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