 } 4351 4352 case vmIntrinsics::_floatToIntBits: { 4353 // two paths (plus control) merge in a wood 4354 RegionNode *r = new RegionNode(3); 4355 Node *phi = new PhiNode(r, TypeInt::INT); 4356 4357 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4358 // Build the boolean node 4359 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4360 4361 // Branch either way. 4362 // NaN case is less traveled, which makes all the difference. 4363 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4364 Node *opt_isnan = _gvn.transform(ifisnan); 4365 assert( opt_isnan->is_If(), "Expect an IfNode"); 4366 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4367 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4368 4369 set_control(iftrue); 4370 4371 static const jint nan_bits = 0x7fc00000; 4372 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4373 phi->init_req(1, _gvn.transform( slow_result )); 4374 r->init_req(1, iftrue); 4375 4376 // Else fall through 4377 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4378 set_control(iffalse); 4379 4380 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4381 r->init_req(2, iffalse); 4382 4383 // Post merge 4384 set_control(_gvn.transform(r)); 4385 record_for_igvn(r); 4386 4387 C->set_has_split_ifs(true); // Has chance for split-if optimization 4388 result = phi; 4389 assert(result->bottom_type()->isa_int(), "must be"); 4390 break; 4391 } 4392 4393 default: 4394 fatal_unexpected_iid(id); 4395 break; 4396 } 4397 set_result(_gvn.transform(result)); 4398 return true; 4399 } 4400 4401 #ifdef _LP64 4402 #define XTOP ,top() /*additional argument*/ 4403 #else //_LP64 4404 #define XTOP /*no additional argument*/ 4405 #endif //_LP64 4406 4407 //----------------------inline_unsafe_copyMemory------------------------- 4408 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4409 bool LibraryCallKit::inline_unsafe_copyMemory() { 4410 if (callee()->is_static()) return false; // caller must have the capability! 4411 null_check_receiver(); // null-check receiver 4412 if (stopped()) return true; 4413 4414 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4415 4416 Node* src_ptr = argument(1); // type: oop 4417 Node* src_off = ConvL2X(argument(2)); // type: long 4418 Node* dst_ptr = argument(4); // type: oop 4419 Node* dst_off = ConvL2X(argument(5)); // type: long 4420 Node* size = ConvL2X(argument(7)); // type: long 4421 4422 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4423 "fieldOffset must be byte-scaled"); 4424 4425 Node* src = make_unsafe_address(src_ptr, src_off); 4426 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4427 4428 // Conservatively insert a memory barrier on all memory slices. 4429 // Do not let writes of the copy source or destination float below the copy. 4430 insert_mem_bar(Op_MemBarCPUOrder); 4431 4432 // Call it. Note that the length argument is not scaled. 4433 make_runtime_call(RC_LEAF|RC_NO_FP, 4434 OptoRuntime::fast_arraycopy_Type(), 4435 StubRoutines::unsafe_arraycopy(), 4436 "unsafe_arraycopy", 4437 TypeRawPtr::BOTTOM, 4438 src, dst, size XTOP); 4439 4440 // Do not let reads of the copy destination float above the copy. 4441 insert_mem_bar(Op_MemBarCPUOrder); 4442 4443 return true; 4444 } 4445 4446 //------------------------clone_coping----------------------------------- 4447 // Helper function for inline_native_clone. 4448 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4449 assert(obj_size != NULL, ""); 4450 Node* raw_obj = alloc_obj->in(1); 4451 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4452 4453 AllocateNode* alloc = NULL; 4454 if (ReduceBulkZeroing) { 4455 // We will be completely responsible for initializing this object - 4456 // mark Initialize node as complete. 4457 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4458 // The object was just allocated - there should be no any stores! 4459 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4460 // Mark as complete_with_arraycopy so that on AllocateNode 4461 // expansion, we know this AllocateNode is initialized by an array 4462 // copy and a StoreStore barrier exists after the array copy. 4463 alloc->initialization()->set_complete_with_arraycopy(); 4464 } 4465 4466 // Copy the fastest available way. 4467 // TODO: generate fields copies for small objects instead. 4468 Node* src = obj; 4469 Node* dest = alloc_obj; 4470 Node* size = _gvn.transform(obj_size); 4471 4472 // Exclude the header but include array length to copy by 8 bytes words. 4473 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4474 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4475 instanceOopDesc::base_offset_in_bytes(); 4476 // base_off: 4477 // 8 - 32-bit VM 4478 // 12 - 64-bit VM, compressed klass 4479 // 16 - 64-bit VM, normal klass 4480 if (base_off % BytesPerLong != 0) { 4481 assert(UseCompressedClassPointers, ""); 4482 if (is_array) { 4483 // Exclude length to copy by 8 bytes words. 4484 base_off += sizeof(int); 4485 } else { 4486 // Include klass to copy by 8 bytes words. 4487 base_off = instanceOopDesc::klass_offset_in_bytes(); 4488 } 4489 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4490 } 4491 src = basic_plus_adr(src, base_off); 4492 dest = basic_plus_adr(dest, base_off); 4493 4494 // Compute the length also, if needed: 4495 Node* countx = size; 4496 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4497 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4498 4499 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4500 4501 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4502 ac->set_clonebasic(); 4503 Node* n = _gvn.transform(ac); 4504 if (n == ac) { 4505 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4506 } else { 4507 set_all_memory(n); 4508 } 4509 4510 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4511 if (card_mark) { 4512 assert(!is_array, ""); 4513 // Put in store barrier for any and all oops we are sticking 4514 // into this object. (We could avoid this if we could prove 4515 // that the object type contains no oop fields at all.) 4516 Node* no_particular_value = NULL; 4517 Node* no_particular_field = NULL; 4518 int raw_adr_idx = Compile::AliasIdxRaw; 4519 post_barrier(control(), 4520 memory(raw_adr_type), 4521 alloc_obj, 4522 no_particular_field, 4523 raw_adr_idx, 4524 no_particular_value, 4525 T_OBJECT, 4526 false); 4527 } 4528 4529 // Do not let reads from the cloned object float above the arraycopy. 4530 if (alloc != NULL) { 4531 // Do not let stores that initialize this object be reordered with 4532 // a subsequent store that would make this object accessible by 4533 // other threads. 4534 // Record what AllocateNode this StoreStore protects so that 4535 // escape analysis can go from the MemBarStoreStoreNode to the 4536 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4537 // based on the escape status of the AllocateNode. 4538 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4539 } else { 4540 insert_mem_bar(Op_MemBarCPUOrder); 4541 } 4542 } 4543 4544 //------------------------inline_native_clone---------------------------- 4545 // protected native Object java.lang.Object.clone(); 4546 // 4547 // Here are the simple edge cases: 4548 // null receiver => normal trap 4549 // virtual and clone was overridden => slow path to out-of-line clone 4550 // not cloneable or finalizer => slow path to out-of-line Object.clone 4551 // 4552 // The general case has two steps, allocation and copying. 4553 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4554 // 4555 // Copying also has two cases, oop arrays and everything else. 4556 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4557 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4558 // 4559 // These steps fold up nicely if and when the cloned object's klass 4560 // can be sharply typed as an object array, a type array, or an instance. 4561 // 4562 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4563 PhiNode* result_val; 4564 4565 // Set the reexecute bit for the interpreter to reexecute 4566 // the bytecode that invokes Object.clone if deoptimization happens. 4567 { PreserveReexecuteState preexecs(this); 4568 jvms()->set_should_reexecute(true); 4569 4570 Node* obj = null_check_receiver(); 4571 if (stopped()) return true; 4572 4573 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4574 4575 // If we are going to clone an instance, we need its exact type to 4576 // know the number and types of fields to convert the clone to 4577 // loads/stores. Maybe a speculative type can help us. 4578 if (!obj_type->klass_is_exact() && 4579 obj_type->speculative_type() != NULL && 4580 obj_type->speculative_type()->is_instance_klass()) { 4581 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4582 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4583 !spec_ik->has_injected_fields()) { 4584 ciKlass* k = obj_type->klass(); 4585 if (!k->is_instance_klass() || 4586 k->as_instance_klass()->is_interface() || 4587 k->as_instance_klass()->has_subklass()) { 4588 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4589 } 4590 } 4591 } 4592 4593 Node* obj_klass = load_object_klass(obj); 4594 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4595 const TypeOopPtr* toop = ((tklass != NULL) 4596 ? tklass->as_instance_type() 4597 : TypeInstPtr::NOTNULL); 4598 4599 // Conservatively insert a memory barrier on all memory slices. 4600 // Do not let writes into the original float below the clone. 4601 insert_mem_bar(Op_MemBarCPUOrder); 4602 4603 // paths into result_reg: 4604 enum { 4605 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4606 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4607 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4608 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4609 PATH_LIMIT 4610 }; 4611 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4612 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4613 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4614 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4615 record_for_igvn(result_reg); 4616 4617 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4618 int raw_adr_idx = Compile::AliasIdxRaw; 4619 4620 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4621 if (array_ctl != NULL) { 4622 // It's an array. 4623 PreserveJVMState pjvms(this); 4624 set_control(array_ctl); 4625 Node* obj_length = load_array_length(obj); 4626 Node* obj_size = NULL; 4627 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4628 4629 if (!use_ReduceInitialCardMarks()) { 4630 // If it is an oop array, it requires very special treatment, 4631 // because card marking is required on each card of the array. 4632 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4633 if (is_obja != NULL) { 4634 PreserveJVMState pjvms2(this); 4635 set_control(is_obja); 4636 // Generate a direct call to the right arraycopy function(s). 4637 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4638 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4639 ac->set_cloneoop(); 4640 Node* n = _gvn.transform(ac); 4641 assert(n == ac, "cannot disappear"); 4642 ac->connect_outputs(this); 4643 4644 result_reg->init_req(_objArray_path, control()); 4645 result_val->init_req(_objArray_path, alloc_obj); 4646 result_i_o ->set_req(_objArray_path, i_o()); 4647 result_mem ->set_req(_objArray_path, reset_memory()); 4648 } 4649 } 4650 // Otherwise, there are no card marks to worry about. 4651 // (We can dispense with card marks if we know the allocation 4652 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4653 // causes the non-eden paths to take compensating steps to 4654 // simulate a fresh allocation, so that no further 4655 // card marks are required in compiled code to initialize 4656 // the object.) 4657 4658 if (!stopped()) { 4659 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4660 4661 // Present the results of the copy. 4662 result_reg->init_req(_array_path, control()); 4663 result_val->init_req(_array_path, alloc_obj); 4664 result_i_o ->set_req(_array_path, i_o()); 4665 result_mem ->set_req(_array_path, reset_memory()); 4666 } 4667 } 4668 4669 // We only go to the instance fast case code if we pass a number of guards. 4670 // The paths which do not pass are accumulated in the slow_region. 4671 RegionNode* slow_region = new RegionNode(1); 4672 record_for_igvn(slow_region); 4673 if (!stopped()) { 4674 // It's an instance (we did array above). Make the slow-path tests. 4675 // If this is a virtual call, we generate a funny guard. We grab 4676 // the vtable entry corresponding to clone() from the target object. 4677 // If the target method which we are calling happens to be the 4678 // Object clone() method, we pass the guard. We do not need this 4679 // guard for non-virtual calls; the caller is known to be the native 4680 // Object clone(). 4681 if (is_virtual) { 4682 generate_virtual_guard(obj_klass, slow_region); 4683 } 4684 4685 // The object must be cloneable and must not have a finalizer. 4686 // Both of these conditions may be checked in a single test. 4687 // We could optimize the cloneable test further, but we don't care. 4688 generate_access_flags_guard(obj_klass, 4689 // Test both conditions: 4690 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4691 // Must be cloneable but not finalizer: 4692 JVM_ACC_IS_CLONEABLE, 4693 slow_region); 4694 } 4695 4696 if (!stopped()) { 4697 // It's an instance, and it passed the slow-path tests. 4698 PreserveJVMState pjvms(this); 4699 Node* obj_size = NULL; 4700 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4701 // is reexecuted if deoptimization occurs and there could be problems when merging 4702 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4703 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4704 4705 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4706 4707 // Present the results of the slow call. 4708 result_reg->init_req(_instance_path, control()); 4709 result_val->init_req(_instance_path, alloc_obj); 4710 result_i_o ->set_req(_instance_path, i_o()); 4711 result_mem ->set_req(_instance_path, reset_memory()); 4712 } 4713 4714 // Generate code for the slow case. We make a call to clone(). 4715 set_control(_gvn.transform(slow_region)); 4716 if (!stopped()) { 4717 PreserveJVMState pjvms(this); 4718 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4719 Node* slow_result = set_results_for_java_call(slow_call); 4720 // this->control() comes from set_results_for_java_call 4721 result_reg->init_req(_slow_path, control()); 4722 result_val->init_req(_slow_path, slow_result); 4723 result_i_o ->set_req(_slow_path, i_o()); 4724 result_mem ->set_req(_slow_path, reset_memory()); 4725 } 4726 4727 // Return the combined state. 4728 set_control( _gvn.transform(result_reg)); 4729 set_i_o( _gvn.transform(result_i_o)); 4730 set_all_memory( _gvn.transform(result_mem)); 4731 } // original reexecute is set back here 4732 4733 set_result(_gvn.transform(result_val)); 4734 return true; 4735 } 4736 4737 // If we have a tighly coupled allocation, the arraycopy may take care 4738 // of the array initialization. If one of the guards we insert between 4739 // the allocation and the arraycopy causes a deoptimization, an 4740 // unitialized array will escape the compiled method. To prevent that 4741 // we set the JVM state for uncommon traps between the allocation and 4742 // the arraycopy to the state before the allocation so, in case of 4743 // deoptimization, we'll reexecute the allocation and the 4744 // initialization. 4745 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4746 if (alloc != NULL) { 4747 ciMethod* trap_method = alloc->jvms()->method(); 4748 int trap_bci = alloc->jvms()->bci(); 4749 4750 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4751 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4752 // Make sure there's no store between the allocation and the 4753 // arraycopy otherwise visible side effects could be rexecuted 4754 // in case of deoptimization and cause incorrect execution. 4755 bool no_interfering_store = true; 4756 Node* mem = alloc->in(TypeFunc::Memory); 4757 if (mem->is_MergeMem()) { 4758 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4759 Node* n = mms.memory(); 4760 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4761 assert(n->is_Store(), "what else?"); 4762 no_interfering_store = false; 4763 break; 4764 } 4765 } 4766 } else { 4767 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4768 Node* n = mms.memory(); 4769 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4770 assert(n->is_Store(), "what else?"); 4771 no_interfering_store = false; 4772 break; 4773 } 4774 } 4775 } 4776 4777 if (no_interfering_store) { 4778 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4779 uint size = alloc->req(); 4780 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4781 old_jvms->set_map(sfpt); 4782 for (uint i = 0; i < size; i++) { 4783 sfpt->init_req(i, alloc->in(i)); 4784 } 4785 // re-push array length for deoptimization 4786 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4787 old_jvms->set_sp(old_jvms->sp()+1); 4788 old_jvms->set_monoff(old_jvms->monoff()+1); 4789 old_jvms->set_scloff(old_jvms->scloff()+1); 4790 old_jvms->set_endoff(old_jvms->endoff()+1); 4791 old_jvms->set_should_reexecute(true); 4792 4793 sfpt->set_i_o(map()->i_o()); 4794 sfpt->set_memory(map()->memory()); 4795 sfpt->set_control(map()->control()); 4796 4797 JVMState* saved_jvms = jvms(); 4798 saved_reexecute_sp = _reexecute_sp; 4799 4800 set_jvms(sfpt->jvms()); 4801 _reexecute_sp = jvms()->sp(); 4802 4803 return saved_jvms; 4804 } 4805 } 4806 } 4807 return NULL; 4808 } 4809 4810 // In case of a deoptimization, we restart execution at the 4811 // allocation, allocating a new array. We would leave an uninitialized 4812 // array in the heap that GCs wouldn't expect. Move the allocation 4813 // after the traps so we don't allocate the array if we 4814 // deoptimize. This is possible because tightly_coupled_allocation() 4815 // guarantees there's no observer of the allocated array at this point 4816 // and the control flow is simple enough. 4817 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4818 if (saved_jvms != NULL && !stopped()) { 4819 assert(alloc != NULL, "only with a tightly coupled allocation"); 4820 // restore JVM state to the state at the arraycopy 4821 saved_jvms->map()->set_control(map()->control()); 4822 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4823 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4824 // If we've improved the types of some nodes (null check) while 4825 // emitting the guards, propagate them to the current state 4826 map()->replaced_nodes().apply(saved_jvms->map()); 4827 set_jvms(saved_jvms); 4828 _reexecute_sp = saved_reexecute_sp; 4829 4830 // Remove the allocation from above the guards 4831 CallProjections callprojs; 4832 alloc->extract_projections(&callprojs, true); 4833 InitializeNode* init = alloc->initialization(); 4834 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4835 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4836 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4837 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4838 4839 // move the allocation here (after the guards) 4840 _gvn.hash_delete(alloc); 4841 alloc->set_req(TypeFunc::Control, control()); 4842 alloc->set_req(TypeFunc::I_O, i_o()); 4843 Node *mem = reset_memory(); 4844 set_all_memory(mem); 4845 alloc->set_req(TypeFunc::Memory, mem); 4846 set_control(init->proj_out(TypeFunc::Control)); 4847 set_i_o(callprojs.fallthrough_ioproj); 4848 4849 // Update memory as done in GraphKit::set_output_for_allocation() 4850 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4851 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4852 if (ary_type->isa_aryptr() && length_type != NULL) { 4853 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4854 } 4855 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4856 int elemidx = C->get_alias_index(telemref); 4857 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4858 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4859 4860 Node* allocx = _gvn.transform(alloc); 4861 assert(allocx == alloc, "where has the allocation gone?"); 4862 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4863 4864 _gvn.hash_delete(dest); 4865 dest->set_req(0, control()); 4866 Node* destx = _gvn.transform(dest); 4867 assert(destx == dest, "where has the allocation result gone?"); 4868 } 4869 } 4870 4871 4872 //------------------------------inline_arraycopy----------------------- 4873 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4874 // Object dest, int destPos, 4875 // int length); 4876 bool LibraryCallKit::inline_arraycopy() { 4877 // Get the arguments. 4878 Node* src = argument(0); // type: oop 4879 Node* src_offset = argument(1); // type: int 4880 Node* dest = argument(2); // type: oop 4881 Node* dest_offset = argument(3); // type: int 4882 Node* length = argument(4); // type: int 4883 4884 4885 // Check for allocation before we add nodes that would confuse 4886 // tightly_coupled_allocation() 4887 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4888 4889 int saved_reexecute_sp = -1; 4890 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4891 // See arraycopy_restore_alloc_state() comment 4892 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4893 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4894 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4895 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4896 4897 // The following tests must be performed 4898 // (1) src and dest are arrays. 4899 // (2) src and dest arrays must have elements of the same BasicType 4900 // (3) src and dest must not be null. 4901 // (4) src_offset must not be negative. 4902 // (5) dest_offset must not be negative. 4903 // (6) length must not be negative. 4904 // (7) src_offset + length must not exceed length of src. 4905 // (8) dest_offset + length must not exceed length of dest. 4906 // (9) each element of an oop array must be assignable 4907 4908 // (3) src and dest must not be null. 4909 // always do this here because we need the JVM state for uncommon traps 4910 Node* null_ctl = top(); 4911 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4912 assert(null_ctl->is_top(), "no null control here"); 4913 dest = null_check(dest, T_ARRAY); 4914 4915 if (!can_emit_guards) { 4916 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4917 // guards but the arraycopy node could still take advantage of a 4918 // tightly allocated allocation. tightly_coupled_allocation() is 4919 // called again to make sure it takes the null check above into 4920 // account: the null check is mandatory and if it caused an 4921 // uncommon trap to be emitted then the allocation can't be 4922 // considered tightly coupled in this context. 4923 alloc = tightly_coupled_allocation(dest, NULL); 4924 } 4925 4926 bool validated = false; 4927 4928 const Type* src_type = _gvn.type(src); 4929 const Type* dest_type = _gvn.type(dest); 4930 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4931 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4932 4933 // Do we have the type of src? 4934 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4935 // Do we have the type of dest? 4936 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4937 // Is the type for src from speculation? 4938 bool src_spec = false; 4939 // Is the type for dest from speculation? 4940 bool dest_spec = false; 4941 4942 if ((!has_src || !has_dest) && can_emit_guards) { 4943 // We don't have sufficient type information, let's see if 4944 // speculative types can help. We need to have types for both src 4945 // and dest so that it pays off. 4946 4947 // Do we already have or could we have type information for src 4948 bool could_have_src = has_src; 4949 // Do we already have or could we have type information for dest 4950 bool could_have_dest = has_dest; 4951 4952 ciKlass* src_k = NULL; 4953 if (!has_src) { 4954 src_k = src_type->speculative_type_not_null(); 4955 if (src_k != NULL && src_k->is_array_klass()) { 4956 could_have_src = true; 4957 } 4958 } 4959 4960 ciKlass* dest_k = NULL; 4961 if (!has_dest) { 4962 dest_k = dest_type->speculative_type_not_null(); 4963 if (dest_k != NULL && dest_k->is_array_klass()) { 4964 could_have_dest = true; 4965 } 4966 } 4967 4968 if (could_have_src && could_have_dest) { 4969 // This is going to pay off so emit the required guards 4970 if (!has_src) { 4971 src = maybe_cast_profiled_obj(src, src_k, true); 4972 src_type = _gvn.type(src); 4973 top_src = src_type->isa_aryptr(); 4974 has_src = (top_src != NULL && top_src->klass() != NULL); 4975 src_spec = true; 4976 } 4977 if (!has_dest) { 4978 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4979 dest_type = _gvn.type(dest); 4980 top_dest = dest_type->isa_aryptr(); 4981 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4982 dest_spec = true; 4983 } 4984 } 4985 } 4986 4987 if (has_src && has_dest && can_emit_guards) { 4988 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4989 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4990 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4991 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4992 4993 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4994 // If both arrays are object arrays then having the exact types 4995 // for both will remove the need for a subtype check at runtime 4996 // before the call and may make it possible to pick a faster copy 4997 // routine (without a subtype check on every element) 4998 // Do we have the exact type of src? 4999 bool could_have_src = src_spec; 5000 // Do we have the exact type of dest? 5001 bool could_have_dest = dest_spec; 5002 ciKlass* src_k = top_src->klass(); 5003 ciKlass* dest_k = top_dest->klass(); 5004 if (!src_spec) { 5005 src_k = src_type->speculative_type_not_null(); 5006 if (src_k != NULL && src_k->is_array_klass()) { 5007 could_have_src = true; 5008 } 5009 } 5010 if (!dest_spec) { 5011 dest_k = dest_type->speculative_type_not_null(); 5012 if (dest_k != NULL && dest_k->is_array_klass()) { 5013 could_have_dest = true; 5014 } 5015 } 5016 if (could_have_src && could_have_dest) { 5017 // If we can have both exact types, emit the missing guards 5018 if (could_have_src && !src_spec) { 5019 src = maybe_cast_profiled_obj(src, src_k, true); 5020 } 5021 if (could_have_dest && !dest_spec) { 5022 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5023 } 5024 } 5025 } 5026 } 5027 5028 ciMethod* trap_method = method(); 5029 int trap_bci = bci(); 5030 if (saved_jvms != NULL) { 5031 trap_method = alloc->jvms()->method(); 5032 trap_bci = alloc->jvms()->bci(); 5033 } 5034 5035 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5036 can_emit_guards && 5037 !src->is_top() && !dest->is_top()) { 5038 // validate arguments: enables transformation the ArrayCopyNode 5039 validated = true; 5040 5041 RegionNode* slow_region = new RegionNode(1); 5042 record_for_igvn(slow_region); 5043 5044 // (1) src and dest are arrays. 5045 generate_non_array_guard(load_object_klass(src), slow_region); 5046 generate_non_array_guard(load_object_klass(dest), slow_region); 5047 5048 // (2) src and dest arrays must have elements of the same BasicType 5049 // done at macro expansion or at Ideal transformation time 5050 5051 // (4) src_offset must not be negative. 5052 generate_negative_guard(src_offset, slow_region); 5053 5054 // (5) dest_offset must not be negative. 5055 generate_negative_guard(dest_offset, slow_region); 5056 5057 // (7) src_offset + length must not exceed length of src. 5058 generate_limit_guard(src_offset, length, 5059 load_array_length(src), 5060 slow_region); 5061 5062 // (8) dest_offset + length must not exceed length of dest. 5063 generate_limit_guard(dest_offset, length, 5064 load_array_length(dest), 5065 slow_region); 5066 5067 // (9) each element of an oop array must be assignable 5068 Node* src_klass = load_object_klass(src); 5069 Node* dest_klass = load_object_klass(dest); 5070 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5071 5072 if (not_subtype_ctrl != top()) { 5073 PreserveJVMState pjvms(this); 5074 set_control(not_subtype_ctrl); 5075 uncommon_trap(Deoptimization::Reason_intrinsic, 5076 Deoptimization::Action_make_not_entrant); 5077 assert(stopped(), "Should be stopped"); 5078 } 5079 { 5080 PreserveJVMState pjvms(this); 5081 set_control(_gvn.transform(slow_region)); 5082 uncommon_trap(Deoptimization::Reason_intrinsic, 5083 Deoptimization::Action_make_not_entrant); 5084 assert(stopped(), "Should be stopped"); 5085 } 5086 } 5087 5088 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5089 5090 if (stopped()) { 5091 return true; 5092 } 5093 5094 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5095 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5096 // so the compiler has a chance to eliminate them: during macro expansion, 5097 // we have to set their control (CastPP nodes are eliminated). 5098 load_object_klass(src), load_object_klass(dest), 5099 load_array_length(src), load_array_length(dest)); 5100 5101 ac->set_arraycopy(validated); 5102 5103 Node* n = _gvn.transform(ac); 5104 if (n == ac) { 5105 ac->connect_outputs(this); 5106 } else { 5107 assert(validated, "shouldn't transform if all arguments not validated"); 5108 set_all_memory(n); 5109 } 5110 5111 return true; 5112 } 5113 5114 5115 // Helper function which determines if an arraycopy immediately follows 5116 // an allocation, with no intervening tests or other escapes for the object. 5117 AllocateArrayNode* 5118 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5119 RegionNode* slow_region) { 5120 if (stopped()) return NULL; // no fast path 5121 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5122 5123 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5124 if (alloc == NULL) return NULL; 5125 5126 Node* rawmem = memory(Compile::AliasIdxRaw); 5127 // Is the allocation's memory state untouched? 5128 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5129 // Bail out if there have been raw-memory effects since the allocation. 5130 // (Example: There might have been a call or safepoint.) 5131 return NULL; 5132 } 5133 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5134 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5135 return NULL; 5136 } 5137 5138 // There must be no unexpected observers of this allocation. 5139 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5140 Node* obs = ptr->fast_out(i); 5141 if (obs != this->map()) { 5142 return NULL; 5143 } 5144 } 5145 5146 // This arraycopy must unconditionally follow the allocation of the ptr. 5147 Node* alloc_ctl = ptr->in(0); 5148 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5149 5150 Node* ctl = control(); 5151 while (ctl != alloc_ctl) { 5152 // There may be guards which feed into the slow_region. 5153 // Any other control flow means that we might not get a chance 5154 // to finish initializing the allocated object. 5155 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5156 IfNode* iff = ctl->in(0)->as_If(); 5157 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5158 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5159 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5160 ctl = iff->in(0); // This test feeds the known slow_region. 5161 continue; 5162 } 5163 // One more try: Various low-level checks bottom out in 5164 // uncommon traps. If the debug-info of the trap omits 5165 // any reference to the allocation, as we've already 5166 // observed, then there can be no objection to the trap. 5167 bool found_trap = false; 5168 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5169 Node* obs = not_ctl->fast_out(j); 5170 if (obs->in(0) == not_ctl && obs->is_Call() && 5171 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5172 found_trap = true; break; 5173 } 5174 } 5175 if (found_trap) { 5176 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5177 continue; 5178 } 5179 } 5180 return NULL; 5181 } 5182 5183 // If we get this far, we have an allocation which immediately 5184 // precedes the arraycopy, and we can take over zeroing the new object. 5185 // The arraycopy will finish the initialization, and provide 5186 // a new control state to which we will anchor the destination pointer. 5187 5188 return alloc; 5189 } 5190 5191 //-------------inline_encodeISOArray----------------------------------- 5192 // encode char[] to byte[] in ISO_8859_1 5193 bool LibraryCallKit::inline_encodeISOArray() { 5194 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5195 // no receiver since it is static method 5196 Node *src = argument(0); 5197 Node *src_offset = argument(1); 5198 Node *dst = argument(2); 5199 Node *dst_offset = argument(3); 5200 Node *length = argument(4); 5201 5202 const Type* src_type = src->Value(&_gvn); 5203 const Type* dst_type = dst->Value(&_gvn); 5204 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5205 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5206 if (top_src == NULL || top_src->klass() == NULL || 5207 top_dest == NULL || top_dest->klass() == NULL) { 5208 // failed array check 5209 return false; 5210 } 5211 5212 // Figure out the size and type of the elements we will be copying. 5213 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5214 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5215 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5216 return false; 5217 } 5218 Node* src_start = array_element_address(src, src_offset, src_elem); 5219 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5220 // 'src_start' points to src array + scaled offset 5221 // 'dst_start' points to dst array + scaled offset 5222 5223 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5224 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5225 enc = _gvn.transform(enc); 5226 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5227 set_memory(res_mem, mtype); 5228 set_result(enc); 5229 return true; 5230 } 5231 5232 //-------------inline_multiplyToLen----------------------------------- 5233 bool LibraryCallKit::inline_multiplyToLen() { 5234 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5235 5236 address stubAddr = StubRoutines::multiplyToLen(); 5237 if (stubAddr == NULL) { 5238 return false; // Intrinsic's stub is not implemented on this platform 5239 } 5240 const char* stubName = "multiplyToLen"; 5241 5242 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5243 5244 Node* x = argument(1); 5245 Node* xlen = argument(2); 5246 Node* y = argument(3); 5247 Node* ylen = argument(4); 5248 Node* z = argument(5); 5249 5250 const Type* x_type = x->Value(&_gvn); 5251 const Type* y_type = y->Value(&_gvn); 5252 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5253 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5254 if (top_x == NULL || top_x->klass() == NULL || 5255 top_y == NULL || top_y->klass() == NULL) { 5256 // failed array check 5257 return false; 5258 } 5259 5260 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5261 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5262 if (x_elem != T_INT || y_elem != T_INT) { 5263 return false; 5264 } 5265 5266 // Set the original stack and the reexecute bit for the interpreter to reexecute 5267 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5268 // on the return from z array allocation in runtime. 5269 { PreserveReexecuteState preexecs(this); 5270 jvms()->set_should_reexecute(true); 5271 5272 Node* x_start = array_element_address(x, intcon(0), x_elem); 5273 Node* y_start = array_element_address(y, intcon(0), y_elem); 5274 // 'x_start' points to x array + scaled xlen 5275 // 'y_start' points to y array + scaled ylen 5276 5277 // Allocate the result array 5278 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5279 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5280 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5281 5282 IdealKit ideal(this); 5283 5284 #define __ ideal. 5285 Node* one = __ ConI(1); 5286 Node* zero = __ ConI(0); 5287 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5288 __ set(need_alloc, zero); 5289 __ set(z_alloc, z); 5290 __ if_then(z, BoolTest::eq, null()); { 5291 __ increment (need_alloc, one); 5292 } __ else_(); { 5293 // Update graphKit memory and control from IdealKit. 5294 sync_kit(ideal); 5295 Node* zlen_arg = load_array_length(z); 5296 // Update IdealKit memory and control from graphKit. 5297 __ sync_kit(this); 5298 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5299 __ increment (need_alloc, one); 5300 } __ end_if(); 5301 } __ end_if(); 5302 5303 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5304 // Update graphKit memory and control from IdealKit. 5305 sync_kit(ideal); 5306 Node * narr = new_array(klass_node, zlen, 1); 5307 // Update IdealKit memory and control from graphKit. 5308 __ sync_kit(this); 5309 __ set(z_alloc, narr); 5310 } __ end_if(); 5311 5312 sync_kit(ideal); 5313 z = __ value(z_alloc); 5314 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5315 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5316 // Final sync IdealKit and GraphKit. 5317 final_sync(ideal); 5318 #undef __ 5319 5320 Node* z_start = array_element_address(z, intcon(0), T_INT); 5321 5322 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5323 OptoRuntime::multiplyToLen_Type(), 5324 stubAddr, stubName, TypePtr::BOTTOM, 5325 x_start, xlen, y_start, ylen, z_start, zlen); 5326 } // original reexecute is set back here 5327 5328 C->set_has_split_ifs(true); // Has chance for split-if optimization 5329 set_result(z); 5330 return true; 5331 } 5332 5333 //-------------inline_squareToLen------------------------------------ 5334 bool LibraryCallKit::inline_squareToLen() { 5335 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5336 5337 address stubAddr = StubRoutines::squareToLen(); 5338 if (stubAddr == NULL) { 5339 return false; // Intrinsic's stub is not implemented on this platform 5340 } 5341 const char* stubName = "squareToLen"; 5342 5343 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5344 5345 Node* x = argument(0); 5346 Node* len = argument(1); 5347 Node* z = argument(2); 5348 Node* zlen = argument(3); 5349 5350 const Type* x_type = x->Value(&_gvn); 5351 const Type* z_type = z->Value(&_gvn); 5352 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5353 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5354 if (top_x == NULL || top_x->klass() == NULL || 5355 top_z == NULL || top_z->klass() == NULL) { 5356 // failed array check 5357 return false; 5358 } 5359 5360 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5361 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5362 if (x_elem != T_INT || z_elem != T_INT) { 5363 return false; 5364 } 5365 5366 5367 Node* x_start = array_element_address(x, intcon(0), x_elem); 5368 Node* z_start = array_element_address(z, intcon(0), z_elem); 5369 5370 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5371 OptoRuntime::squareToLen_Type(), 5372 stubAddr, stubName, TypePtr::BOTTOM, 5373 x_start, len, z_start, zlen); 5374 5375 set_result(z); 5376 return true; 5377 } 5378 5379 //-------------inline_mulAdd------------------------------------------ 5380 bool LibraryCallKit::inline_mulAdd() { 5381 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5382 5383 address stubAddr = StubRoutines::mulAdd(); 5384 if (stubAddr == NULL) { 5385 return false; // Intrinsic's stub is not implemented on this platform 5386 } 5387 const char* stubName = "mulAdd"; 5388 5389 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5390 5391 Node* out = argument(0); 5392 Node* in = argument(1); 5393 Node* offset = argument(2); 5394 Node* len = argument(3); 5395 Node* k = argument(4); 5396 5397 const Type* out_type = out->Value(&_gvn); 5398 const Type* in_type = in->Value(&_gvn); 5399 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5400 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5401 if (top_out == NULL || top_out->klass() == NULL || 5402 top_in == NULL || top_in->klass() == NULL) { 5403 // failed array check 5404 return false; 5405 } 5406 5407 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5408 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5409 if (out_elem != T_INT || in_elem != T_INT) { 5410 return false; 5411 } 5412 5413 Node* outlen = load_array_length(out); 5414 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5415 Node* out_start = array_element_address(out, intcon(0), out_elem); 5416 Node* in_start = array_element_address(in, intcon(0), in_elem); 5417 5418 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5419 OptoRuntime::mulAdd_Type(), 5420 stubAddr, stubName, TypePtr::BOTTOM, 5421 out_start,in_start, new_offset, len, k); 5422 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5423 set_result(result); 5424 return true; 5425 } 5426 5427 5428 /** 5429 * Calculate CRC32 for byte. 5430 * int java.util.zip.CRC32.update(int crc, int b) 5431 */ 5432 bool LibraryCallKit::inline_updateCRC32() { 5433 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5434 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5435 // no receiver since it is static method 5436 Node* crc = argument(0); // type: int 5437 Node* b = argument(1); // type: int 5438 5439 /* 5440 * int c = ~ crc; 5441 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5442 * b = b ^ (c >>> 8); 5443 * crc = ~b; 5444 */ 5445 5446 Node* M1 = intcon(-1); 5447 crc = _gvn.transform(new XorINode(crc, M1)); 5448 Node* result = _gvn.transform(new XorINode(crc, b)); 5449 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5450 5451 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5452 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5453 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5454 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5455 5456 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5457 result = _gvn.transform(new XorINode(crc, result)); 5458 result = _gvn.transform(new XorINode(result, M1)); 5459 set_result(result); 5460 return true; 5461 } 5462 5463 /** 5464 * Calculate CRC32 for byte[] array. 5465 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5466 */ 5467 bool LibraryCallKit::inline_updateBytesCRC32() { 5468 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5469 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5470 // no receiver since it is static method 5471 Node* crc = argument(0); // type: int 5472 Node* src = argument(1); // type: oop 5473 Node* offset = argument(2); // type: int 5474 Node* length = argument(3); // type: int 5475 5476 const Type* src_type = src->Value(&_gvn); 5477 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5478 if (top_src == NULL || top_src->klass() == NULL) { 5479 // failed array check 5480 return false; 5481 } 5482 5483 // Figure out the size and type of the elements we will be copying. 5484 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5485 if (src_elem != T_BYTE) { 5486 return false; 5487 } 5488 5489 // 'src_start' points to src array + scaled offset 5490 Node* src_start = array_element_address(src, offset, src_elem); 5491 5492 // We assume that range check is done by caller. 5493 // TODO: generate range check (offset+length < src.length) in debug VM. 5494 5495 // Call the stub. 5496 address stubAddr = StubRoutines::updateBytesCRC32(); 5497 const char *stubName = "updateBytesCRC32"; 5498 5499 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5500 stubAddr, stubName, TypePtr::BOTTOM, 5501 crc, src_start, length); 5502 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5503 set_result(result); 5504 return true; 5505 } 5506 5507 /** 5508 * Calculate CRC32 for ByteBuffer. 5509 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5510 */ 5511 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5512 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5513 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5514 // no receiver since it is static method 5515 Node* crc = argument(0); // type: int 5516 Node* src = argument(1); // type: long 5517 Node* offset = argument(3); // type: int 5518 Node* length = argument(4); // type: int 5519 5520 src = ConvL2X(src); // adjust Java long to machine word 5521 Node* base = _gvn.transform(new CastX2PNode(src)); 5522 offset = ConvI2X(offset); 5523 5524 // 'src_start' points to src array + scaled offset 5525 Node* src_start = basic_plus_adr(top(), base, offset); 5526 5527 // Call the stub. 5528 address stubAddr = StubRoutines::updateBytesCRC32(); 5529 const char *stubName = "updateBytesCRC32"; 5530 5531 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5532 stubAddr, stubName, TypePtr::BOTTOM, 5533 crc, src_start, length); 5534 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5535 set_result(result); 5536 return true; 5537 } 5538 5539 //----------------------------inline_reference_get---------------------------- 5540 // public T java.lang.ref.Reference.get(); 5541 bool LibraryCallKit::inline_reference_get() { 5542 const int referent_offset = java_lang_ref_Reference::referent_offset; 5543 guarantee(referent_offset > 0, "should have already been set"); 5544 5545 // Get the argument: 5546 Node* reference_obj = null_check_receiver(); 5547 if (stopped()) return true; 5548 5549 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5550 5551 ciInstanceKlass* klass = env()->Object_klass(); 5552 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5553 5554 Node* no_ctrl = NULL; 5555 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5556 5557 // Use the pre-barrier to record the value in the referent field 5558 pre_barrier(false /* do_load */, 5559 control(), 5560 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5561 result /* pre_val */, 5562 T_OBJECT); 5563 5564 // Add memory barrier to prevent commoning reads from this field 5565 // across safepoint since GC can change its value. 5566 insert_mem_bar(Op_MemBarCPUOrder); 5567 5568 set_result(result); 5569 return true; 5570 } 5571 5572 5573 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5574 bool is_exact=true, bool is_static=false) { 5575 5576 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5577 assert(tinst != NULL, "obj is null"); 5578 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5579 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5580 5581 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5582 ciSymbol::make(fieldTypeString), 5583 is_static); 5584 if (field == NULL) return (Node *) NULL; 5585 assert (field != NULL, "undefined field"); 5586 5587 // Next code copied from Parse::do_get_xxx(): 5588 5589 // Compute address and memory type. 5590 int offset = field->offset_in_bytes(); 5591 bool is_vol = field->is_volatile(); 5592 ciType* field_klass = field->type(); 5593 assert(field_klass->is_loaded(), "should be loaded"); 5594 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5595 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5596 BasicType bt = field->layout_type(); 5597 5598 // Build the resultant type of the load 5599 const Type *type; 5600 if (bt == T_OBJECT) { 5601 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5602 } else { 5603 type = Type::get_const_basic_type(bt); 5604 } 5605 5606 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5607 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5608 } 5609 // Build the load. 5610 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5611 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5612 // If reference is volatile, prevent following memory ops from 5613 // floating up past the volatile read. Also prevents commoning 5614 // another volatile read. 5615 if (is_vol) { 5616 // Memory barrier includes bogus read of value to force load BEFORE membar 5617 insert_mem_bar(Op_MemBarAcquire, loadedField); 5618 } 5619 return loadedField; 5620 } 5621 5622 5623 //------------------------------inline_aescrypt_Block----------------------- 5624 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5625 address stubAddr; 5626 const char *stubName; 5627 assert(UseAES, "need AES instruction support"); 5628 5629 switch(id) { 5630 case vmIntrinsics::_aescrypt_encryptBlock: 5631 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5632 stubName = "aescrypt_encryptBlock"; 5633 break; 5634 case vmIntrinsics::_aescrypt_decryptBlock: 5635 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5636 stubName = "aescrypt_decryptBlock"; 5637 break; 5638 } 5639 if (stubAddr == NULL) return false; 5640 5641 Node* aescrypt_object = argument(0); 5642 Node* src = argument(1); 5643 Node* src_offset = argument(2); 5644 Node* dest = argument(3); 5645 Node* dest_offset = argument(4); 5646 5647 // (1) src and dest are arrays. 5648 const Type* src_type = src->Value(&_gvn); 5649 const Type* dest_type = dest->Value(&_gvn); 5650 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5651 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5652 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5653 5654 // for the quick and dirty code we will skip all the checks. 5655 // we are just trying to get the call to be generated. 5656 Node* src_start = src; 5657 Node* dest_start = dest; 5658 if (src_offset != NULL || dest_offset != NULL) { 5659 assert(src_offset != NULL && dest_offset != NULL, ""); 5660 src_start = array_element_address(src, src_offset, T_BYTE); 5661 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5662 } 5663 5664 // now need to get the start of its expanded key array 5665 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5666 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5667 if (k_start == NULL) return false; 5668 5669 if (Matcher::pass_original_key_for_aes()) { 5670 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5671 // compatibility issues between Java key expansion and SPARC crypto instructions 5672 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5673 if (original_k_start == NULL) return false; 5674 5675 // Call the stub. 5676 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5677 stubAddr, stubName, TypePtr::BOTTOM, 5678 src_start, dest_start, k_start, original_k_start); 5679 } else { 5680 // Call the stub. 5681 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5682 stubAddr, stubName, TypePtr::BOTTOM, 5683 src_start, dest_start, k_start); 5684 } 5685 5686 return true; 5687 } 5688 5689 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5690 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5691 address stubAddr; 5692 const char *stubName; 5693 5694 assert(UseAES, "need AES instruction support"); 5695 5696 switch(id) { 5697 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5698 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5699 stubName = "cipherBlockChaining_encryptAESCrypt"; 5700 break; 5701 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5702 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5703 stubName = "cipherBlockChaining_decryptAESCrypt"; 5704 break; 5705 } 5706 if (stubAddr == NULL) return false; 5707 5708 Node* cipherBlockChaining_object = argument(0); 5709 Node* src = argument(1); 5710 Node* src_offset = argument(2); 5711 Node* len = argument(3); 5712 Node* dest = argument(4); 5713 Node* dest_offset = argument(5); 5714 5715 // (1) src and dest are arrays. 5716 const Type* src_type = src->Value(&_gvn); 5717 const Type* dest_type = dest->Value(&_gvn); 5718 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5719 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5720 assert (top_src != NULL && top_src->klass() != NULL 5721 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5722 5723 // checks are the responsibility of the caller 5724 Node* src_start = src; 5725 Node* dest_start = dest; 5726 if (src_offset != NULL || dest_offset != NULL) { 5727 assert(src_offset != NULL && dest_offset != NULL, ""); 5728 src_start = array_element_address(src, src_offset, T_BYTE); 5729 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5730 } 5731 5732 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5733 // (because of the predicated logic executed earlier). 5734 // so we cast it here safely. 5735 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5736 5737 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5738 if (embeddedCipherObj == NULL) return false; 5739 5740 // cast it to what we know it will be at runtime 5741 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5742 assert(tinst != NULL, "CBC obj is null"); 5743 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5744 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5745 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5746 5747 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5748 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5749 const TypeOopPtr* xtype = aklass->as_instance_type(); 5750 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5751 aescrypt_object = _gvn.transform(aescrypt_object); 5752 5753 // we need to get the start of the aescrypt_object's expanded key array 5754 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5755 if (k_start == NULL) return false; 5756 5757 // similarly, get the start address of the r vector 5758 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5759 if (objRvec == NULL) return false; 5760 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5761 5762 Node* cbcCrypt; 5763 if (Matcher::pass_original_key_for_aes()) { 5764 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5765 // compatibility issues between Java key expansion and SPARC crypto instructions 5766 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5767 if (original_k_start == NULL) return false; 5768 5769 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5770 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5771 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5772 stubAddr, stubName, TypePtr::BOTTOM, 5773 src_start, dest_start, k_start, r_start, len, original_k_start); 5774 } else { 5775 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5776 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5777 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5778 stubAddr, stubName, TypePtr::BOTTOM, 5779 src_start, dest_start, k_start, r_start, len); 5780 } 5781 5782 // return cipher length (int) 5783 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5784 set_result(retvalue); 5785 return true; 5786 } 5787 5788 //------------------------------get_key_start_from_aescrypt_object----------------------- 5789 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5790 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5791 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5792 if (objAESCryptKey == NULL) return (Node *) NULL; 5793 5794 // now have the array, need to get the start address of the K array 5795 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5796 return k_start; 5797 } 5798 5799 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5800 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5801 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5802 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5803 if (objAESCryptKey == NULL) return (Node *) NULL; 5804 5805 // now have the array, need to get the start address of the lastKey array 5806 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5807 return original_k_start; 5808 } 5809 5810 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5811 // Return node representing slow path of predicate check. 5812 // the pseudo code we want to emulate with this predicate is: 5813 // for encryption: 5814 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5815 // for decryption: 5816 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5817 // note cipher==plain is more conservative than the original java code but that's OK 5818 // 5819 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5820 // The receiver was checked for NULL already. 5821 Node* objCBC = argument(0); 5822 5823 // Load embeddedCipher field of CipherBlockChaining object. 5824 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5825 5826 // get AESCrypt klass for instanceOf check 5827 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5828 // will have same classloader as CipherBlockChaining object 5829 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5830 assert(tinst != NULL, "CBCobj is null"); 5831 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5832 5833 // we want to do an instanceof comparison against the AESCrypt class 5834 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5835 if (!klass_AESCrypt->is_loaded()) { 5836 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5837 Node* ctrl = control(); 5838 set_control(top()); // no regular fast path 5839 return ctrl; 5840 } 5841 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5842 5843 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5844 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5845 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5846 5847 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5848 5849 // for encryption, we are done 5850 if (!decrypting) 5851 return instof_false; // even if it is NULL 5852 5853 // for decryption, we need to add a further check to avoid 5854 // taking the intrinsic path when cipher and plain are the same 5855 // see the original java code for why. 5856 RegionNode* region = new RegionNode(3); 5857 region->init_req(1, instof_false); 5858 Node* src = argument(1); 5859 Node* dest = argument(4); 5860 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5861 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5862 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5863 region->init_req(2, src_dest_conjoint); 5864 5865 record_for_igvn(region); 5866 return _gvn.transform(region); 5867 } 5868 5869 //------------------------------inline_ghash_processBlocks 5870 bool LibraryCallKit::inline_ghash_processBlocks() { 5871 address stubAddr; 5872 const char *stubName; 5873 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 5874 5875 stubAddr = StubRoutines::ghash_processBlocks(); 5876 stubName = "ghash_processBlocks"; 5877 5878 Node* data = argument(0); 5879 Node* offset = argument(1); 5880 Node* len = argument(2); 5881 Node* state = argument(3); 5882 Node* subkeyH = argument(4); 5883 5884 Node* state_start = array_element_address(state, intcon(0), T_LONG); 5885 assert(state_start, "state is NULL"); 5886 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 5887 assert(subkeyH_start, "subkeyH is NULL"); 5888 Node* data_start = array_element_address(data, offset, T_BYTE); 5889 assert(data_start, "data is NULL"); 5890 5891 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 5892 OptoRuntime::ghash_processBlocks_Type(), 5893 stubAddr, stubName, TypePtr::BOTTOM, 5894 state_start, subkeyH_start, data_start, len); 5895 return true; 5896 } 5897 5898 //------------------------------inline_sha_implCompress----------------------- 5899 // 5900 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 5901 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 5902 // 5903 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 5904 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 5905 // 5906 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 5907 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 5908 // 5909 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 5910 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 5911 5912 Node* sha_obj = argument(0); 5913 Node* src = argument(1); // type oop 5914 Node* ofs = argument(2); // type int 5915 5916 const Type* src_type = src->Value(&_gvn); 5917 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5918 if (top_src == NULL || top_src->klass() == NULL) { 5919 // failed array check 5920 return false; 5921 } 5922 // Figure out the size and type of the elements we will be copying. 5923 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5924 if (src_elem != T_BYTE) { 5925 return false; 5926 } 5927 // 'src_start' points to src array + offset 5928 Node* src_start = array_element_address(src, ofs, src_elem); 5929 Node* state = NULL; 5930 address stubAddr; 5931 const char *stubName; 5932 5933 switch(id) { 5934 case vmIntrinsics::_sha_implCompress: 5935 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 5936 state = get_state_from_sha_object(sha_obj); 5937 stubAddr = StubRoutines::sha1_implCompress(); 5938 stubName = "sha1_implCompress"; 5939 break; 5940 case vmIntrinsics::_sha2_implCompress: 5941 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 5942 state = get_state_from_sha_object(sha_obj); 5943 stubAddr = StubRoutines::sha256_implCompress(); 5944 stubName = "sha256_implCompress"; 5945 break; 5946 case vmIntrinsics::_sha5_implCompress: 5947 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 5948 state = get_state_from_sha5_object(sha_obj); 5949 stubAddr = StubRoutines::sha512_implCompress(); 5950 stubName = "sha512_implCompress"; 5951 break; 5952 default: 5953 fatal_unexpected_iid(id); 5954 return false; 5955 } 5956 if (state == NULL) return false; 5957 5958 // Call the stub. 5959 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 5960 stubAddr, stubName, TypePtr::BOTTOM, 5961 src_start, state); 5962 5963 return true; 5964 } 5965 5966 //------------------------------inline_digestBase_implCompressMB----------------------- 5967 // 5968 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 5969 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 5970 // 5971 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 5972 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5973 "need SHA1/SHA256/SHA512 instruction support"); 5974 assert((uint)predicate < 3, "sanity"); 5975 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 5976 5977 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 5978 Node* src = argument(1); // byte[] array 5979 Node* ofs = argument(2); // type int 5980 Node* limit = argument(3); // type int 5981 5982 const Type* src_type = src->Value(&_gvn); 5983 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5984 if (top_src == NULL || top_src->klass() == NULL) { 5985 // failed array check 5986 return false; 5987 } 5988 // Figure out the size and type of the elements we will be copying. 5989 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5990 if (src_elem != T_BYTE) { 5991 return false; 5992 } 5993 // 'src_start' points to src array + offset 5994 Node* src_start = array_element_address(src, ofs, src_elem); 5995 5996 const char* klass_SHA_name = NULL; 5997 const char* stub_name = NULL; 5998 address stub_addr = NULL; 5999 bool long_state = false; 6000 6001 switch (predicate) { 6002 case 0: 6003 if (UseSHA1Intrinsics) { 6004 klass_SHA_name = "sun/security/provider/SHA"; 6005 stub_name = "sha1_implCompressMB"; 6006 stub_addr = StubRoutines::sha1_implCompressMB(); 6007 } 6008 break; 6009 case 1: 6010 if (UseSHA256Intrinsics) { 6011 klass_SHA_name = "sun/security/provider/SHA2"; 6012 stub_name = "sha256_implCompressMB"; 6013 stub_addr = StubRoutines::sha256_implCompressMB(); 6014 } 6015 break; 6016 case 2: 6017 if (UseSHA512Intrinsics) { 6018 klass_SHA_name = "sun/security/provider/SHA5"; 6019 stub_name = "sha512_implCompressMB"; 6020 stub_addr = StubRoutines::sha512_implCompressMB(); 6021 long_state = true; 6022 } 6023 break; 6024 default: 6025 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6026 } 6027 if (klass_SHA_name != NULL) { 6028 // get DigestBase klass to lookup for SHA klass 6029 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6030 assert(tinst != NULL, "digestBase_obj is not instance???"); 6031 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6032 6033 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6034 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6035 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6036 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6037 } 6038 return false; 6039 } 6040 //------------------------------inline_sha_implCompressMB----------------------- 6041 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6042 bool long_state, address stubAddr, const char *stubName, 6043 Node* src_start, Node* ofs, Node* limit) { 6044 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6045 const TypeOopPtr* xtype = aklass->as_instance_type(); 6046 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6047 sha_obj = _gvn.transform(sha_obj); 6048 6049 Node* state; 6050 if (long_state) { 6051 state = get_state_from_sha5_object(sha_obj); 6052 } else { 6053 state = get_state_from_sha_object(sha_obj); 6054 } 6055 if (state == NULL) return false; 6056 6057 // Call the stub. 6058 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6059 OptoRuntime::digestBase_implCompressMB_Type(), 6060 stubAddr, stubName, TypePtr::BOTTOM, 6061 src_start, state, ofs, limit); 6062 // return ofs (int) 6063 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6064 set_result(result); 6065 6066 return true; 6067 } 6068 6069 //------------------------------get_state_from_sha_object----------------------- 6070 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6071 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6072 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6073 if (sha_state == NULL) return (Node *) NULL; 6074 6075 // now have the array, need to get the start address of the state array 6076 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6077 return state; 6078 } 6079 6080 //------------------------------get_state_from_sha5_object----------------------- 6081 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6082 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6083 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6084 if (sha_state == NULL) return (Node *) NULL; 6085 6086 // now have the array, need to get the start address of the state array 6087 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6088 return state; 6089 } 6090 6091 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6092 // Return node representing slow path of predicate check. 6093 // the pseudo code we want to emulate with this predicate is: 6094 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6095 // 6096 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6097 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6098 "need SHA1/SHA256/SHA512 instruction support"); 6099 assert((uint)predicate < 3, "sanity"); 6100 6101 // The receiver was checked for NULL already. 6102 Node* digestBaseObj = argument(0); 6103 6104 // get DigestBase klass for instanceOf check 6105 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6106 assert(tinst != NULL, "digestBaseObj is null"); 6107 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6108 6109 const char* klass_SHA_name = NULL; 6110 switch (predicate) { 6111 case 0: 6112 if (UseSHA1Intrinsics) { 6113 // we want to do an instanceof comparison against the SHA class 6114 klass_SHA_name = "sun/security/provider/SHA"; 6115 } 6116 break; 6117 case 1: 6118 if (UseSHA256Intrinsics) { 6119 // we want to do an instanceof comparison against the SHA2 class 6120 klass_SHA_name = "sun/security/provider/SHA2"; 6121 } 6122 break; 6123 case 2: 6124 if (UseSHA512Intrinsics) { 6125 // we want to do an instanceof comparison against the SHA5 class 6126 klass_SHA_name = "sun/security/provider/SHA5"; 6127 } 6128 break; 6129 default: 6130 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6131 } 6132 6133 ciKlass* klass_SHA = NULL; 6134 if (klass_SHA_name != NULL) { 6135 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6136 } 6137 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6138 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6139 Node* ctrl = control(); 6140 set_control(top()); // no intrinsic path 6141 return ctrl; 6142 } 6143 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6144 6145 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6146 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6147 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6148 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6149 6150 return instof_false; // even if it is NULL 6151 } 6152 6153 bool LibraryCallKit::inline_profileBoolean() { 6154 Node* counts = argument(1); 6155 const TypeAryPtr* ary = NULL; 6156 ciArray* aobj = NULL; 6157 if (counts->is_Con() 6158 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6159 && (aobj = ary->const_oop()->as_array()) != NULL 6160 && (aobj->length() == 2)) { 6161 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6162 jint false_cnt = aobj->element_value(0).as_int(); 6163 jint true_cnt = aobj->element_value(1).as_int(); 6164 6165 if (C->log() != NULL) { 6166 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6167 false_cnt, true_cnt); 6168 } 6169 6170 if (false_cnt + true_cnt == 0) { 6171 // According to profile, never executed. 6172 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6173 Deoptimization::Action_reinterpret); 6174 return true; 6175 } 6176 6177 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6178 // is a number of each value occurrences. 6179 Node* result = argument(0); 6180 if (false_cnt == 0 || true_cnt == 0) { 6181 // According to profile, one value has been never seen. 6182 int expected_val = (false_cnt == 0) ? 1 : 0; 6183 6184 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6185 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6186 6187 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6188 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6189 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6190 6191 { // Slow path: uncommon trap for never seen value and then reexecute 6192 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6193 // the value has been seen at least once. 6194 PreserveJVMState pjvms(this); 6195 PreserveReexecuteState preexecs(this); 6196 jvms()->set_should_reexecute(true); 6197 6198 set_control(slow_path); 6199 set_i_o(i_o()); 6200 6201 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6202 Deoptimization::Action_reinterpret); 6203 } 6204 // The guard for never seen value enables sharpening of the result and 6205 // returning a constant. It allows to eliminate branches on the same value 6206 // later on. 6207 set_control(fast_path); 6208 result = intcon(expected_val); 6209 } 6210 // Stop profiling. 6211 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6212 // By replacing method body with profile data (represented as ProfileBooleanNode 6213 // on IR level) we effectively disable profiling. 6214 // It enables full speed execution once optimized code is generated. 6215 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6216 C->record_for_igvn(profile); 6217 set_result(profile); 6218 return true; 6219 } else { 6220 // Continue profiling. 6221 // Profile data isn't available at the moment. So, execute method's bytecode version. 6222 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6223 // is compiled and counters aren't available since corresponding MethodHandle 6224 // isn't a compile-time constant. 6225 return false; 6226 } 6227 } 6228 6229 bool LibraryCallKit::inline_isCompileConstant() { 6230 Node* n = argument(0); 6231 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6232 return true; 6233 }