1 /* 2 * Copyright (c) 1999, 2019, 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 "classfile/systemDictionary.hpp" 27 #include "classfile/vmSymbols.hpp" 28 #include "compiler/compileBroker.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "jfr/support/jfrIntrinsics.hpp" 31 #include "oops/objArrayKlass.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/callGenerator.hpp" 34 #include "opto/cfgnode.hpp" 35 #include "opto/connode.hpp" 36 #include "opto/idealKit.hpp" 37 #include "opto/mathexactnode.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/parse.hpp" 40 #include "opto/runtime.hpp" 41 #include "opto/subnode.hpp" 42 #include "prims/nativeLookup.hpp" 43 #include "runtime/sharedRuntime.hpp" 44 #include "utilities/macros.hpp" 45 46 class LibraryIntrinsic : public InlineCallGenerator { 47 // Extend the set of intrinsics known to the runtime: 48 public: 49 private: 50 bool _is_virtual; 51 bool _does_virtual_dispatch; 52 int8_t _predicates_count; // Intrinsic is predicated by several conditions 53 int8_t _last_predicate; // Last generated predicate 54 vmIntrinsics::ID _intrinsic_id; 55 56 public: 57 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 58 : InlineCallGenerator(m), 59 _is_virtual(is_virtual), 60 _does_virtual_dispatch(does_virtual_dispatch), 61 _predicates_count((int8_t)predicates_count), 62 _last_predicate((int8_t)-1), 63 _intrinsic_id(id) 64 { 65 } 66 virtual bool is_intrinsic() const { return true; } 67 virtual bool is_virtual() const { return _is_virtual; } 68 virtual bool is_predicated() const { return _predicates_count > 0; } 69 virtual int predicates_count() const { return _predicates_count; } 70 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 71 virtual JVMState* generate(JVMState* jvms); 72 virtual Node* generate_predicate(JVMState* jvms, int predicate); 73 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 74 }; 75 76 77 // Local helper class for LibraryIntrinsic: 78 class LibraryCallKit : public GraphKit { 79 private: 80 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 81 Node* _result; // the result node, if any 82 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 83 84 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 85 86 public: 87 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 88 : GraphKit(jvms), 89 _intrinsic(intrinsic), 90 _result(NULL) 91 { 92 // Check if this is a root compile. In that case we don't have a caller. 93 if (!jvms->has_method()) { 94 _reexecute_sp = sp(); 95 } else { 96 // Find out how many arguments the interpreter needs when deoptimizing 97 // and save the stack pointer value so it can used by uncommon_trap. 98 // We find the argument count by looking at the declared signature. 99 bool ignored_will_link; 100 ciSignature* declared_signature = NULL; 101 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 102 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 103 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 104 } 105 } 106 107 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 108 109 ciMethod* caller() const { return jvms()->method(); } 110 int bci() const { return jvms()->bci(); } 111 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 112 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 113 ciMethod* callee() const { return _intrinsic->method(); } 114 115 bool try_to_inline(int predicate); 116 Node* try_to_predicate(int predicate); 117 118 void push_result() { 119 // Push the result onto the stack. 120 if (!stopped() && result() != NULL) { 121 BasicType bt = result()->bottom_type()->basic_type(); 122 push_node(bt, result()); 123 } 124 } 125 126 private: 127 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 128 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 129 } 130 131 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 132 void set_result(RegionNode* region, PhiNode* value); 133 Node* result() { return _result; } 134 135 virtual int reexecute_sp() { return _reexecute_sp; } 136 137 // Helper functions to inline natives 138 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 139 Node* generate_slow_guard(Node* test, RegionNode* region); 140 Node* generate_fair_guard(Node* test, RegionNode* region); 141 Node* generate_negative_guard(Node* index, RegionNode* region, 142 // resulting CastII of index: 143 Node* *pos_index = NULL); 144 Node* generate_nonpositive_guard(Node* index, bool never_negative, 145 // resulting CastII of index: 146 Node* *pos_index = NULL); 147 Node* generate_limit_guard(Node* offset, Node* subseq_length, 148 Node* array_length, 149 RegionNode* region); 150 Node* generate_current_thread(Node* &tls_output); 151 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 152 bool disjoint_bases, const char* &name, bool dest_uninitialized); 153 Node* load_mirror_from_klass(Node* klass); 154 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 155 RegionNode* region, int null_path, 156 int offset); 157 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 158 RegionNode* region, int null_path) { 159 int offset = java_lang_Class::klass_offset_in_bytes(); 160 return load_klass_from_mirror_common(mirror, never_see_null, 161 region, null_path, 162 offset); 163 } 164 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 165 RegionNode* region, int null_path) { 166 int offset = java_lang_Class::array_klass_offset_in_bytes(); 167 return load_klass_from_mirror_common(mirror, never_see_null, 168 region, null_path, 169 offset); 170 } 171 Node* generate_access_flags_guard(Node* kls, 172 int modifier_mask, int modifier_bits, 173 RegionNode* region); 174 Node* generate_interface_guard(Node* kls, RegionNode* region); 175 Node* generate_array_guard(Node* kls, RegionNode* region) { 176 return generate_array_guard_common(kls, region, false, false); 177 } 178 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 179 return generate_array_guard_common(kls, region, false, true); 180 } 181 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 182 return generate_array_guard_common(kls, region, true, false); 183 } 184 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 185 return generate_array_guard_common(kls, region, true, true); 186 } 187 Node* generate_array_guard_common(Node* kls, RegionNode* region, 188 bool obj_array, bool not_array); 189 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 190 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 191 bool is_virtual = false, bool is_static = false); 192 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 193 return generate_method_call(method_id, false, true); 194 } 195 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 196 return generate_method_call(method_id, true, false); 197 } 198 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 199 200 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 201 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 202 bool inline_string_compareTo(); 203 bool inline_string_indexOf(); 204 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 205 bool inline_string_equals(); 206 Node* round_double_node(Node* n); 207 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 208 bool inline_math_native(vmIntrinsics::ID id); 209 bool inline_trig(vmIntrinsics::ID id); 210 bool inline_math(vmIntrinsics::ID id); 211 template <typename OverflowOp> 212 bool inline_math_overflow(Node* arg1, Node* arg2); 213 void inline_math_mathExact(Node* math, Node* test); 214 bool inline_math_addExactI(bool is_increment); 215 bool inline_math_addExactL(bool is_increment); 216 bool inline_math_multiplyExactI(); 217 bool inline_math_multiplyExactL(); 218 bool inline_math_negateExactI(); 219 bool inline_math_negateExactL(); 220 bool inline_math_subtractExactI(bool is_decrement); 221 bool inline_math_subtractExactL(bool is_decrement); 222 bool inline_exp(); 223 bool inline_pow(); 224 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 225 bool inline_min_max(vmIntrinsics::ID id); 226 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 227 // This returns Type::AnyPtr, RawPtr, or OopPtr. 228 int classify_unsafe_addr(Node* &base, Node* &offset); 229 Node* make_unsafe_address(Node* base, Node* offset); 230 // Helper for inline_unsafe_access. 231 // Generates the guards that check whether the result of 232 // Unsafe.getObject should be recorded in an SATB log buffer. 233 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 234 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned); 235 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 236 static bool klass_needs_init_guard(Node* kls); 237 bool inline_unsafe_allocate(); 238 bool inline_unsafe_copyMemory(); 239 bool inline_native_currentThread(); 240 #ifdef JFR_HAVE_INTRINSICS 241 bool inline_native_classID(); 242 bool inline_native_getEventWriter(); 243 #endif 244 bool inline_native_time_funcs(address method, const char* funcName); 245 bool inline_native_isInterrupted(); 246 bool inline_native_Class_query(vmIntrinsics::ID id); 247 bool inline_native_subtype_check(); 248 249 bool inline_native_newArray(); 250 bool inline_native_getLength(); 251 bool inline_array_copyOf(bool is_copyOfRange); 252 bool inline_array_equals(); 253 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 254 bool inline_native_clone(bool is_virtual); 255 bool inline_native_Reflection_getCallerClass(); 256 // Helper function for inlining native object hash method 257 bool inline_native_hashcode(bool is_virtual, bool is_static); 258 bool inline_native_getClass(); 259 260 // Helper functions for inlining arraycopy 261 bool inline_arraycopy(); 262 void generate_arraycopy(const TypePtr* adr_type, 263 BasicType basic_elem_type, 264 Node* src, Node* src_offset, 265 Node* dest, Node* dest_offset, 266 Node* copy_length, 267 bool disjoint_bases = false, 268 bool length_never_negative = false, 269 RegionNode* slow_region = NULL); 270 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 271 RegionNode* slow_region); 272 void generate_clear_array(const TypePtr* adr_type, 273 Node* dest, 274 BasicType basic_elem_type, 275 Node* slice_off, 276 Node* slice_len, 277 Node* slice_end); 278 bool generate_block_arraycopy(const TypePtr* adr_type, 279 BasicType basic_elem_type, 280 AllocateNode* alloc, 281 Node* src, Node* src_offset, 282 Node* dest, Node* dest_offset, 283 Node* dest_size, bool dest_uninitialized); 284 void generate_slow_arraycopy(const TypePtr* adr_type, 285 Node* src, Node* src_offset, 286 Node* dest, Node* dest_offset, 287 Node* copy_length, bool dest_uninitialized); 288 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 289 Node* dest_elem_klass, 290 Node* src, Node* src_offset, 291 Node* dest, Node* dest_offset, 292 Node* copy_length, bool dest_uninitialized); 293 Node* generate_generic_arraycopy(const TypePtr* adr_type, 294 Node* src, Node* src_offset, 295 Node* dest, Node* dest_offset, 296 Node* copy_length, bool dest_uninitialized); 297 void generate_unchecked_arraycopy(const TypePtr* adr_type, 298 BasicType basic_elem_type, 299 bool disjoint_bases, 300 Node* src, Node* src_offset, 301 Node* dest, Node* dest_offset, 302 Node* copy_length, bool dest_uninitialized); 303 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 304 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 305 bool inline_unsafe_ordered_store(BasicType type); 306 bool inline_unsafe_fence(vmIntrinsics::ID id); 307 bool inline_fp_conversions(vmIntrinsics::ID id); 308 bool inline_number_methods(vmIntrinsics::ID id); 309 bool inline_reference_get(); 310 bool inline_aescrypt_Block(vmIntrinsics::ID id); 311 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 312 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 313 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 314 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 315 bool inline_ghash_processBlocks(); 316 bool inline_sha_implCompress(vmIntrinsics::ID id); 317 bool inline_digestBase_implCompressMB(int predicate); 318 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 319 bool long_state, address stubAddr, const char *stubName, 320 Node* src_start, Node* ofs, Node* limit); 321 Node* get_state_from_sha_object(Node *sha_object); 322 Node* get_state_from_sha5_object(Node *sha_object); 323 Node* inline_digestBase_implCompressMB_predicate(int predicate); 324 bool inline_encodeISOArray(); 325 bool inline_updateCRC32(); 326 bool inline_updateBytesCRC32(); 327 bool inline_updateByteBufferCRC32(); 328 bool inline_multiplyToLen(); 329 bool inline_squareToLen(); 330 bool inline_mulAdd(); 331 bool inline_montgomeryMultiply(); 332 bool inline_montgomerySquare(); 333 334 bool inline_profileBoolean(); 335 }; 336 337 338 //---------------------------make_vm_intrinsic---------------------------- 339 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 340 vmIntrinsics::ID id = m->intrinsic_id(); 341 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 342 343 ccstr disable_intr = NULL; 344 345 if ((DisableIntrinsic[0] != '\0' 346 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 347 (method_has_option_value("DisableIntrinsic", disable_intr) 348 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 349 // disabled by a user request on the command line: 350 // example: -XX:DisableIntrinsic=_hashCode,_getClass 351 return NULL; 352 } 353 354 if (!m->is_loaded()) { 355 // do not attempt to inline unloaded methods 356 return NULL; 357 } 358 359 // Only a few intrinsics implement a virtual dispatch. 360 // They are expensive calls which are also frequently overridden. 361 if (is_virtual) { 362 switch (id) { 363 case vmIntrinsics::_hashCode: 364 case vmIntrinsics::_clone: 365 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 366 break; 367 default: 368 return NULL; 369 } 370 } 371 372 // -XX:-InlineNatives disables nearly all intrinsics: 373 if (!InlineNatives) { 374 switch (id) { 375 case vmIntrinsics::_indexOf: 376 case vmIntrinsics::_compareTo: 377 case vmIntrinsics::_equals: 378 case vmIntrinsics::_equalsC: 379 case vmIntrinsics::_getAndAddInt: 380 case vmIntrinsics::_getAndAddLong: 381 case vmIntrinsics::_getAndSetInt: 382 case vmIntrinsics::_getAndSetLong: 383 case vmIntrinsics::_getAndSetObject: 384 case vmIntrinsics::_loadFence: 385 case vmIntrinsics::_storeFence: 386 case vmIntrinsics::_fullFence: 387 break; // InlineNatives does not control String.compareTo 388 case vmIntrinsics::_Reference_get: 389 break; // InlineNatives does not control Reference.get 390 default: 391 return NULL; 392 } 393 } 394 395 int predicates = 0; 396 bool does_virtual_dispatch = false; 397 398 switch (id) { 399 case vmIntrinsics::_compareTo: 400 if (!SpecialStringCompareTo) return NULL; 401 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 402 break; 403 case vmIntrinsics::_indexOf: 404 if (!SpecialStringIndexOf) return NULL; 405 break; 406 case vmIntrinsics::_equals: 407 if (!SpecialStringEquals) return NULL; 408 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 409 break; 410 case vmIntrinsics::_equalsC: 411 if (!SpecialArraysEquals) return NULL; 412 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 413 break; 414 case vmIntrinsics::_arraycopy: 415 if (!InlineArrayCopy) return NULL; 416 break; 417 case vmIntrinsics::_copyMemory: 418 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 419 if (!InlineArrayCopy) return NULL; 420 break; 421 case vmIntrinsics::_hashCode: 422 if (!InlineObjectHash) return NULL; 423 does_virtual_dispatch = true; 424 break; 425 case vmIntrinsics::_clone: 426 does_virtual_dispatch = true; 427 case vmIntrinsics::_copyOf: 428 case vmIntrinsics::_copyOfRange: 429 if (!InlineObjectCopy) return NULL; 430 // These also use the arraycopy intrinsic mechanism: 431 if (!InlineArrayCopy) return NULL; 432 break; 433 case vmIntrinsics::_encodeISOArray: 434 if (!SpecialEncodeISOArray) return NULL; 435 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 436 break; 437 case vmIntrinsics::_checkIndex: 438 // We do not intrinsify this. The optimizer does fine with it. 439 return NULL; 440 441 case vmIntrinsics::_getCallerClass: 442 if (!UseNewReflection) return NULL; 443 if (!InlineReflectionGetCallerClass) return NULL; 444 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 445 break; 446 447 case vmIntrinsics::_bitCount_i: 448 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 449 break; 450 451 case vmIntrinsics::_bitCount_l: 452 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 453 break; 454 455 case vmIntrinsics::_numberOfLeadingZeros_i: 456 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 457 break; 458 459 case vmIntrinsics::_numberOfLeadingZeros_l: 460 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 461 break; 462 463 case vmIntrinsics::_numberOfTrailingZeros_i: 464 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 465 break; 466 467 case vmIntrinsics::_numberOfTrailingZeros_l: 468 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 469 break; 470 471 case vmIntrinsics::_reverseBytes_c: 472 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 473 break; 474 case vmIntrinsics::_reverseBytes_s: 475 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 476 break; 477 case vmIntrinsics::_reverseBytes_i: 478 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 479 break; 480 case vmIntrinsics::_reverseBytes_l: 481 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 482 break; 483 484 case vmIntrinsics::_Reference_get: 485 // Use the intrinsic version of Reference.get() so that the value in 486 // the referent field can be registered by the G1 pre-barrier code. 487 // Also add memory barrier to prevent commoning reads from this field 488 // across safepoint since GC can change it value. 489 break; 490 491 case vmIntrinsics::_compareAndSwapObject: 492 #ifdef _LP64 493 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 494 #endif 495 break; 496 497 case vmIntrinsics::_compareAndSwapLong: 498 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 499 break; 500 501 case vmIntrinsics::_getAndAddInt: 502 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 503 break; 504 505 case vmIntrinsics::_getAndAddLong: 506 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 507 break; 508 509 case vmIntrinsics::_getAndSetInt: 510 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 511 break; 512 513 case vmIntrinsics::_getAndSetLong: 514 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 515 break; 516 517 case vmIntrinsics::_getAndSetObject: 518 #ifdef _LP64 519 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 520 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 521 break; 522 #else 523 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 524 break; 525 #endif 526 527 case vmIntrinsics::_aescrypt_encryptBlock: 528 case vmIntrinsics::_aescrypt_decryptBlock: 529 if (!UseAESIntrinsics) return NULL; 530 break; 531 532 case vmIntrinsics::_multiplyToLen: 533 if (!UseMultiplyToLenIntrinsic) return NULL; 534 break; 535 536 case vmIntrinsics::_squareToLen: 537 if (!UseSquareToLenIntrinsic) return NULL; 538 break; 539 540 case vmIntrinsics::_mulAdd: 541 if (!UseMulAddIntrinsic) return NULL; 542 break; 543 544 case vmIntrinsics::_montgomeryMultiply: 545 if (!UseMontgomeryMultiplyIntrinsic) return NULL; 546 break; 547 case vmIntrinsics::_montgomerySquare: 548 if (!UseMontgomerySquareIntrinsic) return NULL; 549 break; 550 551 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 552 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 553 if (!UseAESIntrinsics) return NULL; 554 // these two require the predicated logic 555 predicates = 1; 556 break; 557 558 case vmIntrinsics::_sha_implCompress: 559 if (!UseSHA1Intrinsics) return NULL; 560 break; 561 562 case vmIntrinsics::_sha2_implCompress: 563 if (!UseSHA256Intrinsics) return NULL; 564 break; 565 566 case vmIntrinsics::_sha5_implCompress: 567 if (!UseSHA512Intrinsics) return NULL; 568 break; 569 570 case vmIntrinsics::_digestBase_implCompressMB: 571 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 572 predicates = 3; 573 break; 574 575 case vmIntrinsics::_ghash_processBlocks: 576 if (!UseGHASHIntrinsics) return NULL; 577 break; 578 579 case vmIntrinsics::_updateCRC32: 580 case vmIntrinsics::_updateBytesCRC32: 581 case vmIntrinsics::_updateByteBufferCRC32: 582 if (!UseCRC32Intrinsics) return NULL; 583 break; 584 585 case vmIntrinsics::_incrementExactI: 586 case vmIntrinsics::_addExactI: 587 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 588 break; 589 case vmIntrinsics::_incrementExactL: 590 case vmIntrinsics::_addExactL: 591 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 592 break; 593 case vmIntrinsics::_decrementExactI: 594 case vmIntrinsics::_subtractExactI: 595 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 596 break; 597 case vmIntrinsics::_decrementExactL: 598 case vmIntrinsics::_subtractExactL: 599 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 600 break; 601 case vmIntrinsics::_negateExactI: 602 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 603 break; 604 case vmIntrinsics::_negateExactL: 605 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 606 break; 607 case vmIntrinsics::_multiplyExactI: 608 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 609 break; 610 case vmIntrinsics::_multiplyExactL: 611 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 612 break; 613 614 default: 615 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 616 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 617 break; 618 } 619 620 // -XX:-InlineClassNatives disables natives from the Class class. 621 // The flag applies to all reflective calls, notably Array.newArray 622 // (visible to Java programmers as Array.newInstance). 623 if (m->holder()->name() == ciSymbol::java_lang_Class() || 624 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 625 if (!InlineClassNatives) return NULL; 626 } 627 628 // -XX:-InlineThreadNatives disables natives from the Thread class. 629 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 630 if (!InlineThreadNatives) return NULL; 631 } 632 633 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 634 if (m->holder()->name() == ciSymbol::java_lang_Math() || 635 m->holder()->name() == ciSymbol::java_lang_Float() || 636 m->holder()->name() == ciSymbol::java_lang_Double()) { 637 if (!InlineMathNatives) return NULL; 638 } 639 640 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 641 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 642 if (!InlineUnsafeOps) return NULL; 643 } 644 645 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 646 } 647 648 //----------------------register_library_intrinsics----------------------- 649 // Initialize this file's data structures, for each Compile instance. 650 void Compile::register_library_intrinsics() { 651 // Nothing to do here. 652 } 653 654 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 655 LibraryCallKit kit(jvms, this); 656 Compile* C = kit.C; 657 int nodes = C->unique(); 658 #ifndef PRODUCT 659 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 660 char buf[1000]; 661 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 662 tty->print_cr("Intrinsic %s", str); 663 } 664 #endif 665 ciMethod* callee = kit.callee(); 666 const int bci = kit.bci(); 667 668 // Try to inline the intrinsic. 669 if (kit.try_to_inline(_last_predicate)) { 670 if (C->print_intrinsics() || C->print_inlining()) { 671 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 672 } 673 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 674 if (C->log()) { 675 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 676 vmIntrinsics::name_at(intrinsic_id()), 677 (is_virtual() ? " virtual='1'" : ""), 678 C->unique() - nodes); 679 } 680 // Push the result from the inlined method onto the stack. 681 kit.push_result(); 682 return kit.transfer_exceptions_into_jvms(); 683 } 684 685 // The intrinsic bailed out 686 if (C->print_intrinsics() || C->print_inlining()) { 687 if (jvms->has_method()) { 688 // Not a root compile. 689 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 690 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 691 } else { 692 // Root compile 693 tty->print("Did not generate intrinsic %s%s at bci:%d in", 694 vmIntrinsics::name_at(intrinsic_id()), 695 (is_virtual() ? " (virtual)" : ""), bci); 696 } 697 } 698 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 699 return NULL; 700 } 701 702 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 703 LibraryCallKit kit(jvms, this); 704 Compile* C = kit.C; 705 int nodes = C->unique(); 706 _last_predicate = predicate; 707 #ifndef PRODUCT 708 assert(is_predicated() && predicate < predicates_count(), "sanity"); 709 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 710 char buf[1000]; 711 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 712 tty->print_cr("Predicate for intrinsic %s", str); 713 } 714 #endif 715 ciMethod* callee = kit.callee(); 716 const int bci = kit.bci(); 717 718 Node* slow_ctl = kit.try_to_predicate(predicate); 719 if (!kit.failing()) { 720 if (C->print_intrinsics() || C->print_inlining()) { 721 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 722 } 723 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 724 if (C->log()) { 725 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 726 vmIntrinsics::name_at(intrinsic_id()), 727 (is_virtual() ? " virtual='1'" : ""), 728 C->unique() - nodes); 729 } 730 return slow_ctl; // Could be NULL if the check folds. 731 } 732 733 // The intrinsic bailed out 734 if (C->print_intrinsics() || C->print_inlining()) { 735 if (jvms->has_method()) { 736 // Not a root compile. 737 const char* msg = "failed to generate predicate for intrinsic"; 738 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 739 } else { 740 // Root compile 741 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 742 vmIntrinsics::name_at(intrinsic_id()), 743 (is_virtual() ? " (virtual)" : ""), bci); 744 } 745 } 746 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 747 return NULL; 748 } 749 750 bool LibraryCallKit::try_to_inline(int predicate) { 751 // Handle symbolic names for otherwise undistinguished boolean switches: 752 const bool is_store = true; 753 const bool is_native_ptr = true; 754 const bool is_static = true; 755 const bool is_volatile = true; 756 757 if (!jvms()->has_method()) { 758 // Root JVMState has a null method. 759 assert(map()->memory()->Opcode() == Op_Parm, ""); 760 // Insert the memory aliasing node 761 set_all_memory(reset_memory()); 762 } 763 assert(merged_memory(), ""); 764 765 766 switch (intrinsic_id()) { 767 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 768 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 769 case vmIntrinsics::_getClass: return inline_native_getClass(); 770 771 case vmIntrinsics::_dsin: 772 case vmIntrinsics::_dcos: 773 case vmIntrinsics::_dtan: 774 case vmIntrinsics::_dabs: 775 case vmIntrinsics::_datan2: 776 case vmIntrinsics::_dsqrt: 777 case vmIntrinsics::_dexp: 778 case vmIntrinsics::_dlog: 779 case vmIntrinsics::_dlog10: 780 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 781 782 case vmIntrinsics::_min: 783 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 784 785 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 786 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 787 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 788 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 789 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 790 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 791 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 792 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 793 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 794 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 795 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 796 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 797 798 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 799 800 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 801 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 802 case vmIntrinsics::_equals: return inline_string_equals(); 803 804 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false); 805 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false); 806 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 807 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 808 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 809 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false); 810 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false); 811 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 812 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 813 814 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false); 815 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false); 816 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false); 817 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false); 818 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false); 819 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false); 820 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false); 821 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 822 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 823 824 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 825 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 826 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 827 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false); 828 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false); 829 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 830 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 831 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false); 832 833 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false); 834 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false); 835 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false); 836 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false); 837 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false); 838 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 839 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 840 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false); 841 842 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false); 843 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false); 844 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false); 845 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false); 846 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false); 847 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false); 848 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false); 849 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false); 850 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false); 851 852 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false); 853 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false); 854 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false); 855 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false); 856 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false); 857 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false); 858 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false); 859 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false); 860 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false); 861 862 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 863 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 864 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 865 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 866 867 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 868 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 869 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 870 871 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 872 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 873 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 874 875 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 876 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 877 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 878 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 879 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 880 881 case vmIntrinsics::_loadFence: 882 case vmIntrinsics::_storeFence: 883 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 884 885 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 886 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 887 888 #ifdef JFR_HAVE_INTRINSICS 889 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime"); 890 case vmIntrinsics::_getClassId: return inline_native_classID(); 891 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter(); 892 #endif 893 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 894 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 895 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 896 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 897 case vmIntrinsics::_newArray: return inline_native_newArray(); 898 case vmIntrinsics::_getLength: return inline_native_getLength(); 899 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 900 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 901 case vmIntrinsics::_equalsC: return inline_array_equals(); 902 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 903 904 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 905 906 case vmIntrinsics::_isInstance: 907 case vmIntrinsics::_getModifiers: 908 case vmIntrinsics::_isInterface: 909 case vmIntrinsics::_isArray: 910 case vmIntrinsics::_isPrimitive: 911 case vmIntrinsics::_getSuperclass: 912 case vmIntrinsics::_getComponentType: 913 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 914 915 case vmIntrinsics::_floatToRawIntBits: 916 case vmIntrinsics::_floatToIntBits: 917 case vmIntrinsics::_intBitsToFloat: 918 case vmIntrinsics::_doubleToRawLongBits: 919 case vmIntrinsics::_doubleToLongBits: 920 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 921 922 case vmIntrinsics::_numberOfLeadingZeros_i: 923 case vmIntrinsics::_numberOfLeadingZeros_l: 924 case vmIntrinsics::_numberOfTrailingZeros_i: 925 case vmIntrinsics::_numberOfTrailingZeros_l: 926 case vmIntrinsics::_bitCount_i: 927 case vmIntrinsics::_bitCount_l: 928 case vmIntrinsics::_reverseBytes_i: 929 case vmIntrinsics::_reverseBytes_l: 930 case vmIntrinsics::_reverseBytes_s: 931 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 932 933 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 934 935 case vmIntrinsics::_Reference_get: return inline_reference_get(); 936 937 case vmIntrinsics::_aescrypt_encryptBlock: 938 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 939 940 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 941 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 942 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 943 944 case vmIntrinsics::_sha_implCompress: 945 case vmIntrinsics::_sha2_implCompress: 946 case vmIntrinsics::_sha5_implCompress: 947 return inline_sha_implCompress(intrinsic_id()); 948 949 case vmIntrinsics::_digestBase_implCompressMB: 950 return inline_digestBase_implCompressMB(predicate); 951 952 case vmIntrinsics::_multiplyToLen: 953 return inline_multiplyToLen(); 954 955 case vmIntrinsics::_squareToLen: 956 return inline_squareToLen(); 957 958 case vmIntrinsics::_mulAdd: 959 return inline_mulAdd(); 960 961 case vmIntrinsics::_montgomeryMultiply: 962 return inline_montgomeryMultiply(); 963 case vmIntrinsics::_montgomerySquare: 964 return inline_montgomerySquare(); 965 966 case vmIntrinsics::_ghash_processBlocks: 967 return inline_ghash_processBlocks(); 968 969 case vmIntrinsics::_encodeISOArray: 970 return inline_encodeISOArray(); 971 972 case vmIntrinsics::_updateCRC32: 973 return inline_updateCRC32(); 974 case vmIntrinsics::_updateBytesCRC32: 975 return inline_updateBytesCRC32(); 976 case vmIntrinsics::_updateByteBufferCRC32: 977 return inline_updateByteBufferCRC32(); 978 979 case vmIntrinsics::_profileBoolean: 980 return inline_profileBoolean(); 981 982 default: 983 // If you get here, it may be that someone has added a new intrinsic 984 // to the list in vmSymbols.hpp without implementing it here. 985 #ifndef PRODUCT 986 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 987 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 988 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 989 } 990 #endif 991 return false; 992 } 993 } 994 995 Node* LibraryCallKit::try_to_predicate(int predicate) { 996 if (!jvms()->has_method()) { 997 // Root JVMState has a null method. 998 assert(map()->memory()->Opcode() == Op_Parm, ""); 999 // Insert the memory aliasing node 1000 set_all_memory(reset_memory()); 1001 } 1002 assert(merged_memory(), ""); 1003 1004 switch (intrinsic_id()) { 1005 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 1006 return inline_cipherBlockChaining_AESCrypt_predicate(false); 1007 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 1008 return inline_cipherBlockChaining_AESCrypt_predicate(true); 1009 case vmIntrinsics::_digestBase_implCompressMB: 1010 return inline_digestBase_implCompressMB_predicate(predicate); 1011 1012 default: 1013 // If you get here, it may be that someone has added a new intrinsic 1014 // to the list in vmSymbols.hpp without implementing it here. 1015 #ifndef PRODUCT 1016 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 1017 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 1018 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 1019 } 1020 #endif 1021 Node* slow_ctl = control(); 1022 set_control(top()); // No fast path instrinsic 1023 return slow_ctl; 1024 } 1025 } 1026 1027 //------------------------------set_result------------------------------- 1028 // Helper function for finishing intrinsics. 1029 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 1030 record_for_igvn(region); 1031 set_control(_gvn.transform(region)); 1032 set_result( _gvn.transform(value)); 1033 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 1034 } 1035 1036 //------------------------------generate_guard--------------------------- 1037 // Helper function for generating guarded fast-slow graph structures. 1038 // The given 'test', if true, guards a slow path. If the test fails 1039 // then a fast path can be taken. (We generally hope it fails.) 1040 // In all cases, GraphKit::control() is updated to the fast path. 1041 // The returned value represents the control for the slow path. 1042 // The return value is never 'top'; it is either a valid control 1043 // or NULL if it is obvious that the slow path can never be taken. 1044 // Also, if region and the slow control are not NULL, the slow edge 1045 // is appended to the region. 1046 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 1047 if (stopped()) { 1048 // Already short circuited. 1049 return NULL; 1050 } 1051 1052 // Build an if node and its projections. 1053 // If test is true we take the slow path, which we assume is uncommon. 1054 if (_gvn.type(test) == TypeInt::ZERO) { 1055 // The slow branch is never taken. No need to build this guard. 1056 return NULL; 1057 } 1058 1059 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1060 1061 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff)); 1062 if (if_slow == top()) { 1063 // The slow branch is never taken. No need to build this guard. 1064 return NULL; 1065 } 1066 1067 if (region != NULL) 1068 region->add_req(if_slow); 1069 1070 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff)); 1071 set_control(if_fast); 1072 1073 return if_slow; 1074 } 1075 1076 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1077 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1078 } 1079 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1080 return generate_guard(test, region, PROB_FAIR); 1081 } 1082 1083 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1084 Node* *pos_index) { 1085 if (stopped()) 1086 return NULL; // already stopped 1087 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1088 return NULL; // index is already adequately typed 1089 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0))); 1090 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 1091 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1092 if (is_neg != NULL && pos_index != NULL) { 1093 // Emulate effect of Parse::adjust_map_after_if. 1094 Node* ccast = new (C) CastIINode(index, TypeInt::POS); 1095 ccast->set_req(0, control()); 1096 (*pos_index) = _gvn.transform(ccast); 1097 } 1098 return is_neg; 1099 } 1100 1101 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 1102 Node* *pos_index) { 1103 if (stopped()) 1104 return NULL; // already stopped 1105 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 1106 return NULL; // index is already adequately typed 1107 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0))); 1108 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 1109 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq)); 1110 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 1111 if (is_notp != NULL && pos_index != NULL) { 1112 // Emulate effect of Parse::adjust_map_after_if. 1113 Node* ccast = new (C) CastIINode(index, TypeInt::POS1); 1114 ccast->set_req(0, control()); 1115 (*pos_index) = _gvn.transform(ccast); 1116 } 1117 return is_notp; 1118 } 1119 1120 // Make sure that 'position' is a valid limit index, in [0..length]. 1121 // There are two equivalent plans for checking this: 1122 // A. (offset + copyLength) unsigned<= arrayLength 1123 // B. offset <= (arrayLength - copyLength) 1124 // We require that all of the values above, except for the sum and 1125 // difference, are already known to be non-negative. 1126 // Plan A is robust in the face of overflow, if offset and copyLength 1127 // are both hugely positive. 1128 // 1129 // Plan B is less direct and intuitive, but it does not overflow at 1130 // all, since the difference of two non-negatives is always 1131 // representable. Whenever Java methods must perform the equivalent 1132 // check they generally use Plan B instead of Plan A. 1133 // For the moment we use Plan A. 1134 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1135 Node* subseq_length, 1136 Node* array_length, 1137 RegionNode* region) { 1138 if (stopped()) 1139 return NULL; // already stopped 1140 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1141 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1142 return NULL; // common case of whole-array copy 1143 Node* last = subseq_length; 1144 if (!zero_offset) // last += offset 1145 last = _gvn.transform(new (C) AddINode(last, offset)); 1146 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last)); 1147 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 1148 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1149 return is_over; 1150 } 1151 1152 1153 //--------------------------generate_current_thread-------------------- 1154 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1155 ciKlass* thread_klass = env()->Thread_klass(); 1156 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1157 Node* thread = _gvn.transform(new (C) ThreadLocalNode()); 1158 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1159 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1160 tls_output = thread; 1161 return threadObj; 1162 } 1163 1164 1165 //------------------------------make_string_method_node------------------------ 1166 // Helper method for String intrinsic functions. This version is called 1167 // with str1 and str2 pointing to String object nodes. 1168 // 1169 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1170 Node* no_ctrl = NULL; 1171 1172 // Get start addr of string 1173 Node* str1_value = load_String_value(no_ctrl, str1); 1174 Node* str1_offset = load_String_offset(no_ctrl, str1); 1175 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1176 1177 // Get length of string 1 1178 Node* str1_len = load_String_length(no_ctrl, str1); 1179 1180 Node* str2_value = load_String_value(no_ctrl, str2); 1181 Node* str2_offset = load_String_offset(no_ctrl, str2); 1182 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1183 1184 Node* str2_len = NULL; 1185 Node* result = NULL; 1186 1187 switch (opcode) { 1188 case Op_StrIndexOf: 1189 // Get length of string 2 1190 str2_len = load_String_length(no_ctrl, str2); 1191 1192 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1193 str1_start, str1_len, str2_start, str2_len); 1194 break; 1195 case Op_StrComp: 1196 // Get length of string 2 1197 str2_len = load_String_length(no_ctrl, str2); 1198 1199 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1200 str1_start, str1_len, str2_start, str2_len); 1201 break; 1202 case Op_StrEquals: 1203 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1204 str1_start, str2_start, str1_len); 1205 break; 1206 default: 1207 ShouldNotReachHere(); 1208 return NULL; 1209 } 1210 1211 // All these intrinsics have checks. 1212 C->set_has_split_ifs(true); // Has chance for split-if optimization 1213 1214 return _gvn.transform(result); 1215 } 1216 1217 // Helper method for String intrinsic functions. This version is called 1218 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1219 // to Int nodes containing the lenghts of str1 and str2. 1220 // 1221 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1222 Node* result = NULL; 1223 switch (opcode) { 1224 case Op_StrIndexOf: 1225 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1226 str1_start, cnt1, str2_start, cnt2); 1227 break; 1228 case Op_StrComp: 1229 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1230 str1_start, cnt1, str2_start, cnt2); 1231 break; 1232 case Op_StrEquals: 1233 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1234 str1_start, str2_start, cnt1); 1235 break; 1236 default: 1237 ShouldNotReachHere(); 1238 return NULL; 1239 } 1240 1241 // All these intrinsics have checks. 1242 C->set_has_split_ifs(true); // Has chance for split-if optimization 1243 1244 return _gvn.transform(result); 1245 } 1246 1247 //------------------------------inline_string_compareTo------------------------ 1248 // public int java.lang.String.compareTo(String anotherString); 1249 bool LibraryCallKit::inline_string_compareTo() { 1250 Node* receiver = null_check(argument(0)); 1251 Node* arg = null_check(argument(1)); 1252 if (stopped()) { 1253 return true; 1254 } 1255 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1256 return true; 1257 } 1258 1259 //------------------------------inline_string_equals------------------------ 1260 bool LibraryCallKit::inline_string_equals() { 1261 Node* receiver = null_check_receiver(); 1262 // NOTE: Do not null check argument for String.equals() because spec 1263 // allows to specify NULL as argument. 1264 Node* argument = this->argument(1); 1265 if (stopped()) { 1266 return true; 1267 } 1268 1269 // paths (plus control) merge 1270 RegionNode* region = new (C) RegionNode(5); 1271 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 1272 1273 // does source == target string? 1274 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument)); 1275 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 1276 1277 Node* if_eq = generate_slow_guard(bol, NULL); 1278 if (if_eq != NULL) { 1279 // receiver == argument 1280 phi->init_req(2, intcon(1)); 1281 region->init_req(2, if_eq); 1282 } 1283 1284 // get String klass for instanceOf 1285 ciInstanceKlass* klass = env()->String_klass(); 1286 1287 if (!stopped()) { 1288 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1289 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1))); 1290 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 1291 1292 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1293 //instanceOf == true, fallthrough 1294 1295 if (inst_false != NULL) { 1296 phi->init_req(3, intcon(0)); 1297 region->init_req(3, inst_false); 1298 } 1299 } 1300 1301 if (!stopped()) { 1302 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1303 1304 // Properly cast the argument to String 1305 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type)); 1306 // This path is taken only when argument's type is String:NotNull. 1307 argument = cast_not_null(argument, false); 1308 1309 Node* no_ctrl = NULL; 1310 1311 // Get start addr of receiver 1312 Node* receiver_val = load_String_value(no_ctrl, receiver); 1313 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1314 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1315 1316 // Get length of receiver 1317 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1318 1319 // Get start addr of argument 1320 Node* argument_val = load_String_value(no_ctrl, argument); 1321 Node* argument_offset = load_String_offset(no_ctrl, argument); 1322 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1323 1324 // Get length of argument 1325 Node* argument_cnt = load_String_length(no_ctrl, argument); 1326 1327 // Check for receiver count != argument count 1328 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt)); 1329 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne)); 1330 Node* if_ne = generate_slow_guard(bol, NULL); 1331 if (if_ne != NULL) { 1332 phi->init_req(4, intcon(0)); 1333 region->init_req(4, if_ne); 1334 } 1335 1336 // Check for count == 0 is done by assembler code for StrEquals. 1337 1338 if (!stopped()) { 1339 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1340 phi->init_req(1, equals); 1341 region->init_req(1, control()); 1342 } 1343 } 1344 1345 // post merge 1346 set_control(_gvn.transform(region)); 1347 record_for_igvn(region); 1348 1349 set_result(_gvn.transform(phi)); 1350 return true; 1351 } 1352 1353 //------------------------------inline_array_equals---------------------------- 1354 bool LibraryCallKit::inline_array_equals() { 1355 Node* arg1 = argument(0); 1356 Node* arg2 = argument(1); 1357 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1358 return true; 1359 } 1360 1361 // Java version of String.indexOf(constant string) 1362 // class StringDecl { 1363 // StringDecl(char[] ca) { 1364 // offset = 0; 1365 // count = ca.length; 1366 // value = ca; 1367 // } 1368 // int offset; 1369 // int count; 1370 // char[] value; 1371 // } 1372 // 1373 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1374 // int targetOffset, int cache_i, int md2) { 1375 // int cache = cache_i; 1376 // int sourceOffset = string_object.offset; 1377 // int sourceCount = string_object.count; 1378 // int targetCount = target_object.length; 1379 // 1380 // int targetCountLess1 = targetCount - 1; 1381 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1382 // 1383 // char[] source = string_object.value; 1384 // char[] target = target_object; 1385 // int lastChar = target[targetCountLess1]; 1386 // 1387 // outer_loop: 1388 // for (int i = sourceOffset; i < sourceEnd; ) { 1389 // int src = source[i + targetCountLess1]; 1390 // if (src == lastChar) { 1391 // // With random strings and a 4-character alphabet, 1392 // // reverse matching at this point sets up 0.8% fewer 1393 // // frames, but (paradoxically) makes 0.3% more probes. 1394 // // Since those probes are nearer the lastChar probe, 1395 // // there is may be a net D$ win with reverse matching. 1396 // // But, reversing loop inhibits unroll of inner loop 1397 // // for unknown reason. So, does running outer loop from 1398 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1399 // for (int j = 0; j < targetCountLess1; j++) { 1400 // if (target[targetOffset + j] != source[i+j]) { 1401 // if ((cache & (1 << source[i+j])) == 0) { 1402 // if (md2 < j+1) { 1403 // i += j+1; 1404 // continue outer_loop; 1405 // } 1406 // } 1407 // i += md2; 1408 // continue outer_loop; 1409 // } 1410 // } 1411 // return i - sourceOffset; 1412 // } 1413 // if ((cache & (1 << src)) == 0) { 1414 // i += targetCountLess1; 1415 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1416 // i++; 1417 // } 1418 // return -1; 1419 // } 1420 1421 //------------------------------string_indexOf------------------------ 1422 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1423 jint cache_i, jint md2_i) { 1424 1425 Node* no_ctrl = NULL; 1426 float likely = PROB_LIKELY(0.9); 1427 float unlikely = PROB_UNLIKELY(0.9); 1428 1429 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1430 1431 Node* source = load_String_value(no_ctrl, string_object); 1432 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1433 Node* sourceCount = load_String_length(no_ctrl, string_object); 1434 1435 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1436 jint target_length = target_array->length(); 1437 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1438 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1439 1440 // String.value field is known to be @Stable. 1441 if (UseImplicitStableValues) { 1442 target = cast_array_to_stable(target, target_type); 1443 } 1444 1445 IdealKit kit(this, false, true); 1446 #define __ kit. 1447 Node* zero = __ ConI(0); 1448 Node* one = __ ConI(1); 1449 Node* cache = __ ConI(cache_i); 1450 Node* md2 = __ ConI(md2_i); 1451 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1452 Node* targetCount = __ ConI(target_length); 1453 Node* targetCountLess1 = __ ConI(target_length - 1); 1454 Node* targetOffset = __ ConI(targetOffset_i); 1455 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1456 1457 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1458 Node* outer_loop = __ make_label(2 /* goto */); 1459 Node* return_ = __ make_label(1); 1460 1461 __ set(rtn,__ ConI(-1)); 1462 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1463 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1464 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1465 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1466 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1467 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1468 Node* tpj = __ AddI(targetOffset, __ value(j)); 1469 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1470 Node* ipj = __ AddI(__ value(i), __ value(j)); 1471 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1472 __ if_then(targ, BoolTest::ne, src2); { 1473 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1474 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1475 __ increment(i, __ AddI(__ value(j), one)); 1476 __ goto_(outer_loop); 1477 } __ end_if(); __ dead(j); 1478 }__ end_if(); __ dead(j); 1479 __ increment(i, md2); 1480 __ goto_(outer_loop); 1481 }__ end_if(); 1482 __ increment(j, one); 1483 }__ end_loop(); __ dead(j); 1484 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1485 __ goto_(return_); 1486 }__ end_if(); 1487 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1488 __ increment(i, targetCountLess1); 1489 }__ end_if(); 1490 __ increment(i, one); 1491 __ bind(outer_loop); 1492 }__ end_loop(); __ dead(i); 1493 __ bind(return_); 1494 1495 // Final sync IdealKit and GraphKit. 1496 final_sync(kit); 1497 Node* result = __ value(rtn); 1498 #undef __ 1499 C->set_has_loops(true); 1500 return result; 1501 } 1502 1503 //------------------------------inline_string_indexOf------------------------ 1504 bool LibraryCallKit::inline_string_indexOf() { 1505 Node* receiver = argument(0); 1506 Node* arg = argument(1); 1507 1508 Node* result; 1509 // Disable the use of pcmpestri until it can be guaranteed that 1510 // the load doesn't cross into the uncommited space. 1511 if (Matcher::has_match_rule(Op_StrIndexOf) && 1512 UseSSE42Intrinsics) { 1513 // Generate SSE4.2 version of indexOf 1514 // We currently only have match rules that use SSE4.2 1515 1516 receiver = null_check(receiver); 1517 arg = null_check(arg); 1518 if (stopped()) { 1519 return true; 1520 } 1521 1522 ciInstanceKlass* str_klass = env()->String_klass(); 1523 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1524 1525 // Make the merge point 1526 RegionNode* result_rgn = new (C) RegionNode(4); 1527 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT); 1528 Node* no_ctrl = NULL; 1529 1530 // Get start addr of source string 1531 Node* source = load_String_value(no_ctrl, receiver); 1532 Node* source_offset = load_String_offset(no_ctrl, receiver); 1533 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1534 1535 // Get length of source string 1536 Node* source_cnt = load_String_length(no_ctrl, receiver); 1537 1538 // Get start addr of substring 1539 Node* substr = load_String_value(no_ctrl, arg); 1540 Node* substr_offset = load_String_offset(no_ctrl, arg); 1541 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1542 1543 // Get length of source string 1544 Node* substr_cnt = load_String_length(no_ctrl, arg); 1545 1546 // Check for substr count > string count 1547 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt)); 1548 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt)); 1549 Node* if_gt = generate_slow_guard(bol, NULL); 1550 if (if_gt != NULL) { 1551 result_phi->init_req(2, intcon(-1)); 1552 result_rgn->init_req(2, if_gt); 1553 } 1554 1555 if (!stopped()) { 1556 // Check for substr count == 0 1557 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0))); 1558 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 1559 Node* if_zero = generate_slow_guard(bol, NULL); 1560 if (if_zero != NULL) { 1561 result_phi->init_req(3, intcon(0)); 1562 result_rgn->init_req(3, if_zero); 1563 } 1564 } 1565 1566 if (!stopped()) { 1567 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1568 result_phi->init_req(1, result); 1569 result_rgn->init_req(1, control()); 1570 } 1571 set_control(_gvn.transform(result_rgn)); 1572 record_for_igvn(result_rgn); 1573 result = _gvn.transform(result_phi); 1574 1575 } else { // Use LibraryCallKit::string_indexOf 1576 // don't intrinsify if argument isn't a constant string. 1577 if (!arg->is_Con()) { 1578 return false; 1579 } 1580 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1581 if (str_type == NULL) { 1582 return false; 1583 } 1584 ciInstanceKlass* klass = env()->String_klass(); 1585 ciObject* str_const = str_type->const_oop(); 1586 if (str_const == NULL || str_const->klass() != klass) { 1587 return false; 1588 } 1589 ciInstance* str = str_const->as_instance(); 1590 assert(str != NULL, "must be instance"); 1591 1592 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1593 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1594 1595 int o; 1596 int c; 1597 if (java_lang_String::has_offset_field()) { 1598 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1599 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1600 } else { 1601 o = 0; 1602 c = pat->length(); 1603 } 1604 1605 // constant strings have no offset and count == length which 1606 // simplifies the resulting code somewhat so lets optimize for that. 1607 if (o != 0 || c != pat->length()) { 1608 return false; 1609 } 1610 1611 receiver = null_check(receiver, T_OBJECT); 1612 // NOTE: No null check on the argument is needed since it's a constant String oop. 1613 if (stopped()) { 1614 return true; 1615 } 1616 1617 // The null string as a pattern always returns 0 (match at beginning of string) 1618 if (c == 0) { 1619 set_result(intcon(0)); 1620 return true; 1621 } 1622 1623 // Generate default indexOf 1624 jchar lastChar = pat->char_at(o + (c - 1)); 1625 int cache = 0; 1626 int i; 1627 for (i = 0; i < c - 1; i++) { 1628 assert(i < pat->length(), "out of range"); 1629 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1630 } 1631 1632 int md2 = c; 1633 for (i = 0; i < c - 1; i++) { 1634 assert(i < pat->length(), "out of range"); 1635 if (pat->char_at(o + i) == lastChar) { 1636 md2 = (c - 1) - i; 1637 } 1638 } 1639 1640 result = string_indexOf(receiver, pat, o, cache, md2); 1641 } 1642 set_result(result); 1643 return true; 1644 } 1645 1646 //--------------------------round_double_node-------------------------------- 1647 // Round a double node if necessary. 1648 Node* LibraryCallKit::round_double_node(Node* n) { 1649 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1650 n = _gvn.transform(new (C) RoundDoubleNode(0, n)); 1651 return n; 1652 } 1653 1654 //------------------------------inline_math----------------------------------- 1655 // public static double Math.abs(double) 1656 // public static double Math.sqrt(double) 1657 // public static double Math.log(double) 1658 // public static double Math.log10(double) 1659 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1660 Node* arg = round_double_node(argument(0)); 1661 Node* n = NULL; 1662 switch (id) { 1663 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break; 1664 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break; 1665 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break; 1666 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break; 1667 default: fatal_unexpected_iid(id); break; 1668 } 1669 set_result(_gvn.transform(n)); 1670 return true; 1671 } 1672 1673 //------------------------------inline_trig---------------------------------- 1674 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1675 // argument reduction which will turn into a fast/slow diamond. 1676 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1677 Node* arg = round_double_node(argument(0)); 1678 Node* n = NULL; 1679 1680 switch (id) { 1681 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break; 1682 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break; 1683 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break; 1684 default: fatal_unexpected_iid(id); break; 1685 } 1686 n = _gvn.transform(n); 1687 1688 // Rounding required? Check for argument reduction! 1689 if (Matcher::strict_fp_requires_explicit_rounding) { 1690 static const double pi_4 = 0.7853981633974483; 1691 static const double neg_pi_4 = -0.7853981633974483; 1692 // pi/2 in 80-bit extended precision 1693 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1694 // -pi/2 in 80-bit extended precision 1695 // 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}; 1696 // Cutoff value for using this argument reduction technique 1697 //static const double pi_2_minus_epsilon = 1.564660403643354; 1698 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1699 1700 // Pseudocode for sin: 1701 // if (x <= Math.PI / 4.0) { 1702 // if (x >= -Math.PI / 4.0) return fsin(x); 1703 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1704 // } else { 1705 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1706 // } 1707 // return StrictMath.sin(x); 1708 1709 // Pseudocode for cos: 1710 // if (x <= Math.PI / 4.0) { 1711 // if (x >= -Math.PI / 4.0) return fcos(x); 1712 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1713 // } else { 1714 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1715 // } 1716 // return StrictMath.cos(x); 1717 1718 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1719 // requires a special machine instruction to load it. Instead we'll try 1720 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1721 // probably do the math inside the SIN encoding. 1722 1723 // Make the merge point 1724 RegionNode* r = new (C) RegionNode(3); 1725 Node* phi = new (C) PhiNode(r, Type::DOUBLE); 1726 1727 // Flatten arg so we need only 1 test 1728 Node *abs = _gvn.transform(new (C) AbsDNode(arg)); 1729 // Node for PI/4 constant 1730 Node *pi4 = makecon(TypeD::make(pi_4)); 1731 // Check PI/4 : abs(arg) 1732 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs)); 1733 // Check: If PI/4 < abs(arg) then go slow 1734 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt )); 1735 // Branch either way 1736 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1737 set_control(opt_iff(r,iff)); 1738 1739 // Set fast path result 1740 phi->init_req(2, n); 1741 1742 // Slow path - non-blocking leaf call 1743 Node* call = NULL; 1744 switch (id) { 1745 case vmIntrinsics::_dsin: 1746 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1747 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1748 "Sin", NULL, arg, top()); 1749 break; 1750 case vmIntrinsics::_dcos: 1751 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1752 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1753 "Cos", NULL, arg, top()); 1754 break; 1755 case vmIntrinsics::_dtan: 1756 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1757 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1758 "Tan", NULL, arg, top()); 1759 break; 1760 } 1761 assert(control()->in(0) == call, ""); 1762 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1763 r->init_req(1, control()); 1764 phi->init_req(1, slow_result); 1765 1766 // Post-merge 1767 set_control(_gvn.transform(r)); 1768 record_for_igvn(r); 1769 n = _gvn.transform(phi); 1770 1771 C->set_has_split_ifs(true); // Has chance for split-if optimization 1772 } 1773 set_result(n); 1774 return true; 1775 } 1776 1777 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1778 //------------------- 1779 //result=(result.isNaN())? funcAddr():result; 1780 // Check: If isNaN() by checking result!=result? then either trap 1781 // or go to runtime 1782 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result)); 1783 // Build the boolean node 1784 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq)); 1785 1786 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1787 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1788 // The pow or exp intrinsic returned a NaN, which requires a call 1789 // to the runtime. Recompile with the runtime call. 1790 uncommon_trap(Deoptimization::Reason_intrinsic, 1791 Deoptimization::Action_make_not_entrant); 1792 } 1793 return result; 1794 } else { 1795 // If this inlining ever returned NaN in the past, we compile a call 1796 // to the runtime to properly handle corner cases 1797 1798 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1799 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff)); 1800 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff)); 1801 1802 if (!if_slow->is_top()) { 1803 RegionNode* result_region = new (C) RegionNode(3); 1804 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE); 1805 1806 result_region->init_req(1, if_fast); 1807 result_val->init_req(1, result); 1808 1809 set_control(if_slow); 1810 1811 const TypePtr* no_memory_effects = NULL; 1812 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1813 no_memory_effects, 1814 x, top(), y, y ? top() : NULL); 1815 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0)); 1816 #ifdef ASSERT 1817 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1)); 1818 assert(value_top == top(), "second value must be top"); 1819 #endif 1820 1821 result_region->init_req(2, control()); 1822 result_val->init_req(2, value); 1823 set_control(_gvn.transform(result_region)); 1824 return _gvn.transform(result_val); 1825 } else { 1826 return result; 1827 } 1828 } 1829 } 1830 1831 //------------------------------inline_exp------------------------------------- 1832 // Inline exp instructions, if possible. The Intel hardware only misses 1833 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1834 bool LibraryCallKit::inline_exp() { 1835 Node* arg = round_double_node(argument(0)); 1836 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg)); 1837 1838 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1839 set_result(n); 1840 1841 C->set_has_split_ifs(true); // Has chance for split-if optimization 1842 return true; 1843 } 1844 1845 //------------------------------inline_pow------------------------------------- 1846 // Inline power instructions, if possible. 1847 bool LibraryCallKit::inline_pow() { 1848 // Pseudocode for pow 1849 // if (y == 2) { 1850 // return x * x; 1851 // } else { 1852 // if (x <= 0.0) { 1853 // long longy = (long)y; 1854 // if ((double)longy == y) { // if y is long 1855 // if (y + 1 == y) longy = 0; // huge number: even 1856 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1857 // } else { 1858 // result = NaN; 1859 // } 1860 // } else { 1861 // result = DPow(x,y); 1862 // } 1863 // if (result != result)? { 1864 // result = uncommon_trap() or runtime_call(); 1865 // } 1866 // return result; 1867 // } 1868 1869 Node* x = round_double_node(argument(0)); 1870 Node* y = round_double_node(argument(2)); 1871 1872 Node* result = NULL; 1873 1874 Node* const_two_node = makecon(TypeD::make(2.0)); 1875 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node)); 1876 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq)); 1877 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1878 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node)); 1879 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node)); 1880 1881 RegionNode* region_node = new (C) RegionNode(3); 1882 region_node->init_req(1, if_true); 1883 1884 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE); 1885 // special case for x^y where y == 2, we can convert it to x * x 1886 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x))); 1887 1888 // set control to if_false since we will now process the false branch 1889 set_control(if_false); 1890 1891 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1892 // Short form: skip the fancy tests and just check for NaN result. 1893 result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1894 } else { 1895 // If this inlining ever returned NaN in the past, include all 1896 // checks + call to the runtime. 1897 1898 // Set the merge point for If node with condition of (x <= 0.0) 1899 // There are four possible paths to region node and phi node 1900 RegionNode *r = new (C) RegionNode(4); 1901 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1902 1903 // Build the first if node: if (x <= 0.0) 1904 // Node for 0 constant 1905 Node *zeronode = makecon(TypeD::ZERO); 1906 // Check x:0 1907 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode)); 1908 // Check: If (x<=0) then go complex path 1909 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le )); 1910 // Branch either way 1911 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1912 // Fast path taken; set region slot 3 1913 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1)); 1914 r->init_req(3,fast_taken); // Capture fast-control 1915 1916 // Fast path not-taken, i.e. slow path 1917 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1)); 1918 1919 // Set fast path result 1920 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1921 phi->init_req(3, fast_result); 1922 1923 // Complex path 1924 // Build the second if node (if y is long) 1925 // Node for (long)y 1926 Node *longy = _gvn.transform(new (C) ConvD2LNode(y)); 1927 // Node for (double)((long) y) 1928 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy)); 1929 // Check (double)((long) y) : y 1930 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y)); 1931 // Check if (y isn't long) then go to slow path 1932 1933 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne )); 1934 // Branch either way 1935 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1936 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2)); 1937 1938 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2)); 1939 1940 // Calculate DPow(abs(x), y)*(1 & (long)y) 1941 // Node for constant 1 1942 Node *conone = longcon(1); 1943 // 1& (long)y 1944 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy)); 1945 1946 // A huge number is always even. Detect a huge number by checking 1947 // if y + 1 == y and set integer to be tested for parity to 0. 1948 // Required for corner case: 1949 // (long)9.223372036854776E18 = max_jlong 1950 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1951 // max_jlong is odd but 9.223372036854776E18 is even 1952 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1)))); 1953 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y)); 1954 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq )); 1955 Node* correctedsign = NULL; 1956 if (ConditionalMoveLimit != 0) { 1957 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1958 } else { 1959 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1960 RegionNode *r = new (C) RegionNode(3); 1961 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 1962 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1))); 1963 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1))); 1964 phi->init_req(1, signnode); 1965 phi->init_req(2, longcon(0)); 1966 correctedsign = _gvn.transform(phi); 1967 ylong_path = _gvn.transform(r); 1968 record_for_igvn(r); 1969 } 1970 1971 // zero node 1972 Node *conzero = longcon(0); 1973 // Check (1&(long)y)==0? 1974 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero)); 1975 // Check if (1&(long)y)!=0?, if so the result is negative 1976 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne )); 1977 // abs(x) 1978 Node *absx=_gvn.transform(new (C) AbsDNode(x)); 1979 // abs(x)^y 1980 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y)); 1981 // -abs(x)^y 1982 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy)); 1983 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1984 Node *signresult = NULL; 1985 if (ConditionalMoveLimit != 0) { 1986 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1987 } else { 1988 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1989 RegionNode *r = new (C) RegionNode(3); 1990 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1991 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven))); 1992 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven))); 1993 phi->init_req(1, absxpowy); 1994 phi->init_req(2, negabsxpowy); 1995 signresult = _gvn.transform(phi); 1996 ylong_path = _gvn.transform(r); 1997 record_for_igvn(r); 1998 } 1999 // Set complex path fast result 2000 r->init_req(2, ylong_path); 2001 phi->init_req(2, signresult); 2002 2003 static const jlong nan_bits = CONST64(0x7ff8000000000000); 2004 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 2005 r->init_req(1,slow_path); 2006 phi->init_req(1,slow_result); 2007 2008 // Post merge 2009 set_control(_gvn.transform(r)); 2010 record_for_igvn(r); 2011 result = _gvn.transform(phi); 2012 } 2013 2014 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 2015 2016 // control from finish_pow_exp is now input to the region node 2017 region_node->set_req(2, control()); 2018 // the result from finish_pow_exp is now input to the phi node 2019 phi_node->init_req(2, result); 2020 set_control(_gvn.transform(region_node)); 2021 record_for_igvn(region_node); 2022 set_result(_gvn.transform(phi_node)); 2023 2024 C->set_has_split_ifs(true); // Has chance for split-if optimization 2025 return true; 2026 } 2027 2028 //------------------------------runtime_math----------------------------- 2029 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 2030 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 2031 "must be (DD)D or (D)D type"); 2032 2033 // Inputs 2034 Node* a = round_double_node(argument(0)); 2035 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 2036 2037 const TypePtr* no_memory_effects = NULL; 2038 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 2039 no_memory_effects, 2040 a, top(), b, b ? top() : NULL); 2041 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0)); 2042 #ifdef ASSERT 2043 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1)); 2044 assert(value_top == top(), "second value must be top"); 2045 #endif 2046 2047 set_result(value); 2048 return true; 2049 } 2050 2051 //------------------------------inline_math_native----------------------------- 2052 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 2053 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 2054 switch (id) { 2055 // These intrinsics are not properly supported on all hardware 2056 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 2057 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 2058 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 2059 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2060 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2061 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2062 2063 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2064 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2065 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2066 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2067 2068 // These intrinsics are supported on all hardware 2069 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2070 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2071 2072 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2073 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2074 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2075 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2076 #undef FN_PTR 2077 2078 // These intrinsics are not yet correctly implemented 2079 case vmIntrinsics::_datan2: 2080 return false; 2081 2082 default: 2083 fatal_unexpected_iid(id); 2084 return false; 2085 } 2086 } 2087 2088 static bool is_simple_name(Node* n) { 2089 return (n->req() == 1 // constant 2090 || (n->is_Type() && n->as_Type()->type()->singleton()) 2091 || n->is_Proj() // parameter or return value 2092 || n->is_Phi() // local of some sort 2093 ); 2094 } 2095 2096 //----------------------------inline_min_max----------------------------------- 2097 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2098 set_result(generate_min_max(id, argument(0), argument(1))); 2099 return true; 2100 } 2101 2102 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2103 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) ); 2104 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2105 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check)); 2106 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) ); 2107 2108 { 2109 PreserveJVMState pjvms(this); 2110 PreserveReexecuteState preexecs(this); 2111 jvms()->set_should_reexecute(true); 2112 2113 set_control(slow_path); 2114 set_i_o(i_o()); 2115 2116 uncommon_trap(Deoptimization::Reason_intrinsic, 2117 Deoptimization::Action_none); 2118 } 2119 2120 set_control(fast_path); 2121 set_result(math); 2122 } 2123 2124 template <typename OverflowOp> 2125 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2126 typedef typename OverflowOp::MathOp MathOp; 2127 2128 MathOp* mathOp = new(C) MathOp(arg1, arg2); 2129 Node* operation = _gvn.transform( mathOp ); 2130 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) ); 2131 inline_math_mathExact(operation, ofcheck); 2132 return true; 2133 } 2134 2135 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2136 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2137 } 2138 2139 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2140 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2141 } 2142 2143 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2144 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2145 } 2146 2147 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2148 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2149 } 2150 2151 bool LibraryCallKit::inline_math_negateExactI() { 2152 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2153 } 2154 2155 bool LibraryCallKit::inline_math_negateExactL() { 2156 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2157 } 2158 2159 bool LibraryCallKit::inline_math_multiplyExactI() { 2160 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2161 } 2162 2163 bool LibraryCallKit::inline_math_multiplyExactL() { 2164 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2165 } 2166 2167 Node* 2168 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2169 // These are the candidate return value: 2170 Node* xvalue = x0; 2171 Node* yvalue = y0; 2172 2173 if (xvalue == yvalue) { 2174 return xvalue; 2175 } 2176 2177 bool want_max = (id == vmIntrinsics::_max); 2178 2179 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2180 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2181 if (txvalue == NULL || tyvalue == NULL) return top(); 2182 // This is not really necessary, but it is consistent with a 2183 // hypothetical MaxINode::Value method: 2184 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2185 2186 // %%% This folding logic should (ideally) be in a different place. 2187 // Some should be inside IfNode, and there to be a more reliable 2188 // transformation of ?: style patterns into cmoves. We also want 2189 // more powerful optimizations around cmove and min/max. 2190 2191 // Try to find a dominating comparison of these guys. 2192 // It can simplify the index computation for Arrays.copyOf 2193 // and similar uses of System.arraycopy. 2194 // First, compute the normalized version of CmpI(x, y). 2195 int cmp_op = Op_CmpI; 2196 Node* xkey = xvalue; 2197 Node* ykey = yvalue; 2198 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey)); 2199 if (ideal_cmpxy->is_Cmp()) { 2200 // E.g., if we have CmpI(length - offset, count), 2201 // it might idealize to CmpI(length, count + offset) 2202 cmp_op = ideal_cmpxy->Opcode(); 2203 xkey = ideal_cmpxy->in(1); 2204 ykey = ideal_cmpxy->in(2); 2205 } 2206 2207 // Start by locating any relevant comparisons. 2208 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2209 Node* cmpxy = NULL; 2210 Node* cmpyx = NULL; 2211 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2212 Node* cmp = start_from->fast_out(k); 2213 if (cmp->outcnt() > 0 && // must have prior uses 2214 cmp->in(0) == NULL && // must be context-independent 2215 cmp->Opcode() == cmp_op) { // right kind of compare 2216 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2217 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2218 } 2219 } 2220 2221 const int NCMPS = 2; 2222 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2223 int cmpn; 2224 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2225 if (cmps[cmpn] != NULL) break; // find a result 2226 } 2227 if (cmpn < NCMPS) { 2228 // Look for a dominating test that tells us the min and max. 2229 int depth = 0; // Limit search depth for speed 2230 Node* dom = control(); 2231 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2232 if (++depth >= 100) break; 2233 Node* ifproj = dom; 2234 if (!ifproj->is_Proj()) continue; 2235 Node* iff = ifproj->in(0); 2236 if (!iff->is_If()) continue; 2237 Node* bol = iff->in(1); 2238 if (!bol->is_Bool()) continue; 2239 Node* cmp = bol->in(1); 2240 if (cmp == NULL) continue; 2241 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2242 if (cmps[cmpn] == cmp) break; 2243 if (cmpn == NCMPS) continue; 2244 BoolTest::mask btest = bol->as_Bool()->_test._test; 2245 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2246 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2247 // At this point, we know that 'x btest y' is true. 2248 switch (btest) { 2249 case BoolTest::eq: 2250 // They are proven equal, so we can collapse the min/max. 2251 // Either value is the answer. Choose the simpler. 2252 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2253 return yvalue; 2254 return xvalue; 2255 case BoolTest::lt: // x < y 2256 case BoolTest::le: // x <= y 2257 return (want_max ? yvalue : xvalue); 2258 case BoolTest::gt: // x > y 2259 case BoolTest::ge: // x >= y 2260 return (want_max ? xvalue : yvalue); 2261 } 2262 } 2263 } 2264 2265 // We failed to find a dominating test. 2266 // Let's pick a test that might GVN with prior tests. 2267 Node* best_bol = NULL; 2268 BoolTest::mask best_btest = BoolTest::illegal; 2269 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2270 Node* cmp = cmps[cmpn]; 2271 if (cmp == NULL) continue; 2272 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2273 Node* bol = cmp->fast_out(j); 2274 if (!bol->is_Bool()) continue; 2275 BoolTest::mask btest = bol->as_Bool()->_test._test; 2276 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2277 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2278 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2279 best_bol = bol->as_Bool(); 2280 best_btest = btest; 2281 } 2282 } 2283 } 2284 2285 Node* answer_if_true = NULL; 2286 Node* answer_if_false = NULL; 2287 switch (best_btest) { 2288 default: 2289 if (cmpxy == NULL) 2290 cmpxy = ideal_cmpxy; 2291 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt)); 2292 // and fall through: 2293 case BoolTest::lt: // x < y 2294 case BoolTest::le: // x <= y 2295 answer_if_true = (want_max ? yvalue : xvalue); 2296 answer_if_false = (want_max ? xvalue : yvalue); 2297 break; 2298 case BoolTest::gt: // x > y 2299 case BoolTest::ge: // x >= y 2300 answer_if_true = (want_max ? xvalue : yvalue); 2301 answer_if_false = (want_max ? yvalue : xvalue); 2302 break; 2303 } 2304 2305 jint hi, lo; 2306 if (want_max) { 2307 // We can sharpen the minimum. 2308 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2309 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2310 } else { 2311 // We can sharpen the maximum. 2312 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2313 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2314 } 2315 2316 // Use a flow-free graph structure, to avoid creating excess control edges 2317 // which could hinder other optimizations. 2318 // Since Math.min/max is often used with arraycopy, we want 2319 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2320 Node* cmov = CMoveNode::make(C, NULL, best_bol, 2321 answer_if_false, answer_if_true, 2322 TypeInt::make(lo, hi, widen)); 2323 2324 return _gvn.transform(cmov); 2325 2326 /* 2327 // This is not as desirable as it may seem, since Min and Max 2328 // nodes do not have a full set of optimizations. 2329 // And they would interfere, anyway, with 'if' optimizations 2330 // and with CMoveI canonical forms. 2331 switch (id) { 2332 case vmIntrinsics::_min: 2333 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2334 case vmIntrinsics::_max: 2335 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2336 default: 2337 ShouldNotReachHere(); 2338 } 2339 */ 2340 } 2341 2342 inline int 2343 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2344 const TypePtr* base_type = TypePtr::NULL_PTR; 2345 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2346 if (base_type == NULL) { 2347 // Unknown type. 2348 return Type::AnyPtr; 2349 } else if (base_type == TypePtr::NULL_PTR) { 2350 // Since this is a NULL+long form, we have to switch to a rawptr. 2351 base = _gvn.transform(new (C) CastX2PNode(offset)); 2352 offset = MakeConX(0); 2353 return Type::RawPtr; 2354 } else if (base_type->base() == Type::RawPtr) { 2355 return Type::RawPtr; 2356 } else if (base_type->isa_oopptr()) { 2357 // Base is never null => always a heap address. 2358 if (base_type->ptr() == TypePtr::NotNull) { 2359 return Type::OopPtr; 2360 } 2361 // Offset is small => always a heap address. 2362 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2363 if (offset_type != NULL && 2364 base_type->offset() == 0 && // (should always be?) 2365 offset_type->_lo >= 0 && 2366 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2367 return Type::OopPtr; 2368 } 2369 // Otherwise, it might either be oop+off or NULL+addr. 2370 return Type::AnyPtr; 2371 } else { 2372 // No information: 2373 return Type::AnyPtr; 2374 } 2375 } 2376 2377 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2378 int kind = classify_unsafe_addr(base, offset); 2379 if (kind == Type::RawPtr) { 2380 return basic_plus_adr(top(), base, offset); 2381 } else { 2382 return basic_plus_adr(base, offset); 2383 } 2384 } 2385 2386 //--------------------------inline_number_methods----------------------------- 2387 // inline int Integer.numberOfLeadingZeros(int) 2388 // inline int Long.numberOfLeadingZeros(long) 2389 // 2390 // inline int Integer.numberOfTrailingZeros(int) 2391 // inline int Long.numberOfTrailingZeros(long) 2392 // 2393 // inline int Integer.bitCount(int) 2394 // inline int Long.bitCount(long) 2395 // 2396 // inline char Character.reverseBytes(char) 2397 // inline short Short.reverseBytes(short) 2398 // inline int Integer.reverseBytes(int) 2399 // inline long Long.reverseBytes(long) 2400 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2401 Node* arg = argument(0); 2402 Node* n = NULL; 2403 switch (id) { 2404 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break; 2405 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break; 2406 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break; 2407 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break; 2408 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break; 2409 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break; 2410 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break; 2411 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break; 2412 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break; 2413 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break; 2414 default: fatal_unexpected_iid(id); break; 2415 } 2416 set_result(_gvn.transform(n)); 2417 return true; 2418 } 2419 2420 //----------------------------inline_unsafe_access---------------------------- 2421 2422 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2423 2424 // Helper that guards and inserts a pre-barrier. 2425 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2426 Node* pre_val, bool need_mem_bar) { 2427 // We could be accessing the referent field of a reference object. If so, when G1 2428 // is enabled, we need to log the value in the referent field in an SATB buffer. 2429 // This routine performs some compile time filters and generates suitable 2430 // runtime filters that guard the pre-barrier code. 2431 // Also add memory barrier for non volatile load from the referent field 2432 // to prevent commoning of loads across safepoint. 2433 if (!UseG1GC && !need_mem_bar) 2434 return; 2435 2436 // Some compile time checks. 2437 2438 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2439 const TypeX* otype = offset->find_intptr_t_type(); 2440 if (otype != NULL && otype->is_con() && 2441 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2442 // Constant offset but not the reference_offset so just return 2443 return; 2444 } 2445 2446 // We only need to generate the runtime guards for instances. 2447 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2448 if (btype != NULL) { 2449 if (btype->isa_aryptr()) { 2450 // Array type so nothing to do 2451 return; 2452 } 2453 2454 const TypeInstPtr* itype = btype->isa_instptr(); 2455 if (itype != NULL) { 2456 // Can the klass of base_oop be statically determined to be 2457 // _not_ a sub-class of Reference and _not_ Object? 2458 ciKlass* klass = itype->klass(); 2459 if ( klass->is_loaded() && 2460 !klass->is_subtype_of(env()->Reference_klass()) && 2461 !env()->Object_klass()->is_subtype_of(klass)) { 2462 return; 2463 } 2464 } 2465 } 2466 2467 // The compile time filters did not reject base_oop/offset so 2468 // we need to generate the following runtime filters 2469 // 2470 // if (offset == java_lang_ref_Reference::_reference_offset) { 2471 // if (instance_of(base, java.lang.ref.Reference)) { 2472 // pre_barrier(_, pre_val, ...); 2473 // } 2474 // } 2475 2476 float likely = PROB_LIKELY( 0.999); 2477 float unlikely = PROB_UNLIKELY(0.999); 2478 2479 IdealKit ideal(this); 2480 #define __ ideal. 2481 2482 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2483 2484 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2485 // Update graphKit memory and control from IdealKit. 2486 sync_kit(ideal); 2487 2488 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2489 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2490 2491 // Update IdealKit memory and control from graphKit. 2492 __ sync_kit(this); 2493 2494 Node* one = __ ConI(1); 2495 // is_instof == 0 if base_oop == NULL 2496 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2497 2498 // Update graphKit from IdeakKit. 2499 sync_kit(ideal); 2500 2501 // Use the pre-barrier to record the value in the referent field 2502 pre_barrier(false /* do_load */, 2503 __ ctrl(), 2504 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2505 pre_val /* pre_val */, 2506 T_OBJECT); 2507 if (need_mem_bar) { 2508 // Add memory barrier to prevent commoning reads from this field 2509 // across safepoint since GC can change its value. 2510 insert_mem_bar(Op_MemBarCPUOrder); 2511 } 2512 // Update IdealKit from graphKit. 2513 __ sync_kit(this); 2514 2515 } __ end_if(); // _ref_type != ref_none 2516 } __ end_if(); // offset == referent_offset 2517 2518 // Final sync IdealKit and GraphKit. 2519 final_sync(ideal); 2520 #undef __ 2521 } 2522 2523 2524 // Interpret Unsafe.fieldOffset cookies correctly: 2525 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2526 2527 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2528 // Attempt to infer a sharper value type from the offset and base type. 2529 ciKlass* sharpened_klass = NULL; 2530 2531 // See if it is an instance field, with an object type. 2532 if (alias_type->field() != NULL) { 2533 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2534 if (alias_type->field()->type()->is_klass()) { 2535 sharpened_klass = alias_type->field()->type()->as_klass(); 2536 } 2537 } 2538 2539 // See if it is a narrow oop array. 2540 if (adr_type->isa_aryptr()) { 2541 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2542 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2543 if (elem_type != NULL) { 2544 sharpened_klass = elem_type->klass(); 2545 } 2546 } 2547 } 2548 2549 // The sharpened class might be unloaded if there is no class loader 2550 // contraint in place. 2551 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2552 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2553 2554 #ifndef PRODUCT 2555 if (C->print_intrinsics() || C->print_inlining()) { 2556 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2557 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2558 } 2559 #endif 2560 // Sharpen the value type. 2561 return tjp; 2562 } 2563 return NULL; 2564 } 2565 2566 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) { 2567 if (callee()->is_static()) return false; // caller must have the capability! 2568 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2569 2570 #ifndef PRODUCT 2571 { 2572 ResourceMark rm; 2573 // Check the signatures. 2574 ciSignature* sig = callee()->signature(); 2575 #ifdef ASSERT 2576 if (!is_store) { 2577 // Object getObject(Object base, int/long offset), etc. 2578 BasicType rtype = sig->return_type()->basic_type(); 2579 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2580 rtype = T_ADDRESS; // it is really a C void* 2581 assert(rtype == type, "getter must return the expected value"); 2582 if (!is_native_ptr) { 2583 assert(sig->count() == 2, "oop getter has 2 arguments"); 2584 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2585 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2586 } else { 2587 assert(sig->count() == 1, "native getter has 1 argument"); 2588 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2589 } 2590 } else { 2591 // void putObject(Object base, int/long offset, Object x), etc. 2592 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2593 if (!is_native_ptr) { 2594 assert(sig->count() == 3, "oop putter has 3 arguments"); 2595 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2596 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2597 } else { 2598 assert(sig->count() == 2, "native putter has 2 arguments"); 2599 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2600 } 2601 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2602 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2603 vtype = T_ADDRESS; // it is really a C void* 2604 assert(vtype == type, "putter must accept the expected value"); 2605 } 2606 #endif // ASSERT 2607 } 2608 #endif //PRODUCT 2609 2610 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2611 2612 Node* receiver = argument(0); // type: oop 2613 2614 // Build address expression. See the code in inline_unsafe_prefetch. 2615 Node* adr; 2616 Node* heap_base_oop = top(); 2617 Node* offset = top(); 2618 Node* val; 2619 2620 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2621 Node* base = argument(1); // type: oop 2622 2623 if (!is_native_ptr) { 2624 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2625 offset = argument(2); // type: long 2626 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2627 // to be plain byte offsets, which are also the same as those accepted 2628 // by oopDesc::field_base. 2629 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2630 "fieldOffset must be byte-scaled"); 2631 // 32-bit machines ignore the high half! 2632 offset = ConvL2X(offset); 2633 adr = make_unsafe_address(base, offset); 2634 heap_base_oop = base; 2635 val = is_store ? argument(4) : NULL; 2636 } else { 2637 Node* ptr = argument(1); // type: long 2638 ptr = ConvL2X(ptr); // adjust Java long to machine word 2639 adr = make_unsafe_address(NULL, ptr); 2640 val = is_store ? argument(3) : NULL; 2641 } 2642 2643 if ((_gvn.type(base)->isa_ptr() == TypePtr::NULL_PTR) && type == T_OBJECT) { 2644 return false; // off-heap oop accesses are not supported 2645 } 2646 2647 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2648 2649 // Try to categorize the address. 2650 Compile::AliasType* alias_type = C->alias_type(adr_type); 2651 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2652 2653 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2654 alias_type->adr_type() == TypeAryPtr::RANGE) { 2655 return false; // not supported 2656 } 2657 2658 bool mismatched = false; 2659 BasicType bt = alias_type->basic_type(); 2660 if (bt != T_ILLEGAL) { 2661 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2662 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2663 // Alias type doesn't differentiate between byte[] and boolean[]). 2664 // Use address type to get the element type. 2665 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2666 } 2667 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2668 // accessing an array field with getObject is not a mismatch 2669 bt = T_OBJECT; 2670 } 2671 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2672 // Don't intrinsify mismatched object accesses 2673 return false; 2674 } 2675 mismatched = (bt != type); 2676 } 2677 2678 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2679 2680 // First guess at the value type. 2681 const Type *value_type = Type::get_const_basic_type(type); 2682 2683 // We will need memory barriers unless we can determine a unique 2684 // alias category for this reference. (Note: If for some reason 2685 // the barriers get omitted and the unsafe reference begins to "pollute" 2686 // the alias analysis of the rest of the graph, either Compile::can_alias 2687 // or Compile::must_alias will throw a diagnostic assert.) 2688 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2689 2690 // If we are reading the value of the referent field of a Reference 2691 // object (either by using Unsafe directly or through reflection) 2692 // then, if G1 is enabled, we need to record the referent in an 2693 // SATB log buffer using the pre-barrier mechanism. 2694 // Also we need to add memory barrier to prevent commoning reads 2695 // from this field across safepoint since GC can change its value. 2696 bool need_read_barrier = !is_native_ptr && !is_store && 2697 offset != top() && heap_base_oop != top(); 2698 2699 if (!is_store && type == T_OBJECT) { 2700 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2701 if (tjp != NULL) { 2702 value_type = tjp; 2703 } 2704 } 2705 2706 receiver = null_check(receiver); 2707 if (stopped()) { 2708 return true; 2709 } 2710 // Heap pointers get a null-check from the interpreter, 2711 // as a courtesy. However, this is not guaranteed by Unsafe, 2712 // and it is not possible to fully distinguish unintended nulls 2713 // from intended ones in this API. 2714 2715 if (is_volatile) { 2716 // We need to emit leading and trailing CPU membars (see below) in 2717 // addition to memory membars when is_volatile. This is a little 2718 // too strong, but avoids the need to insert per-alias-type 2719 // volatile membars (for stores; compare Parse::do_put_xxx), which 2720 // we cannot do effectively here because we probably only have a 2721 // rough approximation of type. 2722 need_mem_bar = true; 2723 // For Stores, place a memory ordering barrier now. 2724 if (is_store) { 2725 insert_mem_bar(Op_MemBarRelease); 2726 } else { 2727 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2728 insert_mem_bar(Op_MemBarVolatile); 2729 } 2730 } 2731 } 2732 2733 // Memory barrier to prevent normal and 'unsafe' accesses from 2734 // bypassing each other. Happens after null checks, so the 2735 // exception paths do not take memory state from the memory barrier, 2736 // so there's no problems making a strong assert about mixing users 2737 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2738 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2739 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2740 2741 if (!is_store) { 2742 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2743 // To be valid, unsafe loads may depend on other conditions than 2744 // the one that guards them: pin the Load node 2745 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched); 2746 // load value 2747 switch (type) { 2748 case T_BOOLEAN: 2749 case T_CHAR: 2750 case T_BYTE: 2751 case T_SHORT: 2752 case T_INT: 2753 case T_LONG: 2754 case T_FLOAT: 2755 case T_DOUBLE: 2756 break; 2757 case T_OBJECT: 2758 if (need_read_barrier) { 2759 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2760 } 2761 break; 2762 case T_ADDRESS: 2763 // Cast to an int type. 2764 p = _gvn.transform(new (C) CastP2XNode(NULL, p)); 2765 p = ConvX2UL(p); 2766 break; 2767 default: 2768 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2769 break; 2770 } 2771 // The load node has the control of the preceding MemBarCPUOrder. All 2772 // following nodes will have the control of the MemBarCPUOrder inserted at 2773 // the end of this method. So, pushing the load onto the stack at a later 2774 // point is fine. 2775 set_result(p); 2776 } else { 2777 // place effect of store into memory 2778 switch (type) { 2779 case T_DOUBLE: 2780 val = dstore_rounding(val); 2781 break; 2782 case T_ADDRESS: 2783 // Repackage the long as a pointer. 2784 val = ConvL2X(val); 2785 val = _gvn.transform(new (C) CastX2PNode(val)); 2786 break; 2787 } 2788 2789 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2790 if (type == T_OBJECT ) { 2791 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2792 } else { 2793 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched); 2794 } 2795 } 2796 2797 if (is_volatile) { 2798 if (!is_store) { 2799 insert_mem_bar(Op_MemBarAcquire); 2800 } else { 2801 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2802 insert_mem_bar(Op_MemBarVolatile); 2803 } 2804 } 2805 } 2806 2807 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2808 2809 return true; 2810 } 2811 2812 //----------------------------inline_unsafe_prefetch---------------------------- 2813 2814 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2815 #ifndef PRODUCT 2816 { 2817 ResourceMark rm; 2818 // Check the signatures. 2819 ciSignature* sig = callee()->signature(); 2820 #ifdef ASSERT 2821 // Object getObject(Object base, int/long offset), etc. 2822 BasicType rtype = sig->return_type()->basic_type(); 2823 if (!is_native_ptr) { 2824 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2825 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2826 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2827 } else { 2828 assert(sig->count() == 1, "native prefetch has 1 argument"); 2829 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2830 } 2831 #endif // ASSERT 2832 } 2833 #endif // !PRODUCT 2834 2835 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2836 2837 const int idx = is_static ? 0 : 1; 2838 if (!is_static) { 2839 null_check_receiver(); 2840 if (stopped()) { 2841 return true; 2842 } 2843 } 2844 2845 // Build address expression. See the code in inline_unsafe_access. 2846 Node *adr; 2847 if (!is_native_ptr) { 2848 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2849 Node* base = argument(idx + 0); // type: oop 2850 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2851 Node* offset = argument(idx + 1); // type: long 2852 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2853 // to be plain byte offsets, which are also the same as those accepted 2854 // by oopDesc::field_base. 2855 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2856 "fieldOffset must be byte-scaled"); 2857 // 32-bit machines ignore the high half! 2858 offset = ConvL2X(offset); 2859 adr = make_unsafe_address(base, offset); 2860 } else { 2861 Node* ptr = argument(idx + 0); // type: long 2862 ptr = ConvL2X(ptr); // adjust Java long to machine word 2863 adr = make_unsafe_address(NULL, ptr); 2864 } 2865 2866 // Generate the read or write prefetch 2867 Node *prefetch; 2868 if (is_store) { 2869 prefetch = new (C) PrefetchWriteNode(i_o(), adr); 2870 } else { 2871 prefetch = new (C) PrefetchReadNode(i_o(), adr); 2872 } 2873 prefetch->init_req(0, control()); 2874 set_i_o(_gvn.transform(prefetch)); 2875 2876 return true; 2877 } 2878 2879 //----------------------------inline_unsafe_load_store---------------------------- 2880 // This method serves a couple of different customers (depending on LoadStoreKind): 2881 // 2882 // LS_cmpxchg: 2883 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2884 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2885 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2886 // 2887 // LS_xadd: 2888 // public int getAndAddInt( Object o, long offset, int delta) 2889 // public long getAndAddLong(Object o, long offset, long delta) 2890 // 2891 // LS_xchg: 2892 // int getAndSet(Object o, long offset, int newValue) 2893 // long getAndSet(Object o, long offset, long newValue) 2894 // Object getAndSet(Object o, long offset, Object newValue) 2895 // 2896 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2897 // This basic scheme here is the same as inline_unsafe_access, but 2898 // differs in enough details that combining them would make the code 2899 // overly confusing. (This is a true fact! I originally combined 2900 // them, but even I was confused by it!) As much code/comments as 2901 // possible are retained from inline_unsafe_access though to make 2902 // the correspondences clearer. - dl 2903 2904 if (callee()->is_static()) return false; // caller must have the capability! 2905 2906 #ifndef PRODUCT 2907 BasicType rtype; 2908 { 2909 ResourceMark rm; 2910 // Check the signatures. 2911 ciSignature* sig = callee()->signature(); 2912 rtype = sig->return_type()->basic_type(); 2913 if (kind == LS_xadd || kind == LS_xchg) { 2914 // Check the signatures. 2915 #ifdef ASSERT 2916 assert(rtype == type, "get and set must return the expected type"); 2917 assert(sig->count() == 3, "get and set has 3 arguments"); 2918 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2919 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2920 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2921 #endif // ASSERT 2922 } else if (kind == LS_cmpxchg) { 2923 // Check the signatures. 2924 #ifdef ASSERT 2925 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2926 assert(sig->count() == 4, "CAS has 4 arguments"); 2927 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2928 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2929 #endif // ASSERT 2930 } else { 2931 ShouldNotReachHere(); 2932 } 2933 } 2934 #endif //PRODUCT 2935 2936 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2937 2938 // Get arguments: 2939 Node* receiver = NULL; 2940 Node* base = NULL; 2941 Node* offset = NULL; 2942 Node* oldval = NULL; 2943 Node* newval = NULL; 2944 if (kind == LS_cmpxchg) { 2945 const bool two_slot_type = type2size[type] == 2; 2946 receiver = argument(0); // type: oop 2947 base = argument(1); // type: oop 2948 offset = argument(2); // type: long 2949 oldval = argument(4); // type: oop, int, or long 2950 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2951 } else if (kind == LS_xadd || kind == LS_xchg){ 2952 receiver = argument(0); // type: oop 2953 base = argument(1); // type: oop 2954 offset = argument(2); // type: long 2955 oldval = NULL; 2956 newval = argument(4); // type: oop, int, or long 2957 } 2958 2959 // Build field offset expression. 2960 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2961 // to be plain byte offsets, which are also the same as those accepted 2962 // by oopDesc::field_base. 2963 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2964 // 32-bit machines ignore the high half of long offsets 2965 offset = ConvL2X(offset); 2966 Node* adr = make_unsafe_address(base, offset); 2967 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2968 2969 Compile::AliasType* alias_type = C->alias_type(adr_type); 2970 BasicType bt = alias_type->basic_type(); 2971 if (bt != T_ILLEGAL && 2972 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2973 // Don't intrinsify mismatched object accesses. 2974 return false; 2975 } 2976 2977 // For CAS, unlike inline_unsafe_access, there seems no point in 2978 // trying to refine types. Just use the coarse types here. 2979 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2980 const Type *value_type = Type::get_const_basic_type(type); 2981 2982 if (kind == LS_xchg && type == T_OBJECT) { 2983 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2984 if (tjp != NULL) { 2985 value_type = tjp; 2986 } 2987 } 2988 2989 // Null check receiver. 2990 receiver = null_check(receiver); 2991 if (stopped()) { 2992 return true; 2993 } 2994 2995 int alias_idx = C->get_alias_index(adr_type); 2996 2997 // Memory-model-wise, a LoadStore acts like a little synchronized 2998 // block, so needs barriers on each side. These don't translate 2999 // into actual barriers on most machines, but we still need rest of 3000 // compiler to respect ordering. 3001 3002 insert_mem_bar(Op_MemBarRelease); 3003 insert_mem_bar(Op_MemBarCPUOrder); 3004 3005 // 4984716: MemBars must be inserted before this 3006 // memory node in order to avoid a false 3007 // dependency which will confuse the scheduler. 3008 Node *mem = memory(alias_idx); 3009 3010 // For now, we handle only those cases that actually exist: ints, 3011 // longs, and Object. Adding others should be straightforward. 3012 Node* load_store = NULL; 3013 switch(type) { 3014 case T_INT: 3015 if (kind == LS_xadd) { 3016 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type)); 3017 } else if (kind == LS_xchg) { 3018 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type)); 3019 } else if (kind == LS_cmpxchg) { 3020 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 3021 } else { 3022 ShouldNotReachHere(); 3023 } 3024 break; 3025 case T_LONG: 3026 if (kind == LS_xadd) { 3027 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type)); 3028 } else if (kind == LS_xchg) { 3029 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type)); 3030 } else if (kind == LS_cmpxchg) { 3031 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 3032 } else { 3033 ShouldNotReachHere(); 3034 } 3035 break; 3036 case T_OBJECT: 3037 // Transformation of a value which could be NULL pointer (CastPP #NULL) 3038 // could be delayed during Parse (for example, in adjust_map_after_if()). 3039 // Execute transformation here to avoid barrier generation in such case. 3040 if (_gvn.type(newval) == TypePtr::NULL_PTR) 3041 newval = _gvn.makecon(TypePtr::NULL_PTR); 3042 3043 // Reference stores need a store barrier. 3044 if (kind == LS_xchg) { 3045 // If pre-barrier must execute before the oop store, old value will require do_load here. 3046 if (!can_move_pre_barrier()) { 3047 pre_barrier(true /* do_load*/, 3048 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 3049 NULL /* pre_val*/, 3050 T_OBJECT); 3051 } // Else move pre_barrier to use load_store value, see below. 3052 } else if (kind == LS_cmpxchg) { 3053 // Same as for newval above: 3054 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 3055 oldval = _gvn.makecon(TypePtr::NULL_PTR); 3056 } 3057 // The only known value which might get overwritten is oldval. 3058 pre_barrier(false /* do_load */, 3059 control(), NULL, NULL, max_juint, NULL, NULL, 3060 oldval /* pre_val */, 3061 T_OBJECT); 3062 } else { 3063 ShouldNotReachHere(); 3064 } 3065 3066 #ifdef _LP64 3067 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3068 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 3069 if (kind == LS_xchg) { 3070 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr, 3071 newval_enc, adr_type, value_type->make_narrowoop())); 3072 } else { 3073 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 3074 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 3075 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr, 3076 newval_enc, oldval_enc)); 3077 } 3078 } else 3079 #endif 3080 { 3081 if (kind == LS_xchg) { 3082 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 3083 } else { 3084 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 3085 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 3086 } 3087 } 3088 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 3089 break; 3090 default: 3091 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 3092 break; 3093 } 3094 3095 // SCMemProjNodes represent the memory state of a LoadStore. Their 3096 // main role is to prevent LoadStore nodes from being optimized away 3097 // when their results aren't used. 3098 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store)); 3099 set_memory(proj, alias_idx); 3100 3101 if (type == T_OBJECT && kind == LS_xchg) { 3102 #ifdef _LP64 3103 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3104 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type())); 3105 } 3106 #endif 3107 if (can_move_pre_barrier()) { 3108 // Don't need to load pre_val. The old value is returned by load_store. 3109 // The pre_barrier can execute after the xchg as long as no safepoint 3110 // gets inserted between them. 3111 pre_barrier(false /* do_load */, 3112 control(), NULL, NULL, max_juint, NULL, NULL, 3113 load_store /* pre_val */, 3114 T_OBJECT); 3115 } 3116 } 3117 3118 // Add the trailing membar surrounding the access 3119 insert_mem_bar(Op_MemBarCPUOrder); 3120 insert_mem_bar(Op_MemBarAcquire); 3121 3122 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3123 set_result(load_store); 3124 return true; 3125 } 3126 3127 //----------------------------inline_unsafe_ordered_store---------------------- 3128 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3129 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3130 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3131 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3132 // This is another variant of inline_unsafe_access, differing in 3133 // that it always issues store-store ("release") barrier and ensures 3134 // store-atomicity (which only matters for "long"). 3135 3136 if (callee()->is_static()) return false; // caller must have the capability! 3137 3138 #ifndef PRODUCT 3139 { 3140 ResourceMark rm; 3141 // Check the signatures. 3142 ciSignature* sig = callee()->signature(); 3143 #ifdef ASSERT 3144 BasicType rtype = sig->return_type()->basic_type(); 3145 assert(rtype == T_VOID, "must return void"); 3146 assert(sig->count() == 3, "has 3 arguments"); 3147 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3148 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3149 #endif // ASSERT 3150 } 3151 #endif //PRODUCT 3152 3153 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3154 3155 // Get arguments: 3156 Node* receiver = argument(0); // type: oop 3157 Node* base = argument(1); // type: oop 3158 Node* offset = argument(2); // type: long 3159 Node* val = argument(4); // type: oop, int, or long 3160 3161 // Null check receiver. 3162 receiver = null_check(receiver); 3163 if (stopped()) { 3164 return true; 3165 } 3166 3167 // Build field offset expression. 3168 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3169 // 32-bit machines ignore the high half of long offsets 3170 offset = ConvL2X(offset); 3171 Node* adr = make_unsafe_address(base, offset); 3172 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3173 const Type *value_type = Type::get_const_basic_type(type); 3174 Compile::AliasType* alias_type = C->alias_type(adr_type); 3175 3176 insert_mem_bar(Op_MemBarRelease); 3177 insert_mem_bar(Op_MemBarCPUOrder); 3178 // Ensure that the store is atomic for longs: 3179 const bool require_atomic_access = true; 3180 Node* store; 3181 if (type == T_OBJECT) // reference stores need a store barrier. 3182 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3183 else { 3184 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3185 } 3186 insert_mem_bar(Op_MemBarCPUOrder); 3187 return true; 3188 } 3189 3190 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3191 // Regardless of form, don't allow previous ld/st to move down, 3192 // then issue acquire, release, or volatile mem_bar. 3193 insert_mem_bar(Op_MemBarCPUOrder); 3194 switch(id) { 3195 case vmIntrinsics::_loadFence: 3196 insert_mem_bar(Op_LoadFence); 3197 return true; 3198 case vmIntrinsics::_storeFence: 3199 insert_mem_bar(Op_StoreFence); 3200 return true; 3201 case vmIntrinsics::_fullFence: 3202 insert_mem_bar(Op_MemBarVolatile); 3203 return true; 3204 default: 3205 fatal_unexpected_iid(id); 3206 return false; 3207 } 3208 } 3209 3210 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3211 if (!kls->is_Con()) { 3212 return true; 3213 } 3214 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3215 if (klsptr == NULL) { 3216 return true; 3217 } 3218 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3219 // don't need a guard for a klass that is already initialized 3220 return !ik->is_initialized(); 3221 } 3222 3223 //----------------------------inline_unsafe_allocate--------------------------- 3224 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3225 bool LibraryCallKit::inline_unsafe_allocate() { 3226 if (callee()->is_static()) return false; // caller must have the capability! 3227 3228 null_check_receiver(); // null-check, then ignore 3229 Node* cls = null_check(argument(1)); 3230 if (stopped()) return true; 3231 3232 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3233 kls = null_check(kls); 3234 if (stopped()) return true; // argument was like int.class 3235 3236 Node* test = NULL; 3237 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3238 // Note: The argument might still be an illegal value like 3239 // Serializable.class or Object[].class. The runtime will handle it. 3240 // But we must make an explicit check for initialization. 3241 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3242 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3243 // can generate code to load it as unsigned byte. 3244 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3245 Node* bits = intcon(InstanceKlass::fully_initialized); 3246 test = _gvn.transform(new (C) SubINode(inst, bits)); 3247 // The 'test' is non-zero if we need to take a slow path. 3248 } 3249 3250 Node* obj = new_instance(kls, test); 3251 set_result(obj); 3252 return true; 3253 } 3254 3255 #ifdef JFR_HAVE_INTRINSICS 3256 /* 3257 * oop -> myklass 3258 * myklass->trace_id |= USED 3259 * return myklass->trace_id & ~0x3 3260 */ 3261 bool LibraryCallKit::inline_native_classID() { 3262 Node* cls = null_check(argument(0), T_OBJECT); 3263 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3264 kls = null_check(kls, T_OBJECT); 3265 3266 ByteSize offset = KLASS_TRACE_ID_OFFSET; 3267 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3268 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3269 3270 Node* clsused = longcon(0x01l); // set the class bit 3271 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused)); 3272 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3273 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3274 3275 #ifdef TRACE_ID_META_BITS 3276 Node* mbits = longcon(~TRACE_ID_META_BITS); 3277 tvalue = _gvn.transform(new (C) AndLNode(tvalue, mbits)); 3278 #endif 3279 #ifdef TRACE_ID_SHIFT 3280 Node* cbits = intcon(TRACE_ID_SHIFT); 3281 tvalue = _gvn.transform(new (C) URShiftLNode(tvalue, cbits)); 3282 #endif 3283 3284 set_result(tvalue); 3285 return true; 3286 } 3287 3288 bool LibraryCallKit::inline_native_getEventWriter() { 3289 Node* tls_ptr = _gvn.transform(new (C) ThreadLocalNode()); 3290 3291 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr, 3292 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR) 3293 ); 3294 3295 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered); 3296 3297 Node* jobj_cmp_null = _gvn.transform( new (C) CmpPNode(jobj, null()) ); 3298 Node* test_jobj_eq_null = _gvn.transform( new (C) BoolNode(jobj_cmp_null, BoolTest::eq) ); 3299 3300 IfNode* iff_jobj_null = 3301 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN); 3302 3303 enum { _normal_path = 1, 3304 _null_path = 2, 3305 PATH_LIMIT }; 3306 3307 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT); 3308 PhiNode* result_val = new (C) PhiNode(result_rgn, TypePtr::BOTTOM); 3309 3310 Node* jobj_is_null = _gvn.transform(new (C) IfTrueNode(iff_jobj_null)); 3311 result_rgn->init_req(_null_path, jobj_is_null); 3312 result_val->init_req(_null_path, null()); 3313 3314 Node* jobj_is_not_null = _gvn.transform(new (C) IfFalseNode(iff_jobj_null)); 3315 result_rgn->init_req(_normal_path, jobj_is_not_null); 3316 3317 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered); 3318 result_val->init_req(_normal_path, res); 3319 3320 set_result(result_rgn, result_val); 3321 3322 return true; 3323 } 3324 #endif // JFR_HAVE_INTRINSICS 3325 3326 //------------------------inline_native_time_funcs-------------- 3327 // inline code for System.currentTimeMillis() and System.nanoTime() 3328 // these have the same type and signature 3329 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3330 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3331 const TypePtr* no_memory_effects = NULL; 3332 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3333 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0)); 3334 #ifdef ASSERT 3335 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1)); 3336 assert(value_top == top(), "second value must be top"); 3337 #endif 3338 set_result(value); 3339 return true; 3340 } 3341 3342 //------------------------inline_native_currentThread------------------ 3343 bool LibraryCallKit::inline_native_currentThread() { 3344 Node* junk = NULL; 3345 set_result(generate_current_thread(junk)); 3346 return true; 3347 } 3348 3349 //------------------------inline_native_isInterrupted------------------ 3350 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3351 bool LibraryCallKit::inline_native_isInterrupted() { 3352 // Add a fast path to t.isInterrupted(clear_int): 3353 // (t == Thread.current() && 3354 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3355 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3356 // So, in the common case that the interrupt bit is false, 3357 // we avoid making a call into the VM. Even if the interrupt bit 3358 // is true, if the clear_int argument is false, we avoid the VM call. 3359 // However, if the receiver is not currentThread, we must call the VM, 3360 // because there must be some locking done around the operation. 3361 3362 // We only go to the fast case code if we pass two guards. 3363 // Paths which do not pass are accumulated in the slow_region. 3364 3365 enum { 3366 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3367 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3368 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3369 PATH_LIMIT 3370 }; 3371 3372 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3373 // out of the function. 3374 insert_mem_bar(Op_MemBarCPUOrder); 3375 3376 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT); 3377 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL); 3378 3379 RegionNode* slow_region = new (C) RegionNode(1); 3380 record_for_igvn(slow_region); 3381 3382 // (a) Receiving thread must be the current thread. 3383 Node* rec_thr = argument(0); 3384 Node* tls_ptr = NULL; 3385 Node* cur_thr = generate_current_thread(tls_ptr); 3386 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr)); 3387 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne)); 3388 3389 generate_slow_guard(bol_thr, slow_region); 3390 3391 // (b) Interrupt bit on TLS must be false. 3392 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3393 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3394 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3395 3396 // Set the control input on the field _interrupted read to prevent it floating up. 3397 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3398 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0))); 3399 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne)); 3400 3401 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3402 3403 // First fast path: if (!TLS._interrupted) return false; 3404 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit)); 3405 result_rgn->init_req(no_int_result_path, false_bit); 3406 result_val->init_req(no_int_result_path, intcon(0)); 3407 3408 // drop through to next case 3409 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit))); 3410 3411 #ifndef TARGET_OS_FAMILY_windows 3412 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3413 Node* clr_arg = argument(1); 3414 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0))); 3415 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne)); 3416 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3417 3418 // Second fast path: ... else if (!clear_int) return true; 3419 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg)); 3420 result_rgn->init_req(no_clear_result_path, false_arg); 3421 result_val->init_req(no_clear_result_path, intcon(1)); 3422 3423 // drop through to next case 3424 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg))); 3425 #else 3426 // To return true on Windows you must read the _interrupted field 3427 // and check the the event state i.e. take the slow path. 3428 #endif // TARGET_OS_FAMILY_windows 3429 3430 // (d) Otherwise, go to the slow path. 3431 slow_region->add_req(control()); 3432 set_control( _gvn.transform(slow_region)); 3433 3434 if (stopped()) { 3435 // There is no slow path. 3436 result_rgn->init_req(slow_result_path, top()); 3437 result_val->init_req(slow_result_path, top()); 3438 } else { 3439 // non-virtual because it is a private non-static 3440 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3441 3442 Node* slow_val = set_results_for_java_call(slow_call); 3443 // this->control() comes from set_results_for_java_call 3444 3445 Node* fast_io = slow_call->in(TypeFunc::I_O); 3446 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3447 3448 // These two phis are pre-filled with copies of of the fast IO and Memory 3449 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3450 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3451 3452 result_rgn->init_req(slow_result_path, control()); 3453 result_io ->init_req(slow_result_path, i_o()); 3454 result_mem->init_req(slow_result_path, reset_memory()); 3455 result_val->init_req(slow_result_path, slow_val); 3456 3457 set_all_memory(_gvn.transform(result_mem)); 3458 set_i_o( _gvn.transform(result_io)); 3459 } 3460 3461 C->set_has_split_ifs(true); // Has chance for split-if optimization 3462 set_result(result_rgn, result_val); 3463 return true; 3464 } 3465 3466 //---------------------------load_mirror_from_klass---------------------------- 3467 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3468 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3469 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3470 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3471 } 3472 3473 //-----------------------load_klass_from_mirror_common------------------------- 3474 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3475 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3476 // and branch to the given path on the region. 3477 // If never_see_null, take an uncommon trap on null, so we can optimistically 3478 // compile for the non-null case. 3479 // If the region is NULL, force never_see_null = true. 3480 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3481 bool never_see_null, 3482 RegionNode* region, 3483 int null_path, 3484 int offset) { 3485 if (region == NULL) never_see_null = true; 3486 Node* p = basic_plus_adr(mirror, offset); 3487 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3488 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3489 Node* null_ctl = top(); 3490 kls = null_check_oop(kls, &null_ctl, never_see_null); 3491 if (region != NULL) { 3492 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3493 region->init_req(null_path, null_ctl); 3494 } else { 3495 assert(null_ctl == top(), "no loose ends"); 3496 } 3497 return kls; 3498 } 3499 3500 //--------------------(inline_native_Class_query helpers)--------------------- 3501 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3502 // Fall through if (mods & mask) == bits, take the guard otherwise. 3503 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3504 // Branch around if the given klass has the given modifier bit set. 3505 // Like generate_guard, adds a new path onto the region. 3506 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3507 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3508 Node* mask = intcon(modifier_mask); 3509 Node* bits = intcon(modifier_bits); 3510 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask)); 3511 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits)); 3512 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 3513 return generate_fair_guard(bol, region); 3514 } 3515 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3516 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3517 } 3518 3519 //-------------------------inline_native_Class_query------------------- 3520 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3521 const Type* return_type = TypeInt::BOOL; 3522 Node* prim_return_value = top(); // what happens if it's a primitive class? 3523 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3524 bool expect_prim = false; // most of these guys expect to work on refs 3525 3526 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3527 3528 Node* mirror = argument(0); 3529 Node* obj = top(); 3530 3531 switch (id) { 3532 case vmIntrinsics::_isInstance: 3533 // nothing is an instance of a primitive type 3534 prim_return_value = intcon(0); 3535 obj = argument(1); 3536 break; 3537 case vmIntrinsics::_getModifiers: 3538 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3539 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3540 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3541 break; 3542 case vmIntrinsics::_isInterface: 3543 prim_return_value = intcon(0); 3544 break; 3545 case vmIntrinsics::_isArray: 3546 prim_return_value = intcon(0); 3547 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3548 break; 3549 case vmIntrinsics::_isPrimitive: 3550 prim_return_value = intcon(1); 3551 expect_prim = true; // obviously 3552 break; 3553 case vmIntrinsics::_getSuperclass: 3554 prim_return_value = null(); 3555 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3556 break; 3557 case vmIntrinsics::_getComponentType: 3558 prim_return_value = null(); 3559 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3560 break; 3561 case vmIntrinsics::_getClassAccessFlags: 3562 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3563 return_type = TypeInt::INT; // not bool! 6297094 3564 break; 3565 default: 3566 fatal_unexpected_iid(id); 3567 break; 3568 } 3569 3570 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3571 if (mirror_con == NULL) return false; // cannot happen? 3572 3573 #ifndef PRODUCT 3574 if (C->print_intrinsics() || C->print_inlining()) { 3575 ciType* k = mirror_con->java_mirror_type(); 3576 if (k) { 3577 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3578 k->print_name(); 3579 tty->cr(); 3580 } 3581 } 3582 #endif 3583 3584 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3585 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3586 record_for_igvn(region); 3587 PhiNode* phi = new (C) PhiNode(region, return_type); 3588 3589 // The mirror will never be null of Reflection.getClassAccessFlags, however 3590 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3591 // if it is. See bug 4774291. 3592 3593 // For Reflection.getClassAccessFlags(), the null check occurs in 3594 // the wrong place; see inline_unsafe_access(), above, for a similar 3595 // situation. 3596 mirror = null_check(mirror); 3597 // If mirror or obj is dead, only null-path is taken. 3598 if (stopped()) return true; 3599 3600 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3601 3602 // Now load the mirror's klass metaobject, and null-check it. 3603 // Side-effects region with the control path if the klass is null. 3604 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3605 // If kls is null, we have a primitive mirror. 3606 phi->init_req(_prim_path, prim_return_value); 3607 if (stopped()) { set_result(region, phi); return true; } 3608 bool safe_for_replace = (region->in(_prim_path) == top()); 3609 3610 Node* p; // handy temp 3611 Node* null_ctl; 3612 3613 // Now that we have the non-null klass, we can perform the real query. 3614 // For constant classes, the query will constant-fold in LoadNode::Value. 3615 Node* query_value = top(); 3616 switch (id) { 3617 case vmIntrinsics::_isInstance: 3618 // nothing is an instance of a primitive type 3619 query_value = gen_instanceof(obj, kls, safe_for_replace); 3620 break; 3621 3622 case vmIntrinsics::_getModifiers: 3623 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3624 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3625 break; 3626 3627 case vmIntrinsics::_isInterface: 3628 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3629 if (generate_interface_guard(kls, region) != NULL) 3630 // A guard was added. If the guard is taken, it was an interface. 3631 phi->add_req(intcon(1)); 3632 // If we fall through, it's a plain class. 3633 query_value = intcon(0); 3634 break; 3635 3636 case vmIntrinsics::_isArray: 3637 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3638 if (generate_array_guard(kls, region) != NULL) 3639 // A guard was added. If the guard is taken, it was an array. 3640 phi->add_req(intcon(1)); 3641 // If we fall through, it's a plain class. 3642 query_value = intcon(0); 3643 break; 3644 3645 case vmIntrinsics::_isPrimitive: 3646 query_value = intcon(0); // "normal" path produces false 3647 break; 3648 3649 case vmIntrinsics::_getSuperclass: 3650 // The rules here are somewhat unfortunate, but we can still do better 3651 // with random logic than with a JNI call. 3652 // Interfaces store null or Object as _super, but must report null. 3653 // Arrays store an intermediate super as _super, but must report Object. 3654 // Other types can report the actual _super. 3655 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3656 if (generate_interface_guard(kls, region) != NULL) 3657 // A guard was added. If the guard is taken, it was an interface. 3658 phi->add_req(null()); 3659 if (generate_array_guard(kls, region) != NULL) 3660 // A guard was added. If the guard is taken, it was an array. 3661 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3662 // If we fall through, it's a plain class. Get its _super. 3663 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3664 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3665 null_ctl = top(); 3666 kls = null_check_oop(kls, &null_ctl); 3667 if (null_ctl != top()) { 3668 // If the guard is taken, Object.superClass is null (both klass and mirror). 3669 region->add_req(null_ctl); 3670 phi ->add_req(null()); 3671 } 3672 if (!stopped()) { 3673 query_value = load_mirror_from_klass(kls); 3674 } 3675 break; 3676 3677 case vmIntrinsics::_getComponentType: 3678 if (generate_array_guard(kls, region) != NULL) { 3679 // Be sure to pin the oop load to the guard edge just created: 3680 Node* is_array_ctrl = region->in(region->req()-1); 3681 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); 3682 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3683 phi->add_req(cmo); 3684 } 3685 query_value = null(); // non-array case is null 3686 break; 3687 3688 case vmIntrinsics::_getClassAccessFlags: 3689 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3690 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3691 break; 3692 3693 default: 3694 fatal_unexpected_iid(id); 3695 break; 3696 } 3697 3698 // Fall-through is the normal case of a query to a real class. 3699 phi->init_req(1, query_value); 3700 region->init_req(1, control()); 3701 3702 C->set_has_split_ifs(true); // Has chance for split-if optimization 3703 set_result(region, phi); 3704 return true; 3705 } 3706 3707 //--------------------------inline_native_subtype_check------------------------ 3708 // This intrinsic takes the JNI calls out of the heart of 3709 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3710 bool LibraryCallKit::inline_native_subtype_check() { 3711 // Pull both arguments off the stack. 3712 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3713 args[0] = argument(0); 3714 args[1] = argument(1); 3715 Node* klasses[2]; // corresponding Klasses: superk, subk 3716 klasses[0] = klasses[1] = top(); 3717 3718 enum { 3719 // A full decision tree on {superc is prim, subc is prim}: 3720 _prim_0_path = 1, // {P,N} => false 3721 // {P,P} & superc!=subc => false 3722 _prim_same_path, // {P,P} & superc==subc => true 3723 _prim_1_path, // {N,P} => false 3724 _ref_subtype_path, // {N,N} & subtype check wins => true 3725 _both_ref_path, // {N,N} & subtype check loses => false 3726 PATH_LIMIT 3727 }; 3728 3729 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3730 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 3731 record_for_igvn(region); 3732 3733 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3734 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3735 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3736 3737 // First null-check both mirrors and load each mirror's klass metaobject. 3738 int which_arg; 3739 for (which_arg = 0; which_arg <= 1; which_arg++) { 3740 Node* arg = args[which_arg]; 3741 arg = null_check(arg); 3742 if (stopped()) break; 3743 args[which_arg] = arg; 3744 3745 Node* p = basic_plus_adr(arg, class_klass_offset); 3746 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3747 klasses[which_arg] = _gvn.transform(kls); 3748 } 3749 3750 // Having loaded both klasses, test each for null. 3751 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3752 for (which_arg = 0; which_arg <= 1; which_arg++) { 3753 Node* kls = klasses[which_arg]; 3754 Node* null_ctl = top(); 3755 kls = null_check_oop(kls, &null_ctl, never_see_null); 3756 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3757 region->init_req(prim_path, null_ctl); 3758 if (stopped()) break; 3759 klasses[which_arg] = kls; 3760 } 3761 3762 if (!stopped()) { 3763 // now we have two reference types, in klasses[0..1] 3764 Node* subk = klasses[1]; // the argument to isAssignableFrom 3765 Node* superk = klasses[0]; // the receiver 3766 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3767 // now we have a successful reference subtype check 3768 region->set_req(_ref_subtype_path, control()); 3769 } 3770 3771 // If both operands are primitive (both klasses null), then 3772 // we must return true when they are identical primitives. 3773 // It is convenient to test this after the first null klass check. 3774 set_control(region->in(_prim_0_path)); // go back to first null check 3775 if (!stopped()) { 3776 // Since superc is primitive, make a guard for the superc==subc case. 3777 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1])); 3778 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq)); 3779 generate_guard(bol_eq, region, PROB_FAIR); 3780 if (region->req() == PATH_LIMIT+1) { 3781 // A guard was added. If the added guard is taken, superc==subc. 3782 region->swap_edges(PATH_LIMIT, _prim_same_path); 3783 region->del_req(PATH_LIMIT); 3784 } 3785 region->set_req(_prim_0_path, control()); // Not equal after all. 3786 } 3787 3788 // these are the only paths that produce 'true': 3789 phi->set_req(_prim_same_path, intcon(1)); 3790 phi->set_req(_ref_subtype_path, intcon(1)); 3791 3792 // pull together the cases: 3793 assert(region->req() == PATH_LIMIT, "sane region"); 3794 for (uint i = 1; i < region->req(); i++) { 3795 Node* ctl = region->in(i); 3796 if (ctl == NULL || ctl == top()) { 3797 region->set_req(i, top()); 3798 phi ->set_req(i, top()); 3799 } else if (phi->in(i) == NULL) { 3800 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3801 } 3802 } 3803 3804 set_control(_gvn.transform(region)); 3805 set_result(_gvn.transform(phi)); 3806 return true; 3807 } 3808 3809 //---------------------generate_array_guard_common------------------------ 3810 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3811 bool obj_array, bool not_array) { 3812 // If obj_array/non_array==false/false: 3813 // Branch around if the given klass is in fact an array (either obj or prim). 3814 // If obj_array/non_array==false/true: 3815 // Branch around if the given klass is not an array klass of any kind. 3816 // If obj_array/non_array==true/true: 3817 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3818 // If obj_array/non_array==true/false: 3819 // Branch around if the kls is an oop array (Object[] or subtype) 3820 // 3821 // Like generate_guard, adds a new path onto the region. 3822 jint layout_con = 0; 3823 Node* layout_val = get_layout_helper(kls, layout_con); 3824 if (layout_val == NULL) { 3825 bool query = (obj_array 3826 ? Klass::layout_helper_is_objArray(layout_con) 3827 : Klass::layout_helper_is_array(layout_con)); 3828 if (query == not_array) { 3829 return NULL; // never a branch 3830 } else { // always a branch 3831 Node* always_branch = control(); 3832 if (region != NULL) 3833 region->add_req(always_branch); 3834 set_control(top()); 3835 return always_branch; 3836 } 3837 } 3838 // Now test the correct condition. 3839 jint nval = (obj_array 3840 ? (jint)(Klass::_lh_array_tag_type_value 3841 << Klass::_lh_array_tag_shift) 3842 : Klass::_lh_neutral_value); 3843 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval))); 3844 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3845 // invert the test if we are looking for a non-array 3846 if (not_array) btest = BoolTest(btest).negate(); 3847 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest)); 3848 return generate_fair_guard(bol, region); 3849 } 3850 3851 3852 //-----------------------inline_native_newArray-------------------------- 3853 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3854 bool LibraryCallKit::inline_native_newArray() { 3855 Node* mirror = argument(0); 3856 Node* count_val = argument(1); 3857 3858 mirror = null_check(mirror); 3859 // If mirror or obj is dead, only null-path is taken. 3860 if (stopped()) return true; 3861 3862 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3863 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3864 PhiNode* result_val = new(C) PhiNode(result_reg, 3865 TypeInstPtr::NOTNULL); 3866 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3867 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3868 TypePtr::BOTTOM); 3869 3870 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3871 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3872 result_reg, _slow_path); 3873 Node* normal_ctl = control(); 3874 Node* no_array_ctl = result_reg->in(_slow_path); 3875 3876 // Generate code for the slow case. We make a call to newArray(). 3877 set_control(no_array_ctl); 3878 if (!stopped()) { 3879 // Either the input type is void.class, or else the 3880 // array klass has not yet been cached. Either the 3881 // ensuing call will throw an exception, or else it 3882 // will cache the array klass for next time. 3883 PreserveJVMState pjvms(this); 3884 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3885 Node* slow_result = set_results_for_java_call(slow_call); 3886 // this->control() comes from set_results_for_java_call 3887 result_reg->set_req(_slow_path, control()); 3888 result_val->set_req(_slow_path, slow_result); 3889 result_io ->set_req(_slow_path, i_o()); 3890 result_mem->set_req(_slow_path, reset_memory()); 3891 } 3892 3893 set_control(normal_ctl); 3894 if (!stopped()) { 3895 // Normal case: The array type has been cached in the java.lang.Class. 3896 // The following call works fine even if the array type is polymorphic. 3897 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3898 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3899 result_reg->init_req(_normal_path, control()); 3900 result_val->init_req(_normal_path, obj); 3901 result_io ->init_req(_normal_path, i_o()); 3902 result_mem->init_req(_normal_path, reset_memory()); 3903 } 3904 3905 // Return the combined state. 3906 set_i_o( _gvn.transform(result_io) ); 3907 set_all_memory( _gvn.transform(result_mem)); 3908 3909 C->set_has_split_ifs(true); // Has chance for split-if optimization 3910 set_result(result_reg, result_val); 3911 return true; 3912 } 3913 3914 //----------------------inline_native_getLength-------------------------- 3915 // public static native int java.lang.reflect.Array.getLength(Object array); 3916 bool LibraryCallKit::inline_native_getLength() { 3917 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3918 3919 Node* array = null_check(argument(0)); 3920 // If array is dead, only null-path is taken. 3921 if (stopped()) return true; 3922 3923 // Deoptimize if it is a non-array. 3924 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3925 3926 if (non_array != NULL) { 3927 PreserveJVMState pjvms(this); 3928 set_control(non_array); 3929 uncommon_trap(Deoptimization::Reason_intrinsic, 3930 Deoptimization::Action_maybe_recompile); 3931 } 3932 3933 // If control is dead, only non-array-path is taken. 3934 if (stopped()) return true; 3935 3936 // The works fine even if the array type is polymorphic. 3937 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3938 Node* result = load_array_length(array); 3939 3940 C->set_has_split_ifs(true); // Has chance for split-if optimization 3941 set_result(result); 3942 return true; 3943 } 3944 3945 //------------------------inline_array_copyOf---------------------------- 3946 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3947 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3948 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3949 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3950 3951 // Get the arguments. 3952 Node* original = argument(0); 3953 Node* start = is_copyOfRange? argument(1): intcon(0); 3954 Node* end = is_copyOfRange? argument(2): argument(1); 3955 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3956 3957 Node* newcopy = NULL; 3958 3959 // Set the original stack and the reexecute bit for the interpreter to reexecute 3960 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3961 { PreserveReexecuteState preexecs(this); 3962 jvms()->set_should_reexecute(true); 3963 3964 array_type_mirror = null_check(array_type_mirror); 3965 original = null_check(original); 3966 3967 // Check if a null path was taken unconditionally. 3968 if (stopped()) return true; 3969 3970 Node* orig_length = load_array_length(original); 3971 3972 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3973 klass_node = null_check(klass_node); 3974 3975 RegionNode* bailout = new (C) RegionNode(1); 3976 record_for_igvn(bailout); 3977 3978 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3979 // Bail out if that is so. 3980 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3981 if (not_objArray != NULL) { 3982 // Improve the klass node's type from the new optimistic assumption: 3983 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3984 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3985 Node* cast = new (C) CastPPNode(klass_node, akls); 3986 cast->init_req(0, control()); 3987 klass_node = _gvn.transform(cast); 3988 } 3989 3990 // Bail out if either start or end is negative. 3991 generate_negative_guard(start, bailout, &start); 3992 generate_negative_guard(end, bailout, &end); 3993 3994 Node* length = end; 3995 if (_gvn.type(start) != TypeInt::ZERO) { 3996 length = _gvn.transform(new (C) SubINode(end, start)); 3997 } 3998 3999 // Bail out if length is negative. 4000 // Without this the new_array would throw 4001 // NegativeArraySizeException but IllegalArgumentException is what 4002 // should be thrown 4003 generate_negative_guard(length, bailout, &length); 4004 4005 if (bailout->req() > 1) { 4006 PreserveJVMState pjvms(this); 4007 set_control(_gvn.transform(bailout)); 4008 uncommon_trap(Deoptimization::Reason_intrinsic, 4009 Deoptimization::Action_maybe_recompile); 4010 } 4011 4012 if (!stopped()) { 4013 // How many elements will we copy from the original? 4014 // The answer is MinI(orig_length - start, length). 4015 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start)); 4016 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 4017 4018 newcopy = new_array(klass_node, length, 0); // no argments to push 4019 4020 // Generate a direct call to the right arraycopy function(s). 4021 // We know the copy is disjoint but we might not know if the 4022 // oop stores need checking. 4023 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 4024 // This will fail a store-check if x contains any non-nulls. 4025 bool disjoint_bases = true; 4026 // if start > orig_length then the length of the copy may be 4027 // negative. 4028 bool length_never_negative = !is_copyOfRange; 4029 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4030 original, start, newcopy, intcon(0), moved, 4031 disjoint_bases, length_never_negative); 4032 } 4033 } // original reexecute is set back here 4034 4035 C->set_has_split_ifs(true); // Has chance for split-if optimization 4036 if (!stopped()) { 4037 set_result(newcopy); 4038 } 4039 return true; 4040 } 4041 4042 4043 //----------------------generate_virtual_guard--------------------------- 4044 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4045 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4046 RegionNode* slow_region) { 4047 ciMethod* method = callee(); 4048 int vtable_index = method->vtable_index(); 4049 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4050 err_msg_res("bad index %d", vtable_index)); 4051 // Get the Method* out of the appropriate vtable entry. 4052 int entry_offset = (InstanceKlass::vtable_start_offset() + 4053 vtable_index*vtableEntry::size()) * wordSize + 4054 vtableEntry::method_offset_in_bytes(); 4055 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4056 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4057 4058 // Compare the target method with the expected method (e.g., Object.hashCode). 4059 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4060 4061 Node* native_call = makecon(native_call_addr); 4062 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call)); 4063 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne)); 4064 4065 return generate_slow_guard(test_native, slow_region); 4066 } 4067 4068 //-----------------------generate_method_call---------------------------- 4069 // Use generate_method_call to make a slow-call to the real 4070 // method if the fast path fails. An alternative would be to 4071 // use a stub like OptoRuntime::slow_arraycopy_Java. 4072 // This only works for expanding the current library call, 4073 // not another intrinsic. (E.g., don't use this for making an 4074 // arraycopy call inside of the copyOf intrinsic.) 4075 CallJavaNode* 4076 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4077 // When compiling the intrinsic method itself, do not use this technique. 4078 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4079 4080 ciMethod* method = callee(); 4081 // ensure the JVMS we have will be correct for this call 4082 guarantee(method_id == method->intrinsic_id(), "must match"); 4083 4084 const TypeFunc* tf = TypeFunc::make(method); 4085 CallJavaNode* slow_call; 4086 if (is_static) { 4087 assert(!is_virtual, ""); 4088 slow_call = new(C) CallStaticJavaNode(C, tf, 4089 SharedRuntime::get_resolve_static_call_stub(), 4090 method, bci()); 4091 } else if (is_virtual) { 4092 null_check_receiver(); 4093 int vtable_index = Method::invalid_vtable_index; 4094 if (UseInlineCaches) { 4095 // Suppress the vtable call 4096 } else { 4097 // hashCode and clone are not a miranda methods, 4098 // so the vtable index is fixed. 4099 // No need to use the linkResolver to get it. 4100 vtable_index = method->vtable_index(); 4101 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4102 err_msg_res("bad index %d", vtable_index)); 4103 } 4104 slow_call = new(C) CallDynamicJavaNode(tf, 4105 SharedRuntime::get_resolve_virtual_call_stub(), 4106 method, vtable_index, bci()); 4107 } else { // neither virtual nor static: opt_virtual 4108 null_check_receiver(); 4109 slow_call = new(C) CallStaticJavaNode(C, tf, 4110 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4111 method, bci()); 4112 slow_call->set_optimized_virtual(true); 4113 } 4114 set_arguments_for_java_call(slow_call); 4115 set_edges_for_java_call(slow_call); 4116 return slow_call; 4117 } 4118 4119 4120 /** 4121 * Build special case code for calls to hashCode on an object. This call may 4122 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4123 * slightly different code. 4124 */ 4125 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4126 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4127 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4128 4129 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4130 4131 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4132 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT); 4133 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 4134 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4135 Node* obj = NULL; 4136 if (!is_static) { 4137 // Check for hashing null object 4138 obj = null_check_receiver(); 4139 if (stopped()) return true; // unconditionally null 4140 result_reg->init_req(_null_path, top()); 4141 result_val->init_req(_null_path, top()); 4142 } else { 4143 // Do a null check, and return zero if null. 4144 // System.identityHashCode(null) == 0 4145 obj = argument(0); 4146 Node* null_ctl = top(); 4147 obj = null_check_oop(obj, &null_ctl); 4148 result_reg->init_req(_null_path, null_ctl); 4149 result_val->init_req(_null_path, _gvn.intcon(0)); 4150 } 4151 4152 // Unconditionally null? Then return right away. 4153 if (stopped()) { 4154 set_control( result_reg->in(_null_path)); 4155 if (!stopped()) 4156 set_result(result_val->in(_null_path)); 4157 return true; 4158 } 4159 4160 // We only go to the fast case code if we pass a number of guards. The 4161 // paths which do not pass are accumulated in the slow_region. 4162 RegionNode* slow_region = new (C) RegionNode(1); 4163 record_for_igvn(slow_region); 4164 4165 // If this is a virtual call, we generate a funny guard. We pull out 4166 // the vtable entry corresponding to hashCode() from the target object. 4167 // If the target method which we are calling happens to be the native 4168 // Object hashCode() method, we pass the guard. We do not need this 4169 // guard for non-virtual calls -- the caller is known to be the native 4170 // Object hashCode(). 4171 if (is_virtual) { 4172 // After null check, get the object's klass. 4173 Node* obj_klass = load_object_klass(obj); 4174 generate_virtual_guard(obj_klass, slow_region); 4175 } 4176 4177 // Get the header out of the object, use LoadMarkNode when available 4178 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4179 // The control of the load must be NULL. Otherwise, the load can move before 4180 // the null check after castPP removal. 4181 Node* no_ctrl = NULL; 4182 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4183 4184 // Test the header to see if it is unlocked. 4185 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4186 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask)); 4187 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4188 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val)); 4189 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne)); 4190 4191 generate_slow_guard(test_unlocked, slow_region); 4192 4193 // Get the hash value and check to see that it has been properly assigned. 4194 // We depend on hash_mask being at most 32 bits and avoid the use of 4195 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4196 // vm: see markOop.hpp. 4197 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4198 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4199 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift)); 4200 // This hack lets the hash bits live anywhere in the mark object now, as long 4201 // as the shift drops the relevant bits into the low 32 bits. Note that 4202 // Java spec says that HashCode is an int so there's no point in capturing 4203 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4204 hshifted_header = ConvX2I(hshifted_header); 4205 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask)); 4206 4207 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4208 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val)); 4209 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq)); 4210 4211 generate_slow_guard(test_assigned, slow_region); 4212 4213 Node* init_mem = reset_memory(); 4214 // fill in the rest of the null path: 4215 result_io ->init_req(_null_path, i_o()); 4216 result_mem->init_req(_null_path, init_mem); 4217 4218 result_val->init_req(_fast_path, hash_val); 4219 result_reg->init_req(_fast_path, control()); 4220 result_io ->init_req(_fast_path, i_o()); 4221 result_mem->init_req(_fast_path, init_mem); 4222 4223 // Generate code for the slow case. We make a call to hashCode(). 4224 set_control(_gvn.transform(slow_region)); 4225 if (!stopped()) { 4226 // No need for PreserveJVMState, because we're using up the present state. 4227 set_all_memory(init_mem); 4228 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4229 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4230 Node* slow_result = set_results_for_java_call(slow_call); 4231 // this->control() comes from set_results_for_java_call 4232 result_reg->init_req(_slow_path, control()); 4233 result_val->init_req(_slow_path, slow_result); 4234 result_io ->set_req(_slow_path, i_o()); 4235 result_mem ->set_req(_slow_path, reset_memory()); 4236 } 4237 4238 // Return the combined state. 4239 set_i_o( _gvn.transform(result_io) ); 4240 set_all_memory( _gvn.transform(result_mem)); 4241 4242 set_result(result_reg, result_val); 4243 return true; 4244 } 4245 4246 //---------------------------inline_native_getClass---------------------------- 4247 // public final native Class<?> java.lang.Object.getClass(); 4248 // 4249 // Build special case code for calls to getClass on an object. 4250 bool LibraryCallKit::inline_native_getClass() { 4251 Node* obj = null_check_receiver(); 4252 if (stopped()) return true; 4253 set_result(load_mirror_from_klass(load_object_klass(obj))); 4254 return true; 4255 } 4256 4257 //-----------------inline_native_Reflection_getCallerClass--------------------- 4258 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4259 // 4260 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4261 // 4262 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4263 // in that it must skip particular security frames and checks for 4264 // caller sensitive methods. 4265 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4266 #ifndef PRODUCT 4267 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4268 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4269 } 4270 #endif 4271 4272 if (!jvms()->has_method()) { 4273 #ifndef PRODUCT 4274 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4275 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4276 } 4277 #endif 4278 return false; 4279 } 4280 4281 // Walk back up the JVM state to find the caller at the required 4282 // depth. 4283 JVMState* caller_jvms = jvms(); 4284 4285 // Cf. JVM_GetCallerClass 4286 // NOTE: Start the loop at depth 1 because the current JVM state does 4287 // not include the Reflection.getCallerClass() frame. 4288 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4289 ciMethod* m = caller_jvms->method(); 4290 switch (n) { 4291 case 0: 4292 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4293 break; 4294 case 1: 4295 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4296 if (!m->caller_sensitive()) { 4297 #ifndef PRODUCT 4298 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4299 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4300 } 4301 #endif 4302 return false; // bail-out; let JVM_GetCallerClass do the work 4303 } 4304 break; 4305 default: 4306 if (!m->is_ignored_by_security_stack_walk()) { 4307 // We have reached the desired frame; return the holder class. 4308 // Acquire method holder as java.lang.Class and push as constant. 4309 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4310 ciInstance* caller_mirror = caller_klass->java_mirror(); 4311 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4312 4313 #ifndef PRODUCT 4314 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4315 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()); 4316 tty->print_cr(" JVM state at this point:"); 4317 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4318 ciMethod* m = jvms()->of_depth(i)->method(); 4319 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4320 } 4321 } 4322 #endif 4323 return true; 4324 } 4325 break; 4326 } 4327 } 4328 4329 #ifndef PRODUCT 4330 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4331 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4332 tty->print_cr(" JVM state at this point:"); 4333 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4334 ciMethod* m = jvms()->of_depth(i)->method(); 4335 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4336 } 4337 } 4338 #endif 4339 4340 return false; // bail-out; let JVM_GetCallerClass do the work 4341 } 4342 4343 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4344 Node* arg = argument(0); 4345 Node* result = NULL; 4346 4347 switch (id) { 4348 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break; 4349 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break; 4350 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break; 4351 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break; 4352 4353 case vmIntrinsics::_doubleToLongBits: { 4354 // two paths (plus control) merge in a wood 4355 RegionNode *r = new (C) RegionNode(3); 4356 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 4357 4358 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg)); 4359 // Build the boolean node 4360 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4361 4362 // Branch either way. 4363 // NaN case is less traveled, which makes all the difference. 4364 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4365 Node *opt_isnan = _gvn.transform(ifisnan); 4366 assert( opt_isnan->is_If(), "Expect an IfNode"); 4367 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4368 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4369 4370 set_control(iftrue); 4371 4372 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4373 Node *slow_result = longcon(nan_bits); // return NaN 4374 phi->init_req(1, _gvn.transform( slow_result )); 4375 r->init_req(1, iftrue); 4376 4377 // Else fall through 4378 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4379 set_control(iffalse); 4380 4381 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg))); 4382 r->init_req(2, iffalse); 4383 4384 // Post merge 4385 set_control(_gvn.transform(r)); 4386 record_for_igvn(r); 4387 4388 C->set_has_split_ifs(true); // Has chance for split-if optimization 4389 result = phi; 4390 assert(result->bottom_type()->isa_long(), "must be"); 4391 break; 4392 } 4393 4394 case vmIntrinsics::_floatToIntBits: { 4395 // two paths (plus control) merge in a wood 4396 RegionNode *r = new (C) RegionNode(3); 4397 Node *phi = new (C) PhiNode(r, TypeInt::INT); 4398 4399 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg)); 4400 // Build the boolean node 4401 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4402 4403 // Branch either way. 4404 // NaN case is less traveled, which makes all the difference. 4405 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4406 Node *opt_isnan = _gvn.transform(ifisnan); 4407 assert( opt_isnan->is_If(), "Expect an IfNode"); 4408 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4409 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4410 4411 set_control(iftrue); 4412 4413 static const jint nan_bits = 0x7fc00000; 4414 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4415 phi->init_req(1, _gvn.transform( slow_result )); 4416 r->init_req(1, iftrue); 4417 4418 // Else fall through 4419 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4420 set_control(iffalse); 4421 4422 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg))); 4423 r->init_req(2, iffalse); 4424 4425 // Post merge 4426 set_control(_gvn.transform(r)); 4427 record_for_igvn(r); 4428 4429 C->set_has_split_ifs(true); // Has chance for split-if optimization 4430 result = phi; 4431 assert(result->bottom_type()->isa_int(), "must be"); 4432 break; 4433 } 4434 4435 default: 4436 fatal_unexpected_iid(id); 4437 break; 4438 } 4439 set_result(_gvn.transform(result)); 4440 return true; 4441 } 4442 4443 #ifdef _LP64 4444 #define XTOP ,top() /*additional argument*/ 4445 #else //_LP64 4446 #define XTOP /*no additional argument*/ 4447 #endif //_LP64 4448 4449 //----------------------inline_unsafe_copyMemory------------------------- 4450 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4451 bool LibraryCallKit::inline_unsafe_copyMemory() { 4452 if (callee()->is_static()) return false; // caller must have the capability! 4453 null_check_receiver(); // null-check receiver 4454 if (stopped()) return true; 4455 4456 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4457 4458 Node* src_ptr = argument(1); // type: oop 4459 Node* src_off = ConvL2X(argument(2)); // type: long 4460 Node* dst_ptr = argument(4); // type: oop 4461 Node* dst_off = ConvL2X(argument(5)); // type: long 4462 Node* size = ConvL2X(argument(7)); // type: long 4463 4464 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4465 "fieldOffset must be byte-scaled"); 4466 4467 Node* src = make_unsafe_address(src_ptr, src_off); 4468 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4469 4470 // Conservatively insert a memory barrier on all memory slices. 4471 // Do not let writes of the copy source or destination float below the copy. 4472 insert_mem_bar(Op_MemBarCPUOrder); 4473 4474 // Call it. Note that the length argument is not scaled. 4475 make_runtime_call(RC_LEAF|RC_NO_FP, 4476 OptoRuntime::fast_arraycopy_Type(), 4477 StubRoutines::unsafe_arraycopy(), 4478 "unsafe_arraycopy", 4479 TypeRawPtr::BOTTOM, 4480 src, dst, size XTOP); 4481 4482 // Do not let reads of the copy destination float above the copy. 4483 insert_mem_bar(Op_MemBarCPUOrder); 4484 4485 return true; 4486 } 4487 4488 //------------------------clone_coping----------------------------------- 4489 // Helper function for inline_native_clone. 4490 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4491 assert(obj_size != NULL, ""); 4492 Node* raw_obj = alloc_obj->in(1); 4493 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4494 4495 AllocateNode* alloc = NULL; 4496 if (ReduceBulkZeroing) { 4497 // We will be completely responsible for initializing this object - 4498 // mark Initialize node as complete. 4499 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4500 // The object was just allocated - there should be no any stores! 4501 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4502 // Mark as complete_with_arraycopy so that on AllocateNode 4503 // expansion, we know this AllocateNode is initialized by an array 4504 // copy and a StoreStore barrier exists after the array copy. 4505 alloc->initialization()->set_complete_with_arraycopy(); 4506 } 4507 4508 // Copy the fastest available way. 4509 // TODO: generate fields copies for small objects instead. 4510 Node* src = obj; 4511 Node* dest = alloc_obj; 4512 Node* size = _gvn.transform(obj_size); 4513 4514 // Exclude the header but include array length to copy by 8 bytes words. 4515 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4516 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4517 instanceOopDesc::base_offset_in_bytes(); 4518 // base_off: 4519 // 8 - 32-bit VM 4520 // 12 - 64-bit VM, compressed klass 4521 // 16 - 64-bit VM, normal klass 4522 if (base_off % BytesPerLong != 0) { 4523 assert(UseCompressedClassPointers, ""); 4524 if (is_array) { 4525 // Exclude length to copy by 8 bytes words. 4526 base_off += sizeof(int); 4527 } else { 4528 // Include klass to copy by 8 bytes words. 4529 base_off = instanceOopDesc::klass_offset_in_bytes(); 4530 } 4531 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4532 } 4533 src = basic_plus_adr(src, base_off); 4534 dest = basic_plus_adr(dest, base_off); 4535 4536 // Compute the length also, if needed: 4537 Node* countx = size; 4538 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off))); 4539 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4540 4541 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4542 bool disjoint_bases = true; 4543 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4544 src, NULL, dest, NULL, countx, 4545 /*dest_uninitialized*/true); 4546 4547 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4548 if (card_mark) { 4549 assert(!is_array, ""); 4550 // Put in store barrier for any and all oops we are sticking 4551 // into this object. (We could avoid this if we could prove 4552 // that the object type contains no oop fields at all.) 4553 Node* no_particular_value = NULL; 4554 Node* no_particular_field = NULL; 4555 int raw_adr_idx = Compile::AliasIdxRaw; 4556 post_barrier(control(), 4557 memory(raw_adr_type), 4558 alloc_obj, 4559 no_particular_field, 4560 raw_adr_idx, 4561 no_particular_value, 4562 T_OBJECT, 4563 false); 4564 } 4565 4566 // Do not let reads from the cloned object float above the arraycopy. 4567 if (alloc != NULL) { 4568 // Do not let stores that initialize this object be reordered with 4569 // a subsequent store that would make this object accessible by 4570 // other threads. 4571 // Record what AllocateNode this StoreStore protects so that 4572 // escape analysis can go from the MemBarStoreStoreNode to the 4573 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4574 // based on the escape status of the AllocateNode. 4575 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4576 } else { 4577 insert_mem_bar(Op_MemBarCPUOrder); 4578 } 4579 } 4580 4581 //------------------------inline_native_clone---------------------------- 4582 // protected native Object java.lang.Object.clone(); 4583 // 4584 // Here are the simple edge cases: 4585 // null receiver => normal trap 4586 // virtual and clone was overridden => slow path to out-of-line clone 4587 // not cloneable or finalizer => slow path to out-of-line Object.clone 4588 // 4589 // The general case has two steps, allocation and copying. 4590 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4591 // 4592 // Copying also has two cases, oop arrays and everything else. 4593 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4594 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4595 // 4596 // These steps fold up nicely if and when the cloned object's klass 4597 // can be sharply typed as an object array, a type array, or an instance. 4598 // 4599 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4600 PhiNode* result_val; 4601 4602 // Set the reexecute bit for the interpreter to reexecute 4603 // the bytecode that invokes Object.clone if deoptimization happens. 4604 { PreserveReexecuteState preexecs(this); 4605 jvms()->set_should_reexecute(true); 4606 4607 Node* obj = null_check_receiver(); 4608 if (stopped()) return true; 4609 4610 Node* obj_klass = load_object_klass(obj); 4611 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4612 const TypeOopPtr* toop = ((tklass != NULL) 4613 ? tklass->as_instance_type() 4614 : TypeInstPtr::NOTNULL); 4615 4616 // Conservatively insert a memory barrier on all memory slices. 4617 // Do not let writes into the original float below the clone. 4618 insert_mem_bar(Op_MemBarCPUOrder); 4619 4620 // paths into result_reg: 4621 enum { 4622 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4623 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4624 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4625 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4626 PATH_LIMIT 4627 }; 4628 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4629 result_val = new(C) PhiNode(result_reg, 4630 TypeInstPtr::NOTNULL); 4631 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO); 4632 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 4633 TypePtr::BOTTOM); 4634 record_for_igvn(result_reg); 4635 4636 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4637 int raw_adr_idx = Compile::AliasIdxRaw; 4638 4639 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4640 if (array_ctl != NULL) { 4641 // It's an array. 4642 PreserveJVMState pjvms(this); 4643 set_control(array_ctl); 4644 Node* obj_length = load_array_length(obj); 4645 Node* obj_size = NULL; 4646 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4647 4648 if (!use_ReduceInitialCardMarks()) { 4649 // If it is an oop array, it requires very special treatment, 4650 // because card marking is required on each card of the array. 4651 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4652 if (is_obja != NULL) { 4653 PreserveJVMState pjvms2(this); 4654 set_control(is_obja); 4655 // Generate a direct call to the right arraycopy function(s). 4656 bool disjoint_bases = true; 4657 bool length_never_negative = true; 4658 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4659 obj, intcon(0), alloc_obj, intcon(0), 4660 obj_length, 4661 disjoint_bases, length_never_negative); 4662 result_reg->init_req(_objArray_path, control()); 4663 result_val->init_req(_objArray_path, alloc_obj); 4664 result_i_o ->set_req(_objArray_path, i_o()); 4665 result_mem ->set_req(_objArray_path, reset_memory()); 4666 } 4667 } 4668 // Otherwise, there are no card marks to worry about. 4669 // (We can dispense with card marks if we know the allocation 4670 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4671 // causes the non-eden paths to take compensating steps to 4672 // simulate a fresh allocation, so that no further 4673 // card marks are required in compiled code to initialize 4674 // the object.) 4675 4676 if (!stopped()) { 4677 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4678 4679 // Present the results of the copy. 4680 result_reg->init_req(_array_path, control()); 4681 result_val->init_req(_array_path, alloc_obj); 4682 result_i_o ->set_req(_array_path, i_o()); 4683 result_mem ->set_req(_array_path, reset_memory()); 4684 } 4685 } 4686 4687 // We only go to the instance fast case code if we pass a number of guards. 4688 // The paths which do not pass are accumulated in the slow_region. 4689 RegionNode* slow_region = new (C) RegionNode(1); 4690 record_for_igvn(slow_region); 4691 if (!stopped()) { 4692 // It's an instance (we did array above). Make the slow-path tests. 4693 // If this is a virtual call, we generate a funny guard. We grab 4694 // the vtable entry corresponding to clone() from the target object. 4695 // If the target method which we are calling happens to be the 4696 // Object clone() method, we pass the guard. We do not need this 4697 // guard for non-virtual calls; the caller is known to be the native 4698 // Object clone(). 4699 if (is_virtual) { 4700 generate_virtual_guard(obj_klass, slow_region); 4701 } 4702 4703 // The object must be cloneable and must not have a finalizer. 4704 // Both of these conditions may be checked in a single test. 4705 // We could optimize the cloneable test further, but we don't care. 4706 generate_access_flags_guard(obj_klass, 4707 // Test both conditions: 4708 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4709 // Must be cloneable but not finalizer: 4710 JVM_ACC_IS_CLONEABLE, 4711 slow_region); 4712 } 4713 4714 if (!stopped()) { 4715 // It's an instance, and it passed the slow-path tests. 4716 PreserveJVMState pjvms(this); 4717 Node* obj_size = NULL; 4718 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4719 // is reexecuted if deoptimization occurs and there could be problems when merging 4720 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4721 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4722 4723 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4724 4725 // Present the results of the slow call. 4726 result_reg->init_req(_instance_path, control()); 4727 result_val->init_req(_instance_path, alloc_obj); 4728 result_i_o ->set_req(_instance_path, i_o()); 4729 result_mem ->set_req(_instance_path, reset_memory()); 4730 } 4731 4732 // Generate code for the slow case. We make a call to clone(). 4733 set_control(_gvn.transform(slow_region)); 4734 if (!stopped()) { 4735 PreserveJVMState pjvms(this); 4736 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4737 Node* slow_result = set_results_for_java_call(slow_call); 4738 // this->control() comes from set_results_for_java_call 4739 result_reg->init_req(_slow_path, control()); 4740 result_val->init_req(_slow_path, slow_result); 4741 result_i_o ->set_req(_slow_path, i_o()); 4742 result_mem ->set_req(_slow_path, reset_memory()); 4743 } 4744 4745 // Return the combined state. 4746 set_control( _gvn.transform(result_reg)); 4747 set_i_o( _gvn.transform(result_i_o)); 4748 set_all_memory( _gvn.transform(result_mem)); 4749 } // original reexecute is set back here 4750 4751 set_result(_gvn.transform(result_val)); 4752 return true; 4753 } 4754 4755 //------------------------------basictype2arraycopy---------------------------- 4756 address LibraryCallKit::basictype2arraycopy(BasicType t, 4757 Node* src_offset, 4758 Node* dest_offset, 4759 bool disjoint_bases, 4760 const char* &name, 4761 bool dest_uninitialized) { 4762 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4763 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4764 4765 bool aligned = false; 4766 bool disjoint = disjoint_bases; 4767 4768 // if the offsets are the same, we can treat the memory regions as 4769 // disjoint, because either the memory regions are in different arrays, 4770 // or they are identical (which we can treat as disjoint.) We can also 4771 // treat a copy with a destination index less that the source index 4772 // as disjoint since a low->high copy will work correctly in this case. 4773 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4774 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4775 // both indices are constants 4776 int s_offs = src_offset_inttype->get_con(); 4777 int d_offs = dest_offset_inttype->get_con(); 4778 int element_size = type2aelembytes(t); 4779 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4780 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4781 if (s_offs >= d_offs) disjoint = true; 4782 } else if (src_offset == dest_offset && src_offset != NULL) { 4783 // This can occur if the offsets are identical non-constants. 4784 disjoint = true; 4785 } 4786 4787 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4788 } 4789 4790 4791 //------------------------------inline_arraycopy----------------------- 4792 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4793 // Object dest, int destPos, 4794 // int length); 4795 bool LibraryCallKit::inline_arraycopy() { 4796 // Get the arguments. 4797 Node* src = argument(0); // type: oop 4798 Node* src_offset = argument(1); // type: int 4799 Node* dest = argument(2); // type: oop 4800 Node* dest_offset = argument(3); // type: int 4801 Node* length = argument(4); // type: int 4802 4803 // Compile time checks. If any of these checks cannot be verified at compile time, 4804 // we do not make a fast path for this call. Instead, we let the call remain as it 4805 // is. The checks we choose to mandate at compile time are: 4806 // 4807 // (1) src and dest are arrays. 4808 const Type* src_type = src->Value(&_gvn); 4809 const Type* dest_type = dest->Value(&_gvn); 4810 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4811 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4812 4813 // Do we have the type of src? 4814 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4815 // Do we have the type of dest? 4816 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4817 // Is the type for src from speculation? 4818 bool src_spec = false; 4819 // Is the type for dest from speculation? 4820 bool dest_spec = false; 4821 4822 if (!has_src || !has_dest) { 4823 // We don't have sufficient type information, let's see if 4824 // speculative types can help. We need to have types for both src 4825 // and dest so that it pays off. 4826 4827 // Do we already have or could we have type information for src 4828 bool could_have_src = has_src; 4829 // Do we already have or could we have type information for dest 4830 bool could_have_dest = has_dest; 4831 4832 ciKlass* src_k = NULL; 4833 if (!has_src) { 4834 src_k = src_type->speculative_type(); 4835 if (src_k != NULL && src_k->is_array_klass()) { 4836 could_have_src = true; 4837 } 4838 } 4839 4840 ciKlass* dest_k = NULL; 4841 if (!has_dest) { 4842 dest_k = dest_type->speculative_type(); 4843 if (dest_k != NULL && dest_k->is_array_klass()) { 4844 could_have_dest = true; 4845 } 4846 } 4847 4848 if (could_have_src && could_have_dest) { 4849 // This is going to pay off so emit the required guards 4850 if (!has_src) { 4851 src = maybe_cast_profiled_obj(src, src_k); 4852 src_type = _gvn.type(src); 4853 top_src = src_type->isa_aryptr(); 4854 has_src = (top_src != NULL && top_src->klass() != NULL); 4855 src_spec = true; 4856 } 4857 if (!has_dest) { 4858 dest = maybe_cast_profiled_obj(dest, dest_k); 4859 dest_type = _gvn.type(dest); 4860 top_dest = dest_type->isa_aryptr(); 4861 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4862 dest_spec = true; 4863 } 4864 } 4865 } 4866 4867 if (!has_src || !has_dest) { 4868 // Conservatively insert a memory barrier on all memory slices. 4869 // Do not let writes into the source float below the arraycopy. 4870 insert_mem_bar(Op_MemBarCPUOrder); 4871 4872 // Call StubRoutines::generic_arraycopy stub. 4873 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4874 src, src_offset, dest, dest_offset, length); 4875 4876 // Do not let reads from the destination float above the arraycopy. 4877 // Since we cannot type the arrays, we don't know which slices 4878 // might be affected. We could restrict this barrier only to those 4879 // memory slices which pertain to array elements--but don't bother. 4880 if (!InsertMemBarAfterArraycopy) 4881 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4882 insert_mem_bar(Op_MemBarCPUOrder); 4883 return true; 4884 } 4885 4886 // (2) src and dest arrays must have elements of the same BasicType 4887 // Figure out the size and type of the elements we will be copying. 4888 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4889 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4890 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4891 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4892 4893 if (src_elem != dest_elem || dest_elem == T_VOID) { 4894 // The component types are not the same or are not recognized. Punt. 4895 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4896 generate_slow_arraycopy(TypePtr::BOTTOM, 4897 src, src_offset, dest, dest_offset, length, 4898 /*dest_uninitialized*/false); 4899 return true; 4900 } 4901 4902 if (src_elem == T_OBJECT) { 4903 // If both arrays are object arrays then having the exact types 4904 // for both will remove the need for a subtype check at runtime 4905 // before the call and may make it possible to pick a faster copy 4906 // routine (without a subtype check on every element) 4907 // Do we have the exact type of src? 4908 bool could_have_src = src_spec; 4909 // Do we have the exact type of dest? 4910 bool could_have_dest = dest_spec; 4911 ciKlass* src_k = top_src->klass(); 4912 ciKlass* dest_k = top_dest->klass(); 4913 if (!src_spec) { 4914 src_k = src_type->speculative_type(); 4915 if (src_k != NULL && src_k->is_array_klass()) { 4916 could_have_src = true; 4917 } 4918 } 4919 if (!dest_spec) { 4920 dest_k = dest_type->speculative_type(); 4921 if (dest_k != NULL && dest_k->is_array_klass()) { 4922 could_have_dest = true; 4923 } 4924 } 4925 if (could_have_src && could_have_dest) { 4926 // If we can have both exact types, emit the missing guards 4927 if (could_have_src && !src_spec) { 4928 src = maybe_cast_profiled_obj(src, src_k); 4929 } 4930 if (could_have_dest && !dest_spec) { 4931 dest = maybe_cast_profiled_obj(dest, dest_k); 4932 } 4933 } 4934 } 4935 4936 //--------------------------------------------------------------------------- 4937 // We will make a fast path for this call to arraycopy. 4938 4939 // We have the following tests left to perform: 4940 // 4941 // (3) src and dest must not be null. 4942 // (4) src_offset must not be negative. 4943 // (5) dest_offset must not be negative. 4944 // (6) length must not be negative. 4945 // (7) src_offset + length must not exceed length of src. 4946 // (8) dest_offset + length must not exceed length of dest. 4947 // (9) each element of an oop array must be assignable 4948 4949 RegionNode* slow_region = new (C) RegionNode(1); 4950 record_for_igvn(slow_region); 4951 4952 // (3) operands must not be null 4953 // We currently perform our null checks with the null_check routine. 4954 // This means that the null exceptions will be reported in the caller 4955 // rather than (correctly) reported inside of the native arraycopy call. 4956 // This should be corrected, given time. We do our null check with the 4957 // stack pointer restored. 4958 src = null_check(src, T_ARRAY); 4959 dest = null_check(dest, T_ARRAY); 4960 4961 // (4) src_offset must not be negative. 4962 generate_negative_guard(src_offset, slow_region); 4963 4964 // (5) dest_offset must not be negative. 4965 generate_negative_guard(dest_offset, slow_region); 4966 4967 // (6) length must not be negative (moved to generate_arraycopy()). 4968 // generate_negative_guard(length, slow_region); 4969 4970 // (7) src_offset + length must not exceed length of src. 4971 generate_limit_guard(src_offset, length, 4972 load_array_length(src), 4973 slow_region); 4974 4975 // (8) dest_offset + length must not exceed length of dest. 4976 generate_limit_guard(dest_offset, length, 4977 load_array_length(dest), 4978 slow_region); 4979 4980 // (9) each element of an oop array must be assignable 4981 // The generate_arraycopy subroutine checks this. 4982 4983 // This is where the memory effects are placed: 4984 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4985 generate_arraycopy(adr_type, dest_elem, 4986 src, src_offset, dest, dest_offset, length, 4987 false, false, slow_region); 4988 4989 return true; 4990 } 4991 4992 //-----------------------------generate_arraycopy---------------------- 4993 // Generate an optimized call to arraycopy. 4994 // Caller must guard against non-arrays. 4995 // Caller must determine a common array basic-type for both arrays. 4996 // Caller must validate offsets against array bounds. 4997 // The slow_region has already collected guard failure paths 4998 // (such as out of bounds length or non-conformable array types). 4999 // The generated code has this shape, in general: 5000 // 5001 // if (length == 0) return // via zero_path 5002 // slowval = -1 5003 // if (types unknown) { 5004 // slowval = call generic copy loop 5005 // if (slowval == 0) return // via checked_path 5006 // } else if (indexes in bounds) { 5007 // if ((is object array) && !(array type check)) { 5008 // slowval = call checked copy loop 5009 // if (slowval == 0) return // via checked_path 5010 // } else { 5011 // call bulk copy loop 5012 // return // via fast_path 5013 // } 5014 // } 5015 // // adjust params for remaining work: 5016 // if (slowval != -1) { 5017 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 5018 // } 5019 // slow_region: 5020 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 5021 // return // via slow_call_path 5022 // 5023 // This routine is used from several intrinsics: System.arraycopy, 5024 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 5025 // 5026 void 5027 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 5028 BasicType basic_elem_type, 5029 Node* src, Node* src_offset, 5030 Node* dest, Node* dest_offset, 5031 Node* copy_length, 5032 bool disjoint_bases, 5033 bool length_never_negative, 5034 RegionNode* slow_region) { 5035 5036 if (slow_region == NULL) { 5037 slow_region = new(C) RegionNode(1); 5038 record_for_igvn(slow_region); 5039 } 5040 5041 Node* original_dest = dest; 5042 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 5043 bool dest_uninitialized = false; 5044 5045 // See if this is the initialization of a newly-allocated array. 5046 // If so, we will take responsibility here for initializing it to zero. 5047 // (Note: Because tightly_coupled_allocation performs checks on the 5048 // out-edges of the dest, we need to avoid making derived pointers 5049 // from it until we have checked its uses.) 5050 if (ReduceBulkZeroing 5051 && !ZeroTLAB // pointless if already zeroed 5052 && basic_elem_type != T_CONFLICT // avoid corner case 5053 && !src->eqv_uncast(dest) 5054 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 5055 != NULL) 5056 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 5057 && alloc->maybe_set_complete(&_gvn)) { 5058 // "You break it, you buy it." 5059 InitializeNode* init = alloc->initialization(); 5060 assert(init->is_complete(), "we just did this"); 5061 init->set_complete_with_arraycopy(); 5062 assert(dest->is_CheckCastPP(), "sanity"); 5063 assert(dest->in(0)->in(0) == init, "dest pinned"); 5064 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 5065 // From this point on, every exit path is responsible for 5066 // initializing any non-copied parts of the object to zero. 5067 // Also, if this flag is set we make sure that arraycopy interacts properly 5068 // with G1, eliding pre-barriers. See CR 6627983. 5069 dest_uninitialized = true; 5070 } else { 5071 // No zeroing elimination here. 5072 alloc = NULL; 5073 //original_dest = dest; 5074 //dest_uninitialized = false; 5075 } 5076 5077 // Results are placed here: 5078 enum { fast_path = 1, // normal void-returning assembly stub 5079 checked_path = 2, // special assembly stub with cleanup 5080 slow_call_path = 3, // something went wrong; call the VM 5081 zero_path = 4, // bypass when length of copy is zero 5082 bcopy_path = 5, // copy primitive array by 64-bit blocks 5083 PATH_LIMIT = 6 5084 }; 5085 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT); 5086 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO); 5087 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type); 5088 record_for_igvn(result_region); 5089 _gvn.set_type_bottom(result_i_o); 5090 _gvn.set_type_bottom(result_memory); 5091 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 5092 5093 // The slow_control path: 5094 Node* slow_control; 5095 Node* slow_i_o = i_o(); 5096 Node* slow_mem = memory(adr_type); 5097 debug_only(slow_control = (Node*) badAddress); 5098 5099 // Checked control path: 5100 Node* checked_control = top(); 5101 Node* checked_mem = NULL; 5102 Node* checked_i_o = NULL; 5103 Node* checked_value = NULL; 5104 5105 if (basic_elem_type == T_CONFLICT) { 5106 assert(!dest_uninitialized, ""); 5107 Node* cv = generate_generic_arraycopy(adr_type, 5108 src, src_offset, dest, dest_offset, 5109 copy_length, dest_uninitialized); 5110 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 5111 checked_control = control(); 5112 checked_i_o = i_o(); 5113 checked_mem = memory(adr_type); 5114 checked_value = cv; 5115 set_control(top()); // no fast path 5116 } 5117 5118 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 5119 if (not_pos != NULL) { 5120 PreserveJVMState pjvms(this); 5121 set_control(not_pos); 5122 5123 // (6) length must not be negative. 5124 if (!length_never_negative) { 5125 generate_negative_guard(copy_length, slow_region); 5126 } 5127 5128 // copy_length is 0. 5129 if (!stopped() && dest_uninitialized) { 5130 Node* dest_length = alloc->in(AllocateNode::ALength); 5131 if (copy_length->eqv_uncast(dest_length) 5132 || _gvn.find_int_con(dest_length, 1) <= 0) { 5133 // There is no zeroing to do. No need for a secondary raw memory barrier. 5134 } else { 5135 // Clear the whole thing since there are no source elements to copy. 5136 generate_clear_array(adr_type, dest, basic_elem_type, 5137 intcon(0), NULL, 5138 alloc->in(AllocateNode::AllocSize)); 5139 // Use a secondary InitializeNode as raw memory barrier. 5140 // Currently it is needed only on this path since other 5141 // paths have stub or runtime calls as raw memory barriers. 5142 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 5143 Compile::AliasIdxRaw, 5144 top())->as_Initialize(); 5145 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 5146 } 5147 } 5148 5149 // Present the results of the fast call. 5150 result_region->init_req(zero_path, control()); 5151 result_i_o ->init_req(zero_path, i_o()); 5152 result_memory->init_req(zero_path, memory(adr_type)); 5153 } 5154 5155 if (!stopped() && dest_uninitialized) { 5156 // We have to initialize the *uncopied* part of the array to zero. 5157 // The copy destination is the slice dest[off..off+len]. The other slices 5158 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 5159 Node* dest_size = alloc->in(AllocateNode::AllocSize); 5160 Node* dest_length = alloc->in(AllocateNode::ALength); 5161 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset, 5162 copy_length)); 5163 5164 // If there is a head section that needs zeroing, do it now. 5165 if (find_int_con(dest_offset, -1) != 0) { 5166 generate_clear_array(adr_type, dest, basic_elem_type, 5167 intcon(0), dest_offset, 5168 NULL); 5169 } 5170 5171 // Next, perform a dynamic check on the tail length. 5172 // It is often zero, and we can win big if we prove this. 5173 // There are two wins: Avoid generating the ClearArray 5174 // with its attendant messy index arithmetic, and upgrade 5175 // the copy to a more hardware-friendly word size of 64 bits. 5176 Node* tail_ctl = NULL; 5177 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 5178 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length)); 5179 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt)); 5180 tail_ctl = generate_slow_guard(bol_lt, NULL); 5181 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 5182 } 5183 5184 // At this point, let's assume there is no tail. 5185 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 5186 // There is no tail. Try an upgrade to a 64-bit copy. 5187 bool didit = false; 5188 { PreserveJVMState pjvms(this); 5189 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 5190 src, src_offset, dest, dest_offset, 5191 dest_size, dest_uninitialized); 5192 if (didit) { 5193 // Present the results of the block-copying fast call. 5194 result_region->init_req(bcopy_path, control()); 5195 result_i_o ->init_req(bcopy_path, i_o()); 5196 result_memory->init_req(bcopy_path, memory(adr_type)); 5197 } 5198 } 5199 if (didit) 5200 set_control(top()); // no regular fast path 5201 } 5202 5203 // Clear the tail, if any. 5204 if (tail_ctl != NULL) { 5205 Node* notail_ctl = stopped() ? NULL : control(); 5206 set_control(tail_ctl); 5207 if (notail_ctl == NULL) { 5208 generate_clear_array(adr_type, dest, basic_elem_type, 5209 dest_tail, NULL, 5210 dest_size); 5211 } else { 5212 // Make a local merge. 5213 Node* done_ctl = new(C) RegionNode(3); 5214 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type); 5215 done_ctl->init_req(1, notail_ctl); 5216 done_mem->init_req(1, memory(adr_type)); 5217 generate_clear_array(adr_type, dest, basic_elem_type, 5218 dest_tail, NULL, 5219 dest_size); 5220 done_ctl->init_req(2, control()); 5221 done_mem->init_req(2, memory(adr_type)); 5222 set_control( _gvn.transform(done_ctl)); 5223 set_memory( _gvn.transform(done_mem), adr_type ); 5224 } 5225 } 5226 } 5227 5228 BasicType copy_type = basic_elem_type; 5229 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 5230 if (!stopped() && copy_type == T_OBJECT) { 5231 // If src and dest have compatible element types, we can copy bits. 5232 // Types S[] and D[] are compatible if D is a supertype of S. 5233 // 5234 // If they are not, we will use checked_oop_disjoint_arraycopy, 5235 // which performs a fast optimistic per-oop check, and backs off 5236 // further to JVM_ArrayCopy on the first per-oop check that fails. 5237 // (Actually, we don't move raw bits only; the GC requires card marks.) 5238 5239 // Get the Klass* for both src and dest 5240 Node* src_klass = load_object_klass(src); 5241 Node* dest_klass = load_object_klass(dest); 5242 5243 // Generate the subtype check. 5244 // This might fold up statically, or then again it might not. 5245 // 5246 // Non-static example: Copying List<String>.elements to a new String[]. 5247 // The backing store for a List<String> is always an Object[], 5248 // but its elements are always type String, if the generic types 5249 // are correct at the source level. 5250 // 5251 // Test S[] against D[], not S against D, because (probably) 5252 // the secondary supertype cache is less busy for S[] than S. 5253 // This usually only matters when D is an interface. 5254 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5255 // Plug failing path into checked_oop_disjoint_arraycopy 5256 if (not_subtype_ctrl != top()) { 5257 PreserveJVMState pjvms(this); 5258 set_control(not_subtype_ctrl); 5259 // (At this point we can assume disjoint_bases, since types differ.) 5260 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 5261 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 5262 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM); 5263 Node* dest_elem_klass = _gvn.transform(n1); 5264 Node* cv = generate_checkcast_arraycopy(adr_type, 5265 dest_elem_klass, 5266 src, src_offset, dest, dest_offset, 5267 ConvI2X(copy_length), dest_uninitialized); 5268 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 5269 checked_control = control(); 5270 checked_i_o = i_o(); 5271 checked_mem = memory(adr_type); 5272 checked_value = cv; 5273 } 5274 // At this point we know we do not need type checks on oop stores. 5275 5276 // Let's see if we need card marks: 5277 if (alloc != NULL && use_ReduceInitialCardMarks()) { 5278 // If we do not need card marks, copy using the jint or jlong stub. 5279 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 5280 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 5281 "sizes agree"); 5282 } 5283 } 5284 5285 if (!stopped()) { 5286 // Generate the fast path, if possible. 5287 PreserveJVMState pjvms(this); 5288 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 5289 src, src_offset, dest, dest_offset, 5290 ConvI2X(copy_length), dest_uninitialized); 5291 5292 // Present the results of the fast call. 5293 result_region->init_req(fast_path, control()); 5294 result_i_o ->init_req(fast_path, i_o()); 5295 result_memory->init_req(fast_path, memory(adr_type)); 5296 } 5297 5298 // Here are all the slow paths up to this point, in one bundle: 5299 slow_control = top(); 5300 if (slow_region != NULL) 5301 slow_control = _gvn.transform(slow_region); 5302 DEBUG_ONLY(slow_region = (RegionNode*)badAddress); 5303 5304 set_control(checked_control); 5305 if (!stopped()) { 5306 // Clean up after the checked call. 5307 // The returned value is either 0 or -1^K, 5308 // where K = number of partially transferred array elements. 5309 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0))); 5310 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 5311 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 5312 5313 // If it is 0, we are done, so transfer to the end. 5314 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff)); 5315 result_region->init_req(checked_path, checks_done); 5316 result_i_o ->init_req(checked_path, checked_i_o); 5317 result_memory->init_req(checked_path, checked_mem); 5318 5319 // If it is not zero, merge into the slow call. 5320 set_control( _gvn.transform(new(C) IfFalseNode(iff) )); 5321 RegionNode* slow_reg2 = new(C) RegionNode(3); 5322 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO); 5323 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type); 5324 record_for_igvn(slow_reg2); 5325 slow_reg2 ->init_req(1, slow_control); 5326 slow_i_o2 ->init_req(1, slow_i_o); 5327 slow_mem2 ->init_req(1, slow_mem); 5328 slow_reg2 ->init_req(2, control()); 5329 slow_i_o2 ->init_req(2, checked_i_o); 5330 slow_mem2 ->init_req(2, checked_mem); 5331 5332 slow_control = _gvn.transform(slow_reg2); 5333 slow_i_o = _gvn.transform(slow_i_o2); 5334 slow_mem = _gvn.transform(slow_mem2); 5335 5336 if (alloc != NULL) { 5337 // We'll restart from the very beginning, after zeroing the whole thing. 5338 // This can cause double writes, but that's OK since dest is brand new. 5339 // So we ignore the low 31 bits of the value returned from the stub. 5340 } else { 5341 // We must continue the copy exactly where it failed, or else 5342 // another thread might see the wrong number of writes to dest. 5343 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1))); 5344 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT); 5345 slow_offset->init_req(1, intcon(0)); 5346 slow_offset->init_req(2, checked_offset); 5347 slow_offset = _gvn.transform(slow_offset); 5348 5349 // Adjust the arguments by the conditionally incoming offset. 5350 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset)); 5351 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset)); 5352 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset)); 5353 5354 // Tweak the node variables to adjust the code produced below: 5355 src_offset = src_off_plus; 5356 dest_offset = dest_off_plus; 5357 copy_length = length_minus; 5358 } 5359 } 5360 5361 set_control(slow_control); 5362 if (!stopped()) { 5363 // Generate the slow path, if needed. 5364 PreserveJVMState pjvms(this); // replace_in_map may trash the map 5365 5366 set_memory(slow_mem, adr_type); 5367 set_i_o(slow_i_o); 5368 5369 if (dest_uninitialized) { 5370 generate_clear_array(adr_type, dest, basic_elem_type, 5371 intcon(0), NULL, 5372 alloc->in(AllocateNode::AllocSize)); 5373 } 5374 5375 generate_slow_arraycopy(adr_type, 5376 src, src_offset, dest, dest_offset, 5377 copy_length, /*dest_uninitialized*/false); 5378 5379 result_region->init_req(slow_call_path, control()); 5380 result_i_o ->init_req(slow_call_path, i_o()); 5381 result_memory->init_req(slow_call_path, memory(adr_type)); 5382 } 5383 5384 // Remove unused edges. 5385 for (uint i = 1; i < result_region->req(); i++) { 5386 if (result_region->in(i) == NULL) 5387 result_region->init_req(i, top()); 5388 } 5389 5390 // Finished; return the combined state. 5391 set_control( _gvn.transform(result_region)); 5392 set_i_o( _gvn.transform(result_i_o) ); 5393 set_memory( _gvn.transform(result_memory), adr_type ); 5394 5395 // The memory edges above are precise in order to model effects around 5396 // array copies accurately to allow value numbering of field loads around 5397 // arraycopy. Such field loads, both before and after, are common in Java 5398 // collections and similar classes involving header/array data structures. 5399 // 5400 // But with low number of register or when some registers are used or killed 5401 // by arraycopy calls it causes registers spilling on stack. See 6544710. 5402 // The next memory barrier is added to avoid it. If the arraycopy can be 5403 // optimized away (which it can, sometimes) then we can manually remove 5404 // the membar also. 5405 // 5406 // Do not let reads from the cloned object float above the arraycopy. 5407 if (alloc != NULL) { 5408 // Do not let stores that initialize this object be reordered with 5409 // a subsequent store that would make this object accessible by 5410 // other threads. 5411 // Record what AllocateNode this StoreStore protects so that 5412 // escape analysis can go from the MemBarStoreStoreNode to the 5413 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5414 // based on the escape status of the AllocateNode. 5415 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5416 } else if (InsertMemBarAfterArraycopy) 5417 insert_mem_bar(Op_MemBarCPUOrder); 5418 } 5419 5420 5421 // Helper function which determines if an arraycopy immediately follows 5422 // an allocation, with no intervening tests or other escapes for the object. 5423 AllocateArrayNode* 5424 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5425 RegionNode* slow_region) { 5426 if (stopped()) return NULL; // no fast path 5427 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5428 5429 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5430 if (alloc == NULL) return NULL; 5431 5432 Node* rawmem = memory(Compile::AliasIdxRaw); 5433 // Is the allocation's memory state untouched? 5434 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5435 // Bail out if there have been raw-memory effects since the allocation. 5436 // (Example: There might have been a call or safepoint.) 5437 return NULL; 5438 } 5439 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5440 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5441 return NULL; 5442 } 5443 5444 // There must be no unexpected observers of this allocation. 5445 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5446 Node* obs = ptr->fast_out(i); 5447 if (obs != this->map()) { 5448 return NULL; 5449 } 5450 } 5451 5452 // This arraycopy must unconditionally follow the allocation of the ptr. 5453 Node* alloc_ctl = ptr->in(0); 5454 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5455 5456 Node* ctl = control(); 5457 while (ctl != alloc_ctl) { 5458 // There may be guards which feed into the slow_region. 5459 // Any other control flow means that we might not get a chance 5460 // to finish initializing the allocated object. 5461 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5462 IfNode* iff = ctl->in(0)->as_If(); 5463 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5464 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5465 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5466 ctl = iff->in(0); // This test feeds the known slow_region. 5467 continue; 5468 } 5469 // One more try: Various low-level checks bottom out in 5470 // uncommon traps. If the debug-info of the trap omits 5471 // any reference to the allocation, as we've already 5472 // observed, then there can be no objection to the trap. 5473 bool found_trap = false; 5474 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5475 Node* obs = not_ctl->fast_out(j); 5476 if (obs->in(0) == not_ctl && obs->is_Call() && 5477 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5478 found_trap = true; break; 5479 } 5480 } 5481 if (found_trap) { 5482 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5483 continue; 5484 } 5485 } 5486 return NULL; 5487 } 5488 5489 // If we get this far, we have an allocation which immediately 5490 // precedes the arraycopy, and we can take over zeroing the new object. 5491 // The arraycopy will finish the initialization, and provide 5492 // a new control state to which we will anchor the destination pointer. 5493 5494 return alloc; 5495 } 5496 5497 // Helper for initialization of arrays, creating a ClearArray. 5498 // It writes zero bits in [start..end), within the body of an array object. 5499 // The memory effects are all chained onto the 'adr_type' alias category. 5500 // 5501 // Since the object is otherwise uninitialized, we are free 5502 // to put a little "slop" around the edges of the cleared area, 5503 // as long as it does not go back into the array's header, 5504 // or beyond the array end within the heap. 5505 // 5506 // The lower edge can be rounded down to the nearest jint and the 5507 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5508 // 5509 // Arguments: 5510 // adr_type memory slice where writes are generated 5511 // dest oop of the destination array 5512 // basic_elem_type element type of the destination 5513 // slice_idx array index of first element to store 5514 // slice_len number of elements to store (or NULL) 5515 // dest_size total size in bytes of the array object 5516 // 5517 // Exactly one of slice_len or dest_size must be non-NULL. 5518 // If dest_size is non-NULL, zeroing extends to the end of the object. 5519 // If slice_len is non-NULL, the slice_idx value must be a constant. 5520 void 5521 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5522 Node* dest, 5523 BasicType basic_elem_type, 5524 Node* slice_idx, 5525 Node* slice_len, 5526 Node* dest_size) { 5527 // one or the other but not both of slice_len and dest_size: 5528 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5529 if (slice_len == NULL) slice_len = top(); 5530 if (dest_size == NULL) dest_size = top(); 5531 5532 // operate on this memory slice: 5533 Node* mem = memory(adr_type); // memory slice to operate on 5534 5535 // scaling and rounding of indexes: 5536 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5537 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5538 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5539 int bump_bit = (-1 << scale) & BytesPerInt; 5540 5541 // determine constant starts and ends 5542 const intptr_t BIG_NEG = -128; 5543 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5544 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5545 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5546 if (slice_len_con == 0) { 5547 return; // nothing to do here 5548 } 5549 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5550 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5551 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5552 assert(end_con < 0, "not two cons"); 5553 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5554 BytesPerLong); 5555 } 5556 5557 if (start_con >= 0 && end_con >= 0) { 5558 // Constant start and end. Simple. 5559 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5560 start_con, end_con, &_gvn); 5561 } else if (start_con >= 0 && dest_size != top()) { 5562 // Constant start, pre-rounded end after the tail of the array. 5563 Node* end = dest_size; 5564 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5565 start_con, end, &_gvn); 5566 } else if (start_con >= 0 && slice_len != top()) { 5567 // Constant start, non-constant end. End needs rounding up. 5568 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5569 intptr_t end_base = abase + (slice_idx_con << scale); 5570 int end_round = (-1 << scale) & (BytesPerLong - 1); 5571 Node* end = ConvI2X(slice_len); 5572 if (scale != 0) 5573 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) )); 5574 end_base += end_round; 5575 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base))); 5576 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round))); 5577 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5578 start_con, end, &_gvn); 5579 } else if (start_con < 0 && dest_size != top()) { 5580 // Non-constant start, pre-rounded end after the tail of the array. 5581 // This is almost certainly a "round-to-end" operation. 5582 Node* start = slice_idx; 5583 start = ConvI2X(start); 5584 if (scale != 0) 5585 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) )); 5586 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase))); 5587 if ((bump_bit | clear_low) != 0) { 5588 int to_clear = (bump_bit | clear_low); 5589 // Align up mod 8, then store a jint zero unconditionally 5590 // just before the mod-8 boundary. 5591 if (((abase + bump_bit) & ~to_clear) - bump_bit 5592 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5593 bump_bit = 0; 5594 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5595 } else { 5596 // Bump 'start' up to (or past) the next jint boundary: 5597 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit))); 5598 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5599 } 5600 // Round bumped 'start' down to jlong boundary in body of array. 5601 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear))); 5602 if (bump_bit != 0) { 5603 // Store a zero to the immediately preceding jint: 5604 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit))); 5605 Node* p1 = basic_plus_adr(dest, x1); 5606 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered); 5607 mem = _gvn.transform(mem); 5608 } 5609 } 5610 Node* end = dest_size; // pre-rounded 5611 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5612 start, end, &_gvn); 5613 } else { 5614 // Non-constant start, unrounded non-constant end. 5615 // (Nobody zeroes a random midsection of an array using this routine.) 5616 ShouldNotReachHere(); // fix caller 5617 } 5618 5619 // Done. 5620 set_memory(mem, adr_type); 5621 } 5622 5623 5624 bool 5625 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5626 BasicType basic_elem_type, 5627 AllocateNode* alloc, 5628 Node* src, Node* src_offset, 5629 Node* dest, Node* dest_offset, 5630 Node* dest_size, bool dest_uninitialized) { 5631 // See if there is an advantage from block transfer. 5632 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5633 if (scale >= LogBytesPerLong) 5634 return false; // it is already a block transfer 5635 5636 // Look at the alignment of the starting offsets. 5637 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5638 5639 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5640 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5641 if (src_off_con < 0 || dest_off_con < 0) 5642 // At present, we can only understand constants. 5643 return false; 5644 5645 intptr_t src_off = abase + (src_off_con << scale); 5646 intptr_t dest_off = abase + (dest_off_con << scale); 5647 5648 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5649 // Non-aligned; too bad. 5650 // One more chance: Pick off an initial 32-bit word. 5651 // This is a common case, since abase can be odd mod 8. 5652 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5653 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5654 Node* sptr = basic_plus_adr(src, src_off); 5655 Node* dptr = basic_plus_adr(dest, dest_off); 5656 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 5657 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered); 5658 src_off += BytesPerInt; 5659 dest_off += BytesPerInt; 5660 } else { 5661 return false; 5662 } 5663 } 5664 assert(src_off % BytesPerLong == 0, ""); 5665 assert(dest_off % BytesPerLong == 0, ""); 5666 5667 // Do this copy by giant steps. 5668 Node* sptr = basic_plus_adr(src, src_off); 5669 Node* dptr = basic_plus_adr(dest, dest_off); 5670 Node* countx = dest_size; 5671 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off))); 5672 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong))); 5673 5674 bool disjoint_bases = true; // since alloc != NULL 5675 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5676 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5677 5678 return true; 5679 } 5680 5681 5682 // Helper function; generates code for the slow case. 5683 // We make a call to a runtime method which emulates the native method, 5684 // but without the native wrapper overhead. 5685 void 5686 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5687 Node* src, Node* src_offset, 5688 Node* dest, Node* dest_offset, 5689 Node* copy_length, bool dest_uninitialized) { 5690 assert(!dest_uninitialized, "Invariant"); 5691 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5692 OptoRuntime::slow_arraycopy_Type(), 5693 OptoRuntime::slow_arraycopy_Java(), 5694 "slow_arraycopy", adr_type, 5695 src, src_offset, dest, dest_offset, 5696 copy_length); 5697 5698 // Handle exceptions thrown by this fellow: 5699 make_slow_call_ex(call, env()->Throwable_klass(), false); 5700 } 5701 5702 // Helper function; generates code for cases requiring runtime checks. 5703 Node* 5704 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5705 Node* dest_elem_klass, 5706 Node* src, Node* src_offset, 5707 Node* dest, Node* dest_offset, 5708 Node* copy_length, bool dest_uninitialized) { 5709 if (stopped()) return NULL; 5710 5711 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5712 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5713 return NULL; 5714 } 5715 5716 // Pick out the parameters required to perform a store-check 5717 // for the target array. This is an optimistic check. It will 5718 // look in each non-null element's class, at the desired klass's 5719 // super_check_offset, for the desired klass. 5720 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5721 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5722 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered); 5723 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5724 Node* check_value = dest_elem_klass; 5725 5726 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5727 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5728 5729 // (We know the arrays are never conjoint, because their types differ.) 5730 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5731 OptoRuntime::checkcast_arraycopy_Type(), 5732 copyfunc_addr, "checkcast_arraycopy", adr_type, 5733 // five arguments, of which two are 5734 // intptr_t (jlong in LP64) 5735 src_start, dest_start, 5736 copy_length XTOP, 5737 check_offset XTOP, 5738 check_value); 5739 5740 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5741 } 5742 5743 5744 // Helper function; generates code for cases requiring runtime checks. 5745 Node* 5746 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5747 Node* src, Node* src_offset, 5748 Node* dest, Node* dest_offset, 5749 Node* copy_length, bool dest_uninitialized) { 5750 assert(!dest_uninitialized, "Invariant"); 5751 if (stopped()) return NULL; 5752 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5753 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5754 return NULL; 5755 } 5756 5757 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5758 OptoRuntime::generic_arraycopy_Type(), 5759 copyfunc_addr, "generic_arraycopy", adr_type, 5760 src, src_offset, dest, dest_offset, copy_length); 5761 5762 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5763 } 5764 5765 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5766 void 5767 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5768 BasicType basic_elem_type, 5769 bool disjoint_bases, 5770 Node* src, Node* src_offset, 5771 Node* dest, Node* dest_offset, 5772 Node* copy_length, bool dest_uninitialized) { 5773 if (stopped()) return; // nothing to do 5774 5775 Node* src_start = src; 5776 Node* dest_start = dest; 5777 if (src_offset != NULL || dest_offset != NULL) { 5778 assert(src_offset != NULL && dest_offset != NULL, ""); 5779 src_start = array_element_address(src, src_offset, basic_elem_type); 5780 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5781 } 5782 5783 // Figure out which arraycopy runtime method to call. 5784 const char* copyfunc_name = "arraycopy"; 5785 address copyfunc_addr = 5786 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5787 disjoint_bases, copyfunc_name, dest_uninitialized); 5788 5789 // Call it. Note that the count_ix value is not scaled to a byte-size. 5790 make_runtime_call(RC_LEAF|RC_NO_FP, 5791 OptoRuntime::fast_arraycopy_Type(), 5792 copyfunc_addr, copyfunc_name, adr_type, 5793 src_start, dest_start, copy_length XTOP); 5794 } 5795 5796 //-------------inline_encodeISOArray----------------------------------- 5797 // encode char[] to byte[] in ISO_8859_1 5798 bool LibraryCallKit::inline_encodeISOArray() { 5799 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5800 // no receiver since it is static method 5801 Node *src = argument(0); 5802 Node *src_offset = argument(1); 5803 Node *dst = argument(2); 5804 Node *dst_offset = argument(3); 5805 Node *length = argument(4); 5806 5807 const Type* src_type = src->Value(&_gvn); 5808 const Type* dst_type = dst->Value(&_gvn); 5809 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5810 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5811 if (top_src == NULL || top_src->klass() == NULL || 5812 top_dest == NULL || top_dest->klass() == NULL) { 5813 // failed array check 5814 return false; 5815 } 5816 5817 // Figure out the size and type of the elements we will be copying. 5818 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5819 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5820 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5821 return false; 5822 } 5823 Node* src_start = array_element_address(src, src_offset, src_elem); 5824 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5825 // 'src_start' points to src array + scaled offset 5826 // 'dst_start' points to dst array + scaled offset 5827 5828 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5829 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5830 enc = _gvn.transform(enc); 5831 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc)); 5832 set_memory(res_mem, mtype); 5833 set_result(enc); 5834 return true; 5835 } 5836 5837 //-------------inline_multiplyToLen----------------------------------- 5838 bool LibraryCallKit::inline_multiplyToLen() { 5839 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5840 5841 address stubAddr = StubRoutines::multiplyToLen(); 5842 if (stubAddr == NULL) { 5843 return false; // Intrinsic's stub is not implemented on this platform 5844 } 5845 const char* stubName = "multiplyToLen"; 5846 5847 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5848 5849 // no receiver because it is a static method 5850 Node* x = argument(0); 5851 Node* xlen = argument(1); 5852 Node* y = argument(2); 5853 Node* ylen = argument(3); 5854 Node* z = argument(4); 5855 5856 const Type* x_type = x->Value(&_gvn); 5857 const Type* y_type = y->Value(&_gvn); 5858 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5859 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5860 if (top_x == NULL || top_x->klass() == NULL || 5861 top_y == NULL || top_y->klass() == NULL) { 5862 // failed array check 5863 return false; 5864 } 5865 5866 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5867 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5868 if (x_elem != T_INT || y_elem != T_INT) { 5869 return false; 5870 } 5871 5872 // Set the original stack and the reexecute bit for the interpreter to reexecute 5873 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5874 // on the return from z array allocation in runtime. 5875 { PreserveReexecuteState preexecs(this); 5876 jvms()->set_should_reexecute(true); 5877 5878 Node* x_start = array_element_address(x, intcon(0), x_elem); 5879 Node* y_start = array_element_address(y, intcon(0), y_elem); 5880 // 'x_start' points to x array + scaled xlen 5881 // 'y_start' points to y array + scaled ylen 5882 5883 // Allocate the result array 5884 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen)); 5885 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5886 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5887 5888 IdealKit ideal(this); 5889 5890 #define __ ideal. 5891 Node* one = __ ConI(1); 5892 Node* zero = __ ConI(0); 5893 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5894 __ set(need_alloc, zero); 5895 __ set(z_alloc, z); 5896 __ if_then(z, BoolTest::eq, null()); { 5897 __ increment (need_alloc, one); 5898 } __ else_(); { 5899 // Update graphKit memory and control from IdealKit. 5900 sync_kit(ideal); 5901 Node* zlen_arg = load_array_length(z); 5902 // Update IdealKit memory and control from graphKit. 5903 __ sync_kit(this); 5904 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5905 __ increment (need_alloc, one); 5906 } __ end_if(); 5907 } __ end_if(); 5908 5909 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5910 // Update graphKit memory and control from IdealKit. 5911 sync_kit(ideal); 5912 Node * narr = new_array(klass_node, zlen, 1); 5913 // Update IdealKit memory and control from graphKit. 5914 __ sync_kit(this); 5915 __ set(z_alloc, narr); 5916 } __ end_if(); 5917 5918 sync_kit(ideal); 5919 z = __ value(z_alloc); 5920 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5921 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5922 // Final sync IdealKit and GraphKit. 5923 final_sync(ideal); 5924 #undef __ 5925 5926 Node* z_start = array_element_address(z, intcon(0), T_INT); 5927 5928 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5929 OptoRuntime::multiplyToLen_Type(), 5930 stubAddr, stubName, TypePtr::BOTTOM, 5931 x_start, xlen, y_start, ylen, z_start, zlen); 5932 } // original reexecute is set back here 5933 5934 C->set_has_split_ifs(true); // Has chance for split-if optimization 5935 set_result(z); 5936 return true; 5937 } 5938 5939 //-------------inline_squareToLen------------------------------------ 5940 bool LibraryCallKit::inline_squareToLen() { 5941 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5942 5943 address stubAddr = StubRoutines::squareToLen(); 5944 if (stubAddr == NULL) { 5945 return false; // Intrinsic's stub is not implemented on this platform 5946 } 5947 const char* stubName = "squareToLen"; 5948 5949 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5950 5951 Node* x = argument(0); 5952 Node* len = argument(1); 5953 Node* z = argument(2); 5954 Node* zlen = argument(3); 5955 5956 const Type* x_type = x->Value(&_gvn); 5957 const Type* z_type = z->Value(&_gvn); 5958 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5959 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5960 if (top_x == NULL || top_x->klass() == NULL || 5961 top_z == NULL || top_z->klass() == NULL) { 5962 // failed array check 5963 return false; 5964 } 5965 5966 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5967 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5968 if (x_elem != T_INT || z_elem != T_INT) { 5969 return false; 5970 } 5971 5972 5973 Node* x_start = array_element_address(x, intcon(0), x_elem); 5974 Node* z_start = array_element_address(z, intcon(0), z_elem); 5975 5976 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5977 OptoRuntime::squareToLen_Type(), 5978 stubAddr, stubName, TypePtr::BOTTOM, 5979 x_start, len, z_start, zlen); 5980 5981 set_result(z); 5982 return true; 5983 } 5984 5985 //-------------inline_mulAdd------------------------------------------ 5986 bool LibraryCallKit::inline_mulAdd() { 5987 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5988 5989 address stubAddr = StubRoutines::mulAdd(); 5990 if (stubAddr == NULL) { 5991 return false; // Intrinsic's stub is not implemented on this platform 5992 } 5993 const char* stubName = "mulAdd"; 5994 5995 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5996 5997 Node* out = argument(0); 5998 Node* in = argument(1); 5999 Node* offset = argument(2); 6000 Node* len = argument(3); 6001 Node* k = argument(4); 6002 6003 const Type* out_type = out->Value(&_gvn); 6004 const Type* in_type = in->Value(&_gvn); 6005 const TypeAryPtr* top_out = out_type->isa_aryptr(); 6006 const TypeAryPtr* top_in = in_type->isa_aryptr(); 6007 if (top_out == NULL || top_out->klass() == NULL || 6008 top_in == NULL || top_in->klass() == NULL) { 6009 // failed array check 6010 return false; 6011 } 6012 6013 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6014 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6015 if (out_elem != T_INT || in_elem != T_INT) { 6016 return false; 6017 } 6018 6019 Node* outlen = load_array_length(out); 6020 Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset)); 6021 Node* out_start = array_element_address(out, intcon(0), out_elem); 6022 Node* in_start = array_element_address(in, intcon(0), in_elem); 6023 6024 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6025 OptoRuntime::mulAdd_Type(), 6026 stubAddr, stubName, TypePtr::BOTTOM, 6027 out_start,in_start, new_offset, len, k); 6028 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 6029 set_result(result); 6030 return true; 6031 } 6032 6033 //-------------inline_montgomeryMultiply----------------------------------- 6034 bool LibraryCallKit::inline_montgomeryMultiply() { 6035 address stubAddr = StubRoutines::montgomeryMultiply(); 6036 if (stubAddr == NULL) { 6037 return false; // Intrinsic's stub is not implemented on this platform 6038 } 6039 6040 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 6041 const char* stubName = "montgomery_multiply"; 6042 6043 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 6044 6045 Node* a = argument(0); 6046 Node* b = argument(1); 6047 Node* n = argument(2); 6048 Node* len = argument(3); 6049 Node* inv = argument(4); 6050 Node* m = argument(6); 6051 6052 const Type* a_type = a->Value(&_gvn); 6053 const TypeAryPtr* top_a = a_type->isa_aryptr(); 6054 const Type* b_type = b->Value(&_gvn); 6055 const TypeAryPtr* top_b = b_type->isa_aryptr(); 6056 const Type* n_type = a->Value(&_gvn); 6057 const TypeAryPtr* top_n = n_type->isa_aryptr(); 6058 const Type* m_type = a->Value(&_gvn); 6059 const TypeAryPtr* top_m = m_type->isa_aryptr(); 6060 if (top_a == NULL || top_a->klass() == NULL || 6061 top_b == NULL || top_b->klass() == NULL || 6062 top_n == NULL || top_n->klass() == NULL || 6063 top_m == NULL || top_m->klass() == NULL) { 6064 // failed array check 6065 return false; 6066 } 6067 6068 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6069 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6070 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6071 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6072 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 6073 return false; 6074 } 6075 6076 // Make the call 6077 { 6078 Node* a_start = array_element_address(a, intcon(0), a_elem); 6079 Node* b_start = array_element_address(b, intcon(0), b_elem); 6080 Node* n_start = array_element_address(n, intcon(0), n_elem); 6081 Node* m_start = array_element_address(m, intcon(0), m_elem); 6082 6083 Node* call = NULL; 6084 if (CCallingConventionRequiresIntsAsLongs) { 6085 Node* len_I2L = ConvI2L(len); 6086 call = make_runtime_call(RC_LEAF, 6087 OptoRuntime::montgomeryMultiply_Type(), 6088 stubAddr, stubName, TypePtr::BOTTOM, 6089 a_start, b_start, n_start, len_I2L XTOP, inv, 6090 top(), m_start); 6091 } else { 6092 call = make_runtime_call(RC_LEAF, 6093 OptoRuntime::montgomeryMultiply_Type(), 6094 stubAddr, stubName, TypePtr::BOTTOM, 6095 a_start, b_start, n_start, len, inv, top(), 6096 m_start); 6097 } 6098 set_result(m); 6099 } 6100 6101 return true; 6102 } 6103 6104 bool LibraryCallKit::inline_montgomerySquare() { 6105 address stubAddr = StubRoutines::montgomerySquare(); 6106 if (stubAddr == NULL) { 6107 return false; // Intrinsic's stub is not implemented on this platform 6108 } 6109 6110 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 6111 const char* stubName = "montgomery_square"; 6112 6113 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 6114 6115 Node* a = argument(0); 6116 Node* n = argument(1); 6117 Node* len = argument(2); 6118 Node* inv = argument(3); 6119 Node* m = argument(5); 6120 6121 const Type* a_type = a->Value(&_gvn); 6122 const TypeAryPtr* top_a = a_type->isa_aryptr(); 6123 const Type* n_type = a->Value(&_gvn); 6124 const TypeAryPtr* top_n = n_type->isa_aryptr(); 6125 const Type* m_type = a->Value(&_gvn); 6126 const TypeAryPtr* top_m = m_type->isa_aryptr(); 6127 if (top_a == NULL || top_a->klass() == NULL || 6128 top_n == NULL || top_n->klass() == NULL || 6129 top_m == NULL || top_m->klass() == NULL) { 6130 // failed array check 6131 return false; 6132 } 6133 6134 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6135 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6136 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6137 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 6138 return false; 6139 } 6140 6141 // Make the call 6142 { 6143 Node* a_start = array_element_address(a, intcon(0), a_elem); 6144 Node* n_start = array_element_address(n, intcon(0), n_elem); 6145 Node* m_start = array_element_address(m, intcon(0), m_elem); 6146 6147 Node* call = NULL; 6148 if (CCallingConventionRequiresIntsAsLongs) { 6149 Node* len_I2L = ConvI2L(len); 6150 call = make_runtime_call(RC_LEAF, 6151 OptoRuntime::montgomerySquare_Type(), 6152 stubAddr, stubName, TypePtr::BOTTOM, 6153 a_start, n_start, len_I2L XTOP, inv, top(), 6154 m_start); 6155 } else { 6156 call = make_runtime_call(RC_LEAF, 6157 OptoRuntime::montgomerySquare_Type(), 6158 stubAddr, stubName, TypePtr::BOTTOM, 6159 a_start, n_start, len, inv, top(), 6160 m_start); 6161 } 6162 6163 set_result(m); 6164 } 6165 6166 return true; 6167 } 6168 6169 6170 /** 6171 * Calculate CRC32 for byte. 6172 * int java.util.zip.CRC32.update(int crc, int b) 6173 */ 6174 bool LibraryCallKit::inline_updateCRC32() { 6175 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 6176 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 6177 // no receiver since it is static method 6178 Node* crc = argument(0); // type: int 6179 Node* b = argument(1); // type: int 6180 6181 /* 6182 * int c = ~ crc; 6183 * b = timesXtoThe32[(b ^ c) & 0xFF]; 6184 * b = b ^ (c >>> 8); 6185 * crc = ~b; 6186 */ 6187 6188 Node* M1 = intcon(-1); 6189 crc = _gvn.transform(new (C) XorINode(crc, M1)); 6190 Node* result = _gvn.transform(new (C) XorINode(crc, b)); 6191 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF))); 6192 6193 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 6194 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2))); 6195 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 6196 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 6197 6198 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8))); 6199 result = _gvn.transform(new (C) XorINode(crc, result)); 6200 result = _gvn.transform(new (C) XorINode(result, M1)); 6201 set_result(result); 6202 return true; 6203 } 6204 6205 /** 6206 * Calculate CRC32 for byte[] array. 6207 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 6208 */ 6209 bool LibraryCallKit::inline_updateBytesCRC32() { 6210 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 6211 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 6212 // no receiver since it is static method 6213 Node* crc = argument(0); // type: int 6214 Node* src = argument(1); // type: oop 6215 Node* offset = argument(2); // type: int 6216 Node* length = argument(3); // type: int 6217 6218 const Type* src_type = src->Value(&_gvn); 6219 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6220 if (top_src == NULL || top_src->klass() == NULL) { 6221 // failed array check 6222 return false; 6223 } 6224 6225 // Figure out the size and type of the elements we will be copying. 6226 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6227 if (src_elem != T_BYTE) { 6228 return false; 6229 } 6230 6231 // 'src_start' points to src array + scaled offset 6232 Node* src_start = array_element_address(src, offset, src_elem); 6233 6234 // We assume that range check is done by caller. 6235 // TODO: generate range check (offset+length < src.length) in debug VM. 6236 6237 // Call the stub. 6238 address stubAddr = StubRoutines::updateBytesCRC32(); 6239 const char *stubName = "updateBytesCRC32"; 6240 Node* call; 6241 if (CCallingConventionRequiresIntsAsLongs) { 6242 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6243 stubAddr, stubName, TypePtr::BOTTOM, 6244 crc XTOP, src_start, length XTOP); 6245 } else { 6246 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6247 stubAddr, stubName, TypePtr::BOTTOM, 6248 crc, src_start, length); 6249 } 6250 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 6251 set_result(result); 6252 return true; 6253 } 6254 6255 /** 6256 * Calculate CRC32 for ByteBuffer. 6257 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 6258 */ 6259 bool LibraryCallKit::inline_updateByteBufferCRC32() { 6260 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 6261 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 6262 // no receiver since it is static method 6263 Node* crc = argument(0); // type: int 6264 Node* src = argument(1); // type: long 6265 Node* offset = argument(3); // type: int 6266 Node* length = argument(4); // type: int 6267 6268 src = ConvL2X(src); // adjust Java long to machine word 6269 Node* base = _gvn.transform(new (C) CastX2PNode(src)); 6270 offset = ConvI2X(offset); 6271 6272 // 'src_start' points to src array + scaled offset 6273 Node* src_start = basic_plus_adr(top(), base, offset); 6274 6275 // Call the stub. 6276 address stubAddr = StubRoutines::updateBytesCRC32(); 6277 const char *stubName = "updateBytesCRC32"; 6278 Node* call; 6279 if (CCallingConventionRequiresIntsAsLongs) { 6280 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6281 stubAddr, stubName, TypePtr::BOTTOM, 6282 crc XTOP, src_start, length XTOP); 6283 } else { 6284 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6285 stubAddr, stubName, TypePtr::BOTTOM, 6286 crc, src_start, length); 6287 } 6288 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 6289 set_result(result); 6290 return true; 6291 } 6292 6293 //----------------------------inline_reference_get---------------------------- 6294 // public T java.lang.ref.Reference.get(); 6295 bool LibraryCallKit::inline_reference_get() { 6296 const int referent_offset = java_lang_ref_Reference::referent_offset; 6297 guarantee(referent_offset > 0, "should have already been set"); 6298 6299 // Get the argument: 6300 Node* reference_obj = null_check_receiver(); 6301 if (stopped()) return true; 6302 6303 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 6304 6305 ciInstanceKlass* klass = env()->Object_klass(); 6306 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 6307 6308 Node* no_ctrl = NULL; 6309 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 6310 6311 // Use the pre-barrier to record the value in the referent field 6312 pre_barrier(false /* do_load */, 6313 control(), 6314 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 6315 result /* pre_val */, 6316 T_OBJECT); 6317 6318 // Add memory barrier to prevent commoning reads from this field 6319 // across safepoint since GC can change its value. 6320 insert_mem_bar(Op_MemBarCPUOrder); 6321 6322 set_result(result); 6323 return true; 6324 } 6325 6326 6327 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 6328 bool is_exact=true, bool is_static=false) { 6329 6330 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 6331 assert(tinst != NULL, "obj is null"); 6332 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6333 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 6334 6335 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 6336 ciSymbol::make(fieldTypeString), 6337 is_static); 6338 if (field == NULL) return (Node *) NULL; 6339 assert (field != NULL, "undefined field"); 6340 6341 // Next code copied from Parse::do_get_xxx(): 6342 6343 // Compute address and memory type. 6344 int offset = field->offset_in_bytes(); 6345 bool is_vol = field->is_volatile(); 6346 ciType* field_klass = field->type(); 6347 assert(field_klass->is_loaded(), "should be loaded"); 6348 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 6349 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 6350 BasicType bt = field->layout_type(); 6351 6352 // Build the resultant type of the load 6353 const Type *type; 6354 if (bt == T_OBJECT) { 6355 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 6356 } else { 6357 type = Type::get_const_basic_type(bt); 6358 } 6359 6360 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 6361 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 6362 } 6363 // Build the load. 6364 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 6365 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 6366 // If reference is volatile, prevent following memory ops from 6367 // floating up past the volatile read. Also prevents commoning 6368 // another volatile read. 6369 if (is_vol) { 6370 // Memory barrier includes bogus read of value to force load BEFORE membar 6371 insert_mem_bar(Op_MemBarAcquire, loadedField); 6372 } 6373 return loadedField; 6374 } 6375 6376 6377 //------------------------------inline_aescrypt_Block----------------------- 6378 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 6379 address stubAddr = NULL; 6380 const char *stubName; 6381 assert(UseAES, "need AES instruction support"); 6382 6383 switch(id) { 6384 case vmIntrinsics::_aescrypt_encryptBlock: 6385 stubAddr = StubRoutines::aescrypt_encryptBlock(); 6386 stubName = "aescrypt_encryptBlock"; 6387 break; 6388 case vmIntrinsics::_aescrypt_decryptBlock: 6389 stubAddr = StubRoutines::aescrypt_decryptBlock(); 6390 stubName = "aescrypt_decryptBlock"; 6391 break; 6392 } 6393 if (stubAddr == NULL) return false; 6394 6395 Node* aescrypt_object = argument(0); 6396 Node* src = argument(1); 6397 Node* src_offset = argument(2); 6398 Node* dest = argument(3); 6399 Node* dest_offset = argument(4); 6400 6401 // (1) src and dest are arrays. 6402 const Type* src_type = src->Value(&_gvn); 6403 const Type* dest_type = dest->Value(&_gvn); 6404 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6405 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6406 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6407 6408 // for the quick and dirty code we will skip all the checks. 6409 // we are just trying to get the call to be generated. 6410 Node* src_start = src; 6411 Node* dest_start = dest; 6412 if (src_offset != NULL || dest_offset != NULL) { 6413 assert(src_offset != NULL && dest_offset != NULL, ""); 6414 src_start = array_element_address(src, src_offset, T_BYTE); 6415 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6416 } 6417 6418 // now need to get the start of its expanded key array 6419 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6420 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6421 if (k_start == NULL) return false; 6422 6423 if (Matcher::pass_original_key_for_aes()) { 6424 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6425 // compatibility issues between Java key expansion and SPARC crypto instructions 6426 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6427 if (original_k_start == NULL) return false; 6428 6429 // Call the stub. 6430 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6431 stubAddr, stubName, TypePtr::BOTTOM, 6432 src_start, dest_start, k_start, original_k_start); 6433 } else { 6434 // Call the stub. 6435 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6436 stubAddr, stubName, TypePtr::BOTTOM, 6437 src_start, dest_start, k_start); 6438 } 6439 6440 return true; 6441 } 6442 6443 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 6444 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 6445 address stubAddr = NULL; 6446 const char *stubName = NULL; 6447 6448 assert(UseAES, "need AES instruction support"); 6449 6450 switch(id) { 6451 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6452 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6453 stubName = "cipherBlockChaining_encryptAESCrypt"; 6454 break; 6455 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6456 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6457 stubName = "cipherBlockChaining_decryptAESCrypt"; 6458 break; 6459 } 6460 if (stubAddr == NULL) return false; 6461 6462 Node* cipherBlockChaining_object = argument(0); 6463 Node* src = argument(1); 6464 Node* src_offset = argument(2); 6465 Node* len = argument(3); 6466 Node* dest = argument(4); 6467 Node* dest_offset = argument(5); 6468 6469 // (1) src and dest are arrays. 6470 const Type* src_type = src->Value(&_gvn); 6471 const Type* dest_type = dest->Value(&_gvn); 6472 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6473 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6474 assert (top_src != NULL && top_src->klass() != NULL 6475 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6476 6477 // checks are the responsibility of the caller 6478 Node* src_start = src; 6479 Node* dest_start = dest; 6480 if (src_offset != NULL || dest_offset != NULL) { 6481 assert(src_offset != NULL && dest_offset != NULL, ""); 6482 src_start = array_element_address(src, src_offset, T_BYTE); 6483 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6484 } 6485 6486 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6487 // (because of the predicated logic executed earlier). 6488 // so we cast it here safely. 6489 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6490 6491 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6492 if (embeddedCipherObj == NULL) return false; 6493 6494 // cast it to what we know it will be at runtime 6495 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6496 assert(tinst != NULL, "CBC obj is null"); 6497 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6498 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6499 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6500 6501 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6502 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6503 const TypeOopPtr* xtype = aklass->as_instance_type(); 6504 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype); 6505 aescrypt_object = _gvn.transform(aescrypt_object); 6506 6507 // we need to get the start of the aescrypt_object's expanded key array 6508 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6509 if (k_start == NULL) return false; 6510 6511 // similarly, get the start address of the r vector 6512 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6513 if (objRvec == NULL) return false; 6514 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6515 6516 Node* cbcCrypt; 6517 if (Matcher::pass_original_key_for_aes()) { 6518 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6519 // compatibility issues between Java key expansion and SPARC crypto instructions 6520 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6521 if (original_k_start == NULL) return false; 6522 6523 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6524 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6525 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6526 stubAddr, stubName, TypePtr::BOTTOM, 6527 src_start, dest_start, k_start, r_start, len, original_k_start); 6528 } else { 6529 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6530 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6531 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6532 stubAddr, stubName, TypePtr::BOTTOM, 6533 src_start, dest_start, k_start, r_start, len); 6534 } 6535 6536 // return cipher length (int) 6537 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms)); 6538 set_result(retvalue); 6539 return true; 6540 } 6541 6542 //------------------------------get_key_start_from_aescrypt_object----------------------- 6543 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6544 #ifdef PPC64 6545 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6546 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6547 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6548 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6549 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6550 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6551 if (objSessionK == NULL) { 6552 return (Node *) NULL; 6553 } 6554 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6555 #else 6556 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6557 #endif // PPC64 6558 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6559 if (objAESCryptKey == NULL) return (Node *) NULL; 6560 6561 // now have the array, need to get the start address of the K array 6562 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6563 return k_start; 6564 } 6565 6566 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6567 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6568 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6569 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6570 if (objAESCryptKey == NULL) return (Node *) NULL; 6571 6572 // now have the array, need to get the start address of the lastKey array 6573 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6574 return original_k_start; 6575 } 6576 6577 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6578 // Return node representing slow path of predicate check. 6579 // the pseudo code we want to emulate with this predicate is: 6580 // for encryption: 6581 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6582 // for decryption: 6583 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6584 // note cipher==plain is more conservative than the original java code but that's OK 6585 // 6586 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6587 // The receiver was checked for NULL already. 6588 Node* objCBC = argument(0); 6589 6590 // Load embeddedCipher field of CipherBlockChaining object. 6591 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6592 6593 // get AESCrypt klass for instanceOf check 6594 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6595 // will have same classloader as CipherBlockChaining object 6596 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6597 assert(tinst != NULL, "CBCobj is null"); 6598 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6599 6600 // we want to do an instanceof comparison against the AESCrypt class 6601 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6602 if (!klass_AESCrypt->is_loaded()) { 6603 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6604 Node* ctrl = control(); 6605 set_control(top()); // no regular fast path 6606 return ctrl; 6607 } 6608 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6609 6610 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6611 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1))); 6612 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); 6613 6614 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6615 6616 // for encryption, we are done 6617 if (!decrypting) 6618 return instof_false; // even if it is NULL 6619 6620 // for decryption, we need to add a further check to avoid 6621 // taking the intrinsic path when cipher and plain are the same 6622 // see the original java code for why. 6623 RegionNode* region = new(C) RegionNode(3); 6624 region->init_req(1, instof_false); 6625 Node* src = argument(1); 6626 Node* dest = argument(4); 6627 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest)); 6628 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq)); 6629 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6630 region->init_req(2, src_dest_conjoint); 6631 6632 record_for_igvn(region); 6633 return _gvn.transform(region); 6634 } 6635 6636 //------------------------------inline_ghash_processBlocks 6637 bool LibraryCallKit::inline_ghash_processBlocks() { 6638 address stubAddr; 6639 const char *stubName; 6640 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6641 6642 stubAddr = StubRoutines::ghash_processBlocks(); 6643 stubName = "ghash_processBlocks"; 6644 6645 Node* data = argument(0); 6646 Node* offset = argument(1); 6647 Node* len = argument(2); 6648 Node* state = argument(3); 6649 Node* subkeyH = argument(4); 6650 6651 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6652 assert(state_start, "state is NULL"); 6653 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6654 assert(subkeyH_start, "subkeyH is NULL"); 6655 Node* data_start = array_element_address(data, offset, T_BYTE); 6656 assert(data_start, "data is NULL"); 6657 6658 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6659 OptoRuntime::ghash_processBlocks_Type(), 6660 stubAddr, stubName, TypePtr::BOTTOM, 6661 state_start, subkeyH_start, data_start, len); 6662 return true; 6663 } 6664 6665 //------------------------------inline_sha_implCompress----------------------- 6666 // 6667 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6668 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6669 // 6670 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6671 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6672 // 6673 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6674 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6675 // 6676 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6677 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6678 6679 Node* sha_obj = argument(0); 6680 Node* src = argument(1); // type oop 6681 Node* ofs = argument(2); // type int 6682 6683 const Type* src_type = src->Value(&_gvn); 6684 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6685 if (top_src == NULL || top_src->klass() == NULL) { 6686 // failed array check 6687 return false; 6688 } 6689 // Figure out the size and type of the elements we will be copying. 6690 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6691 if (src_elem != T_BYTE) { 6692 return false; 6693 } 6694 // 'src_start' points to src array + offset 6695 Node* src_start = array_element_address(src, ofs, src_elem); 6696 Node* state = NULL; 6697 address stubAddr; 6698 const char *stubName; 6699 6700 switch(id) { 6701 case vmIntrinsics::_sha_implCompress: 6702 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6703 state = get_state_from_sha_object(sha_obj); 6704 stubAddr = StubRoutines::sha1_implCompress(); 6705 stubName = "sha1_implCompress"; 6706 break; 6707 case vmIntrinsics::_sha2_implCompress: 6708 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6709 state = get_state_from_sha_object(sha_obj); 6710 stubAddr = StubRoutines::sha256_implCompress(); 6711 stubName = "sha256_implCompress"; 6712 break; 6713 case vmIntrinsics::_sha5_implCompress: 6714 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6715 state = get_state_from_sha5_object(sha_obj); 6716 stubAddr = StubRoutines::sha512_implCompress(); 6717 stubName = "sha512_implCompress"; 6718 break; 6719 default: 6720 fatal_unexpected_iid(id); 6721 return false; 6722 } 6723 if (state == NULL) return false; 6724 6725 // Call the stub. 6726 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6727 stubAddr, stubName, TypePtr::BOTTOM, 6728 src_start, state); 6729 6730 return true; 6731 } 6732 6733 //------------------------------inline_digestBase_implCompressMB----------------------- 6734 // 6735 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6736 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6737 // 6738 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6739 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6740 "need SHA1/SHA256/SHA512 instruction support"); 6741 assert((uint)predicate < 3, "sanity"); 6742 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6743 6744 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6745 Node* src = argument(1); // byte[] array 6746 Node* ofs = argument(2); // type int 6747 Node* limit = argument(3); // type int 6748 6749 const Type* src_type = src->Value(&_gvn); 6750 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6751 if (top_src == NULL || top_src->klass() == NULL) { 6752 // failed array check 6753 return false; 6754 } 6755 // Figure out the size and type of the elements we will be copying. 6756 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6757 if (src_elem != T_BYTE) { 6758 return false; 6759 } 6760 // 'src_start' points to src array + offset 6761 Node* src_start = array_element_address(src, ofs, src_elem); 6762 6763 const char* klass_SHA_name = NULL; 6764 const char* stub_name = NULL; 6765 address stub_addr = NULL; 6766 bool long_state = false; 6767 6768 switch (predicate) { 6769 case 0: 6770 if (UseSHA1Intrinsics) { 6771 klass_SHA_name = "sun/security/provider/SHA"; 6772 stub_name = "sha1_implCompressMB"; 6773 stub_addr = StubRoutines::sha1_implCompressMB(); 6774 } 6775 break; 6776 case 1: 6777 if (UseSHA256Intrinsics) { 6778 klass_SHA_name = "sun/security/provider/SHA2"; 6779 stub_name = "sha256_implCompressMB"; 6780 stub_addr = StubRoutines::sha256_implCompressMB(); 6781 } 6782 break; 6783 case 2: 6784 if (UseSHA512Intrinsics) { 6785 klass_SHA_name = "sun/security/provider/SHA5"; 6786 stub_name = "sha512_implCompressMB"; 6787 stub_addr = StubRoutines::sha512_implCompressMB(); 6788 long_state = true; 6789 } 6790 break; 6791 default: 6792 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6793 } 6794 if (klass_SHA_name != NULL) { 6795 // get DigestBase klass to lookup for SHA klass 6796 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6797 assert(tinst != NULL, "digestBase_obj is not instance???"); 6798 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6799 6800 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6801 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6802 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6803 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6804 } 6805 return false; 6806 } 6807 //------------------------------inline_sha_implCompressMB----------------------- 6808 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6809 bool long_state, address stubAddr, const char *stubName, 6810 Node* src_start, Node* ofs, Node* limit) { 6811 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6812 const TypeOopPtr* xtype = aklass->as_instance_type(); 6813 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype); 6814 sha_obj = _gvn.transform(sha_obj); 6815 6816 Node* state; 6817 if (long_state) { 6818 state = get_state_from_sha5_object(sha_obj); 6819 } else { 6820 state = get_state_from_sha_object(sha_obj); 6821 } 6822 if (state == NULL) return false; 6823 6824 // Call the stub. 6825 Node *call; 6826 if (CCallingConventionRequiresIntsAsLongs) { 6827 call = make_runtime_call(RC_LEAF|RC_NO_FP, 6828 OptoRuntime::digestBase_implCompressMB_Type(), 6829 stubAddr, stubName, TypePtr::BOTTOM, 6830 src_start, state, ofs XTOP, limit XTOP); 6831 } else { 6832 call = make_runtime_call(RC_LEAF|RC_NO_FP, 6833 OptoRuntime::digestBase_implCompressMB_Type(), 6834 stubAddr, stubName, TypePtr::BOTTOM, 6835 src_start, state, ofs, limit); 6836 } 6837 // return ofs (int) 6838 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 6839 set_result(result); 6840 6841 return true; 6842 } 6843 6844 //------------------------------get_state_from_sha_object----------------------- 6845 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6846 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6847 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6848 if (sha_state == NULL) return (Node *) NULL; 6849 6850 // now have the array, need to get the start address of the state array 6851 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6852 return state; 6853 } 6854 6855 //------------------------------get_state_from_sha5_object----------------------- 6856 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6857 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6858 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6859 if (sha_state == NULL) return (Node *) NULL; 6860 6861 // now have the array, need to get the start address of the state array 6862 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6863 return state; 6864 } 6865 6866 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6867 // Return node representing slow path of predicate check. 6868 // the pseudo code we want to emulate with this predicate is: 6869 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6870 // 6871 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6872 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6873 "need SHA1/SHA256/SHA512 instruction support"); 6874 assert((uint)predicate < 3, "sanity"); 6875 6876 // The receiver was checked for NULL already. 6877 Node* digestBaseObj = argument(0); 6878 6879 // get DigestBase klass for instanceOf check 6880 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6881 assert(tinst != NULL, "digestBaseObj is null"); 6882 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6883 6884 const char* klass_SHA_name = NULL; 6885 switch (predicate) { 6886 case 0: 6887 if (UseSHA1Intrinsics) { 6888 // we want to do an instanceof comparison against the SHA class 6889 klass_SHA_name = "sun/security/provider/SHA"; 6890 } 6891 break; 6892 case 1: 6893 if (UseSHA256Intrinsics) { 6894 // we want to do an instanceof comparison against the SHA2 class 6895 klass_SHA_name = "sun/security/provider/SHA2"; 6896 } 6897 break; 6898 case 2: 6899 if (UseSHA512Intrinsics) { 6900 // we want to do an instanceof comparison against the SHA5 class 6901 klass_SHA_name = "sun/security/provider/SHA5"; 6902 } 6903 break; 6904 default: 6905 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6906 } 6907 6908 ciKlass* klass_SHA = NULL; 6909 if (klass_SHA_name != NULL) { 6910 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6911 } 6912 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6913 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6914 Node* ctrl = control(); 6915 set_control(top()); // no intrinsic path 6916 return ctrl; 6917 } 6918 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6919 6920 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6921 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1))); 6922 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); 6923 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6924 6925 return instof_false; // even if it is NULL 6926 } 6927 6928 bool LibraryCallKit::inline_profileBoolean() { 6929 Node* counts = argument(1); 6930 const TypeAryPtr* ary = NULL; 6931 ciArray* aobj = NULL; 6932 if (counts->is_Con() 6933 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6934 && (aobj = ary->const_oop()->as_array()) != NULL 6935 && (aobj->length() == 2)) { 6936 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6937 jint false_cnt = aobj->element_value(0).as_int(); 6938 jint true_cnt = aobj->element_value(1).as_int(); 6939 6940 if (C->log() != NULL) { 6941 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6942 false_cnt, true_cnt); 6943 } 6944 6945 if (false_cnt + true_cnt == 0) { 6946 // According to profile, never executed. 6947 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6948 Deoptimization::Action_reinterpret); 6949 return true; 6950 } 6951 6952 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6953 // is a number of each value occurrences. 6954 Node* result = argument(0); 6955 if (false_cnt == 0 || true_cnt == 0) { 6956 // According to profile, one value has been never seen. 6957 int expected_val = (false_cnt == 0) ? 1 : 0; 6958 6959 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val))); 6960 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 6961 6962 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6963 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check)); 6964 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check)); 6965 6966 { // Slow path: uncommon trap for never seen value and then reexecute 6967 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6968 // the value has been seen at least once. 6969 PreserveJVMState pjvms(this); 6970 PreserveReexecuteState preexecs(this); 6971 jvms()->set_should_reexecute(true); 6972 6973 set_control(slow_path); 6974 set_i_o(i_o()); 6975 6976 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6977 Deoptimization::Action_reinterpret); 6978 } 6979 // The guard for never seen value enables sharpening of the result and 6980 // returning a constant. It allows to eliminate branches on the same value 6981 // later on. 6982 set_control(fast_path); 6983 result = intcon(expected_val); 6984 } 6985 // Stop profiling. 6986 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6987 // By replacing method body with profile data (represented as ProfileBooleanNode 6988 // on IR level) we effectively disable profiling. 6989 // It enables full speed execution once optimized code is generated. 6990 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt)); 6991 C->record_for_igvn(profile); 6992 set_result(profile); 6993 return true; 6994 } else { 6995 // Continue profiling. 6996 // Profile data isn't available at the moment. So, execute method's bytecode version. 6997 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6998 // is compiled and counters aren't available since corresponding MethodHandle 6999 // isn't a compile-time constant. 7000 return false; 7001 } 7002 }