library_call.cpp revision 5356:c9ccd7b85f20
121308Sache/* 221308Sache * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved. 321308Sache * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 421308Sache * 521308Sache * This code is free software; you can redistribute it and/or modify it 621308Sache * under the terms of the GNU General Public License version 2 only, as 721308Sache * published by the Free Software Foundation. 821308Sache * 921308Sache * This code is distributed in the hope that it will be useful, but WITHOUT 1058310Sache * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 1121308Sache * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 1221308Sache * version 2 for more details (a copy is included in the LICENSE file that 1321308Sache * accompanied this code). 1421308Sache * 1521308Sache * You should have received a copy of the GNU General Public License version 1621308Sache * 2 along with this work; if not, write to the Free Software Foundation, 1721308Sache * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 1821308Sache * 1921308Sache * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 2021308Sache * or visit www.oracle.com if you need additional information or have any 2158310Sache * questions. 2221308Sache * 2321308Sache */ 2421308Sache 2521308Sache#include "precompiled.hpp" 2621308Sache#include "classfile/systemDictionary.hpp" 2721308Sache#include "classfile/vmSymbols.hpp" 2821308Sache#include "compiler/compileBroker.hpp" 2921308Sache#include "compiler/compileLog.hpp" 3021308Sache#include "oops/objArrayKlass.hpp" 3121308Sache#include "opto/addnode.hpp" 3221308Sache#include "opto/callGenerator.hpp" 3321308Sache#include "opto/cfgnode.hpp" 3421308Sache#include "opto/idealKit.hpp" 3521308Sache#include "opto/mathexactnode.hpp" 3621308Sache#include "opto/mulnode.hpp" 3721308Sache#include "opto/parse.hpp" 3821308Sache#include "opto/runtime.hpp" 3921308Sache#include "opto/subnode.hpp" 4021308Sache#include "prims/nativeLookup.hpp" 4121308Sache#include "runtime/sharedRuntime.hpp" 4221308Sache#include "trace/traceMacros.hpp" 4321308Sache 4421308Sacheclass LibraryIntrinsic : public InlineCallGenerator { 4521308Sache // Extend the set of intrinsics known to the runtime: 4621308Sache public: 4721308Sache private: 4821308Sache bool _is_virtual; 4958310Sache bool _is_predicted; 5058310Sache vmIntrinsics::ID _intrinsic_id; 5158310Sache 5221308Sache public: 5321308Sache LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, vmIntrinsics::ID id) 5421308Sache : InlineCallGenerator(m), 5521308Sache _is_virtual(is_virtual), 5621308Sache _is_predicted(is_predicted), 5721308Sache _intrinsic_id(id) 5875406Sache { 5975406Sache } 6075406Sache virtual bool is_intrinsic() const { return true; } 6175406Sache virtual bool is_virtual() const { return _is_virtual; } 6221308Sache virtual bool is_predicted() const { return _is_predicted; } 6321308Sache virtual JVMState* generate(JVMState* jvms); 6421308Sache virtual Node* generate_predicate(JVMState* jvms); 6521308Sache vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 6621308Sache}; 6721308Sache 6821308Sache 6921308Sache// Local helper class for LibraryIntrinsic: 7021308Sacheclass LibraryCallKit : public GraphKit { 7121308Sache private: 7221308Sache LibraryIntrinsic* _intrinsic; // the library intrinsic being called 7321308Sache Node* _result; // the result node, if any 7421308Sache int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 7521308Sache 7621308Sache const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 7721308Sache 7821308Sache public: 7921308Sache LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 8021308Sache : GraphKit(jvms), 8121308Sache _intrinsic(intrinsic), 8221308Sache _result(NULL) 8321308Sache { 8421308Sache // Check if this is a root compile. In that case we don't have a caller. 8521308Sache if (!jvms->has_method()) { 8621308Sache _reexecute_sp = sp(); 8721308Sache } else { 8821308Sache // Find out how many arguments the interpreter needs when deoptimizing 8921308Sache // and save the stack pointer value so it can used by uncommon_trap. 9021308Sache // We find the argument count by looking at the declared signature. 9121308Sache bool ignored_will_link; 9221308Sache ciSignature* declared_signature = NULL; 9375406Sache ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 9421308Sache const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 9575406Sache _reexecute_sp = sp() + nargs; // "push" arguments back on stack 9621308Sache } 9721308Sache } 9821308Sache 9921308Sache virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 10021308Sache 10121308Sache ciMethod* caller() const { return jvms()->method(); } 10221308Sache int bci() const { return jvms()->bci(); } 103157184Sache LibraryIntrinsic* intrinsic() const { return _intrinsic; } 104157184Sache vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 10575406Sache ciMethod* callee() const { return _intrinsic->method(); } 10621308Sache 10721308Sache bool try_to_inline(); 10875406Sache Node* try_to_predicate(); 10921308Sache 11021308Sache void push_result() { 11121308Sache // Push the result onto the stack. 11221308Sache if (!stopped() && result() != NULL) { 11321308Sache BasicType bt = result()->bottom_type()->basic_type(); 114157184Sache push_node(bt, result()); 115157184Sache } 116165670Sache } 117157184Sache 118157184Sache private: 119157184Sache void fatal_unexpected_iid(vmIntrinsics::ID iid) { 12075406Sache fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 121157184Sache } 12221308Sache 12321308Sache void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 12421308Sache void set_result(RegionNode* region, PhiNode* value); 12521308Sache Node* result() { return _result; } 12621308Sache 12721308Sache virtual int reexecute_sp() { return _reexecute_sp; } 12821308Sache 12921308Sache // Helper functions to inline natives 13021308Sache Node* generate_guard(Node* test, RegionNode* region, float true_prob); 13121308Sache Node* generate_slow_guard(Node* test, RegionNode* region); 13221308Sache Node* generate_fair_guard(Node* test, RegionNode* region); 13375406Sache Node* generate_negative_guard(Node* index, RegionNode* region, 13421308Sache // resulting CastII of index: 13521308Sache Node* *pos_index = NULL); 13621308Sache Node* generate_nonpositive_guard(Node* index, bool never_negative, 13721308Sache // resulting CastII of index: 13821308Sache Node* *pos_index = NULL); 13921308Sache Node* generate_limit_guard(Node* offset, Node* subseq_length, 14021308Sache Node* array_length, 14121308Sache RegionNode* region); 14221308Sache Node* generate_current_thread(Node* &tls_output); 14321308Sache address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 14421308Sache bool disjoint_bases, const char* &name, bool dest_uninitialized); 14575406Sache Node* load_mirror_from_klass(Node* klass); 14675406Sache Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 14721308Sache RegionNode* region, int null_path, 14821308Sache int offset); 14921308Sache Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 15021308Sache RegionNode* region, int null_path) { 15121308Sache int offset = java_lang_Class::klass_offset_in_bytes(); 15275406Sache return load_klass_from_mirror_common(mirror, never_see_null, 15321308Sache region, null_path, 15421308Sache offset); 15521308Sache } 15621308Sache Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 15775406Sache RegionNode* region, int null_path) { 15875406Sache int offset = java_lang_Class::array_klass_offset_in_bytes(); 15975406Sache return load_klass_from_mirror_common(mirror, never_see_null, 16021308Sache region, null_path, 16121308Sache offset); 16221308Sache } 16321308Sache Node* generate_access_flags_guard(Node* kls, 16421308Sache int modifier_mask, int modifier_bits, 16521308Sache RegionNode* region); 16621308Sache Node* generate_interface_guard(Node* kls, RegionNode* region); 16721308Sache Node* generate_array_guard(Node* kls, RegionNode* region) { 16821308Sache return generate_array_guard_common(kls, region, false, false); 16921308Sache } 170119610Sache Node* generate_non_array_guard(Node* kls, RegionNode* region) { 17121308Sache return generate_array_guard_common(kls, region, false, true); 172119610Sache } 17321308Sache Node* generate_objArray_guard(Node* kls, RegionNode* region) { 17421308Sache return generate_array_guard_common(kls, region, true, false); 17521308Sache } 17621308Sache Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 17721308Sache return generate_array_guard_common(kls, region, true, true); 17821308Sache } 17921308Sache Node* generate_array_guard_common(Node* kls, RegionNode* region, 18021308Sache bool obj_array, bool not_array); 18121308Sache Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 18221308Sache CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 18321308Sache bool is_virtual = false, bool is_static = false); 18421308Sache CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 18521308Sache return generate_method_call(method_id, false, true); 18621308Sache } 18721308Sache CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 18821308Sache return generate_method_call(method_id, true, false); 18975406Sache } 19075406Sache Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 19121308Sache 19221308Sache Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 19375406Sache Node* make_string_method_node(int opcode, Node* str1, Node* str2); 19421308Sache bool inline_string_compareTo(); 19521308Sache bool inline_string_indexOf(); 19621308Sache Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 19721308Sache bool inline_string_equals(); 19821308Sache Node* round_double_node(Node* n); 19921308Sache bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 20021308Sache bool inline_math_native(vmIntrinsics::ID id); 20121308Sache bool inline_trig(vmIntrinsics::ID id); 20221308Sache bool inline_math(vmIntrinsics::ID id); 20321308Sache bool inline_math_mathExact(Node* math); 20421308Sache bool inline_math_addExact(); 20521308Sache bool inline_exp(); 206119610Sache bool inline_pow(); 20721308Sache void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 20821308Sache bool inline_min_max(vmIntrinsics::ID id); 20921308Sache Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 21021308Sache // This returns Type::AnyPtr, RawPtr, or OopPtr. 21121308Sache int classify_unsafe_addr(Node* &base, Node* &offset); 21221308Sache Node* make_unsafe_address(Node* base, Node* offset); 21321308Sache // Helper for inline_unsafe_access. 21421308Sache // Generates the guards that check whether the result of 21521308Sache // Unsafe.getObject should be recorded in an SATB log buffer. 21621308Sache void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 21721308Sache bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 21821308Sache bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 21921308Sache static bool klass_needs_init_guard(Node* kls); 22075406Sache bool inline_unsafe_allocate(); 22121308Sache bool inline_unsafe_copyMemory(); 22221308Sache bool inline_native_currentThread(); 22321308Sache#ifdef TRACE_HAVE_INTRINSICS 22421308Sache bool inline_native_classID(); 22521308Sache bool inline_native_threadID(); 22621308Sache#endif 22721308Sache bool inline_native_time_funcs(address method, const char* funcName); 22821308Sache bool inline_native_isInterrupted(); 22921308Sache bool inline_native_Class_query(vmIntrinsics::ID id); 23021308Sache bool inline_native_subtype_check(); 231119610Sache 23221308Sache bool inline_native_newArray(); 23321308Sache bool inline_native_getLength(); 23421308Sache bool inline_array_copyOf(bool is_copyOfRange); 23521308Sache bool inline_array_equals(); 23621308Sache void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 23721308Sache bool inline_native_clone(bool is_virtual); 23821308Sache bool inline_native_Reflection_getCallerClass(); 23921308Sache // Helper function for inlining native object hash method 24075406Sache bool inline_native_hashcode(bool is_virtual, bool is_static); 24121308Sache bool inline_native_getClass(); 24221308Sache 24321308Sache // Helper functions for inlining arraycopy 24421308Sache bool inline_arraycopy(); 24521308Sache void generate_arraycopy(const TypePtr* adr_type, 24621308Sache BasicType basic_elem_type, 24721308Sache Node* src, Node* src_offset, 24821308Sache Node* dest, Node* dest_offset, 24921308Sache Node* copy_length, 25021308Sache bool disjoint_bases = false, 25121308Sache bool length_never_negative = false, 25221308Sache RegionNode* slow_region = NULL); 25321308Sache AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 254119610Sache RegionNode* slow_region); 25521308Sache void generate_clear_array(const TypePtr* adr_type, 25675406Sache Node* dest, 25721308Sache BasicType basic_elem_type, 25821308Sache Node* slice_off, 25921308Sache Node* slice_len, 26021308Sache Node* slice_end); 26121308Sache bool generate_block_arraycopy(const TypePtr* adr_type, 26221308Sache BasicType basic_elem_type, 26321308Sache AllocateNode* alloc, 26421308Sache Node* src, Node* src_offset, 26521308Sache Node* dest, Node* dest_offset, 26621308Sache Node* dest_size, bool dest_uninitialized); 26721308Sache void generate_slow_arraycopy(const TypePtr* adr_type, 26821308Sache Node* src, Node* src_offset, 26921308Sache Node* dest, Node* dest_offset, 27021308Sache Node* copy_length, bool dest_uninitialized); 27121308Sache Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 272 Node* dest_elem_klass, 273 Node* src, Node* src_offset, 274 Node* dest, Node* dest_offset, 275 Node* copy_length, bool dest_uninitialized); 276 Node* generate_generic_arraycopy(const TypePtr* adr_type, 277 Node* src, Node* src_offset, 278 Node* dest, Node* dest_offset, 279 Node* copy_length, bool dest_uninitialized); 280 void generate_unchecked_arraycopy(const TypePtr* adr_type, 281 BasicType basic_elem_type, 282 bool disjoint_bases, 283 Node* src, Node* src_offset, 284 Node* dest, Node* dest_offset, 285 Node* copy_length, bool dest_uninitialized); 286 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 287 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 288 bool inline_unsafe_ordered_store(BasicType type); 289 bool inline_unsafe_fence(vmIntrinsics::ID id); 290 bool inline_fp_conversions(vmIntrinsics::ID id); 291 bool inline_number_methods(vmIntrinsics::ID id); 292 bool inline_reference_get(); 293 bool inline_aescrypt_Block(vmIntrinsics::ID id); 294 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 295 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 296 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 297 bool inline_encodeISOArray(); 298 bool inline_updateCRC32(); 299 bool inline_updateBytesCRC32(); 300 bool inline_updateByteBufferCRC32(); 301}; 302 303 304//---------------------------make_vm_intrinsic---------------------------- 305CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 306 vmIntrinsics::ID id = m->intrinsic_id(); 307 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 308 309 if (DisableIntrinsic[0] != '\0' 310 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 311 // disabled by a user request on the command line: 312 // example: -XX:DisableIntrinsic=_hashCode,_getClass 313 return NULL; 314 } 315 316 if (!m->is_loaded()) { 317 // do not attempt to inline unloaded methods 318 return NULL; 319 } 320 321 // Only a few intrinsics implement a virtual dispatch. 322 // They are expensive calls which are also frequently overridden. 323 if (is_virtual) { 324 switch (id) { 325 case vmIntrinsics::_hashCode: 326 case vmIntrinsics::_clone: 327 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 328 break; 329 default: 330 return NULL; 331 } 332 } 333 334 // -XX:-InlineNatives disables nearly all intrinsics: 335 if (!InlineNatives) { 336 switch (id) { 337 case vmIntrinsics::_indexOf: 338 case vmIntrinsics::_compareTo: 339 case vmIntrinsics::_equals: 340 case vmIntrinsics::_equalsC: 341 case vmIntrinsics::_getAndAddInt: 342 case vmIntrinsics::_getAndAddLong: 343 case vmIntrinsics::_getAndSetInt: 344 case vmIntrinsics::_getAndSetLong: 345 case vmIntrinsics::_getAndSetObject: 346 case vmIntrinsics::_loadFence: 347 case vmIntrinsics::_storeFence: 348 case vmIntrinsics::_fullFence: 349 break; // InlineNatives does not control String.compareTo 350 case vmIntrinsics::_Reference_get: 351 break; // InlineNatives does not control Reference.get 352 default: 353 return NULL; 354 } 355 } 356 357 bool is_predicted = false; 358 359 switch (id) { 360 case vmIntrinsics::_compareTo: 361 if (!SpecialStringCompareTo) return NULL; 362 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 363 break; 364 case vmIntrinsics::_indexOf: 365 if (!SpecialStringIndexOf) return NULL; 366 break; 367 case vmIntrinsics::_equals: 368 if (!SpecialStringEquals) return NULL; 369 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 370 break; 371 case vmIntrinsics::_equalsC: 372 if (!SpecialArraysEquals) return NULL; 373 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 374 break; 375 case vmIntrinsics::_arraycopy: 376 if (!InlineArrayCopy) return NULL; 377 break; 378 case vmIntrinsics::_copyMemory: 379 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 380 if (!InlineArrayCopy) return NULL; 381 break; 382 case vmIntrinsics::_hashCode: 383 if (!InlineObjectHash) return NULL; 384 break; 385 case vmIntrinsics::_clone: 386 case vmIntrinsics::_copyOf: 387 case vmIntrinsics::_copyOfRange: 388 if (!InlineObjectCopy) return NULL; 389 // These also use the arraycopy intrinsic mechanism: 390 if (!InlineArrayCopy) return NULL; 391 break; 392 case vmIntrinsics::_encodeISOArray: 393 if (!SpecialEncodeISOArray) return NULL; 394 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 395 break; 396 case vmIntrinsics::_checkIndex: 397 // We do not intrinsify this. The optimizer does fine with it. 398 return NULL; 399 400 case vmIntrinsics::_getCallerClass: 401 if (!UseNewReflection) return NULL; 402 if (!InlineReflectionGetCallerClass) return NULL; 403 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 404 break; 405 406 case vmIntrinsics::_bitCount_i: 407 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 408 break; 409 410 case vmIntrinsics::_bitCount_l: 411 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 412 break; 413 414 case vmIntrinsics::_numberOfLeadingZeros_i: 415 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 416 break; 417 418 case vmIntrinsics::_numberOfLeadingZeros_l: 419 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 420 break; 421 422 case vmIntrinsics::_numberOfTrailingZeros_i: 423 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 424 break; 425 426 case vmIntrinsics::_numberOfTrailingZeros_l: 427 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 428 break; 429 430 case vmIntrinsics::_reverseBytes_c: 431 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 432 break; 433 case vmIntrinsics::_reverseBytes_s: 434 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 435 break; 436 case vmIntrinsics::_reverseBytes_i: 437 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 438 break; 439 case vmIntrinsics::_reverseBytes_l: 440 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 441 break; 442 443 case vmIntrinsics::_Reference_get: 444 // Use the intrinsic version of Reference.get() so that the value in 445 // the referent field can be registered by the G1 pre-barrier code. 446 // Also add memory barrier to prevent commoning reads from this field 447 // across safepoint since GC can change it value. 448 break; 449 450 case vmIntrinsics::_compareAndSwapObject: 451#ifdef _LP64 452 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 453#endif 454 break; 455 456 case vmIntrinsics::_compareAndSwapLong: 457 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 458 break; 459 460 case vmIntrinsics::_getAndAddInt: 461 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 462 break; 463 464 case vmIntrinsics::_getAndAddLong: 465 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 466 break; 467 468 case vmIntrinsics::_getAndSetInt: 469 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 470 break; 471 472 case vmIntrinsics::_getAndSetLong: 473 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 474 break; 475 476 case vmIntrinsics::_getAndSetObject: 477#ifdef _LP64 478 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 479 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 480 break; 481#else 482 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 483 break; 484#endif 485 486 case vmIntrinsics::_aescrypt_encryptBlock: 487 case vmIntrinsics::_aescrypt_decryptBlock: 488 if (!UseAESIntrinsics) return NULL; 489 break; 490 491 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 492 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 493 if (!UseAESIntrinsics) return NULL; 494 // these two require the predicated logic 495 is_predicted = true; 496 break; 497 498 case vmIntrinsics::_updateCRC32: 499 case vmIntrinsics::_updateBytesCRC32: 500 case vmIntrinsics::_updateByteBufferCRC32: 501 if (!UseCRC32Intrinsics) return NULL; 502 break; 503 504 case vmIntrinsics::_addExact: 505 if (!Matcher::match_rule_supported(Op_AddExactI)) { 506 return NULL; 507 } 508 if (!UseMathExactIntrinsics) { 509 return NULL; 510 } 511 break; 512 513 default: 514 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 515 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 516 break; 517 } 518 519 // -XX:-InlineClassNatives disables natives from the Class class. 520 // The flag applies to all reflective calls, notably Array.newArray 521 // (visible to Java programmers as Array.newInstance). 522 if (m->holder()->name() == ciSymbol::java_lang_Class() || 523 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 524 if (!InlineClassNatives) return NULL; 525 } 526 527 // -XX:-InlineThreadNatives disables natives from the Thread class. 528 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 529 if (!InlineThreadNatives) return NULL; 530 } 531 532 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 533 if (m->holder()->name() == ciSymbol::java_lang_Math() || 534 m->holder()->name() == ciSymbol::java_lang_Float() || 535 m->holder()->name() == ciSymbol::java_lang_Double()) { 536 if (!InlineMathNatives) return NULL; 537 } 538 539 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 540 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 541 if (!InlineUnsafeOps) return NULL; 542 } 543 544 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id); 545} 546 547//----------------------register_library_intrinsics----------------------- 548// Initialize this file's data structures, for each Compile instance. 549void Compile::register_library_intrinsics() { 550 // Nothing to do here. 551} 552 553JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 554 LibraryCallKit kit(jvms, this); 555 Compile* C = kit.C; 556 int nodes = C->unique(); 557#ifndef PRODUCT 558 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 559 char buf[1000]; 560 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 561 tty->print_cr("Intrinsic %s", str); 562 } 563#endif 564 ciMethod* callee = kit.callee(); 565 const int bci = kit.bci(); 566 567 // Try to inline the intrinsic. 568 if (kit.try_to_inline()) { 569 if (C->print_intrinsics() || C->print_inlining()) { 570 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 571 } 572 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 573 if (C->log()) { 574 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 575 vmIntrinsics::name_at(intrinsic_id()), 576 (is_virtual() ? " virtual='1'" : ""), 577 C->unique() - nodes); 578 } 579 // Push the result from the inlined method onto the stack. 580 kit.push_result(); 581 return kit.transfer_exceptions_into_jvms(); 582 } 583 584 // The intrinsic bailed out 585 if (C->print_intrinsics() || C->print_inlining()) { 586 if (jvms->has_method()) { 587 // Not a root compile. 588 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 589 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 590 } else { 591 // Root compile 592 tty->print("Did not generate intrinsic %s%s at bci:%d in", 593 vmIntrinsics::name_at(intrinsic_id()), 594 (is_virtual() ? " (virtual)" : ""), bci); 595 } 596 } 597 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 598 return NULL; 599} 600 601Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) { 602 LibraryCallKit kit(jvms, this); 603 Compile* C = kit.C; 604 int nodes = C->unique(); 605#ifndef PRODUCT 606 assert(is_predicted(), "sanity"); 607 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 608 char buf[1000]; 609 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 610 tty->print_cr("Predicate for intrinsic %s", str); 611 } 612#endif 613 ciMethod* callee = kit.callee(); 614 const int bci = kit.bci(); 615 616 Node* slow_ctl = kit.try_to_predicate(); 617 if (!kit.failing()) { 618 if (C->print_intrinsics() || C->print_inlining()) { 619 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 620 } 621 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 622 if (C->log()) { 623 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 624 vmIntrinsics::name_at(intrinsic_id()), 625 (is_virtual() ? " virtual='1'" : ""), 626 C->unique() - nodes); 627 } 628 return slow_ctl; // Could be NULL if the check folds. 629 } 630 631 // The intrinsic bailed out 632 if (C->print_intrinsics() || C->print_inlining()) { 633 if (jvms->has_method()) { 634 // Not a root compile. 635 const char* msg = "failed to generate predicate for intrinsic"; 636 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 637 } else { 638 // Root compile 639 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 640 vmIntrinsics::name_at(intrinsic_id()), 641 (is_virtual() ? " (virtual)" : ""), bci); 642 } 643 } 644 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 645 return NULL; 646} 647 648bool LibraryCallKit::try_to_inline() { 649 // Handle symbolic names for otherwise undistinguished boolean switches: 650 const bool is_store = true; 651 const bool is_native_ptr = true; 652 const bool is_static = true; 653 const bool is_volatile = true; 654 655 if (!jvms()->has_method()) { 656 // Root JVMState has a null method. 657 assert(map()->memory()->Opcode() == Op_Parm, ""); 658 // Insert the memory aliasing node 659 set_all_memory(reset_memory()); 660 } 661 assert(merged_memory(), ""); 662 663 664 switch (intrinsic_id()) { 665 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 666 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 667 case vmIntrinsics::_getClass: return inline_native_getClass(); 668 669 case vmIntrinsics::_dsin: 670 case vmIntrinsics::_dcos: 671 case vmIntrinsics::_dtan: 672 case vmIntrinsics::_dabs: 673 case vmIntrinsics::_datan2: 674 case vmIntrinsics::_dsqrt: 675 case vmIntrinsics::_dexp: 676 case vmIntrinsics::_dlog: 677 case vmIntrinsics::_dlog10: 678 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 679 680 case vmIntrinsics::_min: 681 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 682 683 case vmIntrinsics::_addExact: return inline_math_addExact(); 684 685 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 686 687 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 688 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 689 case vmIntrinsics::_equals: return inline_string_equals(); 690 691 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 692 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 693 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 694 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 695 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 696 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 697 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 698 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 699 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 700 701 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 702 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 703 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 704 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 705 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 706 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 707 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 708 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 709 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 710 711 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 712 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 713 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 714 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 715 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 716 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 717 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 718 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 719 720 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 721 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 722 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 723 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 724 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 725 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 726 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 727 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 728 729 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 730 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 731 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 732 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 733 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 734 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 735 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 736 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 737 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 738 739 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 740 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 741 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 742 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 743 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 744 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 745 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 746 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 747 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 748 749 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 750 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 751 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 752 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 753 754 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 755 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 756 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 757 758 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 759 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 760 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 761 762 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 763 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 764 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 765 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 766 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 767 768 case vmIntrinsics::_loadFence: 769 case vmIntrinsics::_storeFence: 770 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 771 772 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 773 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 774 775#ifdef TRACE_HAVE_INTRINSICS 776 case vmIntrinsics::_classID: return inline_native_classID(); 777 case vmIntrinsics::_threadID: return inline_native_threadID(); 778 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 779#endif 780 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 781 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 782 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 783 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 784 case vmIntrinsics::_newArray: return inline_native_newArray(); 785 case vmIntrinsics::_getLength: return inline_native_getLength(); 786 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 787 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 788 case vmIntrinsics::_equalsC: return inline_array_equals(); 789 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 790 791 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 792 793 case vmIntrinsics::_isInstance: 794 case vmIntrinsics::_getModifiers: 795 case vmIntrinsics::_isInterface: 796 case vmIntrinsics::_isArray: 797 case vmIntrinsics::_isPrimitive: 798 case vmIntrinsics::_getSuperclass: 799 case vmIntrinsics::_getComponentType: 800 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 801 802 case vmIntrinsics::_floatToRawIntBits: 803 case vmIntrinsics::_floatToIntBits: 804 case vmIntrinsics::_intBitsToFloat: 805 case vmIntrinsics::_doubleToRawLongBits: 806 case vmIntrinsics::_doubleToLongBits: 807 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 808 809 case vmIntrinsics::_numberOfLeadingZeros_i: 810 case vmIntrinsics::_numberOfLeadingZeros_l: 811 case vmIntrinsics::_numberOfTrailingZeros_i: 812 case vmIntrinsics::_numberOfTrailingZeros_l: 813 case vmIntrinsics::_bitCount_i: 814 case vmIntrinsics::_bitCount_l: 815 case vmIntrinsics::_reverseBytes_i: 816 case vmIntrinsics::_reverseBytes_l: 817 case vmIntrinsics::_reverseBytes_s: 818 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 819 820 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 821 822 case vmIntrinsics::_Reference_get: return inline_reference_get(); 823 824 case vmIntrinsics::_aescrypt_encryptBlock: 825 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 826 827 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 828 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 829 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 830 831 case vmIntrinsics::_encodeISOArray: 832 return inline_encodeISOArray(); 833 834 case vmIntrinsics::_updateCRC32: 835 return inline_updateCRC32(); 836 case vmIntrinsics::_updateBytesCRC32: 837 return inline_updateBytesCRC32(); 838 case vmIntrinsics::_updateByteBufferCRC32: 839 return inline_updateByteBufferCRC32(); 840 841 default: 842 // If you get here, it may be that someone has added a new intrinsic 843 // to the list in vmSymbols.hpp without implementing it here. 844#ifndef PRODUCT 845 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 846 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 847 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 848 } 849#endif 850 return false; 851 } 852} 853 854Node* LibraryCallKit::try_to_predicate() { 855 if (!jvms()->has_method()) { 856 // Root JVMState has a null method. 857 assert(map()->memory()->Opcode() == Op_Parm, ""); 858 // Insert the memory aliasing node 859 set_all_memory(reset_memory()); 860 } 861 assert(merged_memory(), ""); 862 863 switch (intrinsic_id()) { 864 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 865 return inline_cipherBlockChaining_AESCrypt_predicate(false); 866 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 867 return inline_cipherBlockChaining_AESCrypt_predicate(true); 868 869 default: 870 // If you get here, it may be that someone has added a new intrinsic 871 // to the list in vmSymbols.hpp without implementing it here. 872#ifndef PRODUCT 873 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 874 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 875 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 876 } 877#endif 878 Node* slow_ctl = control(); 879 set_control(top()); // No fast path instrinsic 880 return slow_ctl; 881 } 882} 883 884//------------------------------set_result------------------------------- 885// Helper function for finishing intrinsics. 886void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 887 record_for_igvn(region); 888 set_control(_gvn.transform(region)); 889 set_result( _gvn.transform(value)); 890 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 891} 892 893//------------------------------generate_guard--------------------------- 894// Helper function for generating guarded fast-slow graph structures. 895// The given 'test', if true, guards a slow path. If the test fails 896// then a fast path can be taken. (We generally hope it fails.) 897// In all cases, GraphKit::control() is updated to the fast path. 898// The returned value represents the control for the slow path. 899// The return value is never 'top'; it is either a valid control 900// or NULL if it is obvious that the slow path can never be taken. 901// Also, if region and the slow control are not NULL, the slow edge 902// is appended to the region. 903Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 904 if (stopped()) { 905 // Already short circuited. 906 return NULL; 907 } 908 909 // Build an if node and its projections. 910 // If test is true we take the slow path, which we assume is uncommon. 911 if (_gvn.type(test) == TypeInt::ZERO) { 912 // The slow branch is never taken. No need to build this guard. 913 return NULL; 914 } 915 916 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 917 918 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff)); 919 if (if_slow == top()) { 920 // The slow branch is never taken. No need to build this guard. 921 return NULL; 922 } 923 924 if (region != NULL) 925 region->add_req(if_slow); 926 927 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff)); 928 set_control(if_fast); 929 930 return if_slow; 931} 932 933inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 934 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 935} 936inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 937 return generate_guard(test, region, PROB_FAIR); 938} 939 940inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 941 Node* *pos_index) { 942 if (stopped()) 943 return NULL; // already stopped 944 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 945 return NULL; // index is already adequately typed 946 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0))); 947 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 948 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 949 if (is_neg != NULL && pos_index != NULL) { 950 // Emulate effect of Parse::adjust_map_after_if. 951 Node* ccast = new (C) CastIINode(index, TypeInt::POS); 952 ccast->set_req(0, control()); 953 (*pos_index) = _gvn.transform(ccast); 954 } 955 return is_neg; 956} 957 958inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 959 Node* *pos_index) { 960 if (stopped()) 961 return NULL; // already stopped 962 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 963 return NULL; // index is already adequately typed 964 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0))); 965 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 966 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq)); 967 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 968 if (is_notp != NULL && pos_index != NULL) { 969 // Emulate effect of Parse::adjust_map_after_if. 970 Node* ccast = new (C) CastIINode(index, TypeInt::POS1); 971 ccast->set_req(0, control()); 972 (*pos_index) = _gvn.transform(ccast); 973 } 974 return is_notp; 975} 976 977// Make sure that 'position' is a valid limit index, in [0..length]. 978// There are two equivalent plans for checking this: 979// A. (offset + copyLength) unsigned<= arrayLength 980// B. offset <= (arrayLength - copyLength) 981// We require that all of the values above, except for the sum and 982// difference, are already known to be non-negative. 983// Plan A is robust in the face of overflow, if offset and copyLength 984// are both hugely positive. 985// 986// Plan B is less direct and intuitive, but it does not overflow at 987// all, since the difference of two non-negatives is always 988// representable. Whenever Java methods must perform the equivalent 989// check they generally use Plan B instead of Plan A. 990// For the moment we use Plan A. 991inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 992 Node* subseq_length, 993 Node* array_length, 994 RegionNode* region) { 995 if (stopped()) 996 return NULL; // already stopped 997 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 998 if (zero_offset && subseq_length->eqv_uncast(array_length)) 999 return NULL; // common case of whole-array copy 1000 Node* last = subseq_length; 1001 if (!zero_offset) // last += offset 1002 last = _gvn.transform(new (C) AddINode(last, offset)); 1003 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last)); 1004 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 1005 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1006 return is_over; 1007} 1008 1009 1010//--------------------------generate_current_thread-------------------- 1011Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1012 ciKlass* thread_klass = env()->Thread_klass(); 1013 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1014 Node* thread = _gvn.transform(new (C) ThreadLocalNode()); 1015 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1016 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 1017 tls_output = thread; 1018 return threadObj; 1019} 1020 1021 1022//------------------------------make_string_method_node------------------------ 1023// Helper method for String intrinsic functions. This version is called 1024// with str1 and str2 pointing to String object nodes. 1025// 1026Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1027 Node* no_ctrl = NULL; 1028 1029 // Get start addr of string 1030 Node* str1_value = load_String_value(no_ctrl, str1); 1031 Node* str1_offset = load_String_offset(no_ctrl, str1); 1032 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1033 1034 // Get length of string 1 1035 Node* str1_len = load_String_length(no_ctrl, str1); 1036 1037 Node* str2_value = load_String_value(no_ctrl, str2); 1038 Node* str2_offset = load_String_offset(no_ctrl, str2); 1039 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1040 1041 Node* str2_len = NULL; 1042 Node* result = NULL; 1043 1044 switch (opcode) { 1045 case Op_StrIndexOf: 1046 // Get length of string 2 1047 str2_len = load_String_length(no_ctrl, str2); 1048 1049 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1050 str1_start, str1_len, str2_start, str2_len); 1051 break; 1052 case Op_StrComp: 1053 // Get length of string 2 1054 str2_len = load_String_length(no_ctrl, str2); 1055 1056 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1057 str1_start, str1_len, str2_start, str2_len); 1058 break; 1059 case Op_StrEquals: 1060 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1061 str1_start, str2_start, str1_len); 1062 break; 1063 default: 1064 ShouldNotReachHere(); 1065 return NULL; 1066 } 1067 1068 // All these intrinsics have checks. 1069 C->set_has_split_ifs(true); // Has chance for split-if optimization 1070 1071 return _gvn.transform(result); 1072} 1073 1074// Helper method for String intrinsic functions. This version is called 1075// with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1076// to Int nodes containing the lenghts of str1 and str2. 1077// 1078Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1079 Node* result = NULL; 1080 switch (opcode) { 1081 case Op_StrIndexOf: 1082 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1083 str1_start, cnt1, str2_start, cnt2); 1084 break; 1085 case Op_StrComp: 1086 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1087 str1_start, cnt1, str2_start, cnt2); 1088 break; 1089 case Op_StrEquals: 1090 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1091 str1_start, str2_start, cnt1); 1092 break; 1093 default: 1094 ShouldNotReachHere(); 1095 return NULL; 1096 } 1097 1098 // All these intrinsics have checks. 1099 C->set_has_split_ifs(true); // Has chance for split-if optimization 1100 1101 return _gvn.transform(result); 1102} 1103 1104//------------------------------inline_string_compareTo------------------------ 1105// public int java.lang.String.compareTo(String anotherString); 1106bool LibraryCallKit::inline_string_compareTo() { 1107 Node* receiver = null_check(argument(0)); 1108 Node* arg = null_check(argument(1)); 1109 if (stopped()) { 1110 return true; 1111 } 1112 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1113 return true; 1114} 1115 1116//------------------------------inline_string_equals------------------------ 1117bool LibraryCallKit::inline_string_equals() { 1118 Node* receiver = null_check_receiver(); 1119 // NOTE: Do not null check argument for String.equals() because spec 1120 // allows to specify NULL as argument. 1121 Node* argument = this->argument(1); 1122 if (stopped()) { 1123 return true; 1124 } 1125 1126 // paths (plus control) merge 1127 RegionNode* region = new (C) RegionNode(5); 1128 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 1129 1130 // does source == target string? 1131 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument)); 1132 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 1133 1134 Node* if_eq = generate_slow_guard(bol, NULL); 1135 if (if_eq != NULL) { 1136 // receiver == argument 1137 phi->init_req(2, intcon(1)); 1138 region->init_req(2, if_eq); 1139 } 1140 1141 // get String klass for instanceOf 1142 ciInstanceKlass* klass = env()->String_klass(); 1143 1144 if (!stopped()) { 1145 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1146 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1))); 1147 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 1148 1149 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1150 //instanceOf == true, fallthrough 1151 1152 if (inst_false != NULL) { 1153 phi->init_req(3, intcon(0)); 1154 region->init_req(3, inst_false); 1155 } 1156 } 1157 1158 if (!stopped()) { 1159 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1160 1161 // Properly cast the argument to String 1162 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type)); 1163 // This path is taken only when argument's type is String:NotNull. 1164 argument = cast_not_null(argument, false); 1165 1166 Node* no_ctrl = NULL; 1167 1168 // Get start addr of receiver 1169 Node* receiver_val = load_String_value(no_ctrl, receiver); 1170 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1171 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1172 1173 // Get length of receiver 1174 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1175 1176 // Get start addr of argument 1177 Node* argument_val = load_String_value(no_ctrl, argument); 1178 Node* argument_offset = load_String_offset(no_ctrl, argument); 1179 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1180 1181 // Get length of argument 1182 Node* argument_cnt = load_String_length(no_ctrl, argument); 1183 1184 // Check for receiver count != argument count 1185 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt)); 1186 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne)); 1187 Node* if_ne = generate_slow_guard(bol, NULL); 1188 if (if_ne != NULL) { 1189 phi->init_req(4, intcon(0)); 1190 region->init_req(4, if_ne); 1191 } 1192 1193 // Check for count == 0 is done by assembler code for StrEquals. 1194 1195 if (!stopped()) { 1196 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1197 phi->init_req(1, equals); 1198 region->init_req(1, control()); 1199 } 1200 } 1201 1202 // post merge 1203 set_control(_gvn.transform(region)); 1204 record_for_igvn(region); 1205 1206 set_result(_gvn.transform(phi)); 1207 return true; 1208} 1209 1210//------------------------------inline_array_equals---------------------------- 1211bool LibraryCallKit::inline_array_equals() { 1212 Node* arg1 = argument(0); 1213 Node* arg2 = argument(1); 1214 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1215 return true; 1216} 1217 1218// Java version of String.indexOf(constant string) 1219// class StringDecl { 1220// StringDecl(char[] ca) { 1221// offset = 0; 1222// count = ca.length; 1223// value = ca; 1224// } 1225// int offset; 1226// int count; 1227// char[] value; 1228// } 1229// 1230// static int string_indexOf_J(StringDecl string_object, char[] target_object, 1231// int targetOffset, int cache_i, int md2) { 1232// int cache = cache_i; 1233// int sourceOffset = string_object.offset; 1234// int sourceCount = string_object.count; 1235// int targetCount = target_object.length; 1236// 1237// int targetCountLess1 = targetCount - 1; 1238// int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1239// 1240// char[] source = string_object.value; 1241// char[] target = target_object; 1242// int lastChar = target[targetCountLess1]; 1243// 1244// outer_loop: 1245// for (int i = sourceOffset; i < sourceEnd; ) { 1246// int src = source[i + targetCountLess1]; 1247// if (src == lastChar) { 1248// // With random strings and a 4-character alphabet, 1249// // reverse matching at this point sets up 0.8% fewer 1250// // frames, but (paradoxically) makes 0.3% more probes. 1251// // Since those probes are nearer the lastChar probe, 1252// // there is may be a net D$ win with reverse matching. 1253// // But, reversing loop inhibits unroll of inner loop 1254// // for unknown reason. So, does running outer loop from 1255// // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1256// for (int j = 0; j < targetCountLess1; j++) { 1257// if (target[targetOffset + j] != source[i+j]) { 1258// if ((cache & (1 << source[i+j])) == 0) { 1259// if (md2 < j+1) { 1260// i += j+1; 1261// continue outer_loop; 1262// } 1263// } 1264// i += md2; 1265// continue outer_loop; 1266// } 1267// } 1268// return i - sourceOffset; 1269// } 1270// if ((cache & (1 << src)) == 0) { 1271// i += targetCountLess1; 1272// } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1273// i++; 1274// } 1275// return -1; 1276// } 1277 1278//------------------------------string_indexOf------------------------ 1279Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1280 jint cache_i, jint md2_i) { 1281 1282 Node* no_ctrl = NULL; 1283 float likely = PROB_LIKELY(0.9); 1284 float unlikely = PROB_UNLIKELY(0.9); 1285 1286 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1287 1288 Node* source = load_String_value(no_ctrl, string_object); 1289 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1290 Node* sourceCount = load_String_length(no_ctrl, string_object); 1291 1292 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1293 jint target_length = target_array->length(); 1294 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1295 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1296 1297 // String.value field is known to be @Stable. 1298 if (UseImplicitStableValues) { 1299 target = cast_array_to_stable(target, target_type); 1300 } 1301 1302 IdealKit kit(this, false, true); 1303#define __ kit. 1304 Node* zero = __ ConI(0); 1305 Node* one = __ ConI(1); 1306 Node* cache = __ ConI(cache_i); 1307 Node* md2 = __ ConI(md2_i); 1308 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1309 Node* targetCount = __ ConI(target_length); 1310 Node* targetCountLess1 = __ ConI(target_length - 1); 1311 Node* targetOffset = __ ConI(targetOffset_i); 1312 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1313 1314 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1315 Node* outer_loop = __ make_label(2 /* goto */); 1316 Node* return_ = __ make_label(1); 1317 1318 __ set(rtn,__ ConI(-1)); 1319 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1320 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1321 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1322 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1323 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1324 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1325 Node* tpj = __ AddI(targetOffset, __ value(j)); 1326 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1327 Node* ipj = __ AddI(__ value(i), __ value(j)); 1328 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1329 __ if_then(targ, BoolTest::ne, src2); { 1330 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1331 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1332 __ increment(i, __ AddI(__ value(j), one)); 1333 __ goto_(outer_loop); 1334 } __ end_if(); __ dead(j); 1335 }__ end_if(); __ dead(j); 1336 __ increment(i, md2); 1337 __ goto_(outer_loop); 1338 }__ end_if(); 1339 __ increment(j, one); 1340 }__ end_loop(); __ dead(j); 1341 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1342 __ goto_(return_); 1343 }__ end_if(); 1344 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1345 __ increment(i, targetCountLess1); 1346 }__ end_if(); 1347 __ increment(i, one); 1348 __ bind(outer_loop); 1349 }__ end_loop(); __ dead(i); 1350 __ bind(return_); 1351 1352 // Final sync IdealKit and GraphKit. 1353 final_sync(kit); 1354 Node* result = __ value(rtn); 1355#undef __ 1356 C->set_has_loops(true); 1357 return result; 1358} 1359 1360//------------------------------inline_string_indexOf------------------------ 1361bool LibraryCallKit::inline_string_indexOf() { 1362 Node* receiver = argument(0); 1363 Node* arg = argument(1); 1364 1365 Node* result; 1366 // Disable the use of pcmpestri until it can be guaranteed that 1367 // the load doesn't cross into the uncommited space. 1368 if (Matcher::has_match_rule(Op_StrIndexOf) && 1369 UseSSE42Intrinsics) { 1370 // Generate SSE4.2 version of indexOf 1371 // We currently only have match rules that use SSE4.2 1372 1373 receiver = null_check(receiver); 1374 arg = null_check(arg); 1375 if (stopped()) { 1376 return true; 1377 } 1378 1379 ciInstanceKlass* str_klass = env()->String_klass(); 1380 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1381 1382 // Make the merge point 1383 RegionNode* result_rgn = new (C) RegionNode(4); 1384 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT); 1385 Node* no_ctrl = NULL; 1386 1387 // Get start addr of source string 1388 Node* source = load_String_value(no_ctrl, receiver); 1389 Node* source_offset = load_String_offset(no_ctrl, receiver); 1390 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1391 1392 // Get length of source string 1393 Node* source_cnt = load_String_length(no_ctrl, receiver); 1394 1395 // Get start addr of substring 1396 Node* substr = load_String_value(no_ctrl, arg); 1397 Node* substr_offset = load_String_offset(no_ctrl, arg); 1398 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1399 1400 // Get length of source string 1401 Node* substr_cnt = load_String_length(no_ctrl, arg); 1402 1403 // Check for substr count > string count 1404 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt)); 1405 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt)); 1406 Node* if_gt = generate_slow_guard(bol, NULL); 1407 if (if_gt != NULL) { 1408 result_phi->init_req(2, intcon(-1)); 1409 result_rgn->init_req(2, if_gt); 1410 } 1411 1412 if (!stopped()) { 1413 // Check for substr count == 0 1414 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0))); 1415 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 1416 Node* if_zero = generate_slow_guard(bol, NULL); 1417 if (if_zero != NULL) { 1418 result_phi->init_req(3, intcon(0)); 1419 result_rgn->init_req(3, if_zero); 1420 } 1421 } 1422 1423 if (!stopped()) { 1424 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1425 result_phi->init_req(1, result); 1426 result_rgn->init_req(1, control()); 1427 } 1428 set_control(_gvn.transform(result_rgn)); 1429 record_for_igvn(result_rgn); 1430 result = _gvn.transform(result_phi); 1431 1432 } else { // Use LibraryCallKit::string_indexOf 1433 // don't intrinsify if argument isn't a constant string. 1434 if (!arg->is_Con()) { 1435 return false; 1436 } 1437 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1438 if (str_type == NULL) { 1439 return false; 1440 } 1441 ciInstanceKlass* klass = env()->String_klass(); 1442 ciObject* str_const = str_type->const_oop(); 1443 if (str_const == NULL || str_const->klass() != klass) { 1444 return false; 1445 } 1446 ciInstance* str = str_const->as_instance(); 1447 assert(str != NULL, "must be instance"); 1448 1449 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1450 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1451 1452 int o; 1453 int c; 1454 if (java_lang_String::has_offset_field()) { 1455 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1456 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1457 } else { 1458 o = 0; 1459 c = pat->length(); 1460 } 1461 1462 // constant strings have no offset and count == length which 1463 // simplifies the resulting code somewhat so lets optimize for that. 1464 if (o != 0 || c != pat->length()) { 1465 return false; 1466 } 1467 1468 receiver = null_check(receiver, T_OBJECT); 1469 // NOTE: No null check on the argument is needed since it's a constant String oop. 1470 if (stopped()) { 1471 return true; 1472 } 1473 1474 // The null string as a pattern always returns 0 (match at beginning of string) 1475 if (c == 0) { 1476 set_result(intcon(0)); 1477 return true; 1478 } 1479 1480 // Generate default indexOf 1481 jchar lastChar = pat->char_at(o + (c - 1)); 1482 int cache = 0; 1483 int i; 1484 for (i = 0; i < c - 1; i++) { 1485 assert(i < pat->length(), "out of range"); 1486 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1487 } 1488 1489 int md2 = c; 1490 for (i = 0; i < c - 1; i++) { 1491 assert(i < pat->length(), "out of range"); 1492 if (pat->char_at(o + i) == lastChar) { 1493 md2 = (c - 1) - i; 1494 } 1495 } 1496 1497 result = string_indexOf(receiver, pat, o, cache, md2); 1498 } 1499 set_result(result); 1500 return true; 1501} 1502 1503//--------------------------round_double_node-------------------------------- 1504// Round a double node if necessary. 1505Node* LibraryCallKit::round_double_node(Node* n) { 1506 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1507 n = _gvn.transform(new (C) RoundDoubleNode(0, n)); 1508 return n; 1509} 1510 1511//------------------------------inline_math----------------------------------- 1512// public static double Math.abs(double) 1513// public static double Math.sqrt(double) 1514// public static double Math.log(double) 1515// public static double Math.log10(double) 1516bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1517 Node* arg = round_double_node(argument(0)); 1518 Node* n; 1519 switch (id) { 1520 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break; 1521 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break; 1522 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break; 1523 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break; 1524 default: fatal_unexpected_iid(id); break; 1525 } 1526 set_result(_gvn.transform(n)); 1527 return true; 1528} 1529 1530//------------------------------inline_trig---------------------------------- 1531// Inline sin/cos/tan instructions, if possible. If rounding is required, do 1532// argument reduction which will turn into a fast/slow diamond. 1533bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1534 Node* arg = round_double_node(argument(0)); 1535 Node* n = NULL; 1536 1537 switch (id) { 1538 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break; 1539 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break; 1540 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break; 1541 default: fatal_unexpected_iid(id); break; 1542 } 1543 n = _gvn.transform(n); 1544 1545 // Rounding required? Check for argument reduction! 1546 if (Matcher::strict_fp_requires_explicit_rounding) { 1547 static const double pi_4 = 0.7853981633974483; 1548 static const double neg_pi_4 = -0.7853981633974483; 1549 // pi/2 in 80-bit extended precision 1550 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1551 // -pi/2 in 80-bit extended precision 1552 // 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}; 1553 // Cutoff value for using this argument reduction technique 1554 //static const double pi_2_minus_epsilon = 1.564660403643354; 1555 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1556 1557 // Pseudocode for sin: 1558 // if (x <= Math.PI / 4.0) { 1559 // if (x >= -Math.PI / 4.0) return fsin(x); 1560 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1561 // } else { 1562 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1563 // } 1564 // return StrictMath.sin(x); 1565 1566 // Pseudocode for cos: 1567 // if (x <= Math.PI / 4.0) { 1568 // if (x >= -Math.PI / 4.0) return fcos(x); 1569 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1570 // } else { 1571 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1572 // } 1573 // return StrictMath.cos(x); 1574 1575 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1576 // requires a special machine instruction to load it. Instead we'll try 1577 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1578 // probably do the math inside the SIN encoding. 1579 1580 // Make the merge point 1581 RegionNode* r = new (C) RegionNode(3); 1582 Node* phi = new (C) PhiNode(r, Type::DOUBLE); 1583 1584 // Flatten arg so we need only 1 test 1585 Node *abs = _gvn.transform(new (C) AbsDNode(arg)); 1586 // Node for PI/4 constant 1587 Node *pi4 = makecon(TypeD::make(pi_4)); 1588 // Check PI/4 : abs(arg) 1589 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs)); 1590 // Check: If PI/4 < abs(arg) then go slow 1591 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt )); 1592 // Branch either way 1593 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1594 set_control(opt_iff(r,iff)); 1595 1596 // Set fast path result 1597 phi->init_req(2, n); 1598 1599 // Slow path - non-blocking leaf call 1600 Node* call = NULL; 1601 switch (id) { 1602 case vmIntrinsics::_dsin: 1603 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1604 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1605 "Sin", NULL, arg, top()); 1606 break; 1607 case vmIntrinsics::_dcos: 1608 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1609 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1610 "Cos", NULL, arg, top()); 1611 break; 1612 case vmIntrinsics::_dtan: 1613 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1614 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1615 "Tan", NULL, arg, top()); 1616 break; 1617 } 1618 assert(control()->in(0) == call, ""); 1619 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1620 r->init_req(1, control()); 1621 phi->init_req(1, slow_result); 1622 1623 // Post-merge 1624 set_control(_gvn.transform(r)); 1625 record_for_igvn(r); 1626 n = _gvn.transform(phi); 1627 1628 C->set_has_split_ifs(true); // Has chance for split-if optimization 1629 } 1630 set_result(n); 1631 return true; 1632} 1633 1634void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1635 //------------------- 1636 //result=(result.isNaN())? funcAddr():result; 1637 // Check: If isNaN() by checking result!=result? then either trap 1638 // or go to runtime 1639 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result)); 1640 // Build the boolean node 1641 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq)); 1642 1643 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1644 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1645 // The pow or exp intrinsic returned a NaN, which requires a call 1646 // to the runtime. Recompile with the runtime call. 1647 uncommon_trap(Deoptimization::Reason_intrinsic, 1648 Deoptimization::Action_make_not_entrant); 1649 } 1650 set_result(result); 1651 } else { 1652 // If this inlining ever returned NaN in the past, we compile a call 1653 // to the runtime to properly handle corner cases 1654 1655 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1656 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff)); 1657 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff)); 1658 1659 if (!if_slow->is_top()) { 1660 RegionNode* result_region = new (C) RegionNode(3); 1661 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE); 1662 1663 result_region->init_req(1, if_fast); 1664 result_val->init_req(1, result); 1665 1666 set_control(if_slow); 1667 1668 const TypePtr* no_memory_effects = NULL; 1669 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1670 no_memory_effects, 1671 x, top(), y, y ? top() : NULL); 1672 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0)); 1673#ifdef ASSERT 1674 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1)); 1675 assert(value_top == top(), "second value must be top"); 1676#endif 1677 1678 result_region->init_req(2, control()); 1679 result_val->init_req(2, value); 1680 set_result(result_region, result_val); 1681 } else { 1682 set_result(result); 1683 } 1684 } 1685} 1686 1687//------------------------------inline_exp------------------------------------- 1688// Inline exp instructions, if possible. The Intel hardware only misses 1689// really odd corner cases (+/- Infinity). Just uncommon-trap them. 1690bool LibraryCallKit::inline_exp() { 1691 Node* arg = round_double_node(argument(0)); 1692 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg)); 1693 1694 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1695 1696 C->set_has_split_ifs(true); // Has chance for split-if optimization 1697 return true; 1698} 1699 1700//------------------------------inline_pow------------------------------------- 1701// Inline power instructions, if possible. 1702bool LibraryCallKit::inline_pow() { 1703 // Pseudocode for pow 1704 // if (x <= 0.0) { 1705 // long longy = (long)y; 1706 // if ((double)longy == y) { // if y is long 1707 // if (y + 1 == y) longy = 0; // huge number: even 1708 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1709 // } else { 1710 // result = NaN; 1711 // } 1712 // } else { 1713 // result = DPow(x,y); 1714 // } 1715 // if (result != result)? { 1716 // result = uncommon_trap() or runtime_call(); 1717 // } 1718 // return result; 1719 1720 Node* x = round_double_node(argument(0)); 1721 Node* y = round_double_node(argument(2)); 1722 1723 Node* result = NULL; 1724 1725 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1726 // Short form: skip the fancy tests and just check for NaN result. 1727 result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1728 } else { 1729 // If this inlining ever returned NaN in the past, include all 1730 // checks + call to the runtime. 1731 1732 // Set the merge point for If node with condition of (x <= 0.0) 1733 // There are four possible paths to region node and phi node 1734 RegionNode *r = new (C) RegionNode(4); 1735 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1736 1737 // Build the first if node: if (x <= 0.0) 1738 // Node for 0 constant 1739 Node *zeronode = makecon(TypeD::ZERO); 1740 // Check x:0 1741 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode)); 1742 // Check: If (x<=0) then go complex path 1743 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le )); 1744 // Branch either way 1745 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1746 // Fast path taken; set region slot 3 1747 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1)); 1748 r->init_req(3,fast_taken); // Capture fast-control 1749 1750 // Fast path not-taken, i.e. slow path 1751 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1)); 1752 1753 // Set fast path result 1754 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1755 phi->init_req(3, fast_result); 1756 1757 // Complex path 1758 // Build the second if node (if y is long) 1759 // Node for (long)y 1760 Node *longy = _gvn.transform(new (C) ConvD2LNode(y)); 1761 // Node for (double)((long) y) 1762 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy)); 1763 // Check (double)((long) y) : y 1764 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y)); 1765 // Check if (y isn't long) then go to slow path 1766 1767 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne )); 1768 // Branch either way 1769 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1770 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2)); 1771 1772 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2)); 1773 1774 // Calculate DPow(abs(x), y)*(1 & (long)y) 1775 // Node for constant 1 1776 Node *conone = longcon(1); 1777 // 1& (long)y 1778 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy)); 1779 1780 // A huge number is always even. Detect a huge number by checking 1781 // if y + 1 == y and set integer to be tested for parity to 0. 1782 // Required for corner case: 1783 // (long)9.223372036854776E18 = max_jlong 1784 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1785 // max_jlong is odd but 9.223372036854776E18 is even 1786 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1)))); 1787 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y)); 1788 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq )); 1789 Node* correctedsign = NULL; 1790 if (ConditionalMoveLimit != 0) { 1791 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1792 } else { 1793 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1794 RegionNode *r = new (C) RegionNode(3); 1795 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 1796 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1))); 1797 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1))); 1798 phi->init_req(1, signnode); 1799 phi->init_req(2, longcon(0)); 1800 correctedsign = _gvn.transform(phi); 1801 ylong_path = _gvn.transform(r); 1802 record_for_igvn(r); 1803 } 1804 1805 // zero node 1806 Node *conzero = longcon(0); 1807 // Check (1&(long)y)==0? 1808 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero)); 1809 // Check if (1&(long)y)!=0?, if so the result is negative 1810 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne )); 1811 // abs(x) 1812 Node *absx=_gvn.transform(new (C) AbsDNode(x)); 1813 // abs(x)^y 1814 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y)); 1815 // -abs(x)^y 1816 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy)); 1817 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1818 Node *signresult = NULL; 1819 if (ConditionalMoveLimit != 0) { 1820 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1821 } else { 1822 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1823 RegionNode *r = new (C) RegionNode(3); 1824 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1825 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven))); 1826 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven))); 1827 phi->init_req(1, absxpowy); 1828 phi->init_req(2, negabsxpowy); 1829 signresult = _gvn.transform(phi); 1830 ylong_path = _gvn.transform(r); 1831 record_for_igvn(r); 1832 } 1833 // Set complex path fast result 1834 r->init_req(2, ylong_path); 1835 phi->init_req(2, signresult); 1836 1837 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1838 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1839 r->init_req(1,slow_path); 1840 phi->init_req(1,slow_result); 1841 1842 // Post merge 1843 set_control(_gvn.transform(r)); 1844 record_for_igvn(r); 1845 result = _gvn.transform(phi); 1846 } 1847 1848 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1849 1850 C->set_has_split_ifs(true); // Has chance for split-if optimization 1851 return true; 1852} 1853 1854//------------------------------runtime_math----------------------------- 1855bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1856 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1857 "must be (DD)D or (D)D type"); 1858 1859 // Inputs 1860 Node* a = round_double_node(argument(0)); 1861 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1862 1863 const TypePtr* no_memory_effects = NULL; 1864 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1865 no_memory_effects, 1866 a, top(), b, b ? top() : NULL); 1867 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0)); 1868#ifdef ASSERT 1869 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1)); 1870 assert(value_top == top(), "second value must be top"); 1871#endif 1872 1873 set_result(value); 1874 return true; 1875} 1876 1877//------------------------------inline_math_native----------------------------- 1878bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1879#define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1880 switch (id) { 1881 // These intrinsics are not properly supported on all hardware 1882 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1883 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1884 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1885 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1886 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1887 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1888 1889 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1890 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1891 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1892 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1893 1894 // These intrinsics are supported on all hardware 1895 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false; 1896 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1897 1898 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1899 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1900 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1901 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1902#undef FN_PTR 1903 1904 // These intrinsics are not yet correctly implemented 1905 case vmIntrinsics::_datan2: 1906 return false; 1907 1908 default: 1909 fatal_unexpected_iid(id); 1910 return false; 1911 } 1912} 1913 1914static bool is_simple_name(Node* n) { 1915 return (n->req() == 1 // constant 1916 || (n->is_Type() && n->as_Type()->type()->singleton()) 1917 || n->is_Proj() // parameter or return value 1918 || n->is_Phi() // local of some sort 1919 ); 1920} 1921 1922//----------------------------inline_min_max----------------------------------- 1923bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1924 set_result(generate_min_max(id, argument(0), argument(1))); 1925 return true; 1926} 1927 1928bool LibraryCallKit::inline_math_mathExact(Node* math) { 1929 Node* result = _gvn.transform( new(C) ProjNode(math, MathExactNode::result_proj_node)); 1930 Node* flags = _gvn.transform( new(C) FlagsProjNode(math, MathExactNode::flags_proj_node)); 1931 1932 Node* bol = _gvn.transform( new (C) BoolNode(flags, BoolTest::overflow) ); 1933 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1934 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check)); 1935 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) ); 1936 1937 { 1938 PreserveJVMState pjvms(this); 1939 PreserveReexecuteState preexecs(this); 1940 jvms()->set_should_reexecute(true); 1941 1942 set_control(slow_path); 1943 set_i_o(i_o()); 1944 1945 uncommon_trap(Deoptimization::Reason_intrinsic, 1946 Deoptimization::Action_none); 1947 } 1948 1949 set_control(fast_path); 1950 set_result(result); 1951 return true; 1952} 1953 1954bool LibraryCallKit::inline_math_addExact() { 1955 Node* arg1 = argument(0); 1956 Node* arg2 = argument(1); 1957 1958 Node* add = _gvn.transform( new(C) AddExactINode(NULL, arg1, arg2) ); 1959 if (add->Opcode() == Op_AddExactI) { 1960 return inline_math_mathExact(add); 1961 } else { 1962 set_result(add); 1963 } 1964 return true; 1965} 1966 1967Node* 1968LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1969 // These are the candidate return value: 1970 Node* xvalue = x0; 1971 Node* yvalue = y0; 1972 1973 if (xvalue == yvalue) { 1974 return xvalue; 1975 } 1976 1977 bool want_max = (id == vmIntrinsics::_max); 1978 1979 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1980 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1981 if (txvalue == NULL || tyvalue == NULL) return top(); 1982 // This is not really necessary, but it is consistent with a 1983 // hypothetical MaxINode::Value method: 1984 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1985 1986 // %%% This folding logic should (ideally) be in a different place. 1987 // Some should be inside IfNode, and there to be a more reliable 1988 // transformation of ?: style patterns into cmoves. We also want 1989 // more powerful optimizations around cmove and min/max. 1990 1991 // Try to find a dominating comparison of these guys. 1992 // It can simplify the index computation for Arrays.copyOf 1993 // and similar uses of System.arraycopy. 1994 // First, compute the normalized version of CmpI(x, y). 1995 int cmp_op = Op_CmpI; 1996 Node* xkey = xvalue; 1997 Node* ykey = yvalue; 1998 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey)); 1999 if (ideal_cmpxy->is_Cmp()) { 2000 // E.g., if we have CmpI(length - offset, count), 2001 // it might idealize to CmpI(length, count + offset) 2002 cmp_op = ideal_cmpxy->Opcode(); 2003 xkey = ideal_cmpxy->in(1); 2004 ykey = ideal_cmpxy->in(2); 2005 } 2006 2007 // Start by locating any relevant comparisons. 2008 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2009 Node* cmpxy = NULL; 2010 Node* cmpyx = NULL; 2011 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2012 Node* cmp = start_from->fast_out(k); 2013 if (cmp->outcnt() > 0 && // must have prior uses 2014 cmp->in(0) == NULL && // must be context-independent 2015 cmp->Opcode() == cmp_op) { // right kind of compare 2016 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2017 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2018 } 2019 } 2020 2021 const int NCMPS = 2; 2022 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2023 int cmpn; 2024 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2025 if (cmps[cmpn] != NULL) break; // find a result 2026 } 2027 if (cmpn < NCMPS) { 2028 // Look for a dominating test that tells us the min and max. 2029 int depth = 0; // Limit search depth for speed 2030 Node* dom = control(); 2031 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2032 if (++depth >= 100) break; 2033 Node* ifproj = dom; 2034 if (!ifproj->is_Proj()) continue; 2035 Node* iff = ifproj->in(0); 2036 if (!iff->is_If()) continue; 2037 Node* bol = iff->in(1); 2038 if (!bol->is_Bool()) continue; 2039 Node* cmp = bol->in(1); 2040 if (cmp == NULL) continue; 2041 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2042 if (cmps[cmpn] == cmp) break; 2043 if (cmpn == NCMPS) continue; 2044 BoolTest::mask btest = bol->as_Bool()->_test._test; 2045 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2046 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2047 // At this point, we know that 'x btest y' is true. 2048 switch (btest) { 2049 case BoolTest::eq: 2050 // They are proven equal, so we can collapse the min/max. 2051 // Either value is the answer. Choose the simpler. 2052 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2053 return yvalue; 2054 return xvalue; 2055 case BoolTest::lt: // x < y 2056 case BoolTest::le: // x <= y 2057 return (want_max ? yvalue : xvalue); 2058 case BoolTest::gt: // x > y 2059 case BoolTest::ge: // x >= y 2060 return (want_max ? xvalue : yvalue); 2061 } 2062 } 2063 } 2064 2065 // We failed to find a dominating test. 2066 // Let's pick a test that might GVN with prior tests. 2067 Node* best_bol = NULL; 2068 BoolTest::mask best_btest = BoolTest::illegal; 2069 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2070 Node* cmp = cmps[cmpn]; 2071 if (cmp == NULL) continue; 2072 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2073 Node* bol = cmp->fast_out(j); 2074 if (!bol->is_Bool()) continue; 2075 BoolTest::mask btest = bol->as_Bool()->_test._test; 2076 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2077 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2078 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2079 best_bol = bol->as_Bool(); 2080 best_btest = btest; 2081 } 2082 } 2083 } 2084 2085 Node* answer_if_true = NULL; 2086 Node* answer_if_false = NULL; 2087 switch (best_btest) { 2088 default: 2089 if (cmpxy == NULL) 2090 cmpxy = ideal_cmpxy; 2091 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt)); 2092 // and fall through: 2093 case BoolTest::lt: // x < y 2094 case BoolTest::le: // x <= y 2095 answer_if_true = (want_max ? yvalue : xvalue); 2096 answer_if_false = (want_max ? xvalue : yvalue); 2097 break; 2098 case BoolTest::gt: // x > y 2099 case BoolTest::ge: // x >= y 2100 answer_if_true = (want_max ? xvalue : yvalue); 2101 answer_if_false = (want_max ? yvalue : xvalue); 2102 break; 2103 } 2104 2105 jint hi, lo; 2106 if (want_max) { 2107 // We can sharpen the minimum. 2108 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2109 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2110 } else { 2111 // We can sharpen the maximum. 2112 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2113 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2114 } 2115 2116 // Use a flow-free graph structure, to avoid creating excess control edges 2117 // which could hinder other optimizations. 2118 // Since Math.min/max is often used with arraycopy, we want 2119 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2120 Node* cmov = CMoveNode::make(C, NULL, best_bol, 2121 answer_if_false, answer_if_true, 2122 TypeInt::make(lo, hi, widen)); 2123 2124 return _gvn.transform(cmov); 2125 2126 /* 2127 // This is not as desirable as it may seem, since Min and Max 2128 // nodes do not have a full set of optimizations. 2129 // And they would interfere, anyway, with 'if' optimizations 2130 // and with CMoveI canonical forms. 2131 switch (id) { 2132 case vmIntrinsics::_min: 2133 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2134 case vmIntrinsics::_max: 2135 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2136 default: 2137 ShouldNotReachHere(); 2138 } 2139 */ 2140} 2141 2142inline int 2143LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2144 const TypePtr* base_type = TypePtr::NULL_PTR; 2145 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2146 if (base_type == NULL) { 2147 // Unknown type. 2148 return Type::AnyPtr; 2149 } else if (base_type == TypePtr::NULL_PTR) { 2150 // Since this is a NULL+long form, we have to switch to a rawptr. 2151 base = _gvn.transform(new (C) CastX2PNode(offset)); 2152 offset = MakeConX(0); 2153 return Type::RawPtr; 2154 } else if (base_type->base() == Type::RawPtr) { 2155 return Type::RawPtr; 2156 } else if (base_type->isa_oopptr()) { 2157 // Base is never null => always a heap address. 2158 if (base_type->ptr() == TypePtr::NotNull) { 2159 return Type::OopPtr; 2160 } 2161 // Offset is small => always a heap address. 2162 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2163 if (offset_type != NULL && 2164 base_type->offset() == 0 && // (should always be?) 2165 offset_type->_lo >= 0 && 2166 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2167 return Type::OopPtr; 2168 } 2169 // Otherwise, it might either be oop+off or NULL+addr. 2170 return Type::AnyPtr; 2171 } else { 2172 // No information: 2173 return Type::AnyPtr; 2174 } 2175} 2176 2177inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2178 int kind = classify_unsafe_addr(base, offset); 2179 if (kind == Type::RawPtr) { 2180 return basic_plus_adr(top(), base, offset); 2181 } else { 2182 return basic_plus_adr(base, offset); 2183 } 2184} 2185 2186//--------------------------inline_number_methods----------------------------- 2187// inline int Integer.numberOfLeadingZeros(int) 2188// inline int Long.numberOfLeadingZeros(long) 2189// 2190// inline int Integer.numberOfTrailingZeros(int) 2191// inline int Long.numberOfTrailingZeros(long) 2192// 2193// inline int Integer.bitCount(int) 2194// inline int Long.bitCount(long) 2195// 2196// inline char Character.reverseBytes(char) 2197// inline short Short.reverseBytes(short) 2198// inline int Integer.reverseBytes(int) 2199// inline long Long.reverseBytes(long) 2200bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2201 Node* arg = argument(0); 2202 Node* n; 2203 switch (id) { 2204 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break; 2205 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break; 2206 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break; 2207 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break; 2208 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break; 2209 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break; 2210 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break; 2211 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break; 2212 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break; 2213 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break; 2214 default: fatal_unexpected_iid(id); break; 2215 } 2216 set_result(_gvn.transform(n)); 2217 return true; 2218} 2219 2220//----------------------------inline_unsafe_access---------------------------- 2221 2222const static BasicType T_ADDRESS_HOLDER = T_LONG; 2223 2224// Helper that guards and inserts a pre-barrier. 2225void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2226 Node* pre_val, bool need_mem_bar) { 2227 // We could be accessing the referent field of a reference object. If so, when G1 2228 // is enabled, we need to log the value in the referent field in an SATB buffer. 2229 // This routine performs some compile time filters and generates suitable 2230 // runtime filters that guard the pre-barrier code. 2231 // Also add memory barrier for non volatile load from the referent field 2232 // to prevent commoning of loads across safepoint. 2233 if (!UseG1GC && !need_mem_bar) 2234 return; 2235 2236 // Some compile time checks. 2237 2238 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2239 const TypeX* otype = offset->find_intptr_t_type(); 2240 if (otype != NULL && otype->is_con() && 2241 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2242 // Constant offset but not the reference_offset so just return 2243 return; 2244 } 2245 2246 // We only need to generate the runtime guards for instances. 2247 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2248 if (btype != NULL) { 2249 if (btype->isa_aryptr()) { 2250 // Array type so nothing to do 2251 return; 2252 } 2253 2254 const TypeInstPtr* itype = btype->isa_instptr(); 2255 if (itype != NULL) { 2256 // Can the klass of base_oop be statically determined to be 2257 // _not_ a sub-class of Reference and _not_ Object? 2258 ciKlass* klass = itype->klass(); 2259 if ( klass->is_loaded() && 2260 !klass->is_subtype_of(env()->Reference_klass()) && 2261 !env()->Object_klass()->is_subtype_of(klass)) { 2262 return; 2263 } 2264 } 2265 } 2266 2267 // The compile time filters did not reject base_oop/offset so 2268 // we need to generate the following runtime filters 2269 // 2270 // if (offset == java_lang_ref_Reference::_reference_offset) { 2271 // if (instance_of(base, java.lang.ref.Reference)) { 2272 // pre_barrier(_, pre_val, ...); 2273 // } 2274 // } 2275 2276 float likely = PROB_LIKELY( 0.999); 2277 float unlikely = PROB_UNLIKELY(0.999); 2278 2279 IdealKit ideal(this); 2280#define __ ideal. 2281 2282 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2283 2284 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2285 // Update graphKit memory and control from IdealKit. 2286 sync_kit(ideal); 2287 2288 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2289 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2290 2291 // Update IdealKit memory and control from graphKit. 2292 __ sync_kit(this); 2293 2294 Node* one = __ ConI(1); 2295 // is_instof == 0 if base_oop == NULL 2296 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2297 2298 // Update graphKit from IdeakKit. 2299 sync_kit(ideal); 2300 2301 // Use the pre-barrier to record the value in the referent field 2302 pre_barrier(false /* do_load */, 2303 __ ctrl(), 2304 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2305 pre_val /* pre_val */, 2306 T_OBJECT); 2307 if (need_mem_bar) { 2308 // Add memory barrier to prevent commoning reads from this field 2309 // across safepoint since GC can change its value. 2310 insert_mem_bar(Op_MemBarCPUOrder); 2311 } 2312 // Update IdealKit from graphKit. 2313 __ sync_kit(this); 2314 2315 } __ end_if(); // _ref_type != ref_none 2316 } __ end_if(); // offset == referent_offset 2317 2318 // Final sync IdealKit and GraphKit. 2319 final_sync(ideal); 2320#undef __ 2321} 2322 2323 2324// Interpret Unsafe.fieldOffset cookies correctly: 2325extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2326 2327const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2328 // Attempt to infer a sharper value type from the offset and base type. 2329 ciKlass* sharpened_klass = NULL; 2330 2331 // See if it is an instance field, with an object type. 2332 if (alias_type->field() != NULL) { 2333 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2334 if (alias_type->field()->type()->is_klass()) { 2335 sharpened_klass = alias_type->field()->type()->as_klass(); 2336 } 2337 } 2338 2339 // See if it is a narrow oop array. 2340 if (adr_type->isa_aryptr()) { 2341 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2342 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2343 if (elem_type != NULL) { 2344 sharpened_klass = elem_type->klass(); 2345 } 2346 } 2347 } 2348 2349 // The sharpened class might be unloaded if there is no class loader 2350 // contraint in place. 2351 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2352 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2353 2354#ifndef PRODUCT 2355 if (C->print_intrinsics() || C->print_inlining()) { 2356 tty->print(" from base type: "); adr_type->dump(); 2357 tty->print(" sharpened value: "); tjp->dump(); 2358 } 2359#endif 2360 // Sharpen the value type. 2361 return tjp; 2362 } 2363 return NULL; 2364} 2365 2366bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2367 if (callee()->is_static()) return false; // caller must have the capability! 2368 2369#ifndef PRODUCT 2370 { 2371 ResourceMark rm; 2372 // Check the signatures. 2373 ciSignature* sig = callee()->signature(); 2374#ifdef ASSERT 2375 if (!is_store) { 2376 // Object getObject(Object base, int/long offset), etc. 2377 BasicType rtype = sig->return_type()->basic_type(); 2378 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2379 rtype = T_ADDRESS; // it is really a C void* 2380 assert(rtype == type, "getter must return the expected value"); 2381 if (!is_native_ptr) { 2382 assert(sig->count() == 2, "oop getter has 2 arguments"); 2383 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2384 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2385 } else { 2386 assert(sig->count() == 1, "native getter has 1 argument"); 2387 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2388 } 2389 } else { 2390 // void putObject(Object base, int/long offset, Object x), etc. 2391 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2392 if (!is_native_ptr) { 2393 assert(sig->count() == 3, "oop putter has 3 arguments"); 2394 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2395 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2396 } else { 2397 assert(sig->count() == 2, "native putter has 2 arguments"); 2398 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2399 } 2400 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2401 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2402 vtype = T_ADDRESS; // it is really a C void* 2403 assert(vtype == type, "putter must accept the expected value"); 2404 } 2405#endif // ASSERT 2406 } 2407#endif //PRODUCT 2408 2409 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2410 2411 Node* receiver = argument(0); // type: oop 2412 2413 // Build address expression. See the code in inline_unsafe_prefetch. 2414 Node* adr; 2415 Node* heap_base_oop = top(); 2416 Node* offset = top(); 2417 Node* val; 2418 2419 if (!is_native_ptr) { 2420 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2421 Node* base = argument(1); // type: oop 2422 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2423 offset = argument(2); // type: long 2424 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2425 // to be plain byte offsets, which are also the same as those accepted 2426 // by oopDesc::field_base. 2427 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2428 "fieldOffset must be byte-scaled"); 2429 // 32-bit machines ignore the high half! 2430 offset = ConvL2X(offset); 2431 adr = make_unsafe_address(base, offset); 2432 heap_base_oop = base; 2433 val = is_store ? argument(4) : NULL; 2434 } else { 2435 Node* ptr = argument(1); // type: long 2436 ptr = ConvL2X(ptr); // adjust Java long to machine word 2437 adr = make_unsafe_address(NULL, ptr); 2438 val = is_store ? argument(3) : NULL; 2439 } 2440 2441 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2442 2443 // First guess at the value type. 2444 const Type *value_type = Type::get_const_basic_type(type); 2445 2446 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2447 // there was not enough information to nail it down. 2448 Compile::AliasType* alias_type = C->alias_type(adr_type); 2449 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2450 2451 // We will need memory barriers unless we can determine a unique 2452 // alias category for this reference. (Note: If for some reason 2453 // the barriers get omitted and the unsafe reference begins to "pollute" 2454 // the alias analysis of the rest of the graph, either Compile::can_alias 2455 // or Compile::must_alias will throw a diagnostic assert.) 2456 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2457 2458 // If we are reading the value of the referent field of a Reference 2459 // object (either by using Unsafe directly or through reflection) 2460 // then, if G1 is enabled, we need to record the referent in an 2461 // SATB log buffer using the pre-barrier mechanism. 2462 // Also we need to add memory barrier to prevent commoning reads 2463 // from this field across safepoint since GC can change its value. 2464 bool need_read_barrier = !is_native_ptr && !is_store && 2465 offset != top() && heap_base_oop != top(); 2466 2467 if (!is_store && type == T_OBJECT) { 2468 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2469 if (tjp != NULL) { 2470 value_type = tjp; 2471 } 2472 } 2473 2474 receiver = null_check(receiver); 2475 if (stopped()) { 2476 return true; 2477 } 2478 // Heap pointers get a null-check from the interpreter, 2479 // as a courtesy. However, this is not guaranteed by Unsafe, 2480 // and it is not possible to fully distinguish unintended nulls 2481 // from intended ones in this API. 2482 2483 if (is_volatile) { 2484 // We need to emit leading and trailing CPU membars (see below) in 2485 // addition to memory membars when is_volatile. This is a little 2486 // too strong, but avoids the need to insert per-alias-type 2487 // volatile membars (for stores; compare Parse::do_put_xxx), which 2488 // we cannot do effectively here because we probably only have a 2489 // rough approximation of type. 2490 need_mem_bar = true; 2491 // For Stores, place a memory ordering barrier now. 2492 if (is_store) 2493 insert_mem_bar(Op_MemBarRelease); 2494 } 2495 2496 // Memory barrier to prevent normal and 'unsafe' accesses from 2497 // bypassing each other. Happens after null checks, so the 2498 // exception paths do not take memory state from the memory barrier, 2499 // so there's no problems making a strong assert about mixing users 2500 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2501 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2502 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2503 2504 if (!is_store) { 2505 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2506 // load value 2507 switch (type) { 2508 case T_BOOLEAN: 2509 case T_CHAR: 2510 case T_BYTE: 2511 case T_SHORT: 2512 case T_INT: 2513 case T_LONG: 2514 case T_FLOAT: 2515 case T_DOUBLE: 2516 break; 2517 case T_OBJECT: 2518 if (need_read_barrier) { 2519 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2520 } 2521 break; 2522 case T_ADDRESS: 2523 // Cast to an int type. 2524 p = _gvn.transform(new (C) CastP2XNode(NULL, p)); 2525 p = ConvX2L(p); 2526 break; 2527 default: 2528 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2529 break; 2530 } 2531 // The load node has the control of the preceding MemBarCPUOrder. All 2532 // following nodes will have the control of the MemBarCPUOrder inserted at 2533 // the end of this method. So, pushing the load onto the stack at a later 2534 // point is fine. 2535 set_result(p); 2536 } else { 2537 // place effect of store into memory 2538 switch (type) { 2539 case T_DOUBLE: 2540 val = dstore_rounding(val); 2541 break; 2542 case T_ADDRESS: 2543 // Repackage the long as a pointer. 2544 val = ConvL2X(val); 2545 val = _gvn.transform(new (C) CastX2PNode(val)); 2546 break; 2547 } 2548 2549 if (type != T_OBJECT ) { 2550 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2551 } else { 2552 // Possibly an oop being stored to Java heap or native memory 2553 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2554 // oop to Java heap. 2555 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2556 } else { 2557 // We can't tell at compile time if we are storing in the Java heap or outside 2558 // of it. So we need to emit code to conditionally do the proper type of 2559 // store. 2560 2561 IdealKit ideal(this); 2562#define __ ideal. 2563 // QQQ who knows what probability is here?? 2564 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2565 // Sync IdealKit and graphKit. 2566 sync_kit(ideal); 2567 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2568 // Update IdealKit memory. 2569 __ sync_kit(this); 2570 } __ else_(); { 2571 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile); 2572 } __ end_if(); 2573 // Final sync IdealKit and GraphKit. 2574 final_sync(ideal); 2575#undef __ 2576 } 2577 } 2578 } 2579 2580 if (is_volatile) { 2581 if (!is_store) 2582 insert_mem_bar(Op_MemBarAcquire); 2583 else 2584 insert_mem_bar(Op_MemBarVolatile); 2585 } 2586 2587 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2588 2589 return true; 2590} 2591 2592//----------------------------inline_unsafe_prefetch---------------------------- 2593 2594bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2595#ifndef PRODUCT 2596 { 2597 ResourceMark rm; 2598 // Check the signatures. 2599 ciSignature* sig = callee()->signature(); 2600#ifdef ASSERT 2601 // Object getObject(Object base, int/long offset), etc. 2602 BasicType rtype = sig->return_type()->basic_type(); 2603 if (!is_native_ptr) { 2604 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2605 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2606 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2607 } else { 2608 assert(sig->count() == 1, "native prefetch has 1 argument"); 2609 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2610 } 2611#endif // ASSERT 2612 } 2613#endif // !PRODUCT 2614 2615 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2616 2617 const int idx = is_static ? 0 : 1; 2618 if (!is_static) { 2619 null_check_receiver(); 2620 if (stopped()) { 2621 return true; 2622 } 2623 } 2624 2625 // Build address expression. See the code in inline_unsafe_access. 2626 Node *adr; 2627 if (!is_native_ptr) { 2628 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2629 Node* base = argument(idx + 0); // type: oop 2630 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2631 Node* offset = argument(idx + 1); // type: long 2632 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2633 // to be plain byte offsets, which are also the same as those accepted 2634 // by oopDesc::field_base. 2635 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2636 "fieldOffset must be byte-scaled"); 2637 // 32-bit machines ignore the high half! 2638 offset = ConvL2X(offset); 2639 adr = make_unsafe_address(base, offset); 2640 } else { 2641 Node* ptr = argument(idx + 0); // type: long 2642 ptr = ConvL2X(ptr); // adjust Java long to machine word 2643 adr = make_unsafe_address(NULL, ptr); 2644 } 2645 2646 // Generate the read or write prefetch 2647 Node *prefetch; 2648 if (is_store) { 2649 prefetch = new (C) PrefetchWriteNode(i_o(), adr); 2650 } else { 2651 prefetch = new (C) PrefetchReadNode(i_o(), adr); 2652 } 2653 prefetch->init_req(0, control()); 2654 set_i_o(_gvn.transform(prefetch)); 2655 2656 return true; 2657} 2658 2659//----------------------------inline_unsafe_load_store---------------------------- 2660// This method serves a couple of different customers (depending on LoadStoreKind): 2661// 2662// LS_cmpxchg: 2663// public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2664// public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2665// public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2666// 2667// LS_xadd: 2668// public int getAndAddInt( Object o, long offset, int delta) 2669// public long getAndAddLong(Object o, long offset, long delta) 2670// 2671// LS_xchg: 2672// int getAndSet(Object o, long offset, int newValue) 2673// long getAndSet(Object o, long offset, long newValue) 2674// Object getAndSet(Object o, long offset, Object newValue) 2675// 2676bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2677 // This basic scheme here is the same as inline_unsafe_access, but 2678 // differs in enough details that combining them would make the code 2679 // overly confusing. (This is a true fact! I originally combined 2680 // them, but even I was confused by it!) As much code/comments as 2681 // possible are retained from inline_unsafe_access though to make 2682 // the correspondences clearer. - dl 2683 2684 if (callee()->is_static()) return false; // caller must have the capability! 2685 2686#ifndef PRODUCT 2687 BasicType rtype; 2688 { 2689 ResourceMark rm; 2690 // Check the signatures. 2691 ciSignature* sig = callee()->signature(); 2692 rtype = sig->return_type()->basic_type(); 2693 if (kind == LS_xadd || kind == LS_xchg) { 2694 // Check the signatures. 2695#ifdef ASSERT 2696 assert(rtype == type, "get and set must return the expected type"); 2697 assert(sig->count() == 3, "get and set has 3 arguments"); 2698 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2699 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2700 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2701#endif // ASSERT 2702 } else if (kind == LS_cmpxchg) { 2703 // Check the signatures. 2704#ifdef ASSERT 2705 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2706 assert(sig->count() == 4, "CAS has 4 arguments"); 2707 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2708 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2709#endif // ASSERT 2710 } else { 2711 ShouldNotReachHere(); 2712 } 2713 } 2714#endif //PRODUCT 2715 2716 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2717 2718 // Get arguments: 2719 Node* receiver = NULL; 2720 Node* base = NULL; 2721 Node* offset = NULL; 2722 Node* oldval = NULL; 2723 Node* newval = NULL; 2724 if (kind == LS_cmpxchg) { 2725 const bool two_slot_type = type2size[type] == 2; 2726 receiver = argument(0); // type: oop 2727 base = argument(1); // type: oop 2728 offset = argument(2); // type: long 2729 oldval = argument(4); // type: oop, int, or long 2730 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2731 } else if (kind == LS_xadd || kind == LS_xchg){ 2732 receiver = argument(0); // type: oop 2733 base = argument(1); // type: oop 2734 offset = argument(2); // type: long 2735 oldval = NULL; 2736 newval = argument(4); // type: oop, int, or long 2737 } 2738 2739 // Null check receiver. 2740 receiver = null_check(receiver); 2741 if (stopped()) { 2742 return true; 2743 } 2744 2745 // Build field offset expression. 2746 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2747 // to be plain byte offsets, which are also the same as those accepted 2748 // by oopDesc::field_base. 2749 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2750 // 32-bit machines ignore the high half of long offsets 2751 offset = ConvL2X(offset); 2752 Node* adr = make_unsafe_address(base, offset); 2753 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2754 2755 // For CAS, unlike inline_unsafe_access, there seems no point in 2756 // trying to refine types. Just use the coarse types here. 2757 const Type *value_type = Type::get_const_basic_type(type); 2758 Compile::AliasType* alias_type = C->alias_type(adr_type); 2759 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2760 2761 if (kind == LS_xchg && type == T_OBJECT) { 2762 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2763 if (tjp != NULL) { 2764 value_type = tjp; 2765 } 2766 } 2767 2768 int alias_idx = C->get_alias_index(adr_type); 2769 2770 // Memory-model-wise, a LoadStore acts like a little synchronized 2771 // block, so needs barriers on each side. These don't translate 2772 // into actual barriers on most machines, but we still need rest of 2773 // compiler to respect ordering. 2774 2775 insert_mem_bar(Op_MemBarRelease); 2776 insert_mem_bar(Op_MemBarCPUOrder); 2777 2778 // 4984716: MemBars must be inserted before this 2779 // memory node in order to avoid a false 2780 // dependency which will confuse the scheduler. 2781 Node *mem = memory(alias_idx); 2782 2783 // For now, we handle only those cases that actually exist: ints, 2784 // longs, and Object. Adding others should be straightforward. 2785 Node* load_store; 2786 switch(type) { 2787 case T_INT: 2788 if (kind == LS_xadd) { 2789 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type)); 2790 } else if (kind == LS_xchg) { 2791 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type)); 2792 } else if (kind == LS_cmpxchg) { 2793 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2794 } else { 2795 ShouldNotReachHere(); 2796 } 2797 break; 2798 case T_LONG: 2799 if (kind == LS_xadd) { 2800 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2801 } else if (kind == LS_xchg) { 2802 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2803 } else if (kind == LS_cmpxchg) { 2804 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2805 } else { 2806 ShouldNotReachHere(); 2807 } 2808 break; 2809 case T_OBJECT: 2810 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2811 // could be delayed during Parse (for example, in adjust_map_after_if()). 2812 // Execute transformation here to avoid barrier generation in such case. 2813 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2814 newval = _gvn.makecon(TypePtr::NULL_PTR); 2815 2816 // Reference stores need a store barrier. 2817 if (kind == LS_xchg) { 2818 // If pre-barrier must execute before the oop store, old value will require do_load here. 2819 if (!can_move_pre_barrier()) { 2820 pre_barrier(true /* do_load*/, 2821 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2822 NULL /* pre_val*/, 2823 T_OBJECT); 2824 } // Else move pre_barrier to use load_store value, see below. 2825 } else if (kind == LS_cmpxchg) { 2826 // Same as for newval above: 2827 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2828 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2829 } 2830 // The only known value which might get overwritten is oldval. 2831 pre_barrier(false /* do_load */, 2832 control(), NULL, NULL, max_juint, NULL, NULL, 2833 oldval /* pre_val */, 2834 T_OBJECT); 2835 } else { 2836 ShouldNotReachHere(); 2837 } 2838 2839#ifdef _LP64 2840 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2841 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2842 if (kind == LS_xchg) { 2843 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr, 2844 newval_enc, adr_type, value_type->make_narrowoop())); 2845 } else { 2846 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2847 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2848 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr, 2849 newval_enc, oldval_enc)); 2850 } 2851 } else 2852#endif 2853 { 2854 if (kind == LS_xchg) { 2855 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2856 } else { 2857 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2858 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2859 } 2860 } 2861 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2862 break; 2863 default: 2864 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2865 break; 2866 } 2867 2868 // SCMemProjNodes represent the memory state of a LoadStore. Their 2869 // main role is to prevent LoadStore nodes from being optimized away 2870 // when their results aren't used. 2871 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store)); 2872 set_memory(proj, alias_idx); 2873 2874 if (type == T_OBJECT && kind == LS_xchg) { 2875#ifdef _LP64 2876 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2877 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type())); 2878 } 2879#endif 2880 if (can_move_pre_barrier()) { 2881 // Don't need to load pre_val. The old value is returned by load_store. 2882 // The pre_barrier can execute after the xchg as long as no safepoint 2883 // gets inserted between them. 2884 pre_barrier(false /* do_load */, 2885 control(), NULL, NULL, max_juint, NULL, NULL, 2886 load_store /* pre_val */, 2887 T_OBJECT); 2888 } 2889 } 2890 2891 // Add the trailing membar surrounding the access 2892 insert_mem_bar(Op_MemBarCPUOrder); 2893 insert_mem_bar(Op_MemBarAcquire); 2894 2895 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2896 set_result(load_store); 2897 return true; 2898} 2899 2900//----------------------------inline_unsafe_ordered_store---------------------- 2901// public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2902// public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 2903// public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 2904bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2905 // This is another variant of inline_unsafe_access, differing in 2906 // that it always issues store-store ("release") barrier and ensures 2907 // store-atomicity (which only matters for "long"). 2908 2909 if (callee()->is_static()) return false; // caller must have the capability! 2910 2911#ifndef PRODUCT 2912 { 2913 ResourceMark rm; 2914 // Check the signatures. 2915 ciSignature* sig = callee()->signature(); 2916#ifdef ASSERT 2917 BasicType rtype = sig->return_type()->basic_type(); 2918 assert(rtype == T_VOID, "must return void"); 2919 assert(sig->count() == 3, "has 3 arguments"); 2920 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2921 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2922#endif // ASSERT 2923 } 2924#endif //PRODUCT 2925 2926 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2927 2928 // Get arguments: 2929 Node* receiver = argument(0); // type: oop 2930 Node* base = argument(1); // type: oop 2931 Node* offset = argument(2); // type: long 2932 Node* val = argument(4); // type: oop, int, or long 2933 2934 // Null check receiver. 2935 receiver = null_check(receiver); 2936 if (stopped()) { 2937 return true; 2938 } 2939 2940 // Build field offset expression. 2941 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2942 // 32-bit machines ignore the high half of long offsets 2943 offset = ConvL2X(offset); 2944 Node* adr = make_unsafe_address(base, offset); 2945 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2946 const Type *value_type = Type::get_const_basic_type(type); 2947 Compile::AliasType* alias_type = C->alias_type(adr_type); 2948 2949 insert_mem_bar(Op_MemBarRelease); 2950 insert_mem_bar(Op_MemBarCPUOrder); 2951 // Ensure that the store is atomic for longs: 2952 const bool require_atomic_access = true; 2953 Node* store; 2954 if (type == T_OBJECT) // reference stores need a store barrier. 2955 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2956 else { 2957 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2958 } 2959 insert_mem_bar(Op_MemBarCPUOrder); 2960 return true; 2961} 2962 2963bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2964 // Regardless of form, don't allow previous ld/st to move down, 2965 // then issue acquire, release, or volatile mem_bar. 2966 insert_mem_bar(Op_MemBarCPUOrder); 2967 switch(id) { 2968 case vmIntrinsics::_loadFence: 2969 insert_mem_bar(Op_MemBarAcquire); 2970 return true; 2971 case vmIntrinsics::_storeFence: 2972 insert_mem_bar(Op_MemBarRelease); 2973 return true; 2974 case vmIntrinsics::_fullFence: 2975 insert_mem_bar(Op_MemBarVolatile); 2976 return true; 2977 default: 2978 fatal_unexpected_iid(id); 2979 return false; 2980 } 2981} 2982 2983bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 2984 if (!kls->is_Con()) { 2985 return true; 2986 } 2987 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 2988 if (klsptr == NULL) { 2989 return true; 2990 } 2991 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 2992 // don't need a guard for a klass that is already initialized 2993 return !ik->is_initialized(); 2994} 2995 2996//----------------------------inline_unsafe_allocate--------------------------- 2997// public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 2998bool LibraryCallKit::inline_unsafe_allocate() { 2999 if (callee()->is_static()) return false; // caller must have the capability! 3000 3001 null_check_receiver(); // null-check, then ignore 3002 Node* cls = null_check(argument(1)); 3003 if (stopped()) return true; 3004 3005 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3006 kls = null_check(kls); 3007 if (stopped()) return true; // argument was like int.class 3008 3009 Node* test = NULL; 3010 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3011 // Note: The argument might still be an illegal value like 3012 // Serializable.class or Object[].class. The runtime will handle it. 3013 // But we must make an explicit check for initialization. 3014 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3015 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3016 // can generate code to load it as unsigned byte. 3017 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN); 3018 Node* bits = intcon(InstanceKlass::fully_initialized); 3019 test = _gvn.transform(new (C) SubINode(inst, bits)); 3020 // The 'test' is non-zero if we need to take a slow path. 3021 } 3022 3023 Node* obj = new_instance(kls, test); 3024 set_result(obj); 3025 return true; 3026} 3027 3028#ifdef TRACE_HAVE_INTRINSICS 3029/* 3030 * oop -> myklass 3031 * myklass->trace_id |= USED 3032 * return myklass->trace_id & ~0x3 3033 */ 3034bool LibraryCallKit::inline_native_classID() { 3035 null_check_receiver(); // null-check, then ignore 3036 Node* cls = null_check(argument(1), T_OBJECT); 3037 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3038 kls = null_check(kls, T_OBJECT); 3039 ByteSize offset = TRACE_ID_OFFSET; 3040 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3041 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG); 3042 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3043 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits)); 3044 Node* clsused = longcon(0x01l); // set the class bit 3045 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused)); 3046 3047 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3048 store_to_memory(control(), insp, orl, T_LONG, adr_type); 3049 set_result(andl); 3050 return true; 3051} 3052 3053bool LibraryCallKit::inline_native_threadID() { 3054 Node* tls_ptr = NULL; 3055 Node* cur_thr = generate_current_thread(tls_ptr); 3056 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3057 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 3058 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3059 3060 Node* threadid = NULL; 3061 size_t thread_id_size = OSThread::thread_id_size(); 3062 if (thread_id_size == (size_t) BytesPerLong) { 3063 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG)); 3064 } else if (thread_id_size == (size_t) BytesPerInt) { 3065 threadid = make_load(control(), p, TypeInt::INT, T_INT); 3066 } else { 3067 ShouldNotReachHere(); 3068 } 3069 set_result(threadid); 3070 return true; 3071} 3072#endif 3073 3074//------------------------inline_native_time_funcs-------------- 3075// inline code for System.currentTimeMillis() and System.nanoTime() 3076// these have the same type and signature 3077bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3078 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3079 const TypePtr* no_memory_effects = NULL; 3080 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3081 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0)); 3082#ifdef ASSERT 3083 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1)); 3084 assert(value_top == top(), "second value must be top"); 3085#endif 3086 set_result(value); 3087 return true; 3088} 3089 3090//------------------------inline_native_currentThread------------------ 3091bool LibraryCallKit::inline_native_currentThread() { 3092 Node* junk = NULL; 3093 set_result(generate_current_thread(junk)); 3094 return true; 3095} 3096 3097//------------------------inline_native_isInterrupted------------------ 3098// private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3099bool LibraryCallKit::inline_native_isInterrupted() { 3100 // Add a fast path to t.isInterrupted(clear_int): 3101 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 3102 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3103 // So, in the common case that the interrupt bit is false, 3104 // we avoid making a call into the VM. Even if the interrupt bit 3105 // is true, if the clear_int argument is false, we avoid the VM call. 3106 // However, if the receiver is not currentThread, we must call the VM, 3107 // because there must be some locking done around the operation. 3108 3109 // We only go to the fast case code if we pass two guards. 3110 // Paths which do not pass are accumulated in the slow_region. 3111 3112 enum { 3113 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3114 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3115 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3116 PATH_LIMIT 3117 }; 3118 3119 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3120 // out of the function. 3121 insert_mem_bar(Op_MemBarCPUOrder); 3122 3123 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT); 3124 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL); 3125 3126 RegionNode* slow_region = new (C) RegionNode(1); 3127 record_for_igvn(slow_region); 3128 3129 // (a) Receiving thread must be the current thread. 3130 Node* rec_thr = argument(0); 3131 Node* tls_ptr = NULL; 3132 Node* cur_thr = generate_current_thread(tls_ptr); 3133 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr)); 3134 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne)); 3135 3136 generate_slow_guard(bol_thr, slow_region); 3137 3138 // (b) Interrupt bit on TLS must be false. 3139 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3140 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 3141 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3142 3143 // Set the control input on the field _interrupted read to prevent it floating up. 3144 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 3145 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0))); 3146 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne)); 3147 3148 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3149 3150 // First fast path: if (!TLS._interrupted) return false; 3151 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit)); 3152 result_rgn->init_req(no_int_result_path, false_bit); 3153 result_val->init_req(no_int_result_path, intcon(0)); 3154 3155 // drop through to next case 3156 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit))); 3157 3158 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3159 Node* clr_arg = argument(1); 3160 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0))); 3161 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne)); 3162 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3163 3164 // Second fast path: ... else if (!clear_int) return true; 3165 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg)); 3166 result_rgn->init_req(no_clear_result_path, false_arg); 3167 result_val->init_req(no_clear_result_path, intcon(1)); 3168 3169 // drop through to next case 3170 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg))); 3171 3172 // (d) Otherwise, go to the slow path. 3173 slow_region->add_req(control()); 3174 set_control( _gvn.transform(slow_region)); 3175 3176 if (stopped()) { 3177 // There is no slow path. 3178 result_rgn->init_req(slow_result_path, top()); 3179 result_val->init_req(slow_result_path, top()); 3180 } else { 3181 // non-virtual because it is a private non-static 3182 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3183 3184 Node* slow_val = set_results_for_java_call(slow_call); 3185 // this->control() comes from set_results_for_java_call 3186 3187 Node* fast_io = slow_call->in(TypeFunc::I_O); 3188 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3189 3190 // These two phis are pre-filled with copies of of the fast IO and Memory 3191 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3192 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3193 3194 result_rgn->init_req(slow_result_path, control()); 3195 result_io ->init_req(slow_result_path, i_o()); 3196 result_mem->init_req(slow_result_path, reset_memory()); 3197 result_val->init_req(slow_result_path, slow_val); 3198 3199 set_all_memory(_gvn.transform(result_mem)); 3200 set_i_o( _gvn.transform(result_io)); 3201 } 3202 3203 C->set_has_split_ifs(true); // Has chance for split-if optimization 3204 set_result(result_rgn, result_val); 3205 return true; 3206} 3207 3208//---------------------------load_mirror_from_klass---------------------------- 3209// Given a klass oop, load its java mirror (a java.lang.Class oop). 3210Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3211 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3212 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 3213} 3214 3215//-----------------------load_klass_from_mirror_common------------------------- 3216// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3217// Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3218// and branch to the given path on the region. 3219// If never_see_null, take an uncommon trap on null, so we can optimistically 3220// compile for the non-null case. 3221// If the region is NULL, force never_see_null = true. 3222Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3223 bool never_see_null, 3224 RegionNode* region, 3225 int null_path, 3226 int offset) { 3227 if (region == NULL) never_see_null = true; 3228 Node* p = basic_plus_adr(mirror, offset); 3229 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3230 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3231 Node* null_ctl = top(); 3232 kls = null_check_oop(kls, &null_ctl, never_see_null); 3233 if (region != NULL) { 3234 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3235 region->init_req(null_path, null_ctl); 3236 } else { 3237 assert(null_ctl == top(), "no loose ends"); 3238 } 3239 return kls; 3240} 3241 3242//--------------------(inline_native_Class_query helpers)--------------------- 3243// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3244// Fall through if (mods & mask) == bits, take the guard otherwise. 3245Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3246 // Branch around if the given klass has the given modifier bit set. 3247 // Like generate_guard, adds a new path onto the region. 3248 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3249 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 3250 Node* mask = intcon(modifier_mask); 3251 Node* bits = intcon(modifier_bits); 3252 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask)); 3253 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits)); 3254 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 3255 return generate_fair_guard(bol, region); 3256} 3257Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3258 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3259} 3260 3261//-------------------------inline_native_Class_query------------------- 3262bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3263 const Type* return_type = TypeInt::BOOL; 3264 Node* prim_return_value = top(); // what happens if it's a primitive class? 3265 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3266 bool expect_prim = false; // most of these guys expect to work on refs 3267 3268 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3269 3270 Node* mirror = argument(0); 3271 Node* obj = top(); 3272 3273 switch (id) { 3274 case vmIntrinsics::_isInstance: 3275 // nothing is an instance of a primitive type 3276 prim_return_value = intcon(0); 3277 obj = argument(1); 3278 break; 3279 case vmIntrinsics::_getModifiers: 3280 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3281 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3282 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3283 break; 3284 case vmIntrinsics::_isInterface: 3285 prim_return_value = intcon(0); 3286 break; 3287 case vmIntrinsics::_isArray: 3288 prim_return_value = intcon(0); 3289 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3290 break; 3291 case vmIntrinsics::_isPrimitive: 3292 prim_return_value = intcon(1); 3293 expect_prim = true; // obviously 3294 break; 3295 case vmIntrinsics::_getSuperclass: 3296 prim_return_value = null(); 3297 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3298 break; 3299 case vmIntrinsics::_getComponentType: 3300 prim_return_value = null(); 3301 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3302 break; 3303 case vmIntrinsics::_getClassAccessFlags: 3304 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3305 return_type = TypeInt::INT; // not bool! 6297094 3306 break; 3307 default: 3308 fatal_unexpected_iid(id); 3309 break; 3310 } 3311 3312 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3313 if (mirror_con == NULL) return false; // cannot happen? 3314 3315#ifndef PRODUCT 3316 if (C->print_intrinsics() || C->print_inlining()) { 3317 ciType* k = mirror_con->java_mirror_type(); 3318 if (k) { 3319 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3320 k->print_name(); 3321 tty->cr(); 3322 } 3323 } 3324#endif 3325 3326 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3327 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3328 record_for_igvn(region); 3329 PhiNode* phi = new (C) PhiNode(region, return_type); 3330 3331 // The mirror will never be null of Reflection.getClassAccessFlags, however 3332 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3333 // if it is. See bug 4774291. 3334 3335 // For Reflection.getClassAccessFlags(), the null check occurs in 3336 // the wrong place; see inline_unsafe_access(), above, for a similar 3337 // situation. 3338 mirror = null_check(mirror); 3339 // If mirror or obj is dead, only null-path is taken. 3340 if (stopped()) return true; 3341 3342 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3343 3344 // Now load the mirror's klass metaobject, and null-check it. 3345 // Side-effects region with the control path if the klass is null. 3346 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3347 // If kls is null, we have a primitive mirror. 3348 phi->init_req(_prim_path, prim_return_value); 3349 if (stopped()) { set_result(region, phi); return true; } 3350 3351 Node* p; // handy temp 3352 Node* null_ctl; 3353 3354 // Now that we have the non-null klass, we can perform the real query. 3355 // For constant classes, the query will constant-fold in LoadNode::Value. 3356 Node* query_value = top(); 3357 switch (id) { 3358 case vmIntrinsics::_isInstance: 3359 // nothing is an instance of a primitive type 3360 query_value = gen_instanceof(obj, kls); 3361 break; 3362 3363 case vmIntrinsics::_getModifiers: 3364 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3365 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3366 break; 3367 3368 case vmIntrinsics::_isInterface: 3369 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3370 if (generate_interface_guard(kls, region) != NULL) 3371 // A guard was added. If the guard is taken, it was an interface. 3372 phi->add_req(intcon(1)); 3373 // If we fall through, it's a plain class. 3374 query_value = intcon(0); 3375 break; 3376 3377 case vmIntrinsics::_isArray: 3378 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3379 if (generate_array_guard(kls, region) != NULL) 3380 // A guard was added. If the guard is taken, it was an array. 3381 phi->add_req(intcon(1)); 3382 // If we fall through, it's a plain class. 3383 query_value = intcon(0); 3384 break; 3385 3386 case vmIntrinsics::_isPrimitive: 3387 query_value = intcon(0); // "normal" path produces false 3388 break; 3389 3390 case vmIntrinsics::_getSuperclass: 3391 // The rules here are somewhat unfortunate, but we can still do better 3392 // with random logic than with a JNI call. 3393 // Interfaces store null or Object as _super, but must report null. 3394 // Arrays store an intermediate super as _super, but must report Object. 3395 // Other types can report the actual _super. 3396 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3397 if (generate_interface_guard(kls, region) != NULL) 3398 // A guard was added. If the guard is taken, it was an interface. 3399 phi->add_req(null()); 3400 if (generate_array_guard(kls, region) != NULL) 3401 // A guard was added. If the guard is taken, it was an array. 3402 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3403 // If we fall through, it's a plain class. Get its _super. 3404 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3405 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3406 null_ctl = top(); 3407 kls = null_check_oop(kls, &null_ctl); 3408 if (null_ctl != top()) { 3409 // If the guard is taken, Object.superClass is null (both klass and mirror). 3410 region->add_req(null_ctl); 3411 phi ->add_req(null()); 3412 } 3413 if (!stopped()) { 3414 query_value = load_mirror_from_klass(kls); 3415 } 3416 break; 3417 3418 case vmIntrinsics::_getComponentType: 3419 if (generate_array_guard(kls, region) != NULL) { 3420 // Be sure to pin the oop load to the guard edge just created: 3421 Node* is_array_ctrl = region->in(region->req()-1); 3422 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); 3423 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 3424 phi->add_req(cmo); 3425 } 3426 query_value = null(); // non-array case is null 3427 break; 3428 3429 case vmIntrinsics::_getClassAccessFlags: 3430 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3431 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3432 break; 3433 3434 default: 3435 fatal_unexpected_iid(id); 3436 break; 3437 } 3438 3439 // Fall-through is the normal case of a query to a real class. 3440 phi->init_req(1, query_value); 3441 region->init_req(1, control()); 3442 3443 C->set_has_split_ifs(true); // Has chance for split-if optimization 3444 set_result(region, phi); 3445 return true; 3446} 3447 3448//--------------------------inline_native_subtype_check------------------------ 3449// This intrinsic takes the JNI calls out of the heart of 3450// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3451bool LibraryCallKit::inline_native_subtype_check() { 3452 // Pull both arguments off the stack. 3453 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3454 args[0] = argument(0); 3455 args[1] = argument(1); 3456 Node* klasses[2]; // corresponding Klasses: superk, subk 3457 klasses[0] = klasses[1] = top(); 3458 3459 enum { 3460 // A full decision tree on {superc is prim, subc is prim}: 3461 _prim_0_path = 1, // {P,N} => false 3462 // {P,P} & superc!=subc => false 3463 _prim_same_path, // {P,P} & superc==subc => true 3464 _prim_1_path, // {N,P} => false 3465 _ref_subtype_path, // {N,N} & subtype check wins => true 3466 _both_ref_path, // {N,N} & subtype check loses => false 3467 PATH_LIMIT 3468 }; 3469 3470 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3471 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 3472 record_for_igvn(region); 3473 3474 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3475 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3476 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3477 3478 // First null-check both mirrors and load each mirror's klass metaobject. 3479 int which_arg; 3480 for (which_arg = 0; which_arg <= 1; which_arg++) { 3481 Node* arg = args[which_arg]; 3482 arg = null_check(arg); 3483 if (stopped()) break; 3484 args[which_arg] = arg; 3485 3486 Node* p = basic_plus_adr(arg, class_klass_offset); 3487 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3488 klasses[which_arg] = _gvn.transform(kls); 3489 } 3490 3491 // Having loaded both klasses, test each for null. 3492 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3493 for (which_arg = 0; which_arg <= 1; which_arg++) { 3494 Node* kls = klasses[which_arg]; 3495 Node* null_ctl = top(); 3496 kls = null_check_oop(kls, &null_ctl, never_see_null); 3497 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3498 region->init_req(prim_path, null_ctl); 3499 if (stopped()) break; 3500 klasses[which_arg] = kls; 3501 } 3502 3503 if (!stopped()) { 3504 // now we have two reference types, in klasses[0..1] 3505 Node* subk = klasses[1]; // the argument to isAssignableFrom 3506 Node* superk = klasses[0]; // the receiver 3507 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3508 // now we have a successful reference subtype check 3509 region->set_req(_ref_subtype_path, control()); 3510 } 3511 3512 // If both operands are primitive (both klasses null), then 3513 // we must return true when they are identical primitives. 3514 // It is convenient to test this after the first null klass check. 3515 set_control(region->in(_prim_0_path)); // go back to first null check 3516 if (!stopped()) { 3517 // Since superc is primitive, make a guard for the superc==subc case. 3518 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1])); 3519 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq)); 3520 generate_guard(bol_eq, region, PROB_FAIR); 3521 if (region->req() == PATH_LIMIT+1) { 3522 // A guard was added. If the added guard is taken, superc==subc. 3523 region->swap_edges(PATH_LIMIT, _prim_same_path); 3524 region->del_req(PATH_LIMIT); 3525 } 3526 region->set_req(_prim_0_path, control()); // Not equal after all. 3527 } 3528 3529 // these are the only paths that produce 'true': 3530 phi->set_req(_prim_same_path, intcon(1)); 3531 phi->set_req(_ref_subtype_path, intcon(1)); 3532 3533 // pull together the cases: 3534 assert(region->req() == PATH_LIMIT, "sane region"); 3535 for (uint i = 1; i < region->req(); i++) { 3536 Node* ctl = region->in(i); 3537 if (ctl == NULL || ctl == top()) { 3538 region->set_req(i, top()); 3539 phi ->set_req(i, top()); 3540 } else if (phi->in(i) == NULL) { 3541 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3542 } 3543 } 3544 3545 set_control(_gvn.transform(region)); 3546 set_result(_gvn.transform(phi)); 3547 return true; 3548} 3549 3550//---------------------generate_array_guard_common------------------------ 3551Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3552 bool obj_array, bool not_array) { 3553 // If obj_array/non_array==false/false: 3554 // Branch around if the given klass is in fact an array (either obj or prim). 3555 // If obj_array/non_array==false/true: 3556 // Branch around if the given klass is not an array klass of any kind. 3557 // If obj_array/non_array==true/true: 3558 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3559 // If obj_array/non_array==true/false: 3560 // Branch around if the kls is an oop array (Object[] or subtype) 3561 // 3562 // Like generate_guard, adds a new path onto the region. 3563 jint layout_con = 0; 3564 Node* layout_val = get_layout_helper(kls, layout_con); 3565 if (layout_val == NULL) { 3566 bool query = (obj_array 3567 ? Klass::layout_helper_is_objArray(layout_con) 3568 : Klass::layout_helper_is_array(layout_con)); 3569 if (query == not_array) { 3570 return NULL; // never a branch 3571 } else { // always a branch 3572 Node* always_branch = control(); 3573 if (region != NULL) 3574 region->add_req(always_branch); 3575 set_control(top()); 3576 return always_branch; 3577 } 3578 } 3579 // Now test the correct condition. 3580 jint nval = (obj_array 3581 ? ((jint)Klass::_lh_array_tag_type_value 3582 << Klass::_lh_array_tag_shift) 3583 : Klass::_lh_neutral_value); 3584 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval))); 3585 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3586 // invert the test if we are looking for a non-array 3587 if (not_array) btest = BoolTest(btest).negate(); 3588 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest)); 3589 return generate_fair_guard(bol, region); 3590} 3591 3592 3593//-----------------------inline_native_newArray-------------------------- 3594// private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3595bool LibraryCallKit::inline_native_newArray() { 3596 Node* mirror = argument(0); 3597 Node* count_val = argument(1); 3598 3599 mirror = null_check(mirror); 3600 // If mirror or obj is dead, only null-path is taken. 3601 if (stopped()) return true; 3602 3603 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3604 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3605 PhiNode* result_val = new(C) PhiNode(result_reg, 3606 TypeInstPtr::NOTNULL); 3607 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3608 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3609 TypePtr::BOTTOM); 3610 3611 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3612 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3613 result_reg, _slow_path); 3614 Node* normal_ctl = control(); 3615 Node* no_array_ctl = result_reg->in(_slow_path); 3616 3617 // Generate code for the slow case. We make a call to newArray(). 3618 set_control(no_array_ctl); 3619 if (!stopped()) { 3620 // Either the input type is void.class, or else the 3621 // array klass has not yet been cached. Either the 3622 // ensuing call will throw an exception, or else it 3623 // will cache the array klass for next time. 3624 PreserveJVMState pjvms(this); 3625 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3626 Node* slow_result = set_results_for_java_call(slow_call); 3627 // this->control() comes from set_results_for_java_call 3628 result_reg->set_req(_slow_path, control()); 3629 result_val->set_req(_slow_path, slow_result); 3630 result_io ->set_req(_slow_path, i_o()); 3631 result_mem->set_req(_slow_path, reset_memory()); 3632 } 3633 3634 set_control(normal_ctl); 3635 if (!stopped()) { 3636 // Normal case: The array type has been cached in the java.lang.Class. 3637 // The following call works fine even if the array type is polymorphic. 3638 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3639 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3640 result_reg->init_req(_normal_path, control()); 3641 result_val->init_req(_normal_path, obj); 3642 result_io ->init_req(_normal_path, i_o()); 3643 result_mem->init_req(_normal_path, reset_memory()); 3644 } 3645 3646 // Return the combined state. 3647 set_i_o( _gvn.transform(result_io) ); 3648 set_all_memory( _gvn.transform(result_mem)); 3649 3650 C->set_has_split_ifs(true); // Has chance for split-if optimization 3651 set_result(result_reg, result_val); 3652 return true; 3653} 3654 3655//----------------------inline_native_getLength-------------------------- 3656// public static native int java.lang.reflect.Array.getLength(Object array); 3657bool LibraryCallKit::inline_native_getLength() { 3658 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3659 3660 Node* array = null_check(argument(0)); 3661 // If array is dead, only null-path is taken. 3662 if (stopped()) return true; 3663 3664 // Deoptimize if it is a non-array. 3665 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3666 3667 if (non_array != NULL) { 3668 PreserveJVMState pjvms(this); 3669 set_control(non_array); 3670 uncommon_trap(Deoptimization::Reason_intrinsic, 3671 Deoptimization::Action_maybe_recompile); 3672 } 3673 3674 // If control is dead, only non-array-path is taken. 3675 if (stopped()) return true; 3676 3677 // The works fine even if the array type is polymorphic. 3678 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3679 Node* result = load_array_length(array); 3680 3681 C->set_has_split_ifs(true); // Has chance for split-if optimization 3682 set_result(result); 3683 return true; 3684} 3685 3686//------------------------inline_array_copyOf---------------------------- 3687// public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3688// public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3689bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3690 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3691 3692 // Get the arguments. 3693 Node* original = argument(0); 3694 Node* start = is_copyOfRange? argument(1): intcon(0); 3695 Node* end = is_copyOfRange? argument(2): argument(1); 3696 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3697 3698 Node* newcopy; 3699 3700 // Set the original stack and the reexecute bit for the interpreter to reexecute 3701 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3702 { PreserveReexecuteState preexecs(this); 3703 jvms()->set_should_reexecute(true); 3704 3705 array_type_mirror = null_check(array_type_mirror); 3706 original = null_check(original); 3707 3708 // Check if a null path was taken unconditionally. 3709 if (stopped()) return true; 3710 3711 Node* orig_length = load_array_length(original); 3712 3713 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3714 klass_node = null_check(klass_node); 3715 3716 RegionNode* bailout = new (C) RegionNode(1); 3717 record_for_igvn(bailout); 3718 3719 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3720 // Bail out if that is so. 3721 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3722 if (not_objArray != NULL) { 3723 // Improve the klass node's type from the new optimistic assumption: 3724 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3725 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3726 Node* cast = new (C) CastPPNode(klass_node, akls); 3727 cast->init_req(0, control()); 3728 klass_node = _gvn.transform(cast); 3729 } 3730 3731 // Bail out if either start or end is negative. 3732 generate_negative_guard(start, bailout, &start); 3733 generate_negative_guard(end, bailout, &end); 3734 3735 Node* length = end; 3736 if (_gvn.type(start) != TypeInt::ZERO) { 3737 length = _gvn.transform(new (C) SubINode(end, start)); 3738 } 3739 3740 // Bail out if length is negative. 3741 // Without this the new_array would throw 3742 // NegativeArraySizeException but IllegalArgumentException is what 3743 // should be thrown 3744 generate_negative_guard(length, bailout, &length); 3745 3746 if (bailout->req() > 1) { 3747 PreserveJVMState pjvms(this); 3748 set_control(_gvn.transform(bailout)); 3749 uncommon_trap(Deoptimization::Reason_intrinsic, 3750 Deoptimization::Action_maybe_recompile); 3751 } 3752 3753 if (!stopped()) { 3754 // How many elements will we copy from the original? 3755 // The answer is MinI(orig_length - start, length). 3756 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start)); 3757 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3758 3759 newcopy = new_array(klass_node, length, 0); // no argments to push 3760 3761 // Generate a direct call to the right arraycopy function(s). 3762 // We know the copy is disjoint but we might not know if the 3763 // oop stores need checking. 3764 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3765 // This will fail a store-check if x contains any non-nulls. 3766 bool disjoint_bases = true; 3767 // if start > orig_length then the length of the copy may be 3768 // negative. 3769 bool length_never_negative = !is_copyOfRange; 3770 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3771 original, start, newcopy, intcon(0), moved, 3772 disjoint_bases, length_never_negative); 3773 } 3774 } // original reexecute is set back here 3775 3776 C->set_has_split_ifs(true); // Has chance for split-if optimization 3777 if (!stopped()) { 3778 set_result(newcopy); 3779 } 3780 return true; 3781} 3782 3783 3784//----------------------generate_virtual_guard--------------------------- 3785// Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3786Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3787 RegionNode* slow_region) { 3788 ciMethod* method = callee(); 3789 int vtable_index = method->vtable_index(); 3790 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3791 err_msg_res("bad index %d", vtable_index)); 3792 // Get the Method* out of the appropriate vtable entry. 3793 int entry_offset = (InstanceKlass::vtable_start_offset() + 3794 vtable_index*vtableEntry::size()) * wordSize + 3795 vtableEntry::method_offset_in_bytes(); 3796 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3797 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS); 3798 3799 // Compare the target method with the expected method (e.g., Object.hashCode). 3800 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3801 3802 Node* native_call = makecon(native_call_addr); 3803 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call)); 3804 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne)); 3805 3806 return generate_slow_guard(test_native, slow_region); 3807} 3808 3809//-----------------------generate_method_call---------------------------- 3810// Use generate_method_call to make a slow-call to the real 3811// method if the fast path fails. An alternative would be to 3812// use a stub like OptoRuntime::slow_arraycopy_Java. 3813// This only works for expanding the current library call, 3814// not another intrinsic. (E.g., don't use this for making an 3815// arraycopy call inside of the copyOf intrinsic.) 3816CallJavaNode* 3817LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3818 // When compiling the intrinsic method itself, do not use this technique. 3819 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3820 3821 ciMethod* method = callee(); 3822 // ensure the JVMS we have will be correct for this call 3823 guarantee(method_id == method->intrinsic_id(), "must match"); 3824 3825 const TypeFunc* tf = TypeFunc::make(method); 3826 CallJavaNode* slow_call; 3827 if (is_static) { 3828 assert(!is_virtual, ""); 3829 slow_call = new(C) CallStaticJavaNode(C, tf, 3830 SharedRuntime::get_resolve_static_call_stub(), 3831 method, bci()); 3832 } else if (is_virtual) { 3833 null_check_receiver(); 3834 int vtable_index = Method::invalid_vtable_index; 3835 if (UseInlineCaches) { 3836 // Suppress the vtable call 3837 } else { 3838 // hashCode and clone are not a miranda methods, 3839 // so the vtable index is fixed. 3840 // No need to use the linkResolver to get it. 3841 vtable_index = method->vtable_index(); 3842 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3843 err_msg_res("bad index %d", vtable_index)); 3844 } 3845 slow_call = new(C) CallDynamicJavaNode(tf, 3846 SharedRuntime::get_resolve_virtual_call_stub(), 3847 method, vtable_index, bci()); 3848 } else { // neither virtual nor static: opt_virtual 3849 null_check_receiver(); 3850 slow_call = new(C) CallStaticJavaNode(C, tf, 3851 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3852 method, bci()); 3853 slow_call->set_optimized_virtual(true); 3854 } 3855 set_arguments_for_java_call(slow_call); 3856 set_edges_for_java_call(slow_call); 3857 return slow_call; 3858} 3859 3860 3861//------------------------------inline_native_hashcode-------------------- 3862// Build special case code for calls to hashCode on an object. 3863bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3864 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3865 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3866 3867 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3868 3869 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3870 PhiNode* result_val = new(C) PhiNode(result_reg, 3871 TypeInt::INT); 3872 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3873 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3874 TypePtr::BOTTOM); 3875 Node* obj = NULL; 3876 if (!is_static) { 3877 // Check for hashing null object 3878 obj = null_check_receiver(); 3879 if (stopped()) return true; // unconditionally null 3880 result_reg->init_req(_null_path, top()); 3881 result_val->init_req(_null_path, top()); 3882 } else { 3883 // Do a null check, and return zero if null. 3884 // System.identityHashCode(null) == 0 3885 obj = argument(0); 3886 Node* null_ctl = top(); 3887 obj = null_check_oop(obj, &null_ctl); 3888 result_reg->init_req(_null_path, null_ctl); 3889 result_val->init_req(_null_path, _gvn.intcon(0)); 3890 } 3891 3892 // Unconditionally null? Then return right away. 3893 if (stopped()) { 3894 set_control( result_reg->in(_null_path)); 3895 if (!stopped()) 3896 set_result(result_val->in(_null_path)); 3897 return true; 3898 } 3899 3900 // After null check, get the object's klass. 3901 Node* obj_klass = load_object_klass(obj); 3902 3903 // This call may be virtual (invokevirtual) or bound (invokespecial). 3904 // For each case we generate slightly different code. 3905 3906 // We only go to the fast case code if we pass a number of guards. The 3907 // paths which do not pass are accumulated in the slow_region. 3908 RegionNode* slow_region = new (C) RegionNode(1); 3909 record_for_igvn(slow_region); 3910 3911 // If this is a virtual call, we generate a funny guard. We pull out 3912 // the vtable entry corresponding to hashCode() from the target object. 3913 // If the target method which we are calling happens to be the native 3914 // Object hashCode() method, we pass the guard. We do not need this 3915 // guard for non-virtual calls -- the caller is known to be the native 3916 // Object hashCode(). 3917 if (is_virtual) { 3918 generate_virtual_guard(obj_klass, slow_region); 3919 } 3920 3921 // Get the header out of the object, use LoadMarkNode when available 3922 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3923 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type()); 3924 3925 // Test the header to see if it is unlocked. 3926 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3927 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask)); 3928 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3929 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val)); 3930 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne)); 3931 3932 generate_slow_guard(test_unlocked, slow_region); 3933 3934 // Get the hash value and check to see that it has been properly assigned. 3935 // We depend on hash_mask being at most 32 bits and avoid the use of 3936 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3937 // vm: see markOop.hpp. 3938 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3939 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3940 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift)); 3941 // This hack lets the hash bits live anywhere in the mark object now, as long 3942 // as the shift drops the relevant bits into the low 32 bits. Note that 3943 // Java spec says that HashCode is an int so there's no point in capturing 3944 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3945 hshifted_header = ConvX2I(hshifted_header); 3946 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask)); 3947 3948 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3949 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val)); 3950 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq)); 3951 3952 generate_slow_guard(test_assigned, slow_region); 3953 3954 Node* init_mem = reset_memory(); 3955 // fill in the rest of the null path: 3956 result_io ->init_req(_null_path, i_o()); 3957 result_mem->init_req(_null_path, init_mem); 3958 3959 result_val->init_req(_fast_path, hash_val); 3960 result_reg->init_req(_fast_path, control()); 3961 result_io ->init_req(_fast_path, i_o()); 3962 result_mem->init_req(_fast_path, init_mem); 3963 3964 // Generate code for the slow case. We make a call to hashCode(). 3965 set_control(_gvn.transform(slow_region)); 3966 if (!stopped()) { 3967 // No need for PreserveJVMState, because we're using up the present state. 3968 set_all_memory(init_mem); 3969 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 3970 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3971 Node* slow_result = set_results_for_java_call(slow_call); 3972 // this->control() comes from set_results_for_java_call 3973 result_reg->init_req(_slow_path, control()); 3974 result_val->init_req(_slow_path, slow_result); 3975 result_io ->set_req(_slow_path, i_o()); 3976 result_mem ->set_req(_slow_path, reset_memory()); 3977 } 3978 3979 // Return the combined state. 3980 set_i_o( _gvn.transform(result_io) ); 3981 set_all_memory( _gvn.transform(result_mem)); 3982 3983 set_result(result_reg, result_val); 3984 return true; 3985} 3986 3987//---------------------------inline_native_getClass---------------------------- 3988// public final native Class<?> java.lang.Object.getClass(); 3989// 3990// Build special case code for calls to getClass on an object. 3991bool LibraryCallKit::inline_native_getClass() { 3992 Node* obj = null_check_receiver(); 3993 if (stopped()) return true; 3994 set_result(load_mirror_from_klass(load_object_klass(obj))); 3995 return true; 3996} 3997 3998//-----------------inline_native_Reflection_getCallerClass--------------------- 3999// public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4000// 4001// In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4002// 4003// NOTE: This code must perform the same logic as JVM_GetCallerClass 4004// in that it must skip particular security frames and checks for 4005// caller sensitive methods. 4006bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4007#ifndef PRODUCT 4008 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4009 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4010 } 4011#endif 4012 4013 if (!jvms()->has_method()) { 4014#ifndef PRODUCT 4015 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4016 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4017 } 4018#endif 4019 return false; 4020 } 4021 4022 // Walk back up the JVM state to find the caller at the required 4023 // depth. 4024 JVMState* caller_jvms = jvms(); 4025 4026 // Cf. JVM_GetCallerClass 4027 // NOTE: Start the loop at depth 1 because the current JVM state does 4028 // not include the Reflection.getCallerClass() frame. 4029 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4030 ciMethod* m = caller_jvms->method(); 4031 switch (n) { 4032 case 0: 4033 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4034 break; 4035 case 1: 4036 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4037 if (!m->caller_sensitive()) { 4038#ifndef PRODUCT 4039 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4040 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4041 } 4042#endif 4043 return false; // bail-out; let JVM_GetCallerClass do the work 4044 } 4045 break; 4046 default: 4047 if (!m->is_ignored_by_security_stack_walk()) { 4048 // We have reached the desired frame; return the holder class. 4049 // Acquire method holder as java.lang.Class and push as constant. 4050 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4051 ciInstance* caller_mirror = caller_klass->java_mirror(); 4052 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4053 4054#ifndef PRODUCT 4055 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4056 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()); 4057 tty->print_cr(" JVM state at this point:"); 4058 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4059 ciMethod* m = jvms()->of_depth(i)->method(); 4060 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4061 } 4062 } 4063#endif 4064 return true; 4065 } 4066 break; 4067 } 4068 } 4069 4070#ifndef PRODUCT 4071 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4072 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4073 tty->print_cr(" JVM state at this point:"); 4074 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4075 ciMethod* m = jvms()->of_depth(i)->method(); 4076 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4077 } 4078 } 4079#endif 4080 4081 return false; // bail-out; let JVM_GetCallerClass do the work 4082} 4083 4084bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4085 Node* arg = argument(0); 4086 Node* result; 4087 4088 switch (id) { 4089 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break; 4090 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break; 4091 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break; 4092 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break; 4093 4094 case vmIntrinsics::_doubleToLongBits: { 4095 // two paths (plus control) merge in a wood 4096 RegionNode *r = new (C) RegionNode(3); 4097 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 4098 4099 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg)); 4100 // Build the boolean node 4101 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4102 4103 // Branch either way. 4104 // NaN case is less traveled, which makes all the difference. 4105 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4106 Node *opt_isnan = _gvn.transform(ifisnan); 4107 assert( opt_isnan->is_If(), "Expect an IfNode"); 4108 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4109 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4110 4111 set_control(iftrue); 4112 4113 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4114 Node *slow_result = longcon(nan_bits); // return NaN 4115 phi->init_req(1, _gvn.transform( slow_result )); 4116 r->init_req(1, iftrue); 4117 4118 // Else fall through 4119 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4120 set_control(iffalse); 4121 4122 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg))); 4123 r->init_req(2, iffalse); 4124 4125 // Post merge 4126 set_control(_gvn.transform(r)); 4127 record_for_igvn(r); 4128 4129 C->set_has_split_ifs(true); // Has chance for split-if optimization 4130 result = phi; 4131 assert(result->bottom_type()->isa_long(), "must be"); 4132 break; 4133 } 4134 4135 case vmIntrinsics::_floatToIntBits: { 4136 // two paths (plus control) merge in a wood 4137 RegionNode *r = new (C) RegionNode(3); 4138 Node *phi = new (C) PhiNode(r, TypeInt::INT); 4139 4140 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg)); 4141 // Build the boolean node 4142 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4143 4144 // Branch either way. 4145 // NaN case is less traveled, which makes all the difference. 4146 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4147 Node *opt_isnan = _gvn.transform(ifisnan); 4148 assert( opt_isnan->is_If(), "Expect an IfNode"); 4149 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4150 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4151 4152 set_control(iftrue); 4153 4154 static const jint nan_bits = 0x7fc00000; 4155 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4156 phi->init_req(1, _gvn.transform( slow_result )); 4157 r->init_req(1, iftrue); 4158 4159 // Else fall through 4160 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4161 set_control(iffalse); 4162 4163 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg))); 4164 r->init_req(2, iffalse); 4165 4166 // Post merge 4167 set_control(_gvn.transform(r)); 4168 record_for_igvn(r); 4169 4170 C->set_has_split_ifs(true); // Has chance for split-if optimization 4171 result = phi; 4172 assert(result->bottom_type()->isa_int(), "must be"); 4173 break; 4174 } 4175 4176 default: 4177 fatal_unexpected_iid(id); 4178 break; 4179 } 4180 set_result(_gvn.transform(result)); 4181 return true; 4182} 4183 4184#ifdef _LP64 4185#define XTOP ,top() /*additional argument*/ 4186#else //_LP64 4187#define XTOP /*no additional argument*/ 4188#endif //_LP64 4189 4190//----------------------inline_unsafe_copyMemory------------------------- 4191// public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4192bool LibraryCallKit::inline_unsafe_copyMemory() { 4193 if (callee()->is_static()) return false; // caller must have the capability! 4194 null_check_receiver(); // null-check receiver 4195 if (stopped()) return true; 4196 4197 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4198 4199 Node* src_ptr = argument(1); // type: oop 4200 Node* src_off = ConvL2X(argument(2)); // type: long 4201 Node* dst_ptr = argument(4); // type: oop 4202 Node* dst_off = ConvL2X(argument(5)); // type: long 4203 Node* size = ConvL2X(argument(7)); // type: long 4204 4205 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4206 "fieldOffset must be byte-scaled"); 4207 4208 Node* src = make_unsafe_address(src_ptr, src_off); 4209 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4210 4211 // Conservatively insert a memory barrier on all memory slices. 4212 // Do not let writes of the copy source or destination float below the copy. 4213 insert_mem_bar(Op_MemBarCPUOrder); 4214 4215 // Call it. Note that the length argument is not scaled. 4216 make_runtime_call(RC_LEAF|RC_NO_FP, 4217 OptoRuntime::fast_arraycopy_Type(), 4218 StubRoutines::unsafe_arraycopy(), 4219 "unsafe_arraycopy", 4220 TypeRawPtr::BOTTOM, 4221 src, dst, size XTOP); 4222 4223 // Do not let reads of the copy destination float above the copy. 4224 insert_mem_bar(Op_MemBarCPUOrder); 4225 4226 return true; 4227} 4228 4229//------------------------clone_coping----------------------------------- 4230// Helper function for inline_native_clone. 4231void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4232 assert(obj_size != NULL, ""); 4233 Node* raw_obj = alloc_obj->in(1); 4234 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4235 4236 AllocateNode* alloc = NULL; 4237 if (ReduceBulkZeroing) { 4238 // We will be completely responsible for initializing this object - 4239 // mark Initialize node as complete. 4240 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4241 // The object was just allocated - there should be no any stores! 4242 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4243 // Mark as complete_with_arraycopy so that on AllocateNode 4244 // expansion, we know this AllocateNode is initialized by an array 4245 // copy and a StoreStore barrier exists after the array copy. 4246 alloc->initialization()->set_complete_with_arraycopy(); 4247 } 4248 4249 // Copy the fastest available way. 4250 // TODO: generate fields copies for small objects instead. 4251 Node* src = obj; 4252 Node* dest = alloc_obj; 4253 Node* size = _gvn.transform(obj_size); 4254 4255 // Exclude the header but include array length to copy by 8 bytes words. 4256 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4257 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4258 instanceOopDesc::base_offset_in_bytes(); 4259 // base_off: 4260 // 8 - 32-bit VM 4261 // 12 - 64-bit VM, compressed klass 4262 // 16 - 64-bit VM, normal klass 4263 if (base_off % BytesPerLong != 0) { 4264 assert(UseCompressedClassPointers, ""); 4265 if (is_array) { 4266 // Exclude length to copy by 8 bytes words. 4267 base_off += sizeof(int); 4268 } else { 4269 // Include klass to copy by 8 bytes words. 4270 base_off = instanceOopDesc::klass_offset_in_bytes(); 4271 } 4272 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4273 } 4274 src = basic_plus_adr(src, base_off); 4275 dest = basic_plus_adr(dest, base_off); 4276 4277 // Compute the length also, if needed: 4278 Node* countx = size; 4279 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off))); 4280 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4281 4282 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4283 bool disjoint_bases = true; 4284 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4285 src, NULL, dest, NULL, countx, 4286 /*dest_uninitialized*/true); 4287 4288 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4289 if (card_mark) { 4290 assert(!is_array, ""); 4291 // Put in store barrier for any and all oops we are sticking 4292 // into this object. (We could avoid this if we could prove 4293 // that the object type contains no oop fields at all.) 4294 Node* no_particular_value = NULL; 4295 Node* no_particular_field = NULL; 4296 int raw_adr_idx = Compile::AliasIdxRaw; 4297 post_barrier(control(), 4298 memory(raw_adr_type), 4299 alloc_obj, 4300 no_particular_field, 4301 raw_adr_idx, 4302 no_particular_value, 4303 T_OBJECT, 4304 false); 4305 } 4306 4307 // Do not let reads from the cloned object float above the arraycopy. 4308 if (alloc != NULL) { 4309 // Do not let stores that initialize this object be reordered with 4310 // a subsequent store that would make this object accessible by 4311 // other threads. 4312 // Record what AllocateNode this StoreStore protects so that 4313 // escape analysis can go from the MemBarStoreStoreNode to the 4314 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4315 // based on the escape status of the AllocateNode. 4316 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4317 } else { 4318 insert_mem_bar(Op_MemBarCPUOrder); 4319 } 4320} 4321 4322//------------------------inline_native_clone---------------------------- 4323// protected native Object java.lang.Object.clone(); 4324// 4325// Here are the simple edge cases: 4326// null receiver => normal trap 4327// virtual and clone was overridden => slow path to out-of-line clone 4328// not cloneable or finalizer => slow path to out-of-line Object.clone 4329// 4330// The general case has two steps, allocation and copying. 4331// Allocation has two cases, and uses GraphKit::new_instance or new_array. 4332// 4333// Copying also has two cases, oop arrays and everything else. 4334// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4335// Everything else uses the tight inline loop supplied by CopyArrayNode. 4336// 4337// These steps fold up nicely if and when the cloned object's klass 4338// can be sharply typed as an object array, a type array, or an instance. 4339// 4340bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4341 PhiNode* result_val; 4342 4343 // Set the reexecute bit for the interpreter to reexecute 4344 // the bytecode that invokes Object.clone if deoptimization happens. 4345 { PreserveReexecuteState preexecs(this); 4346 jvms()->set_should_reexecute(true); 4347 4348 Node* obj = null_check_receiver(); 4349 if (stopped()) return true; 4350 4351 Node* obj_klass = load_object_klass(obj); 4352 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4353 const TypeOopPtr* toop = ((tklass != NULL) 4354 ? tklass->as_instance_type() 4355 : TypeInstPtr::NOTNULL); 4356 4357 // Conservatively insert a memory barrier on all memory slices. 4358 // Do not let writes into the original float below the clone. 4359 insert_mem_bar(Op_MemBarCPUOrder); 4360 4361 // paths into result_reg: 4362 enum { 4363 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4364 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4365 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4366 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4367 PATH_LIMIT 4368 }; 4369 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4370 result_val = new(C) PhiNode(result_reg, 4371 TypeInstPtr::NOTNULL); 4372 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO); 4373 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 4374 TypePtr::BOTTOM); 4375 record_for_igvn(result_reg); 4376 4377 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4378 int raw_adr_idx = Compile::AliasIdxRaw; 4379 4380 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4381 if (array_ctl != NULL) { 4382 // It's an array. 4383 PreserveJVMState pjvms(this); 4384 set_control(array_ctl); 4385 Node* obj_length = load_array_length(obj); 4386 Node* obj_size = NULL; 4387 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4388 4389 if (!use_ReduceInitialCardMarks()) { 4390 // If it is an oop array, it requires very special treatment, 4391 // because card marking is required on each card of the array. 4392 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4393 if (is_obja != NULL) { 4394 PreserveJVMState pjvms2(this); 4395 set_control(is_obja); 4396 // Generate a direct call to the right arraycopy function(s). 4397 bool disjoint_bases = true; 4398 bool length_never_negative = true; 4399 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4400 obj, intcon(0), alloc_obj, intcon(0), 4401 obj_length, 4402 disjoint_bases, length_never_negative); 4403 result_reg->init_req(_objArray_path, control()); 4404 result_val->init_req(_objArray_path, alloc_obj); 4405 result_i_o ->set_req(_objArray_path, i_o()); 4406 result_mem ->set_req(_objArray_path, reset_memory()); 4407 } 4408 } 4409 // Otherwise, there are no card marks to worry about. 4410 // (We can dispense with card marks if we know the allocation 4411 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4412 // causes the non-eden paths to take compensating steps to 4413 // simulate a fresh allocation, so that no further 4414 // card marks are required in compiled code to initialize 4415 // the object.) 4416 4417 if (!stopped()) { 4418 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4419 4420 // Present the results of the copy. 4421 result_reg->init_req(_array_path, control()); 4422 result_val->init_req(_array_path, alloc_obj); 4423 result_i_o ->set_req(_array_path, i_o()); 4424 result_mem ->set_req(_array_path, reset_memory()); 4425 } 4426 } 4427 4428 // We only go to the instance fast case code if we pass a number of guards. 4429 // The paths which do not pass are accumulated in the slow_region. 4430 RegionNode* slow_region = new (C) RegionNode(1); 4431 record_for_igvn(slow_region); 4432 if (!stopped()) { 4433 // It's an instance (we did array above). Make the slow-path tests. 4434 // If this is a virtual call, we generate a funny guard. We grab 4435 // the vtable entry corresponding to clone() from the target object. 4436 // If the target method which we are calling happens to be the 4437 // Object clone() method, we pass the guard. We do not need this 4438 // guard for non-virtual calls; the caller is known to be the native 4439 // Object clone(). 4440 if (is_virtual) { 4441 generate_virtual_guard(obj_klass, slow_region); 4442 } 4443 4444 // The object must be cloneable and must not have a finalizer. 4445 // Both of these conditions may be checked in a single test. 4446 // We could optimize the cloneable test further, but we don't care. 4447 generate_access_flags_guard(obj_klass, 4448 // Test both conditions: 4449 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4450 // Must be cloneable but not finalizer: 4451 JVM_ACC_IS_CLONEABLE, 4452 slow_region); 4453 } 4454 4455 if (!stopped()) { 4456 // It's an instance, and it passed the slow-path tests. 4457 PreserveJVMState pjvms(this); 4458 Node* obj_size = NULL; 4459 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size); 4460 4461 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4462 4463 // Present the results of the slow call. 4464 result_reg->init_req(_instance_path, control()); 4465 result_val->init_req(_instance_path, alloc_obj); 4466 result_i_o ->set_req(_instance_path, i_o()); 4467 result_mem ->set_req(_instance_path, reset_memory()); 4468 } 4469 4470 // Generate code for the slow case. We make a call to clone(). 4471 set_control(_gvn.transform(slow_region)); 4472 if (!stopped()) { 4473 PreserveJVMState pjvms(this); 4474 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4475 Node* slow_result = set_results_for_java_call(slow_call); 4476 // this->control() comes from set_results_for_java_call 4477 result_reg->init_req(_slow_path, control()); 4478 result_val->init_req(_slow_path, slow_result); 4479 result_i_o ->set_req(_slow_path, i_o()); 4480 result_mem ->set_req(_slow_path, reset_memory()); 4481 } 4482 4483 // Return the combined state. 4484 set_control( _gvn.transform(result_reg)); 4485 set_i_o( _gvn.transform(result_i_o)); 4486 set_all_memory( _gvn.transform(result_mem)); 4487 } // original reexecute is set back here 4488 4489 set_result(_gvn.transform(result_val)); 4490 return true; 4491} 4492 4493//------------------------------basictype2arraycopy---------------------------- 4494address LibraryCallKit::basictype2arraycopy(BasicType t, 4495 Node* src_offset, 4496 Node* dest_offset, 4497 bool disjoint_bases, 4498 const char* &name, 4499 bool dest_uninitialized) { 4500 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4501 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4502 4503 bool aligned = false; 4504 bool disjoint = disjoint_bases; 4505 4506 // if the offsets are the same, we can treat the memory regions as 4507 // disjoint, because either the memory regions are in different arrays, 4508 // or they are identical (which we can treat as disjoint.) We can also 4509 // treat a copy with a destination index less that the source index 4510 // as disjoint since a low->high copy will work correctly in this case. 4511 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4512 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4513 // both indices are constants 4514 int s_offs = src_offset_inttype->get_con(); 4515 int d_offs = dest_offset_inttype->get_con(); 4516 int element_size = type2aelembytes(t); 4517 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4518 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4519 if (s_offs >= d_offs) disjoint = true; 4520 } else if (src_offset == dest_offset && src_offset != NULL) { 4521 // This can occur if the offsets are identical non-constants. 4522 disjoint = true; 4523 } 4524 4525 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4526} 4527 4528 4529//------------------------------inline_arraycopy----------------------- 4530// public static native void java.lang.System.arraycopy(Object src, int srcPos, 4531// Object dest, int destPos, 4532// int length); 4533bool LibraryCallKit::inline_arraycopy() { 4534 // Get the arguments. 4535 Node* src = argument(0); // type: oop 4536 Node* src_offset = argument(1); // type: int 4537 Node* dest = argument(2); // type: oop 4538 Node* dest_offset = argument(3); // type: int 4539 Node* length = argument(4); // type: int 4540 4541 // Compile time checks. If any of these checks cannot be verified at compile time, 4542 // we do not make a fast path for this call. Instead, we let the call remain as it 4543 // is. The checks we choose to mandate at compile time are: 4544 // 4545 // (1) src and dest are arrays. 4546 const Type* src_type = src->Value(&_gvn); 4547 const Type* dest_type = dest->Value(&_gvn); 4548 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4549 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4550 if (top_src == NULL || top_src->klass() == NULL || 4551 top_dest == NULL || top_dest->klass() == NULL) { 4552 // Conservatively insert a memory barrier on all memory slices. 4553 // Do not let writes into the source float below the arraycopy. 4554 insert_mem_bar(Op_MemBarCPUOrder); 4555 4556 // Call StubRoutines::generic_arraycopy stub. 4557 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4558 src, src_offset, dest, dest_offset, length); 4559 4560 // Do not let reads from the destination float above the arraycopy. 4561 // Since we cannot type the arrays, we don't know which slices 4562 // might be affected. We could restrict this barrier only to those 4563 // memory slices which pertain to array elements--but don't bother. 4564 if (!InsertMemBarAfterArraycopy) 4565 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4566 insert_mem_bar(Op_MemBarCPUOrder); 4567 return true; 4568 } 4569 4570 // (2) src and dest arrays must have elements of the same BasicType 4571 // Figure out the size and type of the elements we will be copying. 4572 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4573 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4574 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4575 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4576 4577 if (src_elem != dest_elem || dest_elem == T_VOID) { 4578 // The component types are not the same or are not recognized. Punt. 4579 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4580 generate_slow_arraycopy(TypePtr::BOTTOM, 4581 src, src_offset, dest, dest_offset, length, 4582 /*dest_uninitialized*/false); 4583 return true; 4584 } 4585 4586 //--------------------------------------------------------------------------- 4587 // We will make a fast path for this call to arraycopy. 4588 4589 // We have the following tests left to perform: 4590 // 4591 // (3) src and dest must not be null. 4592 // (4) src_offset must not be negative. 4593 // (5) dest_offset must not be negative. 4594 // (6) length must not be negative. 4595 // (7) src_offset + length must not exceed length of src. 4596 // (8) dest_offset + length must not exceed length of dest. 4597 // (9) each element of an oop array must be assignable 4598 4599 RegionNode* slow_region = new (C) RegionNode(1); 4600 record_for_igvn(slow_region); 4601 4602 // (3) operands must not be null 4603 // We currently perform our null checks with the null_check routine. 4604 // This means that the null exceptions will be reported in the caller 4605 // rather than (correctly) reported inside of the native arraycopy call. 4606 // This should be corrected, given time. We do our null check with the 4607 // stack pointer restored. 4608 src = null_check(src, T_ARRAY); 4609 dest = null_check(dest, T_ARRAY); 4610 4611 // (4) src_offset must not be negative. 4612 generate_negative_guard(src_offset, slow_region); 4613 4614 // (5) dest_offset must not be negative. 4615 generate_negative_guard(dest_offset, slow_region); 4616 4617 // (6) length must not be negative (moved to generate_arraycopy()). 4618 // generate_negative_guard(length, slow_region); 4619 4620 // (7) src_offset + length must not exceed length of src. 4621 generate_limit_guard(src_offset, length, 4622 load_array_length(src), 4623 slow_region); 4624 4625 // (8) dest_offset + length must not exceed length of dest. 4626 generate_limit_guard(dest_offset, length, 4627 load_array_length(dest), 4628 slow_region); 4629 4630 // (9) each element of an oop array must be assignable 4631 // The generate_arraycopy subroutine checks this. 4632 4633 // This is where the memory effects are placed: 4634 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4635 generate_arraycopy(adr_type, dest_elem, 4636 src, src_offset, dest, dest_offset, length, 4637 false, false, slow_region); 4638 4639 return true; 4640} 4641 4642//-----------------------------generate_arraycopy---------------------- 4643// Generate an optimized call to arraycopy. 4644// Caller must guard against non-arrays. 4645// Caller must determine a common array basic-type for both arrays. 4646// Caller must validate offsets against array bounds. 4647// The slow_region has already collected guard failure paths 4648// (such as out of bounds length or non-conformable array types). 4649// The generated code has this shape, in general: 4650// 4651// if (length == 0) return // via zero_path 4652// slowval = -1 4653// if (types unknown) { 4654// slowval = call generic copy loop 4655// if (slowval == 0) return // via checked_path 4656// } else if (indexes in bounds) { 4657// if ((is object array) && !(array type check)) { 4658// slowval = call checked copy loop 4659// if (slowval == 0) return // via checked_path 4660// } else { 4661// call bulk copy loop 4662// return // via fast_path 4663// } 4664// } 4665// // adjust params for remaining work: 4666// if (slowval != -1) { 4667// n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4668// } 4669// slow_region: 4670// call slow arraycopy(src, src_offset, dest, dest_offset, length) 4671// return // via slow_call_path 4672// 4673// This routine is used from several intrinsics: System.arraycopy, 4674// Object.clone (the array subcase), and Arrays.copyOf[Range]. 4675// 4676void 4677LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4678 BasicType basic_elem_type, 4679 Node* src, Node* src_offset, 4680 Node* dest, Node* dest_offset, 4681 Node* copy_length, 4682 bool disjoint_bases, 4683 bool length_never_negative, 4684 RegionNode* slow_region) { 4685 4686 if (slow_region == NULL) { 4687 slow_region = new(C) RegionNode(1); 4688 record_for_igvn(slow_region); 4689 } 4690 4691 Node* original_dest = dest; 4692 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4693 bool dest_uninitialized = false; 4694 4695 // See if this is the initialization of a newly-allocated array. 4696 // If so, we will take responsibility here for initializing it to zero. 4697 // (Note: Because tightly_coupled_allocation performs checks on the 4698 // out-edges of the dest, we need to avoid making derived pointers 4699 // from it until we have checked its uses.) 4700 if (ReduceBulkZeroing 4701 && !ZeroTLAB // pointless if already zeroed 4702 && basic_elem_type != T_CONFLICT // avoid corner case 4703 && !src->eqv_uncast(dest) 4704 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4705 != NULL) 4706 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4707 && alloc->maybe_set_complete(&_gvn)) { 4708 // "You break it, you buy it." 4709 InitializeNode* init = alloc->initialization(); 4710 assert(init->is_complete(), "we just did this"); 4711 init->set_complete_with_arraycopy(); 4712 assert(dest->is_CheckCastPP(), "sanity"); 4713 assert(dest->in(0)->in(0) == init, "dest pinned"); 4714 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4715 // From this point on, every exit path is responsible for 4716 // initializing any non-copied parts of the object to zero. 4717 // Also, if this flag is set we make sure that arraycopy interacts properly 4718 // with G1, eliding pre-barriers. See CR 6627983. 4719 dest_uninitialized = true; 4720 } else { 4721 // No zeroing elimination here. 4722 alloc = NULL; 4723 //original_dest = dest; 4724 //dest_uninitialized = false; 4725 } 4726 4727 // Results are placed here: 4728 enum { fast_path = 1, // normal void-returning assembly stub 4729 checked_path = 2, // special assembly stub with cleanup 4730 slow_call_path = 3, // something went wrong; call the VM 4731 zero_path = 4, // bypass when length of copy is zero 4732 bcopy_path = 5, // copy primitive array by 64-bit blocks 4733 PATH_LIMIT = 6 4734 }; 4735 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT); 4736 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO); 4737 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type); 4738 record_for_igvn(result_region); 4739 _gvn.set_type_bottom(result_i_o); 4740 _gvn.set_type_bottom(result_memory); 4741 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4742 4743 // The slow_control path: 4744 Node* slow_control; 4745 Node* slow_i_o = i_o(); 4746 Node* slow_mem = memory(adr_type); 4747 debug_only(slow_control = (Node*) badAddress); 4748 4749 // Checked control path: 4750 Node* checked_control = top(); 4751 Node* checked_mem = NULL; 4752 Node* checked_i_o = NULL; 4753 Node* checked_value = NULL; 4754 4755 if (basic_elem_type == T_CONFLICT) { 4756 assert(!dest_uninitialized, ""); 4757 Node* cv = generate_generic_arraycopy(adr_type, 4758 src, src_offset, dest, dest_offset, 4759 copy_length, dest_uninitialized); 4760 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4761 checked_control = control(); 4762 checked_i_o = i_o(); 4763 checked_mem = memory(adr_type); 4764 checked_value = cv; 4765 set_control(top()); // no fast path 4766 } 4767 4768 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4769 if (not_pos != NULL) { 4770 PreserveJVMState pjvms(this); 4771 set_control(not_pos); 4772 4773 // (6) length must not be negative. 4774 if (!length_never_negative) { 4775 generate_negative_guard(copy_length, slow_region); 4776 } 4777 4778 // copy_length is 0. 4779 if (!stopped() && dest_uninitialized) { 4780 Node* dest_length = alloc->in(AllocateNode::ALength); 4781 if (copy_length->eqv_uncast(dest_length) 4782 || _gvn.find_int_con(dest_length, 1) <= 0) { 4783 // There is no zeroing to do. No need for a secondary raw memory barrier. 4784 } else { 4785 // Clear the whole thing since there are no source elements to copy. 4786 generate_clear_array(adr_type, dest, basic_elem_type, 4787 intcon(0), NULL, 4788 alloc->in(AllocateNode::AllocSize)); 4789 // Use a secondary InitializeNode as raw memory barrier. 4790 // Currently it is needed only on this path since other 4791 // paths have stub or runtime calls as raw memory barriers. 4792 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4793 Compile::AliasIdxRaw, 4794 top())->as_Initialize(); 4795 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4796 } 4797 } 4798 4799 // Present the results of the fast call. 4800 result_region->init_req(zero_path, control()); 4801 result_i_o ->init_req(zero_path, i_o()); 4802 result_memory->init_req(zero_path, memory(adr_type)); 4803 } 4804 4805 if (!stopped() && dest_uninitialized) { 4806 // We have to initialize the *uncopied* part of the array to zero. 4807 // The copy destination is the slice dest[off..off+len]. The other slices 4808 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4809 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4810 Node* dest_length = alloc->in(AllocateNode::ALength); 4811 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset, 4812 copy_length)); 4813 4814 // If there is a head section that needs zeroing, do it now. 4815 if (find_int_con(dest_offset, -1) != 0) { 4816 generate_clear_array(adr_type, dest, basic_elem_type, 4817 intcon(0), dest_offset, 4818 NULL); 4819 } 4820 4821 // Next, perform a dynamic check on the tail length. 4822 // It is often zero, and we can win big if we prove this. 4823 // There are two wins: Avoid generating the ClearArray 4824 // with its attendant messy index arithmetic, and upgrade 4825 // the copy to a more hardware-friendly word size of 64 bits. 4826 Node* tail_ctl = NULL; 4827 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 4828 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length)); 4829 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt)); 4830 tail_ctl = generate_slow_guard(bol_lt, NULL); 4831 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4832 } 4833 4834 // At this point, let's assume there is no tail. 4835 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4836 // There is no tail. Try an upgrade to a 64-bit copy. 4837 bool didit = false; 4838 { PreserveJVMState pjvms(this); 4839 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4840 src, src_offset, dest, dest_offset, 4841 dest_size, dest_uninitialized); 4842 if (didit) { 4843 // Present the results of the block-copying fast call. 4844 result_region->init_req(bcopy_path, control()); 4845 result_i_o ->init_req(bcopy_path, i_o()); 4846 result_memory->init_req(bcopy_path, memory(adr_type)); 4847 } 4848 } 4849 if (didit) 4850 set_control(top()); // no regular fast path 4851 } 4852 4853 // Clear the tail, if any. 4854 if (tail_ctl != NULL) { 4855 Node* notail_ctl = stopped() ? NULL : control(); 4856 set_control(tail_ctl); 4857 if (notail_ctl == NULL) { 4858 generate_clear_array(adr_type, dest, basic_elem_type, 4859 dest_tail, NULL, 4860 dest_size); 4861 } else { 4862 // Make a local merge. 4863 Node* done_ctl = new(C) RegionNode(3); 4864 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type); 4865 done_ctl->init_req(1, notail_ctl); 4866 done_mem->init_req(1, memory(adr_type)); 4867 generate_clear_array(adr_type, dest, basic_elem_type, 4868 dest_tail, NULL, 4869 dest_size); 4870 done_ctl->init_req(2, control()); 4871 done_mem->init_req(2, memory(adr_type)); 4872 set_control( _gvn.transform(done_ctl)); 4873 set_memory( _gvn.transform(done_mem), adr_type ); 4874 } 4875 } 4876 } 4877 4878 BasicType copy_type = basic_elem_type; 4879 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 4880 if (!stopped() && copy_type == T_OBJECT) { 4881 // If src and dest have compatible element types, we can copy bits. 4882 // Types S[] and D[] are compatible if D is a supertype of S. 4883 // 4884 // If they are not, we will use checked_oop_disjoint_arraycopy, 4885 // which performs a fast optimistic per-oop check, and backs off 4886 // further to JVM_ArrayCopy on the first per-oop check that fails. 4887 // (Actually, we don't move raw bits only; the GC requires card marks.) 4888 4889 // Get the Klass* for both src and dest 4890 Node* src_klass = load_object_klass(src); 4891 Node* dest_klass = load_object_klass(dest); 4892 4893 // Generate the subtype check. 4894 // This might fold up statically, or then again it might not. 4895 // 4896 // Non-static example: Copying List<String>.elements to a new String[]. 4897 // The backing store for a List<String> is always an Object[], 4898 // but its elements are always type String, if the generic types 4899 // are correct at the source level. 4900 // 4901 // Test S[] against D[], not S against D, because (probably) 4902 // the secondary supertype cache is less busy for S[] than S. 4903 // This usually only matters when D is an interface. 4904 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4905 // Plug failing path into checked_oop_disjoint_arraycopy 4906 if (not_subtype_ctrl != top()) { 4907 PreserveJVMState pjvms(this); 4908 set_control(not_subtype_ctrl); 4909 // (At this point we can assume disjoint_bases, since types differ.) 4910 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 4911 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 4912 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 4913 Node* dest_elem_klass = _gvn.transform(n1); 4914 Node* cv = generate_checkcast_arraycopy(adr_type, 4915 dest_elem_klass, 4916 src, src_offset, dest, dest_offset, 4917 ConvI2X(copy_length), dest_uninitialized); 4918 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4919 checked_control = control(); 4920 checked_i_o = i_o(); 4921 checked_mem = memory(adr_type); 4922 checked_value = cv; 4923 } 4924 // At this point we know we do not need type checks on oop stores. 4925 4926 // Let's see if we need card marks: 4927 if (alloc != NULL && use_ReduceInitialCardMarks()) { 4928 // If we do not need card marks, copy using the jint or jlong stub. 4929 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 4930 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 4931 "sizes agree"); 4932 } 4933 } 4934 4935 if (!stopped()) { 4936 // Generate the fast path, if possible. 4937 PreserveJVMState pjvms(this); 4938 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 4939 src, src_offset, dest, dest_offset, 4940 ConvI2X(copy_length), dest_uninitialized); 4941 4942 // Present the results of the fast call. 4943 result_region->init_req(fast_path, control()); 4944 result_i_o ->init_req(fast_path, i_o()); 4945 result_memory->init_req(fast_path, memory(adr_type)); 4946 } 4947 4948 // Here are all the slow paths up to this point, in one bundle: 4949 slow_control = top(); 4950 if (slow_region != NULL) 4951 slow_control = _gvn.transform(slow_region); 4952 DEBUG_ONLY(slow_region = (RegionNode*)badAddress); 4953 4954 set_control(checked_control); 4955 if (!stopped()) { 4956 // Clean up after the checked call. 4957 // The returned value is either 0 or -1^K, 4958 // where K = number of partially transferred array elements. 4959 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0))); 4960 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 4961 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4962 4963 // If it is 0, we are done, so transfer to the end. 4964 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff)); 4965 result_region->init_req(checked_path, checks_done); 4966 result_i_o ->init_req(checked_path, checked_i_o); 4967 result_memory->init_req(checked_path, checked_mem); 4968 4969 // If it is not zero, merge into the slow call. 4970 set_control( _gvn.transform(new(C) IfFalseNode(iff) )); 4971 RegionNode* slow_reg2 = new(C) RegionNode(3); 4972 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO); 4973 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type); 4974 record_for_igvn(slow_reg2); 4975 slow_reg2 ->init_req(1, slow_control); 4976 slow_i_o2 ->init_req(1, slow_i_o); 4977 slow_mem2 ->init_req(1, slow_mem); 4978 slow_reg2 ->init_req(2, control()); 4979 slow_i_o2 ->init_req(2, checked_i_o); 4980 slow_mem2 ->init_req(2, checked_mem); 4981 4982 slow_control = _gvn.transform(slow_reg2); 4983 slow_i_o = _gvn.transform(slow_i_o2); 4984 slow_mem = _gvn.transform(slow_mem2); 4985 4986 if (alloc != NULL) { 4987 // We'll restart from the very beginning, after zeroing the whole thing. 4988 // This can cause double writes, but that's OK since dest is brand new. 4989 // So we ignore the low 31 bits of the value returned from the stub. 4990 } else { 4991 // We must continue the copy exactly where it failed, or else 4992 // another thread might see the wrong number of writes to dest. 4993 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1))); 4994 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT); 4995 slow_offset->init_req(1, intcon(0)); 4996 slow_offset->init_req(2, checked_offset); 4997 slow_offset = _gvn.transform(slow_offset); 4998 4999 // Adjust the arguments by the conditionally incoming offset. 5000 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset)); 5001 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset)); 5002 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset)); 5003 5004 // Tweak the node variables to adjust the code produced below: 5005 src_offset = src_off_plus; 5006 dest_offset = dest_off_plus; 5007 copy_length = length_minus; 5008 } 5009 } 5010 5011 set_control(slow_control); 5012 if (!stopped()) { 5013 // Generate the slow path, if needed. 5014 PreserveJVMState pjvms(this); // replace_in_map may trash the map 5015 5016 set_memory(slow_mem, adr_type); 5017 set_i_o(slow_i_o); 5018 5019 if (dest_uninitialized) { 5020 generate_clear_array(adr_type, dest, basic_elem_type, 5021 intcon(0), NULL, 5022 alloc->in(AllocateNode::AllocSize)); 5023 } 5024 5025 generate_slow_arraycopy(adr_type, 5026 src, src_offset, dest, dest_offset, 5027 copy_length, /*dest_uninitialized*/false); 5028 5029 result_region->init_req(slow_call_path, control()); 5030 result_i_o ->init_req(slow_call_path, i_o()); 5031 result_memory->init_req(slow_call_path, memory(adr_type)); 5032 } 5033 5034 // Remove unused edges. 5035 for (uint i = 1; i < result_region->req(); i++) { 5036 if (result_region->in(i) == NULL) 5037 result_region->init_req(i, top()); 5038 } 5039 5040 // Finished; return the combined state. 5041 set_control( _gvn.transform(result_region)); 5042 set_i_o( _gvn.transform(result_i_o) ); 5043 set_memory( _gvn.transform(result_memory), adr_type ); 5044 5045 // The memory edges above are precise in order to model effects around 5046 // array copies accurately to allow value numbering of field loads around 5047 // arraycopy. Such field loads, both before and after, are common in Java 5048 // collections and similar classes involving header/array data structures. 5049 // 5050 // But with low number of register or when some registers are used or killed 5051 // by arraycopy calls it causes registers spilling on stack. See 6544710. 5052 // The next memory barrier is added to avoid it. If the arraycopy can be 5053 // optimized away (which it can, sometimes) then we can manually remove 5054 // the membar also. 5055 // 5056 // Do not let reads from the cloned object float above the arraycopy. 5057 if (alloc != NULL) { 5058 // Do not let stores that initialize this object be reordered with 5059 // a subsequent store that would make this object accessible by 5060 // other threads. 5061 // Record what AllocateNode this StoreStore protects so that 5062 // escape analysis can go from the MemBarStoreStoreNode to the 5063 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5064 // based on the escape status of the AllocateNode. 5065 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5066 } else if (InsertMemBarAfterArraycopy) 5067 insert_mem_bar(Op_MemBarCPUOrder); 5068} 5069 5070 5071// Helper function which determines if an arraycopy immediately follows 5072// an allocation, with no intervening tests or other escapes for the object. 5073AllocateArrayNode* 5074LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5075 RegionNode* slow_region) { 5076 if (stopped()) return NULL; // no fast path 5077 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5078 5079 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5080 if (alloc == NULL) return NULL; 5081 5082 Node* rawmem = memory(Compile::AliasIdxRaw); 5083 // Is the allocation's memory state untouched? 5084 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5085 // Bail out if there have been raw-memory effects since the allocation. 5086 // (Example: There might have been a call or safepoint.) 5087 return NULL; 5088 } 5089 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5090 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5091 return NULL; 5092 } 5093 5094 // There must be no unexpected observers of this allocation. 5095 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5096 Node* obs = ptr->fast_out(i); 5097 if (obs != this->map()) { 5098 return NULL; 5099 } 5100 } 5101 5102 // This arraycopy must unconditionally follow the allocation of the ptr. 5103 Node* alloc_ctl = ptr->in(0); 5104 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5105 5106 Node* ctl = control(); 5107 while (ctl != alloc_ctl) { 5108 // There may be guards which feed into the slow_region. 5109 // Any other control flow means that we might not get a chance 5110 // to finish initializing the allocated object. 5111 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5112 IfNode* iff = ctl->in(0)->as_If(); 5113 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5114 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5115 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5116 ctl = iff->in(0); // This test feeds the known slow_region. 5117 continue; 5118 } 5119 // One more try: Various low-level checks bottom out in 5120 // uncommon traps. If the debug-info of the trap omits 5121 // any reference to the allocation, as we've already 5122 // observed, then there can be no objection to the trap. 5123 bool found_trap = false; 5124 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5125 Node* obs = not_ctl->fast_out(j); 5126 if (obs->in(0) == not_ctl && obs->is_Call() && 5127 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5128 found_trap = true; break; 5129 } 5130 } 5131 if (found_trap) { 5132 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5133 continue; 5134 } 5135 } 5136 return NULL; 5137 } 5138 5139 // If we get this far, we have an allocation which immediately 5140 // precedes the arraycopy, and we can take over zeroing the new object. 5141 // The arraycopy will finish the initialization, and provide 5142 // a new control state to which we will anchor the destination pointer. 5143 5144 return alloc; 5145} 5146 5147// Helper for initialization of arrays, creating a ClearArray. 5148// It writes zero bits in [start..end), within the body of an array object. 5149// The memory effects are all chained onto the 'adr_type' alias category. 5150// 5151// Since the object is otherwise uninitialized, we are free 5152// to put a little "slop" around the edges of the cleared area, 5153// as long as it does not go back into the array's header, 5154// or beyond the array end within the heap. 5155// 5156// The lower edge can be rounded down to the nearest jint and the 5157// upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5158// 5159// Arguments: 5160// adr_type memory slice where writes are generated 5161// dest oop of the destination array 5162// basic_elem_type element type of the destination 5163// slice_idx array index of first element to store 5164// slice_len number of elements to store (or NULL) 5165// dest_size total size in bytes of the array object 5166// 5167// Exactly one of slice_len or dest_size must be non-NULL. 5168// If dest_size is non-NULL, zeroing extends to the end of the object. 5169// If slice_len is non-NULL, the slice_idx value must be a constant. 5170void 5171LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5172 Node* dest, 5173 BasicType basic_elem_type, 5174 Node* slice_idx, 5175 Node* slice_len, 5176 Node* dest_size) { 5177 // one or the other but not both of slice_len and dest_size: 5178 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5179 if (slice_len == NULL) slice_len = top(); 5180 if (dest_size == NULL) dest_size = top(); 5181 5182 // operate on this memory slice: 5183 Node* mem = memory(adr_type); // memory slice to operate on 5184 5185 // scaling and rounding of indexes: 5186 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5187 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5188 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5189 int bump_bit = (-1 << scale) & BytesPerInt; 5190 5191 // determine constant starts and ends 5192 const intptr_t BIG_NEG = -128; 5193 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5194 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5195 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5196 if (slice_len_con == 0) { 5197 return; // nothing to do here 5198 } 5199 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5200 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5201 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5202 assert(end_con < 0, "not two cons"); 5203 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5204 BytesPerLong); 5205 } 5206 5207 if (start_con >= 0 && end_con >= 0) { 5208 // Constant start and end. Simple. 5209 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5210 start_con, end_con, &_gvn); 5211 } else if (start_con >= 0 && dest_size != top()) { 5212 // Constant start, pre-rounded end after the tail of the array. 5213 Node* end = dest_size; 5214 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5215 start_con, end, &_gvn); 5216 } else if (start_con >= 0 && slice_len != top()) { 5217 // Constant start, non-constant end. End needs rounding up. 5218 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5219 intptr_t end_base = abase + (slice_idx_con << scale); 5220 int end_round = (-1 << scale) & (BytesPerLong - 1); 5221 Node* end = ConvI2X(slice_len); 5222 if (scale != 0) 5223 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) )); 5224 end_base += end_round; 5225 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base))); 5226 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round))); 5227 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5228 start_con, end, &_gvn); 5229 } else if (start_con < 0 && dest_size != top()) { 5230 // Non-constant start, pre-rounded end after the tail of the array. 5231 // This is almost certainly a "round-to-end" operation. 5232 Node* start = slice_idx; 5233 start = ConvI2X(start); 5234 if (scale != 0) 5235 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) )); 5236 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase))); 5237 if ((bump_bit | clear_low) != 0) { 5238 int to_clear = (bump_bit | clear_low); 5239 // Align up mod 8, then store a jint zero unconditionally 5240 // just before the mod-8 boundary. 5241 if (((abase + bump_bit) & ~to_clear) - bump_bit 5242 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5243 bump_bit = 0; 5244 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5245 } else { 5246 // Bump 'start' up to (or past) the next jint boundary: 5247 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit))); 5248 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5249 } 5250 // Round bumped 'start' down to jlong boundary in body of array. 5251 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear))); 5252 if (bump_bit != 0) { 5253 // Store a zero to the immediately preceding jint: 5254 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit))); 5255 Node* p1 = basic_plus_adr(dest, x1); 5256 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 5257 mem = _gvn.transform(mem); 5258 } 5259 } 5260 Node* end = dest_size; // pre-rounded 5261 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5262 start, end, &_gvn); 5263 } else { 5264 // Non-constant start, unrounded non-constant end. 5265 // (Nobody zeroes a random midsection of an array using this routine.) 5266 ShouldNotReachHere(); // fix caller 5267 } 5268 5269 // Done. 5270 set_memory(mem, adr_type); 5271} 5272 5273 5274bool 5275LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5276 BasicType basic_elem_type, 5277 AllocateNode* alloc, 5278 Node* src, Node* src_offset, 5279 Node* dest, Node* dest_offset, 5280 Node* dest_size, bool dest_uninitialized) { 5281 // See if there is an advantage from block transfer. 5282 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5283 if (scale >= LogBytesPerLong) 5284 return false; // it is already a block transfer 5285 5286 // Look at the alignment of the starting offsets. 5287 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5288 5289 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5290 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5291 if (src_off_con < 0 || dest_off_con < 0) 5292 // At present, we can only understand constants. 5293 return false; 5294 5295 intptr_t src_off = abase + (src_off_con << scale); 5296 intptr_t dest_off = abase + (dest_off_con << scale); 5297 5298 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5299 // Non-aligned; too bad. 5300 // One more chance: Pick off an initial 32-bit word. 5301 // This is a common case, since abase can be odd mod 8. 5302 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5303 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5304 Node* sptr = basic_plus_adr(src, src_off); 5305 Node* dptr = basic_plus_adr(dest, dest_off); 5306 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 5307 store_to_memory(control(), dptr, sval, T_INT, adr_type); 5308 src_off += BytesPerInt; 5309 dest_off += BytesPerInt; 5310 } else { 5311 return false; 5312 } 5313 } 5314 assert(src_off % BytesPerLong == 0, ""); 5315 assert(dest_off % BytesPerLong == 0, ""); 5316 5317 // Do this copy by giant steps. 5318 Node* sptr = basic_plus_adr(src, src_off); 5319 Node* dptr = basic_plus_adr(dest, dest_off); 5320 Node* countx = dest_size; 5321 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off))); 5322 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong))); 5323 5324 bool disjoint_bases = true; // since alloc != NULL 5325 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5326 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5327 5328 return true; 5329} 5330 5331 5332// Helper function; generates code for the slow case. 5333// We make a call to a runtime method which emulates the native method, 5334// but without the native wrapper overhead. 5335void 5336LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5337 Node* src, Node* src_offset, 5338 Node* dest, Node* dest_offset, 5339 Node* copy_length, bool dest_uninitialized) { 5340 assert(!dest_uninitialized, "Invariant"); 5341 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5342 OptoRuntime::slow_arraycopy_Type(), 5343 OptoRuntime::slow_arraycopy_Java(), 5344 "slow_arraycopy", adr_type, 5345 src, src_offset, dest, dest_offset, 5346 copy_length); 5347 5348 // Handle exceptions thrown by this fellow: 5349 make_slow_call_ex(call, env()->Throwable_klass(), false); 5350} 5351 5352// Helper function; generates code for cases requiring runtime checks. 5353Node* 5354LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5355 Node* dest_elem_klass, 5356 Node* src, Node* src_offset, 5357 Node* dest, Node* dest_offset, 5358 Node* copy_length, bool dest_uninitialized) { 5359 if (stopped()) return NULL; 5360 5361 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5362 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5363 return NULL; 5364 } 5365 5366 // Pick out the parameters required to perform a store-check 5367 // for the target array. This is an optimistic check. It will 5368 // look in each non-null element's class, at the desired klass's 5369 // super_check_offset, for the desired klass. 5370 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5371 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5372 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr()); 5373 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5374 Node* check_value = dest_elem_klass; 5375 5376 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5377 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5378 5379 // (We know the arrays are never conjoint, because their types differ.) 5380 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5381 OptoRuntime::checkcast_arraycopy_Type(), 5382 copyfunc_addr, "checkcast_arraycopy", adr_type, 5383 // five arguments, of which two are 5384 // intptr_t (jlong in LP64) 5385 src_start, dest_start, 5386 copy_length XTOP, 5387 check_offset XTOP, 5388 check_value); 5389 5390 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5391} 5392 5393 5394// Helper function; generates code for cases requiring runtime checks. 5395Node* 5396LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5397 Node* src, Node* src_offset, 5398 Node* dest, Node* dest_offset, 5399 Node* copy_length, bool dest_uninitialized) { 5400 assert(!dest_uninitialized, "Invariant"); 5401 if (stopped()) return NULL; 5402 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5403 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5404 return NULL; 5405 } 5406 5407 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5408 OptoRuntime::generic_arraycopy_Type(), 5409 copyfunc_addr, "generic_arraycopy", adr_type, 5410 src, src_offset, dest, dest_offset, copy_length); 5411 5412 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5413} 5414 5415// Helper function; generates the fast out-of-line call to an arraycopy stub. 5416void 5417LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5418 BasicType basic_elem_type, 5419 bool disjoint_bases, 5420 Node* src, Node* src_offset, 5421 Node* dest, Node* dest_offset, 5422 Node* copy_length, bool dest_uninitialized) { 5423 if (stopped()) return; // nothing to do 5424 5425 Node* src_start = src; 5426 Node* dest_start = dest; 5427 if (src_offset != NULL || dest_offset != NULL) { 5428 assert(src_offset != NULL && dest_offset != NULL, ""); 5429 src_start = array_element_address(src, src_offset, basic_elem_type); 5430 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5431 } 5432 5433 // Figure out which arraycopy runtime method to call. 5434 const char* copyfunc_name = "arraycopy"; 5435 address copyfunc_addr = 5436 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5437 disjoint_bases, copyfunc_name, dest_uninitialized); 5438 5439 // Call it. Note that the count_ix value is not scaled to a byte-size. 5440 make_runtime_call(RC_LEAF|RC_NO_FP, 5441 OptoRuntime::fast_arraycopy_Type(), 5442 copyfunc_addr, copyfunc_name, adr_type, 5443 src_start, dest_start, copy_length XTOP); 5444} 5445 5446//-------------inline_encodeISOArray----------------------------------- 5447// encode char[] to byte[] in ISO_8859_1 5448bool LibraryCallKit::inline_encodeISOArray() { 5449 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5450 // no receiver since it is static method 5451 Node *src = argument(0); 5452 Node *src_offset = argument(1); 5453 Node *dst = argument(2); 5454 Node *dst_offset = argument(3); 5455 Node *length = argument(4); 5456 5457 const Type* src_type = src->Value(&_gvn); 5458 const Type* dst_type = dst->Value(&_gvn); 5459 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5460 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5461 if (top_src == NULL || top_src->klass() == NULL || 5462 top_dest == NULL || top_dest->klass() == NULL) { 5463 // failed array check 5464 return false; 5465 } 5466 5467 // Figure out the size and type of the elements we will be copying. 5468 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5469 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5470 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5471 return false; 5472 } 5473 Node* src_start = array_element_address(src, src_offset, src_elem); 5474 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5475 // 'src_start' points to src array + scaled offset 5476 // 'dst_start' points to dst array + scaled offset 5477 5478 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5479 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5480 enc = _gvn.transform(enc); 5481 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc)); 5482 set_memory(res_mem, mtype); 5483 set_result(enc); 5484 return true; 5485} 5486 5487/** 5488 * Calculate CRC32 for byte. 5489 * int java.util.zip.CRC32.update(int crc, int b) 5490 */ 5491bool LibraryCallKit::inline_updateCRC32() { 5492 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5493 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5494 // no receiver since it is static method 5495 Node* crc = argument(0); // type: int 5496 Node* b = argument(1); // type: int 5497 5498 /* 5499 * int c = ~ crc; 5500 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5501 * b = b ^ (c >>> 8); 5502 * crc = ~b; 5503 */ 5504 5505 Node* M1 = intcon(-1); 5506 crc = _gvn.transform(new (C) XorINode(crc, M1)); 5507 Node* result = _gvn.transform(new (C) XorINode(crc, b)); 5508 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF))); 5509 5510 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5511 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2))); 5512 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5513 result = make_load(control(), adr, TypeInt::INT, T_INT); 5514 5515 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8))); 5516 result = _gvn.transform(new (C) XorINode(crc, result)); 5517 result = _gvn.transform(new (C) XorINode(result, M1)); 5518 set_result(result); 5519 return true; 5520} 5521 5522/** 5523 * Calculate CRC32 for byte[] array. 5524 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5525 */ 5526bool LibraryCallKit::inline_updateBytesCRC32() { 5527 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5528 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5529 // no receiver since it is static method 5530 Node* crc = argument(0); // type: int 5531 Node* src = argument(1); // type: oop 5532 Node* offset = argument(2); // type: int 5533 Node* length = argument(3); // type: int 5534 5535 const Type* src_type = src->Value(&_gvn); 5536 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5537 if (top_src == NULL || top_src->klass() == NULL) { 5538 // failed array check 5539 return false; 5540 } 5541 5542 // Figure out the size and type of the elements we will be copying. 5543 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5544 if (src_elem != T_BYTE) { 5545 return false; 5546 } 5547 5548 // 'src_start' points to src array + scaled offset 5549 Node* src_start = array_element_address(src, offset, src_elem); 5550 5551 // We assume that range check is done by caller. 5552 // TODO: generate range check (offset+length < src.length) in debug VM. 5553 5554 // Call the stub. 5555 address stubAddr = StubRoutines::updateBytesCRC32(); 5556 const char *stubName = "updateBytesCRC32"; 5557 5558 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5559 stubAddr, stubName, TypePtr::BOTTOM, 5560 crc, src_start, length); 5561 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5562 set_result(result); 5563 return true; 5564} 5565 5566/** 5567 * Calculate CRC32 for ByteBuffer. 5568 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5569 */ 5570bool LibraryCallKit::inline_updateByteBufferCRC32() { 5571 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5572 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5573 // no receiver since it is static method 5574 Node* crc = argument(0); // type: int 5575 Node* src = argument(1); // type: long 5576 Node* offset = argument(3); // type: int 5577 Node* length = argument(4); // type: int 5578 5579 src = ConvL2X(src); // adjust Java long to machine word 5580 Node* base = _gvn.transform(new (C) CastX2PNode(src)); 5581 offset = ConvI2X(offset); 5582 5583 // 'src_start' points to src array + scaled offset 5584 Node* src_start = basic_plus_adr(top(), base, offset); 5585 5586 // Call the stub. 5587 address stubAddr = StubRoutines::updateBytesCRC32(); 5588 const char *stubName = "updateBytesCRC32"; 5589 5590 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5591 stubAddr, stubName, TypePtr::BOTTOM, 5592 crc, src_start, length); 5593 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5594 set_result(result); 5595 return true; 5596} 5597 5598//----------------------------inline_reference_get---------------------------- 5599// public T java.lang.ref.Reference.get(); 5600bool LibraryCallKit::inline_reference_get() { 5601 const int referent_offset = java_lang_ref_Reference::referent_offset; 5602 guarantee(referent_offset > 0, "should have already been set"); 5603 5604 // Get the argument: 5605 Node* reference_obj = null_check_receiver(); 5606 if (stopped()) return true; 5607 5608 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5609 5610 ciInstanceKlass* klass = env()->Object_klass(); 5611 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5612 5613 Node* no_ctrl = NULL; 5614 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT); 5615 5616 // Use the pre-barrier to record the value in the referent field 5617 pre_barrier(false /* do_load */, 5618 control(), 5619 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5620 result /* pre_val */, 5621 T_OBJECT); 5622 5623 // Add memory barrier to prevent commoning reads from this field 5624 // across safepoint since GC can change its value. 5625 insert_mem_bar(Op_MemBarCPUOrder); 5626 5627 set_result(result); 5628 return true; 5629} 5630 5631 5632Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5633 bool is_exact=true, bool is_static=false) { 5634 5635 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5636 assert(tinst != NULL, "obj is null"); 5637 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5638 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5639 5640 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5641 ciSymbol::make(fieldTypeString), 5642 is_static); 5643 if (field == NULL) return (Node *) NULL; 5644 assert (field != NULL, "undefined field"); 5645 5646 // Next code copied from Parse::do_get_xxx(): 5647 5648 // Compute address and memory type. 5649 int offset = field->offset_in_bytes(); 5650 bool is_vol = field->is_volatile(); 5651 ciType* field_klass = field->type(); 5652 assert(field_klass->is_loaded(), "should be loaded"); 5653 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5654 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5655 BasicType bt = field->layout_type(); 5656 5657 // Build the resultant type of the load 5658 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5659 5660 // Build the load. 5661 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol); 5662 return loadedField; 5663} 5664 5665 5666//------------------------------inline_aescrypt_Block----------------------- 5667bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5668 address stubAddr; 5669 const char *stubName; 5670 assert(UseAES, "need AES instruction support"); 5671 5672 switch(id) { 5673 case vmIntrinsics::_aescrypt_encryptBlock: 5674 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5675 stubName = "aescrypt_encryptBlock"; 5676 break; 5677 case vmIntrinsics::_aescrypt_decryptBlock: 5678 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5679 stubName = "aescrypt_decryptBlock"; 5680 break; 5681 } 5682 if (stubAddr == NULL) return false; 5683 5684 Node* aescrypt_object = argument(0); 5685 Node* src = argument(1); 5686 Node* src_offset = argument(2); 5687 Node* dest = argument(3); 5688 Node* dest_offset = argument(4); 5689 5690 // (1) src and dest are arrays. 5691 const Type* src_type = src->Value(&_gvn); 5692 const Type* dest_type = dest->Value(&_gvn); 5693 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5694 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5695 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5696 5697 // for the quick and dirty code we will skip all the checks. 5698 // we are just trying to get the call to be generated. 5699 Node* src_start = src; 5700 Node* dest_start = dest; 5701 if (src_offset != NULL || dest_offset != NULL) { 5702 assert(src_offset != NULL && dest_offset != NULL, ""); 5703 src_start = array_element_address(src, src_offset, T_BYTE); 5704 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5705 } 5706 5707 // now need to get the start of its expanded key array 5708 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5709 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5710 if (k_start == NULL) return false; 5711 5712 // Call the stub. 5713 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5714 stubAddr, stubName, TypePtr::BOTTOM, 5715 src_start, dest_start, k_start); 5716 5717 return true; 5718} 5719 5720//------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5721bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5722 address stubAddr; 5723 const char *stubName; 5724 5725 assert(UseAES, "need AES instruction support"); 5726 5727 switch(id) { 5728 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5729 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5730 stubName = "cipherBlockChaining_encryptAESCrypt"; 5731 break; 5732 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5733 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5734 stubName = "cipherBlockChaining_decryptAESCrypt"; 5735 break; 5736 } 5737 if (stubAddr == NULL) return false; 5738 5739 Node* cipherBlockChaining_object = argument(0); 5740 Node* src = argument(1); 5741 Node* src_offset = argument(2); 5742 Node* len = argument(3); 5743 Node* dest = argument(4); 5744 Node* dest_offset = argument(5); 5745 5746 // (1) src and dest are arrays. 5747 const Type* src_type = src->Value(&_gvn); 5748 const Type* dest_type = dest->Value(&_gvn); 5749 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5750 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5751 assert (top_src != NULL && top_src->klass() != NULL 5752 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5753 5754 // checks are the responsibility of the caller 5755 Node* src_start = src; 5756 Node* dest_start = dest; 5757 if (src_offset != NULL || dest_offset != NULL) { 5758 assert(src_offset != NULL && dest_offset != NULL, ""); 5759 src_start = array_element_address(src, src_offset, T_BYTE); 5760 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5761 } 5762 5763 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5764 // (because of the predicated logic executed earlier). 5765 // so we cast it here safely. 5766 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5767 5768 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5769 if (embeddedCipherObj == NULL) return false; 5770 5771 // cast it to what we know it will be at runtime 5772 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5773 assert(tinst != NULL, "CBC obj is null"); 5774 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5775 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5776 if (!klass_AESCrypt->is_loaded()) return false; 5777 5778 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5779 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5780 const TypeOopPtr* xtype = aklass->as_instance_type(); 5781 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype); 5782 aescrypt_object = _gvn.transform(aescrypt_object); 5783 5784 // we need to get the start of the aescrypt_object's expanded key array 5785 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5786 if (k_start == NULL) return false; 5787 5788 // similarly, get the start address of the r vector 5789 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5790 if (objRvec == NULL) return false; 5791 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5792 5793 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5794 make_runtime_call(RC_LEAF|RC_NO_FP, 5795 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5796 stubAddr, stubName, TypePtr::BOTTOM, 5797 src_start, dest_start, k_start, r_start, len); 5798 5799 // return is void so no result needs to be pushed 5800 5801 return true; 5802} 5803 5804//------------------------------get_key_start_from_aescrypt_object----------------------- 5805Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5806 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5807 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5808 if (objAESCryptKey == NULL) return (Node *) NULL; 5809 5810 // now have the array, need to get the start address of the K array 5811 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5812 return k_start; 5813} 5814 5815//----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5816// Return node representing slow path of predicate check. 5817// the pseudo code we want to emulate with this predicate is: 5818// for encryption: 5819// if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5820// for decryption: 5821// if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5822// note cipher==plain is more conservative than the original java code but that's OK 5823// 5824Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5825 // First, check receiver for NULL since it is virtual method. 5826 Node* objCBC = argument(0); 5827 objCBC = null_check(objCBC); 5828 5829 if (stopped()) return NULL; // Always NULL 5830 5831 // Load embeddedCipher field of CipherBlockChaining object. 5832 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5833 5834 // get AESCrypt klass for instanceOf check 5835 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5836 // will have same classloader as CipherBlockChaining object 5837 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5838 assert(tinst != NULL, "CBCobj is null"); 5839 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5840 5841 // we want to do an instanceof comparison against the AESCrypt class 5842 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5843 if (!klass_AESCrypt->is_loaded()) { 5844 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5845 Node* ctrl = control(); 5846 set_control(top()); // no regular fast path 5847 return ctrl; 5848 } 5849 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5850 5851 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5852 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1))); 5853 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); 5854 5855 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5856 5857 // for encryption, we are done 5858 if (!decrypting) 5859 return instof_false; // even if it is NULL 5860 5861 // for decryption, we need to add a further check to avoid 5862 // taking the intrinsic path when cipher and plain are the same 5863 // see the original java code for why. 5864 RegionNode* region = new(C) RegionNode(3); 5865 region->init_req(1, instof_false); 5866 Node* src = argument(1); 5867 Node* dest = argument(4); 5868 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest)); 5869 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq)); 5870 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5871 region->init_req(2, src_dest_conjoint); 5872 5873 record_for_igvn(region); 5874 return _gvn.transform(region); 5875} 5876