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 ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 638 kit.try_to_inline(_last_predicate)) { 639 if (C->print_intrinsics() || C->print_inlining()) { 640 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 641 } 642 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 643 if (C->log()) { 644 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 645 vmIntrinsics::name_at(intrinsic_id()), 646 (is_virtual() ? " virtual='1'" : ""), 647 C->unique() - nodes); 648 } 649 // Push the result from the inlined method onto the stack. 650 kit.push_result(); 651 C->print_inlining_update(this); 652 return kit.transfer_exceptions_into_jvms(); 653 } 654 655 // The intrinsic bailed out 656 if (C->print_intrinsics() || C->print_inlining()) { 657 if (jvms->has_method()) { 658 // Not a root compile. 659 const char* msg; 660 if (callee->intrinsic_candidate()) { 661 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 662 } else { 663 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 664 : "failed to inline (intrinsic), method not annotated"; 665 } 666 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 667 } else { 668 // Root compile 669 tty->print("Did not generate intrinsic %s%s at bci:%d in", 670 vmIntrinsics::name_at(intrinsic_id()), 671 (is_virtual() ? " (virtual)" : ""), bci); 672 } 673 } 674 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 675 C->print_inlining_update(this); 676 return NULL; 677 } 678 679 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 680 LibraryCallKit kit(jvms, this); 681 Compile* C = kit.C; 682 int nodes = C->unique(); 683 _last_predicate = predicate; 684 #ifndef PRODUCT 685 assert(is_predicated() && predicate < predicates_count(), "sanity"); 686 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 687 char buf[1000]; 688 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 689 tty->print_cr("Predicate for intrinsic %s", str); 690 } 691 #endif 692 ciMethod* callee = kit.callee(); 693 const int bci = kit.bci(); 694 695 Node* slow_ctl = kit.try_to_predicate(predicate); 696 if (!kit.failing()) { 697 if (C->print_intrinsics() || C->print_inlining()) { 698 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 699 } 700 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 701 if (C->log()) { 702 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 703 vmIntrinsics::name_at(intrinsic_id()), 704 (is_virtual() ? " virtual='1'" : ""), 705 C->unique() - nodes); 706 } 707 return slow_ctl; // Could be NULL if the check folds. 708 } 709 710 // The intrinsic bailed out 711 if (C->print_intrinsics() || C->print_inlining()) { 712 if (jvms->has_method()) { 713 // Not a root compile. 714 const char* msg = "failed to generate predicate for intrinsic"; 715 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 716 } else { 717 // Root compile 718 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 719 vmIntrinsics::name_at(intrinsic_id()), 720 (is_virtual() ? " (virtual)" : ""), bci); 721 } 722 } 723 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 724 return NULL; 725 } 726 727 bool LibraryCallKit::try_to_inline(int predicate) { 728 // Handle symbolic names for otherwise undistinguished boolean switches: 729 const bool is_store = true; 730 const bool is_native_ptr = true; 731 const bool is_static = true; 732 const bool is_volatile = true; 733 734 if (!jvms()->has_method()) { 735 // Root JVMState has a null method. 736 assert(map()->memory()->Opcode() == Op_Parm, ""); 737 // Insert the memory aliasing node 738 set_all_memory(reset_memory()); 739 } 740 assert(merged_memory(), ""); 741 742 743 switch (intrinsic_id()) { 744 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 745 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 746 case vmIntrinsics::_getClass: return inline_native_getClass(); 747 748 case vmIntrinsics::_dsin: 749 case vmIntrinsics::_dcos: 750 case vmIntrinsics::_dtan: 751 case vmIntrinsics::_dabs: 752 case vmIntrinsics::_datan2: 753 case vmIntrinsics::_dsqrt: 754 case vmIntrinsics::_dexp: 755 case vmIntrinsics::_dlog: 756 case vmIntrinsics::_dlog10: 757 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 758 759 case vmIntrinsics::_min: 760 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 761 762 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 763 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 764 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 765 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 766 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 767 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 768 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 769 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 770 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 771 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 772 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 773 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 774 775 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 776 777 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 778 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 779 case vmIntrinsics::_equals: return inline_string_equals(); 780 781 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 782 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 783 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 784 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 785 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 786 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 787 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 788 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 789 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 790 791 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 792 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 793 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 794 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 795 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 796 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 797 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 798 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 799 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 800 801 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 802 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 803 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 804 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 805 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 806 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 807 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 808 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 809 810 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 811 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 812 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 813 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 814 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 815 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 816 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 817 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 818 819 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 820 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 821 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 822 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 823 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 824 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 825 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 826 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 827 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 828 829 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 830 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 831 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 832 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 833 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 834 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 835 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 836 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 837 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 838 839 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 840 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 841 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 842 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 843 844 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 845 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 846 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 847 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 848 849 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 850 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 851 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 852 853 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 854 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 855 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 856 857 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 858 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 859 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 860 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 861 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 862 863 case vmIntrinsics::_loadFence: 864 case vmIntrinsics::_storeFence: 865 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 866 867 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 868 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 869 870 #ifdef TRACE_HAVE_INTRINSICS 871 case vmIntrinsics::_classID: return inline_native_classID(); 872 case vmIntrinsics::_threadID: return inline_native_threadID(); 873 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 874 #endif 875 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 876 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 877 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 878 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 879 case vmIntrinsics::_newArray: return inline_native_newArray(); 880 case vmIntrinsics::_getLength: return inline_native_getLength(); 881 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 882 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 883 case vmIntrinsics::_equalsC: return inline_array_equals(); 884 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 885 886 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 887 888 case vmIntrinsics::_isInstance: 889 case vmIntrinsics::_getModifiers: 890 case vmIntrinsics::_isInterface: 891 case vmIntrinsics::_isArray: 892 case vmIntrinsics::_isPrimitive: 893 case vmIntrinsics::_getSuperclass: 894 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 895 896 case vmIntrinsics::_floatToRawIntBits: 897 case vmIntrinsics::_floatToIntBits: 898 case vmIntrinsics::_intBitsToFloat: 899 case vmIntrinsics::_doubleToRawLongBits: 900 case vmIntrinsics::_doubleToLongBits: 901 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 902 903 case vmIntrinsics::_numberOfLeadingZeros_i: 904 case vmIntrinsics::_numberOfLeadingZeros_l: 905 case vmIntrinsics::_numberOfTrailingZeros_i: 906 case vmIntrinsics::_numberOfTrailingZeros_l: 907 case vmIntrinsics::_bitCount_i: 908 case vmIntrinsics::_bitCount_l: 909 case vmIntrinsics::_reverseBytes_i: 910 case vmIntrinsics::_reverseBytes_l: 911 case vmIntrinsics::_reverseBytes_s: 912 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 913 914 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 915 916 case vmIntrinsics::_Reference_get: return inline_reference_get(); 917 918 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 919 920 case vmIntrinsics::_aescrypt_encryptBlock: 921 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 922 923 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 924 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 925 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 926 927 case vmIntrinsics::_sha_implCompress: 928 case vmIntrinsics::_sha2_implCompress: 929 case vmIntrinsics::_sha5_implCompress: 930 return inline_sha_implCompress(intrinsic_id()); 931 932 case vmIntrinsics::_digestBase_implCompressMB: 933 return inline_digestBase_implCompressMB(predicate); 934 935 case vmIntrinsics::_multiplyToLen: 936 return inline_multiplyToLen(); 937 938 case vmIntrinsics::_squareToLen: 939 return inline_squareToLen(); 940 941 case vmIntrinsics::_mulAdd: 942 return inline_mulAdd(); 943 944 case vmIntrinsics::_ghash_processBlocks: 945 return inline_ghash_processBlocks(); 946 947 case vmIntrinsics::_encodeISOArray: 948 return inline_encodeISOArray(); 949 950 case vmIntrinsics::_updateCRC32: 951 return inline_updateCRC32(); 952 case vmIntrinsics::_updateBytesCRC32: 953 return inline_updateBytesCRC32(); 954 case vmIntrinsics::_updateByteBufferCRC32: 955 return inline_updateByteBufferCRC32(); 956 957 case vmIntrinsics::_profileBoolean: 958 return inline_profileBoolean(); 959 case vmIntrinsics::_isCompileConstant: 960 return inline_isCompileConstant(); 961 962 default: 963 // If you get here, it may be that someone has added a new intrinsic 964 // to the list in vmSymbols.hpp without implementing it here. 965 #ifndef PRODUCT 966 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 967 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 968 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 969 } 970 #endif 971 return false; 972 } 973 } 974 975 Node* LibraryCallKit::try_to_predicate(int predicate) { 976 if (!jvms()->has_method()) { 977 // Root JVMState has a null method. 978 assert(map()->memory()->Opcode() == Op_Parm, ""); 979 // Insert the memory aliasing node 980 set_all_memory(reset_memory()); 981 } 982 assert(merged_memory(), ""); 983 984 switch (intrinsic_id()) { 985 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 986 return inline_cipherBlockChaining_AESCrypt_predicate(false); 987 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 988 return inline_cipherBlockChaining_AESCrypt_predicate(true); 989 case vmIntrinsics::_digestBase_implCompressMB: 990 return inline_digestBase_implCompressMB_predicate(predicate); 991 992 default: 993 // If you get here, it may be that someone has added a new intrinsic 994 // to the list in vmSymbols.hpp without implementing it here. 995 #ifndef PRODUCT 996 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 997 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 998 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 999 } 1000 #endif 1001 Node* slow_ctl = control(); 1002 set_control(top()); // No fast path instrinsic 1003 return slow_ctl; 1004 } 1005 } 1006 1007 //------------------------------set_result------------------------------- 1008 // Helper function for finishing intrinsics. 1009 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 1010 record_for_igvn(region); 1011 set_control(_gvn.transform(region)); 1012 set_result( _gvn.transform(value)); 1013 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 1014 } 1015 1016 //------------------------------generate_guard--------------------------- 1017 // Helper function for generating guarded fast-slow graph structures. 1018 // The given 'test', if true, guards a slow path. If the test fails 1019 // then a fast path can be taken. (We generally hope it fails.) 1020 // In all cases, GraphKit::control() is updated to the fast path. 1021 // The returned value represents the control for the slow path. 1022 // The return value is never 'top'; it is either a valid control 1023 // or NULL if it is obvious that the slow path can never be taken. 1024 // Also, if region and the slow control are not NULL, the slow edge 1025 // is appended to the region. 1026 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 1027 if (stopped()) { 1028 // Already short circuited. 1029 return NULL; 1030 } 1031 1032 // Build an if node and its projections. 1033 // If test is true we take the slow path, which we assume is uncommon. 1034 if (_gvn.type(test) == TypeInt::ZERO) { 1035 // The slow branch is never taken. No need to build this guard. 1036 return NULL; 1037 } 1038 1039 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1040 1041 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1042 if (if_slow == top()) { 1043 // The slow branch is never taken. No need to build this guard. 1044 return NULL; 1045 } 1046 1047 if (region != NULL) 1048 region->add_req(if_slow); 1049 1050 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1051 set_control(if_fast); 1052 1053 return if_slow; 1054 } 1055 1056 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1057 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1058 } 1059 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1060 return generate_guard(test, region, PROB_FAIR); 1061 } 1062 1063 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1064 Node* *pos_index) { 1065 if (stopped()) 1066 return NULL; // already stopped 1067 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1068 return NULL; // index is already adequately typed 1069 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1070 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1071 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1072 if (is_neg != NULL && pos_index != NULL) { 1073 // Emulate effect of Parse::adjust_map_after_if. 1074 Node* ccast = new CastIINode(index, TypeInt::POS); 1075 ccast->set_req(0, control()); 1076 (*pos_index) = _gvn.transform(ccast); 1077 } 1078 return is_neg; 1079 } 1080 1081 // Make sure that 'position' is a valid limit index, in [0..length]. 1082 // There are two equivalent plans for checking this: 1083 // A. (offset + copyLength) unsigned<= arrayLength 1084 // B. offset <= (arrayLength - copyLength) 1085 // We require that all of the values above, except for the sum and 1086 // difference, are already known to be non-negative. 1087 // Plan A is robust in the face of overflow, if offset and copyLength 1088 // are both hugely positive. 1089 // 1090 // Plan B is less direct and intuitive, but it does not overflow at 1091 // all, since the difference of two non-negatives is always 1092 // representable. Whenever Java methods must perform the equivalent 1093 // check they generally use Plan B instead of Plan A. 1094 // For the moment we use Plan A. 1095 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1096 Node* subseq_length, 1097 Node* array_length, 1098 RegionNode* region) { 1099 if (stopped()) 1100 return NULL; // already stopped 1101 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1102 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1103 return NULL; // common case of whole-array copy 1104 Node* last = subseq_length; 1105 if (!zero_offset) // last += offset 1106 last = _gvn.transform(new AddINode(last, offset)); 1107 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1108 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1109 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1110 return is_over; 1111 } 1112 1113 1114 //--------------------------generate_current_thread-------------------- 1115 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1116 ciKlass* thread_klass = env()->Thread_klass(); 1117 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1118 Node* thread = _gvn.transform(new ThreadLocalNode()); 1119 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1120 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1121 tls_output = thread; 1122 return threadObj; 1123 } 1124 1125 1126 //------------------------------make_string_method_node------------------------ 1127 // Helper method for String intrinsic functions. This version is called 1128 // with str1 and str2 pointing to String object nodes. 1129 // 1130 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1131 Node* no_ctrl = NULL; 1132 1133 // Get start addr of string 1134 Node* str1_value = load_String_value(no_ctrl, str1); 1135 Node* str1_offset = load_String_offset(no_ctrl, str1); 1136 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1137 1138 // Get length of string 1 1139 Node* str1_len = load_String_length(no_ctrl, str1); 1140 1141 Node* str2_value = load_String_value(no_ctrl, str2); 1142 Node* str2_offset = load_String_offset(no_ctrl, str2); 1143 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1144 1145 Node* str2_len = NULL; 1146 Node* result = NULL; 1147 1148 switch (opcode) { 1149 case Op_StrIndexOf: 1150 // Get length of string 2 1151 str2_len = load_String_length(no_ctrl, str2); 1152 1153 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1154 str1_start, str1_len, str2_start, str2_len); 1155 break; 1156 case Op_StrComp: 1157 // Get length of string 2 1158 str2_len = load_String_length(no_ctrl, str2); 1159 1160 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1161 str1_start, str1_len, str2_start, str2_len); 1162 break; 1163 case Op_StrEquals: 1164 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1165 str1_start, str2_start, str1_len); 1166 break; 1167 default: 1168 ShouldNotReachHere(); 1169 return NULL; 1170 } 1171 1172 // All these intrinsics have checks. 1173 C->set_has_split_ifs(true); // Has chance for split-if optimization 1174 1175 return _gvn.transform(result); 1176 } 1177 1178 // Helper method for String intrinsic functions. This version is called 1179 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1180 // to Int nodes containing the lenghts of str1 and str2. 1181 // 1182 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1183 Node* result = NULL; 1184 switch (opcode) { 1185 case Op_StrIndexOf: 1186 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1187 str1_start, cnt1, str2_start, cnt2); 1188 break; 1189 case Op_StrComp: 1190 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1191 str1_start, cnt1, str2_start, cnt2); 1192 break; 1193 case Op_StrEquals: 1194 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1195 str1_start, str2_start, cnt1); 1196 break; 1197 default: 1198 ShouldNotReachHere(); 1199 return NULL; 1200 } 1201 1202 // All these intrinsics have checks. 1203 C->set_has_split_ifs(true); // Has chance for split-if optimization 1204 1205 return _gvn.transform(result); 1206 } 1207 1208 //------------------------------inline_string_compareTo------------------------ 1209 // public int java.lang.String.compareTo(String anotherString); 1210 bool LibraryCallKit::inline_string_compareTo() { 1211 Node* receiver = null_check(argument(0)); 1212 Node* arg = null_check(argument(1)); 1213 if (stopped()) { 1214 return true; 1215 } 1216 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1217 return true; 1218 } 1219 1220 //------------------------------inline_string_equals------------------------ 1221 bool LibraryCallKit::inline_string_equals() { 1222 Node* receiver = null_check_receiver(); 1223 // NOTE: Do not null check argument for String.equals() because spec 1224 // allows to specify NULL as argument. 1225 Node* argument = this->argument(1); 1226 if (stopped()) { 1227 return true; 1228 } 1229 1230 // paths (plus control) merge 1231 RegionNode* region = new RegionNode(5); 1232 Node* phi = new PhiNode(region, TypeInt::BOOL); 1233 1234 // does source == target string? 1235 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1236 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1237 1238 Node* if_eq = generate_slow_guard(bol, NULL); 1239 if (if_eq != NULL) { 1240 // receiver == argument 1241 phi->init_req(2, intcon(1)); 1242 region->init_req(2, if_eq); 1243 } 1244 1245 // get String klass for instanceOf 1246 ciInstanceKlass* klass = env()->String_klass(); 1247 1248 if (!stopped()) { 1249 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1250 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1251 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1252 1253 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1254 //instanceOf == true, fallthrough 1255 1256 if (inst_false != NULL) { 1257 phi->init_req(3, intcon(0)); 1258 region->init_req(3, inst_false); 1259 } 1260 } 1261 1262 if (!stopped()) { 1263 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1264 1265 // Properly cast the argument to String 1266 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1267 // This path is taken only when argument's type is String:NotNull. 1268 argument = cast_not_null(argument, false); 1269 1270 Node* no_ctrl = NULL; 1271 1272 // Get start addr of receiver 1273 Node* receiver_val = load_String_value(no_ctrl, receiver); 1274 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1275 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1276 1277 // Get length of receiver 1278 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1279 1280 // Get start addr of argument 1281 Node* argument_val = load_String_value(no_ctrl, argument); 1282 Node* argument_offset = load_String_offset(no_ctrl, argument); 1283 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1284 1285 // Get length of argument 1286 Node* argument_cnt = load_String_length(no_ctrl, argument); 1287 1288 // Check for receiver count != argument count 1289 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1290 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1291 Node* if_ne = generate_slow_guard(bol, NULL); 1292 if (if_ne != NULL) { 1293 phi->init_req(4, intcon(0)); 1294 region->init_req(4, if_ne); 1295 } 1296 1297 // Check for count == 0 is done by assembler code for StrEquals. 1298 1299 if (!stopped()) { 1300 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1301 phi->init_req(1, equals); 1302 region->init_req(1, control()); 1303 } 1304 } 1305 1306 // post merge 1307 set_control(_gvn.transform(region)); 1308 record_for_igvn(region); 1309 1310 set_result(_gvn.transform(phi)); 1311 return true; 1312 } 1313 1314 //------------------------------inline_array_equals---------------------------- 1315 bool LibraryCallKit::inline_array_equals() { 1316 Node* arg1 = argument(0); 1317 Node* arg2 = argument(1); 1318 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1319 return true; 1320 } 1321 1322 // Java version of String.indexOf(constant string) 1323 // class StringDecl { 1324 // StringDecl(char[] ca) { 1325 // offset = 0; 1326 // count = ca.length; 1327 // value = ca; 1328 // } 1329 // int offset; 1330 // int count; 1331 // char[] value; 1332 // } 1333 // 1334 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1335 // int targetOffset, int cache_i, int md2) { 1336 // int cache = cache_i; 1337 // int sourceOffset = string_object.offset; 1338 // int sourceCount = string_object.count; 1339 // int targetCount = target_object.length; 1340 // 1341 // int targetCountLess1 = targetCount - 1; 1342 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1343 // 1344 // char[] source = string_object.value; 1345 // char[] target = target_object; 1346 // int lastChar = target[targetCountLess1]; 1347 // 1348 // outer_loop: 1349 // for (int i = sourceOffset; i < sourceEnd; ) { 1350 // int src = source[i + targetCountLess1]; 1351 // if (src == lastChar) { 1352 // // With random strings and a 4-character alphabet, 1353 // // reverse matching at this point sets up 0.8% fewer 1354 // // frames, but (paradoxically) makes 0.3% more probes. 1355 // // Since those probes are nearer the lastChar probe, 1356 // // there is may be a net D$ win with reverse matching. 1357 // // But, reversing loop inhibits unroll of inner loop 1358 // // for unknown reason. So, does running outer loop from 1359 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1360 // for (int j = 0; j < targetCountLess1; j++) { 1361 // if (target[targetOffset + j] != source[i+j]) { 1362 // if ((cache & (1 << source[i+j])) == 0) { 1363 // if (md2 < j+1) { 1364 // i += j+1; 1365 // continue outer_loop; 1366 // } 1367 // } 1368 // i += md2; 1369 // continue outer_loop; 1370 // } 1371 // } 1372 // return i - sourceOffset; 1373 // } 1374 // if ((cache & (1 << src)) == 0) { 1375 // i += targetCountLess1; 1376 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1377 // i++; 1378 // } 1379 // return -1; 1380 // } 1381 1382 //------------------------------string_indexOf------------------------ 1383 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1384 jint cache_i, jint md2_i) { 1385 1386 Node* no_ctrl = NULL; 1387 float likely = PROB_LIKELY(0.9); 1388 float unlikely = PROB_UNLIKELY(0.9); 1389 1390 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1391 1392 Node* source = load_String_value(no_ctrl, string_object); 1393 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1394 Node* sourceCount = load_String_length(no_ctrl, string_object); 1395 1396 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1397 jint target_length = target_array->length(); 1398 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1399 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1400 1401 // String.value field is known to be @Stable. 1402 if (UseImplicitStableValues) { 1403 target = cast_array_to_stable(target, target_type); 1404 } 1405 1406 IdealKit kit(this, false, true); 1407 #define __ kit. 1408 Node* zero = __ ConI(0); 1409 Node* one = __ ConI(1); 1410 Node* cache = __ ConI(cache_i); 1411 Node* md2 = __ ConI(md2_i); 1412 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1413 Node* targetCountLess1 = __ ConI(target_length - 1); 1414 Node* targetOffset = __ ConI(targetOffset_i); 1415 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1416 1417 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1418 Node* outer_loop = __ make_label(2 /* goto */); 1419 Node* return_ = __ make_label(1); 1420 1421 __ set(rtn,__ ConI(-1)); 1422 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1423 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1424 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1425 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1426 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1427 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1428 Node* tpj = __ AddI(targetOffset, __ value(j)); 1429 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1430 Node* ipj = __ AddI(__ value(i), __ value(j)); 1431 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1432 __ if_then(targ, BoolTest::ne, src2); { 1433 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1434 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1435 __ increment(i, __ AddI(__ value(j), one)); 1436 __ goto_(outer_loop); 1437 } __ end_if(); __ dead(j); 1438 }__ end_if(); __ dead(j); 1439 __ increment(i, md2); 1440 __ goto_(outer_loop); 1441 }__ end_if(); 1442 __ increment(j, one); 1443 }__ end_loop(); __ dead(j); 1444 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1445 __ goto_(return_); 1446 }__ end_if(); 1447 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1448 __ increment(i, targetCountLess1); 1449 }__ end_if(); 1450 __ increment(i, one); 1451 __ bind(outer_loop); 1452 }__ end_loop(); __ dead(i); 1453 __ bind(return_); 1454 1455 // Final sync IdealKit and GraphKit. 1456 final_sync(kit); 1457 Node* result = __ value(rtn); 1458 #undef __ 1459 C->set_has_loops(true); 1460 return result; 1461 } 1462 1463 //------------------------------inline_string_indexOf------------------------ 1464 bool LibraryCallKit::inline_string_indexOf() { 1465 Node* receiver = argument(0); 1466 Node* arg = argument(1); 1467 1468 Node* result; 1469 if (Matcher::has_match_rule(Op_StrIndexOf) && 1470 UseSSE42Intrinsics) { 1471 // Generate SSE4.2 version of indexOf 1472 // We currently only have match rules that use SSE4.2 1473 1474 receiver = null_check(receiver); 1475 arg = null_check(arg); 1476 if (stopped()) { 1477 return true; 1478 } 1479 1480 // Make the merge point 1481 RegionNode* result_rgn = new RegionNode(4); 1482 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1483 Node* no_ctrl = NULL; 1484 1485 // Get start addr of source string 1486 Node* source = load_String_value(no_ctrl, receiver); 1487 Node* source_offset = load_String_offset(no_ctrl, receiver); 1488 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1489 1490 // Get length of source string 1491 Node* source_cnt = load_String_length(no_ctrl, receiver); 1492 1493 // Get start addr of substring 1494 Node* substr = load_String_value(no_ctrl, arg); 1495 Node* substr_offset = load_String_offset(no_ctrl, arg); 1496 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1497 1498 // Get length of source string 1499 Node* substr_cnt = load_String_length(no_ctrl, arg); 1500 1501 // Check for substr count > string count 1502 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1503 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1504 Node* if_gt = generate_slow_guard(bol, NULL); 1505 if (if_gt != NULL) { 1506 result_phi->init_req(2, intcon(-1)); 1507 result_rgn->init_req(2, if_gt); 1508 } 1509 1510 if (!stopped()) { 1511 // Check for substr count == 0 1512 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1513 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1514 Node* if_zero = generate_slow_guard(bol, NULL); 1515 if (if_zero != NULL) { 1516 result_phi->init_req(3, intcon(0)); 1517 result_rgn->init_req(3, if_zero); 1518 } 1519 } 1520 1521 if (!stopped()) { 1522 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1523 result_phi->init_req(1, result); 1524 result_rgn->init_req(1, control()); 1525 } 1526 set_control(_gvn.transform(result_rgn)); 1527 record_for_igvn(result_rgn); 1528 result = _gvn.transform(result_phi); 1529 1530 } else { // Use LibraryCallKit::string_indexOf 1531 // don't intrinsify if argument isn't a constant string. 1532 if (!arg->is_Con()) { 1533 return false; 1534 } 1535 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1536 if (str_type == NULL) { 1537 return false; 1538 } 1539 ciInstanceKlass* klass = env()->String_klass(); 1540 ciObject* str_const = str_type->const_oop(); 1541 if (str_const == NULL || str_const->klass() != klass) { 1542 return false; 1543 } 1544 ciInstance* str = str_const->as_instance(); 1545 assert(str != NULL, "must be instance"); 1546 1547 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1548 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1549 1550 int o; 1551 int c; 1552 if (java_lang_String::has_offset_field()) { 1553 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1554 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1555 } else { 1556 o = 0; 1557 c = pat->length(); 1558 } 1559 1560 // constant strings have no offset and count == length which 1561 // simplifies the resulting code somewhat so lets optimize for that. 1562 if (o != 0 || c != pat->length()) { 1563 return false; 1564 } 1565 1566 receiver = null_check(receiver, T_OBJECT); 1567 // NOTE: No null check on the argument is needed since it's a constant String oop. 1568 if (stopped()) { 1569 return true; 1570 } 1571 1572 // The null string as a pattern always returns 0 (match at beginning of string) 1573 if (c == 0) { 1574 set_result(intcon(0)); 1575 return true; 1576 } 1577 1578 // Generate default indexOf 1579 jchar lastChar = pat->char_at(o + (c - 1)); 1580 int cache = 0; 1581 int i; 1582 for (i = 0; i < c - 1; i++) { 1583 assert(i < pat->length(), "out of range"); 1584 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1585 } 1586 1587 int md2 = c; 1588 for (i = 0; i < c - 1; i++) { 1589 assert(i < pat->length(), "out of range"); 1590 if (pat->char_at(o + i) == lastChar) { 1591 md2 = (c - 1) - i; 1592 } 1593 } 1594 1595 result = string_indexOf(receiver, pat, o, cache, md2); 1596 } 1597 set_result(result); 1598 return true; 1599 } 1600 1601 //--------------------------round_double_node-------------------------------- 1602 // Round a double node if necessary. 1603 Node* LibraryCallKit::round_double_node(Node* n) { 1604 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1605 n = _gvn.transform(new RoundDoubleNode(0, n)); 1606 return n; 1607 } 1608 1609 //------------------------------inline_math----------------------------------- 1610 // public static double Math.abs(double) 1611 // public static double Math.sqrt(double) 1612 // public static double Math.log(double) 1613 // public static double Math.log10(double) 1614 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1615 Node* arg = round_double_node(argument(0)); 1616 Node* n; 1617 switch (id) { 1618 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1619 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1620 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1621 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1622 default: fatal_unexpected_iid(id); break; 1623 } 1624 set_result(_gvn.transform(n)); 1625 return true; 1626 } 1627 1628 //------------------------------inline_trig---------------------------------- 1629 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1630 // argument reduction which will turn into a fast/slow diamond. 1631 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1632 Node* arg = round_double_node(argument(0)); 1633 Node* n = NULL; 1634 1635 switch (id) { 1636 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1637 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1638 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1639 default: fatal_unexpected_iid(id); break; 1640 } 1641 n = _gvn.transform(n); 1642 1643 // Rounding required? Check for argument reduction! 1644 if (Matcher::strict_fp_requires_explicit_rounding) { 1645 static const double pi_4 = 0.7853981633974483; 1646 static const double neg_pi_4 = -0.7853981633974483; 1647 // pi/2 in 80-bit extended precision 1648 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1649 // -pi/2 in 80-bit extended precision 1650 // 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}; 1651 // Cutoff value for using this argument reduction technique 1652 //static const double pi_2_minus_epsilon = 1.564660403643354; 1653 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1654 1655 // Pseudocode for sin: 1656 // if (x <= Math.PI / 4.0) { 1657 // if (x >= -Math.PI / 4.0) return fsin(x); 1658 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1659 // } else { 1660 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1661 // } 1662 // return StrictMath.sin(x); 1663 1664 // Pseudocode for cos: 1665 // if (x <= Math.PI / 4.0) { 1666 // if (x >= -Math.PI / 4.0) return fcos(x); 1667 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1668 // } else { 1669 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1670 // } 1671 // return StrictMath.cos(x); 1672 1673 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1674 // requires a special machine instruction to load it. Instead we'll try 1675 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1676 // probably do the math inside the SIN encoding. 1677 1678 // Make the merge point 1679 RegionNode* r = new RegionNode(3); 1680 Node* phi = new PhiNode(r, Type::DOUBLE); 1681 1682 // Flatten arg so we need only 1 test 1683 Node *abs = _gvn.transform(new AbsDNode(arg)); 1684 // Node for PI/4 constant 1685 Node *pi4 = makecon(TypeD::make(pi_4)); 1686 // Check PI/4 : abs(arg) 1687 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1688 // Check: If PI/4 < abs(arg) then go slow 1689 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1690 // Branch either way 1691 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1692 set_control(opt_iff(r,iff)); 1693 1694 // Set fast path result 1695 phi->init_req(2, n); 1696 1697 // Slow path - non-blocking leaf call 1698 Node* call = NULL; 1699 switch (id) { 1700 case vmIntrinsics::_dsin: 1701 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1702 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1703 "Sin", NULL, arg, top()); 1704 break; 1705 case vmIntrinsics::_dcos: 1706 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1707 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1708 "Cos", NULL, arg, top()); 1709 break; 1710 case vmIntrinsics::_dtan: 1711 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1712 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1713 "Tan", NULL, arg, top()); 1714 break; 1715 } 1716 assert(control()->in(0) == call, ""); 1717 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1718 r->init_req(1, control()); 1719 phi->init_req(1, slow_result); 1720 1721 // Post-merge 1722 set_control(_gvn.transform(r)); 1723 record_for_igvn(r); 1724 n = _gvn.transform(phi); 1725 1726 C->set_has_split_ifs(true); // Has chance for split-if optimization 1727 } 1728 set_result(n); 1729 return true; 1730 } 1731 1732 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1733 //------------------- 1734 //result=(result.isNaN())? funcAddr():result; 1735 // Check: If isNaN() by checking result!=result? then either trap 1736 // or go to runtime 1737 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1738 // Build the boolean node 1739 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1740 1741 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1742 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1743 // The pow or exp intrinsic returned a NaN, which requires a call 1744 // to the runtime. Recompile with the runtime call. 1745 uncommon_trap(Deoptimization::Reason_intrinsic, 1746 Deoptimization::Action_make_not_entrant); 1747 } 1748 return result; 1749 } else { 1750 // If this inlining ever returned NaN in the past, we compile a call 1751 // to the runtime to properly handle corner cases 1752 1753 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1754 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1755 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1756 1757 if (!if_slow->is_top()) { 1758 RegionNode* result_region = new RegionNode(3); 1759 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1760 1761 result_region->init_req(1, if_fast); 1762 result_val->init_req(1, result); 1763 1764 set_control(if_slow); 1765 1766 const TypePtr* no_memory_effects = NULL; 1767 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1768 no_memory_effects, 1769 x, top(), y, y ? top() : NULL); 1770 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1771 #ifdef ASSERT 1772 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1773 assert(value_top == top(), "second value must be top"); 1774 #endif 1775 1776 result_region->init_req(2, control()); 1777 result_val->init_req(2, value); 1778 set_control(_gvn.transform(result_region)); 1779 return _gvn.transform(result_val); 1780 } else { 1781 return result; 1782 } 1783 } 1784 } 1785 1786 //------------------------------inline_exp------------------------------------- 1787 // Inline exp instructions, if possible. The Intel hardware only misses 1788 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1789 bool LibraryCallKit::inline_exp() { 1790 Node* arg = round_double_node(argument(0)); 1791 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1792 1793 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1794 set_result(n); 1795 1796 C->set_has_split_ifs(true); // Has chance for split-if optimization 1797 return true; 1798 } 1799 1800 //------------------------------inline_pow------------------------------------- 1801 // Inline power instructions, if possible. 1802 bool LibraryCallKit::inline_pow() { 1803 // Pseudocode for pow 1804 // if (y == 2) { 1805 // return x * x; 1806 // } else { 1807 // if (x <= 0.0) { 1808 // long longy = (long)y; 1809 // if ((double)longy == y) { // if y is long 1810 // if (y + 1 == y) longy = 0; // huge number: even 1811 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1812 // } else { 1813 // result = NaN; 1814 // } 1815 // } else { 1816 // result = DPow(x,y); 1817 // } 1818 // if (result != result)? { 1819 // result = uncommon_trap() or runtime_call(); 1820 // } 1821 // return result; 1822 // } 1823 1824 Node* x = round_double_node(argument(0)); 1825 Node* y = round_double_node(argument(2)); 1826 1827 Node* result = NULL; 1828 1829 Node* const_two_node = makecon(TypeD::make(2.0)); 1830 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1831 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1832 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1833 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1834 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1835 1836 RegionNode* region_node = new RegionNode(3); 1837 region_node->init_req(1, if_true); 1838 1839 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1840 // special case for x^y where y == 2, we can convert it to x * x 1841 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1842 1843 // set control to if_false since we will now process the false branch 1844 set_control(if_false); 1845 1846 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1847 // Short form: skip the fancy tests and just check for NaN result. 1848 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1849 } else { 1850 // If this inlining ever returned NaN in the past, include all 1851 // checks + call to the runtime. 1852 1853 // Set the merge point for If node with condition of (x <= 0.0) 1854 // There are four possible paths to region node and phi node 1855 RegionNode *r = new RegionNode(4); 1856 Node *phi = new PhiNode(r, Type::DOUBLE); 1857 1858 // Build the first if node: if (x <= 0.0) 1859 // Node for 0 constant 1860 Node *zeronode = makecon(TypeD::ZERO); 1861 // Check x:0 1862 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1863 // Check: If (x<=0) then go complex path 1864 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1865 // Branch either way 1866 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1867 // Fast path taken; set region slot 3 1868 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1869 r->init_req(3,fast_taken); // Capture fast-control 1870 1871 // Fast path not-taken, i.e. slow path 1872 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1873 1874 // Set fast path result 1875 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1876 phi->init_req(3, fast_result); 1877 1878 // Complex path 1879 // Build the second if node (if y is long) 1880 // Node for (long)y 1881 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1882 // Node for (double)((long) y) 1883 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1884 // Check (double)((long) y) : y 1885 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1886 // Check if (y isn't long) then go to slow path 1887 1888 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1889 // Branch either way 1890 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1891 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1892 1893 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1894 1895 // Calculate DPow(abs(x), y)*(1 & (long)y) 1896 // Node for constant 1 1897 Node *conone = longcon(1); 1898 // 1& (long)y 1899 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1900 1901 // A huge number is always even. Detect a huge number by checking 1902 // if y + 1 == y and set integer to be tested for parity to 0. 1903 // Required for corner case: 1904 // (long)9.223372036854776E18 = max_jlong 1905 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1906 // max_jlong is odd but 9.223372036854776E18 is even 1907 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1908 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1909 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1910 Node* correctedsign = NULL; 1911 if (ConditionalMoveLimit != 0) { 1912 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1913 } else { 1914 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1915 RegionNode *r = new RegionNode(3); 1916 Node *phi = new PhiNode(r, TypeLong::LONG); 1917 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1918 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1919 phi->init_req(1, signnode); 1920 phi->init_req(2, longcon(0)); 1921 correctedsign = _gvn.transform(phi); 1922 ylong_path = _gvn.transform(r); 1923 record_for_igvn(r); 1924 } 1925 1926 // zero node 1927 Node *conzero = longcon(0); 1928 // Check (1&(long)y)==0? 1929 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1930 // Check if (1&(long)y)!=0?, if so the result is negative 1931 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1932 // abs(x) 1933 Node *absx=_gvn.transform(new AbsDNode(x)); 1934 // abs(x)^y 1935 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1936 // -abs(x)^y 1937 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1938 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1939 Node *signresult = NULL; 1940 if (ConditionalMoveLimit != 0) { 1941 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1942 } else { 1943 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1944 RegionNode *r = new RegionNode(3); 1945 Node *phi = new PhiNode(r, Type::DOUBLE); 1946 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1947 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1948 phi->init_req(1, absxpowy); 1949 phi->init_req(2, negabsxpowy); 1950 signresult = _gvn.transform(phi); 1951 ylong_path = _gvn.transform(r); 1952 record_for_igvn(r); 1953 } 1954 // Set complex path fast result 1955 r->init_req(2, ylong_path); 1956 phi->init_req(2, signresult); 1957 1958 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1959 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1960 r->init_req(1,slow_path); 1961 phi->init_req(1,slow_result); 1962 1963 // Post merge 1964 set_control(_gvn.transform(r)); 1965 record_for_igvn(r); 1966 result = _gvn.transform(phi); 1967 } 1968 1969 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1970 1971 // control from finish_pow_exp is now input to the region node 1972 region_node->set_req(2, control()); 1973 // the result from finish_pow_exp is now input to the phi node 1974 phi_node->init_req(2, result); 1975 set_control(_gvn.transform(region_node)); 1976 record_for_igvn(region_node); 1977 set_result(_gvn.transform(phi_node)); 1978 1979 C->set_has_split_ifs(true); // Has chance for split-if optimization 1980 return true; 1981 } 1982 1983 //------------------------------runtime_math----------------------------- 1984 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1985 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1986 "must be (DD)D or (D)D type"); 1987 1988 // Inputs 1989 Node* a = round_double_node(argument(0)); 1990 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1991 1992 const TypePtr* no_memory_effects = NULL; 1993 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1994 no_memory_effects, 1995 a, top(), b, b ? top() : NULL); 1996 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1997 #ifdef ASSERT 1998 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1999 assert(value_top == top(), "second value must be top"); 2000 #endif 2001 2002 set_result(value); 2003 return true; 2004 } 2005 2006 //------------------------------inline_math_native----------------------------- 2007 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 2008 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 2009 switch (id) { 2010 // These intrinsics are not properly supported on all hardware 2011 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 2012 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 2013 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 2014 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2015 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2016 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2017 2018 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2019 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2020 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2021 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2022 2023 // These intrinsics are supported on all hardware 2024 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2025 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2026 2027 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2028 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2029 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2030 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2031 #undef FN_PTR 2032 2033 // These intrinsics are not yet correctly implemented 2034 case vmIntrinsics::_datan2: 2035 return false; 2036 2037 default: 2038 fatal_unexpected_iid(id); 2039 return false; 2040 } 2041 } 2042 2043 static bool is_simple_name(Node* n) { 2044 return (n->req() == 1 // constant 2045 || (n->is_Type() && n->as_Type()->type()->singleton()) 2046 || n->is_Proj() // parameter or return value 2047 || n->is_Phi() // local of some sort 2048 ); 2049 } 2050 2051 //----------------------------inline_min_max----------------------------------- 2052 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2053 set_result(generate_min_max(id, argument(0), argument(1))); 2054 return true; 2055 } 2056 2057 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2058 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2059 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2060 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2061 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2062 2063 { 2064 PreserveJVMState pjvms(this); 2065 PreserveReexecuteState preexecs(this); 2066 jvms()->set_should_reexecute(true); 2067 2068 set_control(slow_path); 2069 set_i_o(i_o()); 2070 2071 uncommon_trap(Deoptimization::Reason_intrinsic, 2072 Deoptimization::Action_none); 2073 } 2074 2075 set_control(fast_path); 2076 set_result(math); 2077 } 2078 2079 template <typename OverflowOp> 2080 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2081 typedef typename OverflowOp::MathOp MathOp; 2082 2083 MathOp* mathOp = new MathOp(arg1, arg2); 2084 Node* operation = _gvn.transform( mathOp ); 2085 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2086 inline_math_mathExact(operation, ofcheck); 2087 return true; 2088 } 2089 2090 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2091 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2092 } 2093 2094 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2095 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2096 } 2097 2098 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2099 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2100 } 2101 2102 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2103 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2104 } 2105 2106 bool LibraryCallKit::inline_math_negateExactI() { 2107 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2108 } 2109 2110 bool LibraryCallKit::inline_math_negateExactL() { 2111 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2112 } 2113 2114 bool LibraryCallKit::inline_math_multiplyExactI() { 2115 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2116 } 2117 2118 bool LibraryCallKit::inline_math_multiplyExactL() { 2119 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2120 } 2121 2122 Node* 2123 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2124 // These are the candidate return value: 2125 Node* xvalue = x0; 2126 Node* yvalue = y0; 2127 2128 if (xvalue == yvalue) { 2129 return xvalue; 2130 } 2131 2132 bool want_max = (id == vmIntrinsics::_max); 2133 2134 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2135 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2136 if (txvalue == NULL || tyvalue == NULL) return top(); 2137 // This is not really necessary, but it is consistent with a 2138 // hypothetical MaxINode::Value method: 2139 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2140 2141 // %%% This folding logic should (ideally) be in a different place. 2142 // Some should be inside IfNode, and there to be a more reliable 2143 // transformation of ?: style patterns into cmoves. We also want 2144 // more powerful optimizations around cmove and min/max. 2145 2146 // Try to find a dominating comparison of these guys. 2147 // It can simplify the index computation for Arrays.copyOf 2148 // and similar uses of System.arraycopy. 2149 // First, compute the normalized version of CmpI(x, y). 2150 int cmp_op = Op_CmpI; 2151 Node* xkey = xvalue; 2152 Node* ykey = yvalue; 2153 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2154 if (ideal_cmpxy->is_Cmp()) { 2155 // E.g., if we have CmpI(length - offset, count), 2156 // it might idealize to CmpI(length, count + offset) 2157 cmp_op = ideal_cmpxy->Opcode(); 2158 xkey = ideal_cmpxy->in(1); 2159 ykey = ideal_cmpxy->in(2); 2160 } 2161 2162 // Start by locating any relevant comparisons. 2163 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2164 Node* cmpxy = NULL; 2165 Node* cmpyx = NULL; 2166 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2167 Node* cmp = start_from->fast_out(k); 2168 if (cmp->outcnt() > 0 && // must have prior uses 2169 cmp->in(0) == NULL && // must be context-independent 2170 cmp->Opcode() == cmp_op) { // right kind of compare 2171 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2172 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2173 } 2174 } 2175 2176 const int NCMPS = 2; 2177 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2178 int cmpn; 2179 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2180 if (cmps[cmpn] != NULL) break; // find a result 2181 } 2182 if (cmpn < NCMPS) { 2183 // Look for a dominating test that tells us the min and max. 2184 int depth = 0; // Limit search depth for speed 2185 Node* dom = control(); 2186 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2187 if (++depth >= 100) break; 2188 Node* ifproj = dom; 2189 if (!ifproj->is_Proj()) continue; 2190 Node* iff = ifproj->in(0); 2191 if (!iff->is_If()) continue; 2192 Node* bol = iff->in(1); 2193 if (!bol->is_Bool()) continue; 2194 Node* cmp = bol->in(1); 2195 if (cmp == NULL) continue; 2196 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2197 if (cmps[cmpn] == cmp) break; 2198 if (cmpn == NCMPS) continue; 2199 BoolTest::mask btest = bol->as_Bool()->_test._test; 2200 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2201 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2202 // At this point, we know that 'x btest y' is true. 2203 switch (btest) { 2204 case BoolTest::eq: 2205 // They are proven equal, so we can collapse the min/max. 2206 // Either value is the answer. Choose the simpler. 2207 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2208 return yvalue; 2209 return xvalue; 2210 case BoolTest::lt: // x < y 2211 case BoolTest::le: // x <= y 2212 return (want_max ? yvalue : xvalue); 2213 case BoolTest::gt: // x > y 2214 case BoolTest::ge: // x >= y 2215 return (want_max ? xvalue : yvalue); 2216 } 2217 } 2218 } 2219 2220 // We failed to find a dominating test. 2221 // Let's pick a test that might GVN with prior tests. 2222 Node* best_bol = NULL; 2223 BoolTest::mask best_btest = BoolTest::illegal; 2224 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2225 Node* cmp = cmps[cmpn]; 2226 if (cmp == NULL) continue; 2227 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2228 Node* bol = cmp->fast_out(j); 2229 if (!bol->is_Bool()) continue; 2230 BoolTest::mask btest = bol->as_Bool()->_test._test; 2231 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2232 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2233 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2234 best_bol = bol->as_Bool(); 2235 best_btest = btest; 2236 } 2237 } 2238 } 2239 2240 Node* answer_if_true = NULL; 2241 Node* answer_if_false = NULL; 2242 switch (best_btest) { 2243 default: 2244 if (cmpxy == NULL) 2245 cmpxy = ideal_cmpxy; 2246 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2247 // and fall through: 2248 case BoolTest::lt: // x < y 2249 case BoolTest::le: // x <= y 2250 answer_if_true = (want_max ? yvalue : xvalue); 2251 answer_if_false = (want_max ? xvalue : yvalue); 2252 break; 2253 case BoolTest::gt: // x > y 2254 case BoolTest::ge: // x >= y 2255 answer_if_true = (want_max ? xvalue : yvalue); 2256 answer_if_false = (want_max ? yvalue : xvalue); 2257 break; 2258 } 2259 2260 jint hi, lo; 2261 if (want_max) { 2262 // We can sharpen the minimum. 2263 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2264 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2265 } else { 2266 // We can sharpen the maximum. 2267 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2268 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2269 } 2270 2271 // Use a flow-free graph structure, to avoid creating excess control edges 2272 // which could hinder other optimizations. 2273 // Since Math.min/max is often used with arraycopy, we want 2274 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2275 Node* cmov = CMoveNode::make(NULL, best_bol, 2276 answer_if_false, answer_if_true, 2277 TypeInt::make(lo, hi, widen)); 2278 2279 return _gvn.transform(cmov); 2280 2281 /* 2282 // This is not as desirable as it may seem, since Min and Max 2283 // nodes do not have a full set of optimizations. 2284 // And they would interfere, anyway, with 'if' optimizations 2285 // and with CMoveI canonical forms. 2286 switch (id) { 2287 case vmIntrinsics::_min: 2288 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2289 case vmIntrinsics::_max: 2290 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2291 default: 2292 ShouldNotReachHere(); 2293 } 2294 */ 2295 } 2296 2297 inline int 2298 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2299 const TypePtr* base_type = TypePtr::NULL_PTR; 2300 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2301 if (base_type == NULL) { 2302 // Unknown type. 2303 return Type::AnyPtr; 2304 } else if (base_type == TypePtr::NULL_PTR) { 2305 // Since this is a NULL+long form, we have to switch to a rawptr. 2306 base = _gvn.transform(new CastX2PNode(offset)); 2307 offset = MakeConX(0); 2308 return Type::RawPtr; 2309 } else if (base_type->base() == Type::RawPtr) { 2310 return Type::RawPtr; 2311 } else if (base_type->isa_oopptr()) { 2312 // Base is never null => always a heap address. 2313 if (base_type->ptr() == TypePtr::NotNull) { 2314 return Type::OopPtr; 2315 } 2316 // Offset is small => always a heap address. 2317 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2318 if (offset_type != NULL && 2319 base_type->offset() == 0 && // (should always be?) 2320 offset_type->_lo >= 0 && 2321 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2322 return Type::OopPtr; 2323 } 2324 // Otherwise, it might either be oop+off or NULL+addr. 2325 return Type::AnyPtr; 2326 } else { 2327 // No information: 2328 return Type::AnyPtr; 2329 } 2330 } 2331 2332 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2333 int kind = classify_unsafe_addr(base, offset); 2334 if (kind == Type::RawPtr) { 2335 return basic_plus_adr(top(), base, offset); 2336 } else { 2337 return basic_plus_adr(base, offset); 2338 } 2339 } 2340 2341 //--------------------------inline_number_methods----------------------------- 2342 // inline int Integer.numberOfLeadingZeros(int) 2343 // inline int Long.numberOfLeadingZeros(long) 2344 // 2345 // inline int Integer.numberOfTrailingZeros(int) 2346 // inline int Long.numberOfTrailingZeros(long) 2347 // 2348 // inline int Integer.bitCount(int) 2349 // inline int Long.bitCount(long) 2350 // 2351 // inline char Character.reverseBytes(char) 2352 // inline short Short.reverseBytes(short) 2353 // inline int Integer.reverseBytes(int) 2354 // inline long Long.reverseBytes(long) 2355 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2356 Node* arg = argument(0); 2357 Node* n; 2358 switch (id) { 2359 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2360 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2361 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2362 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2363 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2364 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2365 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2366 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2367 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2368 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2369 default: fatal_unexpected_iid(id); break; 2370 } 2371 set_result(_gvn.transform(n)); 2372 return true; 2373 } 2374 2375 //----------------------------inline_unsafe_access---------------------------- 2376 2377 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2378 2379 // Helper that guards and inserts a pre-barrier. 2380 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2381 Node* pre_val, bool need_mem_bar) { 2382 // We could be accessing the referent field of a reference object. If so, when G1 2383 // is enabled, we need to log the value in the referent field in an SATB buffer. 2384 // This routine performs some compile time filters and generates suitable 2385 // runtime filters that guard the pre-barrier code. 2386 // Also add memory barrier for non volatile load from the referent field 2387 // to prevent commoning of loads across safepoint. 2388 if (!UseG1GC && !need_mem_bar) 2389 return; 2390 2391 // Some compile time checks. 2392 2393 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2394 const TypeX* otype = offset->find_intptr_t_type(); 2395 if (otype != NULL && otype->is_con() && 2396 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2397 // Constant offset but not the reference_offset so just return 2398 return; 2399 } 2400 2401 // We only need to generate the runtime guards for instances. 2402 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2403 if (btype != NULL) { 2404 if (btype->isa_aryptr()) { 2405 // Array type so nothing to do 2406 return; 2407 } 2408 2409 const TypeInstPtr* itype = btype->isa_instptr(); 2410 if (itype != NULL) { 2411 // Can the klass of base_oop be statically determined to be 2412 // _not_ a sub-class of Reference and _not_ Object? 2413 ciKlass* klass = itype->klass(); 2414 if ( klass->is_loaded() && 2415 !klass->is_subtype_of(env()->Reference_klass()) && 2416 !env()->Object_klass()->is_subtype_of(klass)) { 2417 return; 2418 } 2419 } 2420 } 2421 2422 // The compile time filters did not reject base_oop/offset so 2423 // we need to generate the following runtime filters 2424 // 2425 // if (offset == java_lang_ref_Reference::_reference_offset) { 2426 // if (instance_of(base, java.lang.ref.Reference)) { 2427 // pre_barrier(_, pre_val, ...); 2428 // } 2429 // } 2430 2431 float likely = PROB_LIKELY( 0.999); 2432 float unlikely = PROB_UNLIKELY(0.999); 2433 2434 IdealKit ideal(this); 2435 #define __ ideal. 2436 2437 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2438 2439 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2440 // Update graphKit memory and control from IdealKit. 2441 sync_kit(ideal); 2442 2443 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2444 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2445 2446 // Update IdealKit memory and control from graphKit. 2447 __ sync_kit(this); 2448 2449 Node* one = __ ConI(1); 2450 // is_instof == 0 if base_oop == NULL 2451 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2452 2453 // Update graphKit from IdeakKit. 2454 sync_kit(ideal); 2455 2456 // Use the pre-barrier to record the value in the referent field 2457 pre_barrier(false /* do_load */, 2458 __ ctrl(), 2459 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2460 pre_val /* pre_val */, 2461 T_OBJECT); 2462 if (need_mem_bar) { 2463 // Add memory barrier to prevent commoning reads from this field 2464 // across safepoint since GC can change its value. 2465 insert_mem_bar(Op_MemBarCPUOrder); 2466 } 2467 // Update IdealKit from graphKit. 2468 __ sync_kit(this); 2469 2470 } __ end_if(); // _ref_type != ref_none 2471 } __ end_if(); // offset == referent_offset 2472 2473 // Final sync IdealKit and GraphKit. 2474 final_sync(ideal); 2475 #undef __ 2476 } 2477 2478 2479 // Interpret Unsafe.fieldOffset cookies correctly: 2480 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2481 2482 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2483 // Attempt to infer a sharper value type from the offset and base type. 2484 ciKlass* sharpened_klass = NULL; 2485 2486 // See if it is an instance field, with an object type. 2487 if (alias_type->field() != NULL) { 2488 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2489 if (alias_type->field()->type()->is_klass()) { 2490 sharpened_klass = alias_type->field()->type()->as_klass(); 2491 } 2492 } 2493 2494 // See if it is a narrow oop array. 2495 if (adr_type->isa_aryptr()) { 2496 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2497 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2498 if (elem_type != NULL) { 2499 sharpened_klass = elem_type->klass(); 2500 } 2501 } 2502 } 2503 2504 // The sharpened class might be unloaded if there is no class loader 2505 // contraint in place. 2506 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2507 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2508 2509 #ifndef PRODUCT 2510 if (C->print_intrinsics() || C->print_inlining()) { 2511 tty->print(" from base type: "); adr_type->dump(); 2512 tty->print(" sharpened value: "); tjp->dump(); 2513 } 2514 #endif 2515 // Sharpen the value type. 2516 return tjp; 2517 } 2518 return NULL; 2519 } 2520 2521 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2522 if (callee()->is_static()) return false; // caller must have the capability! 2523 2524 #ifndef PRODUCT 2525 { 2526 ResourceMark rm; 2527 // Check the signatures. 2528 ciSignature* sig = callee()->signature(); 2529 #ifdef ASSERT 2530 if (!is_store) { 2531 // Object getObject(Object base, int/long offset), etc. 2532 BasicType rtype = sig->return_type()->basic_type(); 2533 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2534 rtype = T_ADDRESS; // it is really a C void* 2535 assert(rtype == type, "getter must return the expected value"); 2536 if (!is_native_ptr) { 2537 assert(sig->count() == 2, "oop getter has 2 arguments"); 2538 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2539 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2540 } else { 2541 assert(sig->count() == 1, "native getter has 1 argument"); 2542 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2543 } 2544 } else { 2545 // void putObject(Object base, int/long offset, Object x), etc. 2546 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2547 if (!is_native_ptr) { 2548 assert(sig->count() == 3, "oop putter has 3 arguments"); 2549 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2550 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2551 } else { 2552 assert(sig->count() == 2, "native putter has 2 arguments"); 2553 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2554 } 2555 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2556 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2557 vtype = T_ADDRESS; // it is really a C void* 2558 assert(vtype == type, "putter must accept the expected value"); 2559 } 2560 #endif // ASSERT 2561 } 2562 #endif //PRODUCT 2563 2564 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2565 2566 Node* receiver = argument(0); // type: oop 2567 2568 // Build address expression. 2569 Node* adr; 2570 Node* heap_base_oop = top(); 2571 Node* offset = top(); 2572 Node* val; 2573 2574 if (!is_native_ptr) { 2575 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2576 Node* base = argument(1); // type: oop 2577 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2578 offset = argument(2); // type: long 2579 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2580 // to be plain byte offsets, which are also the same as those accepted 2581 // by oopDesc::field_base. 2582 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2583 "fieldOffset must be byte-scaled"); 2584 // 32-bit machines ignore the high half! 2585 offset = ConvL2X(offset); 2586 adr = make_unsafe_address(base, offset); 2587 heap_base_oop = base; 2588 val = is_store ? argument(4) : NULL; 2589 } else { 2590 Node* ptr = argument(1); // type: long 2591 ptr = ConvL2X(ptr); // adjust Java long to machine word 2592 adr = make_unsafe_address(NULL, ptr); 2593 val = is_store ? argument(3) : NULL; 2594 } 2595 2596 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2597 2598 // First guess at the value type. 2599 const Type *value_type = Type::get_const_basic_type(type); 2600 2601 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2602 // there was not enough information to nail it down. 2603 Compile::AliasType* alias_type = C->alias_type(adr_type); 2604 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2605 2606 // We will need memory barriers unless we can determine a unique 2607 // alias category for this reference. (Note: If for some reason 2608 // the barriers get omitted and the unsafe reference begins to "pollute" 2609 // the alias analysis of the rest of the graph, either Compile::can_alias 2610 // or Compile::must_alias will throw a diagnostic assert.) 2611 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2612 2613 // If we are reading the value of the referent field of a Reference 2614 // object (either by using Unsafe directly or through reflection) 2615 // then, if G1 is enabled, we need to record the referent in an 2616 // SATB log buffer using the pre-barrier mechanism. 2617 // Also we need to add memory barrier to prevent commoning reads 2618 // from this field across safepoint since GC can change its value. 2619 bool need_read_barrier = !is_native_ptr && !is_store && 2620 offset != top() && heap_base_oop != top(); 2621 2622 if (!is_store && type == T_OBJECT) { 2623 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2624 if (tjp != NULL) { 2625 value_type = tjp; 2626 } 2627 } 2628 2629 receiver = null_check(receiver); 2630 if (stopped()) { 2631 return true; 2632 } 2633 // Heap pointers get a null-check from the interpreter, 2634 // as a courtesy. However, this is not guaranteed by Unsafe, 2635 // and it is not possible to fully distinguish unintended nulls 2636 // from intended ones in this API. 2637 2638 if (is_volatile) { 2639 // We need to emit leading and trailing CPU membars (see below) in 2640 // addition to memory membars when is_volatile. This is a little 2641 // too strong, but avoids the need to insert per-alias-type 2642 // volatile membars (for stores; compare Parse::do_put_xxx), which 2643 // we cannot do effectively here because we probably only have a 2644 // rough approximation of type. 2645 need_mem_bar = true; 2646 // For Stores, place a memory ordering barrier now. 2647 if (is_store) { 2648 insert_mem_bar(Op_MemBarRelease); 2649 } else { 2650 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2651 insert_mem_bar(Op_MemBarVolatile); 2652 } 2653 } 2654 } 2655 2656 // Memory barrier to prevent normal and 'unsafe' accesses from 2657 // bypassing each other. Happens after null checks, so the 2658 // exception paths do not take memory state from the memory barrier, 2659 // so there's no problems making a strong assert about mixing users 2660 // of safe & unsafe memory. 2661 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2662 2663 if (!is_store) { 2664 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2665 // To be valid, unsafe loads may depend on other conditions than 2666 // the one that guards them: pin the Load node 2667 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile); 2668 // load value 2669 switch (type) { 2670 case T_BOOLEAN: 2671 case T_CHAR: 2672 case T_BYTE: 2673 case T_SHORT: 2674 case T_INT: 2675 case T_LONG: 2676 case T_FLOAT: 2677 case T_DOUBLE: 2678 break; 2679 case T_OBJECT: 2680 if (need_read_barrier) { 2681 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2682 } 2683 break; 2684 case T_ADDRESS: 2685 // Cast to an int type. 2686 p = _gvn.transform(new CastP2XNode(NULL, p)); 2687 p = ConvX2UL(p); 2688 break; 2689 default: 2690 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2691 break; 2692 } 2693 // The load node has the control of the preceding MemBarCPUOrder. All 2694 // following nodes will have the control of the MemBarCPUOrder inserted at 2695 // the end of this method. So, pushing the load onto the stack at a later 2696 // point is fine. 2697 set_result(p); 2698 } else { 2699 // place effect of store into memory 2700 switch (type) { 2701 case T_DOUBLE: 2702 val = dstore_rounding(val); 2703 break; 2704 case T_ADDRESS: 2705 // Repackage the long as a pointer. 2706 val = ConvL2X(val); 2707 val = _gvn.transform(new CastX2PNode(val)); 2708 break; 2709 } 2710 2711 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2712 if (type != T_OBJECT ) { 2713 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2714 } else { 2715 // Possibly an oop being stored to Java heap or native memory 2716 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2717 // oop to Java heap. 2718 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2719 } else { 2720 // We can't tell at compile time if we are storing in the Java heap or outside 2721 // of it. So we need to emit code to conditionally do the proper type of 2722 // store. 2723 2724 IdealKit ideal(this); 2725 #define __ ideal. 2726 // QQQ who knows what probability is here?? 2727 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2728 // Sync IdealKit and graphKit. 2729 sync_kit(ideal); 2730 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2731 // Update IdealKit memory. 2732 __ sync_kit(this); 2733 } __ else_(); { 2734 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2735 } __ end_if(); 2736 // Final sync IdealKit and GraphKit. 2737 final_sync(ideal); 2738 #undef __ 2739 } 2740 } 2741 } 2742 2743 if (is_volatile) { 2744 if (!is_store) { 2745 insert_mem_bar(Op_MemBarAcquire); 2746 } else { 2747 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2748 insert_mem_bar(Op_MemBarVolatile); 2749 } 2750 } 2751 } 2752 2753 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2754 2755 return true; 2756 } 2757 2758 //----------------------------inline_unsafe_load_store---------------------------- 2759 // This method serves a couple of different customers (depending on LoadStoreKind): 2760 // 2761 // LS_cmpxchg: 2762 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2763 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2764 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2765 // 2766 // LS_xadd: 2767 // public int getAndAddInt( Object o, long offset, int delta) 2768 // public long getAndAddLong(Object o, long offset, long delta) 2769 // 2770 // LS_xchg: 2771 // int getAndSet(Object o, long offset, int newValue) 2772 // long getAndSet(Object o, long offset, long newValue) 2773 // Object getAndSet(Object o, long offset, Object newValue) 2774 // 2775 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2776 // This basic scheme here is the same as inline_unsafe_access, but 2777 // differs in enough details that combining them would make the code 2778 // overly confusing. (This is a true fact! I originally combined 2779 // them, but even I was confused by it!) As much code/comments as 2780 // possible are retained from inline_unsafe_access though to make 2781 // the correspondences clearer. - dl 2782 2783 if (callee()->is_static()) return false; // caller must have the capability! 2784 2785 #ifndef PRODUCT 2786 BasicType rtype; 2787 { 2788 ResourceMark rm; 2789 // Check the signatures. 2790 ciSignature* sig = callee()->signature(); 2791 rtype = sig->return_type()->basic_type(); 2792 if (kind == LS_xadd || kind == LS_xchg) { 2793 // Check the signatures. 2794 #ifdef ASSERT 2795 assert(rtype == type, "get and set must return the expected type"); 2796 assert(sig->count() == 3, "get and set has 3 arguments"); 2797 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2798 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2799 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2800 #endif // ASSERT 2801 } else if (kind == LS_cmpxchg) { 2802 // Check the signatures. 2803 #ifdef ASSERT 2804 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2805 assert(sig->count() == 4, "CAS has 4 arguments"); 2806 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2807 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2808 #endif // ASSERT 2809 } else { 2810 ShouldNotReachHere(); 2811 } 2812 } 2813 #endif //PRODUCT 2814 2815 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2816 2817 // Get arguments: 2818 Node* receiver = NULL; 2819 Node* base = NULL; 2820 Node* offset = NULL; 2821 Node* oldval = NULL; 2822 Node* newval = NULL; 2823 if (kind == LS_cmpxchg) { 2824 const bool two_slot_type = type2size[type] == 2; 2825 receiver = argument(0); // type: oop 2826 base = argument(1); // type: oop 2827 offset = argument(2); // type: long 2828 oldval = argument(4); // type: oop, int, or long 2829 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2830 } else if (kind == LS_xadd || kind == LS_xchg){ 2831 receiver = argument(0); // type: oop 2832 base = argument(1); // type: oop 2833 offset = argument(2); // type: long 2834 oldval = NULL; 2835 newval = argument(4); // type: oop, int, or long 2836 } 2837 2838 // Null check receiver. 2839 receiver = null_check(receiver); 2840 if (stopped()) { 2841 return true; 2842 } 2843 2844 // Build field offset expression. 2845 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2846 // to be plain byte offsets, which are also the same as those accepted 2847 // by oopDesc::field_base. 2848 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2849 // 32-bit machines ignore the high half of long offsets 2850 offset = ConvL2X(offset); 2851 Node* adr = make_unsafe_address(base, offset); 2852 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2853 2854 // For CAS, unlike inline_unsafe_access, there seems no point in 2855 // trying to refine types. Just use the coarse types here. 2856 const Type *value_type = Type::get_const_basic_type(type); 2857 Compile::AliasType* alias_type = C->alias_type(adr_type); 2858 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2859 2860 if (kind == LS_xchg && type == T_OBJECT) { 2861 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2862 if (tjp != NULL) { 2863 value_type = tjp; 2864 } 2865 } 2866 2867 int alias_idx = C->get_alias_index(adr_type); 2868 2869 // Memory-model-wise, a LoadStore acts like a little synchronized 2870 // block, so needs barriers on each side. These don't translate 2871 // into actual barriers on most machines, but we still need rest of 2872 // compiler to respect ordering. 2873 2874 insert_mem_bar(Op_MemBarRelease); 2875 insert_mem_bar(Op_MemBarCPUOrder); 2876 2877 // 4984716: MemBars must be inserted before this 2878 // memory node in order to avoid a false 2879 // dependency which will confuse the scheduler. 2880 Node *mem = memory(alias_idx); 2881 2882 // For now, we handle only those cases that actually exist: ints, 2883 // longs, and Object. Adding others should be straightforward. 2884 Node* load_store; 2885 switch(type) { 2886 case T_INT: 2887 if (kind == LS_xadd) { 2888 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2889 } else if (kind == LS_xchg) { 2890 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2891 } else if (kind == LS_cmpxchg) { 2892 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2893 } else { 2894 ShouldNotReachHere(); 2895 } 2896 break; 2897 case T_LONG: 2898 if (kind == LS_xadd) { 2899 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2900 } else if (kind == LS_xchg) { 2901 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2902 } else if (kind == LS_cmpxchg) { 2903 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2904 } else { 2905 ShouldNotReachHere(); 2906 } 2907 break; 2908 case T_OBJECT: 2909 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2910 // could be delayed during Parse (for example, in adjust_map_after_if()). 2911 // Execute transformation here to avoid barrier generation in such case. 2912 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2913 newval = _gvn.makecon(TypePtr::NULL_PTR); 2914 2915 // Reference stores need a store barrier. 2916 if (kind == LS_xchg) { 2917 // If pre-barrier must execute before the oop store, old value will require do_load here. 2918 if (!can_move_pre_barrier()) { 2919 pre_barrier(true /* do_load*/, 2920 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2921 NULL /* pre_val*/, 2922 T_OBJECT); 2923 } // Else move pre_barrier to use load_store value, see below. 2924 } else if (kind == LS_cmpxchg) { 2925 // Same as for newval above: 2926 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2927 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2928 } 2929 // The only known value which might get overwritten is oldval. 2930 pre_barrier(false /* do_load */, 2931 control(), NULL, NULL, max_juint, NULL, NULL, 2932 oldval /* pre_val */, 2933 T_OBJECT); 2934 } else { 2935 ShouldNotReachHere(); 2936 } 2937 2938 #ifdef _LP64 2939 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2940 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2941 if (kind == LS_xchg) { 2942 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2943 newval_enc, adr_type, value_type->make_narrowoop())); 2944 } else { 2945 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2946 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2947 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2948 newval_enc, oldval_enc)); 2949 } 2950 } else 2951 #endif 2952 { 2953 if (kind == LS_xchg) { 2954 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2955 } else { 2956 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2957 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2958 } 2959 } 2960 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2961 break; 2962 default: 2963 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2964 break; 2965 } 2966 2967 // SCMemProjNodes represent the memory state of a LoadStore. Their 2968 // main role is to prevent LoadStore nodes from being optimized away 2969 // when their results aren't used. 2970 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2971 set_memory(proj, alias_idx); 2972 2973 if (type == T_OBJECT && kind == LS_xchg) { 2974 #ifdef _LP64 2975 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2976 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2977 } 2978 #endif 2979 if (can_move_pre_barrier()) { 2980 // Don't need to load pre_val. The old value is returned by load_store. 2981 // The pre_barrier can execute after the xchg as long as no safepoint 2982 // gets inserted between them. 2983 pre_barrier(false /* do_load */, 2984 control(), NULL, NULL, max_juint, NULL, NULL, 2985 load_store /* pre_val */, 2986 T_OBJECT); 2987 } 2988 } 2989 2990 // Add the trailing membar surrounding the access 2991 insert_mem_bar(Op_MemBarCPUOrder); 2992 insert_mem_bar(Op_MemBarAcquire); 2993 2994 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2995 set_result(load_store); 2996 return true; 2997 } 2998 2999 //----------------------------inline_unsafe_ordered_store---------------------- 3000 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3001 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3002 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3003 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3004 // This is another variant of inline_unsafe_access, differing in 3005 // that it always issues store-store ("release") barrier and ensures 3006 // store-atomicity (which only matters for "long"). 3007 3008 if (callee()->is_static()) return false; // caller must have the capability! 3009 3010 #ifndef PRODUCT 3011 { 3012 ResourceMark rm; 3013 // Check the signatures. 3014 ciSignature* sig = callee()->signature(); 3015 #ifdef ASSERT 3016 BasicType rtype = sig->return_type()->basic_type(); 3017 assert(rtype == T_VOID, "must return void"); 3018 assert(sig->count() == 3, "has 3 arguments"); 3019 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3020 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3021 #endif // ASSERT 3022 } 3023 #endif //PRODUCT 3024 3025 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3026 3027 // Get arguments: 3028 Node* receiver = argument(0); // type: oop 3029 Node* base = argument(1); // type: oop 3030 Node* offset = argument(2); // type: long 3031 Node* val = argument(4); // type: oop, int, or long 3032 3033 // Null check receiver. 3034 receiver = null_check(receiver); 3035 if (stopped()) { 3036 return true; 3037 } 3038 3039 // Build field offset expression. 3040 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3041 // 32-bit machines ignore the high half of long offsets 3042 offset = ConvL2X(offset); 3043 Node* adr = make_unsafe_address(base, offset); 3044 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3045 const Type *value_type = Type::get_const_basic_type(type); 3046 Compile::AliasType* alias_type = C->alias_type(adr_type); 3047 3048 insert_mem_bar(Op_MemBarRelease); 3049 insert_mem_bar(Op_MemBarCPUOrder); 3050 // Ensure that the store is atomic for longs: 3051 const bool require_atomic_access = true; 3052 Node* store; 3053 if (type == T_OBJECT) // reference stores need a store barrier. 3054 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3055 else { 3056 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3057 } 3058 insert_mem_bar(Op_MemBarCPUOrder); 3059 return true; 3060 } 3061 3062 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3063 // Regardless of form, don't allow previous ld/st to move down, 3064 // then issue acquire, release, or volatile mem_bar. 3065 insert_mem_bar(Op_MemBarCPUOrder); 3066 switch(id) { 3067 case vmIntrinsics::_loadFence: 3068 insert_mem_bar(Op_LoadFence); 3069 return true; 3070 case vmIntrinsics::_storeFence: 3071 insert_mem_bar(Op_StoreFence); 3072 return true; 3073 case vmIntrinsics::_fullFence: 3074 insert_mem_bar(Op_MemBarVolatile); 3075 return true; 3076 default: 3077 fatal_unexpected_iid(id); 3078 return false; 3079 } 3080 } 3081 3082 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3083 if (!kls->is_Con()) { 3084 return true; 3085 } 3086 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3087 if (klsptr == NULL) { 3088 return true; 3089 } 3090 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3091 // don't need a guard for a klass that is already initialized 3092 return !ik->is_initialized(); 3093 } 3094 3095 //----------------------------inline_unsafe_allocate--------------------------- 3096 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3097 bool LibraryCallKit::inline_unsafe_allocate() { 3098 if (callee()->is_static()) return false; // caller must have the capability! 3099 3100 null_check_receiver(); // null-check, then ignore 3101 Node* cls = null_check(argument(1)); 3102 if (stopped()) return true; 3103 3104 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3105 kls = null_check(kls); 3106 if (stopped()) return true; // argument was like int.class 3107 3108 Node* test = NULL; 3109 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3110 // Note: The argument might still be an illegal value like 3111 // Serializable.class or Object[].class. The runtime will handle it. 3112 // But we must make an explicit check for initialization. 3113 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3114 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3115 // can generate code to load it as unsigned byte. 3116 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3117 Node* bits = intcon(InstanceKlass::fully_initialized); 3118 test = _gvn.transform(new SubINode(inst, bits)); 3119 // The 'test' is non-zero if we need to take a slow path. 3120 } 3121 3122 Node* obj = new_instance(kls, test); 3123 set_result(obj); 3124 return true; 3125 } 3126 3127 #ifdef TRACE_HAVE_INTRINSICS 3128 /* 3129 * oop -> myklass 3130 * myklass->trace_id |= USED 3131 * return myklass->trace_id & ~0x3 3132 */ 3133 bool LibraryCallKit::inline_native_classID() { 3134 null_check_receiver(); // null-check, then ignore 3135 Node* cls = null_check(argument(1), T_OBJECT); 3136 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3137 kls = null_check(kls, T_OBJECT); 3138 ByteSize offset = TRACE_ID_OFFSET; 3139 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3140 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3141 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3142 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3143 Node* clsused = longcon(0x01l); // set the class bit 3144 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3145 3146 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3147 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3148 set_result(andl); 3149 return true; 3150 } 3151 3152 bool LibraryCallKit::inline_native_threadID() { 3153 Node* tls_ptr = NULL; 3154 Node* cur_thr = generate_current_thread(tls_ptr); 3155 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3156 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3157 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3158 3159 Node* threadid = NULL; 3160 size_t thread_id_size = OSThread::thread_id_size(); 3161 if (thread_id_size == (size_t) BytesPerLong) { 3162 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3163 } else if (thread_id_size == (size_t) BytesPerInt) { 3164 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3165 } else { 3166 ShouldNotReachHere(); 3167 } 3168 set_result(threadid); 3169 return true; 3170 } 3171 #endif 3172 3173 //------------------------inline_native_time_funcs-------------- 3174 // inline code for System.currentTimeMillis() and System.nanoTime() 3175 // these have the same type and signature 3176 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3177 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3178 const TypePtr* no_memory_effects = NULL; 3179 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3180 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3181 #ifdef ASSERT 3182 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3183 assert(value_top == top(), "second value must be top"); 3184 #endif 3185 set_result(value); 3186 return true; 3187 } 3188 3189 //------------------------inline_native_currentThread------------------ 3190 bool LibraryCallKit::inline_native_currentThread() { 3191 Node* junk = NULL; 3192 set_result(generate_current_thread(junk)); 3193 return true; 3194 } 3195 3196 //------------------------inline_native_isInterrupted------------------ 3197 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3198 bool LibraryCallKit::inline_native_isInterrupted() { 3199 // Add a fast path to t.isInterrupted(clear_int): 3200 // (t == Thread.current() && 3201 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3202 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3203 // So, in the common case that the interrupt bit is false, 3204 // we avoid making a call into the VM. Even if the interrupt bit 3205 // is true, if the clear_int argument is false, we avoid the VM call. 3206 // However, if the receiver is not currentThread, we must call the VM, 3207 // because there must be some locking done around the operation. 3208 3209 // We only go to the fast case code if we pass two guards. 3210 // Paths which do not pass are accumulated in the slow_region. 3211 3212 enum { 3213 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3214 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3215 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3216 PATH_LIMIT 3217 }; 3218 3219 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3220 // out of the function. 3221 insert_mem_bar(Op_MemBarCPUOrder); 3222 3223 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3224 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3225 3226 RegionNode* slow_region = new RegionNode(1); 3227 record_for_igvn(slow_region); 3228 3229 // (a) Receiving thread must be the current thread. 3230 Node* rec_thr = argument(0); 3231 Node* tls_ptr = NULL; 3232 Node* cur_thr = generate_current_thread(tls_ptr); 3233 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3234 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3235 3236 generate_slow_guard(bol_thr, slow_region); 3237 3238 // (b) Interrupt bit on TLS must be false. 3239 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3240 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3241 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3242 3243 // Set the control input on the field _interrupted read to prevent it floating up. 3244 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3245 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3246 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3247 3248 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3249 3250 // First fast path: if (!TLS._interrupted) return false; 3251 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3252 result_rgn->init_req(no_int_result_path, false_bit); 3253 result_val->init_req(no_int_result_path, intcon(0)); 3254 3255 // drop through to next case 3256 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3257 3258 #ifndef TARGET_OS_FAMILY_windows 3259 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3260 Node* clr_arg = argument(1); 3261 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3262 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3263 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3264 3265 // Second fast path: ... else if (!clear_int) return true; 3266 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3267 result_rgn->init_req(no_clear_result_path, false_arg); 3268 result_val->init_req(no_clear_result_path, intcon(1)); 3269 3270 // drop through to next case 3271 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3272 #else 3273 // To return true on Windows you must read the _interrupted field 3274 // and check the the event state i.e. take the slow path. 3275 #endif // TARGET_OS_FAMILY_windows 3276 3277 // (d) Otherwise, go to the slow path. 3278 slow_region->add_req(control()); 3279 set_control( _gvn.transform(slow_region)); 3280 3281 if (stopped()) { 3282 // There is no slow path. 3283 result_rgn->init_req(slow_result_path, top()); 3284 result_val->init_req(slow_result_path, top()); 3285 } else { 3286 // non-virtual because it is a private non-static 3287 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3288 3289 Node* slow_val = set_results_for_java_call(slow_call); 3290 // this->control() comes from set_results_for_java_call 3291 3292 Node* fast_io = slow_call->in(TypeFunc::I_O); 3293 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3294 3295 // These two phis are pre-filled with copies of of the fast IO and Memory 3296 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3297 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3298 3299 result_rgn->init_req(slow_result_path, control()); 3300 result_io ->init_req(slow_result_path, i_o()); 3301 result_mem->init_req(slow_result_path, reset_memory()); 3302 result_val->init_req(slow_result_path, slow_val); 3303 3304 set_all_memory(_gvn.transform(result_mem)); 3305 set_i_o( _gvn.transform(result_io)); 3306 } 3307 3308 C->set_has_split_ifs(true); // Has chance for split-if optimization 3309 set_result(result_rgn, result_val); 3310 return true; 3311 } 3312 3313 //---------------------------load_mirror_from_klass---------------------------- 3314 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3315 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3316 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3317 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3318 } 3319 3320 //-----------------------load_klass_from_mirror_common------------------------- 3321 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3322 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3323 // and branch to the given path on the region. 3324 // If never_see_null, take an uncommon trap on null, so we can optimistically 3325 // compile for the non-null case. 3326 // If the region is NULL, force never_see_null = true. 3327 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3328 bool never_see_null, 3329 RegionNode* region, 3330 int null_path, 3331 int offset) { 3332 if (region == NULL) never_see_null = true; 3333 Node* p = basic_plus_adr(mirror, offset); 3334 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3335 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3336 Node* null_ctl = top(); 3337 kls = null_check_oop(kls, &null_ctl, never_see_null); 3338 if (region != NULL) { 3339 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3340 region->init_req(null_path, null_ctl); 3341 } else { 3342 assert(null_ctl == top(), "no loose ends"); 3343 } 3344 return kls; 3345 } 3346 3347 //--------------------(inline_native_Class_query helpers)--------------------- 3348 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3349 // Fall through if (mods & mask) == bits, take the guard otherwise. 3350 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3351 // Branch around if the given klass has the given modifier bit set. 3352 // Like generate_guard, adds a new path onto the region. 3353 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3354 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3355 Node* mask = intcon(modifier_mask); 3356 Node* bits = intcon(modifier_bits); 3357 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3358 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3359 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3360 return generate_fair_guard(bol, region); 3361 } 3362 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3363 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3364 } 3365 3366 //-------------------------inline_native_Class_query------------------- 3367 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3368 const Type* return_type = TypeInt::BOOL; 3369 Node* prim_return_value = top(); // what happens if it's a primitive class? 3370 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3371 bool expect_prim = false; // most of these guys expect to work on refs 3372 3373 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3374 3375 Node* mirror = argument(0); 3376 Node* obj = top(); 3377 3378 switch (id) { 3379 case vmIntrinsics::_isInstance: 3380 // nothing is an instance of a primitive type 3381 prim_return_value = intcon(0); 3382 obj = argument(1); 3383 break; 3384 case vmIntrinsics::_getModifiers: 3385 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3386 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3387 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3388 break; 3389 case vmIntrinsics::_isInterface: 3390 prim_return_value = intcon(0); 3391 break; 3392 case vmIntrinsics::_isArray: 3393 prim_return_value = intcon(0); 3394 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3395 break; 3396 case vmIntrinsics::_isPrimitive: 3397 prim_return_value = intcon(1); 3398 expect_prim = true; // obviously 3399 break; 3400 case vmIntrinsics::_getSuperclass: 3401 prim_return_value = null(); 3402 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3403 break; 3404 case vmIntrinsics::_getClassAccessFlags: 3405 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3406 return_type = TypeInt::INT; // not bool! 6297094 3407 break; 3408 default: 3409 fatal_unexpected_iid(id); 3410 break; 3411 } 3412 3413 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3414 if (mirror_con == NULL) return false; // cannot happen? 3415 3416 #ifndef PRODUCT 3417 if (C->print_intrinsics() || C->print_inlining()) { 3418 ciType* k = mirror_con->java_mirror_type(); 3419 if (k) { 3420 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3421 k->print_name(); 3422 tty->cr(); 3423 } 3424 } 3425 #endif 3426 3427 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3428 RegionNode* region = new RegionNode(PATH_LIMIT); 3429 record_for_igvn(region); 3430 PhiNode* phi = new PhiNode(region, return_type); 3431 3432 // The mirror will never be null of Reflection.getClassAccessFlags, however 3433 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3434 // if it is. See bug 4774291. 3435 3436 // For Reflection.getClassAccessFlags(), the null check occurs in 3437 // the wrong place; see inline_unsafe_access(), above, for a similar 3438 // situation. 3439 mirror = null_check(mirror); 3440 // If mirror or obj is dead, only null-path is taken. 3441 if (stopped()) return true; 3442 3443 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3444 3445 // Now load the mirror's klass metaobject, and null-check it. 3446 // Side-effects region with the control path if the klass is null. 3447 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3448 // If kls is null, we have a primitive mirror. 3449 phi->init_req(_prim_path, prim_return_value); 3450 if (stopped()) { set_result(region, phi); return true; } 3451 bool safe_for_replace = (region->in(_prim_path) == top()); 3452 3453 Node* p; // handy temp 3454 Node* null_ctl; 3455 3456 // Now that we have the non-null klass, we can perform the real query. 3457 // For constant classes, the query will constant-fold in LoadNode::Value. 3458 Node* query_value = top(); 3459 switch (id) { 3460 case vmIntrinsics::_isInstance: 3461 // nothing is an instance of a primitive type 3462 query_value = gen_instanceof(obj, kls, safe_for_replace); 3463 break; 3464 3465 case vmIntrinsics::_getModifiers: 3466 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3467 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3468 break; 3469 3470 case vmIntrinsics::_isInterface: 3471 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3472 if (generate_interface_guard(kls, region) != NULL) 3473 // A guard was added. If the guard is taken, it was an interface. 3474 phi->add_req(intcon(1)); 3475 // If we fall through, it's a plain class. 3476 query_value = intcon(0); 3477 break; 3478 3479 case vmIntrinsics::_isArray: 3480 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3481 if (generate_array_guard(kls, region) != NULL) 3482 // A guard was added. If the guard is taken, it was an array. 3483 phi->add_req(intcon(1)); 3484 // If we fall through, it's a plain class. 3485 query_value = intcon(0); 3486 break; 3487 3488 case vmIntrinsics::_isPrimitive: 3489 query_value = intcon(0); // "normal" path produces false 3490 break; 3491 3492 case vmIntrinsics::_getSuperclass: 3493 // The rules here are somewhat unfortunate, but we can still do better 3494 // with random logic than with a JNI call. 3495 // Interfaces store null or Object as _super, but must report null. 3496 // Arrays store an intermediate super as _super, but must report Object. 3497 // Other types can report the actual _super. 3498 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3499 if (generate_interface_guard(kls, region) != NULL) 3500 // A guard was added. If the guard is taken, it was an interface. 3501 phi->add_req(null()); 3502 if (generate_array_guard(kls, region) != NULL) 3503 // A guard was added. If the guard is taken, it was an array. 3504 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3505 // If we fall through, it's a plain class. Get its _super. 3506 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3507 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3508 null_ctl = top(); 3509 kls = null_check_oop(kls, &null_ctl); 3510 if (null_ctl != top()) { 3511 // If the guard is taken, Object.superClass is null (both klass and mirror). 3512 region->add_req(null_ctl); 3513 phi ->add_req(null()); 3514 } 3515 if (!stopped()) { 3516 query_value = load_mirror_from_klass(kls); 3517 } 3518 break; 3519 3520 case vmIntrinsics::_getClassAccessFlags: 3521 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3522 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3523 break; 3524 3525 default: 3526 fatal_unexpected_iid(id); 3527 break; 3528 } 3529 3530 // Fall-through is the normal case of a query to a real class. 3531 phi->init_req(1, query_value); 3532 region->init_req(1, control()); 3533 3534 C->set_has_split_ifs(true); // Has chance for split-if optimization 3535 set_result(region, phi); 3536 return true; 3537 } 3538 3539 //-------------------------inline_Class_cast------------------- 3540 bool LibraryCallKit::inline_Class_cast() { 3541 Node* mirror = argument(0); // Class 3542 Node* obj = argument(1); 3543 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3544 if (mirror_con == NULL) { 3545 return false; // dead path (mirror->is_top()). 3546 } 3547 if (obj == NULL || obj->is_top()) { 3548 return false; // dead path 3549 } 3550 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3551 3552 // First, see if Class.cast() can be folded statically. 3553 // java_mirror_type() returns non-null for compile-time Class constants. 3554 ciType* tm = mirror_con->java_mirror_type(); 3555 if (tm != NULL && tm->is_klass() && 3556 tp != NULL && tp->klass() != NULL) { 3557 if (!tp->klass()->is_loaded()) { 3558 // Don't use intrinsic when class is not loaded. 3559 return false; 3560 } else { 3561 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3562 if (static_res == Compile::SSC_always_true) { 3563 // isInstance() is true - fold the code. 3564 set_result(obj); 3565 return true; 3566 } else if (static_res == Compile::SSC_always_false) { 3567 // Don't use intrinsic, have to throw ClassCastException. 3568 // If the reference is null, the non-intrinsic bytecode will 3569 // be optimized appropriately. 3570 return false; 3571 } 3572 } 3573 } 3574 3575 // Bailout intrinsic and do normal inlining if exception path is frequent. 3576 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3577 return false; 3578 } 3579 3580 // Generate dynamic checks. 3581 // Class.cast() is java implementation of _checkcast bytecode. 3582 // Do checkcast (Parse::do_checkcast()) optimizations here. 3583 3584 mirror = null_check(mirror); 3585 // If mirror is dead, only null-path is taken. 3586 if (stopped()) { 3587 return true; 3588 } 3589 3590 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3591 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3592 RegionNode* region = new RegionNode(PATH_LIMIT); 3593 record_for_igvn(region); 3594 3595 // Now load the mirror's klass metaobject, and null-check it. 3596 // If kls is null, we have a primitive mirror and 3597 // nothing is an instance of a primitive type. 3598 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3599 3600 Node* res = top(); 3601 if (!stopped()) { 3602 Node* bad_type_ctrl = top(); 3603 // Do checkcast optimizations. 3604 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3605 region->init_req(_bad_type_path, bad_type_ctrl); 3606 } 3607 if (region->in(_prim_path) != top() || 3608 region->in(_bad_type_path) != top()) { 3609 // Let Interpreter throw ClassCastException. 3610 PreserveJVMState pjvms(this); 3611 set_control(_gvn.transform(region)); 3612 uncommon_trap(Deoptimization::Reason_intrinsic, 3613 Deoptimization::Action_maybe_recompile); 3614 } 3615 if (!stopped()) { 3616 set_result(res); 3617 } 3618 return true; 3619 } 3620 3621 3622 //--------------------------inline_native_subtype_check------------------------ 3623 // This intrinsic takes the JNI calls out of the heart of 3624 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3625 bool LibraryCallKit::inline_native_subtype_check() { 3626 // Pull both arguments off the stack. 3627 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3628 args[0] = argument(0); 3629 args[1] = argument(1); 3630 Node* klasses[2]; // corresponding Klasses: superk, subk 3631 klasses[0] = klasses[1] = top(); 3632 3633 enum { 3634 // A full decision tree on {superc is prim, subc is prim}: 3635 _prim_0_path = 1, // {P,N} => false 3636 // {P,P} & superc!=subc => false 3637 _prim_same_path, // {P,P} & superc==subc => true 3638 _prim_1_path, // {N,P} => false 3639 _ref_subtype_path, // {N,N} & subtype check wins => true 3640 _both_ref_path, // {N,N} & subtype check loses => false 3641 PATH_LIMIT 3642 }; 3643 3644 RegionNode* region = new RegionNode(PATH_LIMIT); 3645 Node* phi = new PhiNode(region, TypeInt::BOOL); 3646 record_for_igvn(region); 3647 3648 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3649 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3650 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3651 3652 // First null-check both mirrors and load each mirror's klass metaobject. 3653 int which_arg; 3654 for (which_arg = 0; which_arg <= 1; which_arg++) { 3655 Node* arg = args[which_arg]; 3656 arg = null_check(arg); 3657 if (stopped()) break; 3658 args[which_arg] = arg; 3659 3660 Node* p = basic_plus_adr(arg, class_klass_offset); 3661 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3662 klasses[which_arg] = _gvn.transform(kls); 3663 } 3664 3665 // Having loaded both klasses, test each for null. 3666 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3667 for (which_arg = 0; which_arg <= 1; which_arg++) { 3668 Node* kls = klasses[which_arg]; 3669 Node* null_ctl = top(); 3670 kls = null_check_oop(kls, &null_ctl, never_see_null); 3671 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3672 region->init_req(prim_path, null_ctl); 3673 if (stopped()) break; 3674 klasses[which_arg] = kls; 3675 } 3676 3677 if (!stopped()) { 3678 // now we have two reference types, in klasses[0..1] 3679 Node* subk = klasses[1]; // the argument to isAssignableFrom 3680 Node* superk = klasses[0]; // the receiver 3681 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3682 // now we have a successful reference subtype check 3683 region->set_req(_ref_subtype_path, control()); 3684 } 3685 3686 // If both operands are primitive (both klasses null), then 3687 // we must return true when they are identical primitives. 3688 // It is convenient to test this after the first null klass check. 3689 set_control(region->in(_prim_0_path)); // go back to first null check 3690 if (!stopped()) { 3691 // Since superc is primitive, make a guard for the superc==subc case. 3692 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3693 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3694 generate_guard(bol_eq, region, PROB_FAIR); 3695 if (region->req() == PATH_LIMIT+1) { 3696 // A guard was added. If the added guard is taken, superc==subc. 3697 region->swap_edges(PATH_LIMIT, _prim_same_path); 3698 region->del_req(PATH_LIMIT); 3699 } 3700 region->set_req(_prim_0_path, control()); // Not equal after all. 3701 } 3702 3703 // these are the only paths that produce 'true': 3704 phi->set_req(_prim_same_path, intcon(1)); 3705 phi->set_req(_ref_subtype_path, intcon(1)); 3706 3707 // pull together the cases: 3708 assert(region->req() == PATH_LIMIT, "sane region"); 3709 for (uint i = 1; i < region->req(); i++) { 3710 Node* ctl = region->in(i); 3711 if (ctl == NULL || ctl == top()) { 3712 region->set_req(i, top()); 3713 phi ->set_req(i, top()); 3714 } else if (phi->in(i) == NULL) { 3715 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3716 } 3717 } 3718 3719 set_control(_gvn.transform(region)); 3720 set_result(_gvn.transform(phi)); 3721 return true; 3722 } 3723 3724 //---------------------generate_array_guard_common------------------------ 3725 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3726 bool obj_array, bool not_array) { 3727 3728 if (stopped()) { 3729 return NULL; 3730 } 3731 3732 // If obj_array/non_array==false/false: 3733 // Branch around if the given klass is in fact an array (either obj or prim). 3734 // If obj_array/non_array==false/true: 3735 // Branch around if the given klass is not an array klass of any kind. 3736 // If obj_array/non_array==true/true: 3737 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3738 // If obj_array/non_array==true/false: 3739 // Branch around if the kls is an oop array (Object[] or subtype) 3740 // 3741 // Like generate_guard, adds a new path onto the region. 3742 jint layout_con = 0; 3743 Node* layout_val = get_layout_helper(kls, layout_con); 3744 if (layout_val == NULL) { 3745 bool query = (obj_array 3746 ? Klass::layout_helper_is_objArray(layout_con) 3747 : Klass::layout_helper_is_array(layout_con)); 3748 if (query == not_array) { 3749 return NULL; // never a branch 3750 } else { // always a branch 3751 Node* always_branch = control(); 3752 if (region != NULL) 3753 region->add_req(always_branch); 3754 set_control(top()); 3755 return always_branch; 3756 } 3757 } 3758 // Now test the correct condition. 3759 jint nval = (obj_array 3760 ? ((jint)Klass::_lh_array_tag_type_value 3761 << Klass::_lh_array_tag_shift) 3762 : Klass::_lh_neutral_value); 3763 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3764 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3765 // invert the test if we are looking for a non-array 3766 if (not_array) btest = BoolTest(btest).negate(); 3767 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3768 return generate_fair_guard(bol, region); 3769 } 3770 3771 3772 //-----------------------inline_native_newArray-------------------------- 3773 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3774 bool LibraryCallKit::inline_native_newArray() { 3775 Node* mirror = argument(0); 3776 Node* count_val = argument(1); 3777 3778 mirror = null_check(mirror); 3779 // If mirror or obj is dead, only null-path is taken. 3780 if (stopped()) return true; 3781 3782 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3783 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3784 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3785 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3786 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3787 3788 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3789 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3790 result_reg, _slow_path); 3791 Node* normal_ctl = control(); 3792 Node* no_array_ctl = result_reg->in(_slow_path); 3793 3794 // Generate code for the slow case. We make a call to newArray(). 3795 set_control(no_array_ctl); 3796 if (!stopped()) { 3797 // Either the input type is void.class, or else the 3798 // array klass has not yet been cached. Either the 3799 // ensuing call will throw an exception, or else it 3800 // will cache the array klass for next time. 3801 PreserveJVMState pjvms(this); 3802 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3803 Node* slow_result = set_results_for_java_call(slow_call); 3804 // this->control() comes from set_results_for_java_call 3805 result_reg->set_req(_slow_path, control()); 3806 result_val->set_req(_slow_path, slow_result); 3807 result_io ->set_req(_slow_path, i_o()); 3808 result_mem->set_req(_slow_path, reset_memory()); 3809 } 3810 3811 set_control(normal_ctl); 3812 if (!stopped()) { 3813 // Normal case: The array type has been cached in the java.lang.Class. 3814 // The following call works fine even if the array type is polymorphic. 3815 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3816 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3817 result_reg->init_req(_normal_path, control()); 3818 result_val->init_req(_normal_path, obj); 3819 result_io ->init_req(_normal_path, i_o()); 3820 result_mem->init_req(_normal_path, reset_memory()); 3821 } 3822 3823 // Return the combined state. 3824 set_i_o( _gvn.transform(result_io) ); 3825 set_all_memory( _gvn.transform(result_mem)); 3826 3827 C->set_has_split_ifs(true); // Has chance for split-if optimization 3828 set_result(result_reg, result_val); 3829 return true; 3830 } 3831 3832 //----------------------inline_native_getLength-------------------------- 3833 // public static native int java.lang.reflect.Array.getLength(Object array); 3834 bool LibraryCallKit::inline_native_getLength() { 3835 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3836 3837 Node* array = null_check(argument(0)); 3838 // If array is dead, only null-path is taken. 3839 if (stopped()) return true; 3840 3841 // Deoptimize if it is a non-array. 3842 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3843 3844 if (non_array != NULL) { 3845 PreserveJVMState pjvms(this); 3846 set_control(non_array); 3847 uncommon_trap(Deoptimization::Reason_intrinsic, 3848 Deoptimization::Action_maybe_recompile); 3849 } 3850 3851 // If control is dead, only non-array-path is taken. 3852 if (stopped()) return true; 3853 3854 // The works fine even if the array type is polymorphic. 3855 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3856 Node* result = load_array_length(array); 3857 3858 C->set_has_split_ifs(true); // Has chance for split-if optimization 3859 set_result(result); 3860 return true; 3861 } 3862 3863 //------------------------inline_array_copyOf---------------------------- 3864 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3865 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3866 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3867 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3868 3869 // Get the arguments. 3870 Node* original = argument(0); 3871 Node* start = is_copyOfRange? argument(1): intcon(0); 3872 Node* end = is_copyOfRange? argument(2): argument(1); 3873 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3874 3875 Node* newcopy; 3876 3877 // Set the original stack and the reexecute bit for the interpreter to reexecute 3878 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3879 { PreserveReexecuteState preexecs(this); 3880 jvms()->set_should_reexecute(true); 3881 3882 array_type_mirror = null_check(array_type_mirror); 3883 original = null_check(original); 3884 3885 // Check if a null path was taken unconditionally. 3886 if (stopped()) return true; 3887 3888 Node* orig_length = load_array_length(original); 3889 3890 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3891 klass_node = null_check(klass_node); 3892 3893 RegionNode* bailout = new RegionNode(1); 3894 record_for_igvn(bailout); 3895 3896 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3897 // Bail out if that is so. 3898 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3899 if (not_objArray != NULL) { 3900 // Improve the klass node's type from the new optimistic assumption: 3901 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3902 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3903 Node* cast = new CastPPNode(klass_node, akls); 3904 cast->init_req(0, control()); 3905 klass_node = _gvn.transform(cast); 3906 } 3907 3908 // Bail out if either start or end is negative. 3909 generate_negative_guard(start, bailout, &start); 3910 generate_negative_guard(end, bailout, &end); 3911 3912 Node* length = end; 3913 if (_gvn.type(start) != TypeInt::ZERO) { 3914 length = _gvn.transform(new SubINode(end, start)); 3915 } 3916 3917 // Bail out if length is negative. 3918 // Without this the new_array would throw 3919 // NegativeArraySizeException but IllegalArgumentException is what 3920 // should be thrown 3921 generate_negative_guard(length, bailout, &length); 3922 3923 if (bailout->req() > 1) { 3924 PreserveJVMState pjvms(this); 3925 set_control(_gvn.transform(bailout)); 3926 uncommon_trap(Deoptimization::Reason_intrinsic, 3927 Deoptimization::Action_maybe_recompile); 3928 } 3929 3930 if (!stopped()) { 3931 // How many elements will we copy from the original? 3932 // The answer is MinI(orig_length - start, length). 3933 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3934 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3935 3936 // Generate a direct call to the right arraycopy function(s). 3937 // We know the copy is disjoint but we might not know if the 3938 // oop stores need checking. 3939 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3940 // This will fail a store-check if x contains any non-nulls. 3941 3942 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3943 // loads/stores but it is legal only if we're sure the 3944 // Arrays.copyOf would succeed. So we need all input arguments 3945 // to the copyOf to be validated, including that the copy to the 3946 // new array won't trigger an ArrayStoreException. That subtype 3947 // check can be optimized if we know something on the type of 3948 // the input array from type speculation. 3949 if (_gvn.type(klass_node)->singleton()) { 3950 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3951 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3952 3953 int test = C->static_subtype_check(superk, subk); 3954 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3955 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3956 if (t_original->speculative_type() != NULL) { 3957 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3958 } 3959 } 3960 } 3961 3962 bool validated = false; 3963 // Reason_class_check rather than Reason_intrinsic because we 3964 // want to intrinsify even if this traps. 3965 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3966 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3967 klass_node); 3968 3969 if (not_subtype_ctrl != top()) { 3970 PreserveJVMState pjvms(this); 3971 set_control(not_subtype_ctrl); 3972 uncommon_trap(Deoptimization::Reason_class_check, 3973 Deoptimization::Action_make_not_entrant); 3974 assert(stopped(), "Should be stopped"); 3975 } 3976 validated = true; 3977 } 3978 3979 if (!stopped()) { 3980 newcopy = new_array(klass_node, length, 0); // no arguments to push 3981 3982 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3983 load_object_klass(original), klass_node); 3984 if (!is_copyOfRange) { 3985 ac->set_copyof(validated); 3986 } else { 3987 ac->set_copyofrange(validated); 3988 } 3989 Node* n = _gvn.transform(ac); 3990 if (n == ac) { 3991 ac->connect_outputs(this); 3992 } else { 3993 assert(validated, "shouldn't transform if all arguments not validated"); 3994 set_all_memory(n); 3995 } 3996 } 3997 } 3998 } // original reexecute is set back here 3999 4000 C->set_has_split_ifs(true); // Has chance for split-if optimization 4001 if (!stopped()) { 4002 set_result(newcopy); 4003 } 4004 return true; 4005 } 4006 4007 4008 //----------------------generate_virtual_guard--------------------------- 4009 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4010 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4011 RegionNode* slow_region) { 4012 ciMethod* method = callee(); 4013 int vtable_index = method->vtable_index(); 4014 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4015 err_msg_res("bad index %d", vtable_index)); 4016 // Get the Method* out of the appropriate vtable entry. 4017 int entry_offset = (InstanceKlass::vtable_start_offset() + 4018 vtable_index*vtableEntry::size()) * wordSize + 4019 vtableEntry::method_offset_in_bytes(); 4020 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4021 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4022 4023 // Compare the target method with the expected method (e.g., Object.hashCode). 4024 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4025 4026 Node* native_call = makecon(native_call_addr); 4027 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4028 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4029 4030 return generate_slow_guard(test_native, slow_region); 4031 } 4032 4033 //-----------------------generate_method_call---------------------------- 4034 // Use generate_method_call to make a slow-call to the real 4035 // method if the fast path fails. An alternative would be to 4036 // use a stub like OptoRuntime::slow_arraycopy_Java. 4037 // This only works for expanding the current library call, 4038 // not another intrinsic. (E.g., don't use this for making an 4039 // arraycopy call inside of the copyOf intrinsic.) 4040 CallJavaNode* 4041 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4042 // When compiling the intrinsic method itself, do not use this technique. 4043 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4044 4045 ciMethod* method = callee(); 4046 // ensure the JVMS we have will be correct for this call 4047 guarantee(method_id == method->intrinsic_id(), "must match"); 4048 4049 const TypeFunc* tf = TypeFunc::make(method); 4050 CallJavaNode* slow_call; 4051 if (is_static) { 4052 assert(!is_virtual, ""); 4053 slow_call = new CallStaticJavaNode(C, tf, 4054 SharedRuntime::get_resolve_static_call_stub(), 4055 method, bci()); 4056 } else if (is_virtual) { 4057 null_check_receiver(); 4058 int vtable_index = Method::invalid_vtable_index; 4059 if (UseInlineCaches) { 4060 // Suppress the vtable call 4061 } else { 4062 // hashCode and clone are not a miranda methods, 4063 // so the vtable index is fixed. 4064 // No need to use the linkResolver to get it. 4065 vtable_index = method->vtable_index(); 4066 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4067 err_msg_res("bad index %d", vtable_index)); 4068 } 4069 slow_call = new CallDynamicJavaNode(tf, 4070 SharedRuntime::get_resolve_virtual_call_stub(), 4071 method, vtable_index, bci()); 4072 } else { // neither virtual nor static: opt_virtual 4073 null_check_receiver(); 4074 slow_call = new CallStaticJavaNode(C, tf, 4075 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4076 method, bci()); 4077 slow_call->set_optimized_virtual(true); 4078 } 4079 set_arguments_for_java_call(slow_call); 4080 set_edges_for_java_call(slow_call); 4081 return slow_call; 4082 } 4083 4084 4085 /** 4086 * Build special case code for calls to hashCode on an object. This call may 4087 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4088 * slightly different code. 4089 */ 4090 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4091 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4092 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4093 4094 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4095 4096 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4097 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4098 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4099 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4100 Node* obj = NULL; 4101 if (!is_static) { 4102 // Check for hashing null object 4103 obj = null_check_receiver(); 4104 if (stopped()) return true; // unconditionally null 4105 result_reg->init_req(_null_path, top()); 4106 result_val->init_req(_null_path, top()); 4107 } else { 4108 // Do a null check, and return zero if null. 4109 // System.identityHashCode(null) == 0 4110 obj = argument(0); 4111 Node* null_ctl = top(); 4112 obj = null_check_oop(obj, &null_ctl); 4113 result_reg->init_req(_null_path, null_ctl); 4114 result_val->init_req(_null_path, _gvn.intcon(0)); 4115 } 4116 4117 // Unconditionally null? Then return right away. 4118 if (stopped()) { 4119 set_control( result_reg->in(_null_path)); 4120 if (!stopped()) 4121 set_result(result_val->in(_null_path)); 4122 return true; 4123 } 4124 4125 // We only go to the fast case code if we pass a number of guards. The 4126 // paths which do not pass are accumulated in the slow_region. 4127 RegionNode* slow_region = new RegionNode(1); 4128 record_for_igvn(slow_region); 4129 4130 // If this is a virtual call, we generate a funny guard. We pull out 4131 // the vtable entry corresponding to hashCode() from the target object. 4132 // If the target method which we are calling happens to be the native 4133 // Object hashCode() method, we pass the guard. We do not need this 4134 // guard for non-virtual calls -- the caller is known to be the native 4135 // Object hashCode(). 4136 if (is_virtual) { 4137 // After null check, get the object's klass. 4138 Node* obj_klass = load_object_klass(obj); 4139 generate_virtual_guard(obj_klass, slow_region); 4140 } 4141 4142 // Get the header out of the object, use LoadMarkNode when available 4143 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4144 // The control of the load must be NULL. Otherwise, the load can move before 4145 // the null check after castPP removal. 4146 Node* no_ctrl = NULL; 4147 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4148 4149 // Test the header to see if it is unlocked. 4150 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4151 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4152 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4153 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4154 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4155 4156 generate_slow_guard(test_unlocked, slow_region); 4157 4158 // Get the hash value and check to see that it has been properly assigned. 4159 // We depend on hash_mask being at most 32 bits and avoid the use of 4160 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4161 // vm: see markOop.hpp. 4162 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4163 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4164 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4165 // This hack lets the hash bits live anywhere in the mark object now, as long 4166 // as the shift drops the relevant bits into the low 32 bits. Note that 4167 // Java spec says that HashCode is an int so there's no point in capturing 4168 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4169 hshifted_header = ConvX2I(hshifted_header); 4170 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4171 4172 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4173 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4174 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4175 4176 generate_slow_guard(test_assigned, slow_region); 4177 4178 Node* init_mem = reset_memory(); 4179 // fill in the rest of the null path: 4180 result_io ->init_req(_null_path, i_o()); 4181 result_mem->init_req(_null_path, init_mem); 4182 4183 result_val->init_req(_fast_path, hash_val); 4184 result_reg->init_req(_fast_path, control()); 4185 result_io ->init_req(_fast_path, i_o()); 4186 result_mem->init_req(_fast_path, init_mem); 4187 4188 // Generate code for the slow case. We make a call to hashCode(). 4189 set_control(_gvn.transform(slow_region)); 4190 if (!stopped()) { 4191 // No need for PreserveJVMState, because we're using up the present state. 4192 set_all_memory(init_mem); 4193 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4194 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4195 Node* slow_result = set_results_for_java_call(slow_call); 4196 // this->control() comes from set_results_for_java_call 4197 result_reg->init_req(_slow_path, control()); 4198 result_val->init_req(_slow_path, slow_result); 4199 result_io ->set_req(_slow_path, i_o()); 4200 result_mem ->set_req(_slow_path, reset_memory()); 4201 } 4202 4203 // Return the combined state. 4204 set_i_o( _gvn.transform(result_io) ); 4205 set_all_memory( _gvn.transform(result_mem)); 4206 4207 set_result(result_reg, result_val); 4208 return true; 4209 } 4210 4211 //---------------------------inline_native_getClass---------------------------- 4212 // public final native Class<?> java.lang.Object.getClass(); 4213 // 4214 // Build special case code for calls to getClass on an object. 4215 bool LibraryCallKit::inline_native_getClass() { 4216 Node* obj = null_check_receiver(); 4217 if (stopped()) return true; 4218 set_result(load_mirror_from_klass(load_object_klass(obj))); 4219 return true; 4220 } 4221 4222 //-----------------inline_native_Reflection_getCallerClass--------------------- 4223 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4224 // 4225 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4226 // 4227 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4228 // in that it must skip particular security frames and checks for 4229 // caller sensitive methods. 4230 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4231 #ifndef PRODUCT 4232 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4233 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4234 } 4235 #endif 4236 4237 if (!jvms()->has_method()) { 4238 #ifndef PRODUCT 4239 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4240 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4241 } 4242 #endif 4243 return false; 4244 } 4245 4246 // Walk back up the JVM state to find the caller at the required 4247 // depth. 4248 JVMState* caller_jvms = jvms(); 4249 4250 // Cf. JVM_GetCallerClass 4251 // NOTE: Start the loop at depth 1 because the current JVM state does 4252 // not include the Reflection.getCallerClass() frame. 4253 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4254 ciMethod* m = caller_jvms->method(); 4255 switch (n) { 4256 case 0: 4257 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4258 break; 4259 case 1: 4260 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4261 if (!m->caller_sensitive()) { 4262 #ifndef PRODUCT 4263 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4264 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4265 } 4266 #endif 4267 return false; // bail-out; let JVM_GetCallerClass do the work 4268 } 4269 break; 4270 default: 4271 if (!m->is_ignored_by_security_stack_walk()) { 4272 // We have reached the desired frame; return the holder class. 4273 // Acquire method holder as java.lang.Class and push as constant. 4274 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4275 ciInstance* caller_mirror = caller_klass->java_mirror(); 4276 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4277 4278 #ifndef PRODUCT 4279 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4280 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()); 4281 tty->print_cr(" JVM state at this point:"); 4282 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4283 ciMethod* m = jvms()->of_depth(i)->method(); 4284 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4285 } 4286 } 4287 #endif 4288 return true; 4289 } 4290 break; 4291 } 4292 } 4293 4294 #ifndef PRODUCT 4295 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4296 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4297 tty->print_cr(" JVM state at this point:"); 4298 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4299 ciMethod* m = jvms()->of_depth(i)->method(); 4300 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4301 } 4302 } 4303 #endif 4304 4305 return false; // bail-out; let JVM_GetCallerClass do the work 4306 } 4307 4308 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4309 Node* arg = argument(0); 4310 Node* result; 4311 4312 switch (id) { 4313 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4314 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4315 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4316 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4317 4318 case vmIntrinsics::_doubleToLongBits: { 4319 // two paths (plus control) merge in a wood 4320 RegionNode *r = new RegionNode(3); 4321 Node *phi = new PhiNode(r, TypeLong::LONG); 4322 4323 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4324 // Build the boolean node 4325 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4326 4327 // Branch either way. 4328 // NaN case is less traveled, which makes all the difference. 4329 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4330 Node *opt_isnan = _gvn.transform(ifisnan); 4331 assert( opt_isnan->is_If(), "Expect an IfNode"); 4332 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4333 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4334 4335 set_control(iftrue); 4336 4337 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4338 Node *slow_result = longcon(nan_bits); // return NaN 4339 phi->init_req(1, _gvn.transform( slow_result )); 4340 r->init_req(1, iftrue); 4341 4342 // Else fall through 4343 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4344 set_control(iffalse); 4345 4346 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4347 r->init_req(2, iffalse); 4348 4349 // Post merge 4350 set_control(_gvn.transform(r)); 4351 record_for_igvn(r); 4352 4353 C->set_has_split_ifs(true); // Has chance for split-if optimization 4354 result = phi; 4355 assert(result->bottom_type()->isa_long(), "must be"); 4356 break; 4357 } 4358 4359 case vmIntrinsics::_floatToIntBits: { 4360 // two paths (plus control) merge in a wood 4361 RegionNode *r = new RegionNode(3); 4362 Node *phi = new PhiNode(r, TypeInt::INT); 4363 4364 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4365 // Build the boolean node 4366 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4367 4368 // Branch either way. 4369 // NaN case is less traveled, which makes all the difference. 4370 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4371 Node *opt_isnan = _gvn.transform(ifisnan); 4372 assert( opt_isnan->is_If(), "Expect an IfNode"); 4373 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4374 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4375 4376 set_control(iftrue); 4377 4378 static const jint nan_bits = 0x7fc00000; 4379 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4380 phi->init_req(1, _gvn.transform( slow_result )); 4381 r->init_req(1, iftrue); 4382 4383 // Else fall through 4384 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4385 set_control(iffalse); 4386 4387 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4388 r->init_req(2, iffalse); 4389 4390 // Post merge 4391 set_control(_gvn.transform(r)); 4392 record_for_igvn(r); 4393 4394 C->set_has_split_ifs(true); // Has chance for split-if optimization 4395 result = phi; 4396 assert(result->bottom_type()->isa_int(), "must be"); 4397 break; 4398 } 4399 4400 default: 4401 fatal_unexpected_iid(id); 4402 break; 4403 } 4404 set_result(_gvn.transform(result)); 4405 return true; 4406 } 4407 4408 #ifdef _LP64 4409 #define XTOP ,top() /*additional argument*/ 4410 #else //_LP64 4411 #define XTOP /*no additional argument*/ 4412 #endif //_LP64 4413 4414 //----------------------inline_unsafe_copyMemory------------------------- 4415 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4416 bool LibraryCallKit::inline_unsafe_copyMemory() { 4417 if (callee()->is_static()) return false; // caller must have the capability! 4418 null_check_receiver(); // null-check receiver 4419 if (stopped()) return true; 4420 4421 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4422 4423 Node* src_ptr = argument(1); // type: oop 4424 Node* src_off = ConvL2X(argument(2)); // type: long 4425 Node* dst_ptr = argument(4); // type: oop 4426 Node* dst_off = ConvL2X(argument(5)); // type: long 4427 Node* size = ConvL2X(argument(7)); // type: long 4428 4429 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4430 "fieldOffset must be byte-scaled"); 4431 4432 Node* src = make_unsafe_address(src_ptr, src_off); 4433 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4434 4435 // Conservatively insert a memory barrier on all memory slices. 4436 // Do not let writes of the copy source or destination float below the copy. 4437 insert_mem_bar(Op_MemBarCPUOrder); 4438 4439 // Call it. Note that the length argument is not scaled. 4440 make_runtime_call(RC_LEAF|RC_NO_FP, 4441 OptoRuntime::fast_arraycopy_Type(), 4442 StubRoutines::unsafe_arraycopy(), 4443 "unsafe_arraycopy", 4444 TypeRawPtr::BOTTOM, 4445 src, dst, size XTOP); 4446 4447 // Do not let reads of the copy destination float above the copy. 4448 insert_mem_bar(Op_MemBarCPUOrder); 4449 4450 return true; 4451 } 4452 4453 //------------------------clone_coping----------------------------------- 4454 // Helper function for inline_native_clone. 4455 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4456 assert(obj_size != NULL, ""); 4457 Node* raw_obj = alloc_obj->in(1); 4458 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4459 4460 AllocateNode* alloc = NULL; 4461 if (ReduceBulkZeroing) { 4462 // We will be completely responsible for initializing this object - 4463 // mark Initialize node as complete. 4464 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4465 // The object was just allocated - there should be no any stores! 4466 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4467 // Mark as complete_with_arraycopy so that on AllocateNode 4468 // expansion, we know this AllocateNode is initialized by an array 4469 // copy and a StoreStore barrier exists after the array copy. 4470 alloc->initialization()->set_complete_with_arraycopy(); 4471 } 4472 4473 // Copy the fastest available way. 4474 // TODO: generate fields copies for small objects instead. 4475 Node* src = obj; 4476 Node* dest = alloc_obj; 4477 Node* size = _gvn.transform(obj_size); 4478 4479 // Exclude the header but include array length to copy by 8 bytes words. 4480 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4481 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4482 instanceOopDesc::base_offset_in_bytes(); 4483 // base_off: 4484 // 8 - 32-bit VM 4485 // 12 - 64-bit VM, compressed klass 4486 // 16 - 64-bit VM, normal klass 4487 if (base_off % BytesPerLong != 0) { 4488 assert(UseCompressedClassPointers, ""); 4489 if (is_array) { 4490 // Exclude length to copy by 8 bytes words. 4491 base_off += sizeof(int); 4492 } else { 4493 // Include klass to copy by 8 bytes words. 4494 base_off = instanceOopDesc::klass_offset_in_bytes(); 4495 } 4496 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4497 } 4498 src = basic_plus_adr(src, base_off); 4499 dest = basic_plus_adr(dest, base_off); 4500 4501 // Compute the length also, if needed: 4502 Node* countx = size; 4503 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4504 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4505 4506 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4507 4508 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4509 ac->set_clonebasic(); 4510 Node* n = _gvn.transform(ac); 4511 if (n == ac) { 4512 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4513 } else { 4514 set_all_memory(n); 4515 } 4516 4517 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4518 if (card_mark) { 4519 assert(!is_array, ""); 4520 // Put in store barrier for any and all oops we are sticking 4521 // into this object. (We could avoid this if we could prove 4522 // that the object type contains no oop fields at all.) 4523 Node* no_particular_value = NULL; 4524 Node* no_particular_field = NULL; 4525 int raw_adr_idx = Compile::AliasIdxRaw; 4526 post_barrier(control(), 4527 memory(raw_adr_type), 4528 alloc_obj, 4529 no_particular_field, 4530 raw_adr_idx, 4531 no_particular_value, 4532 T_OBJECT, 4533 false); 4534 } 4535 4536 // Do not let reads from the cloned object float above the arraycopy. 4537 if (alloc != NULL) { 4538 // Do not let stores that initialize this object be reordered with 4539 // a subsequent store that would make this object accessible by 4540 // other threads. 4541 // Record what AllocateNode this StoreStore protects so that 4542 // escape analysis can go from the MemBarStoreStoreNode to the 4543 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4544 // based on the escape status of the AllocateNode. 4545 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4546 } else { 4547 insert_mem_bar(Op_MemBarCPUOrder); 4548 } 4549 } 4550 4551 //------------------------inline_native_clone---------------------------- 4552 // protected native Object java.lang.Object.clone(); 4553 // 4554 // Here are the simple edge cases: 4555 // null receiver => normal trap 4556 // virtual and clone was overridden => slow path to out-of-line clone 4557 // not cloneable or finalizer => slow path to out-of-line Object.clone 4558 // 4559 // The general case has two steps, allocation and copying. 4560 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4561 // 4562 // Copying also has two cases, oop arrays and everything else. 4563 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4564 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4565 // 4566 // These steps fold up nicely if and when the cloned object's klass 4567 // can be sharply typed as an object array, a type array, or an instance. 4568 // 4569 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4570 PhiNode* result_val; 4571 4572 // Set the reexecute bit for the interpreter to reexecute 4573 // the bytecode that invokes Object.clone if deoptimization happens. 4574 { PreserveReexecuteState preexecs(this); 4575 jvms()->set_should_reexecute(true); 4576 4577 Node* obj = null_check_receiver(); 4578 if (stopped()) return true; 4579 4580 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4581 4582 // If we are going to clone an instance, we need its exact type to 4583 // know the number and types of fields to convert the clone to 4584 // loads/stores. Maybe a speculative type can help us. 4585 if (!obj_type->klass_is_exact() && 4586 obj_type->speculative_type() != NULL && 4587 obj_type->speculative_type()->is_instance_klass()) { 4588 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4589 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4590 !spec_ik->has_injected_fields()) { 4591 ciKlass* k = obj_type->klass(); 4592 if (!k->is_instance_klass() || 4593 k->as_instance_klass()->is_interface() || 4594 k->as_instance_klass()->has_subklass()) { 4595 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4596 } 4597 } 4598 } 4599 4600 Node* obj_klass = load_object_klass(obj); 4601 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4602 const TypeOopPtr* toop = ((tklass != NULL) 4603 ? tklass->as_instance_type() 4604 : TypeInstPtr::NOTNULL); 4605 4606 // Conservatively insert a memory barrier on all memory slices. 4607 // Do not let writes into the original float below the clone. 4608 insert_mem_bar(Op_MemBarCPUOrder); 4609 4610 // paths into result_reg: 4611 enum { 4612 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4613 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4614 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4615 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4616 PATH_LIMIT 4617 }; 4618 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4619 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4620 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4621 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4622 record_for_igvn(result_reg); 4623 4624 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4625 int raw_adr_idx = Compile::AliasIdxRaw; 4626 4627 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4628 if (array_ctl != NULL) { 4629 // It's an array. 4630 PreserveJVMState pjvms(this); 4631 set_control(array_ctl); 4632 Node* obj_length = load_array_length(obj); 4633 Node* obj_size = NULL; 4634 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4635 4636 if (!use_ReduceInitialCardMarks()) { 4637 // If it is an oop array, it requires very special treatment, 4638 // because card marking is required on each card of the array. 4639 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4640 if (is_obja != NULL) { 4641 PreserveJVMState pjvms2(this); 4642 set_control(is_obja); 4643 // Generate a direct call to the right arraycopy function(s). 4644 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4645 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4646 ac->set_cloneoop(); 4647 Node* n = _gvn.transform(ac); 4648 assert(n == ac, "cannot disappear"); 4649 ac->connect_outputs(this); 4650 4651 result_reg->init_req(_objArray_path, control()); 4652 result_val->init_req(_objArray_path, alloc_obj); 4653 result_i_o ->set_req(_objArray_path, i_o()); 4654 result_mem ->set_req(_objArray_path, reset_memory()); 4655 } 4656 } 4657 // Otherwise, there are no card marks to worry about. 4658 // (We can dispense with card marks if we know the allocation 4659 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4660 // causes the non-eden paths to take compensating steps to 4661 // simulate a fresh allocation, so that no further 4662 // card marks are required in compiled code to initialize 4663 // the object.) 4664 4665 if (!stopped()) { 4666 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4667 4668 // Present the results of the copy. 4669 result_reg->init_req(_array_path, control()); 4670 result_val->init_req(_array_path, alloc_obj); 4671 result_i_o ->set_req(_array_path, i_o()); 4672 result_mem ->set_req(_array_path, reset_memory()); 4673 } 4674 } 4675 4676 // We only go to the instance fast case code if we pass a number of guards. 4677 // The paths which do not pass are accumulated in the slow_region. 4678 RegionNode* slow_region = new RegionNode(1); 4679 record_for_igvn(slow_region); 4680 if (!stopped()) { 4681 // It's an instance (we did array above). Make the slow-path tests. 4682 // If this is a virtual call, we generate a funny guard. We grab 4683 // the vtable entry corresponding to clone() from the target object. 4684 // If the target method which we are calling happens to be the 4685 // Object clone() method, we pass the guard. We do not need this 4686 // guard for non-virtual calls; the caller is known to be the native 4687 // Object clone(). 4688 if (is_virtual) { 4689 generate_virtual_guard(obj_klass, slow_region); 4690 } 4691 4692 // The object must be cloneable and must not have a finalizer. 4693 // Both of these conditions may be checked in a single test. 4694 // We could optimize the cloneable test further, but we don't care. 4695 generate_access_flags_guard(obj_klass, 4696 // Test both conditions: 4697 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4698 // Must be cloneable but not finalizer: 4699 JVM_ACC_IS_CLONEABLE, 4700 slow_region); 4701 } 4702 4703 if (!stopped()) { 4704 // It's an instance, and it passed the slow-path tests. 4705 PreserveJVMState pjvms(this); 4706 Node* obj_size = NULL; 4707 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4708 // is reexecuted if deoptimization occurs and there could be problems when merging 4709 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4710 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4711 4712 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4713 4714 // Present the results of the slow call. 4715 result_reg->init_req(_instance_path, control()); 4716 result_val->init_req(_instance_path, alloc_obj); 4717 result_i_o ->set_req(_instance_path, i_o()); 4718 result_mem ->set_req(_instance_path, reset_memory()); 4719 } 4720 4721 // Generate code for the slow case. We make a call to clone(). 4722 set_control(_gvn.transform(slow_region)); 4723 if (!stopped()) { 4724 PreserveJVMState pjvms(this); 4725 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4726 Node* slow_result = set_results_for_java_call(slow_call); 4727 // this->control() comes from set_results_for_java_call 4728 result_reg->init_req(_slow_path, control()); 4729 result_val->init_req(_slow_path, slow_result); 4730 result_i_o ->set_req(_slow_path, i_o()); 4731 result_mem ->set_req(_slow_path, reset_memory()); 4732 } 4733 4734 // Return the combined state. 4735 set_control( _gvn.transform(result_reg)); 4736 set_i_o( _gvn.transform(result_i_o)); 4737 set_all_memory( _gvn.transform(result_mem)); 4738 } // original reexecute is set back here 4739 4740 set_result(_gvn.transform(result_val)); 4741 return true; 4742 } 4743 4744 // If we have a tighly coupled allocation, the arraycopy may take care 4745 // of the array initialization. If one of the guards we insert between 4746 // the allocation and the arraycopy causes a deoptimization, an 4747 // unitialized array will escape the compiled method. To prevent that 4748 // we set the JVM state for uncommon traps between the allocation and 4749 // the arraycopy to the state before the allocation so, in case of 4750 // deoptimization, we'll reexecute the allocation and the 4751 // initialization. 4752 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4753 if (alloc != NULL) { 4754 ciMethod* trap_method = alloc->jvms()->method(); 4755 int trap_bci = alloc->jvms()->bci(); 4756 4757 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4758 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4759 // Make sure there's no store between the allocation and the 4760 // arraycopy otherwise visible side effects could be rexecuted 4761 // in case of deoptimization and cause incorrect execution. 4762 bool no_interfering_store = true; 4763 Node* mem = alloc->in(TypeFunc::Memory); 4764 if (mem->is_MergeMem()) { 4765 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4766 Node* n = mms.memory(); 4767 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4768 assert(n->is_Store(), "what else?"); 4769 no_interfering_store = false; 4770 break; 4771 } 4772 } 4773 } else { 4774 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4775 Node* n = mms.memory(); 4776 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4777 assert(n->is_Store(), "what else?"); 4778 no_interfering_store = false; 4779 break; 4780 } 4781 } 4782 } 4783 4784 if (no_interfering_store) { 4785 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4786 uint size = alloc->req(); 4787 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4788 old_jvms->set_map(sfpt); 4789 for (uint i = 0; i < size; i++) { 4790 sfpt->init_req(i, alloc->in(i)); 4791 } 4792 // re-push array length for deoptimization 4793 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4794 old_jvms->set_sp(old_jvms->sp()+1); 4795 old_jvms->set_monoff(old_jvms->monoff()+1); 4796 old_jvms->set_scloff(old_jvms->scloff()+1); 4797 old_jvms->set_endoff(old_jvms->endoff()+1); 4798 old_jvms->set_should_reexecute(true); 4799 4800 sfpt->set_i_o(map()->i_o()); 4801 sfpt->set_memory(map()->memory()); 4802 sfpt->set_control(map()->control()); 4803 4804 JVMState* saved_jvms = jvms(); 4805 saved_reexecute_sp = _reexecute_sp; 4806 4807 set_jvms(sfpt->jvms()); 4808 _reexecute_sp = jvms()->sp(); 4809 4810 return saved_jvms; 4811 } 4812 } 4813 } 4814 return NULL; 4815 } 4816 4817 // In case of a deoptimization, we restart execution at the 4818 // allocation, allocating a new array. We would leave an uninitialized 4819 // array in the heap that GCs wouldn't expect. Move the allocation 4820 // after the traps so we don't allocate the array if we 4821 // deoptimize. This is possible because tightly_coupled_allocation() 4822 // guarantees there's no observer of the allocated array at this point 4823 // and the control flow is simple enough. 4824 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4825 if (saved_jvms != NULL && !stopped()) { 4826 assert(alloc != NULL, "only with a tightly coupled allocation"); 4827 // restore JVM state to the state at the arraycopy 4828 saved_jvms->map()->set_control(map()->control()); 4829 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4830 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4831 // If we've improved the types of some nodes (null check) while 4832 // emitting the guards, propagate them to the current state 4833 map()->replaced_nodes().apply(saved_jvms->map()); 4834 set_jvms(saved_jvms); 4835 _reexecute_sp = saved_reexecute_sp; 4836 4837 // Remove the allocation from above the guards 4838 CallProjections callprojs; 4839 alloc->extract_projections(&callprojs, true); 4840 InitializeNode* init = alloc->initialization(); 4841 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4842 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4843 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4844 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4845 4846 // move the allocation here (after the guards) 4847 _gvn.hash_delete(alloc); 4848 alloc->set_req(TypeFunc::Control, control()); 4849 alloc->set_req(TypeFunc::I_O, i_o()); 4850 Node *mem = reset_memory(); 4851 set_all_memory(mem); 4852 alloc->set_req(TypeFunc::Memory, mem); 4853 set_control(init->proj_out(TypeFunc::Control)); 4854 set_i_o(callprojs.fallthrough_ioproj); 4855 4856 // Update memory as done in GraphKit::set_output_for_allocation() 4857 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4858 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4859 if (ary_type->isa_aryptr() && length_type != NULL) { 4860 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4861 } 4862 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4863 int elemidx = C->get_alias_index(telemref); 4864 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4865 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4866 4867 Node* allocx = _gvn.transform(alloc); 4868 assert(allocx == alloc, "where has the allocation gone?"); 4869 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4870 4871 _gvn.hash_delete(dest); 4872 dest->set_req(0, control()); 4873 Node* destx = _gvn.transform(dest); 4874 assert(destx == dest, "where has the allocation result gone?"); 4875 } 4876 } 4877 4878 4879 //------------------------------inline_arraycopy----------------------- 4880 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4881 // Object dest, int destPos, 4882 // int length); 4883 bool LibraryCallKit::inline_arraycopy() { 4884 // Get the arguments. 4885 Node* src = argument(0); // type: oop 4886 Node* src_offset = argument(1); // type: int 4887 Node* dest = argument(2); // type: oop 4888 Node* dest_offset = argument(3); // type: int 4889 Node* length = argument(4); // type: int 4890 4891 4892 // Check for allocation before we add nodes that would confuse 4893 // tightly_coupled_allocation() 4894 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4895 4896 int saved_reexecute_sp = -1; 4897 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4898 // See arraycopy_restore_alloc_state() comment 4899 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4900 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4901 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4902 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4903 4904 // The following tests must be performed 4905 // (1) src and dest are arrays. 4906 // (2) src and dest arrays must have elements of the same BasicType 4907 // (3) src and dest must not be null. 4908 // (4) src_offset must not be negative. 4909 // (5) dest_offset must not be negative. 4910 // (6) length must not be negative. 4911 // (7) src_offset + length must not exceed length of src. 4912 // (8) dest_offset + length must not exceed length of dest. 4913 // (9) each element of an oop array must be assignable 4914 4915 // (3) src and dest must not be null. 4916 // always do this here because we need the JVM state for uncommon traps 4917 Node* null_ctl = top(); 4918 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4919 assert(null_ctl->is_top(), "no null control here"); 4920 dest = null_check(dest, T_ARRAY); 4921 4922 if (!can_emit_guards) { 4923 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4924 // guards but the arraycopy node could still take advantage of a 4925 // tightly allocated allocation. tightly_coupled_allocation() is 4926 // called again to make sure it takes the null check above into 4927 // account: the null check is mandatory and if it caused an 4928 // uncommon trap to be emitted then the allocation can't be 4929 // considered tightly coupled in this context. 4930 alloc = tightly_coupled_allocation(dest, NULL); 4931 } 4932 4933 bool validated = false; 4934 4935 const Type* src_type = _gvn.type(src); 4936 const Type* dest_type = _gvn.type(dest); 4937 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4938 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4939 4940 // Do we have the type of src? 4941 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4942 // Do we have the type of dest? 4943 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4944 // Is the type for src from speculation? 4945 bool src_spec = false; 4946 // Is the type for dest from speculation? 4947 bool dest_spec = false; 4948 4949 if ((!has_src || !has_dest) && can_emit_guards) { 4950 // We don't have sufficient type information, let's see if 4951 // speculative types can help. We need to have types for both src 4952 // and dest so that it pays off. 4953 4954 // Do we already have or could we have type information for src 4955 bool could_have_src = has_src; 4956 // Do we already have or could we have type information for dest 4957 bool could_have_dest = has_dest; 4958 4959 ciKlass* src_k = NULL; 4960 if (!has_src) { 4961 src_k = src_type->speculative_type_not_null(); 4962 if (src_k != NULL && src_k->is_array_klass()) { 4963 could_have_src = true; 4964 } 4965 } 4966 4967 ciKlass* dest_k = NULL; 4968 if (!has_dest) { 4969 dest_k = dest_type->speculative_type_not_null(); 4970 if (dest_k != NULL && dest_k->is_array_klass()) { 4971 could_have_dest = true; 4972 } 4973 } 4974 4975 if (could_have_src && could_have_dest) { 4976 // This is going to pay off so emit the required guards 4977 if (!has_src) { 4978 src = maybe_cast_profiled_obj(src, src_k, true); 4979 src_type = _gvn.type(src); 4980 top_src = src_type->isa_aryptr(); 4981 has_src = (top_src != NULL && top_src->klass() != NULL); 4982 src_spec = true; 4983 } 4984 if (!has_dest) { 4985 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4986 dest_type = _gvn.type(dest); 4987 top_dest = dest_type->isa_aryptr(); 4988 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4989 dest_spec = true; 4990 } 4991 } 4992 } 4993 4994 if (has_src && has_dest && can_emit_guards) { 4995 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4996 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4997 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4998 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4999 5000 if (src_elem == dest_elem && src_elem == T_OBJECT) { 5001 // If both arrays are object arrays then having the exact types 5002 // for both will remove the need for a subtype check at runtime 5003 // before the call and may make it possible to pick a faster copy 5004 // routine (without a subtype check on every element) 5005 // Do we have the exact type of src? 5006 bool could_have_src = src_spec; 5007 // Do we have the exact type of dest? 5008 bool could_have_dest = dest_spec; 5009 ciKlass* src_k = top_src->klass(); 5010 ciKlass* dest_k = top_dest->klass(); 5011 if (!src_spec) { 5012 src_k = src_type->speculative_type_not_null(); 5013 if (src_k != NULL && src_k->is_array_klass()) { 5014 could_have_src = true; 5015 } 5016 } 5017 if (!dest_spec) { 5018 dest_k = dest_type->speculative_type_not_null(); 5019 if (dest_k != NULL && dest_k->is_array_klass()) { 5020 could_have_dest = true; 5021 } 5022 } 5023 if (could_have_src && could_have_dest) { 5024 // If we can have both exact types, emit the missing guards 5025 if (could_have_src && !src_spec) { 5026 src = maybe_cast_profiled_obj(src, src_k, true); 5027 } 5028 if (could_have_dest && !dest_spec) { 5029 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5030 } 5031 } 5032 } 5033 } 5034 5035 ciMethod* trap_method = method(); 5036 int trap_bci = bci(); 5037 if (saved_jvms != NULL) { 5038 trap_method = alloc->jvms()->method(); 5039 trap_bci = alloc->jvms()->bci(); 5040 } 5041 5042 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5043 can_emit_guards && 5044 !src->is_top() && !dest->is_top()) { 5045 // validate arguments: enables transformation the ArrayCopyNode 5046 validated = true; 5047 5048 RegionNode* slow_region = new RegionNode(1); 5049 record_for_igvn(slow_region); 5050 5051 // (1) src and dest are arrays. 5052 generate_non_array_guard(load_object_klass(src), slow_region); 5053 generate_non_array_guard(load_object_klass(dest), slow_region); 5054 5055 // (2) src and dest arrays must have elements of the same BasicType 5056 // done at macro expansion or at Ideal transformation time 5057 5058 // (4) src_offset must not be negative. 5059 generate_negative_guard(src_offset, slow_region); 5060 5061 // (5) dest_offset must not be negative. 5062 generate_negative_guard(dest_offset, slow_region); 5063 5064 // (7) src_offset + length must not exceed length of src. 5065 generate_limit_guard(src_offset, length, 5066 load_array_length(src), 5067 slow_region); 5068 5069 // (8) dest_offset + length must not exceed length of dest. 5070 generate_limit_guard(dest_offset, length, 5071 load_array_length(dest), 5072 slow_region); 5073 5074 // (9) each element of an oop array must be assignable 5075 Node* src_klass = load_object_klass(src); 5076 Node* dest_klass = load_object_klass(dest); 5077 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5078 5079 if (not_subtype_ctrl != top()) { 5080 PreserveJVMState pjvms(this); 5081 set_control(not_subtype_ctrl); 5082 uncommon_trap(Deoptimization::Reason_intrinsic, 5083 Deoptimization::Action_make_not_entrant); 5084 assert(stopped(), "Should be stopped"); 5085 } 5086 { 5087 PreserveJVMState pjvms(this); 5088 set_control(_gvn.transform(slow_region)); 5089 uncommon_trap(Deoptimization::Reason_intrinsic, 5090 Deoptimization::Action_make_not_entrant); 5091 assert(stopped(), "Should be stopped"); 5092 } 5093 } 5094 5095 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5096 5097 if (stopped()) { 5098 return true; 5099 } 5100 5101 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5102 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5103 // so the compiler has a chance to eliminate them: during macro expansion, 5104 // we have to set their control (CastPP nodes are eliminated). 5105 load_object_klass(src), load_object_klass(dest), 5106 load_array_length(src), load_array_length(dest)); 5107 5108 ac->set_arraycopy(validated); 5109 5110 Node* n = _gvn.transform(ac); 5111 if (n == ac) { 5112 ac->connect_outputs(this); 5113 } else { 5114 assert(validated, "shouldn't transform if all arguments not validated"); 5115 set_all_memory(n); 5116 } 5117 5118 return true; 5119 } 5120 5121 5122 // Helper function which determines if an arraycopy immediately follows 5123 // an allocation, with no intervening tests or other escapes for the object. 5124 AllocateArrayNode* 5125 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5126 RegionNode* slow_region) { 5127 if (stopped()) return NULL; // no fast path 5128 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5129 5130 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5131 if (alloc == NULL) return NULL; 5132 5133 Node* rawmem = memory(Compile::AliasIdxRaw); 5134 // Is the allocation's memory state untouched? 5135 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5136 // Bail out if there have been raw-memory effects since the allocation. 5137 // (Example: There might have been a call or safepoint.) 5138 return NULL; 5139 } 5140 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5141 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5142 return NULL; 5143 } 5144 5145 // There must be no unexpected observers of this allocation. 5146 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5147 Node* obs = ptr->fast_out(i); 5148 if (obs != this->map()) { 5149 return NULL; 5150 } 5151 } 5152 5153 // This arraycopy must unconditionally follow the allocation of the ptr. 5154 Node* alloc_ctl = ptr->in(0); 5155 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5156 5157 Node* ctl = control(); 5158 while (ctl != alloc_ctl) { 5159 // There may be guards which feed into the slow_region. 5160 // Any other control flow means that we might not get a chance 5161 // to finish initializing the allocated object. 5162 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5163 IfNode* iff = ctl->in(0)->as_If(); 5164 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5165 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5166 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5167 ctl = iff->in(0); // This test feeds the known slow_region. 5168 continue; 5169 } 5170 // One more try: Various low-level checks bottom out in 5171 // uncommon traps. If the debug-info of the trap omits 5172 // any reference to the allocation, as we've already 5173 // observed, then there can be no objection to the trap. 5174 bool found_trap = false; 5175 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5176 Node* obs = not_ctl->fast_out(j); 5177 if (obs->in(0) == not_ctl && obs->is_Call() && 5178 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5179 found_trap = true; break; 5180 } 5181 } 5182 if (found_trap) { 5183 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5184 continue; 5185 } 5186 } 5187 return NULL; 5188 } 5189 5190 // If we get this far, we have an allocation which immediately 5191 // precedes the arraycopy, and we can take over zeroing the new object. 5192 // The arraycopy will finish the initialization, and provide 5193 // a new control state to which we will anchor the destination pointer. 5194 5195 return alloc; 5196 } 5197 5198 //-------------inline_encodeISOArray----------------------------------- 5199 // encode char[] to byte[] in ISO_8859_1 5200 bool LibraryCallKit::inline_encodeISOArray() { 5201 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5202 // no receiver since it is static method 5203 Node *src = argument(0); 5204 Node *src_offset = argument(1); 5205 Node *dst = argument(2); 5206 Node *dst_offset = argument(3); 5207 Node *length = argument(4); 5208 5209 const Type* src_type = src->Value(&_gvn); 5210 const Type* dst_type = dst->Value(&_gvn); 5211 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5212 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5213 if (top_src == NULL || top_src->klass() == NULL || 5214 top_dest == NULL || top_dest->klass() == NULL) { 5215 // failed array check 5216 return false; 5217 } 5218 5219 // Figure out the size and type of the elements we will be copying. 5220 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5221 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5222 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5223 return false; 5224 } 5225 Node* src_start = array_element_address(src, src_offset, src_elem); 5226 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5227 // 'src_start' points to src array + scaled offset 5228 // 'dst_start' points to dst array + scaled offset 5229 5230 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5231 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5232 enc = _gvn.transform(enc); 5233 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5234 set_memory(res_mem, mtype); 5235 set_result(enc); 5236 return true; 5237 } 5238 5239 //-------------inline_multiplyToLen----------------------------------- 5240 bool LibraryCallKit::inline_multiplyToLen() { 5241 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5242 5243 address stubAddr = StubRoutines::multiplyToLen(); 5244 if (stubAddr == NULL) { 5245 return false; // Intrinsic's stub is not implemented on this platform 5246 } 5247 const char* stubName = "multiplyToLen"; 5248 5249 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5250 5251 Node* x = argument(1); 5252 Node* xlen = argument(2); 5253 Node* y = argument(3); 5254 Node* ylen = argument(4); 5255 Node* z = argument(5); 5256 5257 const Type* x_type = x->Value(&_gvn); 5258 const Type* y_type = y->Value(&_gvn); 5259 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5260 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5261 if (top_x == NULL || top_x->klass() == NULL || 5262 top_y == NULL || top_y->klass() == NULL) { 5263 // failed array check 5264 return false; 5265 } 5266 5267 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5268 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5269 if (x_elem != T_INT || y_elem != T_INT) { 5270 return false; 5271 } 5272 5273 // Set the original stack and the reexecute bit for the interpreter to reexecute 5274 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5275 // on the return from z array allocation in runtime. 5276 { PreserveReexecuteState preexecs(this); 5277 jvms()->set_should_reexecute(true); 5278 5279 Node* x_start = array_element_address(x, intcon(0), x_elem); 5280 Node* y_start = array_element_address(y, intcon(0), y_elem); 5281 // 'x_start' points to x array + scaled xlen 5282 // 'y_start' points to y array + scaled ylen 5283 5284 // Allocate the result array 5285 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5286 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5287 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5288 5289 IdealKit ideal(this); 5290 5291 #define __ ideal. 5292 Node* one = __ ConI(1); 5293 Node* zero = __ ConI(0); 5294 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5295 __ set(need_alloc, zero); 5296 __ set(z_alloc, z); 5297 __ if_then(z, BoolTest::eq, null()); { 5298 __ increment (need_alloc, one); 5299 } __ else_(); { 5300 // Update graphKit memory and control from IdealKit. 5301 sync_kit(ideal); 5302 Node* zlen_arg = load_array_length(z); 5303 // Update IdealKit memory and control from graphKit. 5304 __ sync_kit(this); 5305 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5306 __ increment (need_alloc, one); 5307 } __ end_if(); 5308 } __ end_if(); 5309 5310 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5311 // Update graphKit memory and control from IdealKit. 5312 sync_kit(ideal); 5313 Node * narr = new_array(klass_node, zlen, 1); 5314 // Update IdealKit memory and control from graphKit. 5315 __ sync_kit(this); 5316 __ set(z_alloc, narr); 5317 } __ end_if(); 5318 5319 sync_kit(ideal); 5320 z = __ value(z_alloc); 5321 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5322 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5323 // Final sync IdealKit and GraphKit. 5324 final_sync(ideal); 5325 #undef __ 5326 5327 Node* z_start = array_element_address(z, intcon(0), T_INT); 5328 5329 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5330 OptoRuntime::multiplyToLen_Type(), 5331 stubAddr, stubName, TypePtr::BOTTOM, 5332 x_start, xlen, y_start, ylen, z_start, zlen); 5333 } // original reexecute is set back here 5334 5335 C->set_has_split_ifs(true); // Has chance for split-if optimization 5336 set_result(z); 5337 return true; 5338 } 5339 5340 //-------------inline_squareToLen------------------------------------ 5341 bool LibraryCallKit::inline_squareToLen() { 5342 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5343 5344 address stubAddr = StubRoutines::squareToLen(); 5345 if (stubAddr == NULL) { 5346 return false; // Intrinsic's stub is not implemented on this platform 5347 } 5348 const char* stubName = "squareToLen"; 5349 5350 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5351 5352 Node* x = argument(0); 5353 Node* len = argument(1); 5354 Node* z = argument(2); 5355 Node* zlen = argument(3); 5356 5357 const Type* x_type = x->Value(&_gvn); 5358 const Type* z_type = z->Value(&_gvn); 5359 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5360 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5361 if (top_x == NULL || top_x->klass() == NULL || 5362 top_z == NULL || top_z->klass() == NULL) { 5363 // failed array check 5364 return false; 5365 } 5366 5367 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5368 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5369 if (x_elem != T_INT || z_elem != T_INT) { 5370 return false; 5371 } 5372 5373 5374 Node* x_start = array_element_address(x, intcon(0), x_elem); 5375 Node* z_start = array_element_address(z, intcon(0), z_elem); 5376 5377 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5378 OptoRuntime::squareToLen_Type(), 5379 stubAddr, stubName, TypePtr::BOTTOM, 5380 x_start, len, z_start, zlen); 5381 5382 set_result(z); 5383 return true; 5384 } 5385 5386 //-------------inline_mulAdd------------------------------------------ 5387 bool LibraryCallKit::inline_mulAdd() { 5388 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5389 5390 address stubAddr = StubRoutines::mulAdd(); 5391 if (stubAddr == NULL) { 5392 return false; // Intrinsic's stub is not implemented on this platform 5393 } 5394 const char* stubName = "mulAdd"; 5395 5396 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5397 5398 Node* out = argument(0); 5399 Node* in = argument(1); 5400 Node* offset = argument(2); 5401 Node* len = argument(3); 5402 Node* k = argument(4); 5403 5404 const Type* out_type = out->Value(&_gvn); 5405 const Type* in_type = in->Value(&_gvn); 5406 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5407 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5408 if (top_out == NULL || top_out->klass() == NULL || 5409 top_in == NULL || top_in->klass() == NULL) { 5410 // failed array check 5411 return false; 5412 } 5413 5414 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5415 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5416 if (out_elem != T_INT || in_elem != T_INT) { 5417 return false; 5418 } 5419 5420 Node* outlen = load_array_length(out); 5421 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5422 Node* out_start = array_element_address(out, intcon(0), out_elem); 5423 Node* in_start = array_element_address(in, intcon(0), in_elem); 5424 5425 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5426 OptoRuntime::mulAdd_Type(), 5427 stubAddr, stubName, TypePtr::BOTTOM, 5428 out_start,in_start, new_offset, len, k); 5429 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5430 set_result(result); 5431 return true; 5432 } 5433 5434 5435 /** 5436 * Calculate CRC32 for byte. 5437 * int java.util.zip.CRC32.update(int crc, int b) 5438 */ 5439 bool LibraryCallKit::inline_updateCRC32() { 5440 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5441 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5442 // no receiver since it is static method 5443 Node* crc = argument(0); // type: int 5444 Node* b = argument(1); // type: int 5445 5446 /* 5447 * int c = ~ crc; 5448 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5449 * b = b ^ (c >>> 8); 5450 * crc = ~b; 5451 */ 5452 5453 Node* M1 = intcon(-1); 5454 crc = _gvn.transform(new XorINode(crc, M1)); 5455 Node* result = _gvn.transform(new XorINode(crc, b)); 5456 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5457 5458 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5459 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5460 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5461 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5462 5463 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5464 result = _gvn.transform(new XorINode(crc, result)); 5465 result = _gvn.transform(new XorINode(result, M1)); 5466 set_result(result); 5467 return true; 5468 } 5469 5470 /** 5471 * Calculate CRC32 for byte[] array. 5472 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5473 */ 5474 bool LibraryCallKit::inline_updateBytesCRC32() { 5475 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5476 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5477 // no receiver since it is static method 5478 Node* crc = argument(0); // type: int 5479 Node* src = argument(1); // type: oop 5480 Node* offset = argument(2); // type: int 5481 Node* length = argument(3); // type: int 5482 5483 const Type* src_type = src->Value(&_gvn); 5484 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5485 if (top_src == NULL || top_src->klass() == NULL) { 5486 // failed array check 5487 return false; 5488 } 5489 5490 // Figure out the size and type of the elements we will be copying. 5491 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5492 if (src_elem != T_BYTE) { 5493 return false; 5494 } 5495 5496 // 'src_start' points to src array + scaled offset 5497 Node* src_start = array_element_address(src, offset, src_elem); 5498 5499 // We assume that range check is done by caller. 5500 // TODO: generate range check (offset+length < src.length) in debug VM. 5501 5502 // Call the stub. 5503 address stubAddr = StubRoutines::updateBytesCRC32(); 5504 const char *stubName = "updateBytesCRC32"; 5505 5506 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5507 stubAddr, stubName, TypePtr::BOTTOM, 5508 crc, src_start, length); 5509 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5510 set_result(result); 5511 return true; 5512 } 5513 5514 /** 5515 * Calculate CRC32 for ByteBuffer. 5516 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5517 */ 5518 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5519 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5520 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5521 // no receiver since it is static method 5522 Node* crc = argument(0); // type: int 5523 Node* src = argument(1); // type: long 5524 Node* offset = argument(3); // type: int 5525 Node* length = argument(4); // type: int 5526 5527 src = ConvL2X(src); // adjust Java long to machine word 5528 Node* base = _gvn.transform(new CastX2PNode(src)); 5529 offset = ConvI2X(offset); 5530 5531 // 'src_start' points to src array + scaled offset 5532 Node* src_start = basic_plus_adr(top(), base, offset); 5533 5534 // Call the stub. 5535 address stubAddr = StubRoutines::updateBytesCRC32(); 5536 const char *stubName = "updateBytesCRC32"; 5537 5538 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5539 stubAddr, stubName, TypePtr::BOTTOM, 5540 crc, src_start, length); 5541 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5542 set_result(result); 5543 return true; 5544 } 5545 5546 //----------------------------inline_reference_get---------------------------- 5547 // public T java.lang.ref.Reference.get(); 5548 bool LibraryCallKit::inline_reference_get() { 5549 const int referent_offset = java_lang_ref_Reference::referent_offset; 5550 guarantee(referent_offset > 0, "should have already been set"); 5551 5552 // Get the argument: 5553 Node* reference_obj = null_check_receiver(); 5554 if (stopped()) return true; 5555 5556 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5557 5558 ciInstanceKlass* klass = env()->Object_klass(); 5559 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5560 5561 Node* no_ctrl = NULL; 5562 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5563 5564 // Use the pre-barrier to record the value in the referent field 5565 pre_barrier(false /* do_load */, 5566 control(), 5567 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5568 result /* pre_val */, 5569 T_OBJECT); 5570 5571 // Add memory barrier to prevent commoning reads from this field 5572 // across safepoint since GC can change its value. 5573 insert_mem_bar(Op_MemBarCPUOrder); 5574 5575 set_result(result); 5576 return true; 5577 } 5578 5579 5580 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5581 bool is_exact=true, bool is_static=false) { 5582 5583 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5584 assert(tinst != NULL, "obj is null"); 5585 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5586 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5587 5588 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5589 ciSymbol::make(fieldTypeString), 5590 is_static); 5591 if (field == NULL) return (Node *) NULL; 5592 assert (field != NULL, "undefined field"); 5593 5594 // Next code copied from Parse::do_get_xxx(): 5595 5596 // Compute address and memory type. 5597 int offset = field->offset_in_bytes(); 5598 bool is_vol = field->is_volatile(); 5599 ciType* field_klass = field->type(); 5600 assert(field_klass->is_loaded(), "should be loaded"); 5601 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5602 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5603 BasicType bt = field->layout_type(); 5604 5605 // Build the resultant type of the load 5606 const Type *type; 5607 if (bt == T_OBJECT) { 5608 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5609 } else { 5610 type = Type::get_const_basic_type(bt); 5611 } 5612 5613 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5614 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5615 } 5616 // Build the load. 5617 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5618 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5619 // If reference is volatile, prevent following memory ops from 5620 // floating up past the volatile read. Also prevents commoning 5621 // another volatile read. 5622 if (is_vol) { 5623 // Memory barrier includes bogus read of value to force load BEFORE membar 5624 insert_mem_bar(Op_MemBarAcquire, loadedField); 5625 } 5626 return loadedField; 5627 } 5628 5629 5630 //------------------------------inline_aescrypt_Block----------------------- 5631 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5632 address stubAddr; 5633 const char *stubName; 5634 assert(UseAES, "need AES instruction support"); 5635 5636 switch(id) { 5637 case vmIntrinsics::_aescrypt_encryptBlock: 5638 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5639 stubName = "aescrypt_encryptBlock"; 5640 break; 5641 case vmIntrinsics::_aescrypt_decryptBlock: 5642 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5643 stubName = "aescrypt_decryptBlock"; 5644 break; 5645 } 5646 if (stubAddr == NULL) return false; 5647 5648 Node* aescrypt_object = argument(0); 5649 Node* src = argument(1); 5650 Node* src_offset = argument(2); 5651 Node* dest = argument(3); 5652 Node* dest_offset = argument(4); 5653 5654 // (1) src and dest are arrays. 5655 const Type* src_type = src->Value(&_gvn); 5656 const Type* dest_type = dest->Value(&_gvn); 5657 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5658 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5659 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5660 5661 // for the quick and dirty code we will skip all the checks. 5662 // we are just trying to get the call to be generated. 5663 Node* src_start = src; 5664 Node* dest_start = dest; 5665 if (src_offset != NULL || dest_offset != NULL) { 5666 assert(src_offset != NULL && dest_offset != NULL, ""); 5667 src_start = array_element_address(src, src_offset, T_BYTE); 5668 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5669 } 5670 5671 // now need to get the start of its expanded key array 5672 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5673 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5674 if (k_start == NULL) return false; 5675 5676 if (Matcher::pass_original_key_for_aes()) { 5677 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5678 // compatibility issues between Java key expansion and SPARC crypto instructions 5679 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5680 if (original_k_start == NULL) return false; 5681 5682 // Call the stub. 5683 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5684 stubAddr, stubName, TypePtr::BOTTOM, 5685 src_start, dest_start, k_start, original_k_start); 5686 } else { 5687 // Call the stub. 5688 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5689 stubAddr, stubName, TypePtr::BOTTOM, 5690 src_start, dest_start, k_start); 5691 } 5692 5693 return true; 5694 } 5695 5696 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5697 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5698 address stubAddr; 5699 const char *stubName; 5700 5701 assert(UseAES, "need AES instruction support"); 5702 5703 switch(id) { 5704 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5705 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5706 stubName = "cipherBlockChaining_encryptAESCrypt"; 5707 break; 5708 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5709 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5710 stubName = "cipherBlockChaining_decryptAESCrypt"; 5711 break; 5712 } 5713 if (stubAddr == NULL) return false; 5714 5715 Node* cipherBlockChaining_object = argument(0); 5716 Node* src = argument(1); 5717 Node* src_offset = argument(2); 5718 Node* len = argument(3); 5719 Node* dest = argument(4); 5720 Node* dest_offset = argument(5); 5721 5722 // (1) src and dest are arrays. 5723 const Type* src_type = src->Value(&_gvn); 5724 const Type* dest_type = dest->Value(&_gvn); 5725 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5726 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5727 assert (top_src != NULL && top_src->klass() != NULL 5728 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5729 5730 // checks are the responsibility of the caller 5731 Node* src_start = src; 5732 Node* dest_start = dest; 5733 if (src_offset != NULL || dest_offset != NULL) { 5734 assert(src_offset != NULL && dest_offset != NULL, ""); 5735 src_start = array_element_address(src, src_offset, T_BYTE); 5736 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5737 } 5738 5739 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5740 // (because of the predicated logic executed earlier). 5741 // so we cast it here safely. 5742 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5743 5744 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5745 if (embeddedCipherObj == NULL) return false; 5746 5747 // cast it to what we know it will be at runtime 5748 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5749 assert(tinst != NULL, "CBC obj is null"); 5750 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5751 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5752 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5753 5754 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5755 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5756 const TypeOopPtr* xtype = aklass->as_instance_type(); 5757 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5758 aescrypt_object = _gvn.transform(aescrypt_object); 5759 5760 // we need to get the start of the aescrypt_object's expanded key array 5761 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5762 if (k_start == NULL) return false; 5763 5764 // similarly, get the start address of the r vector 5765 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5766 if (objRvec == NULL) return false; 5767 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5768 5769 Node* cbcCrypt; 5770 if (Matcher::pass_original_key_for_aes()) { 5771 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5772 // compatibility issues between Java key expansion and SPARC crypto instructions 5773 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5774 if (original_k_start == NULL) return false; 5775 5776 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5777 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5778 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5779 stubAddr, stubName, TypePtr::BOTTOM, 5780 src_start, dest_start, k_start, r_start, len, original_k_start); 5781 } else { 5782 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5783 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5784 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5785 stubAddr, stubName, TypePtr::BOTTOM, 5786 src_start, dest_start, k_start, r_start, len); 5787 } 5788 5789 // return cipher length (int) 5790 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5791 set_result(retvalue); 5792 return true; 5793 } 5794 5795 //------------------------------get_key_start_from_aescrypt_object----------------------- 5796 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5797 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5798 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5799 if (objAESCryptKey == NULL) return (Node *) NULL; 5800 5801 // now have the array, need to get the start address of the K array 5802 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5803 return k_start; 5804 } 5805 5806 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5807 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5808 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5809 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5810 if (objAESCryptKey == NULL) return (Node *) NULL; 5811 5812 // now have the array, need to get the start address of the lastKey array 5813 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5814 return original_k_start; 5815 } 5816 5817 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5818 // Return node representing slow path of predicate check. 5819 // the pseudo code we want to emulate with this predicate is: 5820 // for encryption: 5821 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5822 // for decryption: 5823 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5824 // note cipher==plain is more conservative than the original java code but that's OK 5825 // 5826 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5827 // The receiver was checked for NULL already. 5828 Node* objCBC = argument(0); 5829 5830 // Load embeddedCipher field of CipherBlockChaining object. 5831 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5832 5833 // get AESCrypt klass for instanceOf check 5834 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5835 // will have same classloader as CipherBlockChaining object 5836 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5837 assert(tinst != NULL, "CBCobj is null"); 5838 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5839 5840 // we want to do an instanceof comparison against the AESCrypt class 5841 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5842 if (!klass_AESCrypt->is_loaded()) { 5843 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5844 Node* ctrl = control(); 5845 set_control(top()); // no regular fast path 5846 return ctrl; 5847 } 5848 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5849 5850 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5851 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5852 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5853 5854 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5855 5856 // for encryption, we are done 5857 if (!decrypting) 5858 return instof_false; // even if it is NULL 5859 5860 // for decryption, we need to add a further check to avoid 5861 // taking the intrinsic path when cipher and plain are the same 5862 // see the original java code for why. 5863 RegionNode* region = new RegionNode(3); 5864 region->init_req(1, instof_false); 5865 Node* src = argument(1); 5866 Node* dest = argument(4); 5867 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5868 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5869 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5870 region->init_req(2, src_dest_conjoint); 5871 5872 record_for_igvn(region); 5873 return _gvn.transform(region); 5874 } 5875 5876 //------------------------------inline_ghash_processBlocks 5877 bool LibraryCallKit::inline_ghash_processBlocks() { 5878 address stubAddr; 5879 const char *stubName; 5880 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 5881 5882 stubAddr = StubRoutines::ghash_processBlocks(); 5883 stubName = "ghash_processBlocks"; 5884 5885 Node* data = argument(0); 5886 Node* offset = argument(1); 5887 Node* len = argument(2); 5888 Node* state = argument(3); 5889 Node* subkeyH = argument(4); 5890 5891 Node* state_start = array_element_address(state, intcon(0), T_LONG); 5892 assert(state_start, "state is NULL"); 5893 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 5894 assert(subkeyH_start, "subkeyH is NULL"); 5895 Node* data_start = array_element_address(data, offset, T_BYTE); 5896 assert(data_start, "data is NULL"); 5897 5898 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 5899 OptoRuntime::ghash_processBlocks_Type(), 5900 stubAddr, stubName, TypePtr::BOTTOM, 5901 state_start, subkeyH_start, data_start, len); 5902 return true; 5903 } 5904 5905 //------------------------------inline_sha_implCompress----------------------- 5906 // 5907 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 5908 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 5909 // 5910 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 5911 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 5912 // 5913 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 5914 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 5915 // 5916 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 5917 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 5918 5919 Node* sha_obj = argument(0); 5920 Node* src = argument(1); // type oop 5921 Node* ofs = argument(2); // type int 5922 5923 const Type* src_type = src->Value(&_gvn); 5924 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5925 if (top_src == NULL || top_src->klass() == NULL) { 5926 // failed array check 5927 return false; 5928 } 5929 // Figure out the size and type of the elements we will be copying. 5930 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5931 if (src_elem != T_BYTE) { 5932 return false; 5933 } 5934 // 'src_start' points to src array + offset 5935 Node* src_start = array_element_address(src, ofs, src_elem); 5936 Node* state = NULL; 5937 address stubAddr; 5938 const char *stubName; 5939 5940 switch(id) { 5941 case vmIntrinsics::_sha_implCompress: 5942 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 5943 state = get_state_from_sha_object(sha_obj); 5944 stubAddr = StubRoutines::sha1_implCompress(); 5945 stubName = "sha1_implCompress"; 5946 break; 5947 case vmIntrinsics::_sha2_implCompress: 5948 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 5949 state = get_state_from_sha_object(sha_obj); 5950 stubAddr = StubRoutines::sha256_implCompress(); 5951 stubName = "sha256_implCompress"; 5952 break; 5953 case vmIntrinsics::_sha5_implCompress: 5954 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 5955 state = get_state_from_sha5_object(sha_obj); 5956 stubAddr = StubRoutines::sha512_implCompress(); 5957 stubName = "sha512_implCompress"; 5958 break; 5959 default: 5960 fatal_unexpected_iid(id); 5961 return false; 5962 } 5963 if (state == NULL) return false; 5964 5965 // Call the stub. 5966 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 5967 stubAddr, stubName, TypePtr::BOTTOM, 5968 src_start, state); 5969 5970 return true; 5971 } 5972 5973 //------------------------------inline_digestBase_implCompressMB----------------------- 5974 // 5975 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 5976 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 5977 // 5978 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 5979 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5980 "need SHA1/SHA256/SHA512 instruction support"); 5981 assert((uint)predicate < 3, "sanity"); 5982 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 5983 5984 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 5985 Node* src = argument(1); // byte[] array 5986 Node* ofs = argument(2); // type int 5987 Node* limit = argument(3); // type int 5988 5989 const Type* src_type = src->Value(&_gvn); 5990 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5991 if (top_src == NULL || top_src->klass() == NULL) { 5992 // failed array check 5993 return false; 5994 } 5995 // Figure out the size and type of the elements we will be copying. 5996 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5997 if (src_elem != T_BYTE) { 5998 return false; 5999 } 6000 // 'src_start' points to src array + offset 6001 Node* src_start = array_element_address(src, ofs, src_elem); 6002 6003 const char* klass_SHA_name = NULL; 6004 const char* stub_name = NULL; 6005 address stub_addr = NULL; 6006 bool long_state = false; 6007 6008 switch (predicate) { 6009 case 0: 6010 if (UseSHA1Intrinsics) { 6011 klass_SHA_name = "sun/security/provider/SHA"; 6012 stub_name = "sha1_implCompressMB"; 6013 stub_addr = StubRoutines::sha1_implCompressMB(); 6014 } 6015 break; 6016 case 1: 6017 if (UseSHA256Intrinsics) { 6018 klass_SHA_name = "sun/security/provider/SHA2"; 6019 stub_name = "sha256_implCompressMB"; 6020 stub_addr = StubRoutines::sha256_implCompressMB(); 6021 } 6022 break; 6023 case 2: 6024 if (UseSHA512Intrinsics) { 6025 klass_SHA_name = "sun/security/provider/SHA5"; 6026 stub_name = "sha512_implCompressMB"; 6027 stub_addr = StubRoutines::sha512_implCompressMB(); 6028 long_state = true; 6029 } 6030 break; 6031 default: 6032 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6033 } 6034 if (klass_SHA_name != NULL) { 6035 // get DigestBase klass to lookup for SHA klass 6036 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6037 assert(tinst != NULL, "digestBase_obj is not instance???"); 6038 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6039 6040 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6041 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6042 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6043 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6044 } 6045 return false; 6046 } 6047 //------------------------------inline_sha_implCompressMB----------------------- 6048 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6049 bool long_state, address stubAddr, const char *stubName, 6050 Node* src_start, Node* ofs, Node* limit) { 6051 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6052 const TypeOopPtr* xtype = aklass->as_instance_type(); 6053 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6054 sha_obj = _gvn.transform(sha_obj); 6055 6056 Node* state; 6057 if (long_state) { 6058 state = get_state_from_sha5_object(sha_obj); 6059 } else { 6060 state = get_state_from_sha_object(sha_obj); 6061 } 6062 if (state == NULL) return false; 6063 6064 // Call the stub. 6065 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6066 OptoRuntime::digestBase_implCompressMB_Type(), 6067 stubAddr, stubName, TypePtr::BOTTOM, 6068 src_start, state, ofs, limit); 6069 // return ofs (int) 6070 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6071 set_result(result); 6072 6073 return true; 6074 } 6075 6076 //------------------------------get_state_from_sha_object----------------------- 6077 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6078 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6079 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6080 if (sha_state == NULL) return (Node *) NULL; 6081 6082 // now have the array, need to get the start address of the state array 6083 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6084 return state; 6085 } 6086 6087 //------------------------------get_state_from_sha5_object----------------------- 6088 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6089 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6090 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6091 if (sha_state == NULL) return (Node *) NULL; 6092 6093 // now have the array, need to get the start address of the state array 6094 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6095 return state; 6096 } 6097 6098 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6099 // Return node representing slow path of predicate check. 6100 // the pseudo code we want to emulate with this predicate is: 6101 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6102 // 6103 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6104 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6105 "need SHA1/SHA256/SHA512 instruction support"); 6106 assert((uint)predicate < 3, "sanity"); 6107 6108 // The receiver was checked for NULL already. 6109 Node* digestBaseObj = argument(0); 6110 6111 // get DigestBase klass for instanceOf check 6112 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6113 assert(tinst != NULL, "digestBaseObj is null"); 6114 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6115 6116 const char* klass_SHA_name = NULL; 6117 switch (predicate) { 6118 case 0: 6119 if (UseSHA1Intrinsics) { 6120 // we want to do an instanceof comparison against the SHA class 6121 klass_SHA_name = "sun/security/provider/SHA"; 6122 } 6123 break; 6124 case 1: 6125 if (UseSHA256Intrinsics) { 6126 // we want to do an instanceof comparison against the SHA2 class 6127 klass_SHA_name = "sun/security/provider/SHA2"; 6128 } 6129 break; 6130 case 2: 6131 if (UseSHA512Intrinsics) { 6132 // we want to do an instanceof comparison against the SHA5 class 6133 klass_SHA_name = "sun/security/provider/SHA5"; 6134 } 6135 break; 6136 default: 6137 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6138 } 6139 6140 ciKlass* klass_SHA = NULL; 6141 if (klass_SHA_name != NULL) { 6142 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6143 } 6144 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6145 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6146 Node* ctrl = control(); 6147 set_control(top()); // no intrinsic path 6148 return ctrl; 6149 } 6150 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6151 6152 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6153 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6154 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6155 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6156 6157 return instof_false; // even if it is NULL 6158 } 6159 6160 bool LibraryCallKit::inline_profileBoolean() { 6161 Node* counts = argument(1); 6162 const TypeAryPtr* ary = NULL; 6163 ciArray* aobj = NULL; 6164 if (counts->is_Con() 6165 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6166 && (aobj = ary->const_oop()->as_array()) != NULL 6167 && (aobj->length() == 2)) { 6168 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6169 jint false_cnt = aobj->element_value(0).as_int(); 6170 jint true_cnt = aobj->element_value(1).as_int(); 6171 6172 if (C->log() != NULL) { 6173 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6174 false_cnt, true_cnt); 6175 } 6176 6177 if (false_cnt + true_cnt == 0) { 6178 // According to profile, never executed. 6179 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6180 Deoptimization::Action_reinterpret); 6181 return true; 6182 } 6183 6184 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6185 // is a number of each value occurrences. 6186 Node* result = argument(0); 6187 if (false_cnt == 0 || true_cnt == 0) { 6188 // According to profile, one value has been never seen. 6189 int expected_val = (false_cnt == 0) ? 1 : 0; 6190 6191 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6192 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6193 6194 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6195 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6196 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6197 6198 { // Slow path: uncommon trap for never seen value and then reexecute 6199 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6200 // the value has been seen at least once. 6201 PreserveJVMState pjvms(this); 6202 PreserveReexecuteState preexecs(this); 6203 jvms()->set_should_reexecute(true); 6204 6205 set_control(slow_path); 6206 set_i_o(i_o()); 6207 6208 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6209 Deoptimization::Action_reinterpret); 6210 } 6211 // The guard for never seen value enables sharpening of the result and 6212 // returning a constant. It allows to eliminate branches on the same value 6213 // later on. 6214 set_control(fast_path); 6215 result = intcon(expected_val); 6216 } 6217 // Stop profiling. 6218 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6219 // By replacing method body with profile data (represented as ProfileBooleanNode 6220 // on IR level) we effectively disable profiling. 6221 // It enables full speed execution once optimized code is generated. 6222 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6223 C->record_for_igvn(profile); 6224 set_result(profile); 6225 return true; 6226 } else { 6227 // Continue profiling. 6228 // Profile data isn't available at the moment. So, execute method's bytecode version. 6229 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6230 // is compiled and counters aren't available since corresponding MethodHandle 6231 // isn't a compile-time constant. 6232 return false; 6233 } 6234 } 6235 6236 bool LibraryCallKit::inline_isCompileConstant() { 6237 Node* n = argument(0); 6238 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6239 return true; 6240 }