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/parse.hpp" 45 #include "opto/runtime.hpp" 46 #include "opto/subnode.hpp" 47 #include "prims/nativeLookup.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 #include "trace/traceMacros.hpp" 50 51 class LibraryIntrinsic : public InlineCallGenerator { 52 // Extend the set of intrinsics known to the runtime: 53 public: 54 private: 55 bool _is_virtual; 56 bool _does_virtual_dispatch; 57 int8_t _predicates_count; // Intrinsic is predicated by several conditions 58 int8_t _last_predicate; // Last generated predicate 59 vmIntrinsics::ID _intrinsic_id; 60 61 public: 62 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 63 : InlineCallGenerator(m), 64 _is_virtual(is_virtual), 65 _does_virtual_dispatch(does_virtual_dispatch), 66 _predicates_count((int8_t)predicates_count), 67 _last_predicate((int8_t)-1), 68 _intrinsic_id(id) 69 { 70 } 71 virtual bool is_intrinsic() const { return true; } 72 virtual bool is_virtual() const { return _is_virtual; } 73 virtual bool is_predicated() const { return _predicates_count > 0; } 74 virtual int predicates_count() const { return _predicates_count; } 75 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 76 virtual JVMState* generate(JVMState* jvms); 77 virtual Node* generate_predicate(JVMState* jvms, int predicate); 78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 79 }; 80 81 82 // Local helper class for LibraryIntrinsic: 83 class LibraryCallKit : public GraphKit { 84 private: 85 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 86 Node* _result; // the result node, if any 87 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 88 89 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 90 91 public: 92 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 93 : GraphKit(jvms), 94 _intrinsic(intrinsic), 95 _result(NULL) 96 { 97 // Check if this is a root compile. In that case we don't have a caller. 98 if (!jvms->has_method()) { 99 _reexecute_sp = sp(); 100 } else { 101 // Find out how many arguments the interpreter needs when deoptimizing 102 // and save the stack pointer value so it can used by uncommon_trap. 103 // We find the argument count by looking at the declared signature. 104 bool ignored_will_link; 105 ciSignature* declared_signature = NULL; 106 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 107 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 108 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 109 } 110 } 111 112 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 113 114 ciMethod* caller() const { return jvms()->method(); } 115 int bci() const { return jvms()->bci(); } 116 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 117 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 118 ciMethod* callee() const { return _intrinsic->method(); } 119 120 bool try_to_inline(int predicate); 121 Node* try_to_predicate(int predicate); 122 123 void push_result() { 124 // Push the result onto the stack. 125 if (!stopped() && result() != NULL) { 126 BasicType bt = result()->bottom_type()->basic_type(); 127 push_node(bt, result()); 128 } 129 } 130 131 private: 132 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 133 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 134 } 135 136 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 137 void set_result(RegionNode* region, PhiNode* value); 138 Node* result() { return _result; } 139 140 virtual int reexecute_sp() { return _reexecute_sp; } 141 142 // Helper functions to inline natives 143 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 144 Node* generate_slow_guard(Node* test, RegionNode* region); 145 Node* generate_fair_guard(Node* test, RegionNode* region); 146 Node* generate_negative_guard(Node* index, RegionNode* region, 147 // resulting CastII of index: 148 Node* *pos_index = NULL); 149 Node* generate_nonpositive_guard(Node* index, bool never_negative, 150 // resulting CastII of index: 151 Node* *pos_index = NULL); 152 Node* generate_limit_guard(Node* offset, Node* subseq_length, 153 Node* array_length, 154 RegionNode* region); 155 Node* generate_current_thread(Node* &tls_output); 156 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 157 bool disjoint_bases, const char* &name, bool dest_uninitialized); 158 Node* load_mirror_from_klass(Node* klass); 159 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path, 161 int offset); 162 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 163 RegionNode* region, int null_path) { 164 int offset = java_lang_Class::klass_offset_in_bytes(); 165 return load_klass_from_mirror_common(mirror, never_see_null, 166 region, null_path, 167 offset); 168 } 169 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 170 RegionNode* region, int null_path) { 171 int offset = java_lang_Class::array_klass_offset_in_bytes(); 172 return load_klass_from_mirror_common(mirror, never_see_null, 173 region, null_path, 174 offset); 175 } 176 Node* generate_access_flags_guard(Node* kls, 177 int modifier_mask, int modifier_bits, 178 RegionNode* region); 179 Node* generate_interface_guard(Node* kls, RegionNode* region); 180 Node* generate_array_guard(Node* kls, RegionNode* region) { 181 return generate_array_guard_common(kls, region, false, false); 182 } 183 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 184 return generate_array_guard_common(kls, region, false, true); 185 } 186 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 187 return generate_array_guard_common(kls, region, true, false); 188 } 189 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 190 return generate_array_guard_common(kls, region, true, true); 191 } 192 Node* generate_array_guard_common(Node* kls, RegionNode* region, 193 bool obj_array, bool not_array); 194 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 195 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 196 bool is_virtual = false, bool is_static = false); 197 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 198 return generate_method_call(method_id, false, true); 199 } 200 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 201 return generate_method_call(method_id, true, false); 202 } 203 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 204 205 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 206 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 207 bool inline_string_compareTo(); 208 bool inline_string_indexOf(); 209 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 210 bool inline_string_equals(); 211 Node* round_double_node(Node* n); 212 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 213 bool inline_math_native(vmIntrinsics::ID id); 214 bool inline_trig(vmIntrinsics::ID id); 215 bool inline_math(vmIntrinsics::ID id); 216 template <typename OverflowOp> 217 bool inline_math_overflow(Node* arg1, Node* arg2); 218 void inline_math_mathExact(Node* math, Node* test); 219 bool inline_math_addExactI(bool is_increment); 220 bool inline_math_addExactL(bool is_increment); 221 bool inline_math_multiplyExactI(); 222 bool inline_math_multiplyExactL(); 223 bool inline_math_negateExactI(); 224 bool inline_math_negateExactL(); 225 bool inline_math_subtractExactI(bool is_decrement); 226 bool inline_math_subtractExactL(bool is_decrement); 227 bool inline_exp(); 228 bool inline_pow(); 229 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 230 bool inline_min_max(vmIntrinsics::ID id); 231 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 232 // This returns Type::AnyPtr, RawPtr, or OopPtr. 233 int classify_unsafe_addr(Node* &base, Node* &offset); 234 Node* make_unsafe_address(Node* base, Node* offset); 235 // Helper for inline_unsafe_access. 236 // Generates the guards that check whether the result of 237 // Unsafe.getObject should be recorded in an SATB log buffer. 238 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 239 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 240 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 241 static bool klass_needs_init_guard(Node* kls); 242 bool inline_unsafe_allocate(); 243 bool inline_unsafe_copyMemory(); 244 bool inline_native_currentThread(); 245 #ifdef TRACE_HAVE_INTRINSICS 246 bool inline_native_classID(); 247 bool inline_native_threadID(); 248 #endif 249 bool inline_native_time_funcs(address method, const char* funcName); 250 bool inline_native_isInterrupted(); 251 bool inline_native_Class_query(vmIntrinsics::ID id); 252 bool inline_native_subtype_check(); 253 254 bool inline_native_newArray(); 255 bool inline_native_getLength(); 256 bool inline_array_copyOf(bool is_copyOfRange); 257 bool inline_array_equals(); 258 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 259 bool inline_native_clone(bool is_virtual); 260 bool inline_native_Reflection_getCallerClass(); 261 // Helper function for inlining native object hash method 262 bool inline_native_hashcode(bool is_virtual, bool is_static); 263 bool inline_native_getClass(); 264 265 // Helper functions for inlining arraycopy 266 bool inline_arraycopy(); 267 void generate_arraycopy(const TypePtr* adr_type, 268 BasicType basic_elem_type, 269 Node* src, Node* src_offset, 270 Node* dest, Node* dest_offset, 271 Node* copy_length, 272 bool disjoint_bases = false, 273 bool length_never_negative = false, 274 RegionNode* slow_region = NULL); 275 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 276 RegionNode* slow_region); 277 void generate_clear_array(const TypePtr* adr_type, 278 Node* dest, 279 BasicType basic_elem_type, 280 Node* slice_off, 281 Node* slice_len, 282 Node* slice_end); 283 bool generate_block_arraycopy(const TypePtr* adr_type, 284 BasicType basic_elem_type, 285 AllocateNode* alloc, 286 Node* src, Node* src_offset, 287 Node* dest, Node* dest_offset, 288 Node* dest_size, bool dest_uninitialized); 289 void generate_slow_arraycopy(const TypePtr* adr_type, 290 Node* src, Node* src_offset, 291 Node* dest, Node* dest_offset, 292 Node* copy_length, bool dest_uninitialized); 293 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 294 Node* dest_elem_klass, 295 Node* src, Node* src_offset, 296 Node* dest, Node* dest_offset, 297 Node* copy_length, bool dest_uninitialized); 298 Node* generate_generic_arraycopy(const TypePtr* adr_type, 299 Node* src, Node* src_offset, 300 Node* dest, Node* dest_offset, 301 Node* copy_length, bool dest_uninitialized); 302 void generate_unchecked_arraycopy(const TypePtr* adr_type, 303 BasicType basic_elem_type, 304 bool disjoint_bases, 305 Node* src, Node* src_offset, 306 Node* dest, Node* dest_offset, 307 Node* copy_length, bool dest_uninitialized); 308 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 309 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 310 bool inline_unsafe_ordered_store(BasicType type); 311 bool inline_unsafe_fence(vmIntrinsics::ID id); 312 bool inline_fp_conversions(vmIntrinsics::ID id); 313 bool inline_number_methods(vmIntrinsics::ID id); 314 bool inline_reference_get(); 315 bool inline_aescrypt_Block(vmIntrinsics::ID id); 316 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 317 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 318 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 319 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 320 bool inline_sha_implCompress(vmIntrinsics::ID id); 321 bool inline_digestBase_implCompressMB(int predicate); 322 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 323 bool long_state, address stubAddr, const char *stubName, 324 Node* src_start, Node* ofs, Node* limit); 325 Node* get_state_from_sha_object(Node *sha_object); 326 Node* get_state_from_sha5_object(Node *sha_object); 327 Node* inline_digestBase_implCompressMB_predicate(int predicate); 328 bool inline_encodeISOArray(); 329 bool inline_updateCRC32(); 330 bool inline_updateBytesCRC32(); 331 bool inline_updateByteBufferCRC32(); 332 }; 333 334 335 //---------------------------make_vm_intrinsic---------------------------- 336 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 337 vmIntrinsics::ID id = m->intrinsic_id(); 338 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 339 340 if (DisableIntrinsic[0] != '\0' 341 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 342 // disabled by a user request on the command line: 343 // example: -XX:DisableIntrinsic=_hashCode,_getClass 344 return NULL; 345 } 346 347 if (!m->is_loaded()) { 348 // do not attempt to inline unloaded methods 349 return NULL; 350 } 351 352 // Only a few intrinsics implement a virtual dispatch. 353 // They are expensive calls which are also frequently overridden. 354 if (is_virtual) { 355 switch (id) { 356 case vmIntrinsics::_hashCode: 357 case vmIntrinsics::_clone: 358 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 359 break; 360 default: 361 return NULL; 362 } 363 } 364 365 // -XX:-InlineNatives disables nearly all intrinsics: 366 if (!InlineNatives) { 367 switch (id) { 368 case vmIntrinsics::_indexOf: 369 case vmIntrinsics::_compareTo: 370 case vmIntrinsics::_equals: 371 case vmIntrinsics::_equalsC: 372 case vmIntrinsics::_getAndAddInt: 373 case vmIntrinsics::_getAndAddLong: 374 case vmIntrinsics::_getAndSetInt: 375 case vmIntrinsics::_getAndSetLong: 376 case vmIntrinsics::_getAndSetObject: 377 case vmIntrinsics::_loadFence: 378 case vmIntrinsics::_storeFence: 379 case vmIntrinsics::_fullFence: 380 break; // InlineNatives does not control String.compareTo 381 case vmIntrinsics::_Reference_get: 382 break; // InlineNatives does not control Reference.get 383 default: 384 return NULL; 385 } 386 } 387 388 int predicates = 0; 389 bool does_virtual_dispatch = false; 390 391 switch (id) { 392 case vmIntrinsics::_compareTo: 393 if (!SpecialStringCompareTo) return NULL; 394 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 395 break; 396 case vmIntrinsics::_indexOf: 397 if (!SpecialStringIndexOf) return NULL; 398 break; 399 case vmIntrinsics::_equals: 400 if (!SpecialStringEquals) return NULL; 401 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 402 break; 403 case vmIntrinsics::_equalsC: 404 if (!SpecialArraysEquals) return NULL; 405 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 406 break; 407 case vmIntrinsics::_arraycopy: 408 if (!InlineArrayCopy) return NULL; 409 break; 410 case vmIntrinsics::_copyMemory: 411 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 412 if (!InlineArrayCopy) return NULL; 413 break; 414 case vmIntrinsics::_hashCode: 415 if (!InlineObjectHash) return NULL; 416 does_virtual_dispatch = true; 417 break; 418 case vmIntrinsics::_clone: 419 does_virtual_dispatch = true; 420 case vmIntrinsics::_copyOf: 421 case vmIntrinsics::_copyOfRange: 422 if (!InlineObjectCopy) return NULL; 423 // These also use the arraycopy intrinsic mechanism: 424 if (!InlineArrayCopy) return NULL; 425 break; 426 case vmIntrinsics::_encodeISOArray: 427 if (!SpecialEncodeISOArray) return NULL; 428 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 429 break; 430 case vmIntrinsics::_checkIndex: 431 // We do not intrinsify this. The optimizer does fine with it. 432 return NULL; 433 434 case vmIntrinsics::_getCallerClass: 435 if (!InlineReflectionGetCallerClass) return NULL; 436 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 437 break; 438 439 case vmIntrinsics::_bitCount_i: 440 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 441 break; 442 443 case vmIntrinsics::_bitCount_l: 444 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 445 break; 446 447 case vmIntrinsics::_numberOfLeadingZeros_i: 448 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 449 break; 450 451 case vmIntrinsics::_numberOfLeadingZeros_l: 452 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 453 break; 454 455 case vmIntrinsics::_numberOfTrailingZeros_i: 456 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 457 break; 458 459 case vmIntrinsics::_numberOfTrailingZeros_l: 460 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 461 break; 462 463 case vmIntrinsics::_reverseBytes_c: 464 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 465 break; 466 case vmIntrinsics::_reverseBytes_s: 467 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 468 break; 469 case vmIntrinsics::_reverseBytes_i: 470 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 471 break; 472 case vmIntrinsics::_reverseBytes_l: 473 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 474 break; 475 476 case vmIntrinsics::_Reference_get: 477 // Use the intrinsic version of Reference.get() so that the value in 478 // the referent field can be registered by the G1 pre-barrier code. 479 // Also add memory barrier to prevent commoning reads from this field 480 // across safepoint since GC can change it value. 481 break; 482 483 case vmIntrinsics::_compareAndSwapObject: 484 #ifdef _LP64 485 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 486 #endif 487 break; 488 489 case vmIntrinsics::_compareAndSwapLong: 490 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 491 break; 492 493 case vmIntrinsics::_getAndAddInt: 494 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 495 break; 496 497 case vmIntrinsics::_getAndAddLong: 498 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 499 break; 500 501 case vmIntrinsics::_getAndSetInt: 502 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 503 break; 504 505 case vmIntrinsics::_getAndSetLong: 506 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 507 break; 508 509 case vmIntrinsics::_getAndSetObject: 510 #ifdef _LP64 511 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 512 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 513 break; 514 #else 515 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 516 break; 517 #endif 518 519 case vmIntrinsics::_aescrypt_encryptBlock: 520 case vmIntrinsics::_aescrypt_decryptBlock: 521 if (!UseAESIntrinsics) return NULL; 522 break; 523 524 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 525 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 526 if (!UseAESIntrinsics) return NULL; 527 // these two require the predicated logic 528 predicates = 1; 529 break; 530 531 case vmIntrinsics::_sha_implCompress: 532 if (!UseSHA1Intrinsics) return NULL; 533 break; 534 535 case vmIntrinsics::_sha2_implCompress: 536 if (!UseSHA256Intrinsics) return NULL; 537 break; 538 539 case vmIntrinsics::_sha5_implCompress: 540 if (!UseSHA512Intrinsics) return NULL; 541 break; 542 543 case vmIntrinsics::_digestBase_implCompressMB: 544 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 545 predicates = 3; 546 break; 547 548 case vmIntrinsics::_updateCRC32: 549 case vmIntrinsics::_updateBytesCRC32: 550 case vmIntrinsics::_updateByteBufferCRC32: 551 if (!UseCRC32Intrinsics) return NULL; 552 break; 553 554 case vmIntrinsics::_incrementExactI: 555 case vmIntrinsics::_addExactI: 556 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 557 break; 558 case vmIntrinsics::_incrementExactL: 559 case vmIntrinsics::_addExactL: 560 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 561 break; 562 case vmIntrinsics::_decrementExactI: 563 case vmIntrinsics::_subtractExactI: 564 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 565 break; 566 case vmIntrinsics::_decrementExactL: 567 case vmIntrinsics::_subtractExactL: 568 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 569 break; 570 case vmIntrinsics::_negateExactI: 571 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 572 break; 573 case vmIntrinsics::_negateExactL: 574 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 575 break; 576 case vmIntrinsics::_multiplyExactI: 577 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 578 break; 579 case vmIntrinsics::_multiplyExactL: 580 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 581 break; 582 583 default: 584 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 585 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 586 break; 587 } 588 589 // -XX:-InlineClassNatives disables natives from the Class class. 590 // The flag applies to all reflective calls, notably Array.newArray 591 // (visible to Java programmers as Array.newInstance). 592 if (m->holder()->name() == ciSymbol::java_lang_Class() || 593 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 594 if (!InlineClassNatives) return NULL; 595 } 596 597 // -XX:-InlineThreadNatives disables natives from the Thread class. 598 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 599 if (!InlineThreadNatives) return NULL; 600 } 601 602 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 603 if (m->holder()->name() == ciSymbol::java_lang_Math() || 604 m->holder()->name() == ciSymbol::java_lang_Float() || 605 m->holder()->name() == ciSymbol::java_lang_Double()) { 606 if (!InlineMathNatives) return NULL; 607 } 608 609 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 610 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 611 if (!InlineUnsafeOps) return NULL; 612 } 613 614 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 615 } 616 617 //----------------------register_library_intrinsics----------------------- 618 // Initialize this file's data structures, for each Compile instance. 619 void Compile::register_library_intrinsics() { 620 // Nothing to do here. 621 } 622 623 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 624 LibraryCallKit kit(jvms, this); 625 Compile* C = kit.C; 626 int nodes = C->unique(); 627 #ifndef PRODUCT 628 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 629 char buf[1000]; 630 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 631 tty->print_cr("Intrinsic %s", str); 632 } 633 #endif 634 ciMethod* callee = kit.callee(); 635 const int bci = kit.bci(); 636 637 // Try to inline the intrinsic. 638 if (kit.try_to_inline(_last_predicate)) { 639 if (C->print_intrinsics() || C->print_inlining()) { 640 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 641 } 642 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 643 if (C->log()) { 644 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 645 vmIntrinsics::name_at(intrinsic_id()), 646 (is_virtual() ? " virtual='1'" : ""), 647 C->unique() - nodes); 648 } 649 // Push the result from the inlined method onto the stack. 650 kit.push_result(); 651 C->print_inlining_update(this); 652 return kit.transfer_exceptions_into_jvms(); 653 } 654 655 // The intrinsic bailed out 656 if (C->print_intrinsics() || C->print_inlining()) { 657 if (jvms->has_method()) { 658 // Not a root compile. 659 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 660 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 661 } else { 662 // Root compile 663 tty->print("Did not generate intrinsic %s%s at bci:%d in", 664 vmIntrinsics::name_at(intrinsic_id()), 665 (is_virtual() ? " (virtual)" : ""), bci); 666 } 667 } 668 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 669 C->print_inlining_update(this); 670 return NULL; 671 } 672 673 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 674 LibraryCallKit kit(jvms, this); 675 Compile* C = kit.C; 676 int nodes = C->unique(); 677 _last_predicate = predicate; 678 #ifndef PRODUCT 679 assert(is_predicated() && predicate < predicates_count(), "sanity"); 680 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 681 char buf[1000]; 682 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 683 tty->print_cr("Predicate for intrinsic %s", str); 684 } 685 #endif 686 ciMethod* callee = kit.callee(); 687 const int bci = kit.bci(); 688 689 Node* slow_ctl = kit.try_to_predicate(predicate); 690 if (!kit.failing()) { 691 if (C->print_intrinsics() || C->print_inlining()) { 692 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 693 } 694 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 695 if (C->log()) { 696 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 697 vmIntrinsics::name_at(intrinsic_id()), 698 (is_virtual() ? " virtual='1'" : ""), 699 C->unique() - nodes); 700 } 701 return slow_ctl; // Could be NULL if the check folds. 702 } 703 704 // The intrinsic bailed out 705 if (C->print_intrinsics() || C->print_inlining()) { 706 if (jvms->has_method()) { 707 // Not a root compile. 708 const char* msg = "failed to generate predicate for intrinsic"; 709 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 710 } else { 711 // Root compile 712 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 713 vmIntrinsics::name_at(intrinsic_id()), 714 (is_virtual() ? " (virtual)" : ""), bci); 715 } 716 } 717 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 718 return NULL; 719 } 720 721 bool LibraryCallKit::try_to_inline(int predicate) { 722 // Handle symbolic names for otherwise undistinguished boolean switches: 723 const bool is_store = true; 724 const bool is_native_ptr = true; 725 const bool is_static = true; 726 const bool is_volatile = true; 727 728 if (!jvms()->has_method()) { 729 // Root JVMState has a null method. 730 assert(map()->memory()->Opcode() == Op_Parm, ""); 731 // Insert the memory aliasing node 732 set_all_memory(reset_memory()); 733 } 734 assert(merged_memory(), ""); 735 736 737 switch (intrinsic_id()) { 738 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 739 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 740 case vmIntrinsics::_getClass: return inline_native_getClass(); 741 742 case vmIntrinsics::_dsin: 743 case vmIntrinsics::_dcos: 744 case vmIntrinsics::_dtan: 745 case vmIntrinsics::_dabs: 746 case vmIntrinsics::_datan2: 747 case vmIntrinsics::_dsqrt: 748 case vmIntrinsics::_dexp: 749 case vmIntrinsics::_dlog: 750 case vmIntrinsics::_dlog10: 751 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 752 753 case vmIntrinsics::_min: 754 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 755 756 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 757 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 758 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 759 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 760 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 761 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 762 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 763 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 764 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 765 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 766 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 767 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 768 769 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 770 771 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 772 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 773 case vmIntrinsics::_equals: return inline_string_equals(); 774 775 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 776 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 777 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 778 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 779 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 780 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 781 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 782 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 783 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 784 785 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 786 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 787 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 788 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 789 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 790 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 791 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 792 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 793 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 794 795 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 796 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 797 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 798 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 799 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 800 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 801 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 802 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 803 804 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 805 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 806 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 807 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 808 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 809 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 810 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 811 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 812 813 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 814 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 815 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 816 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 817 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 818 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 819 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 820 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 821 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 822 823 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 824 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 825 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 826 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 827 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 828 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 829 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 830 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 831 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 832 833 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 834 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 835 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 836 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 837 838 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 839 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 840 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 841 842 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 843 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 844 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 845 846 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 847 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 848 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 849 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 850 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 851 852 case vmIntrinsics::_loadFence: 853 case vmIntrinsics::_storeFence: 854 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 855 856 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 857 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 858 859 #ifdef TRACE_HAVE_INTRINSICS 860 case vmIntrinsics::_classID: return inline_native_classID(); 861 case vmIntrinsics::_threadID: return inline_native_threadID(); 862 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 863 #endif 864 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 865 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 866 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 867 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 868 case vmIntrinsics::_newArray: return inline_native_newArray(); 869 case vmIntrinsics::_getLength: return inline_native_getLength(); 870 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 871 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 872 case vmIntrinsics::_equalsC: return inline_array_equals(); 873 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 874 875 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 876 877 case vmIntrinsics::_isInstance: 878 case vmIntrinsics::_getModifiers: 879 case vmIntrinsics::_isInterface: 880 case vmIntrinsics::_isArray: 881 case vmIntrinsics::_isPrimitive: 882 case vmIntrinsics::_getSuperclass: 883 case vmIntrinsics::_getComponentType: 884 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 885 886 case vmIntrinsics::_floatToRawIntBits: 887 case vmIntrinsics::_floatToIntBits: 888 case vmIntrinsics::_intBitsToFloat: 889 case vmIntrinsics::_doubleToRawLongBits: 890 case vmIntrinsics::_doubleToLongBits: 891 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 892 893 case vmIntrinsics::_numberOfLeadingZeros_i: 894 case vmIntrinsics::_numberOfLeadingZeros_l: 895 case vmIntrinsics::_numberOfTrailingZeros_i: 896 case vmIntrinsics::_numberOfTrailingZeros_l: 897 case vmIntrinsics::_bitCount_i: 898 case vmIntrinsics::_bitCount_l: 899 case vmIntrinsics::_reverseBytes_i: 900 case vmIntrinsics::_reverseBytes_l: 901 case vmIntrinsics::_reverseBytes_s: 902 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 903 904 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 905 906 case vmIntrinsics::_Reference_get: return inline_reference_get(); 907 908 case vmIntrinsics::_aescrypt_encryptBlock: 909 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 910 911 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 912 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 913 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 914 915 case vmIntrinsics::_sha_implCompress: 916 case vmIntrinsics::_sha2_implCompress: 917 case vmIntrinsics::_sha5_implCompress: 918 return inline_sha_implCompress(intrinsic_id()); 919 920 case vmIntrinsics::_digestBase_implCompressMB: 921 return inline_digestBase_implCompressMB(predicate); 922 923 case vmIntrinsics::_encodeISOArray: 924 return inline_encodeISOArray(); 925 926 case vmIntrinsics::_updateCRC32: 927 return inline_updateCRC32(); 928 case vmIntrinsics::_updateBytesCRC32: 929 return inline_updateBytesCRC32(); 930 case vmIntrinsics::_updateByteBufferCRC32: 931 return inline_updateByteBufferCRC32(); 932 933 default: 934 // If you get here, it may be that someone has added a new intrinsic 935 // to the list in vmSymbols.hpp without implementing it here. 936 #ifndef PRODUCT 937 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 938 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 939 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 940 } 941 #endif 942 return false; 943 } 944 } 945 946 Node* LibraryCallKit::try_to_predicate(int predicate) { 947 if (!jvms()->has_method()) { 948 // Root JVMState has a null method. 949 assert(map()->memory()->Opcode() == Op_Parm, ""); 950 // Insert the memory aliasing node 951 set_all_memory(reset_memory()); 952 } 953 assert(merged_memory(), ""); 954 955 switch (intrinsic_id()) { 956 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 957 return inline_cipherBlockChaining_AESCrypt_predicate(false); 958 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 959 return inline_cipherBlockChaining_AESCrypt_predicate(true); 960 case vmIntrinsics::_digestBase_implCompressMB: 961 return inline_digestBase_implCompressMB_predicate(predicate); 962 963 default: 964 // If you get here, it may be that someone has added a new intrinsic 965 // to the list in vmSymbols.hpp without implementing it here. 966 #ifndef PRODUCT 967 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 968 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 969 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 970 } 971 #endif 972 Node* slow_ctl = control(); 973 set_control(top()); // No fast path instrinsic 974 return slow_ctl; 975 } 976 } 977 978 //------------------------------set_result------------------------------- 979 // Helper function for finishing intrinsics. 980 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 981 record_for_igvn(region); 982 set_control(_gvn.transform(region)); 983 set_result( _gvn.transform(value)); 984 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 985 } 986 987 //------------------------------generate_guard--------------------------- 988 // Helper function for generating guarded fast-slow graph structures. 989 // The given 'test', if true, guards a slow path. If the test fails 990 // then a fast path can be taken. (We generally hope it fails.) 991 // In all cases, GraphKit::control() is updated to the fast path. 992 // The returned value represents the control for the slow path. 993 // The return value is never 'top'; it is either a valid control 994 // or NULL if it is obvious that the slow path can never be taken. 995 // Also, if region and the slow control are not NULL, the slow edge 996 // is appended to the region. 997 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 998 if (stopped()) { 999 // Already short circuited. 1000 return NULL; 1001 } 1002 1003 // Build an if node and its projections. 1004 // If test is true we take the slow path, which we assume is uncommon. 1005 if (_gvn.type(test) == TypeInt::ZERO) { 1006 // The slow branch is never taken. No need to build this guard. 1007 return NULL; 1008 } 1009 1010 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1011 1012 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1013 if (if_slow == top()) { 1014 // The slow branch is never taken. No need to build this guard. 1015 return NULL; 1016 } 1017 1018 if (region != NULL) 1019 region->add_req(if_slow); 1020 1021 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1022 set_control(if_fast); 1023 1024 return if_slow; 1025 } 1026 1027 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1028 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1029 } 1030 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1031 return generate_guard(test, region, PROB_FAIR); 1032 } 1033 1034 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1035 Node* *pos_index) { 1036 if (stopped()) 1037 return NULL; // already stopped 1038 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1039 return NULL; // index is already adequately typed 1040 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1041 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1042 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1043 if (is_neg != NULL && pos_index != NULL) { 1044 // Emulate effect of Parse::adjust_map_after_if. 1045 Node* ccast = new CastIINode(index, TypeInt::POS); 1046 ccast->set_req(0, control()); 1047 (*pos_index) = _gvn.transform(ccast); 1048 } 1049 return is_neg; 1050 } 1051 1052 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 1053 Node* *pos_index) { 1054 if (stopped()) 1055 return NULL; // already stopped 1056 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 1057 return NULL; // index is already adequately typed 1058 Node* cmp_le = _gvn.transform(new CmpINode(index, intcon(0))); 1059 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 1060 Node* bol_le = _gvn.transform(new BoolNode(cmp_le, le_or_eq)); 1061 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 1062 if (is_notp != NULL && pos_index != NULL) { 1063 // Emulate effect of Parse::adjust_map_after_if. 1064 Node* ccast = new CastIINode(index, TypeInt::POS1); 1065 ccast->set_req(0, control()); 1066 (*pos_index) = _gvn.transform(ccast); 1067 } 1068 return is_notp; 1069 } 1070 1071 // Make sure that 'position' is a valid limit index, in [0..length]. 1072 // There are two equivalent plans for checking this: 1073 // A. (offset + copyLength) unsigned<= arrayLength 1074 // B. offset <= (arrayLength - copyLength) 1075 // We require that all of the values above, except for the sum and 1076 // difference, are already known to be non-negative. 1077 // Plan A is robust in the face of overflow, if offset and copyLength 1078 // are both hugely positive. 1079 // 1080 // Plan B is less direct and intuitive, but it does not overflow at 1081 // all, since the difference of two non-negatives is always 1082 // representable. Whenever Java methods must perform the equivalent 1083 // check they generally use Plan B instead of Plan A. 1084 // For the moment we use Plan A. 1085 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1086 Node* subseq_length, 1087 Node* array_length, 1088 RegionNode* region) { 1089 if (stopped()) 1090 return NULL; // already stopped 1091 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1092 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1093 return NULL; // common case of whole-array copy 1094 Node* last = subseq_length; 1095 if (!zero_offset) // last += offset 1096 last = _gvn.transform(new AddINode(last, offset)); 1097 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1098 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1099 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1100 return is_over; 1101 } 1102 1103 1104 //--------------------------generate_current_thread-------------------- 1105 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1106 ciKlass* thread_klass = env()->Thread_klass(); 1107 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1108 Node* thread = _gvn.transform(new ThreadLocalNode()); 1109 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1110 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1111 tls_output = thread; 1112 return threadObj; 1113 } 1114 1115 1116 //------------------------------make_string_method_node------------------------ 1117 // Helper method for String intrinsic functions. This version is called 1118 // with str1 and str2 pointing to String object nodes. 1119 // 1120 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1121 Node* no_ctrl = NULL; 1122 1123 // Get start addr of string 1124 Node* str1_value = load_String_value(no_ctrl, str1); 1125 Node* str1_offset = load_String_offset(no_ctrl, str1); 1126 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1127 1128 // Get length of string 1 1129 Node* str1_len = load_String_length(no_ctrl, str1); 1130 1131 Node* str2_value = load_String_value(no_ctrl, str2); 1132 Node* str2_offset = load_String_offset(no_ctrl, str2); 1133 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1134 1135 Node* str2_len = NULL; 1136 Node* result = NULL; 1137 1138 switch (opcode) { 1139 case Op_StrIndexOf: 1140 // Get length of string 2 1141 str2_len = load_String_length(no_ctrl, str2); 1142 1143 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1144 str1_start, str1_len, str2_start, str2_len); 1145 break; 1146 case Op_StrComp: 1147 // Get length of string 2 1148 str2_len = load_String_length(no_ctrl, str2); 1149 1150 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1151 str1_start, str1_len, str2_start, str2_len); 1152 break; 1153 case Op_StrEquals: 1154 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1155 str1_start, str2_start, str1_len); 1156 break; 1157 default: 1158 ShouldNotReachHere(); 1159 return NULL; 1160 } 1161 1162 // All these intrinsics have checks. 1163 C->set_has_split_ifs(true); // Has chance for split-if optimization 1164 1165 return _gvn.transform(result); 1166 } 1167 1168 // Helper method for String intrinsic functions. This version is called 1169 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1170 // to Int nodes containing the lenghts of str1 and str2. 1171 // 1172 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1173 Node* result = NULL; 1174 switch (opcode) { 1175 case Op_StrIndexOf: 1176 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1177 str1_start, cnt1, str2_start, cnt2); 1178 break; 1179 case Op_StrComp: 1180 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1181 str1_start, cnt1, str2_start, cnt2); 1182 break; 1183 case Op_StrEquals: 1184 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1185 str1_start, str2_start, cnt1); 1186 break; 1187 default: 1188 ShouldNotReachHere(); 1189 return NULL; 1190 } 1191 1192 // All these intrinsics have checks. 1193 C->set_has_split_ifs(true); // Has chance for split-if optimization 1194 1195 return _gvn.transform(result); 1196 } 1197 1198 //------------------------------inline_string_compareTo------------------------ 1199 // public int java.lang.String.compareTo(String anotherString); 1200 bool LibraryCallKit::inline_string_compareTo() { 1201 Node* receiver = null_check(argument(0)); 1202 Node* arg = null_check(argument(1)); 1203 if (stopped()) { 1204 return true; 1205 } 1206 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1207 return true; 1208 } 1209 1210 //------------------------------inline_string_equals------------------------ 1211 bool LibraryCallKit::inline_string_equals() { 1212 Node* receiver = null_check_receiver(); 1213 // NOTE: Do not null check argument for String.equals() because spec 1214 // allows to specify NULL as argument. 1215 Node* argument = this->argument(1); 1216 if (stopped()) { 1217 return true; 1218 } 1219 1220 // paths (plus control) merge 1221 RegionNode* region = new RegionNode(5); 1222 Node* phi = new PhiNode(region, TypeInt::BOOL); 1223 1224 // does source == target string? 1225 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1226 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1227 1228 Node* if_eq = generate_slow_guard(bol, NULL); 1229 if (if_eq != NULL) { 1230 // receiver == argument 1231 phi->init_req(2, intcon(1)); 1232 region->init_req(2, if_eq); 1233 } 1234 1235 // get String klass for instanceOf 1236 ciInstanceKlass* klass = env()->String_klass(); 1237 1238 if (!stopped()) { 1239 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1240 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1241 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1242 1243 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1244 //instanceOf == true, fallthrough 1245 1246 if (inst_false != NULL) { 1247 phi->init_req(3, intcon(0)); 1248 region->init_req(3, inst_false); 1249 } 1250 } 1251 1252 if (!stopped()) { 1253 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1254 1255 // Properly cast the argument to String 1256 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1257 // This path is taken only when argument's type is String:NotNull. 1258 argument = cast_not_null(argument, false); 1259 1260 Node* no_ctrl = NULL; 1261 1262 // Get start addr of receiver 1263 Node* receiver_val = load_String_value(no_ctrl, receiver); 1264 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1265 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1266 1267 // Get length of receiver 1268 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1269 1270 // Get start addr of argument 1271 Node* argument_val = load_String_value(no_ctrl, argument); 1272 Node* argument_offset = load_String_offset(no_ctrl, argument); 1273 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1274 1275 // Get length of argument 1276 Node* argument_cnt = load_String_length(no_ctrl, argument); 1277 1278 // Check for receiver count != argument count 1279 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1280 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1281 Node* if_ne = generate_slow_guard(bol, NULL); 1282 if (if_ne != NULL) { 1283 phi->init_req(4, intcon(0)); 1284 region->init_req(4, if_ne); 1285 } 1286 1287 // Check for count == 0 is done by assembler code for StrEquals. 1288 1289 if (!stopped()) { 1290 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1291 phi->init_req(1, equals); 1292 region->init_req(1, control()); 1293 } 1294 } 1295 1296 // post merge 1297 set_control(_gvn.transform(region)); 1298 record_for_igvn(region); 1299 1300 set_result(_gvn.transform(phi)); 1301 return true; 1302 } 1303 1304 //------------------------------inline_array_equals---------------------------- 1305 bool LibraryCallKit::inline_array_equals() { 1306 Node* arg1 = argument(0); 1307 Node* arg2 = argument(1); 1308 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1309 return true; 1310 } 1311 1312 // Java version of String.indexOf(constant string) 1313 // class StringDecl { 1314 // StringDecl(char[] ca) { 1315 // offset = 0; 1316 // count = ca.length; 1317 // value = ca; 1318 // } 1319 // int offset; 1320 // int count; 1321 // char[] value; 1322 // } 1323 // 1324 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1325 // int targetOffset, int cache_i, int md2) { 1326 // int cache = cache_i; 1327 // int sourceOffset = string_object.offset; 1328 // int sourceCount = string_object.count; 1329 // int targetCount = target_object.length; 1330 // 1331 // int targetCountLess1 = targetCount - 1; 1332 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1333 // 1334 // char[] source = string_object.value; 1335 // char[] target = target_object; 1336 // int lastChar = target[targetCountLess1]; 1337 // 1338 // outer_loop: 1339 // for (int i = sourceOffset; i < sourceEnd; ) { 1340 // int src = source[i + targetCountLess1]; 1341 // if (src == lastChar) { 1342 // // With random strings and a 4-character alphabet, 1343 // // reverse matching at this point sets up 0.8% fewer 1344 // // frames, but (paradoxically) makes 0.3% more probes. 1345 // // Since those probes are nearer the lastChar probe, 1346 // // there is may be a net D$ win with reverse matching. 1347 // // But, reversing loop inhibits unroll of inner loop 1348 // // for unknown reason. So, does running outer loop from 1349 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1350 // for (int j = 0; j < targetCountLess1; j++) { 1351 // if (target[targetOffset + j] != source[i+j]) { 1352 // if ((cache & (1 << source[i+j])) == 0) { 1353 // if (md2 < j+1) { 1354 // i += j+1; 1355 // continue outer_loop; 1356 // } 1357 // } 1358 // i += md2; 1359 // continue outer_loop; 1360 // } 1361 // } 1362 // return i - sourceOffset; 1363 // } 1364 // if ((cache & (1 << src)) == 0) { 1365 // i += targetCountLess1; 1366 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1367 // i++; 1368 // } 1369 // return -1; 1370 // } 1371 1372 //------------------------------string_indexOf------------------------ 1373 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1374 jint cache_i, jint md2_i) { 1375 1376 Node* no_ctrl = NULL; 1377 float likely = PROB_LIKELY(0.9); 1378 float unlikely = PROB_UNLIKELY(0.9); 1379 1380 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1381 1382 Node* source = load_String_value(no_ctrl, string_object); 1383 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1384 Node* sourceCount = load_String_length(no_ctrl, string_object); 1385 1386 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1387 jint target_length = target_array->length(); 1388 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1389 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1390 1391 // String.value field is known to be @Stable. 1392 if (UseImplicitStableValues) { 1393 target = cast_array_to_stable(target, target_type); 1394 } 1395 1396 IdealKit kit(this, false, true); 1397 #define __ kit. 1398 Node* zero = __ ConI(0); 1399 Node* one = __ ConI(1); 1400 Node* cache = __ ConI(cache_i); 1401 Node* md2 = __ ConI(md2_i); 1402 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1403 Node* targetCount = __ ConI(target_length); 1404 Node* targetCountLess1 = __ ConI(target_length - 1); 1405 Node* targetOffset = __ ConI(targetOffset_i); 1406 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1407 1408 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1409 Node* outer_loop = __ make_label(2 /* goto */); 1410 Node* return_ = __ make_label(1); 1411 1412 __ set(rtn,__ ConI(-1)); 1413 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1414 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1415 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1416 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1417 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1418 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1419 Node* tpj = __ AddI(targetOffset, __ value(j)); 1420 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1421 Node* ipj = __ AddI(__ value(i), __ value(j)); 1422 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1423 __ if_then(targ, BoolTest::ne, src2); { 1424 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1425 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1426 __ increment(i, __ AddI(__ value(j), one)); 1427 __ goto_(outer_loop); 1428 } __ end_if(); __ dead(j); 1429 }__ end_if(); __ dead(j); 1430 __ increment(i, md2); 1431 __ goto_(outer_loop); 1432 }__ end_if(); 1433 __ increment(j, one); 1434 }__ end_loop(); __ dead(j); 1435 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1436 __ goto_(return_); 1437 }__ end_if(); 1438 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1439 __ increment(i, targetCountLess1); 1440 }__ end_if(); 1441 __ increment(i, one); 1442 __ bind(outer_loop); 1443 }__ end_loop(); __ dead(i); 1444 __ bind(return_); 1445 1446 // Final sync IdealKit and GraphKit. 1447 final_sync(kit); 1448 Node* result = __ value(rtn); 1449 #undef __ 1450 C->set_has_loops(true); 1451 return result; 1452 } 1453 1454 //------------------------------inline_string_indexOf------------------------ 1455 bool LibraryCallKit::inline_string_indexOf() { 1456 Node* receiver = argument(0); 1457 Node* arg = argument(1); 1458 1459 Node* result; 1460 // Disable the use of pcmpestri until it can be guaranteed that 1461 // the load doesn't cross into the uncommited space. 1462 if (Matcher::has_match_rule(Op_StrIndexOf) && 1463 UseSSE42Intrinsics) { 1464 // Generate SSE4.2 version of indexOf 1465 // We currently only have match rules that use SSE4.2 1466 1467 receiver = null_check(receiver); 1468 arg = null_check(arg); 1469 if (stopped()) { 1470 return true; 1471 } 1472 1473 ciInstanceKlass* str_klass = env()->String_klass(); 1474 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1475 1476 // Make the merge point 1477 RegionNode* result_rgn = new RegionNode(4); 1478 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1479 Node* no_ctrl = NULL; 1480 1481 // Get start addr of source string 1482 Node* source = load_String_value(no_ctrl, receiver); 1483 Node* source_offset = load_String_offset(no_ctrl, receiver); 1484 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1485 1486 // Get length of source string 1487 Node* source_cnt = load_String_length(no_ctrl, receiver); 1488 1489 // Get start addr of substring 1490 Node* substr = load_String_value(no_ctrl, arg); 1491 Node* substr_offset = load_String_offset(no_ctrl, arg); 1492 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1493 1494 // Get length of source string 1495 Node* substr_cnt = load_String_length(no_ctrl, arg); 1496 1497 // Check for substr count > string count 1498 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1499 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1500 Node* if_gt = generate_slow_guard(bol, NULL); 1501 if (if_gt != NULL) { 1502 result_phi->init_req(2, intcon(-1)); 1503 result_rgn->init_req(2, if_gt); 1504 } 1505 1506 if (!stopped()) { 1507 // Check for substr count == 0 1508 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1509 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1510 Node* if_zero = generate_slow_guard(bol, NULL); 1511 if (if_zero != NULL) { 1512 result_phi->init_req(3, intcon(0)); 1513 result_rgn->init_req(3, if_zero); 1514 } 1515 } 1516 1517 if (!stopped()) { 1518 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1519 result_phi->init_req(1, result); 1520 result_rgn->init_req(1, control()); 1521 } 1522 set_control(_gvn.transform(result_rgn)); 1523 record_for_igvn(result_rgn); 1524 result = _gvn.transform(result_phi); 1525 1526 } else { // Use LibraryCallKit::string_indexOf 1527 // don't intrinsify if argument isn't a constant string. 1528 if (!arg->is_Con()) { 1529 return false; 1530 } 1531 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1532 if (str_type == NULL) { 1533 return false; 1534 } 1535 ciInstanceKlass* klass = env()->String_klass(); 1536 ciObject* str_const = str_type->const_oop(); 1537 if (str_const == NULL || str_const->klass() != klass) { 1538 return false; 1539 } 1540 ciInstance* str = str_const->as_instance(); 1541 assert(str != NULL, "must be instance"); 1542 1543 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1544 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1545 1546 int o; 1547 int c; 1548 if (java_lang_String::has_offset_field()) { 1549 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1550 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1551 } else { 1552 o = 0; 1553 c = pat->length(); 1554 } 1555 1556 // constant strings have no offset and count == length which 1557 // simplifies the resulting code somewhat so lets optimize for that. 1558 if (o != 0 || c != pat->length()) { 1559 return false; 1560 } 1561 1562 receiver = null_check(receiver, T_OBJECT); 1563 // NOTE: No null check on the argument is needed since it's a constant String oop. 1564 if (stopped()) { 1565 return true; 1566 } 1567 1568 // The null string as a pattern always returns 0 (match at beginning of string) 1569 if (c == 0) { 1570 set_result(intcon(0)); 1571 return true; 1572 } 1573 1574 // Generate default indexOf 1575 jchar lastChar = pat->char_at(o + (c - 1)); 1576 int cache = 0; 1577 int i; 1578 for (i = 0; i < c - 1; i++) { 1579 assert(i < pat->length(), "out of range"); 1580 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1581 } 1582 1583 int md2 = c; 1584 for (i = 0; i < c - 1; i++) { 1585 assert(i < pat->length(), "out of range"); 1586 if (pat->char_at(o + i) == lastChar) { 1587 md2 = (c - 1) - i; 1588 } 1589 } 1590 1591 result = string_indexOf(receiver, pat, o, cache, md2); 1592 } 1593 set_result(result); 1594 return true; 1595 } 1596 1597 //--------------------------round_double_node-------------------------------- 1598 // Round a double node if necessary. 1599 Node* LibraryCallKit::round_double_node(Node* n) { 1600 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1601 n = _gvn.transform(new RoundDoubleNode(0, n)); 1602 return n; 1603 } 1604 1605 //------------------------------inline_math----------------------------------- 1606 // public static double Math.abs(double) 1607 // public static double Math.sqrt(double) 1608 // public static double Math.log(double) 1609 // public static double Math.log10(double) 1610 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1611 Node* arg = round_double_node(argument(0)); 1612 Node* n; 1613 switch (id) { 1614 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1615 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1616 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1617 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1618 default: fatal_unexpected_iid(id); break; 1619 } 1620 set_result(_gvn.transform(n)); 1621 return true; 1622 } 1623 1624 //------------------------------inline_trig---------------------------------- 1625 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1626 // argument reduction which will turn into a fast/slow diamond. 1627 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1628 Node* arg = round_double_node(argument(0)); 1629 Node* n = NULL; 1630 1631 switch (id) { 1632 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1633 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1634 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1635 default: fatal_unexpected_iid(id); break; 1636 } 1637 n = _gvn.transform(n); 1638 1639 // Rounding required? Check for argument reduction! 1640 if (Matcher::strict_fp_requires_explicit_rounding) { 1641 static const double pi_4 = 0.7853981633974483; 1642 static const double neg_pi_4 = -0.7853981633974483; 1643 // pi/2 in 80-bit extended precision 1644 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1645 // -pi/2 in 80-bit extended precision 1646 // 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}; 1647 // Cutoff value for using this argument reduction technique 1648 //static const double pi_2_minus_epsilon = 1.564660403643354; 1649 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1650 1651 // Pseudocode for sin: 1652 // if (x <= Math.PI / 4.0) { 1653 // if (x >= -Math.PI / 4.0) return fsin(x); 1654 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1655 // } else { 1656 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1657 // } 1658 // return StrictMath.sin(x); 1659 1660 // Pseudocode for cos: 1661 // if (x <= Math.PI / 4.0) { 1662 // if (x >= -Math.PI / 4.0) return fcos(x); 1663 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1664 // } else { 1665 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1666 // } 1667 // return StrictMath.cos(x); 1668 1669 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1670 // requires a special machine instruction to load it. Instead we'll try 1671 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1672 // probably do the math inside the SIN encoding. 1673 1674 // Make the merge point 1675 RegionNode* r = new RegionNode(3); 1676 Node* phi = new PhiNode(r, Type::DOUBLE); 1677 1678 // Flatten arg so we need only 1 test 1679 Node *abs = _gvn.transform(new AbsDNode(arg)); 1680 // Node for PI/4 constant 1681 Node *pi4 = makecon(TypeD::make(pi_4)); 1682 // Check PI/4 : abs(arg) 1683 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1684 // Check: If PI/4 < abs(arg) then go slow 1685 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1686 // Branch either way 1687 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1688 set_control(opt_iff(r,iff)); 1689 1690 // Set fast path result 1691 phi->init_req(2, n); 1692 1693 // Slow path - non-blocking leaf call 1694 Node* call = NULL; 1695 switch (id) { 1696 case vmIntrinsics::_dsin: 1697 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1698 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1699 "Sin", NULL, arg, top()); 1700 break; 1701 case vmIntrinsics::_dcos: 1702 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1703 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1704 "Cos", NULL, arg, top()); 1705 break; 1706 case vmIntrinsics::_dtan: 1707 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1708 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1709 "Tan", NULL, arg, top()); 1710 break; 1711 } 1712 assert(control()->in(0) == call, ""); 1713 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1714 r->init_req(1, control()); 1715 phi->init_req(1, slow_result); 1716 1717 // Post-merge 1718 set_control(_gvn.transform(r)); 1719 record_for_igvn(r); 1720 n = _gvn.transform(phi); 1721 1722 C->set_has_split_ifs(true); // Has chance for split-if optimization 1723 } 1724 set_result(n); 1725 return true; 1726 } 1727 1728 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1729 //------------------- 1730 //result=(result.isNaN())? funcAddr():result; 1731 // Check: If isNaN() by checking result!=result? then either trap 1732 // or go to runtime 1733 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1734 // Build the boolean node 1735 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1736 1737 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1738 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1739 // The pow or exp intrinsic returned a NaN, which requires a call 1740 // to the runtime. Recompile with the runtime call. 1741 uncommon_trap(Deoptimization::Reason_intrinsic, 1742 Deoptimization::Action_make_not_entrant); 1743 } 1744 return result; 1745 } else { 1746 // If this inlining ever returned NaN in the past, we compile a call 1747 // to the runtime to properly handle corner cases 1748 1749 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1750 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1751 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1752 1753 if (!if_slow->is_top()) { 1754 RegionNode* result_region = new RegionNode(3); 1755 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1756 1757 result_region->init_req(1, if_fast); 1758 result_val->init_req(1, result); 1759 1760 set_control(if_slow); 1761 1762 const TypePtr* no_memory_effects = NULL; 1763 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1764 no_memory_effects, 1765 x, top(), y, y ? top() : NULL); 1766 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1767 #ifdef ASSERT 1768 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1769 assert(value_top == top(), "second value must be top"); 1770 #endif 1771 1772 result_region->init_req(2, control()); 1773 result_val->init_req(2, value); 1774 set_control(_gvn.transform(result_region)); 1775 return _gvn.transform(result_val); 1776 } else { 1777 return result; 1778 } 1779 } 1780 } 1781 1782 //------------------------------inline_exp------------------------------------- 1783 // Inline exp instructions, if possible. The Intel hardware only misses 1784 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1785 bool LibraryCallKit::inline_exp() { 1786 Node* arg = round_double_node(argument(0)); 1787 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1788 1789 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1790 set_result(n); 1791 1792 C->set_has_split_ifs(true); // Has chance for split-if optimization 1793 return true; 1794 } 1795 1796 //------------------------------inline_pow------------------------------------- 1797 // Inline power instructions, if possible. 1798 bool LibraryCallKit::inline_pow() { 1799 // Pseudocode for pow 1800 // if (y == 2) { 1801 // return x * x; 1802 // } else { 1803 // if (x <= 0.0) { 1804 // long longy = (long)y; 1805 // if ((double)longy == y) { // if y is long 1806 // if (y + 1 == y) longy = 0; // huge number: even 1807 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1808 // } else { 1809 // result = NaN; 1810 // } 1811 // } else { 1812 // result = DPow(x,y); 1813 // } 1814 // if (result != result)? { 1815 // result = uncommon_trap() or runtime_call(); 1816 // } 1817 // return result; 1818 // } 1819 1820 Node* x = round_double_node(argument(0)); 1821 Node* y = round_double_node(argument(2)); 1822 1823 Node* result = NULL; 1824 1825 Node* const_two_node = makecon(TypeD::make(2.0)); 1826 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1827 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1828 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1829 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1830 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1831 1832 RegionNode* region_node = new RegionNode(3); 1833 region_node->init_req(1, if_true); 1834 1835 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1836 // special case for x^y where y == 2, we can convert it to x * x 1837 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1838 1839 // set control to if_false since we will now process the false branch 1840 set_control(if_false); 1841 1842 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1843 // Short form: skip the fancy tests and just check for NaN result. 1844 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1845 } else { 1846 // If this inlining ever returned NaN in the past, include all 1847 // checks + call to the runtime. 1848 1849 // Set the merge point for If node with condition of (x <= 0.0) 1850 // There are four possible paths to region node and phi node 1851 RegionNode *r = new RegionNode(4); 1852 Node *phi = new PhiNode(r, Type::DOUBLE); 1853 1854 // Build the first if node: if (x <= 0.0) 1855 // Node for 0 constant 1856 Node *zeronode = makecon(TypeD::ZERO); 1857 // Check x:0 1858 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1859 // Check: If (x<=0) then go complex path 1860 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1861 // Branch either way 1862 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1863 // Fast path taken; set region slot 3 1864 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1865 r->init_req(3,fast_taken); // Capture fast-control 1866 1867 // Fast path not-taken, i.e. slow path 1868 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1869 1870 // Set fast path result 1871 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1872 phi->init_req(3, fast_result); 1873 1874 // Complex path 1875 // Build the second if node (if y is long) 1876 // Node for (long)y 1877 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1878 // Node for (double)((long) y) 1879 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1880 // Check (double)((long) y) : y 1881 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1882 // Check if (y isn't long) then go to slow path 1883 1884 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1885 // Branch either way 1886 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1887 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1888 1889 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1890 1891 // Calculate DPow(abs(x), y)*(1 & (long)y) 1892 // Node for constant 1 1893 Node *conone = longcon(1); 1894 // 1& (long)y 1895 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1896 1897 // A huge number is always even. Detect a huge number by checking 1898 // if y + 1 == y and set integer to be tested for parity to 0. 1899 // Required for corner case: 1900 // (long)9.223372036854776E18 = max_jlong 1901 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1902 // max_jlong is odd but 9.223372036854776E18 is even 1903 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1904 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1905 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1906 Node* correctedsign = NULL; 1907 if (ConditionalMoveLimit != 0) { 1908 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1909 } else { 1910 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1911 RegionNode *r = new RegionNode(3); 1912 Node *phi = new PhiNode(r, TypeLong::LONG); 1913 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1914 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1915 phi->init_req(1, signnode); 1916 phi->init_req(2, longcon(0)); 1917 correctedsign = _gvn.transform(phi); 1918 ylong_path = _gvn.transform(r); 1919 record_for_igvn(r); 1920 } 1921 1922 // zero node 1923 Node *conzero = longcon(0); 1924 // Check (1&(long)y)==0? 1925 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1926 // Check if (1&(long)y)!=0?, if so the result is negative 1927 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1928 // abs(x) 1929 Node *absx=_gvn.transform(new AbsDNode(x)); 1930 // abs(x)^y 1931 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1932 // -abs(x)^y 1933 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1934 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1935 Node *signresult = NULL; 1936 if (ConditionalMoveLimit != 0) { 1937 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1938 } else { 1939 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1940 RegionNode *r = new RegionNode(3); 1941 Node *phi = new PhiNode(r, Type::DOUBLE); 1942 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1943 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1944 phi->init_req(1, absxpowy); 1945 phi->init_req(2, negabsxpowy); 1946 signresult = _gvn.transform(phi); 1947 ylong_path = _gvn.transform(r); 1948 record_for_igvn(r); 1949 } 1950 // Set complex path fast result 1951 r->init_req(2, ylong_path); 1952 phi->init_req(2, signresult); 1953 1954 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1955 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1956 r->init_req(1,slow_path); 1957 phi->init_req(1,slow_result); 1958 1959 // Post merge 1960 set_control(_gvn.transform(r)); 1961 record_for_igvn(r); 1962 result = _gvn.transform(phi); 1963 } 1964 1965 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1966 1967 // control from finish_pow_exp is now input to the region node 1968 region_node->set_req(2, control()); 1969 // the result from finish_pow_exp is now input to the phi node 1970 phi_node->init_req(2, result); 1971 set_control(_gvn.transform(region_node)); 1972 record_for_igvn(region_node); 1973 set_result(_gvn.transform(phi_node)); 1974 1975 C->set_has_split_ifs(true); // Has chance for split-if optimization 1976 return true; 1977 } 1978 1979 //------------------------------runtime_math----------------------------- 1980 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1981 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1982 "must be (DD)D or (D)D type"); 1983 1984 // Inputs 1985 Node* a = round_double_node(argument(0)); 1986 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1987 1988 const TypePtr* no_memory_effects = NULL; 1989 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1990 no_memory_effects, 1991 a, top(), b, b ? top() : NULL); 1992 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1993 #ifdef ASSERT 1994 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1995 assert(value_top == top(), "second value must be top"); 1996 #endif 1997 1998 set_result(value); 1999 return true; 2000 } 2001 2002 //------------------------------inline_math_native----------------------------- 2003 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 2004 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 2005 switch (id) { 2006 // These intrinsics are not properly supported on all hardware 2007 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 2008 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 2009 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 2010 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2011 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2012 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2013 2014 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2015 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2016 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2017 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2018 2019 // These intrinsics are supported on all hardware 2020 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2021 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2022 2023 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2024 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2025 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2026 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2027 #undef FN_PTR 2028 2029 // These intrinsics are not yet correctly implemented 2030 case vmIntrinsics::_datan2: 2031 return false; 2032 2033 default: 2034 fatal_unexpected_iid(id); 2035 return false; 2036 } 2037 } 2038 2039 static bool is_simple_name(Node* n) { 2040 return (n->req() == 1 // constant 2041 || (n->is_Type() && n->as_Type()->type()->singleton()) 2042 || n->is_Proj() // parameter or return value 2043 || n->is_Phi() // local of some sort 2044 ); 2045 } 2046 2047 //----------------------------inline_min_max----------------------------------- 2048 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2049 set_result(generate_min_max(id, argument(0), argument(1))); 2050 return true; 2051 } 2052 2053 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2054 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2055 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2056 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2057 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2058 2059 { 2060 PreserveJVMState pjvms(this); 2061 PreserveReexecuteState preexecs(this); 2062 jvms()->set_should_reexecute(true); 2063 2064 set_control(slow_path); 2065 set_i_o(i_o()); 2066 2067 uncommon_trap(Deoptimization::Reason_intrinsic, 2068 Deoptimization::Action_none); 2069 } 2070 2071 set_control(fast_path); 2072 set_result(math); 2073 } 2074 2075 template <typename OverflowOp> 2076 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2077 typedef typename OverflowOp::MathOp MathOp; 2078 2079 MathOp* mathOp = new MathOp(arg1, arg2); 2080 Node* operation = _gvn.transform( mathOp ); 2081 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2082 inline_math_mathExact(operation, ofcheck); 2083 return true; 2084 } 2085 2086 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2087 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2088 } 2089 2090 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2091 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2092 } 2093 2094 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2095 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2096 } 2097 2098 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2099 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2100 } 2101 2102 bool LibraryCallKit::inline_math_negateExactI() { 2103 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2104 } 2105 2106 bool LibraryCallKit::inline_math_negateExactL() { 2107 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2108 } 2109 2110 bool LibraryCallKit::inline_math_multiplyExactI() { 2111 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2112 } 2113 2114 bool LibraryCallKit::inline_math_multiplyExactL() { 2115 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2116 } 2117 2118 Node* 2119 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2120 // These are the candidate return value: 2121 Node* xvalue = x0; 2122 Node* yvalue = y0; 2123 2124 if (xvalue == yvalue) { 2125 return xvalue; 2126 } 2127 2128 bool want_max = (id == vmIntrinsics::_max); 2129 2130 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2131 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2132 if (txvalue == NULL || tyvalue == NULL) return top(); 2133 // This is not really necessary, but it is consistent with a 2134 // hypothetical MaxINode::Value method: 2135 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2136 2137 // %%% This folding logic should (ideally) be in a different place. 2138 // Some should be inside IfNode, and there to be a more reliable 2139 // transformation of ?: style patterns into cmoves. We also want 2140 // more powerful optimizations around cmove and min/max. 2141 2142 // Try to find a dominating comparison of these guys. 2143 // It can simplify the index computation for Arrays.copyOf 2144 // and similar uses of System.arraycopy. 2145 // First, compute the normalized version of CmpI(x, y). 2146 int cmp_op = Op_CmpI; 2147 Node* xkey = xvalue; 2148 Node* ykey = yvalue; 2149 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2150 if (ideal_cmpxy->is_Cmp()) { 2151 // E.g., if we have CmpI(length - offset, count), 2152 // it might idealize to CmpI(length, count + offset) 2153 cmp_op = ideal_cmpxy->Opcode(); 2154 xkey = ideal_cmpxy->in(1); 2155 ykey = ideal_cmpxy->in(2); 2156 } 2157 2158 // Start by locating any relevant comparisons. 2159 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2160 Node* cmpxy = NULL; 2161 Node* cmpyx = NULL; 2162 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2163 Node* cmp = start_from->fast_out(k); 2164 if (cmp->outcnt() > 0 && // must have prior uses 2165 cmp->in(0) == NULL && // must be context-independent 2166 cmp->Opcode() == cmp_op) { // right kind of compare 2167 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2168 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2169 } 2170 } 2171 2172 const int NCMPS = 2; 2173 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2174 int cmpn; 2175 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2176 if (cmps[cmpn] != NULL) break; // find a result 2177 } 2178 if (cmpn < NCMPS) { 2179 // Look for a dominating test that tells us the min and max. 2180 int depth = 0; // Limit search depth for speed 2181 Node* dom = control(); 2182 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2183 if (++depth >= 100) break; 2184 Node* ifproj = dom; 2185 if (!ifproj->is_Proj()) continue; 2186 Node* iff = ifproj->in(0); 2187 if (!iff->is_If()) continue; 2188 Node* bol = iff->in(1); 2189 if (!bol->is_Bool()) continue; 2190 Node* cmp = bol->in(1); 2191 if (cmp == NULL) continue; 2192 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2193 if (cmps[cmpn] == cmp) break; 2194 if (cmpn == NCMPS) continue; 2195 BoolTest::mask btest = bol->as_Bool()->_test._test; 2196 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2197 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2198 // At this point, we know that 'x btest y' is true. 2199 switch (btest) { 2200 case BoolTest::eq: 2201 // They are proven equal, so we can collapse the min/max. 2202 // Either value is the answer. Choose the simpler. 2203 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2204 return yvalue; 2205 return xvalue; 2206 case BoolTest::lt: // x < y 2207 case BoolTest::le: // x <= y 2208 return (want_max ? yvalue : xvalue); 2209 case BoolTest::gt: // x > y 2210 case BoolTest::ge: // x >= y 2211 return (want_max ? xvalue : yvalue); 2212 } 2213 } 2214 } 2215 2216 // We failed to find a dominating test. 2217 // Let's pick a test that might GVN with prior tests. 2218 Node* best_bol = NULL; 2219 BoolTest::mask best_btest = BoolTest::illegal; 2220 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2221 Node* cmp = cmps[cmpn]; 2222 if (cmp == NULL) continue; 2223 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2224 Node* bol = cmp->fast_out(j); 2225 if (!bol->is_Bool()) continue; 2226 BoolTest::mask btest = bol->as_Bool()->_test._test; 2227 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2228 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2229 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2230 best_bol = bol->as_Bool(); 2231 best_btest = btest; 2232 } 2233 } 2234 } 2235 2236 Node* answer_if_true = NULL; 2237 Node* answer_if_false = NULL; 2238 switch (best_btest) { 2239 default: 2240 if (cmpxy == NULL) 2241 cmpxy = ideal_cmpxy; 2242 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2243 // and fall through: 2244 case BoolTest::lt: // x < y 2245 case BoolTest::le: // x <= y 2246 answer_if_true = (want_max ? yvalue : xvalue); 2247 answer_if_false = (want_max ? xvalue : yvalue); 2248 break; 2249 case BoolTest::gt: // x > y 2250 case BoolTest::ge: // x >= y 2251 answer_if_true = (want_max ? xvalue : yvalue); 2252 answer_if_false = (want_max ? yvalue : xvalue); 2253 break; 2254 } 2255 2256 jint hi, lo; 2257 if (want_max) { 2258 // We can sharpen the minimum. 2259 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2260 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2261 } else { 2262 // We can sharpen the maximum. 2263 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2264 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2265 } 2266 2267 // Use a flow-free graph structure, to avoid creating excess control edges 2268 // which could hinder other optimizations. 2269 // Since Math.min/max is often used with arraycopy, we want 2270 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2271 Node* cmov = CMoveNode::make(NULL, best_bol, 2272 answer_if_false, answer_if_true, 2273 TypeInt::make(lo, hi, widen)); 2274 2275 return _gvn.transform(cmov); 2276 2277 /* 2278 // This is not as desirable as it may seem, since Min and Max 2279 // nodes do not have a full set of optimizations. 2280 // And they would interfere, anyway, with 'if' optimizations 2281 // and with CMoveI canonical forms. 2282 switch (id) { 2283 case vmIntrinsics::_min: 2284 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2285 case vmIntrinsics::_max: 2286 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2287 default: 2288 ShouldNotReachHere(); 2289 } 2290 */ 2291 } 2292 2293 inline int 2294 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2295 const TypePtr* base_type = TypePtr::NULL_PTR; 2296 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2297 if (base_type == NULL) { 2298 // Unknown type. 2299 return Type::AnyPtr; 2300 } else if (base_type == TypePtr::NULL_PTR) { 2301 // Since this is a NULL+long form, we have to switch to a rawptr. 2302 base = _gvn.transform(new CastX2PNode(offset)); 2303 offset = MakeConX(0); 2304 return Type::RawPtr; 2305 } else if (base_type->base() == Type::RawPtr) { 2306 return Type::RawPtr; 2307 } else if (base_type->isa_oopptr()) { 2308 // Base is never null => always a heap address. 2309 if (base_type->ptr() == TypePtr::NotNull) { 2310 return Type::OopPtr; 2311 } 2312 // Offset is small => always a heap address. 2313 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2314 if (offset_type != NULL && 2315 base_type->offset() == 0 && // (should always be?) 2316 offset_type->_lo >= 0 && 2317 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2318 return Type::OopPtr; 2319 } 2320 // Otherwise, it might either be oop+off or NULL+addr. 2321 return Type::AnyPtr; 2322 } else { 2323 // No information: 2324 return Type::AnyPtr; 2325 } 2326 } 2327 2328 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2329 int kind = classify_unsafe_addr(base, offset); 2330 if (kind == Type::RawPtr) { 2331 return basic_plus_adr(top(), base, offset); 2332 } else { 2333 return basic_plus_adr(base, offset); 2334 } 2335 } 2336 2337 //--------------------------inline_number_methods----------------------------- 2338 // inline int Integer.numberOfLeadingZeros(int) 2339 // inline int Long.numberOfLeadingZeros(long) 2340 // 2341 // inline int Integer.numberOfTrailingZeros(int) 2342 // inline int Long.numberOfTrailingZeros(long) 2343 // 2344 // inline int Integer.bitCount(int) 2345 // inline int Long.bitCount(long) 2346 // 2347 // inline char Character.reverseBytes(char) 2348 // inline short Short.reverseBytes(short) 2349 // inline int Integer.reverseBytes(int) 2350 // inline long Long.reverseBytes(long) 2351 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2352 Node* arg = argument(0); 2353 Node* n; 2354 switch (id) { 2355 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2356 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2357 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2358 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2359 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2360 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2361 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2362 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2363 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2364 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2365 default: fatal_unexpected_iid(id); break; 2366 } 2367 set_result(_gvn.transform(n)); 2368 return true; 2369 } 2370 2371 //----------------------------inline_unsafe_access---------------------------- 2372 2373 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2374 2375 // Helper that guards and inserts a pre-barrier. 2376 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2377 Node* pre_val, bool need_mem_bar) { 2378 // We could be accessing the referent field of a reference object. If so, when G1 2379 // is enabled, we need to log the value in the referent field in an SATB buffer. 2380 // This routine performs some compile time filters and generates suitable 2381 // runtime filters that guard the pre-barrier code. 2382 // Also add memory barrier for non volatile load from the referent field 2383 // to prevent commoning of loads across safepoint. 2384 if (!UseG1GC && !need_mem_bar) 2385 return; 2386 2387 // Some compile time checks. 2388 2389 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2390 const TypeX* otype = offset->find_intptr_t_type(); 2391 if (otype != NULL && otype->is_con() && 2392 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2393 // Constant offset but not the reference_offset so just return 2394 return; 2395 } 2396 2397 // We only need to generate the runtime guards for instances. 2398 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2399 if (btype != NULL) { 2400 if (btype->isa_aryptr()) { 2401 // Array type so nothing to do 2402 return; 2403 } 2404 2405 const TypeInstPtr* itype = btype->isa_instptr(); 2406 if (itype != NULL) { 2407 // Can the klass of base_oop be statically determined to be 2408 // _not_ a sub-class of Reference and _not_ Object? 2409 ciKlass* klass = itype->klass(); 2410 if ( klass->is_loaded() && 2411 !klass->is_subtype_of(env()->Reference_klass()) && 2412 !env()->Object_klass()->is_subtype_of(klass)) { 2413 return; 2414 } 2415 } 2416 } 2417 2418 // The compile time filters did not reject base_oop/offset so 2419 // we need to generate the following runtime filters 2420 // 2421 // if (offset == java_lang_ref_Reference::_reference_offset) { 2422 // if (instance_of(base, java.lang.ref.Reference)) { 2423 // pre_barrier(_, pre_val, ...); 2424 // } 2425 // } 2426 2427 float likely = PROB_LIKELY( 0.999); 2428 float unlikely = PROB_UNLIKELY(0.999); 2429 2430 IdealKit ideal(this); 2431 #define __ ideal. 2432 2433 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2434 2435 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2436 // Update graphKit memory and control from IdealKit. 2437 sync_kit(ideal); 2438 2439 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2440 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2441 2442 // Update IdealKit memory and control from graphKit. 2443 __ sync_kit(this); 2444 2445 Node* one = __ ConI(1); 2446 // is_instof == 0 if base_oop == NULL 2447 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2448 2449 // Update graphKit from IdeakKit. 2450 sync_kit(ideal); 2451 2452 // Use the pre-barrier to record the value in the referent field 2453 pre_barrier(false /* do_load */, 2454 __ ctrl(), 2455 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2456 pre_val /* pre_val */, 2457 T_OBJECT); 2458 if (need_mem_bar) { 2459 // Add memory barrier to prevent commoning reads from this field 2460 // across safepoint since GC can change its value. 2461 insert_mem_bar(Op_MemBarCPUOrder); 2462 } 2463 // Update IdealKit from graphKit. 2464 __ sync_kit(this); 2465 2466 } __ end_if(); // _ref_type != ref_none 2467 } __ end_if(); // offset == referent_offset 2468 2469 // Final sync IdealKit and GraphKit. 2470 final_sync(ideal); 2471 #undef __ 2472 } 2473 2474 2475 // Interpret Unsafe.fieldOffset cookies correctly: 2476 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2477 2478 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2479 // Attempt to infer a sharper value type from the offset and base type. 2480 ciKlass* sharpened_klass = NULL; 2481 2482 // See if it is an instance field, with an object type. 2483 if (alias_type->field() != NULL) { 2484 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2485 if (alias_type->field()->type()->is_klass()) { 2486 sharpened_klass = alias_type->field()->type()->as_klass(); 2487 } 2488 } 2489 2490 // See if it is a narrow oop array. 2491 if (adr_type->isa_aryptr()) { 2492 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2493 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2494 if (elem_type != NULL) { 2495 sharpened_klass = elem_type->klass(); 2496 } 2497 } 2498 } 2499 2500 // The sharpened class might be unloaded if there is no class loader 2501 // contraint in place. 2502 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2503 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2504 2505 #ifndef PRODUCT 2506 if (C->print_intrinsics() || C->print_inlining()) { 2507 tty->print(" from base type: "); adr_type->dump(); 2508 tty->print(" sharpened value: "); tjp->dump(); 2509 } 2510 #endif 2511 // Sharpen the value type. 2512 return tjp; 2513 } 2514 return NULL; 2515 } 2516 2517 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2518 if (callee()->is_static()) return false; // caller must have the capability! 2519 2520 #ifndef PRODUCT 2521 { 2522 ResourceMark rm; 2523 // Check the signatures. 2524 ciSignature* sig = callee()->signature(); 2525 #ifdef ASSERT 2526 if (!is_store) { 2527 // Object getObject(Object base, int/long offset), etc. 2528 BasicType rtype = sig->return_type()->basic_type(); 2529 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2530 rtype = T_ADDRESS; // it is really a C void* 2531 assert(rtype == type, "getter must return the expected value"); 2532 if (!is_native_ptr) { 2533 assert(sig->count() == 2, "oop getter has 2 arguments"); 2534 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2535 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2536 } else { 2537 assert(sig->count() == 1, "native getter has 1 argument"); 2538 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2539 } 2540 } else { 2541 // void putObject(Object base, int/long offset, Object x), etc. 2542 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2543 if (!is_native_ptr) { 2544 assert(sig->count() == 3, "oop putter has 3 arguments"); 2545 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2546 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2547 } else { 2548 assert(sig->count() == 2, "native putter has 2 arguments"); 2549 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2550 } 2551 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2552 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2553 vtype = T_ADDRESS; // it is really a C void* 2554 assert(vtype == type, "putter must accept the expected value"); 2555 } 2556 #endif // ASSERT 2557 } 2558 #endif //PRODUCT 2559 2560 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2561 2562 Node* receiver = argument(0); // type: oop 2563 2564 // Build address expression. See the code in inline_unsafe_prefetch. 2565 Node* adr; 2566 Node* heap_base_oop = top(); 2567 Node* offset = top(); 2568 Node* val; 2569 2570 if (!is_native_ptr) { 2571 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2572 Node* base = argument(1); // type: oop 2573 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2574 offset = argument(2); // type: long 2575 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2576 // to be plain byte offsets, which are also the same as those accepted 2577 // by oopDesc::field_base. 2578 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2579 "fieldOffset must be byte-scaled"); 2580 // 32-bit machines ignore the high half! 2581 offset = ConvL2X(offset); 2582 adr = make_unsafe_address(base, offset); 2583 heap_base_oop = base; 2584 val = is_store ? argument(4) : NULL; 2585 } else { 2586 Node* ptr = argument(1); // type: long 2587 ptr = ConvL2X(ptr); // adjust Java long to machine word 2588 adr = make_unsafe_address(NULL, ptr); 2589 val = is_store ? argument(3) : NULL; 2590 } 2591 2592 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2593 2594 // First guess at the value type. 2595 const Type *value_type = Type::get_const_basic_type(type); 2596 2597 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2598 // there was not enough information to nail it down. 2599 Compile::AliasType* alias_type = C->alias_type(adr_type); 2600 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2601 2602 // We will need memory barriers unless we can determine a unique 2603 // alias category for this reference. (Note: If for some reason 2604 // the barriers get omitted and the unsafe reference begins to "pollute" 2605 // the alias analysis of the rest of the graph, either Compile::can_alias 2606 // or Compile::must_alias will throw a diagnostic assert.) 2607 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2608 2609 // If we are reading the value of the referent field of a Reference 2610 // object (either by using Unsafe directly or through reflection) 2611 // then, if G1 is enabled, we need to record the referent in an 2612 // SATB log buffer using the pre-barrier mechanism. 2613 // Also we need to add memory barrier to prevent commoning reads 2614 // from this field across safepoint since GC can change its value. 2615 bool need_read_barrier = !is_native_ptr && !is_store && 2616 offset != top() && heap_base_oop != top(); 2617 2618 if (!is_store && type == T_OBJECT) { 2619 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2620 if (tjp != NULL) { 2621 value_type = tjp; 2622 } 2623 } 2624 2625 receiver = null_check(receiver); 2626 if (stopped()) { 2627 return true; 2628 } 2629 // Heap pointers get a null-check from the interpreter, 2630 // as a courtesy. However, this is not guaranteed by Unsafe, 2631 // and it is not possible to fully distinguish unintended nulls 2632 // from intended ones in this API. 2633 2634 if (is_volatile) { 2635 // We need to emit leading and trailing CPU membars (see below) in 2636 // addition to memory membars when is_volatile. This is a little 2637 // too strong, but avoids the need to insert per-alias-type 2638 // volatile membars (for stores; compare Parse::do_put_xxx), which 2639 // we cannot do effectively here because we probably only have a 2640 // rough approximation of type. 2641 need_mem_bar = true; 2642 // For Stores, place a memory ordering barrier now. 2643 if (is_store) { 2644 insert_mem_bar(Op_MemBarRelease); 2645 } else { 2646 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2647 insert_mem_bar(Op_MemBarVolatile); 2648 } 2649 } 2650 } 2651 2652 // Memory barrier to prevent normal and 'unsafe' accesses from 2653 // bypassing each other. Happens after null checks, so the 2654 // exception paths do not take memory state from the memory barrier, 2655 // so there's no problems making a strong assert about mixing users 2656 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2657 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2658 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2659 2660 if (!is_store) { 2661 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile); 2662 // load value 2663 switch (type) { 2664 case T_BOOLEAN: 2665 case T_CHAR: 2666 case T_BYTE: 2667 case T_SHORT: 2668 case T_INT: 2669 case T_LONG: 2670 case T_FLOAT: 2671 case T_DOUBLE: 2672 break; 2673 case T_OBJECT: 2674 if (need_read_barrier) { 2675 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2676 } 2677 break; 2678 case T_ADDRESS: 2679 // Cast to an int type. 2680 p = _gvn.transform(new CastP2XNode(NULL, p)); 2681 p = ConvX2UL(p); 2682 break; 2683 default: 2684 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2685 break; 2686 } 2687 // The load node has the control of the preceding MemBarCPUOrder. All 2688 // following nodes will have the control of the MemBarCPUOrder inserted at 2689 // the end of this method. So, pushing the load onto the stack at a later 2690 // point is fine. 2691 set_result(p); 2692 } else { 2693 // place effect of store into memory 2694 switch (type) { 2695 case T_DOUBLE: 2696 val = dstore_rounding(val); 2697 break; 2698 case T_ADDRESS: 2699 // Repackage the long as a pointer. 2700 val = ConvL2X(val); 2701 val = _gvn.transform(new CastX2PNode(val)); 2702 break; 2703 } 2704 2705 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2706 if (type != T_OBJECT ) { 2707 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2708 } else { 2709 // Possibly an oop being stored to Java heap or native memory 2710 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2711 // oop to Java heap. 2712 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2713 } else { 2714 // We can't tell at compile time if we are storing in the Java heap or outside 2715 // of it. So we need to emit code to conditionally do the proper type of 2716 // store. 2717 2718 IdealKit ideal(this); 2719 #define __ ideal. 2720 // QQQ who knows what probability is here?? 2721 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2722 // Sync IdealKit and graphKit. 2723 sync_kit(ideal); 2724 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2725 // Update IdealKit memory. 2726 __ sync_kit(this); 2727 } __ else_(); { 2728 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2729 } __ end_if(); 2730 // Final sync IdealKit and GraphKit. 2731 final_sync(ideal); 2732 #undef __ 2733 } 2734 } 2735 } 2736 2737 if (is_volatile) { 2738 if (!is_store) { 2739 insert_mem_bar(Op_MemBarAcquire); 2740 } else { 2741 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2742 insert_mem_bar(Op_MemBarVolatile); 2743 } 2744 } 2745 } 2746 2747 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2748 2749 return true; 2750 } 2751 2752 //----------------------------inline_unsafe_prefetch---------------------------- 2753 2754 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2755 #ifndef PRODUCT 2756 { 2757 ResourceMark rm; 2758 // Check the signatures. 2759 ciSignature* sig = callee()->signature(); 2760 #ifdef ASSERT 2761 // Object getObject(Object base, int/long offset), etc. 2762 BasicType rtype = sig->return_type()->basic_type(); 2763 if (!is_native_ptr) { 2764 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2765 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2766 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2767 } else { 2768 assert(sig->count() == 1, "native prefetch has 1 argument"); 2769 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2770 } 2771 #endif // ASSERT 2772 } 2773 #endif // !PRODUCT 2774 2775 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2776 2777 const int idx = is_static ? 0 : 1; 2778 if (!is_static) { 2779 null_check_receiver(); 2780 if (stopped()) { 2781 return true; 2782 } 2783 } 2784 2785 // Build address expression. See the code in inline_unsafe_access. 2786 Node *adr; 2787 if (!is_native_ptr) { 2788 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2789 Node* base = argument(idx + 0); // type: oop 2790 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2791 Node* offset = argument(idx + 1); // type: long 2792 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2793 // to be plain byte offsets, which are also the same as those accepted 2794 // by oopDesc::field_base. 2795 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2796 "fieldOffset must be byte-scaled"); 2797 // 32-bit machines ignore the high half! 2798 offset = ConvL2X(offset); 2799 adr = make_unsafe_address(base, offset); 2800 } else { 2801 Node* ptr = argument(idx + 0); // type: long 2802 ptr = ConvL2X(ptr); // adjust Java long to machine word 2803 adr = make_unsafe_address(NULL, ptr); 2804 } 2805 2806 // Generate the read or write prefetch 2807 Node *prefetch; 2808 if (is_store) { 2809 prefetch = new PrefetchWriteNode(i_o(), adr); 2810 } else { 2811 prefetch = new PrefetchReadNode(i_o(), adr); 2812 } 2813 prefetch->init_req(0, control()); 2814 set_i_o(_gvn.transform(prefetch)); 2815 2816 return true; 2817 } 2818 2819 //----------------------------inline_unsafe_load_store---------------------------- 2820 // This method serves a couple of different customers (depending on LoadStoreKind): 2821 // 2822 // LS_cmpxchg: 2823 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2824 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2825 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2826 // 2827 // LS_xadd: 2828 // public int getAndAddInt( Object o, long offset, int delta) 2829 // public long getAndAddLong(Object o, long offset, long delta) 2830 // 2831 // LS_xchg: 2832 // int getAndSet(Object o, long offset, int newValue) 2833 // long getAndSet(Object o, long offset, long newValue) 2834 // Object getAndSet(Object o, long offset, Object newValue) 2835 // 2836 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2837 // This basic scheme here is the same as inline_unsafe_access, but 2838 // differs in enough details that combining them would make the code 2839 // overly confusing. (This is a true fact! I originally combined 2840 // them, but even I was confused by it!) As much code/comments as 2841 // possible are retained from inline_unsafe_access though to make 2842 // the correspondences clearer. - dl 2843 2844 if (callee()->is_static()) return false; // caller must have the capability! 2845 2846 #ifndef PRODUCT 2847 BasicType rtype; 2848 { 2849 ResourceMark rm; 2850 // Check the signatures. 2851 ciSignature* sig = callee()->signature(); 2852 rtype = sig->return_type()->basic_type(); 2853 if (kind == LS_xadd || kind == LS_xchg) { 2854 // Check the signatures. 2855 #ifdef ASSERT 2856 assert(rtype == type, "get and set must return the expected type"); 2857 assert(sig->count() == 3, "get and set has 3 arguments"); 2858 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2859 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2860 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2861 #endif // ASSERT 2862 } else if (kind == LS_cmpxchg) { 2863 // Check the signatures. 2864 #ifdef ASSERT 2865 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2866 assert(sig->count() == 4, "CAS has 4 arguments"); 2867 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2868 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2869 #endif // ASSERT 2870 } else { 2871 ShouldNotReachHere(); 2872 } 2873 } 2874 #endif //PRODUCT 2875 2876 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2877 2878 // Get arguments: 2879 Node* receiver = NULL; 2880 Node* base = NULL; 2881 Node* offset = NULL; 2882 Node* oldval = NULL; 2883 Node* newval = NULL; 2884 if (kind == LS_cmpxchg) { 2885 const bool two_slot_type = type2size[type] == 2; 2886 receiver = argument(0); // type: oop 2887 base = argument(1); // type: oop 2888 offset = argument(2); // type: long 2889 oldval = argument(4); // type: oop, int, or long 2890 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2891 } else if (kind == LS_xadd || kind == LS_xchg){ 2892 receiver = argument(0); // type: oop 2893 base = argument(1); // type: oop 2894 offset = argument(2); // type: long 2895 oldval = NULL; 2896 newval = argument(4); // type: oop, int, or long 2897 } 2898 2899 // Null check receiver. 2900 receiver = null_check(receiver); 2901 if (stopped()) { 2902 return true; 2903 } 2904 2905 // Build field offset expression. 2906 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2907 // to be plain byte offsets, which are also the same as those accepted 2908 // by oopDesc::field_base. 2909 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2910 // 32-bit machines ignore the high half of long offsets 2911 offset = ConvL2X(offset); 2912 Node* adr = make_unsafe_address(base, offset); 2913 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2914 2915 // For CAS, unlike inline_unsafe_access, there seems no point in 2916 // trying to refine types. Just use the coarse types here. 2917 const Type *value_type = Type::get_const_basic_type(type); 2918 Compile::AliasType* alias_type = C->alias_type(adr_type); 2919 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2920 2921 if (kind == LS_xchg && type == T_OBJECT) { 2922 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2923 if (tjp != NULL) { 2924 value_type = tjp; 2925 } 2926 } 2927 2928 int alias_idx = C->get_alias_index(adr_type); 2929 2930 // Memory-model-wise, a LoadStore acts like a little synchronized 2931 // block, so needs barriers on each side. These don't translate 2932 // into actual barriers on most machines, but we still need rest of 2933 // compiler to respect ordering. 2934 2935 insert_mem_bar(Op_MemBarRelease); 2936 insert_mem_bar(Op_MemBarCPUOrder); 2937 2938 // 4984716: MemBars must be inserted before this 2939 // memory node in order to avoid a false 2940 // dependency which will confuse the scheduler. 2941 Node *mem = memory(alias_idx); 2942 2943 // For now, we handle only those cases that actually exist: ints, 2944 // longs, and Object. Adding others should be straightforward. 2945 Node* load_store; 2946 switch(type) { 2947 case T_INT: 2948 if (kind == LS_xadd) { 2949 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2950 } else if (kind == LS_xchg) { 2951 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2952 } else if (kind == LS_cmpxchg) { 2953 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2954 } else { 2955 ShouldNotReachHere(); 2956 } 2957 break; 2958 case T_LONG: 2959 if (kind == LS_xadd) { 2960 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2961 } else if (kind == LS_xchg) { 2962 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2963 } else if (kind == LS_cmpxchg) { 2964 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2965 } else { 2966 ShouldNotReachHere(); 2967 } 2968 break; 2969 case T_OBJECT: 2970 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2971 // could be delayed during Parse (for example, in adjust_map_after_if()). 2972 // Execute transformation here to avoid barrier generation in such case. 2973 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2974 newval = _gvn.makecon(TypePtr::NULL_PTR); 2975 2976 // Reference stores need a store barrier. 2977 if (kind == LS_xchg) { 2978 // If pre-barrier must execute before the oop store, old value will require do_load here. 2979 if (!can_move_pre_barrier()) { 2980 pre_barrier(true /* do_load*/, 2981 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2982 NULL /* pre_val*/, 2983 T_OBJECT); 2984 } // Else move pre_barrier to use load_store value, see below. 2985 } else if (kind == LS_cmpxchg) { 2986 // Same as for newval above: 2987 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2988 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2989 } 2990 // The only known value which might get overwritten is oldval. 2991 pre_barrier(false /* do_load */, 2992 control(), NULL, NULL, max_juint, NULL, NULL, 2993 oldval /* pre_val */, 2994 T_OBJECT); 2995 } else { 2996 ShouldNotReachHere(); 2997 } 2998 2999 #ifdef _LP64 3000 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3001 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 3002 if (kind == LS_xchg) { 3003 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 3004 newval_enc, adr_type, value_type->make_narrowoop())); 3005 } else { 3006 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 3007 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 3008 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 3009 newval_enc, oldval_enc)); 3010 } 3011 } else 3012 #endif 3013 { 3014 if (kind == LS_xchg) { 3015 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 3016 } else { 3017 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 3018 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 3019 } 3020 } 3021 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 3022 break; 3023 default: 3024 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 3025 break; 3026 } 3027 3028 // SCMemProjNodes represent the memory state of a LoadStore. Their 3029 // main role is to prevent LoadStore nodes from being optimized away 3030 // when their results aren't used. 3031 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 3032 set_memory(proj, alias_idx); 3033 3034 if (type == T_OBJECT && kind == LS_xchg) { 3035 #ifdef _LP64 3036 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3037 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 3038 } 3039 #endif 3040 if (can_move_pre_barrier()) { 3041 // Don't need to load pre_val. The old value is returned by load_store. 3042 // The pre_barrier can execute after the xchg as long as no safepoint 3043 // gets inserted between them. 3044 pre_barrier(false /* do_load */, 3045 control(), NULL, NULL, max_juint, NULL, NULL, 3046 load_store /* pre_val */, 3047 T_OBJECT); 3048 } 3049 } 3050 3051 // Add the trailing membar surrounding the access 3052 insert_mem_bar(Op_MemBarCPUOrder); 3053 insert_mem_bar(Op_MemBarAcquire); 3054 3055 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3056 set_result(load_store); 3057 return true; 3058 } 3059 3060 //----------------------------inline_unsafe_ordered_store---------------------- 3061 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3062 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3063 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3064 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3065 // This is another variant of inline_unsafe_access, differing in 3066 // that it always issues store-store ("release") barrier and ensures 3067 // store-atomicity (which only matters for "long"). 3068 3069 if (callee()->is_static()) return false; // caller must have the capability! 3070 3071 #ifndef PRODUCT 3072 { 3073 ResourceMark rm; 3074 // Check the signatures. 3075 ciSignature* sig = callee()->signature(); 3076 #ifdef ASSERT 3077 BasicType rtype = sig->return_type()->basic_type(); 3078 assert(rtype == T_VOID, "must return void"); 3079 assert(sig->count() == 3, "has 3 arguments"); 3080 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3081 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3082 #endif // ASSERT 3083 } 3084 #endif //PRODUCT 3085 3086 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3087 3088 // Get arguments: 3089 Node* receiver = argument(0); // type: oop 3090 Node* base = argument(1); // type: oop 3091 Node* offset = argument(2); // type: long 3092 Node* val = argument(4); // type: oop, int, or long 3093 3094 // Null check receiver. 3095 receiver = null_check(receiver); 3096 if (stopped()) { 3097 return true; 3098 } 3099 3100 // Build field offset expression. 3101 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3102 // 32-bit machines ignore the high half of long offsets 3103 offset = ConvL2X(offset); 3104 Node* adr = make_unsafe_address(base, offset); 3105 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3106 const Type *value_type = Type::get_const_basic_type(type); 3107 Compile::AliasType* alias_type = C->alias_type(adr_type); 3108 3109 insert_mem_bar(Op_MemBarRelease); 3110 insert_mem_bar(Op_MemBarCPUOrder); 3111 // Ensure that the store is atomic for longs: 3112 const bool require_atomic_access = true; 3113 Node* store; 3114 if (type == T_OBJECT) // reference stores need a store barrier. 3115 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3116 else { 3117 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3118 } 3119 insert_mem_bar(Op_MemBarCPUOrder); 3120 return true; 3121 } 3122 3123 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3124 // Regardless of form, don't allow previous ld/st to move down, 3125 // then issue acquire, release, or volatile mem_bar. 3126 insert_mem_bar(Op_MemBarCPUOrder); 3127 switch(id) { 3128 case vmIntrinsics::_loadFence: 3129 insert_mem_bar(Op_LoadFence); 3130 return true; 3131 case vmIntrinsics::_storeFence: 3132 insert_mem_bar(Op_StoreFence); 3133 return true; 3134 case vmIntrinsics::_fullFence: 3135 insert_mem_bar(Op_MemBarVolatile); 3136 return true; 3137 default: 3138 fatal_unexpected_iid(id); 3139 return false; 3140 } 3141 } 3142 3143 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3144 if (!kls->is_Con()) { 3145 return true; 3146 } 3147 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3148 if (klsptr == NULL) { 3149 return true; 3150 } 3151 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3152 // don't need a guard for a klass that is already initialized 3153 return !ik->is_initialized(); 3154 } 3155 3156 //----------------------------inline_unsafe_allocate--------------------------- 3157 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3158 bool LibraryCallKit::inline_unsafe_allocate() { 3159 if (callee()->is_static()) return false; // caller must have the capability! 3160 3161 null_check_receiver(); // null-check, then ignore 3162 Node* cls = null_check(argument(1)); 3163 if (stopped()) return true; 3164 3165 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3166 kls = null_check(kls); 3167 if (stopped()) return true; // argument was like int.class 3168 3169 Node* test = NULL; 3170 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3171 // Note: The argument might still be an illegal value like 3172 // Serializable.class or Object[].class. The runtime will handle it. 3173 // But we must make an explicit check for initialization. 3174 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3175 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3176 // can generate code to load it as unsigned byte. 3177 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3178 Node* bits = intcon(InstanceKlass::fully_initialized); 3179 test = _gvn.transform(new SubINode(inst, bits)); 3180 // The 'test' is non-zero if we need to take a slow path. 3181 } 3182 3183 Node* obj = new_instance(kls, test); 3184 set_result(obj); 3185 return true; 3186 } 3187 3188 #ifdef TRACE_HAVE_INTRINSICS 3189 /* 3190 * oop -> myklass 3191 * myklass->trace_id |= USED 3192 * return myklass->trace_id & ~0x3 3193 */ 3194 bool LibraryCallKit::inline_native_classID() { 3195 null_check_receiver(); // null-check, then ignore 3196 Node* cls = null_check(argument(1), T_OBJECT); 3197 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3198 kls = null_check(kls, T_OBJECT); 3199 ByteSize offset = TRACE_ID_OFFSET; 3200 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3201 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3202 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3203 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3204 Node* clsused = longcon(0x01l); // set the class bit 3205 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3206 3207 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3208 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3209 set_result(andl); 3210 return true; 3211 } 3212 3213 bool LibraryCallKit::inline_native_threadID() { 3214 Node* tls_ptr = NULL; 3215 Node* cur_thr = generate_current_thread(tls_ptr); 3216 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3217 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3218 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3219 3220 Node* threadid = NULL; 3221 size_t thread_id_size = OSThread::thread_id_size(); 3222 if (thread_id_size == (size_t) BytesPerLong) { 3223 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3224 } else if (thread_id_size == (size_t) BytesPerInt) { 3225 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3226 } else { 3227 ShouldNotReachHere(); 3228 } 3229 set_result(threadid); 3230 return true; 3231 } 3232 #endif 3233 3234 //------------------------inline_native_time_funcs-------------- 3235 // inline code for System.currentTimeMillis() and System.nanoTime() 3236 // these have the same type and signature 3237 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3238 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3239 const TypePtr* no_memory_effects = NULL; 3240 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3241 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3242 #ifdef ASSERT 3243 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3244 assert(value_top == top(), "second value must be top"); 3245 #endif 3246 set_result(value); 3247 return true; 3248 } 3249 3250 //------------------------inline_native_currentThread------------------ 3251 bool LibraryCallKit::inline_native_currentThread() { 3252 Node* junk = NULL; 3253 set_result(generate_current_thread(junk)); 3254 return true; 3255 } 3256 3257 //------------------------inline_native_isInterrupted------------------ 3258 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3259 bool LibraryCallKit::inline_native_isInterrupted() { 3260 // Add a fast path to t.isInterrupted(clear_int): 3261 // (t == Thread.current() && 3262 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3263 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3264 // So, in the common case that the interrupt bit is false, 3265 // we avoid making a call into the VM. Even if the interrupt bit 3266 // is true, if the clear_int argument is false, we avoid the VM call. 3267 // However, if the receiver is not currentThread, we must call the VM, 3268 // because there must be some locking done around the operation. 3269 3270 // We only go to the fast case code if we pass two guards. 3271 // Paths which do not pass are accumulated in the slow_region. 3272 3273 enum { 3274 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3275 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3276 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3277 PATH_LIMIT 3278 }; 3279 3280 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3281 // out of the function. 3282 insert_mem_bar(Op_MemBarCPUOrder); 3283 3284 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3285 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3286 3287 RegionNode* slow_region = new RegionNode(1); 3288 record_for_igvn(slow_region); 3289 3290 // (a) Receiving thread must be the current thread. 3291 Node* rec_thr = argument(0); 3292 Node* tls_ptr = NULL; 3293 Node* cur_thr = generate_current_thread(tls_ptr); 3294 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3295 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3296 3297 generate_slow_guard(bol_thr, slow_region); 3298 3299 // (b) Interrupt bit on TLS must be false. 3300 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3301 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3302 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3303 3304 // Set the control input on the field _interrupted read to prevent it floating up. 3305 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3306 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3307 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3308 3309 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3310 3311 // First fast path: if (!TLS._interrupted) return false; 3312 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3313 result_rgn->init_req(no_int_result_path, false_bit); 3314 result_val->init_req(no_int_result_path, intcon(0)); 3315 3316 // drop through to next case 3317 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3318 3319 #ifndef TARGET_OS_FAMILY_windows 3320 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3321 Node* clr_arg = argument(1); 3322 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3323 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3324 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3325 3326 // Second fast path: ... else if (!clear_int) return true; 3327 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3328 result_rgn->init_req(no_clear_result_path, false_arg); 3329 result_val->init_req(no_clear_result_path, intcon(1)); 3330 3331 // drop through to next case 3332 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3333 #else 3334 // To return true on Windows you must read the _interrupted field 3335 // and check the the event state i.e. take the slow path. 3336 #endif // TARGET_OS_FAMILY_windows 3337 3338 // (d) Otherwise, go to the slow path. 3339 slow_region->add_req(control()); 3340 set_control( _gvn.transform(slow_region)); 3341 3342 if (stopped()) { 3343 // There is no slow path. 3344 result_rgn->init_req(slow_result_path, top()); 3345 result_val->init_req(slow_result_path, top()); 3346 } else { 3347 // non-virtual because it is a private non-static 3348 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3349 3350 Node* slow_val = set_results_for_java_call(slow_call); 3351 // this->control() comes from set_results_for_java_call 3352 3353 Node* fast_io = slow_call->in(TypeFunc::I_O); 3354 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3355 3356 // These two phis are pre-filled with copies of of the fast IO and Memory 3357 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3358 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3359 3360 result_rgn->init_req(slow_result_path, control()); 3361 result_io ->init_req(slow_result_path, i_o()); 3362 result_mem->init_req(slow_result_path, reset_memory()); 3363 result_val->init_req(slow_result_path, slow_val); 3364 3365 set_all_memory(_gvn.transform(result_mem)); 3366 set_i_o( _gvn.transform(result_io)); 3367 } 3368 3369 C->set_has_split_ifs(true); // Has chance for split-if optimization 3370 set_result(result_rgn, result_val); 3371 return true; 3372 } 3373 3374 //---------------------------load_mirror_from_klass---------------------------- 3375 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3376 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3377 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3378 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3379 } 3380 3381 //-----------------------load_klass_from_mirror_common------------------------- 3382 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3383 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3384 // and branch to the given path on the region. 3385 // If never_see_null, take an uncommon trap on null, so we can optimistically 3386 // compile for the non-null case. 3387 // If the region is NULL, force never_see_null = true. 3388 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3389 bool never_see_null, 3390 RegionNode* region, 3391 int null_path, 3392 int offset) { 3393 if (region == NULL) never_see_null = true; 3394 Node* p = basic_plus_adr(mirror, offset); 3395 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3396 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3397 Node* null_ctl = top(); 3398 kls = null_check_oop(kls, &null_ctl, never_see_null); 3399 if (region != NULL) { 3400 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3401 region->init_req(null_path, null_ctl); 3402 } else { 3403 assert(null_ctl == top(), "no loose ends"); 3404 } 3405 return kls; 3406 } 3407 3408 //--------------------(inline_native_Class_query helpers)--------------------- 3409 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3410 // Fall through if (mods & mask) == bits, take the guard otherwise. 3411 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3412 // Branch around if the given klass has the given modifier bit set. 3413 // Like generate_guard, adds a new path onto the region. 3414 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3415 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3416 Node* mask = intcon(modifier_mask); 3417 Node* bits = intcon(modifier_bits); 3418 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3419 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3420 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3421 return generate_fair_guard(bol, region); 3422 } 3423 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3424 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3425 } 3426 3427 //-------------------------inline_native_Class_query------------------- 3428 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3429 const Type* return_type = TypeInt::BOOL; 3430 Node* prim_return_value = top(); // what happens if it's a primitive class? 3431 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3432 bool expect_prim = false; // most of these guys expect to work on refs 3433 3434 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3435 3436 Node* mirror = argument(0); 3437 Node* obj = top(); 3438 3439 switch (id) { 3440 case vmIntrinsics::_isInstance: 3441 // nothing is an instance of a primitive type 3442 prim_return_value = intcon(0); 3443 obj = argument(1); 3444 break; 3445 case vmIntrinsics::_getModifiers: 3446 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3447 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3448 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3449 break; 3450 case vmIntrinsics::_isInterface: 3451 prim_return_value = intcon(0); 3452 break; 3453 case vmIntrinsics::_isArray: 3454 prim_return_value = intcon(0); 3455 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3456 break; 3457 case vmIntrinsics::_isPrimitive: 3458 prim_return_value = intcon(1); 3459 expect_prim = true; // obviously 3460 break; 3461 case vmIntrinsics::_getSuperclass: 3462 prim_return_value = null(); 3463 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3464 break; 3465 case vmIntrinsics::_getComponentType: 3466 prim_return_value = null(); 3467 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3468 break; 3469 case vmIntrinsics::_getClassAccessFlags: 3470 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3471 return_type = TypeInt::INT; // not bool! 6297094 3472 break; 3473 default: 3474 fatal_unexpected_iid(id); 3475 break; 3476 } 3477 3478 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3479 if (mirror_con == NULL) return false; // cannot happen? 3480 3481 #ifndef PRODUCT 3482 if (C->print_intrinsics() || C->print_inlining()) { 3483 ciType* k = mirror_con->java_mirror_type(); 3484 if (k) { 3485 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3486 k->print_name(); 3487 tty->cr(); 3488 } 3489 } 3490 #endif 3491 3492 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3493 RegionNode* region = new RegionNode(PATH_LIMIT); 3494 record_for_igvn(region); 3495 PhiNode* phi = new PhiNode(region, return_type); 3496 3497 // The mirror will never be null of Reflection.getClassAccessFlags, however 3498 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3499 // if it is. See bug 4774291. 3500 3501 // For Reflection.getClassAccessFlags(), the null check occurs in 3502 // the wrong place; see inline_unsafe_access(), above, for a similar 3503 // situation. 3504 mirror = null_check(mirror); 3505 // If mirror or obj is dead, only null-path is taken. 3506 if (stopped()) return true; 3507 3508 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3509 3510 // Now load the mirror's klass metaobject, and null-check it. 3511 // Side-effects region with the control path if the klass is null. 3512 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3513 // If kls is null, we have a primitive mirror. 3514 phi->init_req(_prim_path, prim_return_value); 3515 if (stopped()) { set_result(region, phi); return true; } 3516 bool safe_for_replace = (region->in(_prim_path) == top()); 3517 3518 Node* p; // handy temp 3519 Node* null_ctl; 3520 3521 // Now that we have the non-null klass, we can perform the real query. 3522 // For constant classes, the query will constant-fold in LoadNode::Value. 3523 Node* query_value = top(); 3524 switch (id) { 3525 case vmIntrinsics::_isInstance: 3526 // nothing is an instance of a primitive type 3527 query_value = gen_instanceof(obj, kls, safe_for_replace); 3528 break; 3529 3530 case vmIntrinsics::_getModifiers: 3531 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3532 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3533 break; 3534 3535 case vmIntrinsics::_isInterface: 3536 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3537 if (generate_interface_guard(kls, region) != NULL) 3538 // A guard was added. If the guard is taken, it was an interface. 3539 phi->add_req(intcon(1)); 3540 // If we fall through, it's a plain class. 3541 query_value = intcon(0); 3542 break; 3543 3544 case vmIntrinsics::_isArray: 3545 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3546 if (generate_array_guard(kls, region) != NULL) 3547 // A guard was added. If the guard is taken, it was an array. 3548 phi->add_req(intcon(1)); 3549 // If we fall through, it's a plain class. 3550 query_value = intcon(0); 3551 break; 3552 3553 case vmIntrinsics::_isPrimitive: 3554 query_value = intcon(0); // "normal" path produces false 3555 break; 3556 3557 case vmIntrinsics::_getSuperclass: 3558 // The rules here are somewhat unfortunate, but we can still do better 3559 // with random logic than with a JNI call. 3560 // Interfaces store null or Object as _super, but must report null. 3561 // Arrays store an intermediate super as _super, but must report Object. 3562 // Other types can report the actual _super. 3563 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3564 if (generate_interface_guard(kls, region) != NULL) 3565 // A guard was added. If the guard is taken, it was an interface. 3566 phi->add_req(null()); 3567 if (generate_array_guard(kls, region) != NULL) 3568 // A guard was added. If the guard is taken, it was an array. 3569 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3570 // If we fall through, it's a plain class. Get its _super. 3571 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3572 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3573 null_ctl = top(); 3574 kls = null_check_oop(kls, &null_ctl); 3575 if (null_ctl != top()) { 3576 // If the guard is taken, Object.superClass is null (both klass and mirror). 3577 region->add_req(null_ctl); 3578 phi ->add_req(null()); 3579 } 3580 if (!stopped()) { 3581 query_value = load_mirror_from_klass(kls); 3582 } 3583 break; 3584 3585 case vmIntrinsics::_getComponentType: 3586 if (generate_array_guard(kls, region) != NULL) { 3587 // Be sure to pin the oop load to the guard edge just created: 3588 Node* is_array_ctrl = region->in(region->req()-1); 3589 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); 3590 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3591 phi->add_req(cmo); 3592 } 3593 query_value = null(); // non-array case is null 3594 break; 3595 3596 case vmIntrinsics::_getClassAccessFlags: 3597 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3598 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3599 break; 3600 3601 default: 3602 fatal_unexpected_iid(id); 3603 break; 3604 } 3605 3606 // Fall-through is the normal case of a query to a real class. 3607 phi->init_req(1, query_value); 3608 region->init_req(1, control()); 3609 3610 C->set_has_split_ifs(true); // Has chance for split-if optimization 3611 set_result(region, phi); 3612 return true; 3613 } 3614 3615 //--------------------------inline_native_subtype_check------------------------ 3616 // This intrinsic takes the JNI calls out of the heart of 3617 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3618 bool LibraryCallKit::inline_native_subtype_check() { 3619 // Pull both arguments off the stack. 3620 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3621 args[0] = argument(0); 3622 args[1] = argument(1); 3623 Node* klasses[2]; // corresponding Klasses: superk, subk 3624 klasses[0] = klasses[1] = top(); 3625 3626 enum { 3627 // A full decision tree on {superc is prim, subc is prim}: 3628 _prim_0_path = 1, // {P,N} => false 3629 // {P,P} & superc!=subc => false 3630 _prim_same_path, // {P,P} & superc==subc => true 3631 _prim_1_path, // {N,P} => false 3632 _ref_subtype_path, // {N,N} & subtype check wins => true 3633 _both_ref_path, // {N,N} & subtype check loses => false 3634 PATH_LIMIT 3635 }; 3636 3637 RegionNode* region = new RegionNode(PATH_LIMIT); 3638 Node* phi = new PhiNode(region, TypeInt::BOOL); 3639 record_for_igvn(region); 3640 3641 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3642 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3643 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3644 3645 // First null-check both mirrors and load each mirror's klass metaobject. 3646 int which_arg; 3647 for (which_arg = 0; which_arg <= 1; which_arg++) { 3648 Node* arg = args[which_arg]; 3649 arg = null_check(arg); 3650 if (stopped()) break; 3651 args[which_arg] = arg; 3652 3653 Node* p = basic_plus_adr(arg, class_klass_offset); 3654 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3655 klasses[which_arg] = _gvn.transform(kls); 3656 } 3657 3658 // Having loaded both klasses, test each for null. 3659 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3660 for (which_arg = 0; which_arg <= 1; which_arg++) { 3661 Node* kls = klasses[which_arg]; 3662 Node* null_ctl = top(); 3663 kls = null_check_oop(kls, &null_ctl, never_see_null); 3664 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3665 region->init_req(prim_path, null_ctl); 3666 if (stopped()) break; 3667 klasses[which_arg] = kls; 3668 } 3669 3670 if (!stopped()) { 3671 // now we have two reference types, in klasses[0..1] 3672 Node* subk = klasses[1]; // the argument to isAssignableFrom 3673 Node* superk = klasses[0]; // the receiver 3674 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3675 // now we have a successful reference subtype check 3676 region->set_req(_ref_subtype_path, control()); 3677 } 3678 3679 // If both operands are primitive (both klasses null), then 3680 // we must return true when they are identical primitives. 3681 // It is convenient to test this after the first null klass check. 3682 set_control(region->in(_prim_0_path)); // go back to first null check 3683 if (!stopped()) { 3684 // Since superc is primitive, make a guard for the superc==subc case. 3685 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3686 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3687 generate_guard(bol_eq, region, PROB_FAIR); 3688 if (region->req() == PATH_LIMIT+1) { 3689 // A guard was added. If the added guard is taken, superc==subc. 3690 region->swap_edges(PATH_LIMIT, _prim_same_path); 3691 region->del_req(PATH_LIMIT); 3692 } 3693 region->set_req(_prim_0_path, control()); // Not equal after all. 3694 } 3695 3696 // these are the only paths that produce 'true': 3697 phi->set_req(_prim_same_path, intcon(1)); 3698 phi->set_req(_ref_subtype_path, intcon(1)); 3699 3700 // pull together the cases: 3701 assert(region->req() == PATH_LIMIT, "sane region"); 3702 for (uint i = 1; i < region->req(); i++) { 3703 Node* ctl = region->in(i); 3704 if (ctl == NULL || ctl == top()) { 3705 region->set_req(i, top()); 3706 phi ->set_req(i, top()); 3707 } else if (phi->in(i) == NULL) { 3708 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3709 } 3710 } 3711 3712 set_control(_gvn.transform(region)); 3713 set_result(_gvn.transform(phi)); 3714 return true; 3715 } 3716 3717 //---------------------generate_array_guard_common------------------------ 3718 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3719 bool obj_array, bool not_array) { 3720 // If obj_array/non_array==false/false: 3721 // Branch around if the given klass is in fact an array (either obj or prim). 3722 // If obj_array/non_array==false/true: 3723 // Branch around if the given klass is not an array klass of any kind. 3724 // If obj_array/non_array==true/true: 3725 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3726 // If obj_array/non_array==true/false: 3727 // Branch around if the kls is an oop array (Object[] or subtype) 3728 // 3729 // Like generate_guard, adds a new path onto the region. 3730 jint layout_con = 0; 3731 Node* layout_val = get_layout_helper(kls, layout_con); 3732 if (layout_val == NULL) { 3733 bool query = (obj_array 3734 ? Klass::layout_helper_is_objArray(layout_con) 3735 : Klass::layout_helper_is_array(layout_con)); 3736 if (query == not_array) { 3737 return NULL; // never a branch 3738 } else { // always a branch 3739 Node* always_branch = control(); 3740 if (region != NULL) 3741 region->add_req(always_branch); 3742 set_control(top()); 3743 return always_branch; 3744 } 3745 } 3746 // Now test the correct condition. 3747 jint nval = (obj_array 3748 ? ((jint)Klass::_lh_array_tag_type_value 3749 << Klass::_lh_array_tag_shift) 3750 : Klass::_lh_neutral_value); 3751 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3752 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3753 // invert the test if we are looking for a non-array 3754 if (not_array) btest = BoolTest(btest).negate(); 3755 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3756 return generate_fair_guard(bol, region); 3757 } 3758 3759 3760 //-----------------------inline_native_newArray-------------------------- 3761 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3762 bool LibraryCallKit::inline_native_newArray() { 3763 Node* mirror = argument(0); 3764 Node* count_val = argument(1); 3765 3766 mirror = null_check(mirror); 3767 // If mirror or obj is dead, only null-path is taken. 3768 if (stopped()) return true; 3769 3770 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3771 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3772 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3773 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3774 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3775 3776 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3777 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3778 result_reg, _slow_path); 3779 Node* normal_ctl = control(); 3780 Node* no_array_ctl = result_reg->in(_slow_path); 3781 3782 // Generate code for the slow case. We make a call to newArray(). 3783 set_control(no_array_ctl); 3784 if (!stopped()) { 3785 // Either the input type is void.class, or else the 3786 // array klass has not yet been cached. Either the 3787 // ensuing call will throw an exception, or else it 3788 // will cache the array klass for next time. 3789 PreserveJVMState pjvms(this); 3790 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3791 Node* slow_result = set_results_for_java_call(slow_call); 3792 // this->control() comes from set_results_for_java_call 3793 result_reg->set_req(_slow_path, control()); 3794 result_val->set_req(_slow_path, slow_result); 3795 result_io ->set_req(_slow_path, i_o()); 3796 result_mem->set_req(_slow_path, reset_memory()); 3797 } 3798 3799 set_control(normal_ctl); 3800 if (!stopped()) { 3801 // Normal case: The array type has been cached in the java.lang.Class. 3802 // The following call works fine even if the array type is polymorphic. 3803 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3804 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3805 result_reg->init_req(_normal_path, control()); 3806 result_val->init_req(_normal_path, obj); 3807 result_io ->init_req(_normal_path, i_o()); 3808 result_mem->init_req(_normal_path, reset_memory()); 3809 } 3810 3811 // Return the combined state. 3812 set_i_o( _gvn.transform(result_io) ); 3813 set_all_memory( _gvn.transform(result_mem)); 3814 3815 C->set_has_split_ifs(true); // Has chance for split-if optimization 3816 set_result(result_reg, result_val); 3817 return true; 3818 } 3819 3820 //----------------------inline_native_getLength-------------------------- 3821 // public static native int java.lang.reflect.Array.getLength(Object array); 3822 bool LibraryCallKit::inline_native_getLength() { 3823 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3824 3825 Node* array = null_check(argument(0)); 3826 // If array is dead, only null-path is taken. 3827 if (stopped()) return true; 3828 3829 // Deoptimize if it is a non-array. 3830 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3831 3832 if (non_array != NULL) { 3833 PreserveJVMState pjvms(this); 3834 set_control(non_array); 3835 uncommon_trap(Deoptimization::Reason_intrinsic, 3836 Deoptimization::Action_maybe_recompile); 3837 } 3838 3839 // If control is dead, only non-array-path is taken. 3840 if (stopped()) return true; 3841 3842 // The works fine even if the array type is polymorphic. 3843 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3844 Node* result = load_array_length(array); 3845 3846 C->set_has_split_ifs(true); // Has chance for split-if optimization 3847 set_result(result); 3848 return true; 3849 } 3850 3851 //------------------------inline_array_copyOf---------------------------- 3852 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3853 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3854 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3855 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3856 3857 // Get the arguments. 3858 Node* original = argument(0); 3859 Node* start = is_copyOfRange? argument(1): intcon(0); 3860 Node* end = is_copyOfRange? argument(2): argument(1); 3861 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3862 3863 Node* newcopy; 3864 3865 // Set the original stack and the reexecute bit for the interpreter to reexecute 3866 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3867 { PreserveReexecuteState preexecs(this); 3868 jvms()->set_should_reexecute(true); 3869 3870 array_type_mirror = null_check(array_type_mirror); 3871 original = null_check(original); 3872 3873 // Check if a null path was taken unconditionally. 3874 if (stopped()) return true; 3875 3876 Node* orig_length = load_array_length(original); 3877 3878 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3879 klass_node = null_check(klass_node); 3880 3881 RegionNode* bailout = new RegionNode(1); 3882 record_for_igvn(bailout); 3883 3884 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3885 // Bail out if that is so. 3886 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3887 if (not_objArray != NULL) { 3888 // Improve the klass node's type from the new optimistic assumption: 3889 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3890 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3891 Node* cast = new CastPPNode(klass_node, akls); 3892 cast->init_req(0, control()); 3893 klass_node = _gvn.transform(cast); 3894 } 3895 3896 // Bail out if either start or end is negative. 3897 generate_negative_guard(start, bailout, &start); 3898 generate_negative_guard(end, bailout, &end); 3899 3900 Node* length = end; 3901 if (_gvn.type(start) != TypeInt::ZERO) { 3902 length = _gvn.transform(new SubINode(end, start)); 3903 } 3904 3905 // Bail out if length is negative. 3906 // Without this the new_array would throw 3907 // NegativeArraySizeException but IllegalArgumentException is what 3908 // should be thrown 3909 generate_negative_guard(length, bailout, &length); 3910 3911 if (bailout->req() > 1) { 3912 PreserveJVMState pjvms(this); 3913 set_control(_gvn.transform(bailout)); 3914 uncommon_trap(Deoptimization::Reason_intrinsic, 3915 Deoptimization::Action_maybe_recompile); 3916 } 3917 3918 if (!stopped()) { 3919 // How many elements will we copy from the original? 3920 // The answer is MinI(orig_length - start, length). 3921 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3922 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3923 3924 newcopy = new_array(klass_node, length, 0); // no argments to push 3925 3926 // Generate a direct call to the right arraycopy function(s). 3927 // We know the copy is disjoint but we might not know if the 3928 // oop stores need checking. 3929 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3930 // This will fail a store-check if x contains any non-nulls. 3931 bool disjoint_bases = true; 3932 // if start > orig_length then the length of the copy may be 3933 // negative. 3934 bool length_never_negative = !is_copyOfRange; 3935 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3936 original, start, newcopy, intcon(0), moved, 3937 disjoint_bases, length_never_negative); 3938 } 3939 } // original reexecute is set back here 3940 3941 C->set_has_split_ifs(true); // Has chance for split-if optimization 3942 if (!stopped()) { 3943 set_result(newcopy); 3944 } 3945 return true; 3946 } 3947 3948 3949 //----------------------generate_virtual_guard--------------------------- 3950 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3951 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3952 RegionNode* slow_region) { 3953 ciMethod* method = callee(); 3954 int vtable_index = method->vtable_index(); 3955 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3956 err_msg_res("bad index %d", vtable_index)); 3957 // Get the Method* out of the appropriate vtable entry. 3958 int entry_offset = (InstanceKlass::vtable_start_offset() + 3959 vtable_index*vtableEntry::size()) * wordSize + 3960 vtableEntry::method_offset_in_bytes(); 3961 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3962 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3963 3964 // Compare the target method with the expected method (e.g., Object.hashCode). 3965 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3966 3967 Node* native_call = makecon(native_call_addr); 3968 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3969 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3970 3971 return generate_slow_guard(test_native, slow_region); 3972 } 3973 3974 //-----------------------generate_method_call---------------------------- 3975 // Use generate_method_call to make a slow-call to the real 3976 // method if the fast path fails. An alternative would be to 3977 // use a stub like OptoRuntime::slow_arraycopy_Java. 3978 // This only works for expanding the current library call, 3979 // not another intrinsic. (E.g., don't use this for making an 3980 // arraycopy call inside of the copyOf intrinsic.) 3981 CallJavaNode* 3982 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3983 // When compiling the intrinsic method itself, do not use this technique. 3984 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3985 3986 ciMethod* method = callee(); 3987 // ensure the JVMS we have will be correct for this call 3988 guarantee(method_id == method->intrinsic_id(), "must match"); 3989 3990 const TypeFunc* tf = TypeFunc::make(method); 3991 CallJavaNode* slow_call; 3992 if (is_static) { 3993 assert(!is_virtual, ""); 3994 slow_call = new CallStaticJavaNode(C, tf, 3995 SharedRuntime::get_resolve_static_call_stub(), 3996 method, bci()); 3997 } else if (is_virtual) { 3998 null_check_receiver(); 3999 int vtable_index = Method::invalid_vtable_index; 4000 if (UseInlineCaches) { 4001 // Suppress the vtable call 4002 } else { 4003 // hashCode and clone are not a miranda methods, 4004 // so the vtable index is fixed. 4005 // No need to use the linkResolver to get it. 4006 vtable_index = method->vtable_index(); 4007 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4008 err_msg_res("bad index %d", vtable_index)); 4009 } 4010 slow_call = new CallDynamicJavaNode(tf, 4011 SharedRuntime::get_resolve_virtual_call_stub(), 4012 method, vtable_index, bci()); 4013 } else { // neither virtual nor static: opt_virtual 4014 null_check_receiver(); 4015 slow_call = new CallStaticJavaNode(C, tf, 4016 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4017 method, bci()); 4018 slow_call->set_optimized_virtual(true); 4019 } 4020 set_arguments_for_java_call(slow_call); 4021 set_edges_for_java_call(slow_call); 4022 return slow_call; 4023 } 4024 4025 4026 /** 4027 * Build special case code for calls to hashCode on an object. This call may 4028 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4029 * slightly different code. 4030 */ 4031 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4032 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4033 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4034 4035 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4036 4037 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4038 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4039 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4040 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4041 Node* obj = NULL; 4042 if (!is_static) { 4043 // Check for hashing null object 4044 obj = null_check_receiver(); 4045 if (stopped()) return true; // unconditionally null 4046 result_reg->init_req(_null_path, top()); 4047 result_val->init_req(_null_path, top()); 4048 } else { 4049 // Do a null check, and return zero if null. 4050 // System.identityHashCode(null) == 0 4051 obj = argument(0); 4052 Node* null_ctl = top(); 4053 obj = null_check_oop(obj, &null_ctl); 4054 result_reg->init_req(_null_path, null_ctl); 4055 result_val->init_req(_null_path, _gvn.intcon(0)); 4056 } 4057 4058 // Unconditionally null? Then return right away. 4059 if (stopped()) { 4060 set_control( result_reg->in(_null_path)); 4061 if (!stopped()) 4062 set_result(result_val->in(_null_path)); 4063 return true; 4064 } 4065 4066 // We only go to the fast case code if we pass a number of guards. The 4067 // paths which do not pass are accumulated in the slow_region. 4068 RegionNode* slow_region = new RegionNode(1); 4069 record_for_igvn(slow_region); 4070 4071 // If this is a virtual call, we generate a funny guard. We pull out 4072 // the vtable entry corresponding to hashCode() from the target object. 4073 // If the target method which we are calling happens to be the native 4074 // Object hashCode() method, we pass the guard. We do not need this 4075 // guard for non-virtual calls -- the caller is known to be the native 4076 // Object hashCode(). 4077 if (is_virtual) { 4078 // After null check, get the object's klass. 4079 Node* obj_klass = load_object_klass(obj); 4080 generate_virtual_guard(obj_klass, slow_region); 4081 } 4082 4083 // Get the header out of the object, use LoadMarkNode when available 4084 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4085 // The control of the load must be NULL. Otherwise, the load can move before 4086 // the null check after castPP removal. 4087 Node* no_ctrl = NULL; 4088 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4089 4090 // Test the header to see if it is unlocked. 4091 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4092 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4093 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4094 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4095 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4096 4097 generate_slow_guard(test_unlocked, slow_region); 4098 4099 // Get the hash value and check to see that it has been properly assigned. 4100 // We depend on hash_mask being at most 32 bits and avoid the use of 4101 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4102 // vm: see markOop.hpp. 4103 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4104 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4105 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4106 // This hack lets the hash bits live anywhere in the mark object now, as long 4107 // as the shift drops the relevant bits into the low 32 bits. Note that 4108 // Java spec says that HashCode is an int so there's no point in capturing 4109 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4110 hshifted_header = ConvX2I(hshifted_header); 4111 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4112 4113 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4114 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4115 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4116 4117 generate_slow_guard(test_assigned, slow_region); 4118 4119 Node* init_mem = reset_memory(); 4120 // fill in the rest of the null path: 4121 result_io ->init_req(_null_path, i_o()); 4122 result_mem->init_req(_null_path, init_mem); 4123 4124 result_val->init_req(_fast_path, hash_val); 4125 result_reg->init_req(_fast_path, control()); 4126 result_io ->init_req(_fast_path, i_o()); 4127 result_mem->init_req(_fast_path, init_mem); 4128 4129 // Generate code for the slow case. We make a call to hashCode(). 4130 set_control(_gvn.transform(slow_region)); 4131 if (!stopped()) { 4132 // No need for PreserveJVMState, because we're using up the present state. 4133 set_all_memory(init_mem); 4134 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4135 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4136 Node* slow_result = set_results_for_java_call(slow_call); 4137 // this->control() comes from set_results_for_java_call 4138 result_reg->init_req(_slow_path, control()); 4139 result_val->init_req(_slow_path, slow_result); 4140 result_io ->set_req(_slow_path, i_o()); 4141 result_mem ->set_req(_slow_path, reset_memory()); 4142 } 4143 4144 // Return the combined state. 4145 set_i_o( _gvn.transform(result_io) ); 4146 set_all_memory( _gvn.transform(result_mem)); 4147 4148 set_result(result_reg, result_val); 4149 return true; 4150 } 4151 4152 //---------------------------inline_native_getClass---------------------------- 4153 // public final native Class<?> java.lang.Object.getClass(); 4154 // 4155 // Build special case code for calls to getClass on an object. 4156 bool LibraryCallKit::inline_native_getClass() { 4157 Node* obj = null_check_receiver(); 4158 if (stopped()) return true; 4159 set_result(load_mirror_from_klass(load_object_klass(obj))); 4160 return true; 4161 } 4162 4163 //-----------------inline_native_Reflection_getCallerClass--------------------- 4164 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4165 // 4166 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4167 // 4168 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4169 // in that it must skip particular security frames and checks for 4170 // caller sensitive methods. 4171 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4172 #ifndef PRODUCT 4173 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4174 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4175 } 4176 #endif 4177 4178 if (!jvms()->has_method()) { 4179 #ifndef PRODUCT 4180 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4181 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4182 } 4183 #endif 4184 return false; 4185 } 4186 4187 // Walk back up the JVM state to find the caller at the required 4188 // depth. 4189 JVMState* caller_jvms = jvms(); 4190 4191 // Cf. JVM_GetCallerClass 4192 // NOTE: Start the loop at depth 1 because the current JVM state does 4193 // not include the Reflection.getCallerClass() frame. 4194 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4195 ciMethod* m = caller_jvms->method(); 4196 switch (n) { 4197 case 0: 4198 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4199 break; 4200 case 1: 4201 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4202 if (!m->caller_sensitive()) { 4203 #ifndef PRODUCT 4204 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4205 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4206 } 4207 #endif 4208 return false; // bail-out; let JVM_GetCallerClass do the work 4209 } 4210 break; 4211 default: 4212 if (!m->is_ignored_by_security_stack_walk()) { 4213 // We have reached the desired frame; return the holder class. 4214 // Acquire method holder as java.lang.Class and push as constant. 4215 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4216 ciInstance* caller_mirror = caller_klass->java_mirror(); 4217 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4218 4219 #ifndef PRODUCT 4220 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4221 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()); 4222 tty->print_cr(" JVM state at this point:"); 4223 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4224 ciMethod* m = jvms()->of_depth(i)->method(); 4225 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4226 } 4227 } 4228 #endif 4229 return true; 4230 } 4231 break; 4232 } 4233 } 4234 4235 #ifndef PRODUCT 4236 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4237 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4238 tty->print_cr(" JVM state at this point:"); 4239 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4240 ciMethod* m = jvms()->of_depth(i)->method(); 4241 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4242 } 4243 } 4244 #endif 4245 4246 return false; // bail-out; let JVM_GetCallerClass do the work 4247 } 4248 4249 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4250 Node* arg = argument(0); 4251 Node* result; 4252 4253 switch (id) { 4254 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4255 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4256 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4257 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4258 4259 case vmIntrinsics::_doubleToLongBits: { 4260 // two paths (plus control) merge in a wood 4261 RegionNode *r = new RegionNode(3); 4262 Node *phi = new PhiNode(r, TypeLong::LONG); 4263 4264 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4265 // Build the boolean node 4266 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4267 4268 // Branch either way. 4269 // NaN case is less traveled, which makes all the difference. 4270 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4271 Node *opt_isnan = _gvn.transform(ifisnan); 4272 assert( opt_isnan->is_If(), "Expect an IfNode"); 4273 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4274 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4275 4276 set_control(iftrue); 4277 4278 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4279 Node *slow_result = longcon(nan_bits); // return NaN 4280 phi->init_req(1, _gvn.transform( slow_result )); 4281 r->init_req(1, iftrue); 4282 4283 // Else fall through 4284 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4285 set_control(iffalse); 4286 4287 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4288 r->init_req(2, iffalse); 4289 4290 // Post merge 4291 set_control(_gvn.transform(r)); 4292 record_for_igvn(r); 4293 4294 C->set_has_split_ifs(true); // Has chance for split-if optimization 4295 result = phi; 4296 assert(result->bottom_type()->isa_long(), "must be"); 4297 break; 4298 } 4299 4300 case vmIntrinsics::_floatToIntBits: { 4301 // two paths (plus control) merge in a wood 4302 RegionNode *r = new RegionNode(3); 4303 Node *phi = new PhiNode(r, TypeInt::INT); 4304 4305 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4306 // Build the boolean node 4307 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4308 4309 // Branch either way. 4310 // NaN case is less traveled, which makes all the difference. 4311 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4312 Node *opt_isnan = _gvn.transform(ifisnan); 4313 assert( opt_isnan->is_If(), "Expect an IfNode"); 4314 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4315 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4316 4317 set_control(iftrue); 4318 4319 static const jint nan_bits = 0x7fc00000; 4320 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4321 phi->init_req(1, _gvn.transform( slow_result )); 4322 r->init_req(1, iftrue); 4323 4324 // Else fall through 4325 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4326 set_control(iffalse); 4327 4328 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4329 r->init_req(2, iffalse); 4330 4331 // Post merge 4332 set_control(_gvn.transform(r)); 4333 record_for_igvn(r); 4334 4335 C->set_has_split_ifs(true); // Has chance for split-if optimization 4336 result = phi; 4337 assert(result->bottom_type()->isa_int(), "must be"); 4338 break; 4339 } 4340 4341 default: 4342 fatal_unexpected_iid(id); 4343 break; 4344 } 4345 set_result(_gvn.transform(result)); 4346 return true; 4347 } 4348 4349 #ifdef _LP64 4350 #define XTOP ,top() /*additional argument*/ 4351 #else //_LP64 4352 #define XTOP /*no additional argument*/ 4353 #endif //_LP64 4354 4355 //----------------------inline_unsafe_copyMemory------------------------- 4356 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4357 bool LibraryCallKit::inline_unsafe_copyMemory() { 4358 if (callee()->is_static()) return false; // caller must have the capability! 4359 null_check_receiver(); // null-check receiver 4360 if (stopped()) return true; 4361 4362 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4363 4364 Node* src_ptr = argument(1); // type: oop 4365 Node* src_off = ConvL2X(argument(2)); // type: long 4366 Node* dst_ptr = argument(4); // type: oop 4367 Node* dst_off = ConvL2X(argument(5)); // type: long 4368 Node* size = ConvL2X(argument(7)); // type: long 4369 4370 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4371 "fieldOffset must be byte-scaled"); 4372 4373 Node* src = make_unsafe_address(src_ptr, src_off); 4374 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4375 4376 // Conservatively insert a memory barrier on all memory slices. 4377 // Do not let writes of the copy source or destination float below the copy. 4378 insert_mem_bar(Op_MemBarCPUOrder); 4379 4380 // Call it. Note that the length argument is not scaled. 4381 make_runtime_call(RC_LEAF|RC_NO_FP, 4382 OptoRuntime::fast_arraycopy_Type(), 4383 StubRoutines::unsafe_arraycopy(), 4384 "unsafe_arraycopy", 4385 TypeRawPtr::BOTTOM, 4386 src, dst, size XTOP); 4387 4388 // Do not let reads of the copy destination float above the copy. 4389 insert_mem_bar(Op_MemBarCPUOrder); 4390 4391 return true; 4392 } 4393 4394 //------------------------clone_coping----------------------------------- 4395 // Helper function for inline_native_clone. 4396 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4397 assert(obj_size != NULL, ""); 4398 Node* raw_obj = alloc_obj->in(1); 4399 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4400 4401 AllocateNode* alloc = NULL; 4402 if (ReduceBulkZeroing) { 4403 // We will be completely responsible for initializing this object - 4404 // mark Initialize node as complete. 4405 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4406 // The object was just allocated - there should be no any stores! 4407 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4408 // Mark as complete_with_arraycopy so that on AllocateNode 4409 // expansion, we know this AllocateNode is initialized by an array 4410 // copy and a StoreStore barrier exists after the array copy. 4411 alloc->initialization()->set_complete_with_arraycopy(); 4412 } 4413 4414 // Copy the fastest available way. 4415 // TODO: generate fields copies for small objects instead. 4416 Node* src = obj; 4417 Node* dest = alloc_obj; 4418 Node* size = _gvn.transform(obj_size); 4419 4420 // Exclude the header but include array length to copy by 8 bytes words. 4421 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4422 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4423 instanceOopDesc::base_offset_in_bytes(); 4424 // base_off: 4425 // 8 - 32-bit VM 4426 // 12 - 64-bit VM, compressed klass 4427 // 16 - 64-bit VM, normal klass 4428 if (base_off % BytesPerLong != 0) { 4429 assert(UseCompressedClassPointers, ""); 4430 if (is_array) { 4431 // Exclude length to copy by 8 bytes words. 4432 base_off += sizeof(int); 4433 } else { 4434 // Include klass to copy by 8 bytes words. 4435 base_off = instanceOopDesc::klass_offset_in_bytes(); 4436 } 4437 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4438 } 4439 src = basic_plus_adr(src, base_off); 4440 dest = basic_plus_adr(dest, base_off); 4441 4442 // Compute the length also, if needed: 4443 Node* countx = size; 4444 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4445 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4446 4447 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4448 bool disjoint_bases = true; 4449 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4450 src, NULL, dest, NULL, countx, 4451 /*dest_uninitialized*/true); 4452 4453 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4454 if (card_mark) { 4455 assert(!is_array, ""); 4456 // Put in store barrier for any and all oops we are sticking 4457 // into this object. (We could avoid this if we could prove 4458 // that the object type contains no oop fields at all.) 4459 Node* no_particular_value = NULL; 4460 Node* no_particular_field = NULL; 4461 int raw_adr_idx = Compile::AliasIdxRaw; 4462 post_barrier(control(), 4463 memory(raw_adr_type), 4464 alloc_obj, 4465 no_particular_field, 4466 raw_adr_idx, 4467 no_particular_value, 4468 T_OBJECT, 4469 false); 4470 } 4471 4472 // Do not let reads from the cloned object float above the arraycopy. 4473 if (alloc != NULL) { 4474 // Do not let stores that initialize this object be reordered with 4475 // a subsequent store that would make this object accessible by 4476 // other threads. 4477 // Record what AllocateNode this StoreStore protects so that 4478 // escape analysis can go from the MemBarStoreStoreNode to the 4479 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4480 // based on the escape status of the AllocateNode. 4481 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4482 } else { 4483 insert_mem_bar(Op_MemBarCPUOrder); 4484 } 4485 } 4486 4487 //------------------------inline_native_clone---------------------------- 4488 // protected native Object java.lang.Object.clone(); 4489 // 4490 // Here are the simple edge cases: 4491 // null receiver => normal trap 4492 // virtual and clone was overridden => slow path to out-of-line clone 4493 // not cloneable or finalizer => slow path to out-of-line Object.clone 4494 // 4495 // The general case has two steps, allocation and copying. 4496 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4497 // 4498 // Copying also has two cases, oop arrays and everything else. 4499 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4500 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4501 // 4502 // These steps fold up nicely if and when the cloned object's klass 4503 // can be sharply typed as an object array, a type array, or an instance. 4504 // 4505 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4506 PhiNode* result_val; 4507 4508 // Set the reexecute bit for the interpreter to reexecute 4509 // the bytecode that invokes Object.clone if deoptimization happens. 4510 { PreserveReexecuteState preexecs(this); 4511 jvms()->set_should_reexecute(true); 4512 4513 Node* obj = null_check_receiver(); 4514 if (stopped()) return true; 4515 4516 Node* obj_klass = load_object_klass(obj); 4517 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4518 const TypeOopPtr* toop = ((tklass != NULL) 4519 ? tklass->as_instance_type() 4520 : TypeInstPtr::NOTNULL); 4521 4522 // Conservatively insert a memory barrier on all memory slices. 4523 // Do not let writes into the original float below the clone. 4524 insert_mem_bar(Op_MemBarCPUOrder); 4525 4526 // paths into result_reg: 4527 enum { 4528 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4529 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4530 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4531 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4532 PATH_LIMIT 4533 }; 4534 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4535 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4536 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4537 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4538 record_for_igvn(result_reg); 4539 4540 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4541 int raw_adr_idx = Compile::AliasIdxRaw; 4542 4543 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4544 if (array_ctl != NULL) { 4545 // It's an array. 4546 PreserveJVMState pjvms(this); 4547 set_control(array_ctl); 4548 Node* obj_length = load_array_length(obj); 4549 Node* obj_size = NULL; 4550 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4551 4552 if (!use_ReduceInitialCardMarks()) { 4553 // If it is an oop array, it requires very special treatment, 4554 // because card marking is required on each card of the array. 4555 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4556 if (is_obja != NULL) { 4557 PreserveJVMState pjvms2(this); 4558 set_control(is_obja); 4559 // Generate a direct call to the right arraycopy function(s). 4560 bool disjoint_bases = true; 4561 bool length_never_negative = true; 4562 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4563 obj, intcon(0), alloc_obj, intcon(0), 4564 obj_length, 4565 disjoint_bases, length_never_negative); 4566 result_reg->init_req(_objArray_path, control()); 4567 result_val->init_req(_objArray_path, alloc_obj); 4568 result_i_o ->set_req(_objArray_path, i_o()); 4569 result_mem ->set_req(_objArray_path, reset_memory()); 4570 } 4571 } 4572 // Otherwise, there are no card marks to worry about. 4573 // (We can dispense with card marks if we know the allocation 4574 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4575 // causes the non-eden paths to take compensating steps to 4576 // simulate a fresh allocation, so that no further 4577 // card marks are required in compiled code to initialize 4578 // the object.) 4579 4580 if (!stopped()) { 4581 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4582 4583 // Present the results of the copy. 4584 result_reg->init_req(_array_path, control()); 4585 result_val->init_req(_array_path, alloc_obj); 4586 result_i_o ->set_req(_array_path, i_o()); 4587 result_mem ->set_req(_array_path, reset_memory()); 4588 } 4589 } 4590 4591 // We only go to the instance fast case code if we pass a number of guards. 4592 // The paths which do not pass are accumulated in the slow_region. 4593 RegionNode* slow_region = new RegionNode(1); 4594 record_for_igvn(slow_region); 4595 if (!stopped()) { 4596 // It's an instance (we did array above). Make the slow-path tests. 4597 // If this is a virtual call, we generate a funny guard. We grab 4598 // the vtable entry corresponding to clone() from the target object. 4599 // If the target method which we are calling happens to be the 4600 // Object clone() method, we pass the guard. We do not need this 4601 // guard for non-virtual calls; the caller is known to be the native 4602 // Object clone(). 4603 if (is_virtual) { 4604 generate_virtual_guard(obj_klass, slow_region); 4605 } 4606 4607 // The object must be cloneable and must not have a finalizer. 4608 // Both of these conditions may be checked in a single test. 4609 // We could optimize the cloneable test further, but we don't care. 4610 generate_access_flags_guard(obj_klass, 4611 // Test both conditions: 4612 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4613 // Must be cloneable but not finalizer: 4614 JVM_ACC_IS_CLONEABLE, 4615 slow_region); 4616 } 4617 4618 if (!stopped()) { 4619 // It's an instance, and it passed the slow-path tests. 4620 PreserveJVMState pjvms(this); 4621 Node* obj_size = NULL; 4622 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4623 // is reexecuted if deoptimization occurs and there could be problems when merging 4624 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4625 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4626 4627 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4628 4629 // Present the results of the slow call. 4630 result_reg->init_req(_instance_path, control()); 4631 result_val->init_req(_instance_path, alloc_obj); 4632 result_i_o ->set_req(_instance_path, i_o()); 4633 result_mem ->set_req(_instance_path, reset_memory()); 4634 } 4635 4636 // Generate code for the slow case. We make a call to clone(). 4637 set_control(_gvn.transform(slow_region)); 4638 if (!stopped()) { 4639 PreserveJVMState pjvms(this); 4640 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4641 Node* slow_result = set_results_for_java_call(slow_call); 4642 // this->control() comes from set_results_for_java_call 4643 result_reg->init_req(_slow_path, control()); 4644 result_val->init_req(_slow_path, slow_result); 4645 result_i_o ->set_req(_slow_path, i_o()); 4646 result_mem ->set_req(_slow_path, reset_memory()); 4647 } 4648 4649 // Return the combined state. 4650 set_control( _gvn.transform(result_reg)); 4651 set_i_o( _gvn.transform(result_i_o)); 4652 set_all_memory( _gvn.transform(result_mem)); 4653 } // original reexecute is set back here 4654 4655 set_result(_gvn.transform(result_val)); 4656 return true; 4657 } 4658 4659 //------------------------------basictype2arraycopy---------------------------- 4660 address LibraryCallKit::basictype2arraycopy(BasicType t, 4661 Node* src_offset, 4662 Node* dest_offset, 4663 bool disjoint_bases, 4664 const char* &name, 4665 bool dest_uninitialized) { 4666 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4667 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4668 4669 bool aligned = false; 4670 bool disjoint = disjoint_bases; 4671 4672 // if the offsets are the same, we can treat the memory regions as 4673 // disjoint, because either the memory regions are in different arrays, 4674 // or they are identical (which we can treat as disjoint.) We can also 4675 // treat a copy with a destination index less that the source index 4676 // as disjoint since a low->high copy will work correctly in this case. 4677 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4678 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4679 // both indices are constants 4680 int s_offs = src_offset_inttype->get_con(); 4681 int d_offs = dest_offset_inttype->get_con(); 4682 int element_size = type2aelembytes(t); 4683 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4684 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4685 if (s_offs >= d_offs) disjoint = true; 4686 } else if (src_offset == dest_offset && src_offset != NULL) { 4687 // This can occur if the offsets are identical non-constants. 4688 disjoint = true; 4689 } 4690 4691 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4692 } 4693 4694 4695 //------------------------------inline_arraycopy----------------------- 4696 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4697 // Object dest, int destPos, 4698 // int length); 4699 bool LibraryCallKit::inline_arraycopy() { 4700 // Get the arguments. 4701 Node* src = argument(0); // type: oop 4702 Node* src_offset = argument(1); // type: int 4703 Node* dest = argument(2); // type: oop 4704 Node* dest_offset = argument(3); // type: int 4705 Node* length = argument(4); // type: int 4706 4707 // Compile time checks. If any of these checks cannot be verified at compile time, 4708 // we do not make a fast path for this call. Instead, we let the call remain as it 4709 // is. The checks we choose to mandate at compile time are: 4710 // 4711 // (1) src and dest are arrays. 4712 const Type* src_type = src->Value(&_gvn); 4713 const Type* dest_type = dest->Value(&_gvn); 4714 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4715 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4716 4717 // Do we have the type of src? 4718 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4719 // Do we have the type of dest? 4720 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4721 // Is the type for src from speculation? 4722 bool src_spec = false; 4723 // Is the type for dest from speculation? 4724 bool dest_spec = false; 4725 4726 if (!has_src || !has_dest) { 4727 // We don't have sufficient type information, let's see if 4728 // speculative types can help. We need to have types for both src 4729 // and dest so that it pays off. 4730 4731 // Do we already have or could we have type information for src 4732 bool could_have_src = has_src; 4733 // Do we already have or could we have type information for dest 4734 bool could_have_dest = has_dest; 4735 4736 ciKlass* src_k = NULL; 4737 if (!has_src) { 4738 src_k = src_type->speculative_type_not_null(); 4739 if (src_k != NULL && src_k->is_array_klass()) { 4740 could_have_src = true; 4741 } 4742 } 4743 4744 ciKlass* dest_k = NULL; 4745 if (!has_dest) { 4746 dest_k = dest_type->speculative_type_not_null(); 4747 if (dest_k != NULL && dest_k->is_array_klass()) { 4748 could_have_dest = true; 4749 } 4750 } 4751 4752 if (could_have_src && could_have_dest) { 4753 // This is going to pay off so emit the required guards 4754 if (!has_src) { 4755 src = maybe_cast_profiled_obj(src, src_k); 4756 src_type = _gvn.type(src); 4757 top_src = src_type->isa_aryptr(); 4758 has_src = (top_src != NULL && top_src->klass() != NULL); 4759 src_spec = true; 4760 } 4761 if (!has_dest) { 4762 dest = maybe_cast_profiled_obj(dest, dest_k); 4763 dest_type = _gvn.type(dest); 4764 top_dest = dest_type->isa_aryptr(); 4765 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4766 dest_spec = true; 4767 } 4768 } 4769 } 4770 4771 if (!has_src || !has_dest) { 4772 // Conservatively insert a memory barrier on all memory slices. 4773 // Do not let writes into the source float below the arraycopy. 4774 insert_mem_bar(Op_MemBarCPUOrder); 4775 4776 // Call StubRoutines::generic_arraycopy stub. 4777 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4778 src, src_offset, dest, dest_offset, length); 4779 4780 // Do not let reads from the destination float above the arraycopy. 4781 // Since we cannot type the arrays, we don't know which slices 4782 // might be affected. We could restrict this barrier only to those 4783 // memory slices which pertain to array elements--but don't bother. 4784 if (!InsertMemBarAfterArraycopy) 4785 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4786 insert_mem_bar(Op_MemBarCPUOrder); 4787 return true; 4788 } 4789 4790 // (2) src and dest arrays must have elements of the same BasicType 4791 // Figure out the size and type of the elements we will be copying. 4792 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4793 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4794 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4795 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4796 4797 if (src_elem != dest_elem || dest_elem == T_VOID) { 4798 // The component types are not the same or are not recognized. Punt. 4799 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4800 generate_slow_arraycopy(TypePtr::BOTTOM, 4801 src, src_offset, dest, dest_offset, length, 4802 /*dest_uninitialized*/false); 4803 return true; 4804 } 4805 4806 if (src_elem == T_OBJECT) { 4807 // If both arrays are object arrays then having the exact types 4808 // for both will remove the need for a subtype check at runtime 4809 // before the call and may make it possible to pick a faster copy 4810 // routine (without a subtype check on every element) 4811 // Do we have the exact type of src? 4812 bool could_have_src = src_spec; 4813 // Do we have the exact type of dest? 4814 bool could_have_dest = dest_spec; 4815 ciKlass* src_k = top_src->klass(); 4816 ciKlass* dest_k = top_dest->klass(); 4817 if (!src_spec) { 4818 src_k = src_type->speculative_type_not_null(); 4819 if (src_k != NULL && src_k->is_array_klass()) { 4820 could_have_src = true; 4821 } 4822 } 4823 if (!dest_spec) { 4824 dest_k = dest_type->speculative_type_not_null(); 4825 if (dest_k != NULL && dest_k->is_array_klass()) { 4826 could_have_dest = true; 4827 } 4828 } 4829 if (could_have_src && could_have_dest) { 4830 // If we can have both exact types, emit the missing guards 4831 if (could_have_src && !src_spec) { 4832 src = maybe_cast_profiled_obj(src, src_k); 4833 } 4834 if (could_have_dest && !dest_spec) { 4835 dest = maybe_cast_profiled_obj(dest, dest_k); 4836 } 4837 } 4838 } 4839 4840 //--------------------------------------------------------------------------- 4841 // We will make a fast path for this call to arraycopy. 4842 4843 // We have the following tests left to perform: 4844 // 4845 // (3) src and dest must not be null. 4846 // (4) src_offset must not be negative. 4847 // (5) dest_offset must not be negative. 4848 // (6) length must not be negative. 4849 // (7) src_offset + length must not exceed length of src. 4850 // (8) dest_offset + length must not exceed length of dest. 4851 // (9) each element of an oop array must be assignable 4852 4853 RegionNode* slow_region = new RegionNode(1); 4854 record_for_igvn(slow_region); 4855 4856 // (3) operands must not be null 4857 // We currently perform our null checks with the null_check routine. 4858 // This means that the null exceptions will be reported in the caller 4859 // rather than (correctly) reported inside of the native arraycopy call. 4860 // This should be corrected, given time. We do our null check with the 4861 // stack pointer restored. 4862 src = null_check(src, T_ARRAY); 4863 dest = null_check(dest, T_ARRAY); 4864 4865 // (4) src_offset must not be negative. 4866 generate_negative_guard(src_offset, slow_region); 4867 4868 // (5) dest_offset must not be negative. 4869 generate_negative_guard(dest_offset, slow_region); 4870 4871 // (6) length must not be negative (moved to generate_arraycopy()). 4872 // generate_negative_guard(length, slow_region); 4873 4874 // (7) src_offset + length must not exceed length of src. 4875 generate_limit_guard(src_offset, length, 4876 load_array_length(src), 4877 slow_region); 4878 4879 // (8) dest_offset + length must not exceed length of dest. 4880 generate_limit_guard(dest_offset, length, 4881 load_array_length(dest), 4882 slow_region); 4883 4884 // (9) each element of an oop array must be assignable 4885 // The generate_arraycopy subroutine checks this. 4886 4887 // This is where the memory effects are placed: 4888 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4889 generate_arraycopy(adr_type, dest_elem, 4890 src, src_offset, dest, dest_offset, length, 4891 false, false, slow_region); 4892 4893 return true; 4894 } 4895 4896 //-----------------------------generate_arraycopy---------------------- 4897 // Generate an optimized call to arraycopy. 4898 // Caller must guard against non-arrays. 4899 // Caller must determine a common array basic-type for both arrays. 4900 // Caller must validate offsets against array bounds. 4901 // The slow_region has already collected guard failure paths 4902 // (such as out of bounds length or non-conformable array types). 4903 // The generated code has this shape, in general: 4904 // 4905 // if (length == 0) return // via zero_path 4906 // slowval = -1 4907 // if (types unknown) { 4908 // slowval = call generic copy loop 4909 // if (slowval == 0) return // via checked_path 4910 // } else if (indexes in bounds) { 4911 // if ((is object array) && !(array type check)) { 4912 // slowval = call checked copy loop 4913 // if (slowval == 0) return // via checked_path 4914 // } else { 4915 // call bulk copy loop 4916 // return // via fast_path 4917 // } 4918 // } 4919 // // adjust params for remaining work: 4920 // if (slowval != -1) { 4921 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4922 // } 4923 // slow_region: 4924 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4925 // return // via slow_call_path 4926 // 4927 // This routine is used from several intrinsics: System.arraycopy, 4928 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4929 // 4930 void 4931 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4932 BasicType basic_elem_type, 4933 Node* src, Node* src_offset, 4934 Node* dest, Node* dest_offset, 4935 Node* copy_length, 4936 bool disjoint_bases, 4937 bool length_never_negative, 4938 RegionNode* slow_region) { 4939 4940 if (slow_region == NULL) { 4941 slow_region = new RegionNode(1); 4942 record_for_igvn(slow_region); 4943 } 4944 4945 Node* original_dest = dest; 4946 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4947 bool dest_uninitialized = false; 4948 4949 // See if this is the initialization of a newly-allocated array. 4950 // If so, we will take responsibility here for initializing it to zero. 4951 // (Note: Because tightly_coupled_allocation performs checks on the 4952 // out-edges of the dest, we need to avoid making derived pointers 4953 // from it until we have checked its uses.) 4954 if (ReduceBulkZeroing 4955 && !ZeroTLAB // pointless if already zeroed 4956 && basic_elem_type != T_CONFLICT // avoid corner case 4957 && !src->eqv_uncast(dest) 4958 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4959 != NULL) 4960 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4961 && alloc->maybe_set_complete(&_gvn)) { 4962 // "You break it, you buy it." 4963 InitializeNode* init = alloc->initialization(); 4964 assert(init->is_complete(), "we just did this"); 4965 init->set_complete_with_arraycopy(); 4966 assert(dest->is_CheckCastPP(), "sanity"); 4967 assert(dest->in(0)->in(0) == init, "dest pinned"); 4968 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4969 // From this point on, every exit path is responsible for 4970 // initializing any non-copied parts of the object to zero. 4971 // Also, if this flag is set we make sure that arraycopy interacts properly 4972 // with G1, eliding pre-barriers. See CR 6627983. 4973 dest_uninitialized = true; 4974 } else { 4975 // No zeroing elimination here. 4976 alloc = NULL; 4977 //original_dest = dest; 4978 //dest_uninitialized = false; 4979 } 4980 4981 // Results are placed here: 4982 enum { fast_path = 1, // normal void-returning assembly stub 4983 checked_path = 2, // special assembly stub with cleanup 4984 slow_call_path = 3, // something went wrong; call the VM 4985 zero_path = 4, // bypass when length of copy is zero 4986 bcopy_path = 5, // copy primitive array by 64-bit blocks 4987 PATH_LIMIT = 6 4988 }; 4989 RegionNode* result_region = new RegionNode(PATH_LIMIT); 4990 PhiNode* result_i_o = new PhiNode(result_region, Type::ABIO); 4991 PhiNode* result_memory = new PhiNode(result_region, Type::MEMORY, adr_type); 4992 record_for_igvn(result_region); 4993 _gvn.set_type_bottom(result_i_o); 4994 _gvn.set_type_bottom(result_memory); 4995 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4996 4997 // The slow_control path: 4998 Node* slow_control; 4999 Node* slow_i_o = i_o(); 5000 Node* slow_mem = memory(adr_type); 5001 debug_only(slow_control = (Node*) badAddress); 5002 5003 // Checked control path: 5004 Node* checked_control = top(); 5005 Node* checked_mem = NULL; 5006 Node* checked_i_o = NULL; 5007 Node* checked_value = NULL; 5008 5009 if (basic_elem_type == T_CONFLICT) { 5010 assert(!dest_uninitialized, ""); 5011 Node* cv = generate_generic_arraycopy(adr_type, 5012 src, src_offset, dest, dest_offset, 5013 copy_length, dest_uninitialized); 5014 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 5015 checked_control = control(); 5016 checked_i_o = i_o(); 5017 checked_mem = memory(adr_type); 5018 checked_value = cv; 5019 set_control(top()); // no fast path 5020 } 5021 5022 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 5023 if (not_pos != NULL) { 5024 PreserveJVMState pjvms(this); 5025 set_control(not_pos); 5026 5027 // (6) length must not be negative. 5028 if (!length_never_negative) { 5029 generate_negative_guard(copy_length, slow_region); 5030 } 5031 5032 // copy_length is 0. 5033 if (!stopped() && dest_uninitialized) { 5034 Node* dest_length = alloc->in(AllocateNode::ALength); 5035 if (copy_length->eqv_uncast(dest_length) 5036 || _gvn.find_int_con(dest_length, 1) <= 0) { 5037 // There is no zeroing to do. No need for a secondary raw memory barrier. 5038 } else { 5039 // Clear the whole thing since there are no source elements to copy. 5040 generate_clear_array(adr_type, dest, basic_elem_type, 5041 intcon(0), NULL, 5042 alloc->in(AllocateNode::AllocSize)); 5043 // Use a secondary InitializeNode as raw memory barrier. 5044 // Currently it is needed only on this path since other 5045 // paths have stub or runtime calls as raw memory barriers. 5046 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 5047 Compile::AliasIdxRaw, 5048 top())->as_Initialize(); 5049 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 5050 } 5051 } 5052 5053 // Present the results of the fast call. 5054 result_region->init_req(zero_path, control()); 5055 result_i_o ->init_req(zero_path, i_o()); 5056 result_memory->init_req(zero_path, memory(adr_type)); 5057 } 5058 5059 if (!stopped() && dest_uninitialized) { 5060 // We have to initialize the *uncopied* part of the array to zero. 5061 // The copy destination is the slice dest[off..off+len]. The other slices 5062 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 5063 Node* dest_size = alloc->in(AllocateNode::AllocSize); 5064 Node* dest_length = alloc->in(AllocateNode::ALength); 5065 Node* dest_tail = _gvn.transform(new AddINode(dest_offset, copy_length)); 5066 5067 // If there is a head section that needs zeroing, do it now. 5068 if (find_int_con(dest_offset, -1) != 0) { 5069 generate_clear_array(adr_type, dest, basic_elem_type, 5070 intcon(0), dest_offset, 5071 NULL); 5072 } 5073 5074 // Next, perform a dynamic check on the tail length. 5075 // It is often zero, and we can win big if we prove this. 5076 // There are two wins: Avoid generating the ClearArray 5077 // with its attendant messy index arithmetic, and upgrade 5078 // the copy to a more hardware-friendly word size of 64 bits. 5079 Node* tail_ctl = NULL; 5080 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 5081 Node* cmp_lt = _gvn.transform(new CmpINode(dest_tail, dest_length)); 5082 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 5083 tail_ctl = generate_slow_guard(bol_lt, NULL); 5084 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 5085 } 5086 5087 // At this point, let's assume there is no tail. 5088 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 5089 // There is no tail. Try an upgrade to a 64-bit copy. 5090 bool didit = false; 5091 { PreserveJVMState pjvms(this); 5092 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 5093 src, src_offset, dest, dest_offset, 5094 dest_size, dest_uninitialized); 5095 if (didit) { 5096 // Present the results of the block-copying fast call. 5097 result_region->init_req(bcopy_path, control()); 5098 result_i_o ->init_req(bcopy_path, i_o()); 5099 result_memory->init_req(bcopy_path, memory(adr_type)); 5100 } 5101 } 5102 if (didit) 5103 set_control(top()); // no regular fast path 5104 } 5105 5106 // Clear the tail, if any. 5107 if (tail_ctl != NULL) { 5108 Node* notail_ctl = stopped() ? NULL : control(); 5109 set_control(tail_ctl); 5110 if (notail_ctl == NULL) { 5111 generate_clear_array(adr_type, dest, basic_elem_type, 5112 dest_tail, NULL, 5113 dest_size); 5114 } else { 5115 // Make a local merge. 5116 Node* done_ctl = new RegionNode(3); 5117 Node* done_mem = new PhiNode(done_ctl, Type::MEMORY, adr_type); 5118 done_ctl->init_req(1, notail_ctl); 5119 done_mem->init_req(1, memory(adr_type)); 5120 generate_clear_array(adr_type, dest, basic_elem_type, 5121 dest_tail, NULL, 5122 dest_size); 5123 done_ctl->init_req(2, control()); 5124 done_mem->init_req(2, memory(adr_type)); 5125 set_control( _gvn.transform(done_ctl)); 5126 set_memory( _gvn.transform(done_mem), adr_type ); 5127 } 5128 } 5129 } 5130 5131 BasicType copy_type = basic_elem_type; 5132 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 5133 if (!stopped() && copy_type == T_OBJECT) { 5134 // If src and dest have compatible element types, we can copy bits. 5135 // Types S[] and D[] are compatible if D is a supertype of S. 5136 // 5137 // If they are not, we will use checked_oop_disjoint_arraycopy, 5138 // which performs a fast optimistic per-oop check, and backs off 5139 // further to JVM_ArrayCopy on the first per-oop check that fails. 5140 // (Actually, we don't move raw bits only; the GC requires card marks.) 5141 5142 // Get the Klass* for both src and dest 5143 Node* src_klass = load_object_klass(src); 5144 Node* dest_klass = load_object_klass(dest); 5145 5146 // Generate the subtype check. 5147 // This might fold up statically, or then again it might not. 5148 // 5149 // Non-static example: Copying List<String>.elements to a new String[]. 5150 // The backing store for a List<String> is always an Object[], 5151 // but its elements are always type String, if the generic types 5152 // are correct at the source level. 5153 // 5154 // Test S[] against D[], not S against D, because (probably) 5155 // the secondary supertype cache is less busy for S[] than S. 5156 // This usually only matters when D is an interface. 5157 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5158 // Plug failing path into checked_oop_disjoint_arraycopy 5159 if (not_subtype_ctrl != top()) { 5160 PreserveJVMState pjvms(this); 5161 set_control(not_subtype_ctrl); 5162 // (At this point we can assume disjoint_bases, since types differ.) 5163 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 5164 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 5165 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 5166 Node* dest_elem_klass = _gvn.transform(n1); 5167 Node* cv = generate_checkcast_arraycopy(adr_type, 5168 dest_elem_klass, 5169 src, src_offset, dest, dest_offset, 5170 ConvI2X(copy_length), dest_uninitialized); 5171 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 5172 checked_control = control(); 5173 checked_i_o = i_o(); 5174 checked_mem = memory(adr_type); 5175 checked_value = cv; 5176 } 5177 // At this point we know we do not need type checks on oop stores. 5178 5179 // Let's see if we need card marks: 5180 if (alloc != NULL && use_ReduceInitialCardMarks()) { 5181 // If we do not need card marks, copy using the jint or jlong stub. 5182 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 5183 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 5184 "sizes agree"); 5185 } 5186 } 5187 5188 if (!stopped()) { 5189 // Generate the fast path, if possible. 5190 PreserveJVMState pjvms(this); 5191 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 5192 src, src_offset, dest, dest_offset, 5193 ConvI2X(copy_length), dest_uninitialized); 5194 5195 // Present the results of the fast call. 5196 result_region->init_req(fast_path, control()); 5197 result_i_o ->init_req(fast_path, i_o()); 5198 result_memory->init_req(fast_path, memory(adr_type)); 5199 } 5200 5201 // Here are all the slow paths up to this point, in one bundle: 5202 slow_control = top(); 5203 if (slow_region != NULL) 5204 slow_control = _gvn.transform(slow_region); 5205 DEBUG_ONLY(slow_region = (RegionNode*)badAddress); 5206 5207 set_control(checked_control); 5208 if (!stopped()) { 5209 // Clean up after the checked call. 5210 // The returned value is either 0 or -1^K, 5211 // where K = number of partially transferred array elements. 5212 Node* cmp = _gvn.transform(new CmpINode(checked_value, intcon(0))); 5213 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 5214 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 5215 5216 // If it is 0, we are done, so transfer to the end. 5217 Node* checks_done = _gvn.transform(new IfTrueNode(iff)); 5218 result_region->init_req(checked_path, checks_done); 5219 result_i_o ->init_req(checked_path, checked_i_o); 5220 result_memory->init_req(checked_path, checked_mem); 5221 5222 // If it is not zero, merge into the slow call. 5223 set_control( _gvn.transform(new IfFalseNode(iff) )); 5224 RegionNode* slow_reg2 = new RegionNode(3); 5225 PhiNode* slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO); 5226 PhiNode* slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type); 5227 record_for_igvn(slow_reg2); 5228 slow_reg2 ->init_req(1, slow_control); 5229 slow_i_o2 ->init_req(1, slow_i_o); 5230 slow_mem2 ->init_req(1, slow_mem); 5231 slow_reg2 ->init_req(2, control()); 5232 slow_i_o2 ->init_req(2, checked_i_o); 5233 slow_mem2 ->init_req(2, checked_mem); 5234 5235 slow_control = _gvn.transform(slow_reg2); 5236 slow_i_o = _gvn.transform(slow_i_o2); 5237 slow_mem = _gvn.transform(slow_mem2); 5238 5239 if (alloc != NULL) { 5240 // We'll restart from the very beginning, after zeroing the whole thing. 5241 // This can cause double writes, but that's OK since dest is brand new. 5242 // So we ignore the low 31 bits of the value returned from the stub. 5243 } else { 5244 // We must continue the copy exactly where it failed, or else 5245 // another thread might see the wrong number of writes to dest. 5246 Node* checked_offset = _gvn.transform(new XorINode(checked_value, intcon(-1))); 5247 Node* slow_offset = new PhiNode(slow_reg2, TypeInt::INT); 5248 slow_offset->init_req(1, intcon(0)); 5249 slow_offset->init_req(2, checked_offset); 5250 slow_offset = _gvn.transform(slow_offset); 5251 5252 // Adjust the arguments by the conditionally incoming offset. 5253 Node* src_off_plus = _gvn.transform(new AddINode(src_offset, slow_offset)); 5254 Node* dest_off_plus = _gvn.transform(new AddINode(dest_offset, slow_offset)); 5255 Node* length_minus = _gvn.transform(new SubINode(copy_length, slow_offset)); 5256 5257 // Tweak the node variables to adjust the code produced below: 5258 src_offset = src_off_plus; 5259 dest_offset = dest_off_plus; 5260 copy_length = length_minus; 5261 } 5262 } 5263 5264 set_control(slow_control); 5265 if (!stopped()) { 5266 // Generate the slow path, if needed. 5267 PreserveJVMState pjvms(this); // replace_in_map may trash the map 5268 5269 set_memory(slow_mem, adr_type); 5270 set_i_o(slow_i_o); 5271 5272 if (dest_uninitialized) { 5273 generate_clear_array(adr_type, dest, basic_elem_type, 5274 intcon(0), NULL, 5275 alloc->in(AllocateNode::AllocSize)); 5276 } 5277 5278 generate_slow_arraycopy(adr_type, 5279 src, src_offset, dest, dest_offset, 5280 copy_length, /*dest_uninitialized*/false); 5281 5282 result_region->init_req(slow_call_path, control()); 5283 result_i_o ->init_req(slow_call_path, i_o()); 5284 result_memory->init_req(slow_call_path, memory(adr_type)); 5285 } 5286 5287 // Remove unused edges. 5288 for (uint i = 1; i < result_region->req(); i++) { 5289 if (result_region->in(i) == NULL) 5290 result_region->init_req(i, top()); 5291 } 5292 5293 // Finished; return the combined state. 5294 set_control( _gvn.transform(result_region)); 5295 set_i_o( _gvn.transform(result_i_o) ); 5296 set_memory( _gvn.transform(result_memory), adr_type ); 5297 5298 // The memory edges above are precise in order to model effects around 5299 // array copies accurately to allow value numbering of field loads around 5300 // arraycopy. Such field loads, both before and after, are common in Java 5301 // collections and similar classes involving header/array data structures. 5302 // 5303 // But with low number of register or when some registers are used or killed 5304 // by arraycopy calls it causes registers spilling on stack. See 6544710. 5305 // The next memory barrier is added to avoid it. If the arraycopy can be 5306 // optimized away (which it can, sometimes) then we can manually remove 5307 // the membar also. 5308 // 5309 // Do not let reads from the cloned object float above the arraycopy. 5310 if (alloc != NULL) { 5311 // Do not let stores that initialize this object be reordered with 5312 // a subsequent store that would make this object accessible by 5313 // other threads. 5314 // Record what AllocateNode this StoreStore protects so that 5315 // escape analysis can go from the MemBarStoreStoreNode to the 5316 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5317 // based on the escape status of the AllocateNode. 5318 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5319 } else if (InsertMemBarAfterArraycopy) 5320 insert_mem_bar(Op_MemBarCPUOrder); 5321 } 5322 5323 5324 // Helper function which determines if an arraycopy immediately follows 5325 // an allocation, with no intervening tests or other escapes for the object. 5326 AllocateArrayNode* 5327 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5328 RegionNode* slow_region) { 5329 if (stopped()) return NULL; // no fast path 5330 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5331 5332 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5333 if (alloc == NULL) return NULL; 5334 5335 Node* rawmem = memory(Compile::AliasIdxRaw); 5336 // Is the allocation's memory state untouched? 5337 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5338 // Bail out if there have been raw-memory effects since the allocation. 5339 // (Example: There might have been a call or safepoint.) 5340 return NULL; 5341 } 5342 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5343 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5344 return NULL; 5345 } 5346 5347 // There must be no unexpected observers of this allocation. 5348 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5349 Node* obs = ptr->fast_out(i); 5350 if (obs != this->map()) { 5351 return NULL; 5352 } 5353 } 5354 5355 // This arraycopy must unconditionally follow the allocation of the ptr. 5356 Node* alloc_ctl = ptr->in(0); 5357 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5358 5359 Node* ctl = control(); 5360 while (ctl != alloc_ctl) { 5361 // There may be guards which feed into the slow_region. 5362 // Any other control flow means that we might not get a chance 5363 // to finish initializing the allocated object. 5364 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5365 IfNode* iff = ctl->in(0)->as_If(); 5366 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5367 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5368 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5369 ctl = iff->in(0); // This test feeds the known slow_region. 5370 continue; 5371 } 5372 // One more try: Various low-level checks bottom out in 5373 // uncommon traps. If the debug-info of the trap omits 5374 // any reference to the allocation, as we've already 5375 // observed, then there can be no objection to the trap. 5376 bool found_trap = false; 5377 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5378 Node* obs = not_ctl->fast_out(j); 5379 if (obs->in(0) == not_ctl && obs->is_Call() && 5380 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5381 found_trap = true; break; 5382 } 5383 } 5384 if (found_trap) { 5385 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5386 continue; 5387 } 5388 } 5389 return NULL; 5390 } 5391 5392 // If we get this far, we have an allocation which immediately 5393 // precedes the arraycopy, and we can take over zeroing the new object. 5394 // The arraycopy will finish the initialization, and provide 5395 // a new control state to which we will anchor the destination pointer. 5396 5397 return alloc; 5398 } 5399 5400 // Helper for initialization of arrays, creating a ClearArray. 5401 // It writes zero bits in [start..end), within the body of an array object. 5402 // The memory effects are all chained onto the 'adr_type' alias category. 5403 // 5404 // Since the object is otherwise uninitialized, we are free 5405 // to put a little "slop" around the edges of the cleared area, 5406 // as long as it does not go back into the array's header, 5407 // or beyond the array end within the heap. 5408 // 5409 // The lower edge can be rounded down to the nearest jint and the 5410 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5411 // 5412 // Arguments: 5413 // adr_type memory slice where writes are generated 5414 // dest oop of the destination array 5415 // basic_elem_type element type of the destination 5416 // slice_idx array index of first element to store 5417 // slice_len number of elements to store (or NULL) 5418 // dest_size total size in bytes of the array object 5419 // 5420 // Exactly one of slice_len or dest_size must be non-NULL. 5421 // If dest_size is non-NULL, zeroing extends to the end of the object. 5422 // If slice_len is non-NULL, the slice_idx value must be a constant. 5423 void 5424 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5425 Node* dest, 5426 BasicType basic_elem_type, 5427 Node* slice_idx, 5428 Node* slice_len, 5429 Node* dest_size) { 5430 // one or the other but not both of slice_len and dest_size: 5431 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5432 if (slice_len == NULL) slice_len = top(); 5433 if (dest_size == NULL) dest_size = top(); 5434 5435 // operate on this memory slice: 5436 Node* mem = memory(adr_type); // memory slice to operate on 5437 5438 // scaling and rounding of indexes: 5439 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5440 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5441 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5442 int bump_bit = (-1 << scale) & BytesPerInt; 5443 5444 // determine constant starts and ends 5445 const intptr_t BIG_NEG = -128; 5446 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5447 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5448 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5449 if (slice_len_con == 0) { 5450 return; // nothing to do here 5451 } 5452 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5453 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5454 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5455 assert(end_con < 0, "not two cons"); 5456 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5457 BytesPerLong); 5458 } 5459 5460 if (start_con >= 0 && end_con >= 0) { 5461 // Constant start and end. Simple. 5462 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5463 start_con, end_con, &_gvn); 5464 } else if (start_con >= 0 && dest_size != top()) { 5465 // Constant start, pre-rounded end after the tail of the array. 5466 Node* end = dest_size; 5467 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5468 start_con, end, &_gvn); 5469 } else if (start_con >= 0 && slice_len != top()) { 5470 // Constant start, non-constant end. End needs rounding up. 5471 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5472 intptr_t end_base = abase + (slice_idx_con << scale); 5473 int end_round = (-1 << scale) & (BytesPerLong - 1); 5474 Node* end = ConvI2X(slice_len); 5475 if (scale != 0) 5476 end = _gvn.transform(new LShiftXNode(end, intcon(scale) )); 5477 end_base += end_round; 5478 end = _gvn.transform(new AddXNode(end, MakeConX(end_base))); 5479 end = _gvn.transform(new AndXNode(end, MakeConX(~end_round))); 5480 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5481 start_con, end, &_gvn); 5482 } else if (start_con < 0 && dest_size != top()) { 5483 // Non-constant start, pre-rounded end after the tail of the array. 5484 // This is almost certainly a "round-to-end" operation. 5485 Node* start = slice_idx; 5486 start = ConvI2X(start); 5487 if (scale != 0) 5488 start = _gvn.transform(new LShiftXNode( start, intcon(scale) )); 5489 start = _gvn.transform(new AddXNode(start, MakeConX(abase))); 5490 if ((bump_bit | clear_low) != 0) { 5491 int to_clear = (bump_bit | clear_low); 5492 // Align up mod 8, then store a jint zero unconditionally 5493 // just before the mod-8 boundary. 5494 if (((abase + bump_bit) & ~to_clear) - bump_bit 5495 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5496 bump_bit = 0; 5497 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5498 } else { 5499 // Bump 'start' up to (or past) the next jint boundary: 5500 start = _gvn.transform(new AddXNode(start, MakeConX(bump_bit))); 5501 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5502 } 5503 // Round bumped 'start' down to jlong boundary in body of array. 5504 start = _gvn.transform(new AndXNode(start, MakeConX(~to_clear))); 5505 if (bump_bit != 0) { 5506 // Store a zero to the immediately preceding jint: 5507 Node* x1 = _gvn.transform(new AddXNode(start, MakeConX(-bump_bit))); 5508 Node* p1 = basic_plus_adr(dest, x1); 5509 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered); 5510 mem = _gvn.transform(mem); 5511 } 5512 } 5513 Node* end = dest_size; // pre-rounded 5514 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5515 start, end, &_gvn); 5516 } else { 5517 // Non-constant start, unrounded non-constant end. 5518 // (Nobody zeroes a random midsection of an array using this routine.) 5519 ShouldNotReachHere(); // fix caller 5520 } 5521 5522 // Done. 5523 set_memory(mem, adr_type); 5524 } 5525 5526 5527 bool 5528 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5529 BasicType basic_elem_type, 5530 AllocateNode* alloc, 5531 Node* src, Node* src_offset, 5532 Node* dest, Node* dest_offset, 5533 Node* dest_size, bool dest_uninitialized) { 5534 // See if there is an advantage from block transfer. 5535 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5536 if (scale >= LogBytesPerLong) 5537 return false; // it is already a block transfer 5538 5539 // Look at the alignment of the starting offsets. 5540 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5541 5542 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5543 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5544 if (src_off_con < 0 || dest_off_con < 0) 5545 // At present, we can only understand constants. 5546 return false; 5547 5548 intptr_t src_off = abase + (src_off_con << scale); 5549 intptr_t dest_off = abase + (dest_off_con << scale); 5550 5551 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5552 // Non-aligned; too bad. 5553 // One more chance: Pick off an initial 32-bit word. 5554 // This is a common case, since abase can be odd mod 8. 5555 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5556 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5557 Node* sptr = basic_plus_adr(src, src_off); 5558 Node* dptr = basic_plus_adr(dest, dest_off); 5559 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 5560 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered); 5561 src_off += BytesPerInt; 5562 dest_off += BytesPerInt; 5563 } else { 5564 return false; 5565 } 5566 } 5567 assert(src_off % BytesPerLong == 0, ""); 5568 assert(dest_off % BytesPerLong == 0, ""); 5569 5570 // Do this copy by giant steps. 5571 Node* sptr = basic_plus_adr(src, src_off); 5572 Node* dptr = basic_plus_adr(dest, dest_off); 5573 Node* countx = dest_size; 5574 countx = _gvn.transform(new SubXNode(countx, MakeConX(dest_off))); 5575 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong))); 5576 5577 bool disjoint_bases = true; // since alloc != NULL 5578 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5579 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5580 5581 return true; 5582 } 5583 5584 5585 // Helper function; generates code for the slow case. 5586 // We make a call to a runtime method which emulates the native method, 5587 // but without the native wrapper overhead. 5588 void 5589 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5590 Node* src, Node* src_offset, 5591 Node* dest, Node* dest_offset, 5592 Node* copy_length, bool dest_uninitialized) { 5593 assert(!dest_uninitialized, "Invariant"); 5594 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5595 OptoRuntime::slow_arraycopy_Type(), 5596 OptoRuntime::slow_arraycopy_Java(), 5597 "slow_arraycopy", adr_type, 5598 src, src_offset, dest, dest_offset, 5599 copy_length); 5600 5601 // Handle exceptions thrown by this fellow: 5602 make_slow_call_ex(call, env()->Throwable_klass(), false); 5603 } 5604 5605 // Helper function; generates code for cases requiring runtime checks. 5606 Node* 5607 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5608 Node* dest_elem_klass, 5609 Node* src, Node* src_offset, 5610 Node* dest, Node* dest_offset, 5611 Node* copy_length, bool dest_uninitialized) { 5612 if (stopped()) return NULL; 5613 5614 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5615 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5616 return NULL; 5617 } 5618 5619 // Pick out the parameters required to perform a store-check 5620 // for the target array. This is an optimistic check. It will 5621 // look in each non-null element's class, at the desired klass's 5622 // super_check_offset, for the desired klass. 5623 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5624 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5625 Node* n3 = new LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered); 5626 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5627 Node* check_value = dest_elem_klass; 5628 5629 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5630 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5631 5632 // (We know the arrays are never conjoint, because their types differ.) 5633 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5634 OptoRuntime::checkcast_arraycopy_Type(), 5635 copyfunc_addr, "checkcast_arraycopy", adr_type, 5636 // five arguments, of which two are 5637 // intptr_t (jlong in LP64) 5638 src_start, dest_start, 5639 copy_length XTOP, 5640 check_offset XTOP, 5641 check_value); 5642 5643 return _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5644 } 5645 5646 5647 // Helper function; generates code for cases requiring runtime checks. 5648 Node* 5649 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5650 Node* src, Node* src_offset, 5651 Node* dest, Node* dest_offset, 5652 Node* copy_length, bool dest_uninitialized) { 5653 assert(!dest_uninitialized, "Invariant"); 5654 if (stopped()) return NULL; 5655 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5656 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5657 return NULL; 5658 } 5659 5660 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5661 OptoRuntime::generic_arraycopy_Type(), 5662 copyfunc_addr, "generic_arraycopy", adr_type, 5663 src, src_offset, dest, dest_offset, copy_length); 5664 5665 return _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5666 } 5667 5668 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5669 void 5670 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5671 BasicType basic_elem_type, 5672 bool disjoint_bases, 5673 Node* src, Node* src_offset, 5674 Node* dest, Node* dest_offset, 5675 Node* copy_length, bool dest_uninitialized) { 5676 if (stopped()) return; // nothing to do 5677 5678 Node* src_start = src; 5679 Node* dest_start = dest; 5680 if (src_offset != NULL || dest_offset != NULL) { 5681 assert(src_offset != NULL && dest_offset != NULL, ""); 5682 src_start = array_element_address(src, src_offset, basic_elem_type); 5683 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5684 } 5685 5686 // Figure out which arraycopy runtime method to call. 5687 const char* copyfunc_name = "arraycopy"; 5688 address copyfunc_addr = 5689 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5690 disjoint_bases, copyfunc_name, dest_uninitialized); 5691 5692 // Call it. Note that the count_ix value is not scaled to a byte-size. 5693 make_runtime_call(RC_LEAF|RC_NO_FP, 5694 OptoRuntime::fast_arraycopy_Type(), 5695 copyfunc_addr, copyfunc_name, adr_type, 5696 src_start, dest_start, copy_length XTOP); 5697 } 5698 5699 //-------------inline_encodeISOArray----------------------------------- 5700 // encode char[] to byte[] in ISO_8859_1 5701 bool LibraryCallKit::inline_encodeISOArray() { 5702 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5703 // no receiver since it is static method 5704 Node *src = argument(0); 5705 Node *src_offset = argument(1); 5706 Node *dst = argument(2); 5707 Node *dst_offset = argument(3); 5708 Node *length = argument(4); 5709 5710 const Type* src_type = src->Value(&_gvn); 5711 const Type* dst_type = dst->Value(&_gvn); 5712 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5713 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5714 if (top_src == NULL || top_src->klass() == NULL || 5715 top_dest == NULL || top_dest->klass() == NULL) { 5716 // failed array check 5717 return false; 5718 } 5719 5720 // Figure out the size and type of the elements we will be copying. 5721 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5722 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5723 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5724 return false; 5725 } 5726 Node* src_start = array_element_address(src, src_offset, src_elem); 5727 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5728 // 'src_start' points to src array + scaled offset 5729 // 'dst_start' points to dst array + scaled offset 5730 5731 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5732 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5733 enc = _gvn.transform(enc); 5734 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5735 set_memory(res_mem, mtype); 5736 set_result(enc); 5737 return true; 5738 } 5739 5740 /** 5741 * Calculate CRC32 for byte. 5742 * int java.util.zip.CRC32.update(int crc, int b) 5743 */ 5744 bool LibraryCallKit::inline_updateCRC32() { 5745 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5746 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5747 // no receiver since it is static method 5748 Node* crc = argument(0); // type: int 5749 Node* b = argument(1); // type: int 5750 5751 /* 5752 * int c = ~ crc; 5753 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5754 * b = b ^ (c >>> 8); 5755 * crc = ~b; 5756 */ 5757 5758 Node* M1 = intcon(-1); 5759 crc = _gvn.transform(new XorINode(crc, M1)); 5760 Node* result = _gvn.transform(new XorINode(crc, b)); 5761 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5762 5763 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5764 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5765 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5766 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5767 5768 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5769 result = _gvn.transform(new XorINode(crc, result)); 5770 result = _gvn.transform(new XorINode(result, M1)); 5771 set_result(result); 5772 return true; 5773 } 5774 5775 /** 5776 * Calculate CRC32 for byte[] array. 5777 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5778 */ 5779 bool LibraryCallKit::inline_updateBytesCRC32() { 5780 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5781 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5782 // no receiver since it is static method 5783 Node* crc = argument(0); // type: int 5784 Node* src = argument(1); // type: oop 5785 Node* offset = argument(2); // type: int 5786 Node* length = argument(3); // type: int 5787 5788 const Type* src_type = src->Value(&_gvn); 5789 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5790 if (top_src == NULL || top_src->klass() == NULL) { 5791 // failed array check 5792 return false; 5793 } 5794 5795 // Figure out the size and type of the elements we will be copying. 5796 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5797 if (src_elem != T_BYTE) { 5798 return false; 5799 } 5800 5801 // 'src_start' points to src array + scaled offset 5802 Node* src_start = array_element_address(src, offset, src_elem); 5803 5804 // We assume that range check is done by caller. 5805 // TODO: generate range check (offset+length < src.length) in debug VM. 5806 5807 // Call the stub. 5808 address stubAddr = StubRoutines::updateBytesCRC32(); 5809 const char *stubName = "updateBytesCRC32"; 5810 5811 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5812 stubAddr, stubName, TypePtr::BOTTOM, 5813 crc, src_start, length); 5814 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5815 set_result(result); 5816 return true; 5817 } 5818 5819 /** 5820 * Calculate CRC32 for ByteBuffer. 5821 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5822 */ 5823 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5824 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5825 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5826 // no receiver since it is static method 5827 Node* crc = argument(0); // type: int 5828 Node* src = argument(1); // type: long 5829 Node* offset = argument(3); // type: int 5830 Node* length = argument(4); // type: int 5831 5832 src = ConvL2X(src); // adjust Java long to machine word 5833 Node* base = _gvn.transform(new CastX2PNode(src)); 5834 offset = ConvI2X(offset); 5835 5836 // 'src_start' points to src array + scaled offset 5837 Node* src_start = basic_plus_adr(top(), base, offset); 5838 5839 // Call the stub. 5840 address stubAddr = StubRoutines::updateBytesCRC32(); 5841 const char *stubName = "updateBytesCRC32"; 5842 5843 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5844 stubAddr, stubName, TypePtr::BOTTOM, 5845 crc, src_start, length); 5846 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5847 set_result(result); 5848 return true; 5849 } 5850 5851 //----------------------------inline_reference_get---------------------------- 5852 // public T java.lang.ref.Reference.get(); 5853 bool LibraryCallKit::inline_reference_get() { 5854 const int referent_offset = java_lang_ref_Reference::referent_offset; 5855 guarantee(referent_offset > 0, "should have already been set"); 5856 5857 // Get the argument: 5858 Node* reference_obj = null_check_receiver(); 5859 if (stopped()) return true; 5860 5861 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5862 5863 ciInstanceKlass* klass = env()->Object_klass(); 5864 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5865 5866 Node* no_ctrl = NULL; 5867 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5868 5869 // Use the pre-barrier to record the value in the referent field 5870 pre_barrier(false /* do_load */, 5871 control(), 5872 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5873 result /* pre_val */, 5874 T_OBJECT); 5875 5876 // Add memory barrier to prevent commoning reads from this field 5877 // across safepoint since GC can change its value. 5878 insert_mem_bar(Op_MemBarCPUOrder); 5879 5880 set_result(result); 5881 return true; 5882 } 5883 5884 5885 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5886 bool is_exact=true, bool is_static=false) { 5887 5888 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5889 assert(tinst != NULL, "obj is null"); 5890 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5891 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5892 5893 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5894 ciSymbol::make(fieldTypeString), 5895 is_static); 5896 if (field == NULL) return (Node *) NULL; 5897 assert (field != NULL, "undefined field"); 5898 5899 // Next code copied from Parse::do_get_xxx(): 5900 5901 // Compute address and memory type. 5902 int offset = field->offset_in_bytes(); 5903 bool is_vol = field->is_volatile(); 5904 ciType* field_klass = field->type(); 5905 assert(field_klass->is_loaded(), "should be loaded"); 5906 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5907 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5908 BasicType bt = field->layout_type(); 5909 5910 // Build the resultant type of the load 5911 const Type *type; 5912 if (bt == T_OBJECT) { 5913 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5914 } else { 5915 type = Type::get_const_basic_type(bt); 5916 } 5917 5918 // Build the load. 5919 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol); 5920 return loadedField; 5921 } 5922 5923 5924 //------------------------------inline_aescrypt_Block----------------------- 5925 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5926 address stubAddr; 5927 const char *stubName; 5928 assert(UseAES, "need AES instruction support"); 5929 5930 switch(id) { 5931 case vmIntrinsics::_aescrypt_encryptBlock: 5932 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5933 stubName = "aescrypt_encryptBlock"; 5934 break; 5935 case vmIntrinsics::_aescrypt_decryptBlock: 5936 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5937 stubName = "aescrypt_decryptBlock"; 5938 break; 5939 } 5940 if (stubAddr == NULL) return false; 5941 5942 Node* aescrypt_object = argument(0); 5943 Node* src = argument(1); 5944 Node* src_offset = argument(2); 5945 Node* dest = argument(3); 5946 Node* dest_offset = argument(4); 5947 5948 // (1) src and dest are arrays. 5949 const Type* src_type = src->Value(&_gvn); 5950 const Type* dest_type = dest->Value(&_gvn); 5951 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5952 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5953 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5954 5955 // for the quick and dirty code we will skip all the checks. 5956 // we are just trying to get the call to be generated. 5957 Node* src_start = src; 5958 Node* dest_start = dest; 5959 if (src_offset != NULL || dest_offset != NULL) { 5960 assert(src_offset != NULL && dest_offset != NULL, ""); 5961 src_start = array_element_address(src, src_offset, T_BYTE); 5962 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5963 } 5964 5965 // now need to get the start of its expanded key array 5966 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5967 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5968 if (k_start == NULL) return false; 5969 5970 if (Matcher::pass_original_key_for_aes()) { 5971 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5972 // compatibility issues between Java key expansion and SPARC crypto instructions 5973 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5974 if (original_k_start == NULL) return false; 5975 5976 // Call the stub. 5977 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5978 stubAddr, stubName, TypePtr::BOTTOM, 5979 src_start, dest_start, k_start, original_k_start); 5980 } else { 5981 // Call the stub. 5982 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5983 stubAddr, stubName, TypePtr::BOTTOM, 5984 src_start, dest_start, k_start); 5985 } 5986 5987 return true; 5988 } 5989 5990 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5991 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5992 address stubAddr; 5993 const char *stubName; 5994 5995 assert(UseAES, "need AES instruction support"); 5996 5997 switch(id) { 5998 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5999 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6000 stubName = "cipherBlockChaining_encryptAESCrypt"; 6001 break; 6002 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6003 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6004 stubName = "cipherBlockChaining_decryptAESCrypt"; 6005 break; 6006 } 6007 if (stubAddr == NULL) return false; 6008 6009 Node* cipherBlockChaining_object = argument(0); 6010 Node* src = argument(1); 6011 Node* src_offset = argument(2); 6012 Node* len = argument(3); 6013 Node* dest = argument(4); 6014 Node* dest_offset = argument(5); 6015 6016 // (1) src and dest are arrays. 6017 const Type* src_type = src->Value(&_gvn); 6018 const Type* dest_type = dest->Value(&_gvn); 6019 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6020 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6021 assert (top_src != NULL && top_src->klass() != NULL 6022 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6023 6024 // checks are the responsibility of the caller 6025 Node* src_start = src; 6026 Node* dest_start = dest; 6027 if (src_offset != NULL || dest_offset != NULL) { 6028 assert(src_offset != NULL && dest_offset != NULL, ""); 6029 src_start = array_element_address(src, src_offset, T_BYTE); 6030 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6031 } 6032 6033 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6034 // (because of the predicated logic executed earlier). 6035 // so we cast it here safely. 6036 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6037 6038 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6039 if (embeddedCipherObj == NULL) return false; 6040 6041 // cast it to what we know it will be at runtime 6042 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6043 assert(tinst != NULL, "CBC obj is null"); 6044 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6045 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6046 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6047 6048 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6049 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6050 const TypeOopPtr* xtype = aklass->as_instance_type(); 6051 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6052 aescrypt_object = _gvn.transform(aescrypt_object); 6053 6054 // we need to get the start of the aescrypt_object's expanded key array 6055 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6056 if (k_start == NULL) return false; 6057 6058 // similarly, get the start address of the r vector 6059 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6060 if (objRvec == NULL) return false; 6061 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6062 6063 Node* cbcCrypt; 6064 if (Matcher::pass_original_key_for_aes()) { 6065 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6066 // compatibility issues between Java key expansion and SPARC crypto instructions 6067 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6068 if (original_k_start == NULL) return false; 6069 6070 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6071 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6072 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6073 stubAddr, stubName, TypePtr::BOTTOM, 6074 src_start, dest_start, k_start, r_start, len, original_k_start); 6075 } else { 6076 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6077 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6078 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6079 stubAddr, stubName, TypePtr::BOTTOM, 6080 src_start, dest_start, k_start, r_start, len); 6081 } 6082 6083 // return cipher length (int) 6084 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6085 set_result(retvalue); 6086 return true; 6087 } 6088 6089 //------------------------------get_key_start_from_aescrypt_object----------------------- 6090 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6091 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6092 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6093 if (objAESCryptKey == NULL) return (Node *) NULL; 6094 6095 // now have the array, need to get the start address of the K array 6096 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6097 return k_start; 6098 } 6099 6100 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6101 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6102 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6103 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6104 if (objAESCryptKey == NULL) return (Node *) NULL; 6105 6106 // now have the array, need to get the start address of the lastKey array 6107 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6108 return original_k_start; 6109 } 6110 6111 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6112 // Return node representing slow path of predicate check. 6113 // the pseudo code we want to emulate with this predicate is: 6114 // for encryption: 6115 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6116 // for decryption: 6117 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6118 // note cipher==plain is more conservative than the original java code but that's OK 6119 // 6120 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6121 // The receiver was checked for NULL already. 6122 Node* objCBC = argument(0); 6123 6124 // Load embeddedCipher field of CipherBlockChaining object. 6125 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6126 6127 // get AESCrypt klass for instanceOf check 6128 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6129 // will have same classloader as CipherBlockChaining object 6130 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6131 assert(tinst != NULL, "CBCobj is null"); 6132 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6133 6134 // we want to do an instanceof comparison against the AESCrypt class 6135 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6136 if (!klass_AESCrypt->is_loaded()) { 6137 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6138 Node* ctrl = control(); 6139 set_control(top()); // no regular fast path 6140 return ctrl; 6141 } 6142 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6143 6144 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6145 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6146 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6147 6148 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6149 6150 // for encryption, we are done 6151 if (!decrypting) 6152 return instof_false; // even if it is NULL 6153 6154 // for decryption, we need to add a further check to avoid 6155 // taking the intrinsic path when cipher and plain are the same 6156 // see the original java code for why. 6157 RegionNode* region = new RegionNode(3); 6158 region->init_req(1, instof_false); 6159 Node* src = argument(1); 6160 Node* dest = argument(4); 6161 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6162 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6163 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6164 region->init_req(2, src_dest_conjoint); 6165 6166 record_for_igvn(region); 6167 return _gvn.transform(region); 6168 } 6169 6170 //------------------------------inline_sha_implCompress----------------------- 6171 // 6172 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6173 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6174 // 6175 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6176 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6177 // 6178 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6179 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6180 // 6181 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6182 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6183 6184 Node* sha_obj = argument(0); 6185 Node* src = argument(1); // type oop 6186 Node* ofs = argument(2); // type int 6187 6188 const Type* src_type = src->Value(&_gvn); 6189 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6190 if (top_src == NULL || top_src->klass() == NULL) { 6191 // failed array check 6192 return false; 6193 } 6194 // Figure out the size and type of the elements we will be copying. 6195 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6196 if (src_elem != T_BYTE) { 6197 return false; 6198 } 6199 // 'src_start' points to src array + offset 6200 Node* src_start = array_element_address(src, ofs, src_elem); 6201 Node* state = NULL; 6202 address stubAddr; 6203 const char *stubName; 6204 6205 switch(id) { 6206 case vmIntrinsics::_sha_implCompress: 6207 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6208 state = get_state_from_sha_object(sha_obj); 6209 stubAddr = StubRoutines::sha1_implCompress(); 6210 stubName = "sha1_implCompress"; 6211 break; 6212 case vmIntrinsics::_sha2_implCompress: 6213 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6214 state = get_state_from_sha_object(sha_obj); 6215 stubAddr = StubRoutines::sha256_implCompress(); 6216 stubName = "sha256_implCompress"; 6217 break; 6218 case vmIntrinsics::_sha5_implCompress: 6219 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6220 state = get_state_from_sha5_object(sha_obj); 6221 stubAddr = StubRoutines::sha512_implCompress(); 6222 stubName = "sha512_implCompress"; 6223 break; 6224 default: 6225 fatal_unexpected_iid(id); 6226 return false; 6227 } 6228 if (state == NULL) return false; 6229 6230 // Call the stub. 6231 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6232 stubAddr, stubName, TypePtr::BOTTOM, 6233 src_start, state); 6234 6235 return true; 6236 } 6237 6238 //------------------------------inline_digestBase_implCompressMB----------------------- 6239 // 6240 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6241 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6242 // 6243 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6244 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6245 "need SHA1/SHA256/SHA512 instruction support"); 6246 assert((uint)predicate < 3, "sanity"); 6247 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6248 6249 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6250 Node* src = argument(1); // byte[] array 6251 Node* ofs = argument(2); // type int 6252 Node* limit = argument(3); // type int 6253 6254 const Type* src_type = src->Value(&_gvn); 6255 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6256 if (top_src == NULL || top_src->klass() == NULL) { 6257 // failed array check 6258 return false; 6259 } 6260 // Figure out the size and type of the elements we will be copying. 6261 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6262 if (src_elem != T_BYTE) { 6263 return false; 6264 } 6265 // 'src_start' points to src array + offset 6266 Node* src_start = array_element_address(src, ofs, src_elem); 6267 6268 const char* klass_SHA_name = NULL; 6269 const char* stub_name = NULL; 6270 address stub_addr = NULL; 6271 bool long_state = false; 6272 6273 switch (predicate) { 6274 case 0: 6275 if (UseSHA1Intrinsics) { 6276 klass_SHA_name = "sun/security/provider/SHA"; 6277 stub_name = "sha1_implCompressMB"; 6278 stub_addr = StubRoutines::sha1_implCompressMB(); 6279 } 6280 break; 6281 case 1: 6282 if (UseSHA256Intrinsics) { 6283 klass_SHA_name = "sun/security/provider/SHA2"; 6284 stub_name = "sha256_implCompressMB"; 6285 stub_addr = StubRoutines::sha256_implCompressMB(); 6286 } 6287 break; 6288 case 2: 6289 if (UseSHA512Intrinsics) { 6290 klass_SHA_name = "sun/security/provider/SHA5"; 6291 stub_name = "sha512_implCompressMB"; 6292 stub_addr = StubRoutines::sha512_implCompressMB(); 6293 long_state = true; 6294 } 6295 break; 6296 default: 6297 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6298 } 6299 if (klass_SHA_name != NULL) { 6300 // get DigestBase klass to lookup for SHA klass 6301 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6302 assert(tinst != NULL, "digestBase_obj is not instance???"); 6303 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6304 6305 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6306 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6307 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6308 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6309 } 6310 return false; 6311 } 6312 //------------------------------inline_sha_implCompressMB----------------------- 6313 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6314 bool long_state, address stubAddr, const char *stubName, 6315 Node* src_start, Node* ofs, Node* limit) { 6316 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6317 const TypeOopPtr* xtype = aklass->as_instance_type(); 6318 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6319 sha_obj = _gvn.transform(sha_obj); 6320 6321 Node* state; 6322 if (long_state) { 6323 state = get_state_from_sha5_object(sha_obj); 6324 } else { 6325 state = get_state_from_sha_object(sha_obj); 6326 } 6327 if (state == NULL) return false; 6328 6329 // Call the stub. 6330 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6331 OptoRuntime::digestBase_implCompressMB_Type(), 6332 stubAddr, stubName, TypePtr::BOTTOM, 6333 src_start, state, ofs, limit); 6334 // return ofs (int) 6335 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6336 set_result(result); 6337 6338 return true; 6339 } 6340 6341 //------------------------------get_state_from_sha_object----------------------- 6342 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6343 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6344 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6345 if (sha_state == NULL) return (Node *) NULL; 6346 6347 // now have the array, need to get the start address of the state array 6348 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6349 return state; 6350 } 6351 6352 //------------------------------get_state_from_sha5_object----------------------- 6353 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6354 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6355 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6356 if (sha_state == NULL) return (Node *) NULL; 6357 6358 // now have the array, need to get the start address of the state array 6359 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6360 return state; 6361 } 6362 6363 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6364 // Return node representing slow path of predicate check. 6365 // the pseudo code we want to emulate with this predicate is: 6366 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6367 // 6368 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6369 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6370 "need SHA1/SHA256/SHA512 instruction support"); 6371 assert((uint)predicate < 3, "sanity"); 6372 6373 // The receiver was checked for NULL already. 6374 Node* digestBaseObj = argument(0); 6375 6376 // get DigestBase klass for instanceOf check 6377 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6378 assert(tinst != NULL, "digestBaseObj is null"); 6379 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6380 6381 const char* klass_SHA_name = NULL; 6382 switch (predicate) { 6383 case 0: 6384 if (UseSHA1Intrinsics) { 6385 // we want to do an instanceof comparison against the SHA class 6386 klass_SHA_name = "sun/security/provider/SHA"; 6387 } 6388 break; 6389 case 1: 6390 if (UseSHA256Intrinsics) { 6391 // we want to do an instanceof comparison against the SHA2 class 6392 klass_SHA_name = "sun/security/provider/SHA2"; 6393 } 6394 break; 6395 case 2: 6396 if (UseSHA512Intrinsics) { 6397 // we want to do an instanceof comparison against the SHA5 class 6398 klass_SHA_name = "sun/security/provider/SHA5"; 6399 } 6400 break; 6401 default: 6402 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6403 } 6404 6405 ciKlass* klass_SHA = NULL; 6406 if (klass_SHA_name != NULL) { 6407 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6408 } 6409 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6410 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6411 Node* ctrl = control(); 6412 set_control(top()); // no intrinsic path 6413 return ctrl; 6414 } 6415 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6416 6417 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6418 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6419 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6420 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6421 6422 return instof_false; // even if it is NULL 6423 }