1 /* 2 * Copyright (c) 1999, 2014, 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/callGenerator.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/cfgnode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/countbitsnode.hpp" 38 #include "opto/intrinsicnode.hpp" 39 #include "opto/idealKit.hpp" 40 #include "opto/mathexactnode.hpp" 41 #include "opto/movenode.hpp" 42 #include "opto/mulnode.hpp" 43 #include "opto/narrowptrnode.hpp" 44 #include "opto/opaquenode.hpp" 45 #include "opto/parse.hpp" 46 #include "opto/runtime.hpp" 47 #include "opto/subnode.hpp" 48 #include "prims/nativeLookup.hpp" 49 #include "runtime/sharedRuntime.hpp" 50 #include "trace/traceMacros.hpp" 51 52 class LibraryIntrinsic : public InlineCallGenerator { 53 // Extend the set of intrinsics known to the runtime: 54 public: 55 private: 56 bool _is_virtual; 57 bool _does_virtual_dispatch; 58 int8_t _predicates_count; // Intrinsic is predicated by several conditions 59 int8_t _last_predicate; // Last generated predicate 60 vmIntrinsics::ID _intrinsic_id; 61 62 public: 63 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 64 : InlineCallGenerator(m), 65 _is_virtual(is_virtual), 66 _does_virtual_dispatch(does_virtual_dispatch), 67 _predicates_count((int8_t)predicates_count), 68 _last_predicate((int8_t)-1), 69 _intrinsic_id(id) 70 { 71 } 72 virtual bool is_intrinsic() const { return true; } 73 virtual bool is_virtual() const { return _is_virtual; } 74 virtual bool is_predicated() const { return _predicates_count > 0; } 75 virtual int predicates_count() const { return _predicates_count; } 76 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 77 virtual JVMState* generate(JVMState* jvms); 78 virtual Node* generate_predicate(JVMState* jvms, int predicate); 79 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 80 }; 81 82 83 // Local helper class for LibraryIntrinsic: 84 class LibraryCallKit : public GraphKit { 85 private: 86 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 87 Node* _result; // the result node, if any 88 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 89 90 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 91 92 public: 93 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 94 : GraphKit(jvms), 95 _intrinsic(intrinsic), 96 _result(NULL) 97 { 98 // Check if this is a root compile. In that case we don't have a caller. 99 if (!jvms->has_method()) { 100 _reexecute_sp = sp(); 101 } else { 102 // Find out how many arguments the interpreter needs when deoptimizing 103 // and save the stack pointer value so it can used by uncommon_trap. 104 // We find the argument count by looking at the declared signature. 105 bool ignored_will_link; 106 ciSignature* declared_signature = NULL; 107 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 108 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 109 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 110 } 111 } 112 113 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 114 115 ciMethod* caller() const { return jvms()->method(); } 116 int bci() const { return jvms()->bci(); } 117 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 118 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 119 ciMethod* callee() const { return _intrinsic->method(); } 120 121 bool try_to_inline(int predicate); 122 Node* try_to_predicate(int predicate); 123 124 void push_result() { 125 // Push the result onto the stack. 126 if (!stopped() && result() != NULL) { 127 BasicType bt = result()->bottom_type()->basic_type(); 128 push_node(bt, result()); 129 } 130 } 131 132 private: 133 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 134 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 135 } 136 137 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 138 void set_result(RegionNode* region, PhiNode* value); 139 Node* result() { return _result; } 140 141 virtual int reexecute_sp() { return _reexecute_sp; } 142 143 // Helper functions to inline natives 144 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 145 Node* generate_slow_guard(Node* test, RegionNode* region); 146 Node* generate_fair_guard(Node* test, RegionNode* region); 147 Node* generate_negative_guard(Node* index, RegionNode* region, 148 // resulting CastII of index: 149 Node* *pos_index = NULL); 150 Node* generate_limit_guard(Node* offset, Node* subseq_length, 151 Node* array_length, 152 RegionNode* region); 153 Node* generate_current_thread(Node* &tls_output); 154 Node* load_mirror_from_klass(Node* klass); 155 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 156 RegionNode* region, int null_path, 157 int offset); 158 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 159 RegionNode* region, int null_path) { 160 int offset = java_lang_Class::klass_offset_in_bytes(); 161 return load_klass_from_mirror_common(mirror, never_see_null, 162 region, null_path, 163 offset); 164 } 165 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 166 RegionNode* region, int null_path) { 167 int offset = java_lang_Class::array_klass_offset_in_bytes(); 168 return load_klass_from_mirror_common(mirror, never_see_null, 169 region, null_path, 170 offset); 171 } 172 Node* generate_access_flags_guard(Node* kls, 173 int modifier_mask, int modifier_bits, 174 RegionNode* region); 175 Node* generate_interface_guard(Node* kls, RegionNode* region); 176 Node* generate_array_guard(Node* kls, RegionNode* region) { 177 return generate_array_guard_common(kls, region, false, false); 178 } 179 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 180 return generate_array_guard_common(kls, region, false, true); 181 } 182 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 183 return generate_array_guard_common(kls, region, true, false); 184 } 185 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 186 return generate_array_guard_common(kls, region, true, true); 187 } 188 Node* generate_array_guard_common(Node* kls, RegionNode* region, 189 bool obj_array, bool not_array); 190 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 191 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 192 bool is_virtual = false, bool is_static = false); 193 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 194 return generate_method_call(method_id, false, true); 195 } 196 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 197 return generate_method_call(method_id, true, false); 198 } 199 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 200 201 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 202 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 203 bool inline_string_compareTo(); 204 bool inline_string_indexOf(); 205 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 206 bool inline_string_equals(); 207 Node* round_double_node(Node* n); 208 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 209 bool inline_math_native(vmIntrinsics::ID id); 210 bool inline_trig(vmIntrinsics::ID id); 211 bool inline_math(vmIntrinsics::ID id); 212 template <typename OverflowOp> 213 bool inline_math_overflow(Node* arg1, Node* arg2); 214 void inline_math_mathExact(Node* math, Node* test); 215 bool inline_math_addExactI(bool is_increment); 216 bool inline_math_addExactL(bool is_increment); 217 bool inline_math_multiplyExactI(); 218 bool inline_math_multiplyExactL(); 219 bool inline_math_negateExactI(); 220 bool inline_math_negateExactL(); 221 bool inline_math_subtractExactI(bool is_decrement); 222 bool inline_math_subtractExactL(bool is_decrement); 223 bool inline_exp(); 224 bool inline_pow(); 225 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 226 bool inline_min_max(vmIntrinsics::ID id); 227 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 228 // This returns Type::AnyPtr, RawPtr, or OopPtr. 229 int classify_unsafe_addr(Node* &base, Node* &offset); 230 Node* make_unsafe_address(Node* base, Node* offset); 231 // Helper for inline_unsafe_access. 232 // Generates the guards that check whether the result of 233 // Unsafe.getObject should be recorded in an SATB log buffer. 234 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 235 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 236 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 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 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 266 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 267 bool inline_unsafe_ordered_store(BasicType type); 268 bool inline_unsafe_fence(vmIntrinsics::ID id); 269 bool inline_fp_conversions(vmIntrinsics::ID id); 270 bool inline_number_methods(vmIntrinsics::ID id); 271 bool inline_reference_get(); 272 bool inline_Class_cast(); 273 bool inline_aescrypt_Block(vmIntrinsics::ID id); 274 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 275 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 276 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 277 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 278 bool inline_sha_implCompress(vmIntrinsics::ID id); 279 bool inline_digestBase_implCompressMB(int predicate); 280 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 281 bool long_state, address stubAddr, const char *stubName, 282 Node* src_start, Node* ofs, Node* limit); 283 Node* get_state_from_sha_object(Node *sha_object); 284 Node* get_state_from_sha5_object(Node *sha_object); 285 Node* inline_digestBase_implCompressMB_predicate(int predicate); 286 bool inline_encodeISOArray(); 287 bool inline_updateCRC32(); 288 bool inline_updateBytesCRC32(); 289 bool inline_updateByteBufferCRC32(); 290 bool inline_multiplyToLen(); 291 292 bool inline_profileBranch(); 293 }; 294 295 296 //---------------------------make_vm_intrinsic---------------------------- 297 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 298 vmIntrinsics::ID id = m->intrinsic_id(); 299 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 300 301 ccstr disable_intr = NULL; 302 303 if ((DisableIntrinsic[0] != '\0' 304 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 305 (method_has_option_value("DisableIntrinsic", disable_intr) 306 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 307 // disabled by a user request on the command line: 308 // example: -XX:DisableIntrinsic=_hashCode,_getClass 309 return NULL; 310 } 311 312 if (!m->is_loaded()) { 313 // do not attempt to inline unloaded methods 314 return NULL; 315 } 316 317 // Only a few intrinsics implement a virtual dispatch. 318 // They are expensive calls which are also frequently overridden. 319 if (is_virtual) { 320 switch (id) { 321 case vmIntrinsics::_hashCode: 322 case vmIntrinsics::_clone: 323 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 324 break; 325 default: 326 return NULL; 327 } 328 } 329 330 // -XX:-InlineNatives disables nearly all intrinsics: 331 if (!InlineNatives) { 332 switch (id) { 333 case vmIntrinsics::_indexOf: 334 case vmIntrinsics::_compareTo: 335 case vmIntrinsics::_equals: 336 case vmIntrinsics::_equalsC: 337 case vmIntrinsics::_getAndAddInt: 338 case vmIntrinsics::_getAndAddLong: 339 case vmIntrinsics::_getAndSetInt: 340 case vmIntrinsics::_getAndSetLong: 341 case vmIntrinsics::_getAndSetObject: 342 case vmIntrinsics::_loadFence: 343 case vmIntrinsics::_storeFence: 344 case vmIntrinsics::_fullFence: 345 break; // InlineNatives does not control String.compareTo 346 case vmIntrinsics::_Reference_get: 347 break; // InlineNatives does not control Reference.get 348 default: 349 return NULL; 350 } 351 } 352 353 int predicates = 0; 354 bool does_virtual_dispatch = false; 355 356 switch (id) { 357 case vmIntrinsics::_compareTo: 358 if (!SpecialStringCompareTo) return NULL; 359 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 360 break; 361 case vmIntrinsics::_indexOf: 362 if (!SpecialStringIndexOf) return NULL; 363 break; 364 case vmIntrinsics::_equals: 365 if (!SpecialStringEquals) return NULL; 366 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 367 break; 368 case vmIntrinsics::_equalsC: 369 if (!SpecialArraysEquals) return NULL; 370 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 371 break; 372 case vmIntrinsics::_arraycopy: 373 if (!InlineArrayCopy) return NULL; 374 break; 375 case vmIntrinsics::_copyMemory: 376 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 377 if (!InlineArrayCopy) return NULL; 378 break; 379 case vmIntrinsics::_hashCode: 380 if (!InlineObjectHash) return NULL; 381 does_virtual_dispatch = true; 382 break; 383 case vmIntrinsics::_clone: 384 does_virtual_dispatch = true; 385 case vmIntrinsics::_copyOf: 386 case vmIntrinsics::_copyOfRange: 387 if (!InlineObjectCopy) return NULL; 388 // These also use the arraycopy intrinsic mechanism: 389 if (!InlineArrayCopy) return NULL; 390 break; 391 case vmIntrinsics::_encodeISOArray: 392 if (!SpecialEncodeISOArray) return NULL; 393 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 394 break; 395 case vmIntrinsics::_checkIndex: 396 // We do not intrinsify this. The optimizer does fine with it. 397 return NULL; 398 399 case vmIntrinsics::_getCallerClass: 400 if (!InlineReflectionGetCallerClass) return NULL; 401 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 402 break; 403 404 case vmIntrinsics::_bitCount_i: 405 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 406 break; 407 408 case vmIntrinsics::_bitCount_l: 409 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 410 break; 411 412 case vmIntrinsics::_numberOfLeadingZeros_i: 413 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 414 break; 415 416 case vmIntrinsics::_numberOfLeadingZeros_l: 417 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 418 break; 419 420 case vmIntrinsics::_numberOfTrailingZeros_i: 421 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 422 break; 423 424 case vmIntrinsics::_numberOfTrailingZeros_l: 425 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 426 break; 427 428 case vmIntrinsics::_reverseBytes_c: 429 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 430 break; 431 case vmIntrinsics::_reverseBytes_s: 432 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 433 break; 434 case vmIntrinsics::_reverseBytes_i: 435 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 436 break; 437 case vmIntrinsics::_reverseBytes_l: 438 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 439 break; 440 441 case vmIntrinsics::_Reference_get: 442 // Use the intrinsic version of Reference.get() so that the value in 443 // the referent field can be registered by the G1 pre-barrier code. 444 // Also add memory barrier to prevent commoning reads from this field 445 // across safepoint since GC can change it value. 446 break; 447 448 case vmIntrinsics::_compareAndSwapObject: 449 #ifdef _LP64 450 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 451 #endif 452 break; 453 454 case vmIntrinsics::_compareAndSwapLong: 455 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 456 break; 457 458 case vmIntrinsics::_getAndAddInt: 459 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 460 break; 461 462 case vmIntrinsics::_getAndAddLong: 463 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 464 break; 465 466 case vmIntrinsics::_getAndSetInt: 467 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 468 break; 469 470 case vmIntrinsics::_getAndSetLong: 471 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 472 break; 473 474 case vmIntrinsics::_getAndSetObject: 475 #ifdef _LP64 476 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 477 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 478 break; 479 #else 480 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 481 break; 482 #endif 483 484 case vmIntrinsics::_aescrypt_encryptBlock: 485 case vmIntrinsics::_aescrypt_decryptBlock: 486 if (!UseAESIntrinsics) return NULL; 487 break; 488 489 case vmIntrinsics::_multiplyToLen: 490 if (!UseMultiplyToLenIntrinsic) return NULL; 491 break; 492 493 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 494 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 495 if (!UseAESIntrinsics) return NULL; 496 // these two require the predicated logic 497 predicates = 1; 498 break; 499 500 case vmIntrinsics::_sha_implCompress: 501 if (!UseSHA1Intrinsics) return NULL; 502 break; 503 504 case vmIntrinsics::_sha2_implCompress: 505 if (!UseSHA256Intrinsics) return NULL; 506 break; 507 508 case vmIntrinsics::_sha5_implCompress: 509 if (!UseSHA512Intrinsics) return NULL; 510 break; 511 512 case vmIntrinsics::_digestBase_implCompressMB: 513 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 514 predicates = 3; 515 break; 516 517 case vmIntrinsics::_updateCRC32: 518 case vmIntrinsics::_updateBytesCRC32: 519 case vmIntrinsics::_updateByteBufferCRC32: 520 if (!UseCRC32Intrinsics) return NULL; 521 break; 522 523 case vmIntrinsics::_incrementExactI: 524 case vmIntrinsics::_addExactI: 525 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 526 break; 527 case vmIntrinsics::_incrementExactL: 528 case vmIntrinsics::_addExactL: 529 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 530 break; 531 case vmIntrinsics::_decrementExactI: 532 case vmIntrinsics::_subtractExactI: 533 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 534 break; 535 case vmIntrinsics::_decrementExactL: 536 case vmIntrinsics::_subtractExactL: 537 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 538 break; 539 case vmIntrinsics::_negateExactI: 540 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 541 break; 542 case vmIntrinsics::_negateExactL: 543 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 544 break; 545 case vmIntrinsics::_multiplyExactI: 546 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 547 break; 548 case vmIntrinsics::_multiplyExactL: 549 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 550 break; 551 552 default: 553 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 554 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 555 break; 556 } 557 558 // -XX:-InlineClassNatives disables natives from the Class class. 559 // The flag applies to all reflective calls, notably Array.newArray 560 // (visible to Java programmers as Array.newInstance). 561 if (m->holder()->name() == ciSymbol::java_lang_Class() || 562 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 563 if (!InlineClassNatives) return NULL; 564 } 565 566 // -XX:-InlineThreadNatives disables natives from the Thread class. 567 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 568 if (!InlineThreadNatives) return NULL; 569 } 570 571 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 572 if (m->holder()->name() == ciSymbol::java_lang_Math() || 573 m->holder()->name() == ciSymbol::java_lang_Float() || 574 m->holder()->name() == ciSymbol::java_lang_Double()) { 575 if (!InlineMathNatives) return NULL; 576 } 577 578 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 579 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 580 if (!InlineUnsafeOps) return NULL; 581 } 582 583 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 584 } 585 586 //----------------------register_library_intrinsics----------------------- 587 // Initialize this file's data structures, for each Compile instance. 588 void Compile::register_library_intrinsics() { 589 // Nothing to do here. 590 } 591 592 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 593 LibraryCallKit kit(jvms, this); 594 Compile* C = kit.C; 595 int nodes = C->unique(); 596 #ifndef PRODUCT 597 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 598 char buf[1000]; 599 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 600 tty->print_cr("Intrinsic %s", str); 601 } 602 #endif 603 ciMethod* callee = kit.callee(); 604 const int bci = kit.bci(); 605 606 // Try to inline the intrinsic. 607 if (kit.try_to_inline(_last_predicate)) { 608 if (C->print_intrinsics() || C->print_inlining()) { 609 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 610 } 611 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 612 if (C->log()) { 613 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 614 vmIntrinsics::name_at(intrinsic_id()), 615 (is_virtual() ? " virtual='1'" : ""), 616 C->unique() - nodes); 617 } 618 // Push the result from the inlined method onto the stack. 619 kit.push_result(); 620 C->print_inlining_update(this); 621 return kit.transfer_exceptions_into_jvms(); 622 } 623 624 // The intrinsic bailed out 625 if (C->print_intrinsics() || C->print_inlining()) { 626 if (jvms->has_method()) { 627 // Not a root compile. 628 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 629 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 630 } else { 631 // Root compile 632 tty->print("Did not generate intrinsic %s%s at bci:%d in", 633 vmIntrinsics::name_at(intrinsic_id()), 634 (is_virtual() ? " (virtual)" : ""), bci); 635 } 636 } 637 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 638 C->print_inlining_update(this); 639 return NULL; 640 } 641 642 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 643 LibraryCallKit kit(jvms, this); 644 Compile* C = kit.C; 645 int nodes = C->unique(); 646 _last_predicate = predicate; 647 #ifndef PRODUCT 648 assert(is_predicated() && predicate < predicates_count(), "sanity"); 649 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 650 char buf[1000]; 651 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 652 tty->print_cr("Predicate for intrinsic %s", str); 653 } 654 #endif 655 ciMethod* callee = kit.callee(); 656 const int bci = kit.bci(); 657 658 Node* slow_ctl = kit.try_to_predicate(predicate); 659 if (!kit.failing()) { 660 if (C->print_intrinsics() || C->print_inlining()) { 661 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 662 } 663 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 664 if (C->log()) { 665 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 666 vmIntrinsics::name_at(intrinsic_id()), 667 (is_virtual() ? " virtual='1'" : ""), 668 C->unique() - nodes); 669 } 670 return slow_ctl; // Could be NULL if the check folds. 671 } 672 673 // The intrinsic bailed out 674 if (C->print_intrinsics() || C->print_inlining()) { 675 if (jvms->has_method()) { 676 // Not a root compile. 677 const char* msg = "failed to generate predicate for intrinsic"; 678 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 679 } else { 680 // Root compile 681 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 682 vmIntrinsics::name_at(intrinsic_id()), 683 (is_virtual() ? " (virtual)" : ""), bci); 684 } 685 } 686 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 687 return NULL; 688 } 689 690 bool LibraryCallKit::try_to_inline(int predicate) { 691 // Handle symbolic names for otherwise undistinguished boolean switches: 692 const bool is_store = true; 693 const bool is_native_ptr = true; 694 const bool is_static = true; 695 const bool is_volatile = true; 696 697 if (!jvms()->has_method()) { 698 // Root JVMState has a null method. 699 assert(map()->memory()->Opcode() == Op_Parm, ""); 700 // Insert the memory aliasing node 701 set_all_memory(reset_memory()); 702 } 703 assert(merged_memory(), ""); 704 705 706 switch (intrinsic_id()) { 707 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 708 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 709 case vmIntrinsics::_getClass: return inline_native_getClass(); 710 711 case vmIntrinsics::_dsin: 712 case vmIntrinsics::_dcos: 713 case vmIntrinsics::_dtan: 714 case vmIntrinsics::_dabs: 715 case vmIntrinsics::_datan2: 716 case vmIntrinsics::_dsqrt: 717 case vmIntrinsics::_dexp: 718 case vmIntrinsics::_dlog: 719 case vmIntrinsics::_dlog10: 720 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 721 722 case vmIntrinsics::_min: 723 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 724 725 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 726 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 727 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 728 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 729 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 730 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 731 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 732 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 733 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 734 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 735 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 736 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 737 738 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 739 740 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 741 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 742 case vmIntrinsics::_equals: return inline_string_equals(); 743 744 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 745 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 746 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 747 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 748 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 749 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 750 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 751 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 752 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 753 754 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 755 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 756 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 757 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 758 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 759 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 760 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 761 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 762 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 763 764 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 765 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 766 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 767 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 768 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 769 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 770 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 771 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 772 773 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 774 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 775 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 776 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 777 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 778 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 779 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 780 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 781 782 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 783 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 784 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 785 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 786 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 787 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 788 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 789 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 790 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 791 792 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 793 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 794 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 795 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 796 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 797 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 798 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 799 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 800 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 801 802 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 803 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 804 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 805 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 806 807 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 808 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 809 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 810 811 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 812 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 813 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 814 815 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 816 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 817 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 818 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 819 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 820 821 case vmIntrinsics::_loadFence: 822 case vmIntrinsics::_storeFence: 823 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 824 825 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 826 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 827 828 #ifdef TRACE_HAVE_INTRINSICS 829 case vmIntrinsics::_classID: return inline_native_classID(); 830 case vmIntrinsics::_threadID: return inline_native_threadID(); 831 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 832 #endif 833 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 834 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 835 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 836 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 837 case vmIntrinsics::_newArray: return inline_native_newArray(); 838 case vmIntrinsics::_getLength: return inline_native_getLength(); 839 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 840 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 841 case vmIntrinsics::_equalsC: return inline_array_equals(); 842 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 843 844 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 845 846 case vmIntrinsics::_isInstance: 847 case vmIntrinsics::_getModifiers: 848 case vmIntrinsics::_isInterface: 849 case vmIntrinsics::_isArray: 850 case vmIntrinsics::_isPrimitive: 851 case vmIntrinsics::_getSuperclass: 852 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 853 854 case vmIntrinsics::_floatToRawIntBits: 855 case vmIntrinsics::_floatToIntBits: 856 case vmIntrinsics::_intBitsToFloat: 857 case vmIntrinsics::_doubleToRawLongBits: 858 case vmIntrinsics::_doubleToLongBits: 859 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 860 861 case vmIntrinsics::_numberOfLeadingZeros_i: 862 case vmIntrinsics::_numberOfLeadingZeros_l: 863 case vmIntrinsics::_numberOfTrailingZeros_i: 864 case vmIntrinsics::_numberOfTrailingZeros_l: 865 case vmIntrinsics::_bitCount_i: 866 case vmIntrinsics::_bitCount_l: 867 case vmIntrinsics::_reverseBytes_i: 868 case vmIntrinsics::_reverseBytes_l: 869 case vmIntrinsics::_reverseBytes_s: 870 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 871 872 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 873 874 case vmIntrinsics::_Reference_get: return inline_reference_get(); 875 876 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 877 878 case vmIntrinsics::_aescrypt_encryptBlock: 879 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 880 881 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 882 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 883 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 884 885 case vmIntrinsics::_sha_implCompress: 886 case vmIntrinsics::_sha2_implCompress: 887 case vmIntrinsics::_sha5_implCompress: 888 return inline_sha_implCompress(intrinsic_id()); 889 890 case vmIntrinsics::_digestBase_implCompressMB: 891 return inline_digestBase_implCompressMB(predicate); 892 893 case vmIntrinsics::_multiplyToLen: 894 return inline_multiplyToLen(); 895 896 case vmIntrinsics::_encodeISOArray: 897 return inline_encodeISOArray(); 898 899 case vmIntrinsics::_updateCRC32: 900 return inline_updateCRC32(); 901 case vmIntrinsics::_updateBytesCRC32: 902 return inline_updateBytesCRC32(); 903 case vmIntrinsics::_updateByteBufferCRC32: 904 return inline_updateByteBufferCRC32(); 905 906 case vmIntrinsics::_profileBranch: 907 return inline_profileBranch(); 908 909 default: 910 // If you get here, it may be that someone has added a new intrinsic 911 // to the list in vmSymbols.hpp without implementing it here. 912 #ifndef PRODUCT 913 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 914 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 915 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 916 } 917 #endif 918 return false; 919 } 920 } 921 922 Node* LibraryCallKit::try_to_predicate(int predicate) { 923 if (!jvms()->has_method()) { 924 // Root JVMState has a null method. 925 assert(map()->memory()->Opcode() == Op_Parm, ""); 926 // Insert the memory aliasing node 927 set_all_memory(reset_memory()); 928 } 929 assert(merged_memory(), ""); 930 931 switch (intrinsic_id()) { 932 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 933 return inline_cipherBlockChaining_AESCrypt_predicate(false); 934 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 935 return inline_cipherBlockChaining_AESCrypt_predicate(true); 936 case vmIntrinsics::_digestBase_implCompressMB: 937 return inline_digestBase_implCompressMB_predicate(predicate); 938 939 default: 940 // If you get here, it may be that someone has added a new intrinsic 941 // to the list in vmSymbols.hpp without implementing it here. 942 #ifndef PRODUCT 943 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 944 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 945 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 946 } 947 #endif 948 Node* slow_ctl = control(); 949 set_control(top()); // No fast path instrinsic 950 return slow_ctl; 951 } 952 } 953 954 //------------------------------set_result------------------------------- 955 // Helper function for finishing intrinsics. 956 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 957 record_for_igvn(region); 958 set_control(_gvn.transform(region)); 959 set_result( _gvn.transform(value)); 960 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 961 } 962 963 //------------------------------generate_guard--------------------------- 964 // Helper function for generating guarded fast-slow graph structures. 965 // The given 'test', if true, guards a slow path. If the test fails 966 // then a fast path can be taken. (We generally hope it fails.) 967 // In all cases, GraphKit::control() is updated to the fast path. 968 // The returned value represents the control for the slow path. 969 // The return value is never 'top'; it is either a valid control 970 // or NULL if it is obvious that the slow path can never be taken. 971 // Also, if region and the slow control are not NULL, the slow edge 972 // is appended to the region. 973 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 974 if (stopped()) { 975 // Already short circuited. 976 return NULL; 977 } 978 979 // Build an if node and its projections. 980 // If test is true we take the slow path, which we assume is uncommon. 981 if (_gvn.type(test) == TypeInt::ZERO) { 982 // The slow branch is never taken. No need to build this guard. 983 return NULL; 984 } 985 986 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 987 988 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 989 if (if_slow == top()) { 990 // The slow branch is never taken. No need to build this guard. 991 return NULL; 992 } 993 994 if (region != NULL) 995 region->add_req(if_slow); 996 997 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 998 set_control(if_fast); 999 1000 return if_slow; 1001 } 1002 1003 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1004 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1005 } 1006 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1007 return generate_guard(test, region, PROB_FAIR); 1008 } 1009 1010 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1011 Node* *pos_index) { 1012 if (stopped()) 1013 return NULL; // already stopped 1014 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1015 return NULL; // index is already adequately typed 1016 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1017 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1018 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1019 if (is_neg != NULL && pos_index != NULL) { 1020 // Emulate effect of Parse::adjust_map_after_if. 1021 Node* ccast = new CastIINode(index, TypeInt::POS); 1022 ccast->set_req(0, control()); 1023 (*pos_index) = _gvn.transform(ccast); 1024 } 1025 return is_neg; 1026 } 1027 1028 // Make sure that 'position' is a valid limit index, in [0..length]. 1029 // There are two equivalent plans for checking this: 1030 // A. (offset + copyLength) unsigned<= arrayLength 1031 // B. offset <= (arrayLength - copyLength) 1032 // We require that all of the values above, except for the sum and 1033 // difference, are already known to be non-negative. 1034 // Plan A is robust in the face of overflow, if offset and copyLength 1035 // are both hugely positive. 1036 // 1037 // Plan B is less direct and intuitive, but it does not overflow at 1038 // all, since the difference of two non-negatives is always 1039 // representable. Whenever Java methods must perform the equivalent 1040 // check they generally use Plan B instead of Plan A. 1041 // For the moment we use Plan A. 1042 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1043 Node* subseq_length, 1044 Node* array_length, 1045 RegionNode* region) { 1046 if (stopped()) 1047 return NULL; // already stopped 1048 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1049 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1050 return NULL; // common case of whole-array copy 1051 Node* last = subseq_length; 1052 if (!zero_offset) // last += offset 1053 last = _gvn.transform(new AddINode(last, offset)); 1054 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1055 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1056 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1057 return is_over; 1058 } 1059 1060 1061 //--------------------------generate_current_thread-------------------- 1062 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1063 ciKlass* thread_klass = env()->Thread_klass(); 1064 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1065 Node* thread = _gvn.transform(new ThreadLocalNode()); 1066 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1067 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1068 tls_output = thread; 1069 return threadObj; 1070 } 1071 1072 1073 //------------------------------make_string_method_node------------------------ 1074 // Helper method for String intrinsic functions. This version is called 1075 // with str1 and str2 pointing to String object nodes. 1076 // 1077 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1078 Node* no_ctrl = NULL; 1079 1080 // Get start addr of string 1081 Node* str1_value = load_String_value(no_ctrl, str1); 1082 Node* str1_offset = load_String_offset(no_ctrl, str1); 1083 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1084 1085 // Get length of string 1 1086 Node* str1_len = load_String_length(no_ctrl, str1); 1087 1088 Node* str2_value = load_String_value(no_ctrl, str2); 1089 Node* str2_offset = load_String_offset(no_ctrl, str2); 1090 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1091 1092 Node* str2_len = NULL; 1093 Node* result = NULL; 1094 1095 switch (opcode) { 1096 case Op_StrIndexOf: 1097 // Get length of string 2 1098 str2_len = load_String_length(no_ctrl, str2); 1099 1100 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1101 str1_start, str1_len, str2_start, str2_len); 1102 break; 1103 case Op_StrComp: 1104 // Get length of string 2 1105 str2_len = load_String_length(no_ctrl, str2); 1106 1107 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1108 str1_start, str1_len, str2_start, str2_len); 1109 break; 1110 case Op_StrEquals: 1111 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1112 str1_start, str2_start, str1_len); 1113 break; 1114 default: 1115 ShouldNotReachHere(); 1116 return NULL; 1117 } 1118 1119 // All these intrinsics have checks. 1120 C->set_has_split_ifs(true); // Has chance for split-if optimization 1121 1122 return _gvn.transform(result); 1123 } 1124 1125 // Helper method for String intrinsic functions. This version is called 1126 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1127 // to Int nodes containing the lenghts of str1 and str2. 1128 // 1129 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1130 Node* result = NULL; 1131 switch (opcode) { 1132 case Op_StrIndexOf: 1133 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1134 str1_start, cnt1, str2_start, cnt2); 1135 break; 1136 case Op_StrComp: 1137 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1138 str1_start, cnt1, str2_start, cnt2); 1139 break; 1140 case Op_StrEquals: 1141 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1142 str1_start, str2_start, cnt1); 1143 break; 1144 default: 1145 ShouldNotReachHere(); 1146 return NULL; 1147 } 1148 1149 // All these intrinsics have checks. 1150 C->set_has_split_ifs(true); // Has chance for split-if optimization 1151 1152 return _gvn.transform(result); 1153 } 1154 1155 //------------------------------inline_string_compareTo------------------------ 1156 // public int java.lang.String.compareTo(String anotherString); 1157 bool LibraryCallKit::inline_string_compareTo() { 1158 Node* receiver = null_check(argument(0)); 1159 Node* arg = null_check(argument(1)); 1160 if (stopped()) { 1161 return true; 1162 } 1163 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1164 return true; 1165 } 1166 1167 //------------------------------inline_string_equals------------------------ 1168 bool LibraryCallKit::inline_string_equals() { 1169 Node* receiver = null_check_receiver(); 1170 // NOTE: Do not null check argument for String.equals() because spec 1171 // allows to specify NULL as argument. 1172 Node* argument = this->argument(1); 1173 if (stopped()) { 1174 return true; 1175 } 1176 1177 // paths (plus control) merge 1178 RegionNode* region = new RegionNode(5); 1179 Node* phi = new PhiNode(region, TypeInt::BOOL); 1180 1181 // does source == target string? 1182 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1183 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1184 1185 Node* if_eq = generate_slow_guard(bol, NULL); 1186 if (if_eq != NULL) { 1187 // receiver == argument 1188 phi->init_req(2, intcon(1)); 1189 region->init_req(2, if_eq); 1190 } 1191 1192 // get String klass for instanceOf 1193 ciInstanceKlass* klass = env()->String_klass(); 1194 1195 if (!stopped()) { 1196 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1197 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1198 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1199 1200 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1201 //instanceOf == true, fallthrough 1202 1203 if (inst_false != NULL) { 1204 phi->init_req(3, intcon(0)); 1205 region->init_req(3, inst_false); 1206 } 1207 } 1208 1209 if (!stopped()) { 1210 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1211 1212 // Properly cast the argument to String 1213 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1214 // This path is taken only when argument's type is String:NotNull. 1215 argument = cast_not_null(argument, false); 1216 1217 Node* no_ctrl = NULL; 1218 1219 // Get start addr of receiver 1220 Node* receiver_val = load_String_value(no_ctrl, receiver); 1221 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1222 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1223 1224 // Get length of receiver 1225 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1226 1227 // Get start addr of argument 1228 Node* argument_val = load_String_value(no_ctrl, argument); 1229 Node* argument_offset = load_String_offset(no_ctrl, argument); 1230 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1231 1232 // Get length of argument 1233 Node* argument_cnt = load_String_length(no_ctrl, argument); 1234 1235 // Check for receiver count != argument count 1236 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1237 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1238 Node* if_ne = generate_slow_guard(bol, NULL); 1239 if (if_ne != NULL) { 1240 phi->init_req(4, intcon(0)); 1241 region->init_req(4, if_ne); 1242 } 1243 1244 // Check for count == 0 is done by assembler code for StrEquals. 1245 1246 if (!stopped()) { 1247 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1248 phi->init_req(1, equals); 1249 region->init_req(1, control()); 1250 } 1251 } 1252 1253 // post merge 1254 set_control(_gvn.transform(region)); 1255 record_for_igvn(region); 1256 1257 set_result(_gvn.transform(phi)); 1258 return true; 1259 } 1260 1261 //------------------------------inline_array_equals---------------------------- 1262 bool LibraryCallKit::inline_array_equals() { 1263 Node* arg1 = argument(0); 1264 Node* arg2 = argument(1); 1265 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1266 return true; 1267 } 1268 1269 // Java version of String.indexOf(constant string) 1270 // class StringDecl { 1271 // StringDecl(char[] ca) { 1272 // offset = 0; 1273 // count = ca.length; 1274 // value = ca; 1275 // } 1276 // int offset; 1277 // int count; 1278 // char[] value; 1279 // } 1280 // 1281 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1282 // int targetOffset, int cache_i, int md2) { 1283 // int cache = cache_i; 1284 // int sourceOffset = string_object.offset; 1285 // int sourceCount = string_object.count; 1286 // int targetCount = target_object.length; 1287 // 1288 // int targetCountLess1 = targetCount - 1; 1289 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1290 // 1291 // char[] source = string_object.value; 1292 // char[] target = target_object; 1293 // int lastChar = target[targetCountLess1]; 1294 // 1295 // outer_loop: 1296 // for (int i = sourceOffset; i < sourceEnd; ) { 1297 // int src = source[i + targetCountLess1]; 1298 // if (src == lastChar) { 1299 // // With random strings and a 4-character alphabet, 1300 // // reverse matching at this point sets up 0.8% fewer 1301 // // frames, but (paradoxically) makes 0.3% more probes. 1302 // // Since those probes are nearer the lastChar probe, 1303 // // there is may be a net D$ win with reverse matching. 1304 // // But, reversing loop inhibits unroll of inner loop 1305 // // for unknown reason. So, does running outer loop from 1306 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1307 // for (int j = 0; j < targetCountLess1; j++) { 1308 // if (target[targetOffset + j] != source[i+j]) { 1309 // if ((cache & (1 << source[i+j])) == 0) { 1310 // if (md2 < j+1) { 1311 // i += j+1; 1312 // continue outer_loop; 1313 // } 1314 // } 1315 // i += md2; 1316 // continue outer_loop; 1317 // } 1318 // } 1319 // return i - sourceOffset; 1320 // } 1321 // if ((cache & (1 << src)) == 0) { 1322 // i += targetCountLess1; 1323 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1324 // i++; 1325 // } 1326 // return -1; 1327 // } 1328 1329 //------------------------------string_indexOf------------------------ 1330 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1331 jint cache_i, jint md2_i) { 1332 1333 Node* no_ctrl = NULL; 1334 float likely = PROB_LIKELY(0.9); 1335 float unlikely = PROB_UNLIKELY(0.9); 1336 1337 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1338 1339 Node* source = load_String_value(no_ctrl, string_object); 1340 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1341 Node* sourceCount = load_String_length(no_ctrl, string_object); 1342 1343 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1344 jint target_length = target_array->length(); 1345 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1346 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1347 1348 // String.value field is known to be @Stable. 1349 if (UseImplicitStableValues) { 1350 target = cast_array_to_stable(target, target_type); 1351 } 1352 1353 IdealKit kit(this, false, true); 1354 #define __ kit. 1355 Node* zero = __ ConI(0); 1356 Node* one = __ ConI(1); 1357 Node* cache = __ ConI(cache_i); 1358 Node* md2 = __ ConI(md2_i); 1359 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1360 Node* targetCount = __ ConI(target_length); 1361 Node* targetCountLess1 = __ ConI(target_length - 1); 1362 Node* targetOffset = __ ConI(targetOffset_i); 1363 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1364 1365 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1366 Node* outer_loop = __ make_label(2 /* goto */); 1367 Node* return_ = __ make_label(1); 1368 1369 __ set(rtn,__ ConI(-1)); 1370 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1371 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1372 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1373 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1374 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1375 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1376 Node* tpj = __ AddI(targetOffset, __ value(j)); 1377 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1378 Node* ipj = __ AddI(__ value(i), __ value(j)); 1379 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1380 __ if_then(targ, BoolTest::ne, src2); { 1381 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1382 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1383 __ increment(i, __ AddI(__ value(j), one)); 1384 __ goto_(outer_loop); 1385 } __ end_if(); __ dead(j); 1386 }__ end_if(); __ dead(j); 1387 __ increment(i, md2); 1388 __ goto_(outer_loop); 1389 }__ end_if(); 1390 __ increment(j, one); 1391 }__ end_loop(); __ dead(j); 1392 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1393 __ goto_(return_); 1394 }__ end_if(); 1395 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1396 __ increment(i, targetCountLess1); 1397 }__ end_if(); 1398 __ increment(i, one); 1399 __ bind(outer_loop); 1400 }__ end_loop(); __ dead(i); 1401 __ bind(return_); 1402 1403 // Final sync IdealKit and GraphKit. 1404 final_sync(kit); 1405 Node* result = __ value(rtn); 1406 #undef __ 1407 C->set_has_loops(true); 1408 return result; 1409 } 1410 1411 //------------------------------inline_string_indexOf------------------------ 1412 bool LibraryCallKit::inline_string_indexOf() { 1413 Node* receiver = argument(0); 1414 Node* arg = argument(1); 1415 1416 Node* result; 1417 // Disable the use of pcmpestri until it can be guaranteed that 1418 // the load doesn't cross into the uncommited space. 1419 if (Matcher::has_match_rule(Op_StrIndexOf) && 1420 UseSSE42Intrinsics) { 1421 // Generate SSE4.2 version of indexOf 1422 // We currently only have match rules that use SSE4.2 1423 1424 receiver = null_check(receiver); 1425 arg = null_check(arg); 1426 if (stopped()) { 1427 return true; 1428 } 1429 1430 ciInstanceKlass* str_klass = env()->String_klass(); 1431 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1432 1433 // Make the merge point 1434 RegionNode* result_rgn = new RegionNode(4); 1435 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1436 Node* no_ctrl = NULL; 1437 1438 // Get start addr of source string 1439 Node* source = load_String_value(no_ctrl, receiver); 1440 Node* source_offset = load_String_offset(no_ctrl, receiver); 1441 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1442 1443 // Get length of source string 1444 Node* source_cnt = load_String_length(no_ctrl, receiver); 1445 1446 // Get start addr of substring 1447 Node* substr = load_String_value(no_ctrl, arg); 1448 Node* substr_offset = load_String_offset(no_ctrl, arg); 1449 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1450 1451 // Get length of source string 1452 Node* substr_cnt = load_String_length(no_ctrl, arg); 1453 1454 // Check for substr count > string count 1455 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1456 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1457 Node* if_gt = generate_slow_guard(bol, NULL); 1458 if (if_gt != NULL) { 1459 result_phi->init_req(2, intcon(-1)); 1460 result_rgn->init_req(2, if_gt); 1461 } 1462 1463 if (!stopped()) { 1464 // Check for substr count == 0 1465 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1466 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1467 Node* if_zero = generate_slow_guard(bol, NULL); 1468 if (if_zero != NULL) { 1469 result_phi->init_req(3, intcon(0)); 1470 result_rgn->init_req(3, if_zero); 1471 } 1472 } 1473 1474 if (!stopped()) { 1475 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1476 result_phi->init_req(1, result); 1477 result_rgn->init_req(1, control()); 1478 } 1479 set_control(_gvn.transform(result_rgn)); 1480 record_for_igvn(result_rgn); 1481 result = _gvn.transform(result_phi); 1482 1483 } else { // Use LibraryCallKit::string_indexOf 1484 // don't intrinsify if argument isn't a constant string. 1485 if (!arg->is_Con()) { 1486 return false; 1487 } 1488 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1489 if (str_type == NULL) { 1490 return false; 1491 } 1492 ciInstanceKlass* klass = env()->String_klass(); 1493 ciObject* str_const = str_type->const_oop(); 1494 if (str_const == NULL || str_const->klass() != klass) { 1495 return false; 1496 } 1497 ciInstance* str = str_const->as_instance(); 1498 assert(str != NULL, "must be instance"); 1499 1500 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1501 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1502 1503 int o; 1504 int c; 1505 if (java_lang_String::has_offset_field()) { 1506 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1507 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1508 } else { 1509 o = 0; 1510 c = pat->length(); 1511 } 1512 1513 // constant strings have no offset and count == length which 1514 // simplifies the resulting code somewhat so lets optimize for that. 1515 if (o != 0 || c != pat->length()) { 1516 return false; 1517 } 1518 1519 receiver = null_check(receiver, T_OBJECT); 1520 // NOTE: No null check on the argument is needed since it's a constant String oop. 1521 if (stopped()) { 1522 return true; 1523 } 1524 1525 // The null string as a pattern always returns 0 (match at beginning of string) 1526 if (c == 0) { 1527 set_result(intcon(0)); 1528 return true; 1529 } 1530 1531 // Generate default indexOf 1532 jchar lastChar = pat->char_at(o + (c - 1)); 1533 int cache = 0; 1534 int i; 1535 for (i = 0; i < c - 1; i++) { 1536 assert(i < pat->length(), "out of range"); 1537 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1538 } 1539 1540 int md2 = c; 1541 for (i = 0; i < c - 1; i++) { 1542 assert(i < pat->length(), "out of range"); 1543 if (pat->char_at(o + i) == lastChar) { 1544 md2 = (c - 1) - i; 1545 } 1546 } 1547 1548 result = string_indexOf(receiver, pat, o, cache, md2); 1549 } 1550 set_result(result); 1551 return true; 1552 } 1553 1554 //--------------------------round_double_node-------------------------------- 1555 // Round a double node if necessary. 1556 Node* LibraryCallKit::round_double_node(Node* n) { 1557 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1558 n = _gvn.transform(new RoundDoubleNode(0, n)); 1559 return n; 1560 } 1561 1562 //------------------------------inline_math----------------------------------- 1563 // public static double Math.abs(double) 1564 // public static double Math.sqrt(double) 1565 // public static double Math.log(double) 1566 // public static double Math.log10(double) 1567 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1568 Node* arg = round_double_node(argument(0)); 1569 Node* n; 1570 switch (id) { 1571 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1572 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1573 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1574 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1575 default: fatal_unexpected_iid(id); break; 1576 } 1577 set_result(_gvn.transform(n)); 1578 return true; 1579 } 1580 1581 //------------------------------inline_trig---------------------------------- 1582 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1583 // argument reduction which will turn into a fast/slow diamond. 1584 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1585 Node* arg = round_double_node(argument(0)); 1586 Node* n = NULL; 1587 1588 switch (id) { 1589 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1590 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1591 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1592 default: fatal_unexpected_iid(id); break; 1593 } 1594 n = _gvn.transform(n); 1595 1596 // Rounding required? Check for argument reduction! 1597 if (Matcher::strict_fp_requires_explicit_rounding) { 1598 static const double pi_4 = 0.7853981633974483; 1599 static const double neg_pi_4 = -0.7853981633974483; 1600 // pi/2 in 80-bit extended precision 1601 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1602 // -pi/2 in 80-bit extended precision 1603 // 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}; 1604 // Cutoff value for using this argument reduction technique 1605 //static const double pi_2_minus_epsilon = 1.564660403643354; 1606 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1607 1608 // Pseudocode for sin: 1609 // if (x <= Math.PI / 4.0) { 1610 // if (x >= -Math.PI / 4.0) return fsin(x); 1611 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1612 // } else { 1613 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1614 // } 1615 // return StrictMath.sin(x); 1616 1617 // Pseudocode for cos: 1618 // if (x <= Math.PI / 4.0) { 1619 // if (x >= -Math.PI / 4.0) return fcos(x); 1620 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1621 // } else { 1622 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1623 // } 1624 // return StrictMath.cos(x); 1625 1626 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1627 // requires a special machine instruction to load it. Instead we'll try 1628 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1629 // probably do the math inside the SIN encoding. 1630 1631 // Make the merge point 1632 RegionNode* r = new RegionNode(3); 1633 Node* phi = new PhiNode(r, Type::DOUBLE); 1634 1635 // Flatten arg so we need only 1 test 1636 Node *abs = _gvn.transform(new AbsDNode(arg)); 1637 // Node for PI/4 constant 1638 Node *pi4 = makecon(TypeD::make(pi_4)); 1639 // Check PI/4 : abs(arg) 1640 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1641 // Check: If PI/4 < abs(arg) then go slow 1642 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1643 // Branch either way 1644 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1645 set_control(opt_iff(r,iff)); 1646 1647 // Set fast path result 1648 phi->init_req(2, n); 1649 1650 // Slow path - non-blocking leaf call 1651 Node* call = NULL; 1652 switch (id) { 1653 case vmIntrinsics::_dsin: 1654 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1655 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1656 "Sin", NULL, arg, top()); 1657 break; 1658 case vmIntrinsics::_dcos: 1659 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1660 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1661 "Cos", NULL, arg, top()); 1662 break; 1663 case vmIntrinsics::_dtan: 1664 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1665 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1666 "Tan", NULL, arg, top()); 1667 break; 1668 } 1669 assert(control()->in(0) == call, ""); 1670 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1671 r->init_req(1, control()); 1672 phi->init_req(1, slow_result); 1673 1674 // Post-merge 1675 set_control(_gvn.transform(r)); 1676 record_for_igvn(r); 1677 n = _gvn.transform(phi); 1678 1679 C->set_has_split_ifs(true); // Has chance for split-if optimization 1680 } 1681 set_result(n); 1682 return true; 1683 } 1684 1685 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1686 //------------------- 1687 //result=(result.isNaN())? funcAddr():result; 1688 // Check: If isNaN() by checking result!=result? then either trap 1689 // or go to runtime 1690 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1691 // Build the boolean node 1692 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1693 1694 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1695 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1696 // The pow or exp intrinsic returned a NaN, which requires a call 1697 // to the runtime. Recompile with the runtime call. 1698 uncommon_trap(Deoptimization::Reason_intrinsic, 1699 Deoptimization::Action_make_not_entrant); 1700 } 1701 return result; 1702 } else { 1703 // If this inlining ever returned NaN in the past, we compile a call 1704 // to the runtime to properly handle corner cases 1705 1706 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1707 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1708 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1709 1710 if (!if_slow->is_top()) { 1711 RegionNode* result_region = new RegionNode(3); 1712 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1713 1714 result_region->init_req(1, if_fast); 1715 result_val->init_req(1, result); 1716 1717 set_control(if_slow); 1718 1719 const TypePtr* no_memory_effects = NULL; 1720 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1721 no_memory_effects, 1722 x, top(), y, y ? top() : NULL); 1723 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1724 #ifdef ASSERT 1725 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1726 assert(value_top == top(), "second value must be top"); 1727 #endif 1728 1729 result_region->init_req(2, control()); 1730 result_val->init_req(2, value); 1731 set_control(_gvn.transform(result_region)); 1732 return _gvn.transform(result_val); 1733 } else { 1734 return result; 1735 } 1736 } 1737 } 1738 1739 //------------------------------inline_exp------------------------------------- 1740 // Inline exp instructions, if possible. The Intel hardware only misses 1741 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1742 bool LibraryCallKit::inline_exp() { 1743 Node* arg = round_double_node(argument(0)); 1744 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1745 1746 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1747 set_result(n); 1748 1749 C->set_has_split_ifs(true); // Has chance for split-if optimization 1750 return true; 1751 } 1752 1753 //------------------------------inline_pow------------------------------------- 1754 // Inline power instructions, if possible. 1755 bool LibraryCallKit::inline_pow() { 1756 // Pseudocode for pow 1757 // if (y == 2) { 1758 // return x * x; 1759 // } else { 1760 // if (x <= 0.0) { 1761 // long longy = (long)y; 1762 // if ((double)longy == y) { // if y is long 1763 // if (y + 1 == y) longy = 0; // huge number: even 1764 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1765 // } else { 1766 // result = NaN; 1767 // } 1768 // } else { 1769 // result = DPow(x,y); 1770 // } 1771 // if (result != result)? { 1772 // result = uncommon_trap() or runtime_call(); 1773 // } 1774 // return result; 1775 // } 1776 1777 Node* x = round_double_node(argument(0)); 1778 Node* y = round_double_node(argument(2)); 1779 1780 Node* result = NULL; 1781 1782 Node* const_two_node = makecon(TypeD::make(2.0)); 1783 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1784 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1785 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1786 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1787 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1788 1789 RegionNode* region_node = new RegionNode(3); 1790 region_node->init_req(1, if_true); 1791 1792 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1793 // special case for x^y where y == 2, we can convert it to x * x 1794 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1795 1796 // set control to if_false since we will now process the false branch 1797 set_control(if_false); 1798 1799 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1800 // Short form: skip the fancy tests and just check for NaN result. 1801 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1802 } else { 1803 // If this inlining ever returned NaN in the past, include all 1804 // checks + call to the runtime. 1805 1806 // Set the merge point for If node with condition of (x <= 0.0) 1807 // There are four possible paths to region node and phi node 1808 RegionNode *r = new RegionNode(4); 1809 Node *phi = new PhiNode(r, Type::DOUBLE); 1810 1811 // Build the first if node: if (x <= 0.0) 1812 // Node for 0 constant 1813 Node *zeronode = makecon(TypeD::ZERO); 1814 // Check x:0 1815 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1816 // Check: If (x<=0) then go complex path 1817 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1818 // Branch either way 1819 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1820 // Fast path taken; set region slot 3 1821 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1822 r->init_req(3,fast_taken); // Capture fast-control 1823 1824 // Fast path not-taken, i.e. slow path 1825 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1826 1827 // Set fast path result 1828 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1829 phi->init_req(3, fast_result); 1830 1831 // Complex path 1832 // Build the second if node (if y is long) 1833 // Node for (long)y 1834 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1835 // Node for (double)((long) y) 1836 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1837 // Check (double)((long) y) : y 1838 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1839 // Check if (y isn't long) then go to slow path 1840 1841 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1842 // Branch either way 1843 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1844 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1845 1846 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1847 1848 // Calculate DPow(abs(x), y)*(1 & (long)y) 1849 // Node for constant 1 1850 Node *conone = longcon(1); 1851 // 1& (long)y 1852 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1853 1854 // A huge number is always even. Detect a huge number by checking 1855 // if y + 1 == y and set integer to be tested for parity to 0. 1856 // Required for corner case: 1857 // (long)9.223372036854776E18 = max_jlong 1858 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1859 // max_jlong is odd but 9.223372036854776E18 is even 1860 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1861 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1862 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1863 Node* correctedsign = NULL; 1864 if (ConditionalMoveLimit != 0) { 1865 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1866 } else { 1867 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1868 RegionNode *r = new RegionNode(3); 1869 Node *phi = new PhiNode(r, TypeLong::LONG); 1870 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1871 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1872 phi->init_req(1, signnode); 1873 phi->init_req(2, longcon(0)); 1874 correctedsign = _gvn.transform(phi); 1875 ylong_path = _gvn.transform(r); 1876 record_for_igvn(r); 1877 } 1878 1879 // zero node 1880 Node *conzero = longcon(0); 1881 // Check (1&(long)y)==0? 1882 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1883 // Check if (1&(long)y)!=0?, if so the result is negative 1884 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1885 // abs(x) 1886 Node *absx=_gvn.transform(new AbsDNode(x)); 1887 // abs(x)^y 1888 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1889 // -abs(x)^y 1890 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1891 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1892 Node *signresult = NULL; 1893 if (ConditionalMoveLimit != 0) { 1894 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1895 } else { 1896 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1897 RegionNode *r = new RegionNode(3); 1898 Node *phi = new PhiNode(r, Type::DOUBLE); 1899 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1900 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1901 phi->init_req(1, absxpowy); 1902 phi->init_req(2, negabsxpowy); 1903 signresult = _gvn.transform(phi); 1904 ylong_path = _gvn.transform(r); 1905 record_for_igvn(r); 1906 } 1907 // Set complex path fast result 1908 r->init_req(2, ylong_path); 1909 phi->init_req(2, signresult); 1910 1911 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1912 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1913 r->init_req(1,slow_path); 1914 phi->init_req(1,slow_result); 1915 1916 // Post merge 1917 set_control(_gvn.transform(r)); 1918 record_for_igvn(r); 1919 result = _gvn.transform(phi); 1920 } 1921 1922 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1923 1924 // control from finish_pow_exp is now input to the region node 1925 region_node->set_req(2, control()); 1926 // the result from finish_pow_exp is now input to the phi node 1927 phi_node->init_req(2, result); 1928 set_control(_gvn.transform(region_node)); 1929 record_for_igvn(region_node); 1930 set_result(_gvn.transform(phi_node)); 1931 1932 C->set_has_split_ifs(true); // Has chance for split-if optimization 1933 return true; 1934 } 1935 1936 //------------------------------runtime_math----------------------------- 1937 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1938 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1939 "must be (DD)D or (D)D type"); 1940 1941 // Inputs 1942 Node* a = round_double_node(argument(0)); 1943 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1944 1945 const TypePtr* no_memory_effects = NULL; 1946 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1947 no_memory_effects, 1948 a, top(), b, b ? top() : NULL); 1949 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1950 #ifdef ASSERT 1951 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1952 assert(value_top == top(), "second value must be top"); 1953 #endif 1954 1955 set_result(value); 1956 return true; 1957 } 1958 1959 //------------------------------inline_math_native----------------------------- 1960 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1961 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1962 switch (id) { 1963 // These intrinsics are not properly supported on all hardware 1964 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1965 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1966 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1967 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1968 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1969 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1970 1971 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1972 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1973 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1974 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1975 1976 // These intrinsics are supported on all hardware 1977 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1978 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1979 1980 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1981 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1982 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1983 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1984 #undef FN_PTR 1985 1986 // These intrinsics are not yet correctly implemented 1987 case vmIntrinsics::_datan2: 1988 return false; 1989 1990 default: 1991 fatal_unexpected_iid(id); 1992 return false; 1993 } 1994 } 1995 1996 static bool is_simple_name(Node* n) { 1997 return (n->req() == 1 // constant 1998 || (n->is_Type() && n->as_Type()->type()->singleton()) 1999 || n->is_Proj() // parameter or return value 2000 || n->is_Phi() // local of some sort 2001 ); 2002 } 2003 2004 //----------------------------inline_min_max----------------------------------- 2005 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2006 set_result(generate_min_max(id, argument(0), argument(1))); 2007 return true; 2008 } 2009 2010 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2011 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2012 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2013 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2014 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2015 2016 { 2017 PreserveJVMState pjvms(this); 2018 PreserveReexecuteState preexecs(this); 2019 jvms()->set_should_reexecute(true); 2020 2021 set_control(slow_path); 2022 set_i_o(i_o()); 2023 2024 uncommon_trap(Deoptimization::Reason_intrinsic, 2025 Deoptimization::Action_none); 2026 } 2027 2028 set_control(fast_path); 2029 set_result(math); 2030 } 2031 2032 template <typename OverflowOp> 2033 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2034 typedef typename OverflowOp::MathOp MathOp; 2035 2036 MathOp* mathOp = new MathOp(arg1, arg2); 2037 Node* operation = _gvn.transform( mathOp ); 2038 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2039 inline_math_mathExact(operation, ofcheck); 2040 return true; 2041 } 2042 2043 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2044 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2045 } 2046 2047 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2048 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2049 } 2050 2051 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2052 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2053 } 2054 2055 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2056 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2057 } 2058 2059 bool LibraryCallKit::inline_math_negateExactI() { 2060 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2061 } 2062 2063 bool LibraryCallKit::inline_math_negateExactL() { 2064 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2065 } 2066 2067 bool LibraryCallKit::inline_math_multiplyExactI() { 2068 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2069 } 2070 2071 bool LibraryCallKit::inline_math_multiplyExactL() { 2072 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2073 } 2074 2075 Node* 2076 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2077 // These are the candidate return value: 2078 Node* xvalue = x0; 2079 Node* yvalue = y0; 2080 2081 if (xvalue == yvalue) { 2082 return xvalue; 2083 } 2084 2085 bool want_max = (id == vmIntrinsics::_max); 2086 2087 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2088 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2089 if (txvalue == NULL || tyvalue == NULL) return top(); 2090 // This is not really necessary, but it is consistent with a 2091 // hypothetical MaxINode::Value method: 2092 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2093 2094 // %%% This folding logic should (ideally) be in a different place. 2095 // Some should be inside IfNode, and there to be a more reliable 2096 // transformation of ?: style patterns into cmoves. We also want 2097 // more powerful optimizations around cmove and min/max. 2098 2099 // Try to find a dominating comparison of these guys. 2100 // It can simplify the index computation for Arrays.copyOf 2101 // and similar uses of System.arraycopy. 2102 // First, compute the normalized version of CmpI(x, y). 2103 int cmp_op = Op_CmpI; 2104 Node* xkey = xvalue; 2105 Node* ykey = yvalue; 2106 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2107 if (ideal_cmpxy->is_Cmp()) { 2108 // E.g., if we have CmpI(length - offset, count), 2109 // it might idealize to CmpI(length, count + offset) 2110 cmp_op = ideal_cmpxy->Opcode(); 2111 xkey = ideal_cmpxy->in(1); 2112 ykey = ideal_cmpxy->in(2); 2113 } 2114 2115 // Start by locating any relevant comparisons. 2116 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2117 Node* cmpxy = NULL; 2118 Node* cmpyx = NULL; 2119 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2120 Node* cmp = start_from->fast_out(k); 2121 if (cmp->outcnt() > 0 && // must have prior uses 2122 cmp->in(0) == NULL && // must be context-independent 2123 cmp->Opcode() == cmp_op) { // right kind of compare 2124 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2125 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2126 } 2127 } 2128 2129 const int NCMPS = 2; 2130 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2131 int cmpn; 2132 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2133 if (cmps[cmpn] != NULL) break; // find a result 2134 } 2135 if (cmpn < NCMPS) { 2136 // Look for a dominating test that tells us the min and max. 2137 int depth = 0; // Limit search depth for speed 2138 Node* dom = control(); 2139 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2140 if (++depth >= 100) break; 2141 Node* ifproj = dom; 2142 if (!ifproj->is_Proj()) continue; 2143 Node* iff = ifproj->in(0); 2144 if (!iff->is_If()) continue; 2145 Node* bol = iff->in(1); 2146 if (!bol->is_Bool()) continue; 2147 Node* cmp = bol->in(1); 2148 if (cmp == NULL) continue; 2149 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2150 if (cmps[cmpn] == cmp) break; 2151 if (cmpn == NCMPS) continue; 2152 BoolTest::mask btest = bol->as_Bool()->_test._test; 2153 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2154 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2155 // At this point, we know that 'x btest y' is true. 2156 switch (btest) { 2157 case BoolTest::eq: 2158 // They are proven equal, so we can collapse the min/max. 2159 // Either value is the answer. Choose the simpler. 2160 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2161 return yvalue; 2162 return xvalue; 2163 case BoolTest::lt: // x < y 2164 case BoolTest::le: // x <= y 2165 return (want_max ? yvalue : xvalue); 2166 case BoolTest::gt: // x > y 2167 case BoolTest::ge: // x >= y 2168 return (want_max ? xvalue : yvalue); 2169 } 2170 } 2171 } 2172 2173 // We failed to find a dominating test. 2174 // Let's pick a test that might GVN with prior tests. 2175 Node* best_bol = NULL; 2176 BoolTest::mask best_btest = BoolTest::illegal; 2177 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2178 Node* cmp = cmps[cmpn]; 2179 if (cmp == NULL) continue; 2180 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2181 Node* bol = cmp->fast_out(j); 2182 if (!bol->is_Bool()) continue; 2183 BoolTest::mask btest = bol->as_Bool()->_test._test; 2184 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2185 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2186 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2187 best_bol = bol->as_Bool(); 2188 best_btest = btest; 2189 } 2190 } 2191 } 2192 2193 Node* answer_if_true = NULL; 2194 Node* answer_if_false = NULL; 2195 switch (best_btest) { 2196 default: 2197 if (cmpxy == NULL) 2198 cmpxy = ideal_cmpxy; 2199 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2200 // and fall through: 2201 case BoolTest::lt: // x < y 2202 case BoolTest::le: // x <= y 2203 answer_if_true = (want_max ? yvalue : xvalue); 2204 answer_if_false = (want_max ? xvalue : yvalue); 2205 break; 2206 case BoolTest::gt: // x > y 2207 case BoolTest::ge: // x >= y 2208 answer_if_true = (want_max ? xvalue : yvalue); 2209 answer_if_false = (want_max ? yvalue : xvalue); 2210 break; 2211 } 2212 2213 jint hi, lo; 2214 if (want_max) { 2215 // We can sharpen the minimum. 2216 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2217 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2218 } else { 2219 // We can sharpen the maximum. 2220 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2221 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2222 } 2223 2224 // Use a flow-free graph structure, to avoid creating excess control edges 2225 // which could hinder other optimizations. 2226 // Since Math.min/max is often used with arraycopy, we want 2227 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2228 Node* cmov = CMoveNode::make(NULL, best_bol, 2229 answer_if_false, answer_if_true, 2230 TypeInt::make(lo, hi, widen)); 2231 2232 return _gvn.transform(cmov); 2233 2234 /* 2235 // This is not as desirable as it may seem, since Min and Max 2236 // nodes do not have a full set of optimizations. 2237 // And they would interfere, anyway, with 'if' optimizations 2238 // and with CMoveI canonical forms. 2239 switch (id) { 2240 case vmIntrinsics::_min: 2241 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2242 case vmIntrinsics::_max: 2243 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2244 default: 2245 ShouldNotReachHere(); 2246 } 2247 */ 2248 } 2249 2250 inline int 2251 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2252 const TypePtr* base_type = TypePtr::NULL_PTR; 2253 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2254 if (base_type == NULL) { 2255 // Unknown type. 2256 return Type::AnyPtr; 2257 } else if (base_type == TypePtr::NULL_PTR) { 2258 // Since this is a NULL+long form, we have to switch to a rawptr. 2259 base = _gvn.transform(new CastX2PNode(offset)); 2260 offset = MakeConX(0); 2261 return Type::RawPtr; 2262 } else if (base_type->base() == Type::RawPtr) { 2263 return Type::RawPtr; 2264 } else if (base_type->isa_oopptr()) { 2265 // Base is never null => always a heap address. 2266 if (base_type->ptr() == TypePtr::NotNull) { 2267 return Type::OopPtr; 2268 } 2269 // Offset is small => always a heap address. 2270 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2271 if (offset_type != NULL && 2272 base_type->offset() == 0 && // (should always be?) 2273 offset_type->_lo >= 0 && 2274 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2275 return Type::OopPtr; 2276 } 2277 // Otherwise, it might either be oop+off or NULL+addr. 2278 return Type::AnyPtr; 2279 } else { 2280 // No information: 2281 return Type::AnyPtr; 2282 } 2283 } 2284 2285 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2286 int kind = classify_unsafe_addr(base, offset); 2287 if (kind == Type::RawPtr) { 2288 return basic_plus_adr(top(), base, offset); 2289 } else { 2290 return basic_plus_adr(base, offset); 2291 } 2292 } 2293 2294 //--------------------------inline_number_methods----------------------------- 2295 // inline int Integer.numberOfLeadingZeros(int) 2296 // inline int Long.numberOfLeadingZeros(long) 2297 // 2298 // inline int Integer.numberOfTrailingZeros(int) 2299 // inline int Long.numberOfTrailingZeros(long) 2300 // 2301 // inline int Integer.bitCount(int) 2302 // inline int Long.bitCount(long) 2303 // 2304 // inline char Character.reverseBytes(char) 2305 // inline short Short.reverseBytes(short) 2306 // inline int Integer.reverseBytes(int) 2307 // inline long Long.reverseBytes(long) 2308 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2309 Node* arg = argument(0); 2310 Node* n; 2311 switch (id) { 2312 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2313 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2314 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2315 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2316 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2317 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2318 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2319 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2320 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2321 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2322 default: fatal_unexpected_iid(id); break; 2323 } 2324 set_result(_gvn.transform(n)); 2325 return true; 2326 } 2327 2328 //----------------------------inline_unsafe_access---------------------------- 2329 2330 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2331 2332 // Helper that guards and inserts a pre-barrier. 2333 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2334 Node* pre_val, bool need_mem_bar) { 2335 // We could be accessing the referent field of a reference object. If so, when G1 2336 // is enabled, we need to log the value in the referent field in an SATB buffer. 2337 // This routine performs some compile time filters and generates suitable 2338 // runtime filters that guard the pre-barrier code. 2339 // Also add memory barrier for non volatile load from the referent field 2340 // to prevent commoning of loads across safepoint. 2341 if (!UseG1GC && !need_mem_bar) 2342 return; 2343 2344 // Some compile time checks. 2345 2346 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2347 const TypeX* otype = offset->find_intptr_t_type(); 2348 if (otype != NULL && otype->is_con() && 2349 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2350 // Constant offset but not the reference_offset so just return 2351 return; 2352 } 2353 2354 // We only need to generate the runtime guards for instances. 2355 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2356 if (btype != NULL) { 2357 if (btype->isa_aryptr()) { 2358 // Array type so nothing to do 2359 return; 2360 } 2361 2362 const TypeInstPtr* itype = btype->isa_instptr(); 2363 if (itype != NULL) { 2364 // Can the klass of base_oop be statically determined to be 2365 // _not_ a sub-class of Reference and _not_ Object? 2366 ciKlass* klass = itype->klass(); 2367 if ( klass->is_loaded() && 2368 !klass->is_subtype_of(env()->Reference_klass()) && 2369 !env()->Object_klass()->is_subtype_of(klass)) { 2370 return; 2371 } 2372 } 2373 } 2374 2375 // The compile time filters did not reject base_oop/offset so 2376 // we need to generate the following runtime filters 2377 // 2378 // if (offset == java_lang_ref_Reference::_reference_offset) { 2379 // if (instance_of(base, java.lang.ref.Reference)) { 2380 // pre_barrier(_, pre_val, ...); 2381 // } 2382 // } 2383 2384 float likely = PROB_LIKELY( 0.999); 2385 float unlikely = PROB_UNLIKELY(0.999); 2386 2387 IdealKit ideal(this); 2388 #define __ ideal. 2389 2390 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2391 2392 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2393 // Update graphKit memory and control from IdealKit. 2394 sync_kit(ideal); 2395 2396 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2397 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2398 2399 // Update IdealKit memory and control from graphKit. 2400 __ sync_kit(this); 2401 2402 Node* one = __ ConI(1); 2403 // is_instof == 0 if base_oop == NULL 2404 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2405 2406 // Update graphKit from IdeakKit. 2407 sync_kit(ideal); 2408 2409 // Use the pre-barrier to record the value in the referent field 2410 pre_barrier(false /* do_load */, 2411 __ ctrl(), 2412 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2413 pre_val /* pre_val */, 2414 T_OBJECT); 2415 if (need_mem_bar) { 2416 // Add memory barrier to prevent commoning reads from this field 2417 // across safepoint since GC can change its value. 2418 insert_mem_bar(Op_MemBarCPUOrder); 2419 } 2420 // Update IdealKit from graphKit. 2421 __ sync_kit(this); 2422 2423 } __ end_if(); // _ref_type != ref_none 2424 } __ end_if(); // offset == referent_offset 2425 2426 // Final sync IdealKit and GraphKit. 2427 final_sync(ideal); 2428 #undef __ 2429 } 2430 2431 2432 // Interpret Unsafe.fieldOffset cookies correctly: 2433 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2434 2435 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2436 // Attempt to infer a sharper value type from the offset and base type. 2437 ciKlass* sharpened_klass = NULL; 2438 2439 // See if it is an instance field, with an object type. 2440 if (alias_type->field() != NULL) { 2441 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2442 if (alias_type->field()->type()->is_klass()) { 2443 sharpened_klass = alias_type->field()->type()->as_klass(); 2444 } 2445 } 2446 2447 // See if it is a narrow oop array. 2448 if (adr_type->isa_aryptr()) { 2449 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2450 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2451 if (elem_type != NULL) { 2452 sharpened_klass = elem_type->klass(); 2453 } 2454 } 2455 } 2456 2457 // The sharpened class might be unloaded if there is no class loader 2458 // contraint in place. 2459 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2460 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2461 2462 #ifndef PRODUCT 2463 if (C->print_intrinsics() || C->print_inlining()) { 2464 tty->print(" from base type: "); adr_type->dump(); 2465 tty->print(" sharpened value: "); tjp->dump(); 2466 } 2467 #endif 2468 // Sharpen the value type. 2469 return tjp; 2470 } 2471 return NULL; 2472 } 2473 2474 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2475 if (callee()->is_static()) return false; // caller must have the capability! 2476 2477 #ifndef PRODUCT 2478 { 2479 ResourceMark rm; 2480 // Check the signatures. 2481 ciSignature* sig = callee()->signature(); 2482 #ifdef ASSERT 2483 if (!is_store) { 2484 // Object getObject(Object base, int/long offset), etc. 2485 BasicType rtype = sig->return_type()->basic_type(); 2486 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2487 rtype = T_ADDRESS; // it is really a C void* 2488 assert(rtype == type, "getter must return the expected value"); 2489 if (!is_native_ptr) { 2490 assert(sig->count() == 2, "oop getter has 2 arguments"); 2491 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2492 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2493 } else { 2494 assert(sig->count() == 1, "native getter has 1 argument"); 2495 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2496 } 2497 } else { 2498 // void putObject(Object base, int/long offset, Object x), etc. 2499 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2500 if (!is_native_ptr) { 2501 assert(sig->count() == 3, "oop putter has 3 arguments"); 2502 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2503 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2504 } else { 2505 assert(sig->count() == 2, "native putter has 2 arguments"); 2506 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2507 } 2508 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2509 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2510 vtype = T_ADDRESS; // it is really a C void* 2511 assert(vtype == type, "putter must accept the expected value"); 2512 } 2513 #endif // ASSERT 2514 } 2515 #endif //PRODUCT 2516 2517 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2518 2519 Node* receiver = argument(0); // type: oop 2520 2521 // Build address expression. See the code in inline_unsafe_prefetch. 2522 Node* adr; 2523 Node* heap_base_oop = top(); 2524 Node* offset = top(); 2525 Node* val; 2526 2527 if (!is_native_ptr) { 2528 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2529 Node* base = argument(1); // type: oop 2530 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2531 offset = argument(2); // type: long 2532 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2533 // to be plain byte offsets, which are also the same as those accepted 2534 // by oopDesc::field_base. 2535 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2536 "fieldOffset must be byte-scaled"); 2537 // 32-bit machines ignore the high half! 2538 offset = ConvL2X(offset); 2539 adr = make_unsafe_address(base, offset); 2540 heap_base_oop = base; 2541 val = is_store ? argument(4) : NULL; 2542 } else { 2543 Node* ptr = argument(1); // type: long 2544 ptr = ConvL2X(ptr); // adjust Java long to machine word 2545 adr = make_unsafe_address(NULL, ptr); 2546 val = is_store ? argument(3) : NULL; 2547 } 2548 2549 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2550 2551 // First guess at the value type. 2552 const Type *value_type = Type::get_const_basic_type(type); 2553 2554 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2555 // there was not enough information to nail it down. 2556 Compile::AliasType* alias_type = C->alias_type(adr_type); 2557 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2558 2559 // We will need memory barriers unless we can determine a unique 2560 // alias category for this reference. (Note: If for some reason 2561 // the barriers get omitted and the unsafe reference begins to "pollute" 2562 // the alias analysis of the rest of the graph, either Compile::can_alias 2563 // or Compile::must_alias will throw a diagnostic assert.) 2564 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2565 2566 // If we are reading the value of the referent field of a Reference 2567 // object (either by using Unsafe directly or through reflection) 2568 // then, if G1 is enabled, we need to record the referent in an 2569 // SATB log buffer using the pre-barrier mechanism. 2570 // Also we need to add memory barrier to prevent commoning reads 2571 // from this field across safepoint since GC can change its value. 2572 bool need_read_barrier = !is_native_ptr && !is_store && 2573 offset != top() && heap_base_oop != top(); 2574 2575 if (!is_store && type == T_OBJECT) { 2576 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2577 if (tjp != NULL) { 2578 value_type = tjp; 2579 } 2580 } 2581 2582 receiver = null_check(receiver); 2583 if (stopped()) { 2584 return true; 2585 } 2586 // Heap pointers get a null-check from the interpreter, 2587 // as a courtesy. However, this is not guaranteed by Unsafe, 2588 // and it is not possible to fully distinguish unintended nulls 2589 // from intended ones in this API. 2590 2591 if (is_volatile) { 2592 // We need to emit leading and trailing CPU membars (see below) in 2593 // addition to memory membars when is_volatile. This is a little 2594 // too strong, but avoids the need to insert per-alias-type 2595 // volatile membars (for stores; compare Parse::do_put_xxx), which 2596 // we cannot do effectively here because we probably only have a 2597 // rough approximation of type. 2598 need_mem_bar = true; 2599 // For Stores, place a memory ordering barrier now. 2600 if (is_store) { 2601 insert_mem_bar(Op_MemBarRelease); 2602 } else { 2603 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2604 insert_mem_bar(Op_MemBarVolatile); 2605 } 2606 } 2607 } 2608 2609 // Memory barrier to prevent normal and 'unsafe' accesses from 2610 // bypassing each other. Happens after null checks, so the 2611 // exception paths do not take memory state from the memory barrier, 2612 // so there's no problems making a strong assert about mixing users 2613 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2614 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2615 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2616 2617 if (!is_store) { 2618 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2619 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile); 2620 // load value 2621 switch (type) { 2622 case T_BOOLEAN: 2623 case T_CHAR: 2624 case T_BYTE: 2625 case T_SHORT: 2626 case T_INT: 2627 case T_LONG: 2628 case T_FLOAT: 2629 case T_DOUBLE: 2630 break; 2631 case T_OBJECT: 2632 if (need_read_barrier) { 2633 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2634 } 2635 break; 2636 case T_ADDRESS: 2637 // Cast to an int type. 2638 p = _gvn.transform(new CastP2XNode(NULL, p)); 2639 p = ConvX2UL(p); 2640 break; 2641 default: 2642 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2643 break; 2644 } 2645 // The load node has the control of the preceding MemBarCPUOrder. All 2646 // following nodes will have the control of the MemBarCPUOrder inserted at 2647 // the end of this method. So, pushing the load onto the stack at a later 2648 // point is fine. 2649 set_result(p); 2650 } else { 2651 // place effect of store into memory 2652 switch (type) { 2653 case T_DOUBLE: 2654 val = dstore_rounding(val); 2655 break; 2656 case T_ADDRESS: 2657 // Repackage the long as a pointer. 2658 val = ConvL2X(val); 2659 val = _gvn.transform(new CastX2PNode(val)); 2660 break; 2661 } 2662 2663 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2664 if (type != T_OBJECT ) { 2665 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2666 } else { 2667 // Possibly an oop being stored to Java heap or native memory 2668 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2669 // oop to Java heap. 2670 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2671 } else { 2672 // We can't tell at compile time if we are storing in the Java heap or outside 2673 // of it. So we need to emit code to conditionally do the proper type of 2674 // store. 2675 2676 IdealKit ideal(this); 2677 #define __ ideal. 2678 // QQQ who knows what probability is here?? 2679 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2680 // Sync IdealKit and graphKit. 2681 sync_kit(ideal); 2682 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2683 // Update IdealKit memory. 2684 __ sync_kit(this); 2685 } __ else_(); { 2686 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2687 } __ end_if(); 2688 // Final sync IdealKit and GraphKit. 2689 final_sync(ideal); 2690 #undef __ 2691 } 2692 } 2693 } 2694 2695 if (is_volatile) { 2696 if (!is_store) { 2697 insert_mem_bar(Op_MemBarAcquire); 2698 } else { 2699 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2700 insert_mem_bar(Op_MemBarVolatile); 2701 } 2702 } 2703 } 2704 2705 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2706 2707 return true; 2708 } 2709 2710 //----------------------------inline_unsafe_prefetch---------------------------- 2711 2712 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2713 #ifndef PRODUCT 2714 { 2715 ResourceMark rm; 2716 // Check the signatures. 2717 ciSignature* sig = callee()->signature(); 2718 #ifdef ASSERT 2719 // Object getObject(Object base, int/long offset), etc. 2720 BasicType rtype = sig->return_type()->basic_type(); 2721 if (!is_native_ptr) { 2722 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2723 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2724 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2725 } else { 2726 assert(sig->count() == 1, "native prefetch has 1 argument"); 2727 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2728 } 2729 #endif // ASSERT 2730 } 2731 #endif // !PRODUCT 2732 2733 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2734 2735 const int idx = is_static ? 0 : 1; 2736 if (!is_static) { 2737 null_check_receiver(); 2738 if (stopped()) { 2739 return true; 2740 } 2741 } 2742 2743 // Build address expression. See the code in inline_unsafe_access. 2744 Node *adr; 2745 if (!is_native_ptr) { 2746 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2747 Node* base = argument(idx + 0); // type: oop 2748 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2749 Node* offset = argument(idx + 1); // type: long 2750 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2751 // to be plain byte offsets, which are also the same as those accepted 2752 // by oopDesc::field_base. 2753 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2754 "fieldOffset must be byte-scaled"); 2755 // 32-bit machines ignore the high half! 2756 offset = ConvL2X(offset); 2757 adr = make_unsafe_address(base, offset); 2758 } else { 2759 Node* ptr = argument(idx + 0); // type: long 2760 ptr = ConvL2X(ptr); // adjust Java long to machine word 2761 adr = make_unsafe_address(NULL, ptr); 2762 } 2763 2764 // Generate the read or write prefetch 2765 Node *prefetch; 2766 if (is_store) { 2767 prefetch = new PrefetchWriteNode(i_o(), adr); 2768 } else { 2769 prefetch = new PrefetchReadNode(i_o(), adr); 2770 } 2771 prefetch->init_req(0, control()); 2772 set_i_o(_gvn.transform(prefetch)); 2773 2774 return true; 2775 } 2776 2777 //----------------------------inline_unsafe_load_store---------------------------- 2778 // This method serves a couple of different customers (depending on LoadStoreKind): 2779 // 2780 // LS_cmpxchg: 2781 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2782 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2783 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2784 // 2785 // LS_xadd: 2786 // public int getAndAddInt( Object o, long offset, int delta) 2787 // public long getAndAddLong(Object o, long offset, long delta) 2788 // 2789 // LS_xchg: 2790 // int getAndSet(Object o, long offset, int newValue) 2791 // long getAndSet(Object o, long offset, long newValue) 2792 // Object getAndSet(Object o, long offset, Object newValue) 2793 // 2794 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2795 // This basic scheme here is the same as inline_unsafe_access, but 2796 // differs in enough details that combining them would make the code 2797 // overly confusing. (This is a true fact! I originally combined 2798 // them, but even I was confused by it!) As much code/comments as 2799 // possible are retained from inline_unsafe_access though to make 2800 // the correspondences clearer. - dl 2801 2802 if (callee()->is_static()) return false; // caller must have the capability! 2803 2804 #ifndef PRODUCT 2805 BasicType rtype; 2806 { 2807 ResourceMark rm; 2808 // Check the signatures. 2809 ciSignature* sig = callee()->signature(); 2810 rtype = sig->return_type()->basic_type(); 2811 if (kind == LS_xadd || kind == LS_xchg) { 2812 // Check the signatures. 2813 #ifdef ASSERT 2814 assert(rtype == type, "get and set must return the expected type"); 2815 assert(sig->count() == 3, "get and set has 3 arguments"); 2816 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2817 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2818 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2819 #endif // ASSERT 2820 } else if (kind == LS_cmpxchg) { 2821 // Check the signatures. 2822 #ifdef ASSERT 2823 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2824 assert(sig->count() == 4, "CAS has 4 arguments"); 2825 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2826 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2827 #endif // ASSERT 2828 } else { 2829 ShouldNotReachHere(); 2830 } 2831 } 2832 #endif //PRODUCT 2833 2834 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2835 2836 // Get arguments: 2837 Node* receiver = NULL; 2838 Node* base = NULL; 2839 Node* offset = NULL; 2840 Node* oldval = NULL; 2841 Node* newval = NULL; 2842 if (kind == LS_cmpxchg) { 2843 const bool two_slot_type = type2size[type] == 2; 2844 receiver = argument(0); // type: oop 2845 base = argument(1); // type: oop 2846 offset = argument(2); // type: long 2847 oldval = argument(4); // type: oop, int, or long 2848 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2849 } else if (kind == LS_xadd || kind == LS_xchg){ 2850 receiver = argument(0); // type: oop 2851 base = argument(1); // type: oop 2852 offset = argument(2); // type: long 2853 oldval = NULL; 2854 newval = argument(4); // type: oop, int, or long 2855 } 2856 2857 // Null check receiver. 2858 receiver = null_check(receiver); 2859 if (stopped()) { 2860 return true; 2861 } 2862 2863 // Build field offset expression. 2864 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2865 // to be plain byte offsets, which are also the same as those accepted 2866 // by oopDesc::field_base. 2867 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2868 // 32-bit machines ignore the high half of long offsets 2869 offset = ConvL2X(offset); 2870 Node* adr = make_unsafe_address(base, offset); 2871 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2872 2873 // For CAS, unlike inline_unsafe_access, there seems no point in 2874 // trying to refine types. Just use the coarse types here. 2875 const Type *value_type = Type::get_const_basic_type(type); 2876 Compile::AliasType* alias_type = C->alias_type(adr_type); 2877 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2878 2879 if (kind == LS_xchg && type == T_OBJECT) { 2880 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2881 if (tjp != NULL) { 2882 value_type = tjp; 2883 } 2884 } 2885 2886 int alias_idx = C->get_alias_index(adr_type); 2887 2888 // Memory-model-wise, a LoadStore acts like a little synchronized 2889 // block, so needs barriers on each side. These don't translate 2890 // into actual barriers on most machines, but we still need rest of 2891 // compiler to respect ordering. 2892 2893 insert_mem_bar(Op_MemBarRelease); 2894 insert_mem_bar(Op_MemBarCPUOrder); 2895 2896 // 4984716: MemBars must be inserted before this 2897 // memory node in order to avoid a false 2898 // dependency which will confuse the scheduler. 2899 Node *mem = memory(alias_idx); 2900 2901 // For now, we handle only those cases that actually exist: ints, 2902 // longs, and Object. Adding others should be straightforward. 2903 Node* load_store; 2904 switch(type) { 2905 case T_INT: 2906 if (kind == LS_xadd) { 2907 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2908 } else if (kind == LS_xchg) { 2909 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2910 } else if (kind == LS_cmpxchg) { 2911 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2912 } else { 2913 ShouldNotReachHere(); 2914 } 2915 break; 2916 case T_LONG: 2917 if (kind == LS_xadd) { 2918 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2919 } else if (kind == LS_xchg) { 2920 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2921 } else if (kind == LS_cmpxchg) { 2922 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2923 } else { 2924 ShouldNotReachHere(); 2925 } 2926 break; 2927 case T_OBJECT: 2928 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2929 // could be delayed during Parse (for example, in adjust_map_after_if()). 2930 // Execute transformation here to avoid barrier generation in such case. 2931 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2932 newval = _gvn.makecon(TypePtr::NULL_PTR); 2933 2934 // Reference stores need a store barrier. 2935 if (kind == LS_xchg) { 2936 // If pre-barrier must execute before the oop store, old value will require do_load here. 2937 if (!can_move_pre_barrier()) { 2938 pre_barrier(true /* do_load*/, 2939 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2940 NULL /* pre_val*/, 2941 T_OBJECT); 2942 } // Else move pre_barrier to use load_store value, see below. 2943 } else if (kind == LS_cmpxchg) { 2944 // Same as for newval above: 2945 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2946 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2947 } 2948 // The only known value which might get overwritten is oldval. 2949 pre_barrier(false /* do_load */, 2950 control(), NULL, NULL, max_juint, NULL, NULL, 2951 oldval /* pre_val */, 2952 T_OBJECT); 2953 } else { 2954 ShouldNotReachHere(); 2955 } 2956 2957 #ifdef _LP64 2958 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2959 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2960 if (kind == LS_xchg) { 2961 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2962 newval_enc, adr_type, value_type->make_narrowoop())); 2963 } else { 2964 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2965 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2966 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2967 newval_enc, oldval_enc)); 2968 } 2969 } else 2970 #endif 2971 { 2972 if (kind == LS_xchg) { 2973 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2974 } else { 2975 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2976 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2977 } 2978 } 2979 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2980 break; 2981 default: 2982 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2983 break; 2984 } 2985 2986 // SCMemProjNodes represent the memory state of a LoadStore. Their 2987 // main role is to prevent LoadStore nodes from being optimized away 2988 // when their results aren't used. 2989 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2990 set_memory(proj, alias_idx); 2991 2992 if (type == T_OBJECT && kind == LS_xchg) { 2993 #ifdef _LP64 2994 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2995 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2996 } 2997 #endif 2998 if (can_move_pre_barrier()) { 2999 // Don't need to load pre_val. The old value is returned by load_store. 3000 // The pre_barrier can execute after the xchg as long as no safepoint 3001 // gets inserted between them. 3002 pre_barrier(false /* do_load */, 3003 control(), NULL, NULL, max_juint, NULL, NULL, 3004 load_store /* pre_val */, 3005 T_OBJECT); 3006 } 3007 } 3008 3009 // Add the trailing membar surrounding the access 3010 insert_mem_bar(Op_MemBarCPUOrder); 3011 insert_mem_bar(Op_MemBarAcquire); 3012 3013 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3014 set_result(load_store); 3015 return true; 3016 } 3017 3018 //----------------------------inline_unsafe_ordered_store---------------------- 3019 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3020 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3021 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3022 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3023 // This is another variant of inline_unsafe_access, differing in 3024 // that it always issues store-store ("release") barrier and ensures 3025 // store-atomicity (which only matters for "long"). 3026 3027 if (callee()->is_static()) return false; // caller must have the capability! 3028 3029 #ifndef PRODUCT 3030 { 3031 ResourceMark rm; 3032 // Check the signatures. 3033 ciSignature* sig = callee()->signature(); 3034 #ifdef ASSERT 3035 BasicType rtype = sig->return_type()->basic_type(); 3036 assert(rtype == T_VOID, "must return void"); 3037 assert(sig->count() == 3, "has 3 arguments"); 3038 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3039 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3040 #endif // ASSERT 3041 } 3042 #endif //PRODUCT 3043 3044 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3045 3046 // Get arguments: 3047 Node* receiver = argument(0); // type: oop 3048 Node* base = argument(1); // type: oop 3049 Node* offset = argument(2); // type: long 3050 Node* val = argument(4); // type: oop, int, or long 3051 3052 // Null check receiver. 3053 receiver = null_check(receiver); 3054 if (stopped()) { 3055 return true; 3056 } 3057 3058 // Build field offset expression. 3059 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3060 // 32-bit machines ignore the high half of long offsets 3061 offset = ConvL2X(offset); 3062 Node* adr = make_unsafe_address(base, offset); 3063 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3064 const Type *value_type = Type::get_const_basic_type(type); 3065 Compile::AliasType* alias_type = C->alias_type(adr_type); 3066 3067 insert_mem_bar(Op_MemBarRelease); 3068 insert_mem_bar(Op_MemBarCPUOrder); 3069 // Ensure that the store is atomic for longs: 3070 const bool require_atomic_access = true; 3071 Node* store; 3072 if (type == T_OBJECT) // reference stores need a store barrier. 3073 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3074 else { 3075 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3076 } 3077 insert_mem_bar(Op_MemBarCPUOrder); 3078 return true; 3079 } 3080 3081 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3082 // Regardless of form, don't allow previous ld/st to move down, 3083 // then issue acquire, release, or volatile mem_bar. 3084 insert_mem_bar(Op_MemBarCPUOrder); 3085 switch(id) { 3086 case vmIntrinsics::_loadFence: 3087 insert_mem_bar(Op_LoadFence); 3088 return true; 3089 case vmIntrinsics::_storeFence: 3090 insert_mem_bar(Op_StoreFence); 3091 return true; 3092 case vmIntrinsics::_fullFence: 3093 insert_mem_bar(Op_MemBarVolatile); 3094 return true; 3095 default: 3096 fatal_unexpected_iid(id); 3097 return false; 3098 } 3099 } 3100 3101 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3102 if (!kls->is_Con()) { 3103 return true; 3104 } 3105 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3106 if (klsptr == NULL) { 3107 return true; 3108 } 3109 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3110 // don't need a guard for a klass that is already initialized 3111 return !ik->is_initialized(); 3112 } 3113 3114 //----------------------------inline_unsafe_allocate--------------------------- 3115 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3116 bool LibraryCallKit::inline_unsafe_allocate() { 3117 if (callee()->is_static()) return false; // caller must have the capability! 3118 3119 null_check_receiver(); // null-check, then ignore 3120 Node* cls = null_check(argument(1)); 3121 if (stopped()) return true; 3122 3123 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3124 kls = null_check(kls); 3125 if (stopped()) return true; // argument was like int.class 3126 3127 Node* test = NULL; 3128 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3129 // Note: The argument might still be an illegal value like 3130 // Serializable.class or Object[].class. The runtime will handle it. 3131 // But we must make an explicit check for initialization. 3132 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3133 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3134 // can generate code to load it as unsigned byte. 3135 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3136 Node* bits = intcon(InstanceKlass::fully_initialized); 3137 test = _gvn.transform(new SubINode(inst, bits)); 3138 // The 'test' is non-zero if we need to take a slow path. 3139 } 3140 3141 Node* obj = new_instance(kls, test); 3142 set_result(obj); 3143 return true; 3144 } 3145 3146 #ifdef TRACE_HAVE_INTRINSICS 3147 /* 3148 * oop -> myklass 3149 * myklass->trace_id |= USED 3150 * return myklass->trace_id & ~0x3 3151 */ 3152 bool LibraryCallKit::inline_native_classID() { 3153 null_check_receiver(); // null-check, then ignore 3154 Node* cls = null_check(argument(1), T_OBJECT); 3155 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3156 kls = null_check(kls, T_OBJECT); 3157 ByteSize offset = TRACE_ID_OFFSET; 3158 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3159 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3160 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3161 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3162 Node* clsused = longcon(0x01l); // set the class bit 3163 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3164 3165 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3166 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3167 set_result(andl); 3168 return true; 3169 } 3170 3171 bool LibraryCallKit::inline_native_threadID() { 3172 Node* tls_ptr = NULL; 3173 Node* cur_thr = generate_current_thread(tls_ptr); 3174 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3175 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3176 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3177 3178 Node* threadid = NULL; 3179 size_t thread_id_size = OSThread::thread_id_size(); 3180 if (thread_id_size == (size_t) BytesPerLong) { 3181 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3182 } else if (thread_id_size == (size_t) BytesPerInt) { 3183 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3184 } else { 3185 ShouldNotReachHere(); 3186 } 3187 set_result(threadid); 3188 return true; 3189 } 3190 #endif 3191 3192 //------------------------inline_native_time_funcs-------------- 3193 // inline code for System.currentTimeMillis() and System.nanoTime() 3194 // these have the same type and signature 3195 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3196 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3197 const TypePtr* no_memory_effects = NULL; 3198 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3199 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3200 #ifdef ASSERT 3201 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3202 assert(value_top == top(), "second value must be top"); 3203 #endif 3204 set_result(value); 3205 return true; 3206 } 3207 3208 //------------------------inline_native_currentThread------------------ 3209 bool LibraryCallKit::inline_native_currentThread() { 3210 Node* junk = NULL; 3211 set_result(generate_current_thread(junk)); 3212 return true; 3213 } 3214 3215 //------------------------inline_native_isInterrupted------------------ 3216 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3217 bool LibraryCallKit::inline_native_isInterrupted() { 3218 // Add a fast path to t.isInterrupted(clear_int): 3219 // (t == Thread.current() && 3220 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3221 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3222 // So, in the common case that the interrupt bit is false, 3223 // we avoid making a call into the VM. Even if the interrupt bit 3224 // is true, if the clear_int argument is false, we avoid the VM call. 3225 // However, if the receiver is not currentThread, we must call the VM, 3226 // because there must be some locking done around the operation. 3227 3228 // We only go to the fast case code if we pass two guards. 3229 // Paths which do not pass are accumulated in the slow_region. 3230 3231 enum { 3232 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3233 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3234 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3235 PATH_LIMIT 3236 }; 3237 3238 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3239 // out of the function. 3240 insert_mem_bar(Op_MemBarCPUOrder); 3241 3242 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3243 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3244 3245 RegionNode* slow_region = new RegionNode(1); 3246 record_for_igvn(slow_region); 3247 3248 // (a) Receiving thread must be the current thread. 3249 Node* rec_thr = argument(0); 3250 Node* tls_ptr = NULL; 3251 Node* cur_thr = generate_current_thread(tls_ptr); 3252 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3253 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3254 3255 generate_slow_guard(bol_thr, slow_region); 3256 3257 // (b) Interrupt bit on TLS must be false. 3258 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3259 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3260 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3261 3262 // Set the control input on the field _interrupted read to prevent it floating up. 3263 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3264 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3265 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3266 3267 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3268 3269 // First fast path: if (!TLS._interrupted) return false; 3270 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3271 result_rgn->init_req(no_int_result_path, false_bit); 3272 result_val->init_req(no_int_result_path, intcon(0)); 3273 3274 // drop through to next case 3275 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3276 3277 #ifndef TARGET_OS_FAMILY_windows 3278 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3279 Node* clr_arg = argument(1); 3280 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3281 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3282 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3283 3284 // Second fast path: ... else if (!clear_int) return true; 3285 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3286 result_rgn->init_req(no_clear_result_path, false_arg); 3287 result_val->init_req(no_clear_result_path, intcon(1)); 3288 3289 // drop through to next case 3290 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3291 #else 3292 // To return true on Windows you must read the _interrupted field 3293 // and check the the event state i.e. take the slow path. 3294 #endif // TARGET_OS_FAMILY_windows 3295 3296 // (d) Otherwise, go to the slow path. 3297 slow_region->add_req(control()); 3298 set_control( _gvn.transform(slow_region)); 3299 3300 if (stopped()) { 3301 // There is no slow path. 3302 result_rgn->init_req(slow_result_path, top()); 3303 result_val->init_req(slow_result_path, top()); 3304 } else { 3305 // non-virtual because it is a private non-static 3306 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3307 3308 Node* slow_val = set_results_for_java_call(slow_call); 3309 // this->control() comes from set_results_for_java_call 3310 3311 Node* fast_io = slow_call->in(TypeFunc::I_O); 3312 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3313 3314 // These two phis are pre-filled with copies of of the fast IO and Memory 3315 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3316 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3317 3318 result_rgn->init_req(slow_result_path, control()); 3319 result_io ->init_req(slow_result_path, i_o()); 3320 result_mem->init_req(slow_result_path, reset_memory()); 3321 result_val->init_req(slow_result_path, slow_val); 3322 3323 set_all_memory(_gvn.transform(result_mem)); 3324 set_i_o( _gvn.transform(result_io)); 3325 } 3326 3327 C->set_has_split_ifs(true); // Has chance for split-if optimization 3328 set_result(result_rgn, result_val); 3329 return true; 3330 } 3331 3332 //---------------------------load_mirror_from_klass---------------------------- 3333 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3334 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3335 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3336 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3337 } 3338 3339 //-----------------------load_klass_from_mirror_common------------------------- 3340 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3341 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3342 // and branch to the given path on the region. 3343 // If never_see_null, take an uncommon trap on null, so we can optimistically 3344 // compile for the non-null case. 3345 // If the region is NULL, force never_see_null = true. 3346 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3347 bool never_see_null, 3348 RegionNode* region, 3349 int null_path, 3350 int offset) { 3351 if (region == NULL) never_see_null = true; 3352 Node* p = basic_plus_adr(mirror, offset); 3353 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3354 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3355 Node* null_ctl = top(); 3356 kls = null_check_oop(kls, &null_ctl, never_see_null); 3357 if (region != NULL) { 3358 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3359 region->init_req(null_path, null_ctl); 3360 } else { 3361 assert(null_ctl == top(), "no loose ends"); 3362 } 3363 return kls; 3364 } 3365 3366 //--------------------(inline_native_Class_query helpers)--------------------- 3367 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3368 // Fall through if (mods & mask) == bits, take the guard otherwise. 3369 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3370 // Branch around if the given klass has the given modifier bit set. 3371 // Like generate_guard, adds a new path onto the region. 3372 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3373 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3374 Node* mask = intcon(modifier_mask); 3375 Node* bits = intcon(modifier_bits); 3376 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3377 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3378 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3379 return generate_fair_guard(bol, region); 3380 } 3381 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3382 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3383 } 3384 3385 //-------------------------inline_native_Class_query------------------- 3386 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3387 const Type* return_type = TypeInt::BOOL; 3388 Node* prim_return_value = top(); // what happens if it's a primitive class? 3389 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3390 bool expect_prim = false; // most of these guys expect to work on refs 3391 3392 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3393 3394 Node* mirror = argument(0); 3395 Node* obj = top(); 3396 3397 switch (id) { 3398 case vmIntrinsics::_isInstance: 3399 // nothing is an instance of a primitive type 3400 prim_return_value = intcon(0); 3401 obj = argument(1); 3402 break; 3403 case vmIntrinsics::_getModifiers: 3404 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3405 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3406 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3407 break; 3408 case vmIntrinsics::_isInterface: 3409 prim_return_value = intcon(0); 3410 break; 3411 case vmIntrinsics::_isArray: 3412 prim_return_value = intcon(0); 3413 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3414 break; 3415 case vmIntrinsics::_isPrimitive: 3416 prim_return_value = intcon(1); 3417 expect_prim = true; // obviously 3418 break; 3419 case vmIntrinsics::_getSuperclass: 3420 prim_return_value = null(); 3421 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3422 break; 3423 case vmIntrinsics::_getClassAccessFlags: 3424 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3425 return_type = TypeInt::INT; // not bool! 6297094 3426 break; 3427 default: 3428 fatal_unexpected_iid(id); 3429 break; 3430 } 3431 3432 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3433 if (mirror_con == NULL) return false; // cannot happen? 3434 3435 #ifndef PRODUCT 3436 if (C->print_intrinsics() || C->print_inlining()) { 3437 ciType* k = mirror_con->java_mirror_type(); 3438 if (k) { 3439 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3440 k->print_name(); 3441 tty->cr(); 3442 } 3443 } 3444 #endif 3445 3446 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3447 RegionNode* region = new RegionNode(PATH_LIMIT); 3448 record_for_igvn(region); 3449 PhiNode* phi = new PhiNode(region, return_type); 3450 3451 // The mirror will never be null of Reflection.getClassAccessFlags, however 3452 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3453 // if it is. See bug 4774291. 3454 3455 // For Reflection.getClassAccessFlags(), the null check occurs in 3456 // the wrong place; see inline_unsafe_access(), above, for a similar 3457 // situation. 3458 mirror = null_check(mirror); 3459 // If mirror or obj is dead, only null-path is taken. 3460 if (stopped()) return true; 3461 3462 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3463 3464 // Now load the mirror's klass metaobject, and null-check it. 3465 // Side-effects region with the control path if the klass is null. 3466 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3467 // If kls is null, we have a primitive mirror. 3468 phi->init_req(_prim_path, prim_return_value); 3469 if (stopped()) { set_result(region, phi); return true; } 3470 bool safe_for_replace = (region->in(_prim_path) == top()); 3471 3472 Node* p; // handy temp 3473 Node* null_ctl; 3474 3475 // Now that we have the non-null klass, we can perform the real query. 3476 // For constant classes, the query will constant-fold in LoadNode::Value. 3477 Node* query_value = top(); 3478 switch (id) { 3479 case vmIntrinsics::_isInstance: 3480 // nothing is an instance of a primitive type 3481 query_value = gen_instanceof(obj, kls, safe_for_replace); 3482 break; 3483 3484 case vmIntrinsics::_getModifiers: 3485 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3486 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3487 break; 3488 3489 case vmIntrinsics::_isInterface: 3490 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3491 if (generate_interface_guard(kls, region) != NULL) 3492 // A guard was added. If the guard is taken, it was an interface. 3493 phi->add_req(intcon(1)); 3494 // If we fall through, it's a plain class. 3495 query_value = intcon(0); 3496 break; 3497 3498 case vmIntrinsics::_isArray: 3499 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3500 if (generate_array_guard(kls, region) != NULL) 3501 // A guard was added. If the guard is taken, it was an array. 3502 phi->add_req(intcon(1)); 3503 // If we fall through, it's a plain class. 3504 query_value = intcon(0); 3505 break; 3506 3507 case vmIntrinsics::_isPrimitive: 3508 query_value = intcon(0); // "normal" path produces false 3509 break; 3510 3511 case vmIntrinsics::_getSuperclass: 3512 // The rules here are somewhat unfortunate, but we can still do better 3513 // with random logic than with a JNI call. 3514 // Interfaces store null or Object as _super, but must report null. 3515 // Arrays store an intermediate super as _super, but must report Object. 3516 // Other types can report the actual _super. 3517 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3518 if (generate_interface_guard(kls, region) != NULL) 3519 // A guard was added. If the guard is taken, it was an interface. 3520 phi->add_req(null()); 3521 if (generate_array_guard(kls, region) != NULL) 3522 // A guard was added. If the guard is taken, it was an array. 3523 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3524 // If we fall through, it's a plain class. Get its _super. 3525 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3526 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3527 null_ctl = top(); 3528 kls = null_check_oop(kls, &null_ctl); 3529 if (null_ctl != top()) { 3530 // If the guard is taken, Object.superClass is null (both klass and mirror). 3531 region->add_req(null_ctl); 3532 phi ->add_req(null()); 3533 } 3534 if (!stopped()) { 3535 query_value = load_mirror_from_klass(kls); 3536 } 3537 break; 3538 3539 case vmIntrinsics::_getClassAccessFlags: 3540 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3541 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3542 break; 3543 3544 default: 3545 fatal_unexpected_iid(id); 3546 break; 3547 } 3548 3549 // Fall-through is the normal case of a query to a real class. 3550 phi->init_req(1, query_value); 3551 region->init_req(1, control()); 3552 3553 C->set_has_split_ifs(true); // Has chance for split-if optimization 3554 set_result(region, phi); 3555 return true; 3556 } 3557 3558 //-------------------------inline_Class_cast------------------- 3559 bool LibraryCallKit::inline_Class_cast() { 3560 Node* mirror = argument(0); // Class 3561 Node* obj = argument(1); 3562 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3563 if (mirror_con == NULL) { 3564 return false; // dead path (mirror->is_top()). 3565 } 3566 if (obj == NULL || obj->is_top()) { 3567 return false; // dead path 3568 } 3569 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3570 3571 // First, see if Class.cast() can be folded statically. 3572 // java_mirror_type() returns non-null for compile-time Class constants. 3573 ciType* tm = mirror_con->java_mirror_type(); 3574 if (tm != NULL && tm->is_klass() && 3575 tp != NULL && tp->klass() != NULL) { 3576 if (!tp->klass()->is_loaded()) { 3577 // Don't use intrinsic when class is not loaded. 3578 return false; 3579 } else { 3580 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3581 if (static_res == Compile::SSC_always_true) { 3582 // isInstance() is true - fold the code. 3583 set_result(obj); 3584 return true; 3585 } else if (static_res == Compile::SSC_always_false) { 3586 // Don't use intrinsic, have to throw ClassCastException. 3587 // If the reference is null, the non-intrinsic bytecode will 3588 // be optimized appropriately. 3589 return false; 3590 } 3591 } 3592 } 3593 3594 // Bailout intrinsic and do normal inlining if exception path is frequent. 3595 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3596 return false; 3597 } 3598 3599 // Generate dynamic checks. 3600 // Class.cast() is java implementation of _checkcast bytecode. 3601 // Do checkcast (Parse::do_checkcast()) optimizations here. 3602 3603 mirror = null_check(mirror); 3604 // If mirror is dead, only null-path is taken. 3605 if (stopped()) { 3606 return true; 3607 } 3608 3609 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3610 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3611 RegionNode* region = new RegionNode(PATH_LIMIT); 3612 record_for_igvn(region); 3613 3614 // Now load the mirror's klass metaobject, and null-check it. 3615 // If kls is null, we have a primitive mirror and 3616 // nothing is an instance of a primitive type. 3617 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3618 3619 Node* res = top(); 3620 if (!stopped()) { 3621 Node* bad_type_ctrl = top(); 3622 // Do checkcast optimizations. 3623 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3624 region->init_req(_bad_type_path, bad_type_ctrl); 3625 } 3626 if (region->in(_prim_path) != top() || 3627 region->in(_bad_type_path) != top()) { 3628 // Let Interpreter throw ClassCastException. 3629 PreserveJVMState pjvms(this); 3630 set_control(_gvn.transform(region)); 3631 uncommon_trap(Deoptimization::Reason_intrinsic, 3632 Deoptimization::Action_maybe_recompile); 3633 } 3634 if (!stopped()) { 3635 set_result(res); 3636 } 3637 return true; 3638 } 3639 3640 3641 //--------------------------inline_native_subtype_check------------------------ 3642 // This intrinsic takes the JNI calls out of the heart of 3643 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3644 bool LibraryCallKit::inline_native_subtype_check() { 3645 // Pull both arguments off the stack. 3646 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3647 args[0] = argument(0); 3648 args[1] = argument(1); 3649 Node* klasses[2]; // corresponding Klasses: superk, subk 3650 klasses[0] = klasses[1] = top(); 3651 3652 enum { 3653 // A full decision tree on {superc is prim, subc is prim}: 3654 _prim_0_path = 1, // {P,N} => false 3655 // {P,P} & superc!=subc => false 3656 _prim_same_path, // {P,P} & superc==subc => true 3657 _prim_1_path, // {N,P} => false 3658 _ref_subtype_path, // {N,N} & subtype check wins => true 3659 _both_ref_path, // {N,N} & subtype check loses => false 3660 PATH_LIMIT 3661 }; 3662 3663 RegionNode* region = new RegionNode(PATH_LIMIT); 3664 Node* phi = new PhiNode(region, TypeInt::BOOL); 3665 record_for_igvn(region); 3666 3667 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3668 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3669 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3670 3671 // First null-check both mirrors and load each mirror's klass metaobject. 3672 int which_arg; 3673 for (which_arg = 0; which_arg <= 1; which_arg++) { 3674 Node* arg = args[which_arg]; 3675 arg = null_check(arg); 3676 if (stopped()) break; 3677 args[which_arg] = arg; 3678 3679 Node* p = basic_plus_adr(arg, class_klass_offset); 3680 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3681 klasses[which_arg] = _gvn.transform(kls); 3682 } 3683 3684 // Having loaded both klasses, test each for null. 3685 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3686 for (which_arg = 0; which_arg <= 1; which_arg++) { 3687 Node* kls = klasses[which_arg]; 3688 Node* null_ctl = top(); 3689 kls = null_check_oop(kls, &null_ctl, never_see_null); 3690 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3691 region->init_req(prim_path, null_ctl); 3692 if (stopped()) break; 3693 klasses[which_arg] = kls; 3694 } 3695 3696 if (!stopped()) { 3697 // now we have two reference types, in klasses[0..1] 3698 Node* subk = klasses[1]; // the argument to isAssignableFrom 3699 Node* superk = klasses[0]; // the receiver 3700 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3701 // now we have a successful reference subtype check 3702 region->set_req(_ref_subtype_path, control()); 3703 } 3704 3705 // If both operands are primitive (both klasses null), then 3706 // we must return true when they are identical primitives. 3707 // It is convenient to test this after the first null klass check. 3708 set_control(region->in(_prim_0_path)); // go back to first null check 3709 if (!stopped()) { 3710 // Since superc is primitive, make a guard for the superc==subc case. 3711 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3712 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3713 generate_guard(bol_eq, region, PROB_FAIR); 3714 if (region->req() == PATH_LIMIT+1) { 3715 // A guard was added. If the added guard is taken, superc==subc. 3716 region->swap_edges(PATH_LIMIT, _prim_same_path); 3717 region->del_req(PATH_LIMIT); 3718 } 3719 region->set_req(_prim_0_path, control()); // Not equal after all. 3720 } 3721 3722 // these are the only paths that produce 'true': 3723 phi->set_req(_prim_same_path, intcon(1)); 3724 phi->set_req(_ref_subtype_path, intcon(1)); 3725 3726 // pull together the cases: 3727 assert(region->req() == PATH_LIMIT, "sane region"); 3728 for (uint i = 1; i < region->req(); i++) { 3729 Node* ctl = region->in(i); 3730 if (ctl == NULL || ctl == top()) { 3731 region->set_req(i, top()); 3732 phi ->set_req(i, top()); 3733 } else if (phi->in(i) == NULL) { 3734 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3735 } 3736 } 3737 3738 set_control(_gvn.transform(region)); 3739 set_result(_gvn.transform(phi)); 3740 return true; 3741 } 3742 3743 //---------------------generate_array_guard_common------------------------ 3744 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3745 bool obj_array, bool not_array) { 3746 // If obj_array/non_array==false/false: 3747 // Branch around if the given klass is in fact an array (either obj or prim). 3748 // If obj_array/non_array==false/true: 3749 // Branch around if the given klass is not an array klass of any kind. 3750 // If obj_array/non_array==true/true: 3751 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3752 // If obj_array/non_array==true/false: 3753 // Branch around if the kls is an oop array (Object[] or subtype) 3754 // 3755 // Like generate_guard, adds a new path onto the region. 3756 jint layout_con = 0; 3757 Node* layout_val = get_layout_helper(kls, layout_con); 3758 if (layout_val == NULL) { 3759 bool query = (obj_array 3760 ? Klass::layout_helper_is_objArray(layout_con) 3761 : Klass::layout_helper_is_array(layout_con)); 3762 if (query == not_array) { 3763 return NULL; // never a branch 3764 } else { // always a branch 3765 Node* always_branch = control(); 3766 if (region != NULL) 3767 region->add_req(always_branch); 3768 set_control(top()); 3769 return always_branch; 3770 } 3771 } 3772 // Now test the correct condition. 3773 jint nval = (obj_array 3774 ? ((jint)Klass::_lh_array_tag_type_value 3775 << Klass::_lh_array_tag_shift) 3776 : Klass::_lh_neutral_value); 3777 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3778 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3779 // invert the test if we are looking for a non-array 3780 if (not_array) btest = BoolTest(btest).negate(); 3781 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3782 return generate_fair_guard(bol, region); 3783 } 3784 3785 3786 //-----------------------inline_native_newArray-------------------------- 3787 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3788 bool LibraryCallKit::inline_native_newArray() { 3789 Node* mirror = argument(0); 3790 Node* count_val = argument(1); 3791 3792 mirror = null_check(mirror); 3793 // If mirror or obj is dead, only null-path is taken. 3794 if (stopped()) return true; 3795 3796 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3797 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3798 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3799 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3800 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3801 3802 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3803 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3804 result_reg, _slow_path); 3805 Node* normal_ctl = control(); 3806 Node* no_array_ctl = result_reg->in(_slow_path); 3807 3808 // Generate code for the slow case. We make a call to newArray(). 3809 set_control(no_array_ctl); 3810 if (!stopped()) { 3811 // Either the input type is void.class, or else the 3812 // array klass has not yet been cached. Either the 3813 // ensuing call will throw an exception, or else it 3814 // will cache the array klass for next time. 3815 PreserveJVMState pjvms(this); 3816 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3817 Node* slow_result = set_results_for_java_call(slow_call); 3818 // this->control() comes from set_results_for_java_call 3819 result_reg->set_req(_slow_path, control()); 3820 result_val->set_req(_slow_path, slow_result); 3821 result_io ->set_req(_slow_path, i_o()); 3822 result_mem->set_req(_slow_path, reset_memory()); 3823 } 3824 3825 set_control(normal_ctl); 3826 if (!stopped()) { 3827 // Normal case: The array type has been cached in the java.lang.Class. 3828 // The following call works fine even if the array type is polymorphic. 3829 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3830 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3831 result_reg->init_req(_normal_path, control()); 3832 result_val->init_req(_normal_path, obj); 3833 result_io ->init_req(_normal_path, i_o()); 3834 result_mem->init_req(_normal_path, reset_memory()); 3835 } 3836 3837 // Return the combined state. 3838 set_i_o( _gvn.transform(result_io) ); 3839 set_all_memory( _gvn.transform(result_mem)); 3840 3841 C->set_has_split_ifs(true); // Has chance for split-if optimization 3842 set_result(result_reg, result_val); 3843 return true; 3844 } 3845 3846 //----------------------inline_native_getLength-------------------------- 3847 // public static native int java.lang.reflect.Array.getLength(Object array); 3848 bool LibraryCallKit::inline_native_getLength() { 3849 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3850 3851 Node* array = null_check(argument(0)); 3852 // If array is dead, only null-path is taken. 3853 if (stopped()) return true; 3854 3855 // Deoptimize if it is a non-array. 3856 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3857 3858 if (non_array != NULL) { 3859 PreserveJVMState pjvms(this); 3860 set_control(non_array); 3861 uncommon_trap(Deoptimization::Reason_intrinsic, 3862 Deoptimization::Action_maybe_recompile); 3863 } 3864 3865 // If control is dead, only non-array-path is taken. 3866 if (stopped()) return true; 3867 3868 // The works fine even if the array type is polymorphic. 3869 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3870 Node* result = load_array_length(array); 3871 3872 C->set_has_split_ifs(true); // Has chance for split-if optimization 3873 set_result(result); 3874 return true; 3875 } 3876 3877 //------------------------inline_array_copyOf---------------------------- 3878 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3879 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3880 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3881 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3882 3883 // Get the arguments. 3884 Node* original = argument(0); 3885 Node* start = is_copyOfRange? argument(1): intcon(0); 3886 Node* end = is_copyOfRange? argument(2): argument(1); 3887 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3888 3889 Node* newcopy; 3890 3891 // Set the original stack and the reexecute bit for the interpreter to reexecute 3892 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3893 { PreserveReexecuteState preexecs(this); 3894 jvms()->set_should_reexecute(true); 3895 3896 array_type_mirror = null_check(array_type_mirror); 3897 original = null_check(original); 3898 3899 // Check if a null path was taken unconditionally. 3900 if (stopped()) return true; 3901 3902 Node* orig_length = load_array_length(original); 3903 3904 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3905 klass_node = null_check(klass_node); 3906 3907 RegionNode* bailout = new RegionNode(1); 3908 record_for_igvn(bailout); 3909 3910 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3911 // Bail out if that is so. 3912 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3913 if (not_objArray != NULL) { 3914 // Improve the klass node's type from the new optimistic assumption: 3915 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3916 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3917 Node* cast = new CastPPNode(klass_node, akls); 3918 cast->init_req(0, control()); 3919 klass_node = _gvn.transform(cast); 3920 } 3921 3922 // Bail out if either start or end is negative. 3923 generate_negative_guard(start, bailout, &start); 3924 generate_negative_guard(end, bailout, &end); 3925 3926 Node* length = end; 3927 if (_gvn.type(start) != TypeInt::ZERO) { 3928 length = _gvn.transform(new SubINode(end, start)); 3929 } 3930 3931 // Bail out if length is negative. 3932 // Without this the new_array would throw 3933 // NegativeArraySizeException but IllegalArgumentException is what 3934 // should be thrown 3935 generate_negative_guard(length, bailout, &length); 3936 3937 if (bailout->req() > 1) { 3938 PreserveJVMState pjvms(this); 3939 set_control(_gvn.transform(bailout)); 3940 uncommon_trap(Deoptimization::Reason_intrinsic, 3941 Deoptimization::Action_maybe_recompile); 3942 } 3943 3944 if (!stopped()) { 3945 // How many elements will we copy from the original? 3946 // The answer is MinI(orig_length - start, length). 3947 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3948 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3949 3950 newcopy = new_array(klass_node, length, 0); // no arguments to push 3951 3952 // Generate a direct call to the right arraycopy function(s). 3953 // We know the copy is disjoint but we might not know if the 3954 // oop stores need checking. 3955 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3956 // This will fail a store-check if x contains any non-nulls. 3957 3958 Node* alloc = tightly_coupled_allocation(newcopy, NULL); 3959 3960 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, alloc != NULL, 3961 load_object_klass(original), klass_node); 3962 if (!is_copyOfRange) { 3963 ac->set_copyof(); 3964 } else { 3965 ac->set_copyofrange(); 3966 } 3967 Node* n = _gvn.transform(ac); 3968 assert(n == ac, "cannot disappear"); 3969 ac->connect_outputs(this); 3970 } 3971 } // original reexecute is set back here 3972 3973 C->set_has_split_ifs(true); // Has chance for split-if optimization 3974 if (!stopped()) { 3975 set_result(newcopy); 3976 } 3977 return true; 3978 } 3979 3980 3981 //----------------------generate_virtual_guard--------------------------- 3982 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3983 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3984 RegionNode* slow_region) { 3985 ciMethod* method = callee(); 3986 int vtable_index = method->vtable_index(); 3987 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3988 err_msg_res("bad index %d", vtable_index)); 3989 // Get the Method* out of the appropriate vtable entry. 3990 int entry_offset = (InstanceKlass::vtable_start_offset() + 3991 vtable_index*vtableEntry::size()) * wordSize + 3992 vtableEntry::method_offset_in_bytes(); 3993 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3994 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3995 3996 // Compare the target method with the expected method (e.g., Object.hashCode). 3997 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3998 3999 Node* native_call = makecon(native_call_addr); 4000 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4001 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4002 4003 return generate_slow_guard(test_native, slow_region); 4004 } 4005 4006 //-----------------------generate_method_call---------------------------- 4007 // Use generate_method_call to make a slow-call to the real 4008 // method if the fast path fails. An alternative would be to 4009 // use a stub like OptoRuntime::slow_arraycopy_Java. 4010 // This only works for expanding the current library call, 4011 // not another intrinsic. (E.g., don't use this for making an 4012 // arraycopy call inside of the copyOf intrinsic.) 4013 CallJavaNode* 4014 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4015 // When compiling the intrinsic method itself, do not use this technique. 4016 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4017 4018 ciMethod* method = callee(); 4019 // ensure the JVMS we have will be correct for this call 4020 guarantee(method_id == method->intrinsic_id(), "must match"); 4021 4022 const TypeFunc* tf = TypeFunc::make(method); 4023 CallJavaNode* slow_call; 4024 if (is_static) { 4025 assert(!is_virtual, ""); 4026 slow_call = new CallStaticJavaNode(C, tf, 4027 SharedRuntime::get_resolve_static_call_stub(), 4028 method, bci()); 4029 } else if (is_virtual) { 4030 null_check_receiver(); 4031 int vtable_index = Method::invalid_vtable_index; 4032 if (UseInlineCaches) { 4033 // Suppress the vtable call 4034 } else { 4035 // hashCode and clone are not a miranda methods, 4036 // so the vtable index is fixed. 4037 // No need to use the linkResolver to get it. 4038 vtable_index = method->vtable_index(); 4039 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4040 err_msg_res("bad index %d", vtable_index)); 4041 } 4042 slow_call = new CallDynamicJavaNode(tf, 4043 SharedRuntime::get_resolve_virtual_call_stub(), 4044 method, vtable_index, bci()); 4045 } else { // neither virtual nor static: opt_virtual 4046 null_check_receiver(); 4047 slow_call = new CallStaticJavaNode(C, tf, 4048 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4049 method, bci()); 4050 slow_call->set_optimized_virtual(true); 4051 } 4052 set_arguments_for_java_call(slow_call); 4053 set_edges_for_java_call(slow_call); 4054 return slow_call; 4055 } 4056 4057 4058 /** 4059 * Build special case code for calls to hashCode on an object. This call may 4060 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4061 * slightly different code. 4062 */ 4063 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4064 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4065 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4066 4067 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4068 4069 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4070 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4071 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4072 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4073 Node* obj = NULL; 4074 if (!is_static) { 4075 // Check for hashing null object 4076 obj = null_check_receiver(); 4077 if (stopped()) return true; // unconditionally null 4078 result_reg->init_req(_null_path, top()); 4079 result_val->init_req(_null_path, top()); 4080 } else { 4081 // Do a null check, and return zero if null. 4082 // System.identityHashCode(null) == 0 4083 obj = argument(0); 4084 Node* null_ctl = top(); 4085 obj = null_check_oop(obj, &null_ctl); 4086 result_reg->init_req(_null_path, null_ctl); 4087 result_val->init_req(_null_path, _gvn.intcon(0)); 4088 } 4089 4090 // Unconditionally null? Then return right away. 4091 if (stopped()) { 4092 set_control( result_reg->in(_null_path)); 4093 if (!stopped()) 4094 set_result(result_val->in(_null_path)); 4095 return true; 4096 } 4097 4098 // We only go to the fast case code if we pass a number of guards. The 4099 // paths which do not pass are accumulated in the slow_region. 4100 RegionNode* slow_region = new RegionNode(1); 4101 record_for_igvn(slow_region); 4102 4103 // If this is a virtual call, we generate a funny guard. We pull out 4104 // the vtable entry corresponding to hashCode() from the target object. 4105 // If the target method which we are calling happens to be the native 4106 // Object hashCode() method, we pass the guard. We do not need this 4107 // guard for non-virtual calls -- the caller is known to be the native 4108 // Object hashCode(). 4109 if (is_virtual) { 4110 // After null check, get the object's klass. 4111 Node* obj_klass = load_object_klass(obj); 4112 generate_virtual_guard(obj_klass, slow_region); 4113 } 4114 4115 // Get the header out of the object, use LoadMarkNode when available 4116 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4117 // The control of the load must be NULL. Otherwise, the load can move before 4118 // the null check after castPP removal. 4119 Node* no_ctrl = NULL; 4120 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4121 4122 // Test the header to see if it is unlocked. 4123 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4124 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4125 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4126 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4127 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4128 4129 generate_slow_guard(test_unlocked, slow_region); 4130 4131 // Get the hash value and check to see that it has been properly assigned. 4132 // We depend on hash_mask being at most 32 bits and avoid the use of 4133 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4134 // vm: see markOop.hpp. 4135 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4136 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4137 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4138 // This hack lets the hash bits live anywhere in the mark object now, as long 4139 // as the shift drops the relevant bits into the low 32 bits. Note that 4140 // Java spec says that HashCode is an int so there's no point in capturing 4141 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4142 hshifted_header = ConvX2I(hshifted_header); 4143 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4144 4145 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4146 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4147 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4148 4149 generate_slow_guard(test_assigned, slow_region); 4150 4151 Node* init_mem = reset_memory(); 4152 // fill in the rest of the null path: 4153 result_io ->init_req(_null_path, i_o()); 4154 result_mem->init_req(_null_path, init_mem); 4155 4156 result_val->init_req(_fast_path, hash_val); 4157 result_reg->init_req(_fast_path, control()); 4158 result_io ->init_req(_fast_path, i_o()); 4159 result_mem->init_req(_fast_path, init_mem); 4160 4161 // Generate code for the slow case. We make a call to hashCode(). 4162 set_control(_gvn.transform(slow_region)); 4163 if (!stopped()) { 4164 // No need for PreserveJVMState, because we're using up the present state. 4165 set_all_memory(init_mem); 4166 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4167 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4168 Node* slow_result = set_results_for_java_call(slow_call); 4169 // this->control() comes from set_results_for_java_call 4170 result_reg->init_req(_slow_path, control()); 4171 result_val->init_req(_slow_path, slow_result); 4172 result_io ->set_req(_slow_path, i_o()); 4173 result_mem ->set_req(_slow_path, reset_memory()); 4174 } 4175 4176 // Return the combined state. 4177 set_i_o( _gvn.transform(result_io) ); 4178 set_all_memory( _gvn.transform(result_mem)); 4179 4180 set_result(result_reg, result_val); 4181 return true; 4182 } 4183 4184 //---------------------------inline_native_getClass---------------------------- 4185 // public final native Class<?> java.lang.Object.getClass(); 4186 // 4187 // Build special case code for calls to getClass on an object. 4188 bool LibraryCallKit::inline_native_getClass() { 4189 Node* obj = null_check_receiver(); 4190 if (stopped()) return true; 4191 set_result(load_mirror_from_klass(load_object_klass(obj))); 4192 return true; 4193 } 4194 4195 //-----------------inline_native_Reflection_getCallerClass--------------------- 4196 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4197 // 4198 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4199 // 4200 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4201 // in that it must skip particular security frames and checks for 4202 // caller sensitive methods. 4203 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4204 #ifndef PRODUCT 4205 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4206 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4207 } 4208 #endif 4209 4210 if (!jvms()->has_method()) { 4211 #ifndef PRODUCT 4212 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4213 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4214 } 4215 #endif 4216 return false; 4217 } 4218 4219 // Walk back up the JVM state to find the caller at the required 4220 // depth. 4221 JVMState* caller_jvms = jvms(); 4222 4223 // Cf. JVM_GetCallerClass 4224 // NOTE: Start the loop at depth 1 because the current JVM state does 4225 // not include the Reflection.getCallerClass() frame. 4226 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4227 ciMethod* m = caller_jvms->method(); 4228 switch (n) { 4229 case 0: 4230 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4231 break; 4232 case 1: 4233 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4234 if (!m->caller_sensitive()) { 4235 #ifndef PRODUCT 4236 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4237 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4238 } 4239 #endif 4240 return false; // bail-out; let JVM_GetCallerClass do the work 4241 } 4242 break; 4243 default: 4244 if (!m->is_ignored_by_security_stack_walk()) { 4245 // We have reached the desired frame; return the holder class. 4246 // Acquire method holder as java.lang.Class and push as constant. 4247 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4248 ciInstance* caller_mirror = caller_klass->java_mirror(); 4249 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4250 4251 #ifndef PRODUCT 4252 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4253 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()); 4254 tty->print_cr(" JVM state at this point:"); 4255 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4256 ciMethod* m = jvms()->of_depth(i)->method(); 4257 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4258 } 4259 } 4260 #endif 4261 return true; 4262 } 4263 break; 4264 } 4265 } 4266 4267 #ifndef PRODUCT 4268 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4269 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4270 tty->print_cr(" JVM state at this point:"); 4271 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4272 ciMethod* m = jvms()->of_depth(i)->method(); 4273 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4274 } 4275 } 4276 #endif 4277 4278 return false; // bail-out; let JVM_GetCallerClass do the work 4279 } 4280 4281 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4282 Node* arg = argument(0); 4283 Node* result; 4284 4285 switch (id) { 4286 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4287 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4288 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4289 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4290 4291 case vmIntrinsics::_doubleToLongBits: { 4292 // two paths (plus control) merge in a wood 4293 RegionNode *r = new RegionNode(3); 4294 Node *phi = new PhiNode(r, TypeLong::LONG); 4295 4296 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4297 // Build the boolean node 4298 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4299 4300 // Branch either way. 4301 // NaN case is less traveled, which makes all the difference. 4302 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4303 Node *opt_isnan = _gvn.transform(ifisnan); 4304 assert( opt_isnan->is_If(), "Expect an IfNode"); 4305 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4306 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4307 4308 set_control(iftrue); 4309 4310 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4311 Node *slow_result = longcon(nan_bits); // return NaN 4312 phi->init_req(1, _gvn.transform( slow_result )); 4313 r->init_req(1, iftrue); 4314 4315 // Else fall through 4316 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4317 set_control(iffalse); 4318 4319 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4320 r->init_req(2, iffalse); 4321 4322 // Post merge 4323 set_control(_gvn.transform(r)); 4324 record_for_igvn(r); 4325 4326 C->set_has_split_ifs(true); // Has chance for split-if optimization 4327 result = phi; 4328 assert(result->bottom_type()->isa_long(), "must be"); 4329 break; 4330 } 4331 4332 case vmIntrinsics::_floatToIntBits: { 4333 // two paths (plus control) merge in a wood 4334 RegionNode *r = new RegionNode(3); 4335 Node *phi = new PhiNode(r, TypeInt::INT); 4336 4337 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4338 // Build the boolean node 4339 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4340 4341 // Branch either way. 4342 // NaN case is less traveled, which makes all the difference. 4343 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4344 Node *opt_isnan = _gvn.transform(ifisnan); 4345 assert( opt_isnan->is_If(), "Expect an IfNode"); 4346 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4347 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4348 4349 set_control(iftrue); 4350 4351 static const jint nan_bits = 0x7fc00000; 4352 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4353 phi->init_req(1, _gvn.transform( slow_result )); 4354 r->init_req(1, iftrue); 4355 4356 // Else fall through 4357 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4358 set_control(iffalse); 4359 4360 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4361 r->init_req(2, iffalse); 4362 4363 // Post merge 4364 set_control(_gvn.transform(r)); 4365 record_for_igvn(r); 4366 4367 C->set_has_split_ifs(true); // Has chance for split-if optimization 4368 result = phi; 4369 assert(result->bottom_type()->isa_int(), "must be"); 4370 break; 4371 } 4372 4373 default: 4374 fatal_unexpected_iid(id); 4375 break; 4376 } 4377 set_result(_gvn.transform(result)); 4378 return true; 4379 } 4380 4381 #ifdef _LP64 4382 #define XTOP ,top() /*additional argument*/ 4383 #else //_LP64 4384 #define XTOP /*no additional argument*/ 4385 #endif //_LP64 4386 4387 //----------------------inline_unsafe_copyMemory------------------------- 4388 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4389 bool LibraryCallKit::inline_unsafe_copyMemory() { 4390 if (callee()->is_static()) return false; // caller must have the capability! 4391 null_check_receiver(); // null-check receiver 4392 if (stopped()) return true; 4393 4394 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4395 4396 Node* src_ptr = argument(1); // type: oop 4397 Node* src_off = ConvL2X(argument(2)); // type: long 4398 Node* dst_ptr = argument(4); // type: oop 4399 Node* dst_off = ConvL2X(argument(5)); // type: long 4400 Node* size = ConvL2X(argument(7)); // type: long 4401 4402 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4403 "fieldOffset must be byte-scaled"); 4404 4405 Node* src = make_unsafe_address(src_ptr, src_off); 4406 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4407 4408 // Conservatively insert a memory barrier on all memory slices. 4409 // Do not let writes of the copy source or destination float below the copy. 4410 insert_mem_bar(Op_MemBarCPUOrder); 4411 4412 // Call it. Note that the length argument is not scaled. 4413 make_runtime_call(RC_LEAF|RC_NO_FP, 4414 OptoRuntime::fast_arraycopy_Type(), 4415 StubRoutines::unsafe_arraycopy(), 4416 "unsafe_arraycopy", 4417 TypeRawPtr::BOTTOM, 4418 src, dst, size XTOP); 4419 4420 // Do not let reads of the copy destination float above the copy. 4421 insert_mem_bar(Op_MemBarCPUOrder); 4422 4423 return true; 4424 } 4425 4426 //------------------------clone_coping----------------------------------- 4427 // Helper function for inline_native_clone. 4428 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4429 assert(obj_size != NULL, ""); 4430 Node* raw_obj = alloc_obj->in(1); 4431 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4432 4433 AllocateNode* alloc = NULL; 4434 if (ReduceBulkZeroing) { 4435 // We will be completely responsible for initializing this object - 4436 // mark Initialize node as complete. 4437 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4438 // The object was just allocated - there should be no any stores! 4439 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4440 // Mark as complete_with_arraycopy so that on AllocateNode 4441 // expansion, we know this AllocateNode is initialized by an array 4442 // copy and a StoreStore barrier exists after the array copy. 4443 alloc->initialization()->set_complete_with_arraycopy(); 4444 } 4445 4446 // Copy the fastest available way. 4447 // TODO: generate fields copies for small objects instead. 4448 Node* src = obj; 4449 Node* dest = alloc_obj; 4450 Node* size = _gvn.transform(obj_size); 4451 4452 // Exclude the header but include array length to copy by 8 bytes words. 4453 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4454 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4455 instanceOopDesc::base_offset_in_bytes(); 4456 // base_off: 4457 // 8 - 32-bit VM 4458 // 12 - 64-bit VM, compressed klass 4459 // 16 - 64-bit VM, normal klass 4460 if (base_off % BytesPerLong != 0) { 4461 assert(UseCompressedClassPointers, ""); 4462 if (is_array) { 4463 // Exclude length to copy by 8 bytes words. 4464 base_off += sizeof(int); 4465 } else { 4466 // Include klass to copy by 8 bytes words. 4467 base_off = instanceOopDesc::klass_offset_in_bytes(); 4468 } 4469 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4470 } 4471 src = basic_plus_adr(src, base_off); 4472 dest = basic_plus_adr(dest, base_off); 4473 4474 // Compute the length also, if needed: 4475 Node* countx = size; 4476 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4477 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4478 4479 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4480 4481 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4482 ac->set_clonebasic(); 4483 Node* n = _gvn.transform(ac); 4484 assert(n == ac, "cannot disappear"); 4485 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4486 4487 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4488 if (card_mark) { 4489 assert(!is_array, ""); 4490 // Put in store barrier for any and all oops we are sticking 4491 // into this object. (We could avoid this if we could prove 4492 // that the object type contains no oop fields at all.) 4493 Node* no_particular_value = NULL; 4494 Node* no_particular_field = NULL; 4495 int raw_adr_idx = Compile::AliasIdxRaw; 4496 post_barrier(control(), 4497 memory(raw_adr_type), 4498 alloc_obj, 4499 no_particular_field, 4500 raw_adr_idx, 4501 no_particular_value, 4502 T_OBJECT, 4503 false); 4504 } 4505 4506 // Do not let reads from the cloned object float above the arraycopy. 4507 if (alloc != NULL) { 4508 // Do not let stores that initialize this object be reordered with 4509 // a subsequent store that would make this object accessible by 4510 // other threads. 4511 // Record what AllocateNode this StoreStore protects so that 4512 // escape analysis can go from the MemBarStoreStoreNode to the 4513 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4514 // based on the escape status of the AllocateNode. 4515 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4516 } else { 4517 insert_mem_bar(Op_MemBarCPUOrder); 4518 } 4519 } 4520 4521 //------------------------inline_native_clone---------------------------- 4522 // protected native Object java.lang.Object.clone(); 4523 // 4524 // Here are the simple edge cases: 4525 // null receiver => normal trap 4526 // virtual and clone was overridden => slow path to out-of-line clone 4527 // not cloneable or finalizer => slow path to out-of-line Object.clone 4528 // 4529 // The general case has two steps, allocation and copying. 4530 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4531 // 4532 // Copying also has two cases, oop arrays and everything else. 4533 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4534 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4535 // 4536 // These steps fold up nicely if and when the cloned object's klass 4537 // can be sharply typed as an object array, a type array, or an instance. 4538 // 4539 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4540 PhiNode* result_val; 4541 4542 // Set the reexecute bit for the interpreter to reexecute 4543 // the bytecode that invokes Object.clone if deoptimization happens. 4544 { PreserveReexecuteState preexecs(this); 4545 jvms()->set_should_reexecute(true); 4546 4547 Node* obj = null_check_receiver(); 4548 if (stopped()) return true; 4549 4550 Node* obj_klass = load_object_klass(obj); 4551 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4552 const TypeOopPtr* toop = ((tklass != NULL) 4553 ? tklass->as_instance_type() 4554 : TypeInstPtr::NOTNULL); 4555 4556 // Conservatively insert a memory barrier on all memory slices. 4557 // Do not let writes into the original float below the clone. 4558 insert_mem_bar(Op_MemBarCPUOrder); 4559 4560 // paths into result_reg: 4561 enum { 4562 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4563 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4564 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4565 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4566 PATH_LIMIT 4567 }; 4568 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4569 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4570 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4571 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4572 record_for_igvn(result_reg); 4573 4574 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4575 int raw_adr_idx = Compile::AliasIdxRaw; 4576 4577 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4578 if (array_ctl != NULL) { 4579 // It's an array. 4580 PreserveJVMState pjvms(this); 4581 set_control(array_ctl); 4582 Node* obj_length = load_array_length(obj); 4583 Node* obj_size = NULL; 4584 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4585 4586 if (!use_ReduceInitialCardMarks()) { 4587 // If it is an oop array, it requires very special treatment, 4588 // because card marking is required on each card of the array. 4589 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4590 if (is_obja != NULL) { 4591 PreserveJVMState pjvms2(this); 4592 set_control(is_obja); 4593 // Generate a direct call to the right arraycopy function(s). 4594 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4595 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4596 ac->set_cloneoop(); 4597 Node* n = _gvn.transform(ac); 4598 assert(n == ac, "cannot disappear"); 4599 ac->connect_outputs(this); 4600 4601 result_reg->init_req(_objArray_path, control()); 4602 result_val->init_req(_objArray_path, alloc_obj); 4603 result_i_o ->set_req(_objArray_path, i_o()); 4604 result_mem ->set_req(_objArray_path, reset_memory()); 4605 } 4606 } 4607 // Otherwise, there are no card marks to worry about. 4608 // (We can dispense with card marks if we know the allocation 4609 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4610 // causes the non-eden paths to take compensating steps to 4611 // simulate a fresh allocation, so that no further 4612 // card marks are required in compiled code to initialize 4613 // the object.) 4614 4615 if (!stopped()) { 4616 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4617 4618 // Present the results of the copy. 4619 result_reg->init_req(_array_path, control()); 4620 result_val->init_req(_array_path, alloc_obj); 4621 result_i_o ->set_req(_array_path, i_o()); 4622 result_mem ->set_req(_array_path, reset_memory()); 4623 } 4624 } 4625 4626 // We only go to the instance fast case code if we pass a number of guards. 4627 // The paths which do not pass are accumulated in the slow_region. 4628 RegionNode* slow_region = new RegionNode(1); 4629 record_for_igvn(slow_region); 4630 if (!stopped()) { 4631 // It's an instance (we did array above). Make the slow-path tests. 4632 // If this is a virtual call, we generate a funny guard. We grab 4633 // the vtable entry corresponding to clone() from the target object. 4634 // If the target method which we are calling happens to be the 4635 // Object clone() method, we pass the guard. We do not need this 4636 // guard for non-virtual calls; the caller is known to be the native 4637 // Object clone(). 4638 if (is_virtual) { 4639 generate_virtual_guard(obj_klass, slow_region); 4640 } 4641 4642 // The object must be cloneable and must not have a finalizer. 4643 // Both of these conditions may be checked in a single test. 4644 // We could optimize the cloneable test further, but we don't care. 4645 generate_access_flags_guard(obj_klass, 4646 // Test both conditions: 4647 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4648 // Must be cloneable but not finalizer: 4649 JVM_ACC_IS_CLONEABLE, 4650 slow_region); 4651 } 4652 4653 if (!stopped()) { 4654 // It's an instance, and it passed the slow-path tests. 4655 PreserveJVMState pjvms(this); 4656 Node* obj_size = NULL; 4657 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4658 // is reexecuted if deoptimization occurs and there could be problems when merging 4659 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4660 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4661 4662 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4663 4664 // Present the results of the slow call. 4665 result_reg->init_req(_instance_path, control()); 4666 result_val->init_req(_instance_path, alloc_obj); 4667 result_i_o ->set_req(_instance_path, i_o()); 4668 result_mem ->set_req(_instance_path, reset_memory()); 4669 } 4670 4671 // Generate code for the slow case. We make a call to clone(). 4672 set_control(_gvn.transform(slow_region)); 4673 if (!stopped()) { 4674 PreserveJVMState pjvms(this); 4675 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4676 Node* slow_result = set_results_for_java_call(slow_call); 4677 // this->control() comes from set_results_for_java_call 4678 result_reg->init_req(_slow_path, control()); 4679 result_val->init_req(_slow_path, slow_result); 4680 result_i_o ->set_req(_slow_path, i_o()); 4681 result_mem ->set_req(_slow_path, reset_memory()); 4682 } 4683 4684 // Return the combined state. 4685 set_control( _gvn.transform(result_reg)); 4686 set_i_o( _gvn.transform(result_i_o)); 4687 set_all_memory( _gvn.transform(result_mem)); 4688 } // original reexecute is set back here 4689 4690 set_result(_gvn.transform(result_val)); 4691 return true; 4692 } 4693 4694 //------------------------------inline_arraycopy----------------------- 4695 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4696 // Object dest, int destPos, 4697 // int length); 4698 bool LibraryCallKit::inline_arraycopy() { 4699 // Get the arguments. 4700 Node* src = argument(0); // type: oop 4701 Node* src_offset = argument(1); // type: int 4702 Node* dest = argument(2); // type: oop 4703 Node* dest_offset = argument(3); // type: int 4704 Node* length = argument(4); // type: int 4705 4706 // Check for allocation before we add nodes that would confuse 4707 // tightly_coupled_allocation() 4708 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4709 4710 // The following tests must be performed 4711 // (1) src and dest are arrays. 4712 // (2) src and dest arrays must have elements of the same BasicType 4713 // (3) src and dest must not be null. 4714 // (4) src_offset must not be negative. 4715 // (5) dest_offset must not be negative. 4716 // (6) length must not be negative. 4717 // (7) src_offset + length must not exceed length of src. 4718 // (8) dest_offset + length must not exceed length of dest. 4719 // (9) each element of an oop array must be assignable 4720 4721 // (3) src and dest must not be null. 4722 // always do this here because we need the JVM state for uncommon traps 4723 src = null_check(src, T_ARRAY); 4724 dest = null_check(dest, T_ARRAY); 4725 4726 bool notest = false; 4727 4728 const Type* src_type = _gvn.type(src); 4729 const Type* dest_type = _gvn.type(dest); 4730 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4731 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4732 4733 // Do we have the type of src? 4734 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4735 // Do we have the type of dest? 4736 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4737 // Is the type for src from speculation? 4738 bool src_spec = false; 4739 // Is the type for dest from speculation? 4740 bool dest_spec = false; 4741 4742 if (!has_src || !has_dest) { 4743 // We don't have sufficient type information, let's see if 4744 // speculative types can help. We need to have types for both src 4745 // and dest so that it pays off. 4746 4747 // Do we already have or could we have type information for src 4748 bool could_have_src = has_src; 4749 // Do we already have or could we have type information for dest 4750 bool could_have_dest = has_dest; 4751 4752 ciKlass* src_k = NULL; 4753 if (!has_src) { 4754 src_k = src_type->speculative_type_not_null(); 4755 if (src_k != NULL && src_k->is_array_klass()) { 4756 could_have_src = true; 4757 } 4758 } 4759 4760 ciKlass* dest_k = NULL; 4761 if (!has_dest) { 4762 dest_k = dest_type->speculative_type_not_null(); 4763 if (dest_k != NULL && dest_k->is_array_klass()) { 4764 could_have_dest = true; 4765 } 4766 } 4767 4768 if (could_have_src && could_have_dest) { 4769 // This is going to pay off so emit the required guards 4770 if (!has_src) { 4771 src = maybe_cast_profiled_obj(src, src_k); 4772 src_type = _gvn.type(src); 4773 top_src = src_type->isa_aryptr(); 4774 has_src = (top_src != NULL && top_src->klass() != NULL); 4775 src_spec = true; 4776 } 4777 if (!has_dest) { 4778 dest = maybe_cast_profiled_obj(dest, dest_k); 4779 dest_type = _gvn.type(dest); 4780 top_dest = dest_type->isa_aryptr(); 4781 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4782 dest_spec = true; 4783 } 4784 } 4785 } 4786 4787 if (has_src && has_dest) { 4788 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4789 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4790 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4791 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4792 4793 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4794 // If both arrays are object arrays then having the exact types 4795 // for both will remove the need for a subtype check at runtime 4796 // before the call and may make it possible to pick a faster copy 4797 // routine (without a subtype check on every element) 4798 // Do we have the exact type of src? 4799 bool could_have_src = src_spec; 4800 // Do we have the exact type of dest? 4801 bool could_have_dest = dest_spec; 4802 ciKlass* src_k = top_src->klass(); 4803 ciKlass* dest_k = top_dest->klass(); 4804 if (!src_spec) { 4805 src_k = src_type->speculative_type_not_null(); 4806 if (src_k != NULL && src_k->is_array_klass()) { 4807 could_have_src = true; 4808 } 4809 } 4810 if (!dest_spec) { 4811 dest_k = dest_type->speculative_type_not_null(); 4812 if (dest_k != NULL && dest_k->is_array_klass()) { 4813 could_have_dest = true; 4814 } 4815 } 4816 if (could_have_src && could_have_dest) { 4817 // If we can have both exact types, emit the missing guards 4818 if (could_have_src && !src_spec) { 4819 src = maybe_cast_profiled_obj(src, src_k); 4820 } 4821 if (could_have_dest && !dest_spec) { 4822 dest = maybe_cast_profiled_obj(dest, dest_k); 4823 } 4824 } 4825 } 4826 } 4827 4828 if (!too_many_traps(Deoptimization::Reason_intrinsic) && !src->is_top() && !dest->is_top()) { 4829 // validate arguments: enables transformation the ArrayCopyNode 4830 notest = true; 4831 4832 RegionNode* slow_region = new RegionNode(1); 4833 record_for_igvn(slow_region); 4834 4835 // (1) src and dest are arrays. 4836 generate_non_array_guard(load_object_klass(src), slow_region); 4837 generate_non_array_guard(load_object_klass(dest), slow_region); 4838 4839 // (2) src and dest arrays must have elements of the same BasicType 4840 // done at macro expansion or at Ideal transformation time 4841 4842 // (4) src_offset must not be negative. 4843 generate_negative_guard(src_offset, slow_region); 4844 4845 // (5) dest_offset must not be negative. 4846 generate_negative_guard(dest_offset, slow_region); 4847 4848 // (7) src_offset + length must not exceed length of src. 4849 generate_limit_guard(src_offset, length, 4850 load_array_length(src), 4851 slow_region); 4852 4853 // (8) dest_offset + length must not exceed length of dest. 4854 generate_limit_guard(dest_offset, length, 4855 load_array_length(dest), 4856 slow_region); 4857 4858 // (9) each element of an oop array must be assignable 4859 Node* src_klass = load_object_klass(src); 4860 Node* dest_klass = load_object_klass(dest); 4861 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4862 4863 if (not_subtype_ctrl != top()) { 4864 PreserveJVMState pjvms(this); 4865 set_control(not_subtype_ctrl); 4866 uncommon_trap(Deoptimization::Reason_intrinsic, 4867 Deoptimization::Action_make_not_entrant); 4868 assert(stopped(), "Should be stopped"); 4869 } 4870 { 4871 PreserveJVMState pjvms(this); 4872 set_control(_gvn.transform(slow_region)); 4873 uncommon_trap(Deoptimization::Reason_intrinsic, 4874 Deoptimization::Action_make_not_entrant); 4875 assert(stopped(), "Should be stopped"); 4876 } 4877 } 4878 4879 if (stopped()) { 4880 return true; 4881 } 4882 4883 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 4884 // Create LoadRange and LoadKlass nodes for use during macro expansion here 4885 // so the compiler has a chance to eliminate them: during macro expansion, 4886 // we have to set their control (CastPP nodes are eliminated). 4887 load_object_klass(src), load_object_klass(dest), 4888 load_array_length(src), load_array_length(dest)); 4889 4890 if (notest) { 4891 ac->set_arraycopy_notest(); 4892 } 4893 4894 Node* n = _gvn.transform(ac); 4895 assert(n == ac, "cannot disappear"); 4896 ac->connect_outputs(this); 4897 4898 return true; 4899 } 4900 4901 4902 // Helper function which determines if an arraycopy immediately follows 4903 // an allocation, with no intervening tests or other escapes for the object. 4904 AllocateArrayNode* 4905 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4906 RegionNode* slow_region) { 4907 if (stopped()) return NULL; // no fast path 4908 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4909 4910 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4911 if (alloc == NULL) return NULL; 4912 4913 Node* rawmem = memory(Compile::AliasIdxRaw); 4914 // Is the allocation's memory state untouched? 4915 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4916 // Bail out if there have been raw-memory effects since the allocation. 4917 // (Example: There might have been a call or safepoint.) 4918 return NULL; 4919 } 4920 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4921 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4922 return NULL; 4923 } 4924 4925 // There must be no unexpected observers of this allocation. 4926 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4927 Node* obs = ptr->fast_out(i); 4928 if (obs != this->map()) { 4929 return NULL; 4930 } 4931 } 4932 4933 // This arraycopy must unconditionally follow the allocation of the ptr. 4934 Node* alloc_ctl = ptr->in(0); 4935 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4936 4937 Node* ctl = control(); 4938 while (ctl != alloc_ctl) { 4939 // There may be guards which feed into the slow_region. 4940 // Any other control flow means that we might not get a chance 4941 // to finish initializing the allocated object. 4942 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4943 IfNode* iff = ctl->in(0)->as_If(); 4944 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 4945 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4946 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4947 ctl = iff->in(0); // This test feeds the known slow_region. 4948 continue; 4949 } 4950 // One more try: Various low-level checks bottom out in 4951 // uncommon traps. If the debug-info of the trap omits 4952 // any reference to the allocation, as we've already 4953 // observed, then there can be no objection to the trap. 4954 bool found_trap = false; 4955 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 4956 Node* obs = not_ctl->fast_out(j); 4957 if (obs->in(0) == not_ctl && obs->is_Call() && 4958 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 4959 found_trap = true; break; 4960 } 4961 } 4962 if (found_trap) { 4963 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 4964 continue; 4965 } 4966 } 4967 return NULL; 4968 } 4969 4970 // If we get this far, we have an allocation which immediately 4971 // precedes the arraycopy, and we can take over zeroing the new object. 4972 // The arraycopy will finish the initialization, and provide 4973 // a new control state to which we will anchor the destination pointer. 4974 4975 return alloc; 4976 } 4977 4978 //-------------inline_encodeISOArray----------------------------------- 4979 // encode char[] to byte[] in ISO_8859_1 4980 bool LibraryCallKit::inline_encodeISOArray() { 4981 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 4982 // no receiver since it is static method 4983 Node *src = argument(0); 4984 Node *src_offset = argument(1); 4985 Node *dst = argument(2); 4986 Node *dst_offset = argument(3); 4987 Node *length = argument(4); 4988 4989 const Type* src_type = src->Value(&_gvn); 4990 const Type* dst_type = dst->Value(&_gvn); 4991 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4992 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 4993 if (top_src == NULL || top_src->klass() == NULL || 4994 top_dest == NULL || top_dest->klass() == NULL) { 4995 // failed array check 4996 return false; 4997 } 4998 4999 // Figure out the size and type of the elements we will be copying. 5000 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5001 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5002 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5003 return false; 5004 } 5005 Node* src_start = array_element_address(src, src_offset, src_elem); 5006 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5007 // 'src_start' points to src array + scaled offset 5008 // 'dst_start' points to dst array + scaled offset 5009 5010 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5011 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5012 enc = _gvn.transform(enc); 5013 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5014 set_memory(res_mem, mtype); 5015 set_result(enc); 5016 return true; 5017 } 5018 5019 //-------------inline_multiplyToLen----------------------------------- 5020 bool LibraryCallKit::inline_multiplyToLen() { 5021 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5022 5023 address stubAddr = StubRoutines::multiplyToLen(); 5024 if (stubAddr == NULL) { 5025 return false; // Intrinsic's stub is not implemented on this platform 5026 } 5027 const char* stubName = "multiplyToLen"; 5028 5029 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5030 5031 Node* x = argument(1); 5032 Node* xlen = argument(2); 5033 Node* y = argument(3); 5034 Node* ylen = argument(4); 5035 Node* z = argument(5); 5036 5037 const Type* x_type = x->Value(&_gvn); 5038 const Type* y_type = y->Value(&_gvn); 5039 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5040 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5041 if (top_x == NULL || top_x->klass() == NULL || 5042 top_y == NULL || top_y->klass() == NULL) { 5043 // failed array check 5044 return false; 5045 } 5046 5047 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5048 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5049 if (x_elem != T_INT || y_elem != T_INT) { 5050 return false; 5051 } 5052 5053 // Set the original stack and the reexecute bit for the interpreter to reexecute 5054 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5055 // on the return from z array allocation in runtime. 5056 { PreserveReexecuteState preexecs(this); 5057 jvms()->set_should_reexecute(true); 5058 5059 Node* x_start = array_element_address(x, intcon(0), x_elem); 5060 Node* y_start = array_element_address(y, intcon(0), y_elem); 5061 // 'x_start' points to x array + scaled xlen 5062 // 'y_start' points to y array + scaled ylen 5063 5064 // Allocate the result array 5065 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5066 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5067 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5068 5069 IdealKit ideal(this); 5070 5071 #define __ ideal. 5072 Node* one = __ ConI(1); 5073 Node* zero = __ ConI(0); 5074 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5075 __ set(need_alloc, zero); 5076 __ set(z_alloc, z); 5077 __ if_then(z, BoolTest::eq, null()); { 5078 __ increment (need_alloc, one); 5079 } __ else_(); { 5080 // Update graphKit memory and control from IdealKit. 5081 sync_kit(ideal); 5082 Node* zlen_arg = load_array_length(z); 5083 // Update IdealKit memory and control from graphKit. 5084 __ sync_kit(this); 5085 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5086 __ increment (need_alloc, one); 5087 } __ end_if(); 5088 } __ end_if(); 5089 5090 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5091 // Update graphKit memory and control from IdealKit. 5092 sync_kit(ideal); 5093 Node * narr = new_array(klass_node, zlen, 1); 5094 // Update IdealKit memory and control from graphKit. 5095 __ sync_kit(this); 5096 __ set(z_alloc, narr); 5097 } __ end_if(); 5098 5099 sync_kit(ideal); 5100 z = __ value(z_alloc); 5101 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5102 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5103 // Final sync IdealKit and GraphKit. 5104 final_sync(ideal); 5105 #undef __ 5106 5107 Node* z_start = array_element_address(z, intcon(0), T_INT); 5108 5109 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5110 OptoRuntime::multiplyToLen_Type(), 5111 stubAddr, stubName, TypePtr::BOTTOM, 5112 x_start, xlen, y_start, ylen, z_start, zlen); 5113 } // original reexecute is set back here 5114 5115 C->set_has_split_ifs(true); // Has chance for split-if optimization 5116 set_result(z); 5117 return true; 5118 } 5119 5120 5121 /** 5122 * Calculate CRC32 for byte. 5123 * int java.util.zip.CRC32.update(int crc, int b) 5124 */ 5125 bool LibraryCallKit::inline_updateCRC32() { 5126 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5127 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5128 // no receiver since it is static method 5129 Node* crc = argument(0); // type: int 5130 Node* b = argument(1); // type: int 5131 5132 /* 5133 * int c = ~ crc; 5134 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5135 * b = b ^ (c >>> 8); 5136 * crc = ~b; 5137 */ 5138 5139 Node* M1 = intcon(-1); 5140 crc = _gvn.transform(new XorINode(crc, M1)); 5141 Node* result = _gvn.transform(new XorINode(crc, b)); 5142 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5143 5144 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5145 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5146 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5147 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5148 5149 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5150 result = _gvn.transform(new XorINode(crc, result)); 5151 result = _gvn.transform(new XorINode(result, M1)); 5152 set_result(result); 5153 return true; 5154 } 5155 5156 /** 5157 * Calculate CRC32 for byte[] array. 5158 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5159 */ 5160 bool LibraryCallKit::inline_updateBytesCRC32() { 5161 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5162 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5163 // no receiver since it is static method 5164 Node* crc = argument(0); // type: int 5165 Node* src = argument(1); // type: oop 5166 Node* offset = argument(2); // type: int 5167 Node* length = argument(3); // type: int 5168 5169 const Type* src_type = src->Value(&_gvn); 5170 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5171 if (top_src == NULL || top_src->klass() == NULL) { 5172 // failed array check 5173 return false; 5174 } 5175 5176 // Figure out the size and type of the elements we will be copying. 5177 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5178 if (src_elem != T_BYTE) { 5179 return false; 5180 } 5181 5182 // 'src_start' points to src array + scaled offset 5183 Node* src_start = array_element_address(src, offset, src_elem); 5184 5185 // We assume that range check is done by caller. 5186 // TODO: generate range check (offset+length < src.length) in debug VM. 5187 5188 // Call the stub. 5189 address stubAddr = StubRoutines::updateBytesCRC32(); 5190 const char *stubName = "updateBytesCRC32"; 5191 5192 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5193 stubAddr, stubName, TypePtr::BOTTOM, 5194 crc, src_start, length); 5195 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5196 set_result(result); 5197 return true; 5198 } 5199 5200 /** 5201 * Calculate CRC32 for ByteBuffer. 5202 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5203 */ 5204 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5205 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5206 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5207 // no receiver since it is static method 5208 Node* crc = argument(0); // type: int 5209 Node* src = argument(1); // type: long 5210 Node* offset = argument(3); // type: int 5211 Node* length = argument(4); // type: int 5212 5213 src = ConvL2X(src); // adjust Java long to machine word 5214 Node* base = _gvn.transform(new CastX2PNode(src)); 5215 offset = ConvI2X(offset); 5216 5217 // 'src_start' points to src array + scaled offset 5218 Node* src_start = basic_plus_adr(top(), base, offset); 5219 5220 // Call the stub. 5221 address stubAddr = StubRoutines::updateBytesCRC32(); 5222 const char *stubName = "updateBytesCRC32"; 5223 5224 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5225 stubAddr, stubName, TypePtr::BOTTOM, 5226 crc, src_start, length); 5227 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5228 set_result(result); 5229 return true; 5230 } 5231 5232 //----------------------------inline_reference_get---------------------------- 5233 // public T java.lang.ref.Reference.get(); 5234 bool LibraryCallKit::inline_reference_get() { 5235 const int referent_offset = java_lang_ref_Reference::referent_offset; 5236 guarantee(referent_offset > 0, "should have already been set"); 5237 5238 // Get the argument: 5239 Node* reference_obj = null_check_receiver(); 5240 if (stopped()) return true; 5241 5242 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5243 5244 ciInstanceKlass* klass = env()->Object_klass(); 5245 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5246 5247 Node* no_ctrl = NULL; 5248 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5249 5250 // Use the pre-barrier to record the value in the referent field 5251 pre_barrier(false /* do_load */, 5252 control(), 5253 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5254 result /* pre_val */, 5255 T_OBJECT); 5256 5257 // Add memory barrier to prevent commoning reads from this field 5258 // across safepoint since GC can change its value. 5259 insert_mem_bar(Op_MemBarCPUOrder); 5260 5261 set_result(result); 5262 return true; 5263 } 5264 5265 5266 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5267 bool is_exact=true, bool is_static=false) { 5268 5269 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5270 assert(tinst != NULL, "obj is null"); 5271 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5272 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5273 5274 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5275 ciSymbol::make(fieldTypeString), 5276 is_static); 5277 if (field == NULL) return (Node *) NULL; 5278 assert (field != NULL, "undefined field"); 5279 5280 // Next code copied from Parse::do_get_xxx(): 5281 5282 // Compute address and memory type. 5283 int offset = field->offset_in_bytes(); 5284 bool is_vol = field->is_volatile(); 5285 ciType* field_klass = field->type(); 5286 assert(field_klass->is_loaded(), "should be loaded"); 5287 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5288 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5289 BasicType bt = field->layout_type(); 5290 5291 // Build the resultant type of the load 5292 const Type *type; 5293 if (bt == T_OBJECT) { 5294 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5295 } else { 5296 type = Type::get_const_basic_type(bt); 5297 } 5298 5299 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5300 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5301 } 5302 // Build the load. 5303 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5304 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol); 5305 // If reference is volatile, prevent following memory ops from 5306 // floating up past the volatile read. Also prevents commoning 5307 // another volatile read. 5308 if (is_vol) { 5309 // Memory barrier includes bogus read of value to force load BEFORE membar 5310 insert_mem_bar(Op_MemBarAcquire, loadedField); 5311 } 5312 return loadedField; 5313 } 5314 5315 5316 //------------------------------inline_aescrypt_Block----------------------- 5317 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5318 address stubAddr; 5319 const char *stubName; 5320 assert(UseAES, "need AES instruction support"); 5321 5322 switch(id) { 5323 case vmIntrinsics::_aescrypt_encryptBlock: 5324 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5325 stubName = "aescrypt_encryptBlock"; 5326 break; 5327 case vmIntrinsics::_aescrypt_decryptBlock: 5328 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5329 stubName = "aescrypt_decryptBlock"; 5330 break; 5331 } 5332 if (stubAddr == NULL) return false; 5333 5334 Node* aescrypt_object = argument(0); 5335 Node* src = argument(1); 5336 Node* src_offset = argument(2); 5337 Node* dest = argument(3); 5338 Node* dest_offset = argument(4); 5339 5340 // (1) src and dest are arrays. 5341 const Type* src_type = src->Value(&_gvn); 5342 const Type* dest_type = dest->Value(&_gvn); 5343 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5344 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5345 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5346 5347 // for the quick and dirty code we will skip all the checks. 5348 // we are just trying to get the call to be generated. 5349 Node* src_start = src; 5350 Node* dest_start = dest; 5351 if (src_offset != NULL || dest_offset != NULL) { 5352 assert(src_offset != NULL && dest_offset != NULL, ""); 5353 src_start = array_element_address(src, src_offset, T_BYTE); 5354 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5355 } 5356 5357 // now need to get the start of its expanded key array 5358 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5359 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5360 if (k_start == NULL) return false; 5361 5362 if (Matcher::pass_original_key_for_aes()) { 5363 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5364 // compatibility issues between Java key expansion and SPARC crypto instructions 5365 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5366 if (original_k_start == NULL) return false; 5367 5368 // Call the stub. 5369 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5370 stubAddr, stubName, TypePtr::BOTTOM, 5371 src_start, dest_start, k_start, original_k_start); 5372 } else { 5373 // Call the stub. 5374 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5375 stubAddr, stubName, TypePtr::BOTTOM, 5376 src_start, dest_start, k_start); 5377 } 5378 5379 return true; 5380 } 5381 5382 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5383 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5384 address stubAddr; 5385 const char *stubName; 5386 5387 assert(UseAES, "need AES instruction support"); 5388 5389 switch(id) { 5390 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5391 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5392 stubName = "cipherBlockChaining_encryptAESCrypt"; 5393 break; 5394 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5395 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5396 stubName = "cipherBlockChaining_decryptAESCrypt"; 5397 break; 5398 } 5399 if (stubAddr == NULL) return false; 5400 5401 Node* cipherBlockChaining_object = argument(0); 5402 Node* src = argument(1); 5403 Node* src_offset = argument(2); 5404 Node* len = argument(3); 5405 Node* dest = argument(4); 5406 Node* dest_offset = argument(5); 5407 5408 // (1) src and dest are arrays. 5409 const Type* src_type = src->Value(&_gvn); 5410 const Type* dest_type = dest->Value(&_gvn); 5411 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5412 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5413 assert (top_src != NULL && top_src->klass() != NULL 5414 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5415 5416 // checks are the responsibility of the caller 5417 Node* src_start = src; 5418 Node* dest_start = dest; 5419 if (src_offset != NULL || dest_offset != NULL) { 5420 assert(src_offset != NULL && dest_offset != NULL, ""); 5421 src_start = array_element_address(src, src_offset, T_BYTE); 5422 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5423 } 5424 5425 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5426 // (because of the predicated logic executed earlier). 5427 // so we cast it here safely. 5428 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5429 5430 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5431 if (embeddedCipherObj == NULL) return false; 5432 5433 // cast it to what we know it will be at runtime 5434 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5435 assert(tinst != NULL, "CBC obj is null"); 5436 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5437 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5438 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5439 5440 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5441 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5442 const TypeOopPtr* xtype = aklass->as_instance_type(); 5443 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5444 aescrypt_object = _gvn.transform(aescrypt_object); 5445 5446 // we need to get the start of the aescrypt_object's expanded key array 5447 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5448 if (k_start == NULL) return false; 5449 5450 // similarly, get the start address of the r vector 5451 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5452 if (objRvec == NULL) return false; 5453 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5454 5455 Node* cbcCrypt; 5456 if (Matcher::pass_original_key_for_aes()) { 5457 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5458 // compatibility issues between Java key expansion and SPARC crypto instructions 5459 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5460 if (original_k_start == NULL) return false; 5461 5462 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5463 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5464 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5465 stubAddr, stubName, TypePtr::BOTTOM, 5466 src_start, dest_start, k_start, r_start, len, original_k_start); 5467 } else { 5468 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5469 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5470 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5471 stubAddr, stubName, TypePtr::BOTTOM, 5472 src_start, dest_start, k_start, r_start, len); 5473 } 5474 5475 // return cipher length (int) 5476 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5477 set_result(retvalue); 5478 return true; 5479 } 5480 5481 //------------------------------get_key_start_from_aescrypt_object----------------------- 5482 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5483 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5484 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5485 if (objAESCryptKey == NULL) return (Node *) NULL; 5486 5487 // now have the array, need to get the start address of the K array 5488 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5489 return k_start; 5490 } 5491 5492 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5493 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5494 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5495 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5496 if (objAESCryptKey == NULL) return (Node *) NULL; 5497 5498 // now have the array, need to get the start address of the lastKey array 5499 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5500 return original_k_start; 5501 } 5502 5503 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5504 // Return node representing slow path of predicate check. 5505 // the pseudo code we want to emulate with this predicate is: 5506 // for encryption: 5507 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5508 // for decryption: 5509 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5510 // note cipher==plain is more conservative than the original java code but that's OK 5511 // 5512 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5513 // The receiver was checked for NULL already. 5514 Node* objCBC = argument(0); 5515 5516 // Load embeddedCipher field of CipherBlockChaining object. 5517 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5518 5519 // get AESCrypt klass for instanceOf check 5520 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5521 // will have same classloader as CipherBlockChaining object 5522 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5523 assert(tinst != NULL, "CBCobj is null"); 5524 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5525 5526 // we want to do an instanceof comparison against the AESCrypt class 5527 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5528 if (!klass_AESCrypt->is_loaded()) { 5529 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5530 Node* ctrl = control(); 5531 set_control(top()); // no regular fast path 5532 return ctrl; 5533 } 5534 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5535 5536 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5537 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5538 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5539 5540 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5541 5542 // for encryption, we are done 5543 if (!decrypting) 5544 return instof_false; // even if it is NULL 5545 5546 // for decryption, we need to add a further check to avoid 5547 // taking the intrinsic path when cipher and plain are the same 5548 // see the original java code for why. 5549 RegionNode* region = new RegionNode(3); 5550 region->init_req(1, instof_false); 5551 Node* src = argument(1); 5552 Node* dest = argument(4); 5553 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5554 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5555 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5556 region->init_req(2, src_dest_conjoint); 5557 5558 record_for_igvn(region); 5559 return _gvn.transform(region); 5560 } 5561 5562 //------------------------------inline_sha_implCompress----------------------- 5563 // 5564 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 5565 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 5566 // 5567 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 5568 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 5569 // 5570 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 5571 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 5572 // 5573 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 5574 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 5575 5576 Node* sha_obj = argument(0); 5577 Node* src = argument(1); // type oop 5578 Node* ofs = argument(2); // type int 5579 5580 const Type* src_type = src->Value(&_gvn); 5581 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5582 if (top_src == NULL || top_src->klass() == NULL) { 5583 // failed array check 5584 return false; 5585 } 5586 // Figure out the size and type of the elements we will be copying. 5587 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5588 if (src_elem != T_BYTE) { 5589 return false; 5590 } 5591 // 'src_start' points to src array + offset 5592 Node* src_start = array_element_address(src, ofs, src_elem); 5593 Node* state = NULL; 5594 address stubAddr; 5595 const char *stubName; 5596 5597 switch(id) { 5598 case vmIntrinsics::_sha_implCompress: 5599 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 5600 state = get_state_from_sha_object(sha_obj); 5601 stubAddr = StubRoutines::sha1_implCompress(); 5602 stubName = "sha1_implCompress"; 5603 break; 5604 case vmIntrinsics::_sha2_implCompress: 5605 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 5606 state = get_state_from_sha_object(sha_obj); 5607 stubAddr = StubRoutines::sha256_implCompress(); 5608 stubName = "sha256_implCompress"; 5609 break; 5610 case vmIntrinsics::_sha5_implCompress: 5611 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 5612 state = get_state_from_sha5_object(sha_obj); 5613 stubAddr = StubRoutines::sha512_implCompress(); 5614 stubName = "sha512_implCompress"; 5615 break; 5616 default: 5617 fatal_unexpected_iid(id); 5618 return false; 5619 } 5620 if (state == NULL) return false; 5621 5622 // Call the stub. 5623 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 5624 stubAddr, stubName, TypePtr::BOTTOM, 5625 src_start, state); 5626 5627 return true; 5628 } 5629 5630 //------------------------------inline_digestBase_implCompressMB----------------------- 5631 // 5632 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 5633 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 5634 // 5635 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 5636 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5637 "need SHA1/SHA256/SHA512 instruction support"); 5638 assert((uint)predicate < 3, "sanity"); 5639 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 5640 5641 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 5642 Node* src = argument(1); // byte[] array 5643 Node* ofs = argument(2); // type int 5644 Node* limit = argument(3); // type int 5645 5646 const Type* src_type = src->Value(&_gvn); 5647 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5648 if (top_src == NULL || top_src->klass() == NULL) { 5649 // failed array check 5650 return false; 5651 } 5652 // Figure out the size and type of the elements we will be copying. 5653 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5654 if (src_elem != T_BYTE) { 5655 return false; 5656 } 5657 // 'src_start' points to src array + offset 5658 Node* src_start = array_element_address(src, ofs, src_elem); 5659 5660 const char* klass_SHA_name = NULL; 5661 const char* stub_name = NULL; 5662 address stub_addr = NULL; 5663 bool long_state = false; 5664 5665 switch (predicate) { 5666 case 0: 5667 if (UseSHA1Intrinsics) { 5668 klass_SHA_name = "sun/security/provider/SHA"; 5669 stub_name = "sha1_implCompressMB"; 5670 stub_addr = StubRoutines::sha1_implCompressMB(); 5671 } 5672 break; 5673 case 1: 5674 if (UseSHA256Intrinsics) { 5675 klass_SHA_name = "sun/security/provider/SHA2"; 5676 stub_name = "sha256_implCompressMB"; 5677 stub_addr = StubRoutines::sha256_implCompressMB(); 5678 } 5679 break; 5680 case 2: 5681 if (UseSHA512Intrinsics) { 5682 klass_SHA_name = "sun/security/provider/SHA5"; 5683 stub_name = "sha512_implCompressMB"; 5684 stub_addr = StubRoutines::sha512_implCompressMB(); 5685 long_state = true; 5686 } 5687 break; 5688 default: 5689 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5690 } 5691 if (klass_SHA_name != NULL) { 5692 // get DigestBase klass to lookup for SHA klass 5693 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 5694 assert(tinst != NULL, "digestBase_obj is not instance???"); 5695 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5696 5697 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5698 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 5699 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5700 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 5701 } 5702 return false; 5703 } 5704 //------------------------------inline_sha_implCompressMB----------------------- 5705 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 5706 bool long_state, address stubAddr, const char *stubName, 5707 Node* src_start, Node* ofs, Node* limit) { 5708 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 5709 const TypeOopPtr* xtype = aklass->as_instance_type(); 5710 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 5711 sha_obj = _gvn.transform(sha_obj); 5712 5713 Node* state; 5714 if (long_state) { 5715 state = get_state_from_sha5_object(sha_obj); 5716 } else { 5717 state = get_state_from_sha_object(sha_obj); 5718 } 5719 if (state == NULL) return false; 5720 5721 // Call the stub. 5722 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5723 OptoRuntime::digestBase_implCompressMB_Type(), 5724 stubAddr, stubName, TypePtr::BOTTOM, 5725 src_start, state, ofs, limit); 5726 // return ofs (int) 5727 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5728 set_result(result); 5729 5730 return true; 5731 } 5732 5733 //------------------------------get_state_from_sha_object----------------------- 5734 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 5735 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 5736 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 5737 if (sha_state == NULL) return (Node *) NULL; 5738 5739 // now have the array, need to get the start address of the state array 5740 Node* state = array_element_address(sha_state, intcon(0), T_INT); 5741 return state; 5742 } 5743 5744 //------------------------------get_state_from_sha5_object----------------------- 5745 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 5746 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 5747 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 5748 if (sha_state == NULL) return (Node *) NULL; 5749 5750 // now have the array, need to get the start address of the state array 5751 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 5752 return state; 5753 } 5754 5755 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 5756 // Return node representing slow path of predicate check. 5757 // the pseudo code we want to emulate with this predicate is: 5758 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 5759 // 5760 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 5761 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5762 "need SHA1/SHA256/SHA512 instruction support"); 5763 assert((uint)predicate < 3, "sanity"); 5764 5765 // The receiver was checked for NULL already. 5766 Node* digestBaseObj = argument(0); 5767 5768 // get DigestBase klass for instanceOf check 5769 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 5770 assert(tinst != NULL, "digestBaseObj is null"); 5771 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5772 5773 const char* klass_SHA_name = NULL; 5774 switch (predicate) { 5775 case 0: 5776 if (UseSHA1Intrinsics) { 5777 // we want to do an instanceof comparison against the SHA class 5778 klass_SHA_name = "sun/security/provider/SHA"; 5779 } 5780 break; 5781 case 1: 5782 if (UseSHA256Intrinsics) { 5783 // we want to do an instanceof comparison against the SHA2 class 5784 klass_SHA_name = "sun/security/provider/SHA2"; 5785 } 5786 break; 5787 case 2: 5788 if (UseSHA512Intrinsics) { 5789 // we want to do an instanceof comparison against the SHA5 class 5790 klass_SHA_name = "sun/security/provider/SHA5"; 5791 } 5792 break; 5793 default: 5794 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5795 } 5796 5797 ciKlass* klass_SHA = NULL; 5798 if (klass_SHA_name != NULL) { 5799 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5800 } 5801 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 5802 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 5803 Node* ctrl = control(); 5804 set_control(top()); // no intrinsic path 5805 return ctrl; 5806 } 5807 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5808 5809 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 5810 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 5811 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5812 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5813 5814 return instof_false; // even if it is NULL 5815 } 5816 5817 bool LibraryCallKit::inline_profileBranch() { 5818 Node* counts = argument(1); 5819 const TypeAryPtr* ary = NULL; 5820 ciArray* aobj = NULL; 5821 if (counts->is_Con() 5822 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 5823 && (aobj = ary->const_oop()->as_array()) != NULL) { 5824 assert(aobj->length() == 2, ""); 5825 jint taken = aobj->element_value(1).as_int(); 5826 jint not_taken = aobj->element_value(0).as_int(); 5827 5828 if (C->log() != NULL) { 5829 C->log()->elem("profile_branch taken='%d' not_taken='%d'", 5830 taken, not_taken); 5831 } 5832 5833 if (taken + not_taken == 0) { 5834 // According to profile, never executed. 5835 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 5836 Deoptimization::Action_reinterpret); 5837 return true; 5838 } 5839 // Stop profiling 5840 Node* opq = _gvn.transform(new Opaque4Node(argument(0), taken, not_taken)); 5841 C->record_for_igvn(opq); 5842 set_result(opq); 5843 return true; 5844 } else { 5845 // Continue profiling 5846 return false; 5847 } 5848 }