1 /* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "oops/objArrayKlass.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/callGenerator.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/convertnode.hpp" 38 #include "opto/countbitsnode.hpp" 39 #include "opto/intrinsicnode.hpp" 40 #include "opto/idealKit.hpp" 41 #include "opto/mathexactnode.hpp" 42 #include "opto/movenode.hpp" 43 #include "opto/mulnode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/opaquenode.hpp" 46 #include "opto/parse.hpp" 47 #include "opto/runtime.hpp" 48 #include "opto/subnode.hpp" 49 #include "prims/nativeLookup.hpp" 50 #include "runtime/sharedRuntime.hpp" 51 #include "trace/traceMacros.hpp" 52 53 class LibraryIntrinsic : public InlineCallGenerator { 54 // Extend the set of intrinsics known to the runtime: 55 public: 56 private: 57 bool _is_virtual; 58 bool _does_virtual_dispatch; 59 int8_t _predicates_count; // Intrinsic is predicated by several conditions 60 int8_t _last_predicate; // Last generated predicate 61 vmIntrinsics::ID _intrinsic_id; 62 63 public: 64 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 65 : InlineCallGenerator(m), 66 _is_virtual(is_virtual), 67 _does_virtual_dispatch(does_virtual_dispatch), 68 _predicates_count((int8_t)predicates_count), 69 _last_predicate((int8_t)-1), 70 _intrinsic_id(id) 71 { 72 } 73 virtual bool is_intrinsic() const { return true; } 74 virtual bool is_virtual() const { return _is_virtual; } 75 virtual bool is_predicated() const { return _predicates_count > 0; } 76 virtual int predicates_count() const { return _predicates_count; } 77 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 78 virtual JVMState* generate(JVMState* jvms); 79 virtual Node* generate_predicate(JVMState* jvms, int predicate); 80 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 81 }; 82 83 84 // Local helper class for LibraryIntrinsic: 85 class LibraryCallKit : public GraphKit { 86 private: 87 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 88 Node* _result; // the result node, if any 89 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 90 91 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 92 93 public: 94 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 95 : GraphKit(jvms), 96 _intrinsic(intrinsic), 97 _result(NULL) 98 { 99 // Check if this is a root compile. In that case we don't have a caller. 100 if (!jvms->has_method()) { 101 _reexecute_sp = sp(); 102 } else { 103 // Find out how many arguments the interpreter needs when deoptimizing 104 // and save the stack pointer value so it can used by uncommon_trap. 105 // We find the argument count by looking at the declared signature. 106 bool ignored_will_link; 107 ciSignature* declared_signature = NULL; 108 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 109 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 110 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 111 } 112 } 113 114 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 115 116 ciMethod* caller() const { return jvms()->method(); } 117 int bci() const { return jvms()->bci(); } 118 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 119 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 120 ciMethod* callee() const { return _intrinsic->method(); } 121 122 bool try_to_inline(int predicate); 123 Node* try_to_predicate(int predicate); 124 125 void push_result() { 126 // Push the result onto the stack. 127 if (!stopped() && result() != NULL) { 128 BasicType bt = result()->bottom_type()->basic_type(); 129 push_node(bt, result()); 130 } 131 } 132 133 private: 134 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 135 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 136 } 137 138 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 139 void set_result(RegionNode* region, PhiNode* value); 140 Node* result() { return _result; } 141 142 virtual int reexecute_sp() { return _reexecute_sp; } 143 144 // Helper functions to inline natives 145 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 146 Node* generate_slow_guard(Node* test, RegionNode* region); 147 Node* generate_fair_guard(Node* test, RegionNode* region); 148 Node* generate_negative_guard(Node* index, RegionNode* region, 149 // resulting CastII of index: 150 Node* *pos_index = NULL); 151 Node* generate_limit_guard(Node* offset, Node* subseq_length, 152 Node* array_length, 153 RegionNode* region); 154 Node* generate_current_thread(Node* &tls_output); 155 Node* load_mirror_from_klass(Node* klass); 156 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 157 RegionNode* region, int null_path, 158 int offset); 159 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path) { 161 int offset = java_lang_Class::klass_offset_in_bytes(); 162 return load_klass_from_mirror_common(mirror, never_see_null, 163 region, null_path, 164 offset); 165 } 166 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 167 RegionNode* region, int null_path) { 168 int offset = java_lang_Class::array_klass_offset_in_bytes(); 169 return load_klass_from_mirror_common(mirror, never_see_null, 170 region, null_path, 171 offset); 172 } 173 Node* generate_access_flags_guard(Node* kls, 174 int modifier_mask, int modifier_bits, 175 RegionNode* region); 176 Node* generate_interface_guard(Node* kls, RegionNode* region); 177 Node* generate_array_guard(Node* kls, RegionNode* region) { 178 return generate_array_guard_common(kls, region, false, false); 179 } 180 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 181 return generate_array_guard_common(kls, region, false, true); 182 } 183 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 184 return generate_array_guard_common(kls, region, true, false); 185 } 186 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 187 return generate_array_guard_common(kls, region, true, true); 188 } 189 Node* generate_array_guard_common(Node* kls, RegionNode* region, 190 bool obj_array, bool not_array); 191 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 192 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 193 bool is_virtual = false, bool is_static = false); 194 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 195 return generate_method_call(method_id, false, true); 196 } 197 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 198 return generate_method_call(method_id, true, false); 199 } 200 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 201 202 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 203 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 204 bool inline_string_compareTo(); 205 bool inline_string_indexOf(); 206 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 207 bool inline_string_equals(); 208 Node* round_double_node(Node* n); 209 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 210 bool inline_math_native(vmIntrinsics::ID id); 211 bool inline_trig(vmIntrinsics::ID id); 212 bool inline_math(vmIntrinsics::ID id); 213 template <typename OverflowOp> 214 bool inline_math_overflow(Node* arg1, Node* arg2); 215 void inline_math_mathExact(Node* math, Node* test); 216 bool inline_math_addExactI(bool is_increment); 217 bool inline_math_addExactL(bool is_increment); 218 bool inline_math_multiplyExactI(); 219 bool inline_math_multiplyExactL(); 220 bool inline_math_negateExactI(); 221 bool inline_math_negateExactL(); 222 bool inline_math_subtractExactI(bool is_decrement); 223 bool inline_math_subtractExactL(bool is_decrement); 224 bool inline_exp(); 225 bool inline_pow(); 226 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 227 bool inline_min_max(vmIntrinsics::ID id); 228 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 229 // This returns Type::AnyPtr, RawPtr, or OopPtr. 230 int classify_unsafe_addr(Node* &base, Node* &offset); 231 Node* make_unsafe_address(Node* base, Node* offset); 232 // Helper for inline_unsafe_access. 233 // Generates the guards that check whether the result of 234 // Unsafe.getObject should be recorded in an SATB log buffer. 235 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 236 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 237 static bool klass_needs_init_guard(Node* kls); 238 bool inline_unsafe_allocate(); 239 bool inline_unsafe_copyMemory(); 240 bool inline_native_currentThread(); 241 #ifdef TRACE_HAVE_INTRINSICS 242 bool inline_native_classID(); 243 bool inline_native_threadID(); 244 #endif 245 bool inline_native_time_funcs(address method, const char* funcName); 246 bool inline_native_isInterrupted(); 247 bool inline_native_Class_query(vmIntrinsics::ID id); 248 bool inline_native_subtype_check(); 249 250 bool inline_native_newArray(); 251 bool inline_native_getLength(); 252 bool inline_array_copyOf(bool is_copyOfRange); 253 bool inline_array_equals(); 254 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 255 bool inline_native_clone(bool is_virtual); 256 bool inline_native_Reflection_getCallerClass(); 257 // Helper function for inlining native object hash method 258 bool inline_native_hashcode(bool is_virtual, bool is_static); 259 bool inline_native_getClass(); 260 261 // Helper functions for inlining arraycopy 262 bool inline_arraycopy(); 263 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 264 RegionNode* slow_region); 265 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 266 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 267 268 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 269 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 270 bool inline_unsafe_ordered_store(BasicType type); 271 bool inline_unsafe_fence(vmIntrinsics::ID id); 272 bool inline_fp_conversions(vmIntrinsics::ID id); 273 bool inline_number_methods(vmIntrinsics::ID id); 274 bool inline_reference_get(); 275 bool inline_Class_cast(); 276 bool inline_aescrypt_Block(vmIntrinsics::ID id); 277 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 278 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 279 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 280 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 281 bool inline_sha_implCompress(vmIntrinsics::ID id); 282 bool inline_digestBase_implCompressMB(int predicate); 283 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 284 bool long_state, address stubAddr, const char *stubName, 285 Node* src_start, Node* ofs, Node* limit); 286 Node* get_state_from_sha_object(Node *sha_object); 287 Node* get_state_from_sha5_object(Node *sha_object); 288 Node* inline_digestBase_implCompressMB_predicate(int predicate); 289 bool inline_encodeISOArray(); 290 bool inline_updateCRC32(); 291 bool inline_updateBytesCRC32(); 292 bool inline_updateByteBufferCRC32(); 293 bool inline_multiplyToLen(); 294 295 bool inline_profileBoolean(); 296 bool inline_isCompileConstant(); 297 }; 298 299 300 //---------------------------make_vm_intrinsic---------------------------- 301 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 302 vmIntrinsics::ID id = m->intrinsic_id(); 303 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 304 305 ccstr disable_intr = NULL; 306 307 if ((DisableIntrinsic[0] != '\0' 308 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 309 (method_has_option_value("DisableIntrinsic", disable_intr) 310 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 311 // disabled by a user request on the command line: 312 // example: -XX:DisableIntrinsic=_hashCode,_getClass 313 return NULL; 314 } 315 316 if (!m->is_loaded()) { 317 // do not attempt to inline unloaded methods 318 return NULL; 319 } 320 321 // Only a few intrinsics implement a virtual dispatch. 322 // They are expensive calls which are also frequently overridden. 323 if (is_virtual) { 324 switch (id) { 325 case vmIntrinsics::_hashCode: 326 case vmIntrinsics::_clone: 327 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 328 break; 329 default: 330 return NULL; 331 } 332 } 333 334 // -XX:-InlineNatives disables nearly all intrinsics: 335 if (!InlineNatives) { 336 switch (id) { 337 case vmIntrinsics::_indexOf: 338 case vmIntrinsics::_compareTo: 339 case vmIntrinsics::_equals: 340 case vmIntrinsics::_equalsC: 341 case vmIntrinsics::_getAndAddInt: 342 case vmIntrinsics::_getAndAddLong: 343 case vmIntrinsics::_getAndSetInt: 344 case vmIntrinsics::_getAndSetLong: 345 case vmIntrinsics::_getAndSetObject: 346 case vmIntrinsics::_loadFence: 347 case vmIntrinsics::_storeFence: 348 case vmIntrinsics::_fullFence: 349 break; // InlineNatives does not control String.compareTo 350 case vmIntrinsics::_Reference_get: 351 break; // InlineNatives does not control Reference.get 352 default: 353 return NULL; 354 } 355 } 356 357 int predicates = 0; 358 bool does_virtual_dispatch = false; 359 360 switch (id) { 361 case vmIntrinsics::_compareTo: 362 if (!SpecialStringCompareTo) return NULL; 363 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 364 break; 365 case vmIntrinsics::_indexOf: 366 if (!SpecialStringIndexOf) return NULL; 367 break; 368 case vmIntrinsics::_equals: 369 if (!SpecialStringEquals) return NULL; 370 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 371 break; 372 case vmIntrinsics::_equalsC: 373 if (!SpecialArraysEquals) return NULL; 374 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 375 break; 376 case vmIntrinsics::_arraycopy: 377 if (!InlineArrayCopy) return NULL; 378 break; 379 case vmIntrinsics::_copyMemory: 380 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 381 if (!InlineArrayCopy) return NULL; 382 break; 383 case vmIntrinsics::_hashCode: 384 if (!InlineObjectHash) return NULL; 385 does_virtual_dispatch = true; 386 break; 387 case vmIntrinsics::_clone: 388 does_virtual_dispatch = true; 389 case vmIntrinsics::_copyOf: 390 case vmIntrinsics::_copyOfRange: 391 if (!InlineObjectCopy) return NULL; 392 // These also use the arraycopy intrinsic mechanism: 393 if (!InlineArrayCopy) return NULL; 394 break; 395 case vmIntrinsics::_encodeISOArray: 396 if (!SpecialEncodeISOArray) return NULL; 397 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 398 break; 399 case vmIntrinsics::_checkIndex: 400 // We do not intrinsify this. The optimizer does fine with it. 401 return NULL; 402 403 case vmIntrinsics::_getCallerClass: 404 if (!InlineReflectionGetCallerClass) return NULL; 405 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 406 break; 407 408 case vmIntrinsics::_bitCount_i: 409 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 410 break; 411 412 case vmIntrinsics::_bitCount_l: 413 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 414 break; 415 416 case vmIntrinsics::_numberOfLeadingZeros_i: 417 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 418 break; 419 420 case vmIntrinsics::_numberOfLeadingZeros_l: 421 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 422 break; 423 424 case vmIntrinsics::_numberOfTrailingZeros_i: 425 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 426 break; 427 428 case vmIntrinsics::_numberOfTrailingZeros_l: 429 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 430 break; 431 432 case vmIntrinsics::_reverseBytes_c: 433 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 434 break; 435 case vmIntrinsics::_reverseBytes_s: 436 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 437 break; 438 case vmIntrinsics::_reverseBytes_i: 439 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 440 break; 441 case vmIntrinsics::_reverseBytes_l: 442 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 443 break; 444 445 case vmIntrinsics::_Reference_get: 446 // Use the intrinsic version of Reference.get() so that the value in 447 // the referent field can be registered by the G1 pre-barrier code. 448 // Also add memory barrier to prevent commoning reads from this field 449 // across safepoint since GC can change it value. 450 break; 451 452 case vmIntrinsics::_compareAndSwapObject: 453 #ifdef _LP64 454 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 455 #endif 456 break; 457 458 case vmIntrinsics::_compareAndSwapLong: 459 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 460 break; 461 462 case vmIntrinsics::_getAndAddInt: 463 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 464 break; 465 466 case vmIntrinsics::_getAndAddLong: 467 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 468 break; 469 470 case vmIntrinsics::_getAndSetInt: 471 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 472 break; 473 474 case vmIntrinsics::_getAndSetLong: 475 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 476 break; 477 478 case vmIntrinsics::_getAndSetObject: 479 #ifdef _LP64 480 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 481 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 482 break; 483 #else 484 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 485 break; 486 #endif 487 488 case vmIntrinsics::_aescrypt_encryptBlock: 489 case vmIntrinsics::_aescrypt_decryptBlock: 490 if (!UseAESIntrinsics) return NULL; 491 break; 492 493 case vmIntrinsics::_multiplyToLen: 494 if (!UseMultiplyToLenIntrinsic) return NULL; 495 break; 496 497 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 498 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 499 if (!UseAESIntrinsics) return NULL; 500 // these two require the predicated logic 501 predicates = 1; 502 break; 503 504 case vmIntrinsics::_sha_implCompress: 505 if (!UseSHA1Intrinsics) return NULL; 506 break; 507 508 case vmIntrinsics::_sha2_implCompress: 509 if (!UseSHA256Intrinsics) return NULL; 510 break; 511 512 case vmIntrinsics::_sha5_implCompress: 513 if (!UseSHA512Intrinsics) return NULL; 514 break; 515 516 case vmIntrinsics::_digestBase_implCompressMB: 517 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 518 predicates = 3; 519 break; 520 521 case vmIntrinsics::_updateCRC32: 522 case vmIntrinsics::_updateBytesCRC32: 523 case vmIntrinsics::_updateByteBufferCRC32: 524 if (!UseCRC32Intrinsics) return NULL; 525 break; 526 527 case vmIntrinsics::_incrementExactI: 528 case vmIntrinsics::_addExactI: 529 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 530 break; 531 case vmIntrinsics::_incrementExactL: 532 case vmIntrinsics::_addExactL: 533 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 534 break; 535 case vmIntrinsics::_decrementExactI: 536 case vmIntrinsics::_subtractExactI: 537 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 538 break; 539 case vmIntrinsics::_decrementExactL: 540 case vmIntrinsics::_subtractExactL: 541 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 542 break; 543 case vmIntrinsics::_negateExactI: 544 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 545 break; 546 case vmIntrinsics::_negateExactL: 547 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 548 break; 549 case vmIntrinsics::_multiplyExactI: 550 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 551 break; 552 case vmIntrinsics::_multiplyExactL: 553 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 554 break; 555 556 case vmIntrinsics::_getShortUnaligned: 557 case vmIntrinsics::_getCharUnaligned: 558 case vmIntrinsics::_getIntUnaligned: 559 case vmIntrinsics::_getLongUnaligned: 560 case vmIntrinsics::_putShortUnaligned: 561 case vmIntrinsics::_putCharUnaligned: 562 case vmIntrinsics::_putIntUnaligned: 563 case vmIntrinsics::_putLongUnaligned: 564 if (!UseUnalignedAccesses) return NULL; 565 break; 566 567 default: 568 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 569 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 570 break; 571 } 572 573 // -XX:-InlineClassNatives disables natives from the Class class. 574 // The flag applies to all reflective calls, notably Array.newArray 575 // (visible to Java programmers as Array.newInstance). 576 if (m->holder()->name() == ciSymbol::java_lang_Class() || 577 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 578 if (!InlineClassNatives) return NULL; 579 } 580 581 // -XX:-InlineThreadNatives disables natives from the Thread class. 582 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 583 if (!InlineThreadNatives) return NULL; 584 } 585 586 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 587 if (m->holder()->name() == ciSymbol::java_lang_Math() || 588 m->holder()->name() == ciSymbol::java_lang_Float() || 589 m->holder()->name() == ciSymbol::java_lang_Double()) { 590 if (!InlineMathNatives) return NULL; 591 } 592 593 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 594 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 595 if (!InlineUnsafeOps) return NULL; 596 } 597 598 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 599 } 600 601 //----------------------register_library_intrinsics----------------------- 602 // Initialize this file's data structures, for each Compile instance. 603 void Compile::register_library_intrinsics() { 604 // Nothing to do here. 605 } 606 607 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 608 LibraryCallKit kit(jvms, this); 609 Compile* C = kit.C; 610 int nodes = C->unique(); 611 #ifndef PRODUCT 612 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 613 char buf[1000]; 614 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 615 tty->print_cr("Intrinsic %s", str); 616 } 617 #endif 618 ciMethod* callee = kit.callee(); 619 const int bci = kit.bci(); 620 621 // Try to inline the intrinsic. 622 if (kit.try_to_inline(_last_predicate)) { 623 if (C->print_intrinsics() || C->print_inlining()) { 624 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 625 } 626 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 627 if (C->log()) { 628 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 629 vmIntrinsics::name_at(intrinsic_id()), 630 (is_virtual() ? " virtual='1'" : ""), 631 C->unique() - nodes); 632 } 633 // Push the result from the inlined method onto the stack. 634 kit.push_result(); 635 C->print_inlining_update(this); 636 return kit.transfer_exceptions_into_jvms(); 637 } 638 639 // The intrinsic bailed out 640 if (C->print_intrinsics() || C->print_inlining()) { 641 if (jvms->has_method()) { 642 // Not a root compile. 643 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 644 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 645 } else { 646 // Root compile 647 tty->print("Did not generate intrinsic %s%s at bci:%d in", 648 vmIntrinsics::name_at(intrinsic_id()), 649 (is_virtual() ? " (virtual)" : ""), bci); 650 } 651 } 652 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 653 C->print_inlining_update(this); 654 return NULL; 655 } 656 657 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 658 LibraryCallKit kit(jvms, this); 659 Compile* C = kit.C; 660 int nodes = C->unique(); 661 _last_predicate = predicate; 662 #ifndef PRODUCT 663 assert(is_predicated() && predicate < predicates_count(), "sanity"); 664 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 665 char buf[1000]; 666 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 667 tty->print_cr("Predicate for intrinsic %s", str); 668 } 669 #endif 670 ciMethod* callee = kit.callee(); 671 const int bci = kit.bci(); 672 673 Node* slow_ctl = kit.try_to_predicate(predicate); 674 if (!kit.failing()) { 675 if (C->print_intrinsics() || C->print_inlining()) { 676 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 677 } 678 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 679 if (C->log()) { 680 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 681 vmIntrinsics::name_at(intrinsic_id()), 682 (is_virtual() ? " virtual='1'" : ""), 683 C->unique() - nodes); 684 } 685 return slow_ctl; // Could be NULL if the check folds. 686 } 687 688 // The intrinsic bailed out 689 if (C->print_intrinsics() || C->print_inlining()) { 690 if (jvms->has_method()) { 691 // Not a root compile. 692 const char* msg = "failed to generate predicate for intrinsic"; 693 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 694 } else { 695 // Root compile 696 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 697 vmIntrinsics::name_at(intrinsic_id()), 698 (is_virtual() ? " (virtual)" : ""), bci); 699 } 700 } 701 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 702 return NULL; 703 } 704 705 bool LibraryCallKit::try_to_inline(int predicate) { 706 // Handle symbolic names for otherwise undistinguished boolean switches: 707 const bool is_store = true; 708 const bool is_native_ptr = true; 709 const bool is_static = true; 710 const bool is_volatile = true; 711 712 if (!jvms()->has_method()) { 713 // Root JVMState has a null method. 714 assert(map()->memory()->Opcode() == Op_Parm, ""); 715 // Insert the memory aliasing node 716 set_all_memory(reset_memory()); 717 } 718 assert(merged_memory(), ""); 719 720 721 switch (intrinsic_id()) { 722 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 723 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 724 case vmIntrinsics::_getClass: return inline_native_getClass(); 725 726 case vmIntrinsics::_dsin: 727 case vmIntrinsics::_dcos: 728 case vmIntrinsics::_dtan: 729 case vmIntrinsics::_dabs: 730 case vmIntrinsics::_datan2: 731 case vmIntrinsics::_dsqrt: 732 case vmIntrinsics::_dexp: 733 case vmIntrinsics::_dlog: 734 case vmIntrinsics::_dlog10: 735 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 736 737 case vmIntrinsics::_min: 738 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 739 740 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 741 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 742 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 743 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 744 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 745 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 746 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 747 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 748 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 749 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 750 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 751 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 752 753 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 754 755 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 756 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 757 case vmIntrinsics::_equals: return inline_string_equals(); 758 759 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 760 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 761 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 762 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 763 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 764 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 765 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 766 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 767 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 768 769 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 770 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 771 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 772 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 773 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 774 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 775 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 776 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 777 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 778 779 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 780 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 781 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 782 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 783 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 784 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 785 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 786 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 787 788 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 789 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 790 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 791 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 792 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 793 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 794 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 795 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 796 797 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 798 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 799 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 800 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 801 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 802 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 803 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 804 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 805 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 806 807 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 808 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 809 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 810 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 811 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 812 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 813 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 814 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 815 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 816 817 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 818 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 819 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 820 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 821 822 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 823 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 824 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 825 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 826 827 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 828 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 829 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 830 831 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 832 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 833 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 834 835 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 836 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 837 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 838 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 839 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 840 841 case vmIntrinsics::_loadFence: 842 case vmIntrinsics::_storeFence: 843 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 844 845 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 846 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 847 848 #ifdef TRACE_HAVE_INTRINSICS 849 case vmIntrinsics::_classID: return inline_native_classID(); 850 case vmIntrinsics::_threadID: return inline_native_threadID(); 851 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 852 #endif 853 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 854 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 855 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 856 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 857 case vmIntrinsics::_newArray: return inline_native_newArray(); 858 case vmIntrinsics::_getLength: return inline_native_getLength(); 859 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 860 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 861 case vmIntrinsics::_equalsC: return inline_array_equals(); 862 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 863 864 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 865 866 case vmIntrinsics::_isInstance: 867 case vmIntrinsics::_getModifiers: 868 case vmIntrinsics::_isInterface: 869 case vmIntrinsics::_isArray: 870 case vmIntrinsics::_isPrimitive: 871 case vmIntrinsics::_getSuperclass: 872 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 873 874 case vmIntrinsics::_floatToRawIntBits: 875 case vmIntrinsics::_floatToIntBits: 876 case vmIntrinsics::_intBitsToFloat: 877 case vmIntrinsics::_doubleToRawLongBits: 878 case vmIntrinsics::_doubleToLongBits: 879 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 880 881 case vmIntrinsics::_numberOfLeadingZeros_i: 882 case vmIntrinsics::_numberOfLeadingZeros_l: 883 case vmIntrinsics::_numberOfTrailingZeros_i: 884 case vmIntrinsics::_numberOfTrailingZeros_l: 885 case vmIntrinsics::_bitCount_i: 886 case vmIntrinsics::_bitCount_l: 887 case vmIntrinsics::_reverseBytes_i: 888 case vmIntrinsics::_reverseBytes_l: 889 case vmIntrinsics::_reverseBytes_s: 890 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 891 892 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 893 894 case vmIntrinsics::_Reference_get: return inline_reference_get(); 895 896 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 897 898 case vmIntrinsics::_aescrypt_encryptBlock: 899 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 900 901 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 902 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 903 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 904 905 case vmIntrinsics::_sha_implCompress: 906 case vmIntrinsics::_sha2_implCompress: 907 case vmIntrinsics::_sha5_implCompress: 908 return inline_sha_implCompress(intrinsic_id()); 909 910 case vmIntrinsics::_digestBase_implCompressMB: 911 return inline_digestBase_implCompressMB(predicate); 912 913 case vmIntrinsics::_multiplyToLen: 914 return inline_multiplyToLen(); 915 916 case vmIntrinsics::_encodeISOArray: 917 return inline_encodeISOArray(); 918 919 case vmIntrinsics::_updateCRC32: 920 return inline_updateCRC32(); 921 case vmIntrinsics::_updateBytesCRC32: 922 return inline_updateBytesCRC32(); 923 case vmIntrinsics::_updateByteBufferCRC32: 924 return inline_updateByteBufferCRC32(); 925 926 case vmIntrinsics::_profileBoolean: 927 return inline_profileBoolean(); 928 case vmIntrinsics::_isCompileConstant: 929 return inline_isCompileConstant(); 930 931 default: 932 // If you get here, it may be that someone has added a new intrinsic 933 // to the list in vmSymbols.hpp without implementing it here. 934 #ifndef PRODUCT 935 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 936 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 937 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 938 } 939 #endif 940 return false; 941 } 942 } 943 944 Node* LibraryCallKit::try_to_predicate(int predicate) { 945 if (!jvms()->has_method()) { 946 // Root JVMState has a null method. 947 assert(map()->memory()->Opcode() == Op_Parm, ""); 948 // Insert the memory aliasing node 949 set_all_memory(reset_memory()); 950 } 951 assert(merged_memory(), ""); 952 953 switch (intrinsic_id()) { 954 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 955 return inline_cipherBlockChaining_AESCrypt_predicate(false); 956 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 957 return inline_cipherBlockChaining_AESCrypt_predicate(true); 958 case vmIntrinsics::_digestBase_implCompressMB: 959 return inline_digestBase_implCompressMB_predicate(predicate); 960 961 default: 962 // If you get here, it may be that someone has added a new intrinsic 963 // to the list in vmSymbols.hpp without implementing it here. 964 #ifndef PRODUCT 965 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 966 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 967 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 968 } 969 #endif 970 Node* slow_ctl = control(); 971 set_control(top()); // No fast path instrinsic 972 return slow_ctl; 973 } 974 } 975 976 //------------------------------set_result------------------------------- 977 // Helper function for finishing intrinsics. 978 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 979 record_for_igvn(region); 980 set_control(_gvn.transform(region)); 981 set_result( _gvn.transform(value)); 982 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 983 } 984 985 //------------------------------generate_guard--------------------------- 986 // Helper function for generating guarded fast-slow graph structures. 987 // The given 'test', if true, guards a slow path. If the test fails 988 // then a fast path can be taken. (We generally hope it fails.) 989 // In all cases, GraphKit::control() is updated to the fast path. 990 // The returned value represents the control for the slow path. 991 // The return value is never 'top'; it is either a valid control 992 // or NULL if it is obvious that the slow path can never be taken. 993 // Also, if region and the slow control are not NULL, the slow edge 994 // is appended to the region. 995 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 996 if (stopped()) { 997 // Already short circuited. 998 return NULL; 999 } 1000 1001 // Build an if node and its projections. 1002 // If test is true we take the slow path, which we assume is uncommon. 1003 if (_gvn.type(test) == TypeInt::ZERO) { 1004 // The slow branch is never taken. No need to build this guard. 1005 return NULL; 1006 } 1007 1008 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1009 1010 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1011 if (if_slow == top()) { 1012 // The slow branch is never taken. No need to build this guard. 1013 return NULL; 1014 } 1015 1016 if (region != NULL) 1017 region->add_req(if_slow); 1018 1019 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1020 set_control(if_fast); 1021 1022 return if_slow; 1023 } 1024 1025 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1026 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1027 } 1028 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1029 return generate_guard(test, region, PROB_FAIR); 1030 } 1031 1032 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1033 Node* *pos_index) { 1034 if (stopped()) 1035 return NULL; // already stopped 1036 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1037 return NULL; // index is already adequately typed 1038 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1039 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1040 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1041 if (is_neg != NULL && pos_index != NULL) { 1042 // Emulate effect of Parse::adjust_map_after_if. 1043 Node* ccast = new CastIINode(index, TypeInt::POS); 1044 ccast->set_req(0, control()); 1045 (*pos_index) = _gvn.transform(ccast); 1046 } 1047 return is_neg; 1048 } 1049 1050 // Make sure that 'position' is a valid limit index, in [0..length]. 1051 // There are two equivalent plans for checking this: 1052 // A. (offset + copyLength) unsigned<= arrayLength 1053 // B. offset <= (arrayLength - copyLength) 1054 // We require that all of the values above, except for the sum and 1055 // difference, are already known to be non-negative. 1056 // Plan A is robust in the face of overflow, if offset and copyLength 1057 // are both hugely positive. 1058 // 1059 // Plan B is less direct and intuitive, but it does not overflow at 1060 // all, since the difference of two non-negatives is always 1061 // representable. Whenever Java methods must perform the equivalent 1062 // check they generally use Plan B instead of Plan A. 1063 // For the moment we use Plan A. 1064 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1065 Node* subseq_length, 1066 Node* array_length, 1067 RegionNode* region) { 1068 if (stopped()) 1069 return NULL; // already stopped 1070 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1071 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1072 return NULL; // common case of whole-array copy 1073 Node* last = subseq_length; 1074 if (!zero_offset) // last += offset 1075 last = _gvn.transform(new AddINode(last, offset)); 1076 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1077 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1078 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1079 return is_over; 1080 } 1081 1082 1083 //--------------------------generate_current_thread-------------------- 1084 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1085 ciKlass* thread_klass = env()->Thread_klass(); 1086 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1087 Node* thread = _gvn.transform(new ThreadLocalNode()); 1088 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1089 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1090 tls_output = thread; 1091 return threadObj; 1092 } 1093 1094 1095 //------------------------------make_string_method_node------------------------ 1096 // Helper method for String intrinsic functions. This version is called 1097 // with str1 and str2 pointing to String object nodes. 1098 // 1099 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1100 Node* no_ctrl = NULL; 1101 1102 // Get start addr of string 1103 Node* str1_value = load_String_value(no_ctrl, str1); 1104 Node* str1_offset = load_String_offset(no_ctrl, str1); 1105 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1106 1107 // Get length of string 1 1108 Node* str1_len = load_String_length(no_ctrl, str1); 1109 1110 Node* str2_value = load_String_value(no_ctrl, str2); 1111 Node* str2_offset = load_String_offset(no_ctrl, str2); 1112 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1113 1114 Node* str2_len = NULL; 1115 Node* result = NULL; 1116 1117 switch (opcode) { 1118 case Op_StrIndexOf: 1119 // Get length of string 2 1120 str2_len = load_String_length(no_ctrl, str2); 1121 1122 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1123 str1_start, str1_len, str2_start, str2_len); 1124 break; 1125 case Op_StrComp: 1126 // Get length of string 2 1127 str2_len = load_String_length(no_ctrl, str2); 1128 1129 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1130 str1_start, str1_len, str2_start, str2_len); 1131 break; 1132 case Op_StrEquals: 1133 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1134 str1_start, str2_start, str1_len); 1135 break; 1136 default: 1137 ShouldNotReachHere(); 1138 return NULL; 1139 } 1140 1141 // All these intrinsics have checks. 1142 C->set_has_split_ifs(true); // Has chance for split-if optimization 1143 1144 return _gvn.transform(result); 1145 } 1146 1147 // Helper method for String intrinsic functions. This version is called 1148 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1149 // to Int nodes containing the lenghts of str1 and str2. 1150 // 1151 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1152 Node* result = NULL; 1153 switch (opcode) { 1154 case Op_StrIndexOf: 1155 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1156 str1_start, cnt1, str2_start, cnt2); 1157 break; 1158 case Op_StrComp: 1159 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1160 str1_start, cnt1, str2_start, cnt2); 1161 break; 1162 case Op_StrEquals: 1163 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1164 str1_start, str2_start, cnt1); 1165 break; 1166 default: 1167 ShouldNotReachHere(); 1168 return NULL; 1169 } 1170 1171 // All these intrinsics have checks. 1172 C->set_has_split_ifs(true); // Has chance for split-if optimization 1173 1174 return _gvn.transform(result); 1175 } 1176 1177 //------------------------------inline_string_compareTo------------------------ 1178 // public int java.lang.String.compareTo(String anotherString); 1179 bool LibraryCallKit::inline_string_compareTo() { 1180 Node* receiver = null_check(argument(0)); 1181 Node* arg = null_check(argument(1)); 1182 if (stopped()) { 1183 return true; 1184 } 1185 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1186 return true; 1187 } 1188 1189 //------------------------------inline_string_equals------------------------ 1190 bool LibraryCallKit::inline_string_equals() { 1191 Node* receiver = null_check_receiver(); 1192 // NOTE: Do not null check argument for String.equals() because spec 1193 // allows to specify NULL as argument. 1194 Node* argument = this->argument(1); 1195 if (stopped()) { 1196 return true; 1197 } 1198 1199 // paths (plus control) merge 1200 RegionNode* region = new RegionNode(5); 1201 Node* phi = new PhiNode(region, TypeInt::BOOL); 1202 1203 // does source == target string? 1204 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1205 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1206 1207 Node* if_eq = generate_slow_guard(bol, NULL); 1208 if (if_eq != NULL) { 1209 // receiver == argument 1210 phi->init_req(2, intcon(1)); 1211 region->init_req(2, if_eq); 1212 } 1213 1214 // get String klass for instanceOf 1215 ciInstanceKlass* klass = env()->String_klass(); 1216 1217 if (!stopped()) { 1218 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1219 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1220 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1221 1222 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1223 //instanceOf == true, fallthrough 1224 1225 if (inst_false != NULL) { 1226 phi->init_req(3, intcon(0)); 1227 region->init_req(3, inst_false); 1228 } 1229 } 1230 1231 if (!stopped()) { 1232 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1233 1234 // Properly cast the argument to String 1235 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1236 // This path is taken only when argument's type is String:NotNull. 1237 argument = cast_not_null(argument, false); 1238 1239 Node* no_ctrl = NULL; 1240 1241 // Get start addr of receiver 1242 Node* receiver_val = load_String_value(no_ctrl, receiver); 1243 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1244 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1245 1246 // Get length of receiver 1247 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1248 1249 // Get start addr of argument 1250 Node* argument_val = load_String_value(no_ctrl, argument); 1251 Node* argument_offset = load_String_offset(no_ctrl, argument); 1252 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1253 1254 // Get length of argument 1255 Node* argument_cnt = load_String_length(no_ctrl, argument); 1256 1257 // Check for receiver count != argument count 1258 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1259 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1260 Node* if_ne = generate_slow_guard(bol, NULL); 1261 if (if_ne != NULL) { 1262 phi->init_req(4, intcon(0)); 1263 region->init_req(4, if_ne); 1264 } 1265 1266 // Check for count == 0 is done by assembler code for StrEquals. 1267 1268 if (!stopped()) { 1269 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1270 phi->init_req(1, equals); 1271 region->init_req(1, control()); 1272 } 1273 } 1274 1275 // post merge 1276 set_control(_gvn.transform(region)); 1277 record_for_igvn(region); 1278 1279 set_result(_gvn.transform(phi)); 1280 return true; 1281 } 1282 1283 //------------------------------inline_array_equals---------------------------- 1284 bool LibraryCallKit::inline_array_equals() { 1285 Node* arg1 = argument(0); 1286 Node* arg2 = argument(1); 1287 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1288 return true; 1289 } 1290 1291 // Java version of String.indexOf(constant string) 1292 // class StringDecl { 1293 // StringDecl(char[] ca) { 1294 // offset = 0; 1295 // count = ca.length; 1296 // value = ca; 1297 // } 1298 // int offset; 1299 // int count; 1300 // char[] value; 1301 // } 1302 // 1303 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1304 // int targetOffset, int cache_i, int md2) { 1305 // int cache = cache_i; 1306 // int sourceOffset = string_object.offset; 1307 // int sourceCount = string_object.count; 1308 // int targetCount = target_object.length; 1309 // 1310 // int targetCountLess1 = targetCount - 1; 1311 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1312 // 1313 // char[] source = string_object.value; 1314 // char[] target = target_object; 1315 // int lastChar = target[targetCountLess1]; 1316 // 1317 // outer_loop: 1318 // for (int i = sourceOffset; i < sourceEnd; ) { 1319 // int src = source[i + targetCountLess1]; 1320 // if (src == lastChar) { 1321 // // With random strings and a 4-character alphabet, 1322 // // reverse matching at this point sets up 0.8% fewer 1323 // // frames, but (paradoxically) makes 0.3% more probes. 1324 // // Since those probes are nearer the lastChar probe, 1325 // // there is may be a net D$ win with reverse matching. 1326 // // But, reversing loop inhibits unroll of inner loop 1327 // // for unknown reason. So, does running outer loop from 1328 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1329 // for (int j = 0; j < targetCountLess1; j++) { 1330 // if (target[targetOffset + j] != source[i+j]) { 1331 // if ((cache & (1 << source[i+j])) == 0) { 1332 // if (md2 < j+1) { 1333 // i += j+1; 1334 // continue outer_loop; 1335 // } 1336 // } 1337 // i += md2; 1338 // continue outer_loop; 1339 // } 1340 // } 1341 // return i - sourceOffset; 1342 // } 1343 // if ((cache & (1 << src)) == 0) { 1344 // i += targetCountLess1; 1345 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1346 // i++; 1347 // } 1348 // return -1; 1349 // } 1350 1351 //------------------------------string_indexOf------------------------ 1352 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1353 jint cache_i, jint md2_i) { 1354 1355 Node* no_ctrl = NULL; 1356 float likely = PROB_LIKELY(0.9); 1357 float unlikely = PROB_UNLIKELY(0.9); 1358 1359 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1360 1361 Node* source = load_String_value(no_ctrl, string_object); 1362 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1363 Node* sourceCount = load_String_length(no_ctrl, string_object); 1364 1365 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1366 jint target_length = target_array->length(); 1367 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1368 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1369 1370 // String.value field is known to be @Stable. 1371 if (UseImplicitStableValues) { 1372 target = cast_array_to_stable(target, target_type); 1373 } 1374 1375 IdealKit kit(this, false, true); 1376 #define __ kit. 1377 Node* zero = __ ConI(0); 1378 Node* one = __ ConI(1); 1379 Node* cache = __ ConI(cache_i); 1380 Node* md2 = __ ConI(md2_i); 1381 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1382 Node* targetCountLess1 = __ ConI(target_length - 1); 1383 Node* targetOffset = __ ConI(targetOffset_i); 1384 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1385 1386 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1387 Node* outer_loop = __ make_label(2 /* goto */); 1388 Node* return_ = __ make_label(1); 1389 1390 __ set(rtn,__ ConI(-1)); 1391 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1392 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1393 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1394 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1395 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1396 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1397 Node* tpj = __ AddI(targetOffset, __ value(j)); 1398 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1399 Node* ipj = __ AddI(__ value(i), __ value(j)); 1400 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1401 __ if_then(targ, BoolTest::ne, src2); { 1402 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1403 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1404 __ increment(i, __ AddI(__ value(j), one)); 1405 __ goto_(outer_loop); 1406 } __ end_if(); __ dead(j); 1407 }__ end_if(); __ dead(j); 1408 __ increment(i, md2); 1409 __ goto_(outer_loop); 1410 }__ end_if(); 1411 __ increment(j, one); 1412 }__ end_loop(); __ dead(j); 1413 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1414 __ goto_(return_); 1415 }__ end_if(); 1416 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1417 __ increment(i, targetCountLess1); 1418 }__ end_if(); 1419 __ increment(i, one); 1420 __ bind(outer_loop); 1421 }__ end_loop(); __ dead(i); 1422 __ bind(return_); 1423 1424 // Final sync IdealKit and GraphKit. 1425 final_sync(kit); 1426 Node* result = __ value(rtn); 1427 #undef __ 1428 C->set_has_loops(true); 1429 return result; 1430 } 1431 1432 //------------------------------inline_string_indexOf------------------------ 1433 bool LibraryCallKit::inline_string_indexOf() { 1434 Node* receiver = argument(0); 1435 Node* arg = argument(1); 1436 1437 Node* result; 1438 if (Matcher::has_match_rule(Op_StrIndexOf) && 1439 UseSSE42Intrinsics) { 1440 // Generate SSE4.2 version of indexOf 1441 // We currently only have match rules that use SSE4.2 1442 1443 receiver = null_check(receiver); 1444 arg = null_check(arg); 1445 if (stopped()) { 1446 return true; 1447 } 1448 1449 // Make the merge point 1450 RegionNode* result_rgn = new RegionNode(4); 1451 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1452 Node* no_ctrl = NULL; 1453 1454 // Get start addr of source string 1455 Node* source = load_String_value(no_ctrl, receiver); 1456 Node* source_offset = load_String_offset(no_ctrl, receiver); 1457 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1458 1459 // Get length of source string 1460 Node* source_cnt = load_String_length(no_ctrl, receiver); 1461 1462 // Get start addr of substring 1463 Node* substr = load_String_value(no_ctrl, arg); 1464 Node* substr_offset = load_String_offset(no_ctrl, arg); 1465 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1466 1467 // Get length of source string 1468 Node* substr_cnt = load_String_length(no_ctrl, arg); 1469 1470 // Check for substr count > string count 1471 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1472 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1473 Node* if_gt = generate_slow_guard(bol, NULL); 1474 if (if_gt != NULL) { 1475 result_phi->init_req(2, intcon(-1)); 1476 result_rgn->init_req(2, if_gt); 1477 } 1478 1479 if (!stopped()) { 1480 // Check for substr count == 0 1481 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1482 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1483 Node* if_zero = generate_slow_guard(bol, NULL); 1484 if (if_zero != NULL) { 1485 result_phi->init_req(3, intcon(0)); 1486 result_rgn->init_req(3, if_zero); 1487 } 1488 } 1489 1490 if (!stopped()) { 1491 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1492 result_phi->init_req(1, result); 1493 result_rgn->init_req(1, control()); 1494 } 1495 set_control(_gvn.transform(result_rgn)); 1496 record_for_igvn(result_rgn); 1497 result = _gvn.transform(result_phi); 1498 1499 } else { // Use LibraryCallKit::string_indexOf 1500 // don't intrinsify if argument isn't a constant string. 1501 if (!arg->is_Con()) { 1502 return false; 1503 } 1504 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1505 if (str_type == NULL) { 1506 return false; 1507 } 1508 ciInstanceKlass* klass = env()->String_klass(); 1509 ciObject* str_const = str_type->const_oop(); 1510 if (str_const == NULL || str_const->klass() != klass) { 1511 return false; 1512 } 1513 ciInstance* str = str_const->as_instance(); 1514 assert(str != NULL, "must be instance"); 1515 1516 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1517 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1518 1519 int o; 1520 int c; 1521 if (java_lang_String::has_offset_field()) { 1522 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1523 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1524 } else { 1525 o = 0; 1526 c = pat->length(); 1527 } 1528 1529 // constant strings have no offset and count == length which 1530 // simplifies the resulting code somewhat so lets optimize for that. 1531 if (o != 0 || c != pat->length()) { 1532 return false; 1533 } 1534 1535 receiver = null_check(receiver, T_OBJECT); 1536 // NOTE: No null check on the argument is needed since it's a constant String oop. 1537 if (stopped()) { 1538 return true; 1539 } 1540 1541 // The null string as a pattern always returns 0 (match at beginning of string) 1542 if (c == 0) { 1543 set_result(intcon(0)); 1544 return true; 1545 } 1546 1547 // Generate default indexOf 1548 jchar lastChar = pat->char_at(o + (c - 1)); 1549 int cache = 0; 1550 int i; 1551 for (i = 0; i < c - 1; i++) { 1552 assert(i < pat->length(), "out of range"); 1553 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1554 } 1555 1556 int md2 = c; 1557 for (i = 0; i < c - 1; i++) { 1558 assert(i < pat->length(), "out of range"); 1559 if (pat->char_at(o + i) == lastChar) { 1560 md2 = (c - 1) - i; 1561 } 1562 } 1563 1564 result = string_indexOf(receiver, pat, o, cache, md2); 1565 } 1566 set_result(result); 1567 return true; 1568 } 1569 1570 //--------------------------round_double_node-------------------------------- 1571 // Round a double node if necessary. 1572 Node* LibraryCallKit::round_double_node(Node* n) { 1573 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1574 n = _gvn.transform(new RoundDoubleNode(0, n)); 1575 return n; 1576 } 1577 1578 //------------------------------inline_math----------------------------------- 1579 // public static double Math.abs(double) 1580 // public static double Math.sqrt(double) 1581 // public static double Math.log(double) 1582 // public static double Math.log10(double) 1583 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1584 Node* arg = round_double_node(argument(0)); 1585 Node* n; 1586 switch (id) { 1587 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1588 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1589 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1590 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1591 default: fatal_unexpected_iid(id); break; 1592 } 1593 set_result(_gvn.transform(n)); 1594 return true; 1595 } 1596 1597 //------------------------------inline_trig---------------------------------- 1598 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1599 // argument reduction which will turn into a fast/slow diamond. 1600 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1601 Node* arg = round_double_node(argument(0)); 1602 Node* n = NULL; 1603 1604 switch (id) { 1605 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1606 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1607 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1608 default: fatal_unexpected_iid(id); break; 1609 } 1610 n = _gvn.transform(n); 1611 1612 // Rounding required? Check for argument reduction! 1613 if (Matcher::strict_fp_requires_explicit_rounding) { 1614 static const double pi_4 = 0.7853981633974483; 1615 static const double neg_pi_4 = -0.7853981633974483; 1616 // pi/2 in 80-bit extended precision 1617 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1618 // -pi/2 in 80-bit extended precision 1619 // 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}; 1620 // Cutoff value for using this argument reduction technique 1621 //static const double pi_2_minus_epsilon = 1.564660403643354; 1622 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1623 1624 // Pseudocode for sin: 1625 // if (x <= Math.PI / 4.0) { 1626 // if (x >= -Math.PI / 4.0) return fsin(x); 1627 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1628 // } else { 1629 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1630 // } 1631 // return StrictMath.sin(x); 1632 1633 // Pseudocode for cos: 1634 // if (x <= Math.PI / 4.0) { 1635 // if (x >= -Math.PI / 4.0) return fcos(x); 1636 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1637 // } else { 1638 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1639 // } 1640 // return StrictMath.cos(x); 1641 1642 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1643 // requires a special machine instruction to load it. Instead we'll try 1644 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1645 // probably do the math inside the SIN encoding. 1646 1647 // Make the merge point 1648 RegionNode* r = new RegionNode(3); 1649 Node* phi = new PhiNode(r, Type::DOUBLE); 1650 1651 // Flatten arg so we need only 1 test 1652 Node *abs = _gvn.transform(new AbsDNode(arg)); 1653 // Node for PI/4 constant 1654 Node *pi4 = makecon(TypeD::make(pi_4)); 1655 // Check PI/4 : abs(arg) 1656 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1657 // Check: If PI/4 < abs(arg) then go slow 1658 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1659 // Branch either way 1660 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1661 set_control(opt_iff(r,iff)); 1662 1663 // Set fast path result 1664 phi->init_req(2, n); 1665 1666 // Slow path - non-blocking leaf call 1667 Node* call = NULL; 1668 switch (id) { 1669 case vmIntrinsics::_dsin: 1670 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1671 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1672 "Sin", NULL, arg, top()); 1673 break; 1674 case vmIntrinsics::_dcos: 1675 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1676 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1677 "Cos", NULL, arg, top()); 1678 break; 1679 case vmIntrinsics::_dtan: 1680 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1681 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1682 "Tan", NULL, arg, top()); 1683 break; 1684 } 1685 assert(control()->in(0) == call, ""); 1686 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1687 r->init_req(1, control()); 1688 phi->init_req(1, slow_result); 1689 1690 // Post-merge 1691 set_control(_gvn.transform(r)); 1692 record_for_igvn(r); 1693 n = _gvn.transform(phi); 1694 1695 C->set_has_split_ifs(true); // Has chance for split-if optimization 1696 } 1697 set_result(n); 1698 return true; 1699 } 1700 1701 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1702 //------------------- 1703 //result=(result.isNaN())? funcAddr():result; 1704 // Check: If isNaN() by checking result!=result? then either trap 1705 // or go to runtime 1706 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1707 // Build the boolean node 1708 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1709 1710 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1711 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1712 // The pow or exp intrinsic returned a NaN, which requires a call 1713 // to the runtime. Recompile with the runtime call. 1714 uncommon_trap(Deoptimization::Reason_intrinsic, 1715 Deoptimization::Action_make_not_entrant); 1716 } 1717 return result; 1718 } else { 1719 // If this inlining ever returned NaN in the past, we compile a call 1720 // to the runtime to properly handle corner cases 1721 1722 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1723 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1724 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1725 1726 if (!if_slow->is_top()) { 1727 RegionNode* result_region = new RegionNode(3); 1728 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1729 1730 result_region->init_req(1, if_fast); 1731 result_val->init_req(1, result); 1732 1733 set_control(if_slow); 1734 1735 const TypePtr* no_memory_effects = NULL; 1736 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1737 no_memory_effects, 1738 x, top(), y, y ? top() : NULL); 1739 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1740 #ifdef ASSERT 1741 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1742 assert(value_top == top(), "second value must be top"); 1743 #endif 1744 1745 result_region->init_req(2, control()); 1746 result_val->init_req(2, value); 1747 set_control(_gvn.transform(result_region)); 1748 return _gvn.transform(result_val); 1749 } else { 1750 return result; 1751 } 1752 } 1753 } 1754 1755 //------------------------------inline_exp------------------------------------- 1756 // Inline exp instructions, if possible. The Intel hardware only misses 1757 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1758 bool LibraryCallKit::inline_exp() { 1759 Node* arg = round_double_node(argument(0)); 1760 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1761 1762 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1763 set_result(n); 1764 1765 C->set_has_split_ifs(true); // Has chance for split-if optimization 1766 return true; 1767 } 1768 1769 //------------------------------inline_pow------------------------------------- 1770 // Inline power instructions, if possible. 1771 bool LibraryCallKit::inline_pow() { 1772 // Pseudocode for pow 1773 // if (y == 2) { 1774 // return x * x; 1775 // } else { 1776 // if (x <= 0.0) { 1777 // long longy = (long)y; 1778 // if ((double)longy == y) { // if y is long 1779 // if (y + 1 == y) longy = 0; // huge number: even 1780 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1781 // } else { 1782 // result = NaN; 1783 // } 1784 // } else { 1785 // result = DPow(x,y); 1786 // } 1787 // if (result != result)? { 1788 // result = uncommon_trap() or runtime_call(); 1789 // } 1790 // return result; 1791 // } 1792 1793 Node* x = round_double_node(argument(0)); 1794 Node* y = round_double_node(argument(2)); 1795 1796 Node* result = NULL; 1797 1798 Node* const_two_node = makecon(TypeD::make(2.0)); 1799 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1800 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1801 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1802 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1803 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1804 1805 RegionNode* region_node = new RegionNode(3); 1806 region_node->init_req(1, if_true); 1807 1808 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1809 // special case for x^y where y == 2, we can convert it to x * x 1810 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1811 1812 // set control to if_false since we will now process the false branch 1813 set_control(if_false); 1814 1815 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1816 // Short form: skip the fancy tests and just check for NaN result. 1817 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1818 } else { 1819 // If this inlining ever returned NaN in the past, include all 1820 // checks + call to the runtime. 1821 1822 // Set the merge point for If node with condition of (x <= 0.0) 1823 // There are four possible paths to region node and phi node 1824 RegionNode *r = new RegionNode(4); 1825 Node *phi = new PhiNode(r, Type::DOUBLE); 1826 1827 // Build the first if node: if (x <= 0.0) 1828 // Node for 0 constant 1829 Node *zeronode = makecon(TypeD::ZERO); 1830 // Check x:0 1831 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1832 // Check: If (x<=0) then go complex path 1833 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1834 // Branch either way 1835 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1836 // Fast path taken; set region slot 3 1837 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1838 r->init_req(3,fast_taken); // Capture fast-control 1839 1840 // Fast path not-taken, i.e. slow path 1841 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1842 1843 // Set fast path result 1844 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1845 phi->init_req(3, fast_result); 1846 1847 // Complex path 1848 // Build the second if node (if y is long) 1849 // Node for (long)y 1850 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1851 // Node for (double)((long) y) 1852 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1853 // Check (double)((long) y) : y 1854 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1855 // Check if (y isn't long) then go to slow path 1856 1857 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1858 // Branch either way 1859 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1860 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1861 1862 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1863 1864 // Calculate DPow(abs(x), y)*(1 & (long)y) 1865 // Node for constant 1 1866 Node *conone = longcon(1); 1867 // 1& (long)y 1868 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1869 1870 // A huge number is always even. Detect a huge number by checking 1871 // if y + 1 == y and set integer to be tested for parity to 0. 1872 // Required for corner case: 1873 // (long)9.223372036854776E18 = max_jlong 1874 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1875 // max_jlong is odd but 9.223372036854776E18 is even 1876 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1877 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1878 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1879 Node* correctedsign = NULL; 1880 if (ConditionalMoveLimit != 0) { 1881 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1882 } else { 1883 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1884 RegionNode *r = new RegionNode(3); 1885 Node *phi = new PhiNode(r, TypeLong::LONG); 1886 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1887 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1888 phi->init_req(1, signnode); 1889 phi->init_req(2, longcon(0)); 1890 correctedsign = _gvn.transform(phi); 1891 ylong_path = _gvn.transform(r); 1892 record_for_igvn(r); 1893 } 1894 1895 // zero node 1896 Node *conzero = longcon(0); 1897 // Check (1&(long)y)==0? 1898 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1899 // Check if (1&(long)y)!=0?, if so the result is negative 1900 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1901 // abs(x) 1902 Node *absx=_gvn.transform(new AbsDNode(x)); 1903 // abs(x)^y 1904 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1905 // -abs(x)^y 1906 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1907 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1908 Node *signresult = NULL; 1909 if (ConditionalMoveLimit != 0) { 1910 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1911 } else { 1912 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1913 RegionNode *r = new RegionNode(3); 1914 Node *phi = new PhiNode(r, Type::DOUBLE); 1915 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1916 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1917 phi->init_req(1, absxpowy); 1918 phi->init_req(2, negabsxpowy); 1919 signresult = _gvn.transform(phi); 1920 ylong_path = _gvn.transform(r); 1921 record_for_igvn(r); 1922 } 1923 // Set complex path fast result 1924 r->init_req(2, ylong_path); 1925 phi->init_req(2, signresult); 1926 1927 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1928 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1929 r->init_req(1,slow_path); 1930 phi->init_req(1,slow_result); 1931 1932 // Post merge 1933 set_control(_gvn.transform(r)); 1934 record_for_igvn(r); 1935 result = _gvn.transform(phi); 1936 } 1937 1938 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1939 1940 // control from finish_pow_exp is now input to the region node 1941 region_node->set_req(2, control()); 1942 // the result from finish_pow_exp is now input to the phi node 1943 phi_node->init_req(2, result); 1944 set_control(_gvn.transform(region_node)); 1945 record_for_igvn(region_node); 1946 set_result(_gvn.transform(phi_node)); 1947 1948 C->set_has_split_ifs(true); // Has chance for split-if optimization 1949 return true; 1950 } 1951 1952 //------------------------------runtime_math----------------------------- 1953 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1954 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1955 "must be (DD)D or (D)D type"); 1956 1957 // Inputs 1958 Node* a = round_double_node(argument(0)); 1959 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1960 1961 const TypePtr* no_memory_effects = NULL; 1962 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1963 no_memory_effects, 1964 a, top(), b, b ? top() : NULL); 1965 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1966 #ifdef ASSERT 1967 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1968 assert(value_top == top(), "second value must be top"); 1969 #endif 1970 1971 set_result(value); 1972 return true; 1973 } 1974 1975 //------------------------------inline_math_native----------------------------- 1976 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1977 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1978 switch (id) { 1979 // These intrinsics are not properly supported on all hardware 1980 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1981 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1982 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1983 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1984 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1985 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1986 1987 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1988 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1989 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1990 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1991 1992 // These intrinsics are supported on all hardware 1993 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1994 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1995 1996 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1997 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1998 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1999 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2000 #undef FN_PTR 2001 2002 // These intrinsics are not yet correctly implemented 2003 case vmIntrinsics::_datan2: 2004 return false; 2005 2006 default: 2007 fatal_unexpected_iid(id); 2008 return false; 2009 } 2010 } 2011 2012 static bool is_simple_name(Node* n) { 2013 return (n->req() == 1 // constant 2014 || (n->is_Type() && n->as_Type()->type()->singleton()) 2015 || n->is_Proj() // parameter or return value 2016 || n->is_Phi() // local of some sort 2017 ); 2018 } 2019 2020 //----------------------------inline_min_max----------------------------------- 2021 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2022 set_result(generate_min_max(id, argument(0), argument(1))); 2023 return true; 2024 } 2025 2026 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2027 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2028 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2029 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2030 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2031 2032 { 2033 PreserveJVMState pjvms(this); 2034 PreserveReexecuteState preexecs(this); 2035 jvms()->set_should_reexecute(true); 2036 2037 set_control(slow_path); 2038 set_i_o(i_o()); 2039 2040 uncommon_trap(Deoptimization::Reason_intrinsic, 2041 Deoptimization::Action_none); 2042 } 2043 2044 set_control(fast_path); 2045 set_result(math); 2046 } 2047 2048 template <typename OverflowOp> 2049 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2050 typedef typename OverflowOp::MathOp MathOp; 2051 2052 MathOp* mathOp = new MathOp(arg1, arg2); 2053 Node* operation = _gvn.transform( mathOp ); 2054 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2055 inline_math_mathExact(operation, ofcheck); 2056 return true; 2057 } 2058 2059 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2060 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2061 } 2062 2063 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2064 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2065 } 2066 2067 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2068 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2069 } 2070 2071 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2072 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2073 } 2074 2075 bool LibraryCallKit::inline_math_negateExactI() { 2076 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2077 } 2078 2079 bool LibraryCallKit::inline_math_negateExactL() { 2080 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2081 } 2082 2083 bool LibraryCallKit::inline_math_multiplyExactI() { 2084 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2085 } 2086 2087 bool LibraryCallKit::inline_math_multiplyExactL() { 2088 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2089 } 2090 2091 Node* 2092 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2093 // These are the candidate return value: 2094 Node* xvalue = x0; 2095 Node* yvalue = y0; 2096 2097 if (xvalue == yvalue) { 2098 return xvalue; 2099 } 2100 2101 bool want_max = (id == vmIntrinsics::_max); 2102 2103 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2104 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2105 if (txvalue == NULL || tyvalue == NULL) return top(); 2106 // This is not really necessary, but it is consistent with a 2107 // hypothetical MaxINode::Value method: 2108 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2109 2110 // %%% This folding logic should (ideally) be in a different place. 2111 // Some should be inside IfNode, and there to be a more reliable 2112 // transformation of ?: style patterns into cmoves. We also want 2113 // more powerful optimizations around cmove and min/max. 2114 2115 // Try to find a dominating comparison of these guys. 2116 // It can simplify the index computation for Arrays.copyOf 2117 // and similar uses of System.arraycopy. 2118 // First, compute the normalized version of CmpI(x, y). 2119 int cmp_op = Op_CmpI; 2120 Node* xkey = xvalue; 2121 Node* ykey = yvalue; 2122 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2123 if (ideal_cmpxy->is_Cmp()) { 2124 // E.g., if we have CmpI(length - offset, count), 2125 // it might idealize to CmpI(length, count + offset) 2126 cmp_op = ideal_cmpxy->Opcode(); 2127 xkey = ideal_cmpxy->in(1); 2128 ykey = ideal_cmpxy->in(2); 2129 } 2130 2131 // Start by locating any relevant comparisons. 2132 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2133 Node* cmpxy = NULL; 2134 Node* cmpyx = NULL; 2135 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2136 Node* cmp = start_from->fast_out(k); 2137 if (cmp->outcnt() > 0 && // must have prior uses 2138 cmp->in(0) == NULL && // must be context-independent 2139 cmp->Opcode() == cmp_op) { // right kind of compare 2140 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2141 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2142 } 2143 } 2144 2145 const int NCMPS = 2; 2146 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2147 int cmpn; 2148 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2149 if (cmps[cmpn] != NULL) break; // find a result 2150 } 2151 if (cmpn < NCMPS) { 2152 // Look for a dominating test that tells us the min and max. 2153 int depth = 0; // Limit search depth for speed 2154 Node* dom = control(); 2155 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2156 if (++depth >= 100) break; 2157 Node* ifproj = dom; 2158 if (!ifproj->is_Proj()) continue; 2159 Node* iff = ifproj->in(0); 2160 if (!iff->is_If()) continue; 2161 Node* bol = iff->in(1); 2162 if (!bol->is_Bool()) continue; 2163 Node* cmp = bol->in(1); 2164 if (cmp == NULL) continue; 2165 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2166 if (cmps[cmpn] == cmp) break; 2167 if (cmpn == NCMPS) continue; 2168 BoolTest::mask btest = bol->as_Bool()->_test._test; 2169 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2170 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2171 // At this point, we know that 'x btest y' is true. 2172 switch (btest) { 2173 case BoolTest::eq: 2174 // They are proven equal, so we can collapse the min/max. 2175 // Either value is the answer. Choose the simpler. 2176 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2177 return yvalue; 2178 return xvalue; 2179 case BoolTest::lt: // x < y 2180 case BoolTest::le: // x <= y 2181 return (want_max ? yvalue : xvalue); 2182 case BoolTest::gt: // x > y 2183 case BoolTest::ge: // x >= y 2184 return (want_max ? xvalue : yvalue); 2185 } 2186 } 2187 } 2188 2189 // We failed to find a dominating test. 2190 // Let's pick a test that might GVN with prior tests. 2191 Node* best_bol = NULL; 2192 BoolTest::mask best_btest = BoolTest::illegal; 2193 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2194 Node* cmp = cmps[cmpn]; 2195 if (cmp == NULL) continue; 2196 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2197 Node* bol = cmp->fast_out(j); 2198 if (!bol->is_Bool()) continue; 2199 BoolTest::mask btest = bol->as_Bool()->_test._test; 2200 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2201 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2202 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2203 best_bol = bol->as_Bool(); 2204 best_btest = btest; 2205 } 2206 } 2207 } 2208 2209 Node* answer_if_true = NULL; 2210 Node* answer_if_false = NULL; 2211 switch (best_btest) { 2212 default: 2213 if (cmpxy == NULL) 2214 cmpxy = ideal_cmpxy; 2215 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2216 // and fall through: 2217 case BoolTest::lt: // x < y 2218 case BoolTest::le: // x <= y 2219 answer_if_true = (want_max ? yvalue : xvalue); 2220 answer_if_false = (want_max ? xvalue : yvalue); 2221 break; 2222 case BoolTest::gt: // x > y 2223 case BoolTest::ge: // x >= y 2224 answer_if_true = (want_max ? xvalue : yvalue); 2225 answer_if_false = (want_max ? yvalue : xvalue); 2226 break; 2227 } 2228 2229 jint hi, lo; 2230 if (want_max) { 2231 // We can sharpen the minimum. 2232 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2233 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2234 } else { 2235 // We can sharpen the maximum. 2236 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2237 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2238 } 2239 2240 // Use a flow-free graph structure, to avoid creating excess control edges 2241 // which could hinder other optimizations. 2242 // Since Math.min/max is often used with arraycopy, we want 2243 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2244 Node* cmov = CMoveNode::make(NULL, best_bol, 2245 answer_if_false, answer_if_true, 2246 TypeInt::make(lo, hi, widen)); 2247 2248 return _gvn.transform(cmov); 2249 2250 /* 2251 // This is not as desirable as it may seem, since Min and Max 2252 // nodes do not have a full set of optimizations. 2253 // And they would interfere, anyway, with 'if' optimizations 2254 // and with CMoveI canonical forms. 2255 switch (id) { 2256 case vmIntrinsics::_min: 2257 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2258 case vmIntrinsics::_max: 2259 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2260 default: 2261 ShouldNotReachHere(); 2262 } 2263 */ 2264 } 2265 2266 inline int 2267 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2268 const TypePtr* base_type = TypePtr::NULL_PTR; 2269 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2270 if (base_type == NULL) { 2271 // Unknown type. 2272 return Type::AnyPtr; 2273 } else if (base_type == TypePtr::NULL_PTR) { 2274 // Since this is a NULL+long form, we have to switch to a rawptr. 2275 base = _gvn.transform(new CastX2PNode(offset)); 2276 offset = MakeConX(0); 2277 return Type::RawPtr; 2278 } else if (base_type->base() == Type::RawPtr) { 2279 return Type::RawPtr; 2280 } else if (base_type->isa_oopptr()) { 2281 // Base is never null => always a heap address. 2282 if (base_type->ptr() == TypePtr::NotNull) { 2283 return Type::OopPtr; 2284 } 2285 // Offset is small => always a heap address. 2286 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2287 if (offset_type != NULL && 2288 base_type->offset() == 0 && // (should always be?) 2289 offset_type->_lo >= 0 && 2290 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2291 return Type::OopPtr; 2292 } 2293 // Otherwise, it might either be oop+off or NULL+addr. 2294 return Type::AnyPtr; 2295 } else { 2296 // No information: 2297 return Type::AnyPtr; 2298 } 2299 } 2300 2301 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2302 int kind = classify_unsafe_addr(base, offset); 2303 if (kind == Type::RawPtr) { 2304 return basic_plus_adr(top(), base, offset); 2305 } else { 2306 return basic_plus_adr(base, offset); 2307 } 2308 } 2309 2310 //--------------------------inline_number_methods----------------------------- 2311 // inline int Integer.numberOfLeadingZeros(int) 2312 // inline int Long.numberOfLeadingZeros(long) 2313 // 2314 // inline int Integer.numberOfTrailingZeros(int) 2315 // inline int Long.numberOfTrailingZeros(long) 2316 // 2317 // inline int Integer.bitCount(int) 2318 // inline int Long.bitCount(long) 2319 // 2320 // inline char Character.reverseBytes(char) 2321 // inline short Short.reverseBytes(short) 2322 // inline int Integer.reverseBytes(int) 2323 // inline long Long.reverseBytes(long) 2324 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2325 Node* arg = argument(0); 2326 Node* n; 2327 switch (id) { 2328 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2329 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2330 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2331 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2332 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2333 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2334 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2335 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2336 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2337 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2338 default: fatal_unexpected_iid(id); break; 2339 } 2340 set_result(_gvn.transform(n)); 2341 return true; 2342 } 2343 2344 //----------------------------inline_unsafe_access---------------------------- 2345 2346 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2347 2348 // Helper that guards and inserts a pre-barrier. 2349 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2350 Node* pre_val, bool need_mem_bar) { 2351 // We could be accessing the referent field of a reference object. If so, when G1 2352 // is enabled, we need to log the value in the referent field in an SATB buffer. 2353 // This routine performs some compile time filters and generates suitable 2354 // runtime filters that guard the pre-barrier code. 2355 // Also add memory barrier for non volatile load from the referent field 2356 // to prevent commoning of loads across safepoint. 2357 if (!UseG1GC && !need_mem_bar) 2358 return; 2359 2360 // Some compile time checks. 2361 2362 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2363 const TypeX* otype = offset->find_intptr_t_type(); 2364 if (otype != NULL && otype->is_con() && 2365 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2366 // Constant offset but not the reference_offset so just return 2367 return; 2368 } 2369 2370 // We only need to generate the runtime guards for instances. 2371 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2372 if (btype != NULL) { 2373 if (btype->isa_aryptr()) { 2374 // Array type so nothing to do 2375 return; 2376 } 2377 2378 const TypeInstPtr* itype = btype->isa_instptr(); 2379 if (itype != NULL) { 2380 // Can the klass of base_oop be statically determined to be 2381 // _not_ a sub-class of Reference and _not_ Object? 2382 ciKlass* klass = itype->klass(); 2383 if ( klass->is_loaded() && 2384 !klass->is_subtype_of(env()->Reference_klass()) && 2385 !env()->Object_klass()->is_subtype_of(klass)) { 2386 return; 2387 } 2388 } 2389 } 2390 2391 // The compile time filters did not reject base_oop/offset so 2392 // we need to generate the following runtime filters 2393 // 2394 // if (offset == java_lang_ref_Reference::_reference_offset) { 2395 // if (instance_of(base, java.lang.ref.Reference)) { 2396 // pre_barrier(_, pre_val, ...); 2397 // } 2398 // } 2399 2400 float likely = PROB_LIKELY( 0.999); 2401 float unlikely = PROB_UNLIKELY(0.999); 2402 2403 IdealKit ideal(this); 2404 #define __ ideal. 2405 2406 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2407 2408 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2409 // Update graphKit memory and control from IdealKit. 2410 sync_kit(ideal); 2411 2412 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2413 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2414 2415 // Update IdealKit memory and control from graphKit. 2416 __ sync_kit(this); 2417 2418 Node* one = __ ConI(1); 2419 // is_instof == 0 if base_oop == NULL 2420 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2421 2422 // Update graphKit from IdeakKit. 2423 sync_kit(ideal); 2424 2425 // Use the pre-barrier to record the value in the referent field 2426 pre_barrier(false /* do_load */, 2427 __ ctrl(), 2428 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2429 pre_val /* pre_val */, 2430 T_OBJECT); 2431 if (need_mem_bar) { 2432 // Add memory barrier to prevent commoning reads from this field 2433 // across safepoint since GC can change its value. 2434 insert_mem_bar(Op_MemBarCPUOrder); 2435 } 2436 // Update IdealKit from graphKit. 2437 __ sync_kit(this); 2438 2439 } __ end_if(); // _ref_type != ref_none 2440 } __ end_if(); // offset == referent_offset 2441 2442 // Final sync IdealKit and GraphKit. 2443 final_sync(ideal); 2444 #undef __ 2445 } 2446 2447 2448 // Interpret Unsafe.fieldOffset cookies correctly: 2449 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2450 2451 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2452 // Attempt to infer a sharper value type from the offset and base type. 2453 ciKlass* sharpened_klass = NULL; 2454 2455 // See if it is an instance field, with an object type. 2456 if (alias_type->field() != NULL) { 2457 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2458 if (alias_type->field()->type()->is_klass()) { 2459 sharpened_klass = alias_type->field()->type()->as_klass(); 2460 } 2461 } 2462 2463 // See if it is a narrow oop array. 2464 if (adr_type->isa_aryptr()) { 2465 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2466 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2467 if (elem_type != NULL) { 2468 sharpened_klass = elem_type->klass(); 2469 } 2470 } 2471 } 2472 2473 // The sharpened class might be unloaded if there is no class loader 2474 // contraint in place. 2475 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2476 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2477 2478 #ifndef PRODUCT 2479 if (C->print_intrinsics() || C->print_inlining()) { 2480 tty->print(" from base type: "); adr_type->dump(); 2481 tty->print(" sharpened value: "); tjp->dump(); 2482 } 2483 #endif 2484 // Sharpen the value type. 2485 return tjp; 2486 } 2487 return NULL; 2488 } 2489 2490 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2491 if (callee()->is_static()) return false; // caller must have the capability! 2492 2493 #ifndef PRODUCT 2494 { 2495 ResourceMark rm; 2496 // Check the signatures. 2497 ciSignature* sig = callee()->signature(); 2498 #ifdef ASSERT 2499 if (!is_store) { 2500 // Object getObject(Object base, int/long offset), etc. 2501 BasicType rtype = sig->return_type()->basic_type(); 2502 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2503 rtype = T_ADDRESS; // it is really a C void* 2504 assert(rtype == type, "getter must return the expected value"); 2505 if (!is_native_ptr) { 2506 assert(sig->count() == 2, "oop getter has 2 arguments"); 2507 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2508 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2509 } else { 2510 assert(sig->count() == 1, "native getter has 1 argument"); 2511 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2512 } 2513 } else { 2514 // void putObject(Object base, int/long offset, Object x), etc. 2515 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2516 if (!is_native_ptr) { 2517 assert(sig->count() == 3, "oop putter has 3 arguments"); 2518 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2519 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2520 } else { 2521 assert(sig->count() == 2, "native putter has 2 arguments"); 2522 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2523 } 2524 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2525 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2526 vtype = T_ADDRESS; // it is really a C void* 2527 assert(vtype == type, "putter must accept the expected value"); 2528 } 2529 #endif // ASSERT 2530 } 2531 #endif //PRODUCT 2532 2533 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2534 2535 Node* receiver = argument(0); // type: oop 2536 2537 // Build address expression. 2538 Node* adr; 2539 Node* heap_base_oop = top(); 2540 Node* offset = top(); 2541 Node* val; 2542 2543 if (!is_native_ptr) { 2544 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2545 Node* base = argument(1); // type: oop 2546 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2547 offset = argument(2); // type: long 2548 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2549 // to be plain byte offsets, which are also the same as those accepted 2550 // by oopDesc::field_base. 2551 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2552 "fieldOffset must be byte-scaled"); 2553 // 32-bit machines ignore the high half! 2554 offset = ConvL2X(offset); 2555 adr = make_unsafe_address(base, offset); 2556 heap_base_oop = base; 2557 val = is_store ? argument(4) : NULL; 2558 } else { 2559 Node* ptr = argument(1); // type: long 2560 ptr = ConvL2X(ptr); // adjust Java long to machine word 2561 adr = make_unsafe_address(NULL, ptr); 2562 val = is_store ? argument(3) : NULL; 2563 } 2564 2565 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2566 2567 // First guess at the value type. 2568 const Type *value_type = Type::get_const_basic_type(type); 2569 2570 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2571 // there was not enough information to nail it down. 2572 Compile::AliasType* alias_type = C->alias_type(adr_type); 2573 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2574 2575 // We will need memory barriers unless we can determine a unique 2576 // alias category for this reference. (Note: If for some reason 2577 // the barriers get omitted and the unsafe reference begins to "pollute" 2578 // the alias analysis of the rest of the graph, either Compile::can_alias 2579 // or Compile::must_alias will throw a diagnostic assert.) 2580 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2581 2582 // If we are reading the value of the referent field of a Reference 2583 // object (either by using Unsafe directly or through reflection) 2584 // then, if G1 is enabled, we need to record the referent in an 2585 // SATB log buffer using the pre-barrier mechanism. 2586 // Also we need to add memory barrier to prevent commoning reads 2587 // from this field across safepoint since GC can change its value. 2588 bool need_read_barrier = !is_native_ptr && !is_store && 2589 offset != top() && heap_base_oop != top(); 2590 2591 if (!is_store && type == T_OBJECT) { 2592 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2593 if (tjp != NULL) { 2594 value_type = tjp; 2595 } 2596 } 2597 2598 receiver = null_check(receiver); 2599 if (stopped()) { 2600 return true; 2601 } 2602 // Heap pointers get a null-check from the interpreter, 2603 // as a courtesy. However, this is not guaranteed by Unsafe, 2604 // and it is not possible to fully distinguish unintended nulls 2605 // from intended ones in this API. 2606 2607 if (is_volatile) { 2608 // We need to emit leading and trailing CPU membars (see below) in 2609 // addition to memory membars when is_volatile. This is a little 2610 // too strong, but avoids the need to insert per-alias-type 2611 // volatile membars (for stores; compare Parse::do_put_xxx), which 2612 // we cannot do effectively here because we probably only have a 2613 // rough approximation of type. 2614 need_mem_bar = true; 2615 // For Stores, place a memory ordering barrier now. 2616 if (is_store) { 2617 insert_mem_bar(Op_MemBarRelease); 2618 } else { 2619 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2620 insert_mem_bar(Op_MemBarVolatile); 2621 } 2622 } 2623 } 2624 2625 // Memory barrier to prevent normal and 'unsafe' accesses from 2626 // bypassing each other. Happens after null checks, so the 2627 // exception paths do not take memory state from the memory barrier, 2628 // so there's no problems making a strong assert about mixing users 2629 // of safe & unsafe memory. 2630 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2631 2632 if (!is_store) { 2633 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2634 // To be valid, unsafe loads may depend on other conditions than 2635 // the one that guards them: pin the Load node 2636 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile); 2637 // load value 2638 switch (type) { 2639 case T_BOOLEAN: 2640 case T_CHAR: 2641 case T_BYTE: 2642 case T_SHORT: 2643 case T_INT: 2644 case T_LONG: 2645 case T_FLOAT: 2646 case T_DOUBLE: 2647 break; 2648 case T_OBJECT: 2649 if (need_read_barrier) { 2650 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2651 } 2652 break; 2653 case T_ADDRESS: 2654 // Cast to an int type. 2655 p = _gvn.transform(new CastP2XNode(NULL, p)); 2656 p = ConvX2UL(p); 2657 break; 2658 default: 2659 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2660 break; 2661 } 2662 // The load node has the control of the preceding MemBarCPUOrder. All 2663 // following nodes will have the control of the MemBarCPUOrder inserted at 2664 // the end of this method. So, pushing the load onto the stack at a later 2665 // point is fine. 2666 set_result(p); 2667 } else { 2668 // place effect of store into memory 2669 switch (type) { 2670 case T_DOUBLE: 2671 val = dstore_rounding(val); 2672 break; 2673 case T_ADDRESS: 2674 // Repackage the long as a pointer. 2675 val = ConvL2X(val); 2676 val = _gvn.transform(new CastX2PNode(val)); 2677 break; 2678 } 2679 2680 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2681 if (type != T_OBJECT ) { 2682 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2683 } else { 2684 // Possibly an oop being stored to Java heap or native memory 2685 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2686 // oop to Java heap. 2687 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2688 } else { 2689 // We can't tell at compile time if we are storing in the Java heap or outside 2690 // of it. So we need to emit code to conditionally do the proper type of 2691 // store. 2692 2693 IdealKit ideal(this); 2694 #define __ ideal. 2695 // QQQ who knows what probability is here?? 2696 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2697 // Sync IdealKit and graphKit. 2698 sync_kit(ideal); 2699 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2700 // Update IdealKit memory. 2701 __ sync_kit(this); 2702 } __ else_(); { 2703 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2704 } __ end_if(); 2705 // Final sync IdealKit and GraphKit. 2706 final_sync(ideal); 2707 #undef __ 2708 } 2709 } 2710 } 2711 2712 if (is_volatile) { 2713 if (!is_store) { 2714 insert_mem_bar(Op_MemBarAcquire); 2715 } else { 2716 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2717 insert_mem_bar(Op_MemBarVolatile); 2718 } 2719 } 2720 } 2721 2722 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2723 2724 return true; 2725 } 2726 2727 //----------------------------inline_unsafe_load_store---------------------------- 2728 // This method serves a couple of different customers (depending on LoadStoreKind): 2729 // 2730 // LS_cmpxchg: 2731 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2732 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2733 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2734 // 2735 // LS_xadd: 2736 // public int getAndAddInt( Object o, long offset, int delta) 2737 // public long getAndAddLong(Object o, long offset, long delta) 2738 // 2739 // LS_xchg: 2740 // int getAndSet(Object o, long offset, int newValue) 2741 // long getAndSet(Object o, long offset, long newValue) 2742 // Object getAndSet(Object o, long offset, Object newValue) 2743 // 2744 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2745 // This basic scheme here is the same as inline_unsafe_access, but 2746 // differs in enough details that combining them would make the code 2747 // overly confusing. (This is a true fact! I originally combined 2748 // them, but even I was confused by it!) As much code/comments as 2749 // possible are retained from inline_unsafe_access though to make 2750 // the correspondences clearer. - dl 2751 2752 if (callee()->is_static()) return false; // caller must have the capability! 2753 2754 #ifndef PRODUCT 2755 BasicType rtype; 2756 { 2757 ResourceMark rm; 2758 // Check the signatures. 2759 ciSignature* sig = callee()->signature(); 2760 rtype = sig->return_type()->basic_type(); 2761 if (kind == LS_xadd || kind == LS_xchg) { 2762 // Check the signatures. 2763 #ifdef ASSERT 2764 assert(rtype == type, "get and set must return the expected type"); 2765 assert(sig->count() == 3, "get and set has 3 arguments"); 2766 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2767 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2768 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2769 #endif // ASSERT 2770 } else if (kind == LS_cmpxchg) { 2771 // Check the signatures. 2772 #ifdef ASSERT 2773 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2774 assert(sig->count() == 4, "CAS has 4 arguments"); 2775 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2776 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2777 #endif // ASSERT 2778 } else { 2779 ShouldNotReachHere(); 2780 } 2781 } 2782 #endif //PRODUCT 2783 2784 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2785 2786 // Get arguments: 2787 Node* receiver = NULL; 2788 Node* base = NULL; 2789 Node* offset = NULL; 2790 Node* oldval = NULL; 2791 Node* newval = NULL; 2792 if (kind == LS_cmpxchg) { 2793 const bool two_slot_type = type2size[type] == 2; 2794 receiver = argument(0); // type: oop 2795 base = argument(1); // type: oop 2796 offset = argument(2); // type: long 2797 oldval = argument(4); // type: oop, int, or long 2798 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2799 } else if (kind == LS_xadd || kind == LS_xchg){ 2800 receiver = argument(0); // type: oop 2801 base = argument(1); // type: oop 2802 offset = argument(2); // type: long 2803 oldval = NULL; 2804 newval = argument(4); // type: oop, int, or long 2805 } 2806 2807 // Null check receiver. 2808 receiver = null_check(receiver); 2809 if (stopped()) { 2810 return true; 2811 } 2812 2813 // Build field offset expression. 2814 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2815 // to be plain byte offsets, which are also the same as those accepted 2816 // by oopDesc::field_base. 2817 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2818 // 32-bit machines ignore the high half of long offsets 2819 offset = ConvL2X(offset); 2820 Node* adr = make_unsafe_address(base, offset); 2821 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2822 2823 // For CAS, unlike inline_unsafe_access, there seems no point in 2824 // trying to refine types. Just use the coarse types here. 2825 const Type *value_type = Type::get_const_basic_type(type); 2826 Compile::AliasType* alias_type = C->alias_type(adr_type); 2827 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2828 2829 if (kind == LS_xchg && type == T_OBJECT) { 2830 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2831 if (tjp != NULL) { 2832 value_type = tjp; 2833 } 2834 } 2835 2836 int alias_idx = C->get_alias_index(adr_type); 2837 2838 // Memory-model-wise, a LoadStore acts like a little synchronized 2839 // block, so needs barriers on each side. These don't translate 2840 // into actual barriers on most machines, but we still need rest of 2841 // compiler to respect ordering. 2842 2843 insert_mem_bar(Op_MemBarRelease); 2844 insert_mem_bar(Op_MemBarCPUOrder); 2845 2846 // 4984716: MemBars must be inserted before this 2847 // memory node in order to avoid a false 2848 // dependency which will confuse the scheduler. 2849 Node *mem = memory(alias_idx); 2850 2851 // For now, we handle only those cases that actually exist: ints, 2852 // longs, and Object. Adding others should be straightforward. 2853 Node* load_store; 2854 switch(type) { 2855 case T_INT: 2856 if (kind == LS_xadd) { 2857 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2858 } else if (kind == LS_xchg) { 2859 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2860 } else if (kind == LS_cmpxchg) { 2861 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2862 } else { 2863 ShouldNotReachHere(); 2864 } 2865 break; 2866 case T_LONG: 2867 if (kind == LS_xadd) { 2868 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2869 } else if (kind == LS_xchg) { 2870 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2871 } else if (kind == LS_cmpxchg) { 2872 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2873 } else { 2874 ShouldNotReachHere(); 2875 } 2876 break; 2877 case T_OBJECT: 2878 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2879 // could be delayed during Parse (for example, in adjust_map_after_if()). 2880 // Execute transformation here to avoid barrier generation in such case. 2881 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2882 newval = _gvn.makecon(TypePtr::NULL_PTR); 2883 2884 // Reference stores need a store barrier. 2885 if (kind == LS_xchg) { 2886 // If pre-barrier must execute before the oop store, old value will require do_load here. 2887 if (!can_move_pre_barrier()) { 2888 pre_barrier(true /* do_load*/, 2889 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2890 NULL /* pre_val*/, 2891 T_OBJECT); 2892 } // Else move pre_barrier to use load_store value, see below. 2893 } else if (kind == LS_cmpxchg) { 2894 // Same as for newval above: 2895 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2896 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2897 } 2898 // The only known value which might get overwritten is oldval. 2899 pre_barrier(false /* do_load */, 2900 control(), NULL, NULL, max_juint, NULL, NULL, 2901 oldval /* pre_val */, 2902 T_OBJECT); 2903 } else { 2904 ShouldNotReachHere(); 2905 } 2906 2907 #ifdef _LP64 2908 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2909 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2910 if (kind == LS_xchg) { 2911 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2912 newval_enc, adr_type, value_type->make_narrowoop())); 2913 } else { 2914 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2915 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2916 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2917 newval_enc, oldval_enc)); 2918 } 2919 } else 2920 #endif 2921 { 2922 if (kind == LS_xchg) { 2923 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2924 } else { 2925 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2926 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2927 } 2928 } 2929 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2930 break; 2931 default: 2932 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2933 break; 2934 } 2935 2936 // SCMemProjNodes represent the memory state of a LoadStore. Their 2937 // main role is to prevent LoadStore nodes from being optimized away 2938 // when their results aren't used. 2939 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2940 set_memory(proj, alias_idx); 2941 2942 if (type == T_OBJECT && kind == LS_xchg) { 2943 #ifdef _LP64 2944 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2945 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2946 } 2947 #endif 2948 if (can_move_pre_barrier()) { 2949 // Don't need to load pre_val. The old value is returned by load_store. 2950 // The pre_barrier can execute after the xchg as long as no safepoint 2951 // gets inserted between them. 2952 pre_barrier(false /* do_load */, 2953 control(), NULL, NULL, max_juint, NULL, NULL, 2954 load_store /* pre_val */, 2955 T_OBJECT); 2956 } 2957 } 2958 2959 // Add the trailing membar surrounding the access 2960 insert_mem_bar(Op_MemBarCPUOrder); 2961 insert_mem_bar(Op_MemBarAcquire); 2962 2963 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2964 set_result(load_store); 2965 return true; 2966 } 2967 2968 //----------------------------inline_unsafe_ordered_store---------------------- 2969 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2970 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 2971 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 2972 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2973 // This is another variant of inline_unsafe_access, differing in 2974 // that it always issues store-store ("release") barrier and ensures 2975 // store-atomicity (which only matters for "long"). 2976 2977 if (callee()->is_static()) return false; // caller must have the capability! 2978 2979 #ifndef PRODUCT 2980 { 2981 ResourceMark rm; 2982 // Check the signatures. 2983 ciSignature* sig = callee()->signature(); 2984 #ifdef ASSERT 2985 BasicType rtype = sig->return_type()->basic_type(); 2986 assert(rtype == T_VOID, "must return void"); 2987 assert(sig->count() == 3, "has 3 arguments"); 2988 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2989 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2990 #endif // ASSERT 2991 } 2992 #endif //PRODUCT 2993 2994 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2995 2996 // Get arguments: 2997 Node* receiver = argument(0); // type: oop 2998 Node* base = argument(1); // type: oop 2999 Node* offset = argument(2); // type: long 3000 Node* val = argument(4); // type: oop, int, or long 3001 3002 // Null check receiver. 3003 receiver = null_check(receiver); 3004 if (stopped()) { 3005 return true; 3006 } 3007 3008 // Build field offset expression. 3009 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3010 // 32-bit machines ignore the high half of long offsets 3011 offset = ConvL2X(offset); 3012 Node* adr = make_unsafe_address(base, offset); 3013 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3014 const Type *value_type = Type::get_const_basic_type(type); 3015 Compile::AliasType* alias_type = C->alias_type(adr_type); 3016 3017 insert_mem_bar(Op_MemBarRelease); 3018 insert_mem_bar(Op_MemBarCPUOrder); 3019 // Ensure that the store is atomic for longs: 3020 const bool require_atomic_access = true; 3021 Node* store; 3022 if (type == T_OBJECT) // reference stores need a store barrier. 3023 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3024 else { 3025 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3026 } 3027 insert_mem_bar(Op_MemBarCPUOrder); 3028 return true; 3029 } 3030 3031 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3032 // Regardless of form, don't allow previous ld/st to move down, 3033 // then issue acquire, release, or volatile mem_bar. 3034 insert_mem_bar(Op_MemBarCPUOrder); 3035 switch(id) { 3036 case vmIntrinsics::_loadFence: 3037 insert_mem_bar(Op_LoadFence); 3038 return true; 3039 case vmIntrinsics::_storeFence: 3040 insert_mem_bar(Op_StoreFence); 3041 return true; 3042 case vmIntrinsics::_fullFence: 3043 insert_mem_bar(Op_MemBarVolatile); 3044 return true; 3045 default: 3046 fatal_unexpected_iid(id); 3047 return false; 3048 } 3049 } 3050 3051 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3052 if (!kls->is_Con()) { 3053 return true; 3054 } 3055 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3056 if (klsptr == NULL) { 3057 return true; 3058 } 3059 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3060 // don't need a guard for a klass that is already initialized 3061 return !ik->is_initialized(); 3062 } 3063 3064 //----------------------------inline_unsafe_allocate--------------------------- 3065 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3066 bool LibraryCallKit::inline_unsafe_allocate() { 3067 if (callee()->is_static()) return false; // caller must have the capability! 3068 3069 null_check_receiver(); // null-check, then ignore 3070 Node* cls = null_check(argument(1)); 3071 if (stopped()) return true; 3072 3073 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3074 kls = null_check(kls); 3075 if (stopped()) return true; // argument was like int.class 3076 3077 Node* test = NULL; 3078 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3079 // Note: The argument might still be an illegal value like 3080 // Serializable.class or Object[].class. The runtime will handle it. 3081 // But we must make an explicit check for initialization. 3082 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3083 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3084 // can generate code to load it as unsigned byte. 3085 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3086 Node* bits = intcon(InstanceKlass::fully_initialized); 3087 test = _gvn.transform(new SubINode(inst, bits)); 3088 // The 'test' is non-zero if we need to take a slow path. 3089 } 3090 3091 Node* obj = new_instance(kls, test); 3092 set_result(obj); 3093 return true; 3094 } 3095 3096 #ifdef TRACE_HAVE_INTRINSICS 3097 /* 3098 * oop -> myklass 3099 * myklass->trace_id |= USED 3100 * return myklass->trace_id & ~0x3 3101 */ 3102 bool LibraryCallKit::inline_native_classID() { 3103 null_check_receiver(); // null-check, then ignore 3104 Node* cls = null_check(argument(1), T_OBJECT); 3105 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3106 kls = null_check(kls, T_OBJECT); 3107 ByteSize offset = TRACE_ID_OFFSET; 3108 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3109 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3110 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3111 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3112 Node* clsused = longcon(0x01l); // set the class bit 3113 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3114 3115 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3116 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3117 set_result(andl); 3118 return true; 3119 } 3120 3121 bool LibraryCallKit::inline_native_threadID() { 3122 Node* tls_ptr = NULL; 3123 Node* cur_thr = generate_current_thread(tls_ptr); 3124 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3125 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3126 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3127 3128 Node* threadid = NULL; 3129 size_t thread_id_size = OSThread::thread_id_size(); 3130 if (thread_id_size == (size_t) BytesPerLong) { 3131 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3132 } else if (thread_id_size == (size_t) BytesPerInt) { 3133 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3134 } else { 3135 ShouldNotReachHere(); 3136 } 3137 set_result(threadid); 3138 return true; 3139 } 3140 #endif 3141 3142 //------------------------inline_native_time_funcs-------------- 3143 // inline code for System.currentTimeMillis() and System.nanoTime() 3144 // these have the same type and signature 3145 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3146 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3147 const TypePtr* no_memory_effects = NULL; 3148 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3149 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3150 #ifdef ASSERT 3151 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3152 assert(value_top == top(), "second value must be top"); 3153 #endif 3154 set_result(value); 3155 return true; 3156 } 3157 3158 //------------------------inline_native_currentThread------------------ 3159 bool LibraryCallKit::inline_native_currentThread() { 3160 Node* junk = NULL; 3161 set_result(generate_current_thread(junk)); 3162 return true; 3163 } 3164 3165 //------------------------inline_native_isInterrupted------------------ 3166 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3167 bool LibraryCallKit::inline_native_isInterrupted() { 3168 // Add a fast path to t.isInterrupted(clear_int): 3169 // (t == Thread.current() && 3170 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3171 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3172 // So, in the common case that the interrupt bit is false, 3173 // we avoid making a call into the VM. Even if the interrupt bit 3174 // is true, if the clear_int argument is false, we avoid the VM call. 3175 // However, if the receiver is not currentThread, we must call the VM, 3176 // because there must be some locking done around the operation. 3177 3178 // We only go to the fast case code if we pass two guards. 3179 // Paths which do not pass are accumulated in the slow_region. 3180 3181 enum { 3182 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3183 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3184 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3185 PATH_LIMIT 3186 }; 3187 3188 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3189 // out of the function. 3190 insert_mem_bar(Op_MemBarCPUOrder); 3191 3192 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3193 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3194 3195 RegionNode* slow_region = new RegionNode(1); 3196 record_for_igvn(slow_region); 3197 3198 // (a) Receiving thread must be the current thread. 3199 Node* rec_thr = argument(0); 3200 Node* tls_ptr = NULL; 3201 Node* cur_thr = generate_current_thread(tls_ptr); 3202 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3203 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3204 3205 generate_slow_guard(bol_thr, slow_region); 3206 3207 // (b) Interrupt bit on TLS must be false. 3208 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3209 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3210 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3211 3212 // Set the control input on the field _interrupted read to prevent it floating up. 3213 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3214 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3215 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3216 3217 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3218 3219 // First fast path: if (!TLS._interrupted) return false; 3220 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3221 result_rgn->init_req(no_int_result_path, false_bit); 3222 result_val->init_req(no_int_result_path, intcon(0)); 3223 3224 // drop through to next case 3225 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3226 3227 #ifndef TARGET_OS_FAMILY_windows 3228 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3229 Node* clr_arg = argument(1); 3230 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3231 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3232 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3233 3234 // Second fast path: ... else if (!clear_int) return true; 3235 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3236 result_rgn->init_req(no_clear_result_path, false_arg); 3237 result_val->init_req(no_clear_result_path, intcon(1)); 3238 3239 // drop through to next case 3240 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3241 #else 3242 // To return true on Windows you must read the _interrupted field 3243 // and check the the event state i.e. take the slow path. 3244 #endif // TARGET_OS_FAMILY_windows 3245 3246 // (d) Otherwise, go to the slow path. 3247 slow_region->add_req(control()); 3248 set_control( _gvn.transform(slow_region)); 3249 3250 if (stopped()) { 3251 // There is no slow path. 3252 result_rgn->init_req(slow_result_path, top()); 3253 result_val->init_req(slow_result_path, top()); 3254 } else { 3255 // non-virtual because it is a private non-static 3256 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3257 3258 Node* slow_val = set_results_for_java_call(slow_call); 3259 // this->control() comes from set_results_for_java_call 3260 3261 Node* fast_io = slow_call->in(TypeFunc::I_O); 3262 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3263 3264 // These two phis are pre-filled with copies of of the fast IO and Memory 3265 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3266 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3267 3268 result_rgn->init_req(slow_result_path, control()); 3269 result_io ->init_req(slow_result_path, i_o()); 3270 result_mem->init_req(slow_result_path, reset_memory()); 3271 result_val->init_req(slow_result_path, slow_val); 3272 3273 set_all_memory(_gvn.transform(result_mem)); 3274 set_i_o( _gvn.transform(result_io)); 3275 } 3276 3277 C->set_has_split_ifs(true); // Has chance for split-if optimization 3278 set_result(result_rgn, result_val); 3279 return true; 3280 } 3281 3282 //---------------------------load_mirror_from_klass---------------------------- 3283 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3284 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3285 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3286 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3287 } 3288 3289 //-----------------------load_klass_from_mirror_common------------------------- 3290 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3291 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3292 // and branch to the given path on the region. 3293 // If never_see_null, take an uncommon trap on null, so we can optimistically 3294 // compile for the non-null case. 3295 // If the region is NULL, force never_see_null = true. 3296 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3297 bool never_see_null, 3298 RegionNode* region, 3299 int null_path, 3300 int offset) { 3301 if (region == NULL) never_see_null = true; 3302 Node* p = basic_plus_adr(mirror, offset); 3303 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3304 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3305 Node* null_ctl = top(); 3306 kls = null_check_oop(kls, &null_ctl, never_see_null); 3307 if (region != NULL) { 3308 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3309 region->init_req(null_path, null_ctl); 3310 } else { 3311 assert(null_ctl == top(), "no loose ends"); 3312 } 3313 return kls; 3314 } 3315 3316 //--------------------(inline_native_Class_query helpers)--------------------- 3317 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3318 // Fall through if (mods & mask) == bits, take the guard otherwise. 3319 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3320 // Branch around if the given klass has the given modifier bit set. 3321 // Like generate_guard, adds a new path onto the region. 3322 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3323 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3324 Node* mask = intcon(modifier_mask); 3325 Node* bits = intcon(modifier_bits); 3326 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3327 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3328 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3329 return generate_fair_guard(bol, region); 3330 } 3331 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3332 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3333 } 3334 3335 //-------------------------inline_native_Class_query------------------- 3336 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3337 const Type* return_type = TypeInt::BOOL; 3338 Node* prim_return_value = top(); // what happens if it's a primitive class? 3339 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3340 bool expect_prim = false; // most of these guys expect to work on refs 3341 3342 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3343 3344 Node* mirror = argument(0); 3345 Node* obj = top(); 3346 3347 switch (id) { 3348 case vmIntrinsics::_isInstance: 3349 // nothing is an instance of a primitive type 3350 prim_return_value = intcon(0); 3351 obj = argument(1); 3352 break; 3353 case vmIntrinsics::_getModifiers: 3354 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3355 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3356 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3357 break; 3358 case vmIntrinsics::_isInterface: 3359 prim_return_value = intcon(0); 3360 break; 3361 case vmIntrinsics::_isArray: 3362 prim_return_value = intcon(0); 3363 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3364 break; 3365 case vmIntrinsics::_isPrimitive: 3366 prim_return_value = intcon(1); 3367 expect_prim = true; // obviously 3368 break; 3369 case vmIntrinsics::_getSuperclass: 3370 prim_return_value = null(); 3371 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3372 break; 3373 case vmIntrinsics::_getClassAccessFlags: 3374 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3375 return_type = TypeInt::INT; // not bool! 6297094 3376 break; 3377 default: 3378 fatal_unexpected_iid(id); 3379 break; 3380 } 3381 3382 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3383 if (mirror_con == NULL) return false; // cannot happen? 3384 3385 #ifndef PRODUCT 3386 if (C->print_intrinsics() || C->print_inlining()) { 3387 ciType* k = mirror_con->java_mirror_type(); 3388 if (k) { 3389 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3390 k->print_name(); 3391 tty->cr(); 3392 } 3393 } 3394 #endif 3395 3396 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3397 RegionNode* region = new RegionNode(PATH_LIMIT); 3398 record_for_igvn(region); 3399 PhiNode* phi = new PhiNode(region, return_type); 3400 3401 // The mirror will never be null of Reflection.getClassAccessFlags, however 3402 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3403 // if it is. See bug 4774291. 3404 3405 // For Reflection.getClassAccessFlags(), the null check occurs in 3406 // the wrong place; see inline_unsafe_access(), above, for a similar 3407 // situation. 3408 mirror = null_check(mirror); 3409 // If mirror or obj is dead, only null-path is taken. 3410 if (stopped()) return true; 3411 3412 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3413 3414 // Now load the mirror's klass metaobject, and null-check it. 3415 // Side-effects region with the control path if the klass is null. 3416 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3417 // If kls is null, we have a primitive mirror. 3418 phi->init_req(_prim_path, prim_return_value); 3419 if (stopped()) { set_result(region, phi); return true; } 3420 bool safe_for_replace = (region->in(_prim_path) == top()); 3421 3422 Node* p; // handy temp 3423 Node* null_ctl; 3424 3425 // Now that we have the non-null klass, we can perform the real query. 3426 // For constant classes, the query will constant-fold in LoadNode::Value. 3427 Node* query_value = top(); 3428 switch (id) { 3429 case vmIntrinsics::_isInstance: 3430 // nothing is an instance of a primitive type 3431 query_value = gen_instanceof(obj, kls, safe_for_replace); 3432 break; 3433 3434 case vmIntrinsics::_getModifiers: 3435 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3436 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3437 break; 3438 3439 case vmIntrinsics::_isInterface: 3440 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3441 if (generate_interface_guard(kls, region) != NULL) 3442 // A guard was added. If the guard is taken, it was an interface. 3443 phi->add_req(intcon(1)); 3444 // If we fall through, it's a plain class. 3445 query_value = intcon(0); 3446 break; 3447 3448 case vmIntrinsics::_isArray: 3449 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3450 if (generate_array_guard(kls, region) != NULL) 3451 // A guard was added. If the guard is taken, it was an array. 3452 phi->add_req(intcon(1)); 3453 // If we fall through, it's a plain class. 3454 query_value = intcon(0); 3455 break; 3456 3457 case vmIntrinsics::_isPrimitive: 3458 query_value = intcon(0); // "normal" path produces false 3459 break; 3460 3461 case vmIntrinsics::_getSuperclass: 3462 // The rules here are somewhat unfortunate, but we can still do better 3463 // with random logic than with a JNI call. 3464 // Interfaces store null or Object as _super, but must report null. 3465 // Arrays store an intermediate super as _super, but must report Object. 3466 // Other types can report the actual _super. 3467 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3468 if (generate_interface_guard(kls, region) != NULL) 3469 // A guard was added. If the guard is taken, it was an interface. 3470 phi->add_req(null()); 3471 if (generate_array_guard(kls, region) != NULL) 3472 // A guard was added. If the guard is taken, it was an array. 3473 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3474 // If we fall through, it's a plain class. Get its _super. 3475 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3476 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3477 null_ctl = top(); 3478 kls = null_check_oop(kls, &null_ctl); 3479 if (null_ctl != top()) { 3480 // If the guard is taken, Object.superClass is null (both klass and mirror). 3481 region->add_req(null_ctl); 3482 phi ->add_req(null()); 3483 } 3484 if (!stopped()) { 3485 query_value = load_mirror_from_klass(kls); 3486 } 3487 break; 3488 3489 case vmIntrinsics::_getClassAccessFlags: 3490 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3491 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3492 break; 3493 3494 default: 3495 fatal_unexpected_iid(id); 3496 break; 3497 } 3498 3499 // Fall-through is the normal case of a query to a real class. 3500 phi->init_req(1, query_value); 3501 region->init_req(1, control()); 3502 3503 C->set_has_split_ifs(true); // Has chance for split-if optimization 3504 set_result(region, phi); 3505 return true; 3506 } 3507 3508 //-------------------------inline_Class_cast------------------- 3509 bool LibraryCallKit::inline_Class_cast() { 3510 Node* mirror = argument(0); // Class 3511 Node* obj = argument(1); 3512 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3513 if (mirror_con == NULL) { 3514 return false; // dead path (mirror->is_top()). 3515 } 3516 if (obj == NULL || obj->is_top()) { 3517 return false; // dead path 3518 } 3519 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3520 3521 // First, see if Class.cast() can be folded statically. 3522 // java_mirror_type() returns non-null for compile-time Class constants. 3523 ciType* tm = mirror_con->java_mirror_type(); 3524 if (tm != NULL && tm->is_klass() && 3525 tp != NULL && tp->klass() != NULL) { 3526 if (!tp->klass()->is_loaded()) { 3527 // Don't use intrinsic when class is not loaded. 3528 return false; 3529 } else { 3530 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3531 if (static_res == Compile::SSC_always_true) { 3532 // isInstance() is true - fold the code. 3533 set_result(obj); 3534 return true; 3535 } else if (static_res == Compile::SSC_always_false) { 3536 // Don't use intrinsic, have to throw ClassCastException. 3537 // If the reference is null, the non-intrinsic bytecode will 3538 // be optimized appropriately. 3539 return false; 3540 } 3541 } 3542 } 3543 3544 // Bailout intrinsic and do normal inlining if exception path is frequent. 3545 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3546 return false; 3547 } 3548 3549 // Generate dynamic checks. 3550 // Class.cast() is java implementation of _checkcast bytecode. 3551 // Do checkcast (Parse::do_checkcast()) optimizations here. 3552 3553 mirror = null_check(mirror); 3554 // If mirror is dead, only null-path is taken. 3555 if (stopped()) { 3556 return true; 3557 } 3558 3559 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3560 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3561 RegionNode* region = new RegionNode(PATH_LIMIT); 3562 record_for_igvn(region); 3563 3564 // Now load the mirror's klass metaobject, and null-check it. 3565 // If kls is null, we have a primitive mirror and 3566 // nothing is an instance of a primitive type. 3567 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3568 3569 Node* res = top(); 3570 if (!stopped()) { 3571 Node* bad_type_ctrl = top(); 3572 // Do checkcast optimizations. 3573 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3574 region->init_req(_bad_type_path, bad_type_ctrl); 3575 } 3576 if (region->in(_prim_path) != top() || 3577 region->in(_bad_type_path) != top()) { 3578 // Let Interpreter throw ClassCastException. 3579 PreserveJVMState pjvms(this); 3580 set_control(_gvn.transform(region)); 3581 uncommon_trap(Deoptimization::Reason_intrinsic, 3582 Deoptimization::Action_maybe_recompile); 3583 } 3584 if (!stopped()) { 3585 set_result(res); 3586 } 3587 return true; 3588 } 3589 3590 3591 //--------------------------inline_native_subtype_check------------------------ 3592 // This intrinsic takes the JNI calls out of the heart of 3593 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3594 bool LibraryCallKit::inline_native_subtype_check() { 3595 // Pull both arguments off the stack. 3596 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3597 args[0] = argument(0); 3598 args[1] = argument(1); 3599 Node* klasses[2]; // corresponding Klasses: superk, subk 3600 klasses[0] = klasses[1] = top(); 3601 3602 enum { 3603 // A full decision tree on {superc is prim, subc is prim}: 3604 _prim_0_path = 1, // {P,N} => false 3605 // {P,P} & superc!=subc => false 3606 _prim_same_path, // {P,P} & superc==subc => true 3607 _prim_1_path, // {N,P} => false 3608 _ref_subtype_path, // {N,N} & subtype check wins => true 3609 _both_ref_path, // {N,N} & subtype check loses => false 3610 PATH_LIMIT 3611 }; 3612 3613 RegionNode* region = new RegionNode(PATH_LIMIT); 3614 Node* phi = new PhiNode(region, TypeInt::BOOL); 3615 record_for_igvn(region); 3616 3617 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3618 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3619 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3620 3621 // First null-check both mirrors and load each mirror's klass metaobject. 3622 int which_arg; 3623 for (which_arg = 0; which_arg <= 1; which_arg++) { 3624 Node* arg = args[which_arg]; 3625 arg = null_check(arg); 3626 if (stopped()) break; 3627 args[which_arg] = arg; 3628 3629 Node* p = basic_plus_adr(arg, class_klass_offset); 3630 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3631 klasses[which_arg] = _gvn.transform(kls); 3632 } 3633 3634 // Having loaded both klasses, test each for null. 3635 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3636 for (which_arg = 0; which_arg <= 1; which_arg++) { 3637 Node* kls = klasses[which_arg]; 3638 Node* null_ctl = top(); 3639 kls = null_check_oop(kls, &null_ctl, never_see_null); 3640 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3641 region->init_req(prim_path, null_ctl); 3642 if (stopped()) break; 3643 klasses[which_arg] = kls; 3644 } 3645 3646 if (!stopped()) { 3647 // now we have two reference types, in klasses[0..1] 3648 Node* subk = klasses[1]; // the argument to isAssignableFrom 3649 Node* superk = klasses[0]; // the receiver 3650 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3651 // now we have a successful reference subtype check 3652 region->set_req(_ref_subtype_path, control()); 3653 } 3654 3655 // If both operands are primitive (both klasses null), then 3656 // we must return true when they are identical primitives. 3657 // It is convenient to test this after the first null klass check. 3658 set_control(region->in(_prim_0_path)); // go back to first null check 3659 if (!stopped()) { 3660 // Since superc is primitive, make a guard for the superc==subc case. 3661 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3662 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3663 generate_guard(bol_eq, region, PROB_FAIR); 3664 if (region->req() == PATH_LIMIT+1) { 3665 // A guard was added. If the added guard is taken, superc==subc. 3666 region->swap_edges(PATH_LIMIT, _prim_same_path); 3667 region->del_req(PATH_LIMIT); 3668 } 3669 region->set_req(_prim_0_path, control()); // Not equal after all. 3670 } 3671 3672 // these are the only paths that produce 'true': 3673 phi->set_req(_prim_same_path, intcon(1)); 3674 phi->set_req(_ref_subtype_path, intcon(1)); 3675 3676 // pull together the cases: 3677 assert(region->req() == PATH_LIMIT, "sane region"); 3678 for (uint i = 1; i < region->req(); i++) { 3679 Node* ctl = region->in(i); 3680 if (ctl == NULL || ctl == top()) { 3681 region->set_req(i, top()); 3682 phi ->set_req(i, top()); 3683 } else if (phi->in(i) == NULL) { 3684 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3685 } 3686 } 3687 3688 set_control(_gvn.transform(region)); 3689 set_result(_gvn.transform(phi)); 3690 return true; 3691 } 3692 3693 //---------------------generate_array_guard_common------------------------ 3694 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3695 bool obj_array, bool not_array) { 3696 3697 if (stopped()) { 3698 return NULL; 3699 } 3700 3701 // If obj_array/non_array==false/false: 3702 // Branch around if the given klass is in fact an array (either obj or prim). 3703 // If obj_array/non_array==false/true: 3704 // Branch around if the given klass is not an array klass of any kind. 3705 // If obj_array/non_array==true/true: 3706 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3707 // If obj_array/non_array==true/false: 3708 // Branch around if the kls is an oop array (Object[] or subtype) 3709 // 3710 // Like generate_guard, adds a new path onto the region. 3711 jint layout_con = 0; 3712 Node* layout_val = get_layout_helper(kls, layout_con); 3713 if (layout_val == NULL) { 3714 bool query = (obj_array 3715 ? Klass::layout_helper_is_objArray(layout_con) 3716 : Klass::layout_helper_is_array(layout_con)); 3717 if (query == not_array) { 3718 return NULL; // never a branch 3719 } else { // always a branch 3720 Node* always_branch = control(); 3721 if (region != NULL) 3722 region->add_req(always_branch); 3723 set_control(top()); 3724 return always_branch; 3725 } 3726 } 3727 // Now test the correct condition. 3728 jint nval = (obj_array 3729 ? ((jint)Klass::_lh_array_tag_type_value 3730 << Klass::_lh_array_tag_shift) 3731 : Klass::_lh_neutral_value); 3732 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3733 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3734 // invert the test if we are looking for a non-array 3735 if (not_array) btest = BoolTest(btest).negate(); 3736 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3737 return generate_fair_guard(bol, region); 3738 } 3739 3740 3741 //-----------------------inline_native_newArray-------------------------- 3742 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3743 bool LibraryCallKit::inline_native_newArray() { 3744 Node* mirror = argument(0); 3745 Node* count_val = argument(1); 3746 3747 mirror = null_check(mirror); 3748 // If mirror or obj is dead, only null-path is taken. 3749 if (stopped()) return true; 3750 3751 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3752 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3753 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3754 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3755 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3756 3757 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3758 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3759 result_reg, _slow_path); 3760 Node* normal_ctl = control(); 3761 Node* no_array_ctl = result_reg->in(_slow_path); 3762 3763 // Generate code for the slow case. We make a call to newArray(). 3764 set_control(no_array_ctl); 3765 if (!stopped()) { 3766 // Either the input type is void.class, or else the 3767 // array klass has not yet been cached. Either the 3768 // ensuing call will throw an exception, or else it 3769 // will cache the array klass for next time. 3770 PreserveJVMState pjvms(this); 3771 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3772 Node* slow_result = set_results_for_java_call(slow_call); 3773 // this->control() comes from set_results_for_java_call 3774 result_reg->set_req(_slow_path, control()); 3775 result_val->set_req(_slow_path, slow_result); 3776 result_io ->set_req(_slow_path, i_o()); 3777 result_mem->set_req(_slow_path, reset_memory()); 3778 } 3779 3780 set_control(normal_ctl); 3781 if (!stopped()) { 3782 // Normal case: The array type has been cached in the java.lang.Class. 3783 // The following call works fine even if the array type is polymorphic. 3784 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3785 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3786 result_reg->init_req(_normal_path, control()); 3787 result_val->init_req(_normal_path, obj); 3788 result_io ->init_req(_normal_path, i_o()); 3789 result_mem->init_req(_normal_path, reset_memory()); 3790 } 3791 3792 // Return the combined state. 3793 set_i_o( _gvn.transform(result_io) ); 3794 set_all_memory( _gvn.transform(result_mem)); 3795 3796 C->set_has_split_ifs(true); // Has chance for split-if optimization 3797 set_result(result_reg, result_val); 3798 return true; 3799 } 3800 3801 //----------------------inline_native_getLength-------------------------- 3802 // public static native int java.lang.reflect.Array.getLength(Object array); 3803 bool LibraryCallKit::inline_native_getLength() { 3804 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3805 3806 Node* array = null_check(argument(0)); 3807 // If array is dead, only null-path is taken. 3808 if (stopped()) return true; 3809 3810 // Deoptimize if it is a non-array. 3811 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3812 3813 if (non_array != NULL) { 3814 PreserveJVMState pjvms(this); 3815 set_control(non_array); 3816 uncommon_trap(Deoptimization::Reason_intrinsic, 3817 Deoptimization::Action_maybe_recompile); 3818 } 3819 3820 // If control is dead, only non-array-path is taken. 3821 if (stopped()) return true; 3822 3823 // The works fine even if the array type is polymorphic. 3824 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3825 Node* result = load_array_length(array); 3826 3827 C->set_has_split_ifs(true); // Has chance for split-if optimization 3828 set_result(result); 3829 return true; 3830 } 3831 3832 //------------------------inline_array_copyOf---------------------------- 3833 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3834 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3835 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3836 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3837 3838 // Get the arguments. 3839 Node* original = argument(0); 3840 Node* start = is_copyOfRange? argument(1): intcon(0); 3841 Node* end = is_copyOfRange? argument(2): argument(1); 3842 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3843 3844 Node* newcopy; 3845 3846 // Set the original stack and the reexecute bit for the interpreter to reexecute 3847 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3848 { PreserveReexecuteState preexecs(this); 3849 jvms()->set_should_reexecute(true); 3850 3851 array_type_mirror = null_check(array_type_mirror); 3852 original = null_check(original); 3853 3854 // Check if a null path was taken unconditionally. 3855 if (stopped()) return true; 3856 3857 Node* orig_length = load_array_length(original); 3858 3859 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3860 klass_node = null_check(klass_node); 3861 3862 RegionNode* bailout = new RegionNode(1); 3863 record_for_igvn(bailout); 3864 3865 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3866 // Bail out if that is so. 3867 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3868 if (not_objArray != NULL) { 3869 // Improve the klass node's type from the new optimistic assumption: 3870 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3871 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3872 Node* cast = new CastPPNode(klass_node, akls); 3873 cast->init_req(0, control()); 3874 klass_node = _gvn.transform(cast); 3875 } 3876 3877 // Bail out if either start or end is negative. 3878 generate_negative_guard(start, bailout, &start); 3879 generate_negative_guard(end, bailout, &end); 3880 3881 Node* length = end; 3882 if (_gvn.type(start) != TypeInt::ZERO) { 3883 length = _gvn.transform(new SubINode(end, start)); 3884 } 3885 3886 // Bail out if length is negative. 3887 // Without this the new_array would throw 3888 // NegativeArraySizeException but IllegalArgumentException is what 3889 // should be thrown 3890 generate_negative_guard(length, bailout, &length); 3891 3892 if (bailout->req() > 1) { 3893 PreserveJVMState pjvms(this); 3894 set_control(_gvn.transform(bailout)); 3895 uncommon_trap(Deoptimization::Reason_intrinsic, 3896 Deoptimization::Action_maybe_recompile); 3897 } 3898 3899 if (!stopped()) { 3900 // How many elements will we copy from the original? 3901 // The answer is MinI(orig_length - start, length). 3902 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3903 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3904 3905 // Generate a direct call to the right arraycopy function(s). 3906 // We know the copy is disjoint but we might not know if the 3907 // oop stores need checking. 3908 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3909 // This will fail a store-check if x contains any non-nulls. 3910 3911 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3912 // loads/stores but it is legal only if we're sure the 3913 // Arrays.copyOf would succeed. So we need all input arguments 3914 // to the copyOf to be validated, including that the copy to the 3915 // new array won't trigger an ArrayStoreException. That subtype 3916 // check can be optimized if we know something on the type of 3917 // the input array from type speculation. 3918 if (_gvn.type(klass_node)->singleton()) { 3919 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3920 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3921 3922 int test = C->static_subtype_check(superk, subk); 3923 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3924 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3925 if (t_original->speculative_type() != NULL) { 3926 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3927 } 3928 } 3929 } 3930 3931 bool validated = false; 3932 // Reason_class_check rather than Reason_intrinsic because we 3933 // want to intrinsify even if this traps. 3934 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3935 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3936 klass_node); 3937 3938 if (not_subtype_ctrl != top()) { 3939 PreserveJVMState pjvms(this); 3940 set_control(not_subtype_ctrl); 3941 uncommon_trap(Deoptimization::Reason_class_check, 3942 Deoptimization::Action_make_not_entrant); 3943 assert(stopped(), "Should be stopped"); 3944 } 3945 validated = true; 3946 } 3947 3948 if (!stopped()) { 3949 newcopy = new_array(klass_node, length, 0); // no arguments to push 3950 3951 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3952 load_object_klass(original), klass_node); 3953 if (!is_copyOfRange) { 3954 ac->set_copyof(validated); 3955 } else { 3956 ac->set_copyofrange(validated); 3957 } 3958 Node* n = _gvn.transform(ac); 3959 if (n == ac) { 3960 ac->connect_outputs(this); 3961 } else { 3962 assert(validated, "shouldn't transform if all arguments not validated"); 3963 set_all_memory(n); 3964 } 3965 } 3966 } 3967 } // original reexecute is set back here 3968 3969 C->set_has_split_ifs(true); // Has chance for split-if optimization 3970 if (!stopped()) { 3971 set_result(newcopy); 3972 } 3973 return true; 3974 } 3975 3976 3977 //----------------------generate_virtual_guard--------------------------- 3978 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3979 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3980 RegionNode* slow_region) { 3981 ciMethod* method = callee(); 3982 int vtable_index = method->vtable_index(); 3983 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3984 err_msg_res("bad index %d", vtable_index)); 3985 // Get the Method* out of the appropriate vtable entry. 3986 int entry_offset = (InstanceKlass::vtable_start_offset() + 3987 vtable_index*vtableEntry::size()) * wordSize + 3988 vtableEntry::method_offset_in_bytes(); 3989 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3990 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3991 3992 // Compare the target method with the expected method (e.g., Object.hashCode). 3993 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3994 3995 Node* native_call = makecon(native_call_addr); 3996 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3997 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3998 3999 return generate_slow_guard(test_native, slow_region); 4000 } 4001 4002 //-----------------------generate_method_call---------------------------- 4003 // Use generate_method_call to make a slow-call to the real 4004 // method if the fast path fails. An alternative would be to 4005 // use a stub like OptoRuntime::slow_arraycopy_Java. 4006 // This only works for expanding the current library call, 4007 // not another intrinsic. (E.g., don't use this for making an 4008 // arraycopy call inside of the copyOf intrinsic.) 4009 CallJavaNode* 4010 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4011 // When compiling the intrinsic method itself, do not use this technique. 4012 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4013 4014 ciMethod* method = callee(); 4015 // ensure the JVMS we have will be correct for this call 4016 guarantee(method_id == method->intrinsic_id(), "must match"); 4017 4018 const TypeFunc* tf = TypeFunc::make(method); 4019 CallJavaNode* slow_call; 4020 if (is_static) { 4021 assert(!is_virtual, ""); 4022 slow_call = new CallStaticJavaNode(C, tf, 4023 SharedRuntime::get_resolve_static_call_stub(), 4024 method, bci()); 4025 } else if (is_virtual) { 4026 null_check_receiver(); 4027 int vtable_index = Method::invalid_vtable_index; 4028 if (UseInlineCaches) { 4029 // Suppress the vtable call 4030 } else { 4031 // hashCode and clone are not a miranda methods, 4032 // so the vtable index is fixed. 4033 // No need to use the linkResolver to get it. 4034 vtable_index = method->vtable_index(); 4035 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4036 err_msg_res("bad index %d", vtable_index)); 4037 } 4038 slow_call = new CallDynamicJavaNode(tf, 4039 SharedRuntime::get_resolve_virtual_call_stub(), 4040 method, vtable_index, bci()); 4041 } else { // neither virtual nor static: opt_virtual 4042 null_check_receiver(); 4043 slow_call = new CallStaticJavaNode(C, tf, 4044 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4045 method, bci()); 4046 slow_call->set_optimized_virtual(true); 4047 } 4048 set_arguments_for_java_call(slow_call); 4049 set_edges_for_java_call(slow_call); 4050 return slow_call; 4051 } 4052 4053 4054 /** 4055 * Build special case code for calls to hashCode on an object. This call may 4056 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4057 * slightly different code. 4058 */ 4059 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4060 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4061 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4062 4063 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4064 4065 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4066 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4067 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4068 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4069 Node* obj = NULL; 4070 if (!is_static) { 4071 // Check for hashing null object 4072 obj = null_check_receiver(); 4073 if (stopped()) return true; // unconditionally null 4074 result_reg->init_req(_null_path, top()); 4075 result_val->init_req(_null_path, top()); 4076 } else { 4077 // Do a null check, and return zero if null. 4078 // System.identityHashCode(null) == 0 4079 obj = argument(0); 4080 Node* null_ctl = top(); 4081 obj = null_check_oop(obj, &null_ctl); 4082 result_reg->init_req(_null_path, null_ctl); 4083 result_val->init_req(_null_path, _gvn.intcon(0)); 4084 } 4085 4086 // Unconditionally null? Then return right away. 4087 if (stopped()) { 4088 set_control( result_reg->in(_null_path)); 4089 if (!stopped()) 4090 set_result(result_val->in(_null_path)); 4091 return true; 4092 } 4093 4094 // We only go to the fast case code if we pass a number of guards. The 4095 // paths which do not pass are accumulated in the slow_region. 4096 RegionNode* slow_region = new RegionNode(1); 4097 record_for_igvn(slow_region); 4098 4099 // If this is a virtual call, we generate a funny guard. We pull out 4100 // the vtable entry corresponding to hashCode() from the target object. 4101 // If the target method which we are calling happens to be the native 4102 // Object hashCode() method, we pass the guard. We do not need this 4103 // guard for non-virtual calls -- the caller is known to be the native 4104 // Object hashCode(). 4105 if (is_virtual) { 4106 // After null check, get the object's klass. 4107 Node* obj_klass = load_object_klass(obj); 4108 generate_virtual_guard(obj_klass, slow_region); 4109 } 4110 4111 // Get the header out of the object, use LoadMarkNode when available 4112 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4113 // The control of the load must be NULL. Otherwise, the load can move before 4114 // the null check after castPP removal. 4115 Node* no_ctrl = NULL; 4116 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4117 4118 // Test the header to see if it is unlocked. 4119 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4120 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4121 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4122 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4123 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4124 4125 generate_slow_guard(test_unlocked, slow_region); 4126 4127 // Get the hash value and check to see that it has been properly assigned. 4128 // We depend on hash_mask being at most 32 bits and avoid the use of 4129 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4130 // vm: see markOop.hpp. 4131 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4132 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4133 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4134 // This hack lets the hash bits live anywhere in the mark object now, as long 4135 // as the shift drops the relevant bits into the low 32 bits. Note that 4136 // Java spec says that HashCode is an int so there's no point in capturing 4137 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4138 hshifted_header = ConvX2I(hshifted_header); 4139 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4140 4141 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4142 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4143 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4144 4145 generate_slow_guard(test_assigned, slow_region); 4146 4147 Node* init_mem = reset_memory(); 4148 // fill in the rest of the null path: 4149 result_io ->init_req(_null_path, i_o()); 4150 result_mem->init_req(_null_path, init_mem); 4151 4152 result_val->init_req(_fast_path, hash_val); 4153 result_reg->init_req(_fast_path, control()); 4154 result_io ->init_req(_fast_path, i_o()); 4155 result_mem->init_req(_fast_path, init_mem); 4156 4157 // Generate code for the slow case. We make a call to hashCode(). 4158 set_control(_gvn.transform(slow_region)); 4159 if (!stopped()) { 4160 // No need for PreserveJVMState, because we're using up the present state. 4161 set_all_memory(init_mem); 4162 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4163 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4164 Node* slow_result = set_results_for_java_call(slow_call); 4165 // this->control() comes from set_results_for_java_call 4166 result_reg->init_req(_slow_path, control()); 4167 result_val->init_req(_slow_path, slow_result); 4168 result_io ->set_req(_slow_path, i_o()); 4169 result_mem ->set_req(_slow_path, reset_memory()); 4170 } 4171 4172 // Return the combined state. 4173 set_i_o( _gvn.transform(result_io) ); 4174 set_all_memory( _gvn.transform(result_mem)); 4175 4176 set_result(result_reg, result_val); 4177 return true; 4178 } 4179 4180 //---------------------------inline_native_getClass---------------------------- 4181 // public final native Class<?> java.lang.Object.getClass(); 4182 // 4183 // Build special case code for calls to getClass on an object. 4184 bool LibraryCallKit::inline_native_getClass() { 4185 Node* obj = null_check_receiver(); 4186 if (stopped()) return true; 4187 set_result(load_mirror_from_klass(load_object_klass(obj))); 4188 return true; 4189 } 4190 4191 //-----------------inline_native_Reflection_getCallerClass--------------------- 4192 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4193 // 4194 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4195 // 4196 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4197 // in that it must skip particular security frames and checks for 4198 // caller sensitive methods. 4199 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4200 #ifndef PRODUCT 4201 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4202 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4203 } 4204 #endif 4205 4206 if (!jvms()->has_method()) { 4207 #ifndef PRODUCT 4208 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4209 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4210 } 4211 #endif 4212 return false; 4213 } 4214 4215 // Walk back up the JVM state to find the caller at the required 4216 // depth. 4217 JVMState* caller_jvms = jvms(); 4218 4219 // Cf. JVM_GetCallerClass 4220 // NOTE: Start the loop at depth 1 because the current JVM state does 4221 // not include the Reflection.getCallerClass() frame. 4222 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4223 ciMethod* m = caller_jvms->method(); 4224 switch (n) { 4225 case 0: 4226 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4227 break; 4228 case 1: 4229 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4230 if (!m->caller_sensitive()) { 4231 #ifndef PRODUCT 4232 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4233 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4234 } 4235 #endif 4236 return false; // bail-out; let JVM_GetCallerClass do the work 4237 } 4238 break; 4239 default: 4240 if (!m->is_ignored_by_security_stack_walk()) { 4241 // We have reached the desired frame; return the holder class. 4242 // Acquire method holder as java.lang.Class and push as constant. 4243 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4244 ciInstance* caller_mirror = caller_klass->java_mirror(); 4245 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4246 4247 #ifndef PRODUCT 4248 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4249 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()); 4250 tty->print_cr(" JVM state at this point:"); 4251 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4252 ciMethod* m = jvms()->of_depth(i)->method(); 4253 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4254 } 4255 } 4256 #endif 4257 return true; 4258 } 4259 break; 4260 } 4261 } 4262 4263 #ifndef PRODUCT 4264 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4265 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4266 tty->print_cr(" JVM state at this point:"); 4267 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4268 ciMethod* m = jvms()->of_depth(i)->method(); 4269 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4270 } 4271 } 4272 #endif 4273 4274 return false; // bail-out; let JVM_GetCallerClass do the work 4275 } 4276 4277 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4278 Node* arg = argument(0); 4279 Node* result; 4280 4281 switch (id) { 4282 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4283 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4284 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4285 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4286 4287 case vmIntrinsics::_doubleToLongBits: { 4288 // two paths (plus control) merge in a wood 4289 RegionNode *r = new RegionNode(3); 4290 Node *phi = new PhiNode(r, TypeLong::LONG); 4291 4292 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4293 // Build the boolean node 4294 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4295 4296 // Branch either way. 4297 // NaN case is less traveled, which makes all the difference. 4298 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4299 Node *opt_isnan = _gvn.transform(ifisnan); 4300 assert( opt_isnan->is_If(), "Expect an IfNode"); 4301 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4302 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4303 4304 set_control(iftrue); 4305 4306 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4307 Node *slow_result = longcon(nan_bits); // return NaN 4308 phi->init_req(1, _gvn.transform( slow_result )); 4309 r->init_req(1, iftrue); 4310 4311 // Else fall through 4312 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4313 set_control(iffalse); 4314 4315 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4316 r->init_req(2, iffalse); 4317 4318 // Post merge 4319 set_control(_gvn.transform(r)); 4320 record_for_igvn(r); 4321 4322 C->set_has_split_ifs(true); // Has chance for split-if optimization 4323 result = phi; 4324 assert(result->bottom_type()->isa_long(), "must be"); 4325 break; 4326 } 4327 4328 case vmIntrinsics::_floatToIntBits: { 4329 // two paths (plus control) merge in a wood 4330 RegionNode *r = new RegionNode(3); 4331 Node *phi = new PhiNode(r, TypeInt::INT); 4332 4333 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4334 // Build the boolean node 4335 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4336 4337 // Branch either way. 4338 // NaN case is less traveled, which makes all the difference. 4339 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4340 Node *opt_isnan = _gvn.transform(ifisnan); 4341 assert( opt_isnan->is_If(), "Expect an IfNode"); 4342 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4343 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4344 4345 set_control(iftrue); 4346 4347 static const jint nan_bits = 0x7fc00000; 4348 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4349 phi->init_req(1, _gvn.transform( slow_result )); 4350 r->init_req(1, iftrue); 4351 4352 // Else fall through 4353 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4354 set_control(iffalse); 4355 4356 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4357 r->init_req(2, iffalse); 4358 4359 // Post merge 4360 set_control(_gvn.transform(r)); 4361 record_for_igvn(r); 4362 4363 C->set_has_split_ifs(true); // Has chance for split-if optimization 4364 result = phi; 4365 assert(result->bottom_type()->isa_int(), "must be"); 4366 break; 4367 } 4368 4369 default: 4370 fatal_unexpected_iid(id); 4371 break; 4372 } 4373 set_result(_gvn.transform(result)); 4374 return true; 4375 } 4376 4377 #ifdef _LP64 4378 #define XTOP ,top() /*additional argument*/ 4379 #else //_LP64 4380 #define XTOP /*no additional argument*/ 4381 #endif //_LP64 4382 4383 //----------------------inline_unsafe_copyMemory------------------------- 4384 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4385 bool LibraryCallKit::inline_unsafe_copyMemory() { 4386 if (callee()->is_static()) return false; // caller must have the capability! 4387 null_check_receiver(); // null-check receiver 4388 if (stopped()) return true; 4389 4390 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4391 4392 Node* src_ptr = argument(1); // type: oop 4393 Node* src_off = ConvL2X(argument(2)); // type: long 4394 Node* dst_ptr = argument(4); // type: oop 4395 Node* dst_off = ConvL2X(argument(5)); // type: long 4396 Node* size = ConvL2X(argument(7)); // type: long 4397 4398 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4399 "fieldOffset must be byte-scaled"); 4400 4401 Node* src = make_unsafe_address(src_ptr, src_off); 4402 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4403 4404 // Conservatively insert a memory barrier on all memory slices. 4405 // Do not let writes of the copy source or destination float below the copy. 4406 insert_mem_bar(Op_MemBarCPUOrder); 4407 4408 // Call it. Note that the length argument is not scaled. 4409 make_runtime_call(RC_LEAF|RC_NO_FP, 4410 OptoRuntime::fast_arraycopy_Type(), 4411 StubRoutines::unsafe_arraycopy(), 4412 "unsafe_arraycopy", 4413 TypeRawPtr::BOTTOM, 4414 src, dst, size XTOP); 4415 4416 // Do not let reads of the copy destination float above the copy. 4417 insert_mem_bar(Op_MemBarCPUOrder); 4418 4419 return true; 4420 } 4421 4422 //------------------------clone_coping----------------------------------- 4423 // Helper function for inline_native_clone. 4424 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4425 assert(obj_size != NULL, ""); 4426 Node* raw_obj = alloc_obj->in(1); 4427 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4428 4429 AllocateNode* alloc = NULL; 4430 if (ReduceBulkZeroing) { 4431 // We will be completely responsible for initializing this object - 4432 // mark Initialize node as complete. 4433 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4434 // The object was just allocated - there should be no any stores! 4435 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4436 // Mark as complete_with_arraycopy so that on AllocateNode 4437 // expansion, we know this AllocateNode is initialized by an array 4438 // copy and a StoreStore barrier exists after the array copy. 4439 alloc->initialization()->set_complete_with_arraycopy(); 4440 } 4441 4442 // Copy the fastest available way. 4443 // TODO: generate fields copies for small objects instead. 4444 Node* src = obj; 4445 Node* dest = alloc_obj; 4446 Node* size = _gvn.transform(obj_size); 4447 4448 // Exclude the header but include array length to copy by 8 bytes words. 4449 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4450 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4451 instanceOopDesc::base_offset_in_bytes(); 4452 // base_off: 4453 // 8 - 32-bit VM 4454 // 12 - 64-bit VM, compressed klass 4455 // 16 - 64-bit VM, normal klass 4456 if (base_off % BytesPerLong != 0) { 4457 assert(UseCompressedClassPointers, ""); 4458 if (is_array) { 4459 // Exclude length to copy by 8 bytes words. 4460 base_off += sizeof(int); 4461 } else { 4462 // Include klass to copy by 8 bytes words. 4463 base_off = instanceOopDesc::klass_offset_in_bytes(); 4464 } 4465 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4466 } 4467 src = basic_plus_adr(src, base_off); 4468 dest = basic_plus_adr(dest, base_off); 4469 4470 // Compute the length also, if needed: 4471 Node* countx = size; 4472 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4473 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4474 4475 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4476 4477 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4478 ac->set_clonebasic(); 4479 Node* n = _gvn.transform(ac); 4480 if (n == ac) { 4481 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4482 } else { 4483 set_all_memory(n); 4484 } 4485 4486 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4487 if (card_mark) { 4488 assert(!is_array, ""); 4489 // Put in store barrier for any and all oops we are sticking 4490 // into this object. (We could avoid this if we could prove 4491 // that the object type contains no oop fields at all.) 4492 Node* no_particular_value = NULL; 4493 Node* no_particular_field = NULL; 4494 int raw_adr_idx = Compile::AliasIdxRaw; 4495 post_barrier(control(), 4496 memory(raw_adr_type), 4497 alloc_obj, 4498 no_particular_field, 4499 raw_adr_idx, 4500 no_particular_value, 4501 T_OBJECT, 4502 false); 4503 } 4504 4505 // Do not let reads from the cloned object float above the arraycopy. 4506 if (alloc != NULL) { 4507 // Do not let stores that initialize this object be reordered with 4508 // a subsequent store that would make this object accessible by 4509 // other threads. 4510 // Record what AllocateNode this StoreStore protects so that 4511 // escape analysis can go from the MemBarStoreStoreNode to the 4512 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4513 // based on the escape status of the AllocateNode. 4514 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4515 } else { 4516 insert_mem_bar(Op_MemBarCPUOrder); 4517 } 4518 } 4519 4520 //------------------------inline_native_clone---------------------------- 4521 // protected native Object java.lang.Object.clone(); 4522 // 4523 // Here are the simple edge cases: 4524 // null receiver => normal trap 4525 // virtual and clone was overridden => slow path to out-of-line clone 4526 // not cloneable or finalizer => slow path to out-of-line Object.clone 4527 // 4528 // The general case has two steps, allocation and copying. 4529 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4530 // 4531 // Copying also has two cases, oop arrays and everything else. 4532 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4533 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4534 // 4535 // These steps fold up nicely if and when the cloned object's klass 4536 // can be sharply typed as an object array, a type array, or an instance. 4537 // 4538 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4539 PhiNode* result_val; 4540 4541 // Set the reexecute bit for the interpreter to reexecute 4542 // the bytecode that invokes Object.clone if deoptimization happens. 4543 { PreserveReexecuteState preexecs(this); 4544 jvms()->set_should_reexecute(true); 4545 4546 Node* obj = null_check_receiver(); 4547 if (stopped()) return true; 4548 4549 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4550 4551 // If we are going to clone an instance, we need its exact type to 4552 // know the number and types of fields to convert the clone to 4553 // loads/stores. Maybe a speculative type can help us. 4554 if (!obj_type->klass_is_exact() && 4555 obj_type->speculative_type() != NULL && 4556 obj_type->speculative_type()->is_instance_klass()) { 4557 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4558 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4559 !spec_ik->has_injected_fields()) { 4560 ciKlass* k = obj_type->klass(); 4561 if (!k->is_instance_klass() || 4562 k->as_instance_klass()->is_interface() || 4563 k->as_instance_klass()->has_subklass()) { 4564 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4565 } 4566 } 4567 } 4568 4569 Node* obj_klass = load_object_klass(obj); 4570 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4571 const TypeOopPtr* toop = ((tklass != NULL) 4572 ? tklass->as_instance_type() 4573 : TypeInstPtr::NOTNULL); 4574 4575 // Conservatively insert a memory barrier on all memory slices. 4576 // Do not let writes into the original float below the clone. 4577 insert_mem_bar(Op_MemBarCPUOrder); 4578 4579 // paths into result_reg: 4580 enum { 4581 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4582 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4583 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4584 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4585 PATH_LIMIT 4586 }; 4587 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4588 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4589 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4590 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4591 record_for_igvn(result_reg); 4592 4593 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4594 int raw_adr_idx = Compile::AliasIdxRaw; 4595 4596 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4597 if (array_ctl != NULL) { 4598 // It's an array. 4599 PreserveJVMState pjvms(this); 4600 set_control(array_ctl); 4601 Node* obj_length = load_array_length(obj); 4602 Node* obj_size = NULL; 4603 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4604 4605 if (!use_ReduceInitialCardMarks()) { 4606 // If it is an oop array, it requires very special treatment, 4607 // because card marking is required on each card of the array. 4608 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4609 if (is_obja != NULL) { 4610 PreserveJVMState pjvms2(this); 4611 set_control(is_obja); 4612 // Generate a direct call to the right arraycopy function(s). 4613 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4614 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4615 ac->set_cloneoop(); 4616 Node* n = _gvn.transform(ac); 4617 assert(n == ac, "cannot disappear"); 4618 ac->connect_outputs(this); 4619 4620 result_reg->init_req(_objArray_path, control()); 4621 result_val->init_req(_objArray_path, alloc_obj); 4622 result_i_o ->set_req(_objArray_path, i_o()); 4623 result_mem ->set_req(_objArray_path, reset_memory()); 4624 } 4625 } 4626 // Otherwise, there are no card marks to worry about. 4627 // (We can dispense with card marks if we know the allocation 4628 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4629 // causes the non-eden paths to take compensating steps to 4630 // simulate a fresh allocation, so that no further 4631 // card marks are required in compiled code to initialize 4632 // the object.) 4633 4634 if (!stopped()) { 4635 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4636 4637 // Present the results of the copy. 4638 result_reg->init_req(_array_path, control()); 4639 result_val->init_req(_array_path, alloc_obj); 4640 result_i_o ->set_req(_array_path, i_o()); 4641 result_mem ->set_req(_array_path, reset_memory()); 4642 } 4643 } 4644 4645 // We only go to the instance fast case code if we pass a number of guards. 4646 // The paths which do not pass are accumulated in the slow_region. 4647 RegionNode* slow_region = new RegionNode(1); 4648 record_for_igvn(slow_region); 4649 if (!stopped()) { 4650 // It's an instance (we did array above). Make the slow-path tests. 4651 // If this is a virtual call, we generate a funny guard. We grab 4652 // the vtable entry corresponding to clone() from the target object. 4653 // If the target method which we are calling happens to be the 4654 // Object clone() method, we pass the guard. We do not need this 4655 // guard for non-virtual calls; the caller is known to be the native 4656 // Object clone(). 4657 if (is_virtual) { 4658 generate_virtual_guard(obj_klass, slow_region); 4659 } 4660 4661 // The object must be cloneable and must not have a finalizer. 4662 // Both of these conditions may be checked in a single test. 4663 // We could optimize the cloneable test further, but we don't care. 4664 generate_access_flags_guard(obj_klass, 4665 // Test both conditions: 4666 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4667 // Must be cloneable but not finalizer: 4668 JVM_ACC_IS_CLONEABLE, 4669 slow_region); 4670 } 4671 4672 if (!stopped()) { 4673 // It's an instance, and it passed the slow-path tests. 4674 PreserveJVMState pjvms(this); 4675 Node* obj_size = NULL; 4676 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4677 // is reexecuted if deoptimization occurs and there could be problems when merging 4678 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4679 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4680 4681 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4682 4683 // Present the results of the slow call. 4684 result_reg->init_req(_instance_path, control()); 4685 result_val->init_req(_instance_path, alloc_obj); 4686 result_i_o ->set_req(_instance_path, i_o()); 4687 result_mem ->set_req(_instance_path, reset_memory()); 4688 } 4689 4690 // Generate code for the slow case. We make a call to clone(). 4691 set_control(_gvn.transform(slow_region)); 4692 if (!stopped()) { 4693 PreserveJVMState pjvms(this); 4694 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4695 Node* slow_result = set_results_for_java_call(slow_call); 4696 // this->control() comes from set_results_for_java_call 4697 result_reg->init_req(_slow_path, control()); 4698 result_val->init_req(_slow_path, slow_result); 4699 result_i_o ->set_req(_slow_path, i_o()); 4700 result_mem ->set_req(_slow_path, reset_memory()); 4701 } 4702 4703 // Return the combined state. 4704 set_control( _gvn.transform(result_reg)); 4705 set_i_o( _gvn.transform(result_i_o)); 4706 set_all_memory( _gvn.transform(result_mem)); 4707 } // original reexecute is set back here 4708 4709 set_result(_gvn.transform(result_val)); 4710 return true; 4711 } 4712 4713 // If we have a tighly coupled allocation, the arraycopy may take care 4714 // of the array initialization. If one of the guards we insert between 4715 // the allocation and the arraycopy causes a deoptimization, an 4716 // unitialized array will escape the compiled method. To prevent that 4717 // we set the JVM state for uncommon traps between the allocation and 4718 // the arraycopy to the state before the allocation so, in case of 4719 // deoptimization, we'll reexecute the allocation and the 4720 // initialization. 4721 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4722 if (alloc != NULL) { 4723 ciMethod* trap_method = alloc->jvms()->method(); 4724 int trap_bci = alloc->jvms()->bci(); 4725 4726 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4727 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4728 // Make sure there's no store between the allocation and the 4729 // arraycopy otherwise visible side effects could be rexecuted 4730 // in case of deoptimization and cause incorrect execution. 4731 bool no_interfering_store = true; 4732 Node* mem = alloc->in(TypeFunc::Memory); 4733 if (mem->is_MergeMem()) { 4734 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4735 Node* n = mms.memory(); 4736 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4737 assert(n->is_Store(), "what else?"); 4738 no_interfering_store = false; 4739 break; 4740 } 4741 } 4742 } else { 4743 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4744 Node* n = mms.memory(); 4745 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4746 assert(n->is_Store(), "what else?"); 4747 no_interfering_store = false; 4748 break; 4749 } 4750 } 4751 } 4752 4753 if (no_interfering_store) { 4754 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4755 uint size = alloc->req(); 4756 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4757 old_jvms->set_map(sfpt); 4758 for (uint i = 0; i < size; i++) { 4759 sfpt->init_req(i, alloc->in(i)); 4760 } 4761 // re-push array length for deoptimization 4762 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4763 old_jvms->set_sp(old_jvms->sp()+1); 4764 old_jvms->set_monoff(old_jvms->monoff()+1); 4765 old_jvms->set_scloff(old_jvms->scloff()+1); 4766 old_jvms->set_endoff(old_jvms->endoff()+1); 4767 old_jvms->set_should_reexecute(true); 4768 4769 sfpt->set_i_o(map()->i_o()); 4770 sfpt->set_memory(map()->memory()); 4771 sfpt->set_control(map()->control()); 4772 4773 JVMState* saved_jvms = jvms(); 4774 saved_reexecute_sp = _reexecute_sp; 4775 4776 set_jvms(sfpt->jvms()); 4777 _reexecute_sp = jvms()->sp(); 4778 4779 return saved_jvms; 4780 } 4781 } 4782 } 4783 return NULL; 4784 } 4785 4786 // In case of a deoptimization, we restart execution at the 4787 // allocation, allocating a new array. We would leave an uninitialized 4788 // array in the heap that GCs wouldn't expect. Move the allocation 4789 // after the traps so we don't allocate the array if we 4790 // deoptimize. This is possible because tightly_coupled_allocation() 4791 // guarantees there's no observer of the allocated array at this point 4792 // and the control flow is simple enough. 4793 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4794 if (saved_jvms != NULL && !stopped()) { 4795 assert(alloc != NULL, "only with a tightly coupled allocation"); 4796 // restore JVM state to the state at the arraycopy 4797 saved_jvms->map()->set_control(map()->control()); 4798 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4799 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4800 // If we've improved the types of some nodes (null check) while 4801 // emitting the guards, propagate them to the current state 4802 map()->replaced_nodes().apply(saved_jvms->map()); 4803 set_jvms(saved_jvms); 4804 _reexecute_sp = saved_reexecute_sp; 4805 4806 // Remove the allocation from above the guards 4807 CallProjections callprojs; 4808 alloc->extract_projections(&callprojs, true); 4809 InitializeNode* init = alloc->initialization(); 4810 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4811 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4812 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4813 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4814 4815 // move the allocation here (after the guards) 4816 _gvn.hash_delete(alloc); 4817 alloc->set_req(TypeFunc::Control, control()); 4818 alloc->set_req(TypeFunc::I_O, i_o()); 4819 Node *mem = reset_memory(); 4820 set_all_memory(mem); 4821 alloc->set_req(TypeFunc::Memory, mem); 4822 set_control(init->proj_out(TypeFunc::Control)); 4823 set_i_o(callprojs.fallthrough_ioproj); 4824 4825 // Update memory as done in GraphKit::set_output_for_allocation() 4826 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4827 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4828 if (ary_type->isa_aryptr() && length_type != NULL) { 4829 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4830 } 4831 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4832 int elemidx = C->get_alias_index(telemref); 4833 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4834 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4835 4836 Node* allocx = _gvn.transform(alloc); 4837 assert(allocx == alloc, "where has the allocation gone?"); 4838 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4839 4840 _gvn.hash_delete(dest); 4841 dest->set_req(0, control()); 4842 Node* destx = _gvn.transform(dest); 4843 assert(destx == dest, "where has the allocation result gone?"); 4844 } 4845 } 4846 4847 4848 //------------------------------inline_arraycopy----------------------- 4849 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4850 // Object dest, int destPos, 4851 // int length); 4852 bool LibraryCallKit::inline_arraycopy() { 4853 // Get the arguments. 4854 Node* src = argument(0); // type: oop 4855 Node* src_offset = argument(1); // type: int 4856 Node* dest = argument(2); // type: oop 4857 Node* dest_offset = argument(3); // type: int 4858 Node* length = argument(4); // type: int 4859 4860 4861 // Check for allocation before we add nodes that would confuse 4862 // tightly_coupled_allocation() 4863 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4864 4865 int saved_reexecute_sp = -1; 4866 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4867 // See arraycopy_restore_alloc_state() comment 4868 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4869 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4870 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4871 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4872 4873 // The following tests must be performed 4874 // (1) src and dest are arrays. 4875 // (2) src and dest arrays must have elements of the same BasicType 4876 // (3) src and dest must not be null. 4877 // (4) src_offset must not be negative. 4878 // (5) dest_offset must not be negative. 4879 // (6) length must not be negative. 4880 // (7) src_offset + length must not exceed length of src. 4881 // (8) dest_offset + length must not exceed length of dest. 4882 // (9) each element of an oop array must be assignable 4883 4884 // (3) src and dest must not be null. 4885 // always do this here because we need the JVM state for uncommon traps 4886 Node* null_ctl = top(); 4887 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4888 assert(null_ctl->is_top(), "no null control here"); 4889 dest = null_check(dest, T_ARRAY); 4890 4891 if (!can_emit_guards) { 4892 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4893 // guards but the arraycopy node could still take advantage of a 4894 // tightly allocated allocation. tightly_coupled_allocation() is 4895 // called again to make sure it takes the null check above into 4896 // account: the null check is mandatory and if it caused an 4897 // uncommon trap to be emitted then the allocation can't be 4898 // considered tightly coupled in this context. 4899 alloc = tightly_coupled_allocation(dest, NULL); 4900 } 4901 4902 bool validated = false; 4903 4904 const Type* src_type = _gvn.type(src); 4905 const Type* dest_type = _gvn.type(dest); 4906 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4907 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4908 4909 // Do we have the type of src? 4910 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4911 // Do we have the type of dest? 4912 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4913 // Is the type for src from speculation? 4914 bool src_spec = false; 4915 // Is the type for dest from speculation? 4916 bool dest_spec = false; 4917 4918 if ((!has_src || !has_dest) && can_emit_guards) { 4919 // We don't have sufficient type information, let's see if 4920 // speculative types can help. We need to have types for both src 4921 // and dest so that it pays off. 4922 4923 // Do we already have or could we have type information for src 4924 bool could_have_src = has_src; 4925 // Do we already have or could we have type information for dest 4926 bool could_have_dest = has_dest; 4927 4928 ciKlass* src_k = NULL; 4929 if (!has_src) { 4930 src_k = src_type->speculative_type_not_null(); 4931 if (src_k != NULL && src_k->is_array_klass()) { 4932 could_have_src = true; 4933 } 4934 } 4935 4936 ciKlass* dest_k = NULL; 4937 if (!has_dest) { 4938 dest_k = dest_type->speculative_type_not_null(); 4939 if (dest_k != NULL && dest_k->is_array_klass()) { 4940 could_have_dest = true; 4941 } 4942 } 4943 4944 if (could_have_src && could_have_dest) { 4945 // This is going to pay off so emit the required guards 4946 if (!has_src) { 4947 src = maybe_cast_profiled_obj(src, src_k, true); 4948 src_type = _gvn.type(src); 4949 top_src = src_type->isa_aryptr(); 4950 has_src = (top_src != NULL && top_src->klass() != NULL); 4951 src_spec = true; 4952 } 4953 if (!has_dest) { 4954 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4955 dest_type = _gvn.type(dest); 4956 top_dest = dest_type->isa_aryptr(); 4957 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4958 dest_spec = true; 4959 } 4960 } 4961 } 4962 4963 if (has_src && has_dest && can_emit_guards) { 4964 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4965 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4966 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4967 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4968 4969 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4970 // If both arrays are object arrays then having the exact types 4971 // for both will remove the need for a subtype check at runtime 4972 // before the call and may make it possible to pick a faster copy 4973 // routine (without a subtype check on every element) 4974 // Do we have the exact type of src? 4975 bool could_have_src = src_spec; 4976 // Do we have the exact type of dest? 4977 bool could_have_dest = dest_spec; 4978 ciKlass* src_k = top_src->klass(); 4979 ciKlass* dest_k = top_dest->klass(); 4980 if (!src_spec) { 4981 src_k = src_type->speculative_type_not_null(); 4982 if (src_k != NULL && src_k->is_array_klass()) { 4983 could_have_src = true; 4984 } 4985 } 4986 if (!dest_spec) { 4987 dest_k = dest_type->speculative_type_not_null(); 4988 if (dest_k != NULL && dest_k->is_array_klass()) { 4989 could_have_dest = true; 4990 } 4991 } 4992 if (could_have_src && could_have_dest) { 4993 // If we can have both exact types, emit the missing guards 4994 if (could_have_src && !src_spec) { 4995 src = maybe_cast_profiled_obj(src, src_k, true); 4996 } 4997 if (could_have_dest && !dest_spec) { 4998 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4999 } 5000 } 5001 } 5002 } 5003 5004 ciMethod* trap_method = method(); 5005 int trap_bci = bci(); 5006 if (saved_jvms != NULL) { 5007 trap_method = alloc->jvms()->method(); 5008 trap_bci = alloc->jvms()->bci(); 5009 } 5010 5011 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5012 can_emit_guards && 5013 !src->is_top() && !dest->is_top()) { 5014 // validate arguments: enables transformation the ArrayCopyNode 5015 validated = true; 5016 5017 RegionNode* slow_region = new RegionNode(1); 5018 record_for_igvn(slow_region); 5019 5020 // (1) src and dest are arrays. 5021 generate_non_array_guard(load_object_klass(src), slow_region); 5022 generate_non_array_guard(load_object_klass(dest), slow_region); 5023 5024 // (2) src and dest arrays must have elements of the same BasicType 5025 // done at macro expansion or at Ideal transformation time 5026 5027 // (4) src_offset must not be negative. 5028 generate_negative_guard(src_offset, slow_region); 5029 5030 // (5) dest_offset must not be negative. 5031 generate_negative_guard(dest_offset, slow_region); 5032 5033 // (7) src_offset + length must not exceed length of src. 5034 generate_limit_guard(src_offset, length, 5035 load_array_length(src), 5036 slow_region); 5037 5038 // (8) dest_offset + length must not exceed length of dest. 5039 generate_limit_guard(dest_offset, length, 5040 load_array_length(dest), 5041 slow_region); 5042 5043 // (9) each element of an oop array must be assignable 5044 Node* src_klass = load_object_klass(src); 5045 Node* dest_klass = load_object_klass(dest); 5046 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5047 5048 if (not_subtype_ctrl != top()) { 5049 PreserveJVMState pjvms(this); 5050 set_control(not_subtype_ctrl); 5051 uncommon_trap(Deoptimization::Reason_intrinsic, 5052 Deoptimization::Action_make_not_entrant); 5053 assert(stopped(), "Should be stopped"); 5054 } 5055 { 5056 PreserveJVMState pjvms(this); 5057 set_control(_gvn.transform(slow_region)); 5058 uncommon_trap(Deoptimization::Reason_intrinsic, 5059 Deoptimization::Action_make_not_entrant); 5060 assert(stopped(), "Should be stopped"); 5061 } 5062 } 5063 5064 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5065 5066 if (stopped()) { 5067 return true; 5068 } 5069 5070 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5071 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5072 // so the compiler has a chance to eliminate them: during macro expansion, 5073 // we have to set their control (CastPP nodes are eliminated). 5074 load_object_klass(src), load_object_klass(dest), 5075 load_array_length(src), load_array_length(dest)); 5076 5077 ac->set_arraycopy(validated); 5078 5079 Node* n = _gvn.transform(ac); 5080 if (n == ac) { 5081 ac->connect_outputs(this); 5082 } else { 5083 assert(validated, "shouldn't transform if all arguments not validated"); 5084 set_all_memory(n); 5085 } 5086 5087 return true; 5088 } 5089 5090 5091 // Helper function which determines if an arraycopy immediately follows 5092 // an allocation, with no intervening tests or other escapes for the object. 5093 AllocateArrayNode* 5094 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5095 RegionNode* slow_region) { 5096 if (stopped()) return NULL; // no fast path 5097 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5098 5099 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5100 if (alloc == NULL) return NULL; 5101 5102 Node* rawmem = memory(Compile::AliasIdxRaw); 5103 // Is the allocation's memory state untouched? 5104 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5105 // Bail out if there have been raw-memory effects since the allocation. 5106 // (Example: There might have been a call or safepoint.) 5107 return NULL; 5108 } 5109 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5110 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5111 return NULL; 5112 } 5113 5114 // There must be no unexpected observers of this allocation. 5115 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5116 Node* obs = ptr->fast_out(i); 5117 if (obs != this->map()) { 5118 return NULL; 5119 } 5120 } 5121 5122 // This arraycopy must unconditionally follow the allocation of the ptr. 5123 Node* alloc_ctl = ptr->in(0); 5124 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5125 5126 Node* ctl = control(); 5127 while (ctl != alloc_ctl) { 5128 // There may be guards which feed into the slow_region. 5129 // Any other control flow means that we might not get a chance 5130 // to finish initializing the allocated object. 5131 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5132 IfNode* iff = ctl->in(0)->as_If(); 5133 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5134 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5135 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5136 ctl = iff->in(0); // This test feeds the known slow_region. 5137 continue; 5138 } 5139 // One more try: Various low-level checks bottom out in 5140 // uncommon traps. If the debug-info of the trap omits 5141 // any reference to the allocation, as we've already 5142 // observed, then there can be no objection to the trap. 5143 bool found_trap = false; 5144 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5145 Node* obs = not_ctl->fast_out(j); 5146 if (obs->in(0) == not_ctl && obs->is_Call() && 5147 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5148 found_trap = true; break; 5149 } 5150 } 5151 if (found_trap) { 5152 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5153 continue; 5154 } 5155 } 5156 return NULL; 5157 } 5158 5159 // If we get this far, we have an allocation which immediately 5160 // precedes the arraycopy, and we can take over zeroing the new object. 5161 // The arraycopy will finish the initialization, and provide 5162 // a new control state to which we will anchor the destination pointer. 5163 5164 return alloc; 5165 } 5166 5167 //-------------inline_encodeISOArray----------------------------------- 5168 // encode char[] to byte[] in ISO_8859_1 5169 bool LibraryCallKit::inline_encodeISOArray() { 5170 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5171 // no receiver since it is static method 5172 Node *src = argument(0); 5173 Node *src_offset = argument(1); 5174 Node *dst = argument(2); 5175 Node *dst_offset = argument(3); 5176 Node *length = argument(4); 5177 5178 const Type* src_type = src->Value(&_gvn); 5179 const Type* dst_type = dst->Value(&_gvn); 5180 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5181 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5182 if (top_src == NULL || top_src->klass() == NULL || 5183 top_dest == NULL || top_dest->klass() == NULL) { 5184 // failed array check 5185 return false; 5186 } 5187 5188 // Figure out the size and type of the elements we will be copying. 5189 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5190 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5191 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5192 return false; 5193 } 5194 Node* src_start = array_element_address(src, src_offset, src_elem); 5195 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5196 // 'src_start' points to src array + scaled offset 5197 // 'dst_start' points to dst array + scaled offset 5198 5199 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5200 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5201 enc = _gvn.transform(enc); 5202 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5203 set_memory(res_mem, mtype); 5204 set_result(enc); 5205 return true; 5206 } 5207 5208 //-------------inline_multiplyToLen----------------------------------- 5209 bool LibraryCallKit::inline_multiplyToLen() { 5210 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5211 5212 address stubAddr = StubRoutines::multiplyToLen(); 5213 if (stubAddr == NULL) { 5214 return false; // Intrinsic's stub is not implemented on this platform 5215 } 5216 const char* stubName = "multiplyToLen"; 5217 5218 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5219 5220 Node* x = argument(1); 5221 Node* xlen = argument(2); 5222 Node* y = argument(3); 5223 Node* ylen = argument(4); 5224 Node* z = argument(5); 5225 5226 const Type* x_type = x->Value(&_gvn); 5227 const Type* y_type = y->Value(&_gvn); 5228 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5229 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5230 if (top_x == NULL || top_x->klass() == NULL || 5231 top_y == NULL || top_y->klass() == NULL) { 5232 // failed array check 5233 return false; 5234 } 5235 5236 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5237 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5238 if (x_elem != T_INT || y_elem != T_INT) { 5239 return false; 5240 } 5241 5242 // Set the original stack and the reexecute bit for the interpreter to reexecute 5243 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5244 // on the return from z array allocation in runtime. 5245 { PreserveReexecuteState preexecs(this); 5246 jvms()->set_should_reexecute(true); 5247 5248 Node* x_start = array_element_address(x, intcon(0), x_elem); 5249 Node* y_start = array_element_address(y, intcon(0), y_elem); 5250 // 'x_start' points to x array + scaled xlen 5251 // 'y_start' points to y array + scaled ylen 5252 5253 // Allocate the result array 5254 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5255 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5256 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5257 5258 IdealKit ideal(this); 5259 5260 #define __ ideal. 5261 Node* one = __ ConI(1); 5262 Node* zero = __ ConI(0); 5263 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5264 __ set(need_alloc, zero); 5265 __ set(z_alloc, z); 5266 __ if_then(z, BoolTest::eq, null()); { 5267 __ increment (need_alloc, one); 5268 } __ else_(); { 5269 // Update graphKit memory and control from IdealKit. 5270 sync_kit(ideal); 5271 Node* zlen_arg = load_array_length(z); 5272 // Update IdealKit memory and control from graphKit. 5273 __ sync_kit(this); 5274 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5275 __ increment (need_alloc, one); 5276 } __ end_if(); 5277 } __ end_if(); 5278 5279 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5280 // Update graphKit memory and control from IdealKit. 5281 sync_kit(ideal); 5282 Node * narr = new_array(klass_node, zlen, 1); 5283 // Update IdealKit memory and control from graphKit. 5284 __ sync_kit(this); 5285 __ set(z_alloc, narr); 5286 } __ end_if(); 5287 5288 sync_kit(ideal); 5289 z = __ value(z_alloc); 5290 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5291 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5292 // Final sync IdealKit and GraphKit. 5293 final_sync(ideal); 5294 #undef __ 5295 5296 Node* z_start = array_element_address(z, intcon(0), T_INT); 5297 5298 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5299 OptoRuntime::multiplyToLen_Type(), 5300 stubAddr, stubName, TypePtr::BOTTOM, 5301 x_start, xlen, y_start, ylen, z_start, zlen); 5302 } // original reexecute is set back here 5303 5304 C->set_has_split_ifs(true); // Has chance for split-if optimization 5305 set_result(z); 5306 return true; 5307 } 5308 5309 5310 /** 5311 * Calculate CRC32 for byte. 5312 * int java.util.zip.CRC32.update(int crc, int b) 5313 */ 5314 bool LibraryCallKit::inline_updateCRC32() { 5315 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5316 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5317 // no receiver since it is static method 5318 Node* crc = argument(0); // type: int 5319 Node* b = argument(1); // type: int 5320 5321 /* 5322 * int c = ~ crc; 5323 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5324 * b = b ^ (c >>> 8); 5325 * crc = ~b; 5326 */ 5327 5328 Node* M1 = intcon(-1); 5329 crc = _gvn.transform(new XorINode(crc, M1)); 5330 Node* result = _gvn.transform(new XorINode(crc, b)); 5331 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5332 5333 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5334 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5335 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5336 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5337 5338 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5339 result = _gvn.transform(new XorINode(crc, result)); 5340 result = _gvn.transform(new XorINode(result, M1)); 5341 set_result(result); 5342 return true; 5343 } 5344 5345 /** 5346 * Calculate CRC32 for byte[] array. 5347 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5348 */ 5349 bool LibraryCallKit::inline_updateBytesCRC32() { 5350 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5351 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5352 // no receiver since it is static method 5353 Node* crc = argument(0); // type: int 5354 Node* src = argument(1); // type: oop 5355 Node* offset = argument(2); // type: int 5356 Node* length = argument(3); // type: int 5357 5358 const Type* src_type = src->Value(&_gvn); 5359 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5360 if (top_src == NULL || top_src->klass() == NULL) { 5361 // failed array check 5362 return false; 5363 } 5364 5365 // Figure out the size and type of the elements we will be copying. 5366 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5367 if (src_elem != T_BYTE) { 5368 return false; 5369 } 5370 5371 // 'src_start' points to src array + scaled offset 5372 Node* src_start = array_element_address(src, offset, src_elem); 5373 5374 // We assume that range check is done by caller. 5375 // TODO: generate range check (offset+length < src.length) in debug VM. 5376 5377 // Call the stub. 5378 address stubAddr = StubRoutines::updateBytesCRC32(); 5379 const char *stubName = "updateBytesCRC32"; 5380 5381 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5382 stubAddr, stubName, TypePtr::BOTTOM, 5383 crc, src_start, length); 5384 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5385 set_result(result); 5386 return true; 5387 } 5388 5389 /** 5390 * Calculate CRC32 for ByteBuffer. 5391 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5392 */ 5393 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5394 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5395 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5396 // no receiver since it is static method 5397 Node* crc = argument(0); // type: int 5398 Node* src = argument(1); // type: long 5399 Node* offset = argument(3); // type: int 5400 Node* length = argument(4); // type: int 5401 5402 src = ConvL2X(src); // adjust Java long to machine word 5403 Node* base = _gvn.transform(new CastX2PNode(src)); 5404 offset = ConvI2X(offset); 5405 5406 // 'src_start' points to src array + scaled offset 5407 Node* src_start = basic_plus_adr(top(), base, offset); 5408 5409 // Call the stub. 5410 address stubAddr = StubRoutines::updateBytesCRC32(); 5411 const char *stubName = "updateBytesCRC32"; 5412 5413 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5414 stubAddr, stubName, TypePtr::BOTTOM, 5415 crc, src_start, length); 5416 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5417 set_result(result); 5418 return true; 5419 } 5420 5421 //----------------------------inline_reference_get---------------------------- 5422 // public T java.lang.ref.Reference.get(); 5423 bool LibraryCallKit::inline_reference_get() { 5424 const int referent_offset = java_lang_ref_Reference::referent_offset; 5425 guarantee(referent_offset > 0, "should have already been set"); 5426 5427 // Get the argument: 5428 Node* reference_obj = null_check_receiver(); 5429 if (stopped()) return true; 5430 5431 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5432 5433 ciInstanceKlass* klass = env()->Object_klass(); 5434 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5435 5436 Node* no_ctrl = NULL; 5437 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5438 5439 // Use the pre-barrier to record the value in the referent field 5440 pre_barrier(false /* do_load */, 5441 control(), 5442 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5443 result /* pre_val */, 5444 T_OBJECT); 5445 5446 // Add memory barrier to prevent commoning reads from this field 5447 // across safepoint since GC can change its value. 5448 insert_mem_bar(Op_MemBarCPUOrder); 5449 5450 set_result(result); 5451 return true; 5452 } 5453 5454 5455 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5456 bool is_exact=true, bool is_static=false) { 5457 5458 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5459 assert(tinst != NULL, "obj is null"); 5460 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5461 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5462 5463 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5464 ciSymbol::make(fieldTypeString), 5465 is_static); 5466 if (field == NULL) return (Node *) NULL; 5467 assert (field != NULL, "undefined field"); 5468 5469 // Next code copied from Parse::do_get_xxx(): 5470 5471 // Compute address and memory type. 5472 int offset = field->offset_in_bytes(); 5473 bool is_vol = field->is_volatile(); 5474 ciType* field_klass = field->type(); 5475 assert(field_klass->is_loaded(), "should be loaded"); 5476 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5477 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5478 BasicType bt = field->layout_type(); 5479 5480 // Build the resultant type of the load 5481 const Type *type; 5482 if (bt == T_OBJECT) { 5483 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5484 } else { 5485 type = Type::get_const_basic_type(bt); 5486 } 5487 5488 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5489 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5490 } 5491 // Build the load. 5492 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5493 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5494 // If reference is volatile, prevent following memory ops from 5495 // floating up past the volatile read. Also prevents commoning 5496 // another volatile read. 5497 if (is_vol) { 5498 // Memory barrier includes bogus read of value to force load BEFORE membar 5499 insert_mem_bar(Op_MemBarAcquire, loadedField); 5500 } 5501 return loadedField; 5502 } 5503 5504 5505 //------------------------------inline_aescrypt_Block----------------------- 5506 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5507 address stubAddr; 5508 const char *stubName; 5509 assert(UseAES, "need AES instruction support"); 5510 5511 switch(id) { 5512 case vmIntrinsics::_aescrypt_encryptBlock: 5513 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5514 stubName = "aescrypt_encryptBlock"; 5515 break; 5516 case vmIntrinsics::_aescrypt_decryptBlock: 5517 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5518 stubName = "aescrypt_decryptBlock"; 5519 break; 5520 } 5521 if (stubAddr == NULL) return false; 5522 5523 Node* aescrypt_object = argument(0); 5524 Node* src = argument(1); 5525 Node* src_offset = argument(2); 5526 Node* dest = argument(3); 5527 Node* dest_offset = argument(4); 5528 5529 // (1) src and dest are arrays. 5530 const Type* src_type = src->Value(&_gvn); 5531 const Type* dest_type = dest->Value(&_gvn); 5532 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5533 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5534 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5535 5536 // for the quick and dirty code we will skip all the checks. 5537 // we are just trying to get the call to be generated. 5538 Node* src_start = src; 5539 Node* dest_start = dest; 5540 if (src_offset != NULL || dest_offset != NULL) { 5541 assert(src_offset != NULL && dest_offset != NULL, ""); 5542 src_start = array_element_address(src, src_offset, T_BYTE); 5543 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5544 } 5545 5546 // now need to get the start of its expanded key array 5547 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5548 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5549 if (k_start == NULL) return false; 5550 5551 if (Matcher::pass_original_key_for_aes()) { 5552 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5553 // compatibility issues between Java key expansion and SPARC crypto instructions 5554 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5555 if (original_k_start == NULL) return false; 5556 5557 // Call the stub. 5558 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5559 stubAddr, stubName, TypePtr::BOTTOM, 5560 src_start, dest_start, k_start, original_k_start); 5561 } else { 5562 // Call the stub. 5563 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5564 stubAddr, stubName, TypePtr::BOTTOM, 5565 src_start, dest_start, k_start); 5566 } 5567 5568 return true; 5569 } 5570 5571 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5572 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5573 address stubAddr; 5574 const char *stubName; 5575 5576 assert(UseAES, "need AES instruction support"); 5577 5578 switch(id) { 5579 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5580 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5581 stubName = "cipherBlockChaining_encryptAESCrypt"; 5582 break; 5583 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5584 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5585 stubName = "cipherBlockChaining_decryptAESCrypt"; 5586 break; 5587 } 5588 if (stubAddr == NULL) return false; 5589 5590 Node* cipherBlockChaining_object = argument(0); 5591 Node* src = argument(1); 5592 Node* src_offset = argument(2); 5593 Node* len = argument(3); 5594 Node* dest = argument(4); 5595 Node* dest_offset = argument(5); 5596 5597 // (1) src and dest are arrays. 5598 const Type* src_type = src->Value(&_gvn); 5599 const Type* dest_type = dest->Value(&_gvn); 5600 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5601 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5602 assert (top_src != NULL && top_src->klass() != NULL 5603 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5604 5605 // checks are the responsibility of the caller 5606 Node* src_start = src; 5607 Node* dest_start = dest; 5608 if (src_offset != NULL || dest_offset != NULL) { 5609 assert(src_offset != NULL && dest_offset != NULL, ""); 5610 src_start = array_element_address(src, src_offset, T_BYTE); 5611 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5612 } 5613 5614 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5615 // (because of the predicated logic executed earlier). 5616 // so we cast it here safely. 5617 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5618 5619 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5620 if (embeddedCipherObj == NULL) return false; 5621 5622 // cast it to what we know it will be at runtime 5623 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5624 assert(tinst != NULL, "CBC obj is null"); 5625 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5626 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5627 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5628 5629 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5630 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5631 const TypeOopPtr* xtype = aklass->as_instance_type(); 5632 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5633 aescrypt_object = _gvn.transform(aescrypt_object); 5634 5635 // we need to get the start of the aescrypt_object's expanded key array 5636 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5637 if (k_start == NULL) return false; 5638 5639 // similarly, get the start address of the r vector 5640 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5641 if (objRvec == NULL) return false; 5642 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5643 5644 Node* cbcCrypt; 5645 if (Matcher::pass_original_key_for_aes()) { 5646 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5647 // compatibility issues between Java key expansion and SPARC crypto instructions 5648 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5649 if (original_k_start == NULL) return false; 5650 5651 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5652 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5653 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5654 stubAddr, stubName, TypePtr::BOTTOM, 5655 src_start, dest_start, k_start, r_start, len, original_k_start); 5656 } else { 5657 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5658 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5659 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5660 stubAddr, stubName, TypePtr::BOTTOM, 5661 src_start, dest_start, k_start, r_start, len); 5662 } 5663 5664 // return cipher length (int) 5665 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5666 set_result(retvalue); 5667 return true; 5668 } 5669 5670 //------------------------------get_key_start_from_aescrypt_object----------------------- 5671 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5672 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5673 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5674 if (objAESCryptKey == NULL) return (Node *) NULL; 5675 5676 // now have the array, need to get the start address of the K array 5677 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5678 return k_start; 5679 } 5680 5681 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5682 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5683 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5684 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5685 if (objAESCryptKey == NULL) return (Node *) NULL; 5686 5687 // now have the array, need to get the start address of the lastKey array 5688 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5689 return original_k_start; 5690 } 5691 5692 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5693 // Return node representing slow path of predicate check. 5694 // the pseudo code we want to emulate with this predicate is: 5695 // for encryption: 5696 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5697 // for decryption: 5698 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5699 // note cipher==plain is more conservative than the original java code but that's OK 5700 // 5701 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5702 // The receiver was checked for NULL already. 5703 Node* objCBC = argument(0); 5704 5705 // Load embeddedCipher field of CipherBlockChaining object. 5706 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5707 5708 // get AESCrypt klass for instanceOf check 5709 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5710 // will have same classloader as CipherBlockChaining object 5711 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5712 assert(tinst != NULL, "CBCobj is null"); 5713 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5714 5715 // we want to do an instanceof comparison against the AESCrypt class 5716 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5717 if (!klass_AESCrypt->is_loaded()) { 5718 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5719 Node* ctrl = control(); 5720 set_control(top()); // no regular fast path 5721 return ctrl; 5722 } 5723 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5724 5725 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5726 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5727 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5728 5729 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5730 5731 // for encryption, we are done 5732 if (!decrypting) 5733 return instof_false; // even if it is NULL 5734 5735 // for decryption, we need to add a further check to avoid 5736 // taking the intrinsic path when cipher and plain are the same 5737 // see the original java code for why. 5738 RegionNode* region = new RegionNode(3); 5739 region->init_req(1, instof_false); 5740 Node* src = argument(1); 5741 Node* dest = argument(4); 5742 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5743 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5744 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5745 region->init_req(2, src_dest_conjoint); 5746 5747 record_for_igvn(region); 5748 return _gvn.transform(region); 5749 } 5750 5751 //------------------------------inline_sha_implCompress----------------------- 5752 // 5753 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 5754 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 5755 // 5756 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 5757 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 5758 // 5759 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 5760 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 5761 // 5762 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 5763 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 5764 5765 Node* sha_obj = argument(0); 5766 Node* src = argument(1); // type oop 5767 Node* ofs = argument(2); // type int 5768 5769 const Type* src_type = src->Value(&_gvn); 5770 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5771 if (top_src == NULL || top_src->klass() == NULL) { 5772 // failed array check 5773 return false; 5774 } 5775 // Figure out the size and type of the elements we will be copying. 5776 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5777 if (src_elem != T_BYTE) { 5778 return false; 5779 } 5780 // 'src_start' points to src array + offset 5781 Node* src_start = array_element_address(src, ofs, src_elem); 5782 Node* state = NULL; 5783 address stubAddr; 5784 const char *stubName; 5785 5786 switch(id) { 5787 case vmIntrinsics::_sha_implCompress: 5788 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 5789 state = get_state_from_sha_object(sha_obj); 5790 stubAddr = StubRoutines::sha1_implCompress(); 5791 stubName = "sha1_implCompress"; 5792 break; 5793 case vmIntrinsics::_sha2_implCompress: 5794 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 5795 state = get_state_from_sha_object(sha_obj); 5796 stubAddr = StubRoutines::sha256_implCompress(); 5797 stubName = "sha256_implCompress"; 5798 break; 5799 case vmIntrinsics::_sha5_implCompress: 5800 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 5801 state = get_state_from_sha5_object(sha_obj); 5802 stubAddr = StubRoutines::sha512_implCompress(); 5803 stubName = "sha512_implCompress"; 5804 break; 5805 default: 5806 fatal_unexpected_iid(id); 5807 return false; 5808 } 5809 if (state == NULL) return false; 5810 5811 // Call the stub. 5812 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 5813 stubAddr, stubName, TypePtr::BOTTOM, 5814 src_start, state); 5815 5816 return true; 5817 } 5818 5819 //------------------------------inline_digestBase_implCompressMB----------------------- 5820 // 5821 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 5822 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 5823 // 5824 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 5825 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5826 "need SHA1/SHA256/SHA512 instruction support"); 5827 assert((uint)predicate < 3, "sanity"); 5828 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 5829 5830 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 5831 Node* src = argument(1); // byte[] array 5832 Node* ofs = argument(2); // type int 5833 Node* limit = argument(3); // type int 5834 5835 const Type* src_type = src->Value(&_gvn); 5836 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5837 if (top_src == NULL || top_src->klass() == NULL) { 5838 // failed array check 5839 return false; 5840 } 5841 // Figure out the size and type of the elements we will be copying. 5842 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5843 if (src_elem != T_BYTE) { 5844 return false; 5845 } 5846 // 'src_start' points to src array + offset 5847 Node* src_start = array_element_address(src, ofs, src_elem); 5848 5849 const char* klass_SHA_name = NULL; 5850 const char* stub_name = NULL; 5851 address stub_addr = NULL; 5852 bool long_state = false; 5853 5854 switch (predicate) { 5855 case 0: 5856 if (UseSHA1Intrinsics) { 5857 klass_SHA_name = "sun/security/provider/SHA"; 5858 stub_name = "sha1_implCompressMB"; 5859 stub_addr = StubRoutines::sha1_implCompressMB(); 5860 } 5861 break; 5862 case 1: 5863 if (UseSHA256Intrinsics) { 5864 klass_SHA_name = "sun/security/provider/SHA2"; 5865 stub_name = "sha256_implCompressMB"; 5866 stub_addr = StubRoutines::sha256_implCompressMB(); 5867 } 5868 break; 5869 case 2: 5870 if (UseSHA512Intrinsics) { 5871 klass_SHA_name = "sun/security/provider/SHA5"; 5872 stub_name = "sha512_implCompressMB"; 5873 stub_addr = StubRoutines::sha512_implCompressMB(); 5874 long_state = true; 5875 } 5876 break; 5877 default: 5878 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5879 } 5880 if (klass_SHA_name != NULL) { 5881 // get DigestBase klass to lookup for SHA klass 5882 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 5883 assert(tinst != NULL, "digestBase_obj is not instance???"); 5884 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5885 5886 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5887 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 5888 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5889 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 5890 } 5891 return false; 5892 } 5893 //------------------------------inline_sha_implCompressMB----------------------- 5894 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 5895 bool long_state, address stubAddr, const char *stubName, 5896 Node* src_start, Node* ofs, Node* limit) { 5897 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 5898 const TypeOopPtr* xtype = aklass->as_instance_type(); 5899 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 5900 sha_obj = _gvn.transform(sha_obj); 5901 5902 Node* state; 5903 if (long_state) { 5904 state = get_state_from_sha5_object(sha_obj); 5905 } else { 5906 state = get_state_from_sha_object(sha_obj); 5907 } 5908 if (state == NULL) return false; 5909 5910 // Call the stub. 5911 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5912 OptoRuntime::digestBase_implCompressMB_Type(), 5913 stubAddr, stubName, TypePtr::BOTTOM, 5914 src_start, state, ofs, limit); 5915 // return ofs (int) 5916 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5917 set_result(result); 5918 5919 return true; 5920 } 5921 5922 //------------------------------get_state_from_sha_object----------------------- 5923 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 5924 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 5925 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 5926 if (sha_state == NULL) return (Node *) NULL; 5927 5928 // now have the array, need to get the start address of the state array 5929 Node* state = array_element_address(sha_state, intcon(0), T_INT); 5930 return state; 5931 } 5932 5933 //------------------------------get_state_from_sha5_object----------------------- 5934 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 5935 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 5936 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 5937 if (sha_state == NULL) return (Node *) NULL; 5938 5939 // now have the array, need to get the start address of the state array 5940 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 5941 return state; 5942 } 5943 5944 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 5945 // Return node representing slow path of predicate check. 5946 // the pseudo code we want to emulate with this predicate is: 5947 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 5948 // 5949 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 5950 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5951 "need SHA1/SHA256/SHA512 instruction support"); 5952 assert((uint)predicate < 3, "sanity"); 5953 5954 // The receiver was checked for NULL already. 5955 Node* digestBaseObj = argument(0); 5956 5957 // get DigestBase klass for instanceOf check 5958 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 5959 assert(tinst != NULL, "digestBaseObj is null"); 5960 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5961 5962 const char* klass_SHA_name = NULL; 5963 switch (predicate) { 5964 case 0: 5965 if (UseSHA1Intrinsics) { 5966 // we want to do an instanceof comparison against the SHA class 5967 klass_SHA_name = "sun/security/provider/SHA"; 5968 } 5969 break; 5970 case 1: 5971 if (UseSHA256Intrinsics) { 5972 // we want to do an instanceof comparison against the SHA2 class 5973 klass_SHA_name = "sun/security/provider/SHA2"; 5974 } 5975 break; 5976 case 2: 5977 if (UseSHA512Intrinsics) { 5978 // we want to do an instanceof comparison against the SHA5 class 5979 klass_SHA_name = "sun/security/provider/SHA5"; 5980 } 5981 break; 5982 default: 5983 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5984 } 5985 5986 ciKlass* klass_SHA = NULL; 5987 if (klass_SHA_name != NULL) { 5988 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5989 } 5990 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 5991 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 5992 Node* ctrl = control(); 5993 set_control(top()); // no intrinsic path 5994 return ctrl; 5995 } 5996 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5997 5998 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 5999 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6000 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6001 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6002 6003 return instof_false; // even if it is NULL 6004 } 6005 6006 bool LibraryCallKit::inline_profileBoolean() { 6007 Node* counts = argument(1); 6008 const TypeAryPtr* ary = NULL; 6009 ciArray* aobj = NULL; 6010 if (counts->is_Con() 6011 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6012 && (aobj = ary->const_oop()->as_array()) != NULL 6013 && (aobj->length() == 2)) { 6014 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6015 jint false_cnt = aobj->element_value(0).as_int(); 6016 jint true_cnt = aobj->element_value(1).as_int(); 6017 6018 if (C->log() != NULL) { 6019 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6020 false_cnt, true_cnt); 6021 } 6022 6023 if (false_cnt + true_cnt == 0) { 6024 // According to profile, never executed. 6025 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6026 Deoptimization::Action_reinterpret); 6027 return true; 6028 } 6029 6030 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6031 // is a number of each value occurrences. 6032 Node* result = argument(0); 6033 if (false_cnt == 0 || true_cnt == 0) { 6034 // According to profile, one value has been never seen. 6035 int expected_val = (false_cnt == 0) ? 1 : 0; 6036 6037 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6038 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6039 6040 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6041 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6042 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6043 6044 { // Slow path: uncommon trap for never seen value and then reexecute 6045 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6046 // the value has been seen at least once. 6047 PreserveJVMState pjvms(this); 6048 PreserveReexecuteState preexecs(this); 6049 jvms()->set_should_reexecute(true); 6050 6051 set_control(slow_path); 6052 set_i_o(i_o()); 6053 6054 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6055 Deoptimization::Action_reinterpret); 6056 } 6057 // The guard for never seen value enables sharpening of the result and 6058 // returning a constant. It allows to eliminate branches on the same value 6059 // later on. 6060 set_control(fast_path); 6061 result = intcon(expected_val); 6062 } 6063 // Stop profiling. 6064 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6065 // By replacing method body with profile data (represented as ProfileBooleanNode 6066 // on IR level) we effectively disable profiling. 6067 // It enables full speed execution once optimized code is generated. 6068 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6069 C->record_for_igvn(profile); 6070 set_result(profile); 6071 return true; 6072 } else { 6073 // Continue profiling. 6074 // Profile data isn't available at the moment. So, execute method's bytecode version. 6075 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6076 // is compiled and counters aren't available since corresponding MethodHandle 6077 // isn't a compile-time constant. 6078 return false; 6079 } 6080 } 6081 6082 bool LibraryCallKit::inline_isCompileConstant() { 6083 Node* n = argument(0); 6084 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6085 return true; 6086 }