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