1 /* 2 * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "classfile/vmSymbols.hpp" 28 #include "compiler/compileBroker.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "oops/objArrayKlass.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/callGenerator.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/idealKit.hpp" 35 #include "opto/mathexactnode.hpp" 36 #include "opto/mulnode.hpp" 37 #include "opto/parse.hpp" 38 #include "opto/runtime.hpp" 39 #include "opto/subnode.hpp" 40 #include "prims/nativeLookup.hpp" 41 #include "runtime/sharedRuntime.hpp" 42 #include "trace/traceMacros.hpp" 43 44 class LibraryIntrinsic : public InlineCallGenerator { 45 // Extend the set of intrinsics known to the runtime: 46 public: 47 private: 48 bool _is_virtual; 49 bool _is_predicted; 50 bool _does_virtual_dispatch; 51 vmIntrinsics::ID _intrinsic_id; 52 53 public: 54 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id) 55 : InlineCallGenerator(m), 56 _is_virtual(is_virtual), 57 _is_predicted(is_predicted), 58 _does_virtual_dispatch(does_virtual_dispatch), 59 _intrinsic_id(id) 60 { 61 } 62 virtual bool is_intrinsic() const { return true; } 63 virtual bool is_virtual() const { return _is_virtual; } 64 virtual bool is_predicted() const { return _is_predicted; } 65 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 66 virtual JVMState* generate(JVMState* jvms, Parse* parent_parser); 67 virtual Node* generate_predicate(JVMState* jvms); 68 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 69 }; 70 71 72 // Local helper class for LibraryIntrinsic: 73 class LibraryCallKit : public GraphKit { 74 private: 75 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 76 Node* _result; // the result node, if any 77 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 78 79 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 80 81 public: 82 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 83 : GraphKit(jvms), 84 _intrinsic(intrinsic), 85 _result(NULL) 86 { 87 // Check if this is a root compile. In that case we don't have a caller. 88 if (!jvms->has_method()) { 89 _reexecute_sp = sp(); 90 } else { 91 // Find out how many arguments the interpreter needs when deoptimizing 92 // and save the stack pointer value so it can used by uncommon_trap. 93 // We find the argument count by looking at the declared signature. 94 bool ignored_will_link; 95 ciSignature* declared_signature = NULL; 96 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 97 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 98 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 99 } 100 } 101 102 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 103 104 ciMethod* caller() const { return jvms()->method(); } 105 int bci() const { return jvms()->bci(); } 106 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 107 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 108 ciMethod* callee() const { return _intrinsic->method(); } 109 110 bool try_to_inline(); 111 Node* try_to_predicate(); 112 113 void push_result() { 114 // Push the result onto the stack. 115 if (!stopped() && result() != NULL) { 116 BasicType bt = result()->bottom_type()->basic_type(); 117 push_node(bt, result()); 118 } 119 } 120 121 private: 122 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 123 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 124 } 125 126 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 127 void set_result(RegionNode* region, PhiNode* value); 128 Node* result() { return _result; } 129 130 virtual int reexecute_sp() { return _reexecute_sp; } 131 132 // Helper functions to inline natives 133 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 134 Node* generate_slow_guard(Node* test, RegionNode* region); 135 Node* generate_fair_guard(Node* test, RegionNode* region); 136 Node* generate_negative_guard(Node* index, RegionNode* region, 137 // resulting CastII of index: 138 Node* *pos_index = NULL); 139 Node* generate_nonpositive_guard(Node* index, bool never_negative, 140 // resulting CastII of index: 141 Node* *pos_index = NULL); 142 Node* generate_limit_guard(Node* offset, Node* subseq_length, 143 Node* array_length, 144 RegionNode* region); 145 Node* generate_current_thread(Node* &tls_output); 146 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 147 bool disjoint_bases, const char* &name, bool dest_uninitialized); 148 Node* load_mirror_from_klass(Node* klass); 149 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 150 RegionNode* region, int null_path, 151 int offset); 152 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 153 RegionNode* region, int null_path) { 154 int offset = java_lang_Class::klass_offset_in_bytes(); 155 return load_klass_from_mirror_common(mirror, never_see_null, 156 region, null_path, 157 offset); 158 } 159 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path) { 161 int offset = java_lang_Class::array_klass_offset_in_bytes(); 162 return load_klass_from_mirror_common(mirror, never_see_null, 163 region, null_path, 164 offset); 165 } 166 Node* generate_access_flags_guard(Node* kls, 167 int modifier_mask, int modifier_bits, 168 RegionNode* region); 169 Node* generate_interface_guard(Node* kls, RegionNode* region); 170 Node* generate_array_guard(Node* kls, RegionNode* region) { 171 return generate_array_guard_common(kls, region, false, false); 172 } 173 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 174 return generate_array_guard_common(kls, region, false, true); 175 } 176 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 177 return generate_array_guard_common(kls, region, true, false); 178 } 179 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 180 return generate_array_guard_common(kls, region, true, true); 181 } 182 Node* generate_array_guard_common(Node* kls, RegionNode* region, 183 bool obj_array, bool not_array); 184 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 185 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 186 bool is_virtual = false, bool is_static = false); 187 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 188 return generate_method_call(method_id, false, true); 189 } 190 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 191 return generate_method_call(method_id, true, false); 192 } 193 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 194 195 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 196 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 197 bool inline_string_compareTo(); 198 bool inline_string_indexOf(); 199 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 200 bool inline_string_equals(); 201 Node* round_double_node(Node* n); 202 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 203 bool inline_math_native(vmIntrinsics::ID id); 204 bool inline_trig(vmIntrinsics::ID id); 205 bool inline_math(vmIntrinsics::ID id); 206 template <typename OverflowOp> 207 bool inline_math_overflow(Node* arg1, Node* arg2); 208 void inline_math_mathExact(Node* math, Node* test); 209 bool inline_math_addExactI(bool is_increment); 210 bool inline_math_addExactL(bool is_increment); 211 bool inline_math_multiplyExactI(); 212 bool inline_math_multiplyExactL(); 213 bool inline_math_negateExactI(); 214 bool inline_math_negateExactL(); 215 bool inline_math_subtractExactI(bool is_decrement); 216 bool inline_math_subtractExactL(bool is_decrement); 217 bool inline_exp(); 218 bool inline_pow(); 219 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 220 bool inline_min_max(vmIntrinsics::ID id); 221 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 222 // This returns Type::AnyPtr, RawPtr, or OopPtr. 223 int classify_unsafe_addr(Node* &base, Node* &offset); 224 Node* make_unsafe_address(Node* base, Node* offset); 225 // Helper for inline_unsafe_access. 226 // Generates the guards that check whether the result of 227 // Unsafe.getObject should be recorded in an SATB log buffer. 228 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 229 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 230 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 231 static bool klass_needs_init_guard(Node* kls); 232 bool inline_unsafe_allocate(); 233 bool inline_unsafe_copyMemory(); 234 bool inline_native_currentThread(); 235 #ifdef TRACE_HAVE_INTRINSICS 236 bool inline_native_classID(); 237 bool inline_native_threadID(); 238 #endif 239 bool inline_native_time_funcs(address method, const char* funcName); 240 bool inline_native_isInterrupted(); 241 bool inline_native_Class_query(vmIntrinsics::ID id); 242 bool inline_native_subtype_check(); 243 244 bool inline_native_newArray(); 245 bool inline_native_getLength(); 246 bool inline_array_copyOf(bool is_copyOfRange); 247 bool inline_array_equals(); 248 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 249 bool inline_native_clone(bool is_virtual); 250 bool inline_native_Reflection_getCallerClass(); 251 // Helper function for inlining native object hash method 252 bool inline_native_hashcode(bool is_virtual, bool is_static); 253 bool inline_native_getClass(); 254 255 // Helper functions for inlining arraycopy 256 bool inline_arraycopy(); 257 void generate_arraycopy(const TypePtr* adr_type, 258 BasicType basic_elem_type, 259 Node* src, Node* src_offset, 260 Node* dest, Node* dest_offset, 261 Node* copy_length, 262 bool disjoint_bases = false, 263 bool length_never_negative = false, 264 RegionNode* slow_region = NULL); 265 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 266 RegionNode* slow_region); 267 void generate_clear_array(const TypePtr* adr_type, 268 Node* dest, 269 BasicType basic_elem_type, 270 Node* slice_off, 271 Node* slice_len, 272 Node* slice_end); 273 bool generate_block_arraycopy(const TypePtr* adr_type, 274 BasicType basic_elem_type, 275 AllocateNode* alloc, 276 Node* src, Node* src_offset, 277 Node* dest, Node* dest_offset, 278 Node* dest_size, bool dest_uninitialized); 279 void generate_slow_arraycopy(const TypePtr* adr_type, 280 Node* src, Node* src_offset, 281 Node* dest, Node* dest_offset, 282 Node* copy_length, bool dest_uninitialized); 283 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 284 Node* dest_elem_klass, 285 Node* src, Node* src_offset, 286 Node* dest, Node* dest_offset, 287 Node* copy_length, bool dest_uninitialized); 288 Node* generate_generic_arraycopy(const TypePtr* adr_type, 289 Node* src, Node* src_offset, 290 Node* dest, Node* dest_offset, 291 Node* copy_length, bool dest_uninitialized); 292 void generate_unchecked_arraycopy(const TypePtr* adr_type, 293 BasicType basic_elem_type, 294 bool disjoint_bases, 295 Node* src, Node* src_offset, 296 Node* dest, Node* dest_offset, 297 Node* copy_length, bool dest_uninitialized); 298 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 299 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 300 bool inline_unsafe_ordered_store(BasicType type); 301 bool inline_unsafe_fence(vmIntrinsics::ID id); 302 bool inline_fp_conversions(vmIntrinsics::ID id); 303 bool inline_number_methods(vmIntrinsics::ID id); 304 bool inline_reference_get(); 305 bool inline_aescrypt_Block(vmIntrinsics::ID id); 306 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 307 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 308 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 309 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 310 bool inline_encodeISOArray(); 311 bool inline_updateCRC32(); 312 bool inline_updateBytesCRC32(); 313 bool inline_updateByteBufferCRC32(); 314 }; 315 316 317 //---------------------------make_vm_intrinsic---------------------------- 318 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 319 vmIntrinsics::ID id = m->intrinsic_id(); 320 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 321 322 if (DisableIntrinsic[0] != '\0' 323 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 324 // disabled by a user request on the command line: 325 // example: -XX:DisableIntrinsic=_hashCode,_getClass 326 return NULL; 327 } 328 329 if (!m->is_loaded()) { 330 // do not attempt to inline unloaded methods 331 return NULL; 332 } 333 334 // Only a few intrinsics implement a virtual dispatch. 335 // They are expensive calls which are also frequently overridden. 336 if (is_virtual) { 337 switch (id) { 338 case vmIntrinsics::_hashCode: 339 case vmIntrinsics::_clone: 340 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 341 break; 342 default: 343 return NULL; 344 } 345 } 346 347 // -XX:-InlineNatives disables nearly all intrinsics: 348 if (!InlineNatives) { 349 switch (id) { 350 case vmIntrinsics::_indexOf: 351 case vmIntrinsics::_compareTo: 352 case vmIntrinsics::_equals: 353 case vmIntrinsics::_equalsC: 354 case vmIntrinsics::_getAndAddInt: 355 case vmIntrinsics::_getAndAddLong: 356 case vmIntrinsics::_getAndSetInt: 357 case vmIntrinsics::_getAndSetLong: 358 case vmIntrinsics::_getAndSetObject: 359 case vmIntrinsics::_loadFence: 360 case vmIntrinsics::_storeFence: 361 case vmIntrinsics::_fullFence: 362 break; // InlineNatives does not control String.compareTo 363 case vmIntrinsics::_Reference_get: 364 break; // InlineNatives does not control Reference.get 365 default: 366 return NULL; 367 } 368 } 369 370 bool is_predicted = false; 371 bool does_virtual_dispatch = false; 372 373 switch (id) { 374 case vmIntrinsics::_compareTo: 375 if (!SpecialStringCompareTo) return NULL; 376 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 377 break; 378 case vmIntrinsics::_indexOf: 379 if (!SpecialStringIndexOf) return NULL; 380 break; 381 case vmIntrinsics::_equals: 382 if (!SpecialStringEquals) return NULL; 383 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 384 break; 385 case vmIntrinsics::_equalsC: 386 if (!SpecialArraysEquals) return NULL; 387 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 388 break; 389 case vmIntrinsics::_arraycopy: 390 if (!InlineArrayCopy) return NULL; 391 break; 392 case vmIntrinsics::_copyMemory: 393 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 394 if (!InlineArrayCopy) return NULL; 395 break; 396 case vmIntrinsics::_hashCode: 397 if (!InlineObjectHash) return NULL; 398 does_virtual_dispatch = true; 399 break; 400 case vmIntrinsics::_clone: 401 does_virtual_dispatch = true; 402 case vmIntrinsics::_copyOf: 403 case vmIntrinsics::_copyOfRange: 404 if (!InlineObjectCopy) return NULL; 405 // These also use the arraycopy intrinsic mechanism: 406 if (!InlineArrayCopy) return NULL; 407 break; 408 case vmIntrinsics::_encodeISOArray: 409 if (!SpecialEncodeISOArray) return NULL; 410 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 411 break; 412 case vmIntrinsics::_checkIndex: 413 // We do not intrinsify this. The optimizer does fine with it. 414 return NULL; 415 416 case vmIntrinsics::_getCallerClass: 417 if (!UseNewReflection) return NULL; 418 if (!InlineReflectionGetCallerClass) return NULL; 419 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 420 break; 421 422 case vmIntrinsics::_bitCount_i: 423 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 424 break; 425 426 case vmIntrinsics::_bitCount_l: 427 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 428 break; 429 430 case vmIntrinsics::_numberOfLeadingZeros_i: 431 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 432 break; 433 434 case vmIntrinsics::_numberOfLeadingZeros_l: 435 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 436 break; 437 438 case vmIntrinsics::_numberOfTrailingZeros_i: 439 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 440 break; 441 442 case vmIntrinsics::_numberOfTrailingZeros_l: 443 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 444 break; 445 446 case vmIntrinsics::_reverseBytes_c: 447 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 448 break; 449 case vmIntrinsics::_reverseBytes_s: 450 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 451 break; 452 case vmIntrinsics::_reverseBytes_i: 453 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 454 break; 455 case vmIntrinsics::_reverseBytes_l: 456 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 457 break; 458 459 case vmIntrinsics::_Reference_get: 460 // Use the intrinsic version of Reference.get() so that the value in 461 // the referent field can be registered by the G1 pre-barrier code. 462 // Also add memory barrier to prevent commoning reads from this field 463 // across safepoint since GC can change it value. 464 break; 465 466 case vmIntrinsics::_compareAndSwapObject: 467 #ifdef _LP64 468 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 469 #endif 470 break; 471 472 case vmIntrinsics::_compareAndSwapLong: 473 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 474 break; 475 476 case vmIntrinsics::_getAndAddInt: 477 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 478 break; 479 480 case vmIntrinsics::_getAndAddLong: 481 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 482 break; 483 484 case vmIntrinsics::_getAndSetInt: 485 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 486 break; 487 488 case vmIntrinsics::_getAndSetLong: 489 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 490 break; 491 492 case vmIntrinsics::_getAndSetObject: 493 #ifdef _LP64 494 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 495 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 496 break; 497 #else 498 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 499 break; 500 #endif 501 502 case vmIntrinsics::_aescrypt_encryptBlock: 503 case vmIntrinsics::_aescrypt_decryptBlock: 504 if (!UseAESIntrinsics) return NULL; 505 break; 506 507 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 508 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 509 if (!UseAESIntrinsics) return NULL; 510 // these two require the predicated logic 511 is_predicted = true; 512 break; 513 514 case vmIntrinsics::_updateCRC32: 515 case vmIntrinsics::_updateBytesCRC32: 516 case vmIntrinsics::_updateByteBufferCRC32: 517 if (!UseCRC32Intrinsics) return NULL; 518 break; 519 520 case vmIntrinsics::_incrementExactI: 521 case vmIntrinsics::_addExactI: 522 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 523 break; 524 case vmIntrinsics::_incrementExactL: 525 case vmIntrinsics::_addExactL: 526 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 527 break; 528 case vmIntrinsics::_decrementExactI: 529 case vmIntrinsics::_subtractExactI: 530 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 531 break; 532 case vmIntrinsics::_decrementExactL: 533 case vmIntrinsics::_subtractExactL: 534 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 535 break; 536 case vmIntrinsics::_negateExactI: 537 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 538 break; 539 case vmIntrinsics::_negateExactL: 540 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 541 break; 542 case vmIntrinsics::_multiplyExactI: 543 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 544 break; 545 case vmIntrinsics::_multiplyExactL: 546 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 547 break; 548 549 default: 550 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 551 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 552 break; 553 } 554 555 // -XX:-InlineClassNatives disables natives from the Class class. 556 // The flag applies to all reflective calls, notably Array.newArray 557 // (visible to Java programmers as Array.newInstance). 558 if (m->holder()->name() == ciSymbol::java_lang_Class() || 559 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 560 if (!InlineClassNatives) return NULL; 561 } 562 563 // -XX:-InlineThreadNatives disables natives from the Thread class. 564 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 565 if (!InlineThreadNatives) return NULL; 566 } 567 568 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 569 if (m->holder()->name() == ciSymbol::java_lang_Math() || 570 m->holder()->name() == ciSymbol::java_lang_Float() || 571 m->holder()->name() == ciSymbol::java_lang_Double()) { 572 if (!InlineMathNatives) return NULL; 573 } 574 575 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 576 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 577 if (!InlineUnsafeOps) return NULL; 578 } 579 580 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id); 581 } 582 583 //----------------------register_library_intrinsics----------------------- 584 // Initialize this file's data structures, for each Compile instance. 585 void Compile::register_library_intrinsics() { 586 // Nothing to do here. 587 } 588 589 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) { 590 LibraryCallKit kit(jvms, this); 591 Compile* C = kit.C; 592 int nodes = C->unique(); 593 #ifndef PRODUCT 594 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 595 char buf[1000]; 596 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 597 tty->print_cr("Intrinsic %s", str); 598 } 599 #endif 600 ciMethod* callee = kit.callee(); 601 const int bci = kit.bci(); 602 603 // Try to inline the intrinsic. 604 if (kit.try_to_inline()) { 605 if (C->print_intrinsics() || C->print_inlining()) { 606 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 607 } 608 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 609 if (C->log()) { 610 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 611 vmIntrinsics::name_at(intrinsic_id()), 612 (is_virtual() ? " virtual='1'" : ""), 613 C->unique() - nodes); 614 } 615 // Push the result from the inlined method onto the stack. 616 kit.push_result(); 617 return kit.transfer_exceptions_into_jvms(); 618 } 619 620 // The intrinsic bailed out 621 if (C->print_intrinsics() || C->print_inlining()) { 622 if (jvms->has_method()) { 623 // Not a root compile. 624 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 625 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 626 } else { 627 // Root compile 628 tty->print("Did not generate intrinsic %s%s at bci:%d in", 629 vmIntrinsics::name_at(intrinsic_id()), 630 (is_virtual() ? " (virtual)" : ""), bci); 631 } 632 } 633 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 634 return NULL; 635 } 636 637 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) { 638 LibraryCallKit kit(jvms, this); 639 Compile* C = kit.C; 640 int nodes = C->unique(); 641 #ifndef PRODUCT 642 assert(is_predicted(), "sanity"); 643 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 644 char buf[1000]; 645 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 646 tty->print_cr("Predicate for intrinsic %s", str); 647 } 648 #endif 649 ciMethod* callee = kit.callee(); 650 const int bci = kit.bci(); 651 652 Node* slow_ctl = kit.try_to_predicate(); 653 if (!kit.failing()) { 654 if (C->print_intrinsics() || C->print_inlining()) { 655 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 656 } 657 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 658 if (C->log()) { 659 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 660 vmIntrinsics::name_at(intrinsic_id()), 661 (is_virtual() ? " virtual='1'" : ""), 662 C->unique() - nodes); 663 } 664 return slow_ctl; // Could be NULL if the check folds. 665 } 666 667 // The intrinsic bailed out 668 if (C->print_intrinsics() || C->print_inlining()) { 669 if (jvms->has_method()) { 670 // Not a root compile. 671 const char* msg = "failed to generate predicate for intrinsic"; 672 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 673 } else { 674 // Root compile 675 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 676 vmIntrinsics::name_at(intrinsic_id()), 677 (is_virtual() ? " (virtual)" : ""), bci); 678 } 679 } 680 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 681 return NULL; 682 } 683 684 bool LibraryCallKit::try_to_inline() { 685 // Handle symbolic names for otherwise undistinguished boolean switches: 686 const bool is_store = true; 687 const bool is_native_ptr = true; 688 const bool is_static = true; 689 const bool is_volatile = true; 690 691 if (!jvms()->has_method()) { 692 // Root JVMState has a null method. 693 assert(map()->memory()->Opcode() == Op_Parm, ""); 694 // Insert the memory aliasing node 695 set_all_memory(reset_memory()); 696 } 697 assert(merged_memory(), ""); 698 699 700 switch (intrinsic_id()) { 701 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 702 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 703 case vmIntrinsics::_getClass: return inline_native_getClass(); 704 705 case vmIntrinsics::_dsin: 706 case vmIntrinsics::_dcos: 707 case vmIntrinsics::_dtan: 708 case vmIntrinsics::_dabs: 709 case vmIntrinsics::_datan2: 710 case vmIntrinsics::_dsqrt: 711 case vmIntrinsics::_dexp: 712 case vmIntrinsics::_dlog: 713 case vmIntrinsics::_dlog10: 714 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 715 716 case vmIntrinsics::_min: 717 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 718 719 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 720 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 721 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 722 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 723 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 724 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 725 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 726 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 727 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 728 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 729 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 730 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 731 732 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 733 734 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 735 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 736 case vmIntrinsics::_equals: return inline_string_equals(); 737 738 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 739 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 740 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 741 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 742 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 743 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 744 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 745 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 746 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 747 748 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 749 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 750 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 751 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 752 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 753 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 754 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 755 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 756 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 757 758 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 759 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 760 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 761 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 762 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 763 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 764 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 765 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 766 767 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 768 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 769 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 770 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 771 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 772 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 773 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 774 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 775 776 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 777 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 778 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 779 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 780 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 781 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 782 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 783 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 784 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 785 786 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 787 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 788 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 789 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 790 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 791 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 792 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 793 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 794 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 795 796 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 797 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 798 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 799 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 800 801 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 802 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 803 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 804 805 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 806 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 807 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 808 809 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 810 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 811 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 812 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 813 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 814 815 case vmIntrinsics::_loadFence: 816 case vmIntrinsics::_storeFence: 817 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 818 819 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 820 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 821 822 #ifdef TRACE_HAVE_INTRINSICS 823 case vmIntrinsics::_classID: return inline_native_classID(); 824 case vmIntrinsics::_threadID: return inline_native_threadID(); 825 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 826 #endif 827 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 828 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 829 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 830 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 831 case vmIntrinsics::_newArray: return inline_native_newArray(); 832 case vmIntrinsics::_getLength: return inline_native_getLength(); 833 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 834 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 835 case vmIntrinsics::_equalsC: return inline_array_equals(); 836 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 837 838 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 839 840 case vmIntrinsics::_isInstance: 841 case vmIntrinsics::_getModifiers: 842 case vmIntrinsics::_isInterface: 843 case vmIntrinsics::_isArray: 844 case vmIntrinsics::_isPrimitive: 845 case vmIntrinsics::_getSuperclass: 846 case vmIntrinsics::_getComponentType: 847 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 848 849 case vmIntrinsics::_floatToRawIntBits: 850 case vmIntrinsics::_floatToIntBits: 851 case vmIntrinsics::_intBitsToFloat: 852 case vmIntrinsics::_doubleToRawLongBits: 853 case vmIntrinsics::_doubleToLongBits: 854 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 855 856 case vmIntrinsics::_numberOfLeadingZeros_i: 857 case vmIntrinsics::_numberOfLeadingZeros_l: 858 case vmIntrinsics::_numberOfTrailingZeros_i: 859 case vmIntrinsics::_numberOfTrailingZeros_l: 860 case vmIntrinsics::_bitCount_i: 861 case vmIntrinsics::_bitCount_l: 862 case vmIntrinsics::_reverseBytes_i: 863 case vmIntrinsics::_reverseBytes_l: 864 case vmIntrinsics::_reverseBytes_s: 865 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 866 867 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 868 869 case vmIntrinsics::_Reference_get: return inline_reference_get(); 870 871 case vmIntrinsics::_aescrypt_encryptBlock: 872 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 873 874 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 875 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 876 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 877 878 case vmIntrinsics::_encodeISOArray: 879 return inline_encodeISOArray(); 880 881 case vmIntrinsics::_updateCRC32: 882 return inline_updateCRC32(); 883 case vmIntrinsics::_updateBytesCRC32: 884 return inline_updateBytesCRC32(); 885 case vmIntrinsics::_updateByteBufferCRC32: 886 return inline_updateByteBufferCRC32(); 887 888 default: 889 // If you get here, it may be that someone has added a new intrinsic 890 // to the list in vmSymbols.hpp without implementing it here. 891 #ifndef PRODUCT 892 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 893 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 894 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 895 } 896 #endif 897 return false; 898 } 899 } 900 901 Node* LibraryCallKit::try_to_predicate() { 902 if (!jvms()->has_method()) { 903 // Root JVMState has a null method. 904 assert(map()->memory()->Opcode() == Op_Parm, ""); 905 // Insert the memory aliasing node 906 set_all_memory(reset_memory()); 907 } 908 assert(merged_memory(), ""); 909 910 switch (intrinsic_id()) { 911 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 912 return inline_cipherBlockChaining_AESCrypt_predicate(false); 913 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 914 return inline_cipherBlockChaining_AESCrypt_predicate(true); 915 916 default: 917 // If you get here, it may be that someone has added a new intrinsic 918 // to the list in vmSymbols.hpp without implementing it here. 919 #ifndef PRODUCT 920 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 921 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 922 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 923 } 924 #endif 925 Node* slow_ctl = control(); 926 set_control(top()); // No fast path instrinsic 927 return slow_ctl; 928 } 929 } 930 931 //------------------------------set_result------------------------------- 932 // Helper function for finishing intrinsics. 933 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 934 record_for_igvn(region); 935 set_control(_gvn.transform(region)); 936 set_result( _gvn.transform(value)); 937 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 938 } 939 940 //------------------------------generate_guard--------------------------- 941 // Helper function for generating guarded fast-slow graph structures. 942 // The given 'test', if true, guards a slow path. If the test fails 943 // then a fast path can be taken. (We generally hope it fails.) 944 // In all cases, GraphKit::control() is updated to the fast path. 945 // The returned value represents the control for the slow path. 946 // The return value is never 'top'; it is either a valid control 947 // or NULL if it is obvious that the slow path can never be taken. 948 // Also, if region and the slow control are not NULL, the slow edge 949 // is appended to the region. 950 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 951 if (stopped()) { 952 // Already short circuited. 953 return NULL; 954 } 955 956 // Build an if node and its projections. 957 // If test is true we take the slow path, which we assume is uncommon. 958 if (_gvn.type(test) == TypeInt::ZERO) { 959 // The slow branch is never taken. No need to build this guard. 960 return NULL; 961 } 962 963 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 964 965 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff)); 966 if (if_slow == top()) { 967 // The slow branch is never taken. No need to build this guard. 968 return NULL; 969 } 970 971 if (region != NULL) 972 region->add_req(if_slow); 973 974 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff)); 975 set_control(if_fast); 976 977 return if_slow; 978 } 979 980 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 981 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 982 } 983 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 984 return generate_guard(test, region, PROB_FAIR); 985 } 986 987 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 988 Node* *pos_index) { 989 if (stopped()) 990 return NULL; // already stopped 991 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 992 return NULL; // index is already adequately typed 993 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0))); 994 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 995 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 996 if (is_neg != NULL && pos_index != NULL) { 997 // Emulate effect of Parse::adjust_map_after_if. 998 Node* ccast = new (C) CastIINode(index, TypeInt::POS); 999 ccast->set_req(0, control()); 1000 (*pos_index) = _gvn.transform(ccast); 1001 } 1002 return is_neg; 1003 } 1004 1005 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 1006 Node* *pos_index) { 1007 if (stopped()) 1008 return NULL; // already stopped 1009 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 1010 return NULL; // index is already adequately typed 1011 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0))); 1012 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 1013 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq)); 1014 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 1015 if (is_notp != NULL && pos_index != NULL) { 1016 // Emulate effect of Parse::adjust_map_after_if. 1017 Node* ccast = new (C) CastIINode(index, TypeInt::POS1); 1018 ccast->set_req(0, control()); 1019 (*pos_index) = _gvn.transform(ccast); 1020 } 1021 return is_notp; 1022 } 1023 1024 // Make sure that 'position' is a valid limit index, in [0..length]. 1025 // There are two equivalent plans for checking this: 1026 // A. (offset + copyLength) unsigned<= arrayLength 1027 // B. offset <= (arrayLength - copyLength) 1028 // We require that all of the values above, except for the sum and 1029 // difference, are already known to be non-negative. 1030 // Plan A is robust in the face of overflow, if offset and copyLength 1031 // are both hugely positive. 1032 // 1033 // Plan B is less direct and intuitive, but it does not overflow at 1034 // all, since the difference of two non-negatives is always 1035 // representable. Whenever Java methods must perform the equivalent 1036 // check they generally use Plan B instead of Plan A. 1037 // For the moment we use Plan A. 1038 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1039 Node* subseq_length, 1040 Node* array_length, 1041 RegionNode* region) { 1042 if (stopped()) 1043 return NULL; // already stopped 1044 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1045 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1046 return NULL; // common case of whole-array copy 1047 Node* last = subseq_length; 1048 if (!zero_offset) // last += offset 1049 last = _gvn.transform(new (C) AddINode(last, offset)); 1050 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last)); 1051 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 1052 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1053 return is_over; 1054 } 1055 1056 1057 //--------------------------generate_current_thread-------------------- 1058 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1059 ciKlass* thread_klass = env()->Thread_klass(); 1060 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1061 Node* thread = _gvn.transform(new (C) ThreadLocalNode()); 1062 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1063 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1064 tls_output = thread; 1065 return threadObj; 1066 } 1067 1068 1069 //------------------------------make_string_method_node------------------------ 1070 // Helper method for String intrinsic functions. This version is called 1071 // with str1 and str2 pointing to String object nodes. 1072 // 1073 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1074 Node* no_ctrl = NULL; 1075 1076 // Get start addr of string 1077 Node* str1_value = load_String_value(no_ctrl, str1); 1078 Node* str1_offset = load_String_offset(no_ctrl, str1); 1079 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1080 1081 // Get length of string 1 1082 Node* str1_len = load_String_length(no_ctrl, str1); 1083 1084 Node* str2_value = load_String_value(no_ctrl, str2); 1085 Node* str2_offset = load_String_offset(no_ctrl, str2); 1086 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1087 1088 Node* str2_len = NULL; 1089 Node* result = NULL; 1090 1091 switch (opcode) { 1092 case Op_StrIndexOf: 1093 // Get length of string 2 1094 str2_len = load_String_length(no_ctrl, str2); 1095 1096 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1097 str1_start, str1_len, str2_start, str2_len); 1098 break; 1099 case Op_StrComp: 1100 // Get length of string 2 1101 str2_len = load_String_length(no_ctrl, str2); 1102 1103 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1104 str1_start, str1_len, str2_start, str2_len); 1105 break; 1106 case Op_StrEquals: 1107 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1108 str1_start, str2_start, str1_len); 1109 break; 1110 default: 1111 ShouldNotReachHere(); 1112 return NULL; 1113 } 1114 1115 // All these intrinsics have checks. 1116 C->set_has_split_ifs(true); // Has chance for split-if optimization 1117 1118 return _gvn.transform(result); 1119 } 1120 1121 // Helper method for String intrinsic functions. This version is called 1122 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1123 // to Int nodes containing the lenghts of str1 and str2. 1124 // 1125 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1126 Node* result = NULL; 1127 switch (opcode) { 1128 case Op_StrIndexOf: 1129 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1130 str1_start, cnt1, str2_start, cnt2); 1131 break; 1132 case Op_StrComp: 1133 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1134 str1_start, cnt1, str2_start, cnt2); 1135 break; 1136 case Op_StrEquals: 1137 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1138 str1_start, str2_start, cnt1); 1139 break; 1140 default: 1141 ShouldNotReachHere(); 1142 return NULL; 1143 } 1144 1145 // All these intrinsics have checks. 1146 C->set_has_split_ifs(true); // Has chance for split-if optimization 1147 1148 return _gvn.transform(result); 1149 } 1150 1151 //------------------------------inline_string_compareTo------------------------ 1152 // public int java.lang.String.compareTo(String anotherString); 1153 bool LibraryCallKit::inline_string_compareTo() { 1154 Node* receiver = null_check(argument(0)); 1155 Node* arg = null_check(argument(1)); 1156 if (stopped()) { 1157 return true; 1158 } 1159 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1160 return true; 1161 } 1162 1163 //------------------------------inline_string_equals------------------------ 1164 bool LibraryCallKit::inline_string_equals() { 1165 Node* receiver = null_check_receiver(); 1166 // NOTE: Do not null check argument for String.equals() because spec 1167 // allows to specify NULL as argument. 1168 Node* argument = this->argument(1); 1169 if (stopped()) { 1170 return true; 1171 } 1172 1173 // paths (plus control) merge 1174 RegionNode* region = new (C) RegionNode(5); 1175 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 1176 1177 // does source == target string? 1178 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument)); 1179 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 1180 1181 Node* if_eq = generate_slow_guard(bol, NULL); 1182 if (if_eq != NULL) { 1183 // receiver == argument 1184 phi->init_req(2, intcon(1)); 1185 region->init_req(2, if_eq); 1186 } 1187 1188 // get String klass for instanceOf 1189 ciInstanceKlass* klass = env()->String_klass(); 1190 1191 if (!stopped()) { 1192 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1193 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1))); 1194 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 1195 1196 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1197 //instanceOf == true, fallthrough 1198 1199 if (inst_false != NULL) { 1200 phi->init_req(3, intcon(0)); 1201 region->init_req(3, inst_false); 1202 } 1203 } 1204 1205 if (!stopped()) { 1206 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1207 1208 // Properly cast the argument to String 1209 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type)); 1210 // This path is taken only when argument's type is String:NotNull. 1211 argument = cast_not_null(argument, false); 1212 1213 Node* no_ctrl = NULL; 1214 1215 // Get start addr of receiver 1216 Node* receiver_val = load_String_value(no_ctrl, receiver); 1217 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1218 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1219 1220 // Get length of receiver 1221 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1222 1223 // Get start addr of argument 1224 Node* argument_val = load_String_value(no_ctrl, argument); 1225 Node* argument_offset = load_String_offset(no_ctrl, argument); 1226 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1227 1228 // Get length of argument 1229 Node* argument_cnt = load_String_length(no_ctrl, argument); 1230 1231 // Check for receiver count != argument count 1232 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt)); 1233 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne)); 1234 Node* if_ne = generate_slow_guard(bol, NULL); 1235 if (if_ne != NULL) { 1236 phi->init_req(4, intcon(0)); 1237 region->init_req(4, if_ne); 1238 } 1239 1240 // Check for count == 0 is done by assembler code for StrEquals. 1241 1242 if (!stopped()) { 1243 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1244 phi->init_req(1, equals); 1245 region->init_req(1, control()); 1246 } 1247 } 1248 1249 // post merge 1250 set_control(_gvn.transform(region)); 1251 record_for_igvn(region); 1252 1253 set_result(_gvn.transform(phi)); 1254 return true; 1255 } 1256 1257 //------------------------------inline_array_equals---------------------------- 1258 bool LibraryCallKit::inline_array_equals() { 1259 Node* arg1 = argument(0); 1260 Node* arg2 = argument(1); 1261 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1262 return true; 1263 } 1264 1265 // Java version of String.indexOf(constant string) 1266 // class StringDecl { 1267 // StringDecl(char[] ca) { 1268 // offset = 0; 1269 // count = ca.length; 1270 // value = ca; 1271 // } 1272 // int offset; 1273 // int count; 1274 // char[] value; 1275 // } 1276 // 1277 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1278 // int targetOffset, int cache_i, int md2) { 1279 // int cache = cache_i; 1280 // int sourceOffset = string_object.offset; 1281 // int sourceCount = string_object.count; 1282 // int targetCount = target_object.length; 1283 // 1284 // int targetCountLess1 = targetCount - 1; 1285 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1286 // 1287 // char[] source = string_object.value; 1288 // char[] target = target_object; 1289 // int lastChar = target[targetCountLess1]; 1290 // 1291 // outer_loop: 1292 // for (int i = sourceOffset; i < sourceEnd; ) { 1293 // int src = source[i + targetCountLess1]; 1294 // if (src == lastChar) { 1295 // // With random strings and a 4-character alphabet, 1296 // // reverse matching at this point sets up 0.8% fewer 1297 // // frames, but (paradoxically) makes 0.3% more probes. 1298 // // Since those probes are nearer the lastChar probe, 1299 // // there is may be a net D$ win with reverse matching. 1300 // // But, reversing loop inhibits unroll of inner loop 1301 // // for unknown reason. So, does running outer loop from 1302 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1303 // for (int j = 0; j < targetCountLess1; j++) { 1304 // if (target[targetOffset + j] != source[i+j]) { 1305 // if ((cache & (1 << source[i+j])) == 0) { 1306 // if (md2 < j+1) { 1307 // i += j+1; 1308 // continue outer_loop; 1309 // } 1310 // } 1311 // i += md2; 1312 // continue outer_loop; 1313 // } 1314 // } 1315 // return i - sourceOffset; 1316 // } 1317 // if ((cache & (1 << src)) == 0) { 1318 // i += targetCountLess1; 1319 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1320 // i++; 1321 // } 1322 // return -1; 1323 // } 1324 1325 //------------------------------string_indexOf------------------------ 1326 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1327 jint cache_i, jint md2_i) { 1328 1329 Node* no_ctrl = NULL; 1330 float likely = PROB_LIKELY(0.9); 1331 float unlikely = PROB_UNLIKELY(0.9); 1332 1333 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1334 1335 Node* source = load_String_value(no_ctrl, string_object); 1336 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1337 Node* sourceCount = load_String_length(no_ctrl, string_object); 1338 1339 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1340 jint target_length = target_array->length(); 1341 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1342 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1343 1344 // String.value field is known to be @Stable. 1345 if (UseImplicitStableValues) { 1346 target = cast_array_to_stable(target, target_type); 1347 } 1348 1349 IdealKit kit(this, false, true); 1350 #define __ kit. 1351 Node* zero = __ ConI(0); 1352 Node* one = __ ConI(1); 1353 Node* cache = __ ConI(cache_i); 1354 Node* md2 = __ ConI(md2_i); 1355 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1356 Node* targetCount = __ ConI(target_length); 1357 Node* targetCountLess1 = __ ConI(target_length - 1); 1358 Node* targetOffset = __ ConI(targetOffset_i); 1359 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1360 1361 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1362 Node* outer_loop = __ make_label(2 /* goto */); 1363 Node* return_ = __ make_label(1); 1364 1365 __ set(rtn,__ ConI(-1)); 1366 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1367 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1368 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1369 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1370 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1371 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1372 Node* tpj = __ AddI(targetOffset, __ value(j)); 1373 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1374 Node* ipj = __ AddI(__ value(i), __ value(j)); 1375 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1376 __ if_then(targ, BoolTest::ne, src2); { 1377 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1378 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1379 __ increment(i, __ AddI(__ value(j), one)); 1380 __ goto_(outer_loop); 1381 } __ end_if(); __ dead(j); 1382 }__ end_if(); __ dead(j); 1383 __ increment(i, md2); 1384 __ goto_(outer_loop); 1385 }__ end_if(); 1386 __ increment(j, one); 1387 }__ end_loop(); __ dead(j); 1388 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1389 __ goto_(return_); 1390 }__ end_if(); 1391 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1392 __ increment(i, targetCountLess1); 1393 }__ end_if(); 1394 __ increment(i, one); 1395 __ bind(outer_loop); 1396 }__ end_loop(); __ dead(i); 1397 __ bind(return_); 1398 1399 // Final sync IdealKit and GraphKit. 1400 final_sync(kit); 1401 Node* result = __ value(rtn); 1402 #undef __ 1403 C->set_has_loops(true); 1404 return result; 1405 } 1406 1407 //------------------------------inline_string_indexOf------------------------ 1408 bool LibraryCallKit::inline_string_indexOf() { 1409 Node* receiver = argument(0); 1410 Node* arg = argument(1); 1411 1412 Node* result; 1413 // Disable the use of pcmpestri until it can be guaranteed that 1414 // the load doesn't cross into the uncommited space. 1415 if (Matcher::has_match_rule(Op_StrIndexOf) && 1416 UseSSE42Intrinsics) { 1417 // Generate SSE4.2 version of indexOf 1418 // We currently only have match rules that use SSE4.2 1419 1420 receiver = null_check(receiver); 1421 arg = null_check(arg); 1422 if (stopped()) { 1423 return true; 1424 } 1425 1426 ciInstanceKlass* str_klass = env()->String_klass(); 1427 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1428 1429 // Make the merge point 1430 RegionNode* result_rgn = new (C) RegionNode(4); 1431 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT); 1432 Node* no_ctrl = NULL; 1433 1434 // Get start addr of source string 1435 Node* source = load_String_value(no_ctrl, receiver); 1436 Node* source_offset = load_String_offset(no_ctrl, receiver); 1437 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1438 1439 // Get length of source string 1440 Node* source_cnt = load_String_length(no_ctrl, receiver); 1441 1442 // Get start addr of substring 1443 Node* substr = load_String_value(no_ctrl, arg); 1444 Node* substr_offset = load_String_offset(no_ctrl, arg); 1445 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1446 1447 // Get length of source string 1448 Node* substr_cnt = load_String_length(no_ctrl, arg); 1449 1450 // Check for substr count > string count 1451 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt)); 1452 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt)); 1453 Node* if_gt = generate_slow_guard(bol, NULL); 1454 if (if_gt != NULL) { 1455 result_phi->init_req(2, intcon(-1)); 1456 result_rgn->init_req(2, if_gt); 1457 } 1458 1459 if (!stopped()) { 1460 // Check for substr count == 0 1461 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0))); 1462 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 1463 Node* if_zero = generate_slow_guard(bol, NULL); 1464 if (if_zero != NULL) { 1465 result_phi->init_req(3, intcon(0)); 1466 result_rgn->init_req(3, if_zero); 1467 } 1468 } 1469 1470 if (!stopped()) { 1471 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1472 result_phi->init_req(1, result); 1473 result_rgn->init_req(1, control()); 1474 } 1475 set_control(_gvn.transform(result_rgn)); 1476 record_for_igvn(result_rgn); 1477 result = _gvn.transform(result_phi); 1478 1479 } else { // Use LibraryCallKit::string_indexOf 1480 // don't intrinsify if argument isn't a constant string. 1481 if (!arg->is_Con()) { 1482 return false; 1483 } 1484 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1485 if (str_type == NULL) { 1486 return false; 1487 } 1488 ciInstanceKlass* klass = env()->String_klass(); 1489 ciObject* str_const = str_type->const_oop(); 1490 if (str_const == NULL || str_const->klass() != klass) { 1491 return false; 1492 } 1493 ciInstance* str = str_const->as_instance(); 1494 assert(str != NULL, "must be instance"); 1495 1496 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1497 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1498 1499 int o; 1500 int c; 1501 if (java_lang_String::has_offset_field()) { 1502 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1503 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1504 } else { 1505 o = 0; 1506 c = pat->length(); 1507 } 1508 1509 // constant strings have no offset and count == length which 1510 // simplifies the resulting code somewhat so lets optimize for that. 1511 if (o != 0 || c != pat->length()) { 1512 return false; 1513 } 1514 1515 receiver = null_check(receiver, T_OBJECT); 1516 // NOTE: No null check on the argument is needed since it's a constant String oop. 1517 if (stopped()) { 1518 return true; 1519 } 1520 1521 // The null string as a pattern always returns 0 (match at beginning of string) 1522 if (c == 0) { 1523 set_result(intcon(0)); 1524 return true; 1525 } 1526 1527 // Generate default indexOf 1528 jchar lastChar = pat->char_at(o + (c - 1)); 1529 int cache = 0; 1530 int i; 1531 for (i = 0; i < c - 1; i++) { 1532 assert(i < pat->length(), "out of range"); 1533 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1534 } 1535 1536 int md2 = c; 1537 for (i = 0; i < c - 1; i++) { 1538 assert(i < pat->length(), "out of range"); 1539 if (pat->char_at(o + i) == lastChar) { 1540 md2 = (c - 1) - i; 1541 } 1542 } 1543 1544 result = string_indexOf(receiver, pat, o, cache, md2); 1545 } 1546 set_result(result); 1547 return true; 1548 } 1549 1550 //--------------------------round_double_node-------------------------------- 1551 // Round a double node if necessary. 1552 Node* LibraryCallKit::round_double_node(Node* n) { 1553 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1554 n = _gvn.transform(new (C) RoundDoubleNode(0, n)); 1555 return n; 1556 } 1557 1558 //------------------------------inline_math----------------------------------- 1559 // public static double Math.abs(double) 1560 // public static double Math.sqrt(double) 1561 // public static double Math.log(double) 1562 // public static double Math.log10(double) 1563 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1564 Node* arg = round_double_node(argument(0)); 1565 Node* n; 1566 switch (id) { 1567 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break; 1568 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break; 1569 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break; 1570 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break; 1571 default: fatal_unexpected_iid(id); break; 1572 } 1573 set_result(_gvn.transform(n)); 1574 return true; 1575 } 1576 1577 //------------------------------inline_trig---------------------------------- 1578 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1579 // argument reduction which will turn into a fast/slow diamond. 1580 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1581 Node* arg = round_double_node(argument(0)); 1582 Node* n = NULL; 1583 1584 switch (id) { 1585 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break; 1586 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break; 1587 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break; 1588 default: fatal_unexpected_iid(id); break; 1589 } 1590 n = _gvn.transform(n); 1591 1592 // Rounding required? Check for argument reduction! 1593 if (Matcher::strict_fp_requires_explicit_rounding) { 1594 static const double pi_4 = 0.7853981633974483; 1595 static const double neg_pi_4 = -0.7853981633974483; 1596 // pi/2 in 80-bit extended precision 1597 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1598 // -pi/2 in 80-bit extended precision 1599 // 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}; 1600 // Cutoff value for using this argument reduction technique 1601 //static const double pi_2_minus_epsilon = 1.564660403643354; 1602 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1603 1604 // Pseudocode for sin: 1605 // if (x <= Math.PI / 4.0) { 1606 // if (x >= -Math.PI / 4.0) return fsin(x); 1607 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1608 // } else { 1609 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1610 // } 1611 // return StrictMath.sin(x); 1612 1613 // Pseudocode for cos: 1614 // if (x <= Math.PI / 4.0) { 1615 // if (x >= -Math.PI / 4.0) return fcos(x); 1616 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1617 // } else { 1618 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1619 // } 1620 // return StrictMath.cos(x); 1621 1622 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1623 // requires a special machine instruction to load it. Instead we'll try 1624 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1625 // probably do the math inside the SIN encoding. 1626 1627 // Make the merge point 1628 RegionNode* r = new (C) RegionNode(3); 1629 Node* phi = new (C) PhiNode(r, Type::DOUBLE); 1630 1631 // Flatten arg so we need only 1 test 1632 Node *abs = _gvn.transform(new (C) AbsDNode(arg)); 1633 // Node for PI/4 constant 1634 Node *pi4 = makecon(TypeD::make(pi_4)); 1635 // Check PI/4 : abs(arg) 1636 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs)); 1637 // Check: If PI/4 < abs(arg) then go slow 1638 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt )); 1639 // Branch either way 1640 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1641 set_control(opt_iff(r,iff)); 1642 1643 // Set fast path result 1644 phi->init_req(2, n); 1645 1646 // Slow path - non-blocking leaf call 1647 Node* call = NULL; 1648 switch (id) { 1649 case vmIntrinsics::_dsin: 1650 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1651 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1652 "Sin", NULL, arg, top()); 1653 break; 1654 case vmIntrinsics::_dcos: 1655 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1656 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1657 "Cos", NULL, arg, top()); 1658 break; 1659 case vmIntrinsics::_dtan: 1660 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1661 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1662 "Tan", NULL, arg, top()); 1663 break; 1664 } 1665 assert(control()->in(0) == call, ""); 1666 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1667 r->init_req(1, control()); 1668 phi->init_req(1, slow_result); 1669 1670 // Post-merge 1671 set_control(_gvn.transform(r)); 1672 record_for_igvn(r); 1673 n = _gvn.transform(phi); 1674 1675 C->set_has_split_ifs(true); // Has chance for split-if optimization 1676 } 1677 set_result(n); 1678 return true; 1679 } 1680 1681 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1682 //------------------- 1683 //result=(result.isNaN())? funcAddr():result; 1684 // Check: If isNaN() by checking result!=result? then either trap 1685 // or go to runtime 1686 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result)); 1687 // Build the boolean node 1688 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq)); 1689 1690 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1691 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1692 // The pow or exp intrinsic returned a NaN, which requires a call 1693 // to the runtime. Recompile with the runtime call. 1694 uncommon_trap(Deoptimization::Reason_intrinsic, 1695 Deoptimization::Action_make_not_entrant); 1696 } 1697 return result; 1698 } else { 1699 // If this inlining ever returned NaN in the past, we compile a call 1700 // to the runtime to properly handle corner cases 1701 1702 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1703 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff)); 1704 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff)); 1705 1706 if (!if_slow->is_top()) { 1707 RegionNode* result_region = new (C) RegionNode(3); 1708 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE); 1709 1710 result_region->init_req(1, if_fast); 1711 result_val->init_req(1, result); 1712 1713 set_control(if_slow); 1714 1715 const TypePtr* no_memory_effects = NULL; 1716 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1717 no_memory_effects, 1718 x, top(), y, y ? top() : NULL); 1719 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0)); 1720 #ifdef ASSERT 1721 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1)); 1722 assert(value_top == top(), "second value must be top"); 1723 #endif 1724 1725 result_region->init_req(2, control()); 1726 result_val->init_req(2, value); 1727 set_control(_gvn.transform(result_region)); 1728 return _gvn.transform(result_val); 1729 } else { 1730 return result; 1731 } 1732 } 1733 } 1734 1735 //------------------------------inline_exp------------------------------------- 1736 // Inline exp instructions, if possible. The Intel hardware only misses 1737 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1738 bool LibraryCallKit::inline_exp() { 1739 Node* arg = round_double_node(argument(0)); 1740 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg)); 1741 1742 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1743 set_result(n); 1744 1745 C->set_has_split_ifs(true); // Has chance for split-if optimization 1746 return true; 1747 } 1748 1749 //------------------------------inline_pow------------------------------------- 1750 // Inline power instructions, if possible. 1751 bool LibraryCallKit::inline_pow() { 1752 // Pseudocode for pow 1753 // if (y == 2) { 1754 // return x * x; 1755 // } else { 1756 // if (x <= 0.0) { 1757 // long longy = (long)y; 1758 // if ((double)longy == y) { // if y is long 1759 // if (y + 1 == y) longy = 0; // huge number: even 1760 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1761 // } else { 1762 // result = NaN; 1763 // } 1764 // } else { 1765 // result = DPow(x,y); 1766 // } 1767 // if (result != result)? { 1768 // result = uncommon_trap() or runtime_call(); 1769 // } 1770 // return result; 1771 // } 1772 1773 Node* x = round_double_node(argument(0)); 1774 Node* y = round_double_node(argument(2)); 1775 1776 Node* result = NULL; 1777 1778 Node* const_two_node = makecon(TypeD::make(2.0)); 1779 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node)); 1780 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq)); 1781 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1782 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node)); 1783 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node)); 1784 1785 RegionNode* region_node = new (C) RegionNode(3); 1786 region_node->init_req(1, if_true); 1787 1788 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE); 1789 // special case for x^y where y == 2, we can convert it to x * x 1790 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x))); 1791 1792 // set control to if_false since we will now process the false branch 1793 set_control(if_false); 1794 1795 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1796 // Short form: skip the fancy tests and just check for NaN result. 1797 result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1798 } else { 1799 // If this inlining ever returned NaN in the past, include all 1800 // checks + call to the runtime. 1801 1802 // Set the merge point for If node with condition of (x <= 0.0) 1803 // There are four possible paths to region node and phi node 1804 RegionNode *r = new (C) RegionNode(4); 1805 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1806 1807 // Build the first if node: if (x <= 0.0) 1808 // Node for 0 constant 1809 Node *zeronode = makecon(TypeD::ZERO); 1810 // Check x:0 1811 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode)); 1812 // Check: If (x<=0) then go complex path 1813 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le )); 1814 // Branch either way 1815 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1816 // Fast path taken; set region slot 3 1817 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1)); 1818 r->init_req(3,fast_taken); // Capture fast-control 1819 1820 // Fast path not-taken, i.e. slow path 1821 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1)); 1822 1823 // Set fast path result 1824 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1825 phi->init_req(3, fast_result); 1826 1827 // Complex path 1828 // Build the second if node (if y is long) 1829 // Node for (long)y 1830 Node *longy = _gvn.transform(new (C) ConvD2LNode(y)); 1831 // Node for (double)((long) y) 1832 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy)); 1833 // Check (double)((long) y) : y 1834 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y)); 1835 // Check if (y isn't long) then go to slow path 1836 1837 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne )); 1838 // Branch either way 1839 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1840 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2)); 1841 1842 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2)); 1843 1844 // Calculate DPow(abs(x), y)*(1 & (long)y) 1845 // Node for constant 1 1846 Node *conone = longcon(1); 1847 // 1& (long)y 1848 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy)); 1849 1850 // A huge number is always even. Detect a huge number by checking 1851 // if y + 1 == y and set integer to be tested for parity to 0. 1852 // Required for corner case: 1853 // (long)9.223372036854776E18 = max_jlong 1854 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1855 // max_jlong is odd but 9.223372036854776E18 is even 1856 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1)))); 1857 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y)); 1858 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq )); 1859 Node* correctedsign = NULL; 1860 if (ConditionalMoveLimit != 0) { 1861 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1862 } else { 1863 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1864 RegionNode *r = new (C) RegionNode(3); 1865 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 1866 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1))); 1867 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1))); 1868 phi->init_req(1, signnode); 1869 phi->init_req(2, longcon(0)); 1870 correctedsign = _gvn.transform(phi); 1871 ylong_path = _gvn.transform(r); 1872 record_for_igvn(r); 1873 } 1874 1875 // zero node 1876 Node *conzero = longcon(0); 1877 // Check (1&(long)y)==0? 1878 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero)); 1879 // Check if (1&(long)y)!=0?, if so the result is negative 1880 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne )); 1881 // abs(x) 1882 Node *absx=_gvn.transform(new (C) AbsDNode(x)); 1883 // abs(x)^y 1884 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y)); 1885 // -abs(x)^y 1886 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy)); 1887 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1888 Node *signresult = NULL; 1889 if (ConditionalMoveLimit != 0) { 1890 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1891 } else { 1892 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1893 RegionNode *r = new (C) RegionNode(3); 1894 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1895 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven))); 1896 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven))); 1897 phi->init_req(1, absxpowy); 1898 phi->init_req(2, negabsxpowy); 1899 signresult = _gvn.transform(phi); 1900 ylong_path = _gvn.transform(r); 1901 record_for_igvn(r); 1902 } 1903 // Set complex path fast result 1904 r->init_req(2, ylong_path); 1905 phi->init_req(2, signresult); 1906 1907 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1908 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1909 r->init_req(1,slow_path); 1910 phi->init_req(1,slow_result); 1911 1912 // Post merge 1913 set_control(_gvn.transform(r)); 1914 record_for_igvn(r); 1915 result = _gvn.transform(phi); 1916 } 1917 1918 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1919 1920 // control from finish_pow_exp is now input to the region node 1921 region_node->set_req(2, control()); 1922 // the result from finish_pow_exp is now input to the phi node 1923 phi_node->init_req(2, result); 1924 set_control(_gvn.transform(region_node)); 1925 record_for_igvn(region_node); 1926 set_result(_gvn.transform(phi_node)); 1927 1928 C->set_has_split_ifs(true); // Has chance for split-if optimization 1929 return true; 1930 } 1931 1932 //------------------------------runtime_math----------------------------- 1933 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1934 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1935 "must be (DD)D or (D)D type"); 1936 1937 // Inputs 1938 Node* a = round_double_node(argument(0)); 1939 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1940 1941 const TypePtr* no_memory_effects = NULL; 1942 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1943 no_memory_effects, 1944 a, top(), b, b ? top() : NULL); 1945 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0)); 1946 #ifdef ASSERT 1947 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1)); 1948 assert(value_top == top(), "second value must be top"); 1949 #endif 1950 1951 set_result(value); 1952 return true; 1953 } 1954 1955 //------------------------------inline_math_native----------------------------- 1956 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1957 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1958 switch (id) { 1959 // These intrinsics are not properly supported on all hardware 1960 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1961 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1962 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1963 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1964 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1965 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1966 1967 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1968 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1969 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1970 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1971 1972 // These intrinsics are supported on all hardware 1973 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1974 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1975 1976 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1977 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1978 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1979 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1980 #undef FN_PTR 1981 1982 // These intrinsics are not yet correctly implemented 1983 case vmIntrinsics::_datan2: 1984 return false; 1985 1986 default: 1987 fatal_unexpected_iid(id); 1988 return false; 1989 } 1990 } 1991 1992 static bool is_simple_name(Node* n) { 1993 return (n->req() == 1 // constant 1994 || (n->is_Type() && n->as_Type()->type()->singleton()) 1995 || n->is_Proj() // parameter or return value 1996 || n->is_Phi() // local of some sort 1997 ); 1998 } 1999 2000 //----------------------------inline_min_max----------------------------------- 2001 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2002 set_result(generate_min_max(id, argument(0), argument(1))); 2003 return true; 2004 } 2005 2006 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2007 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) ); 2008 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2009 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check)); 2010 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) ); 2011 2012 { 2013 PreserveJVMState pjvms(this); 2014 PreserveReexecuteState preexecs(this); 2015 jvms()->set_should_reexecute(true); 2016 2017 set_control(slow_path); 2018 set_i_o(i_o()); 2019 2020 uncommon_trap(Deoptimization::Reason_intrinsic, 2021 Deoptimization::Action_none); 2022 } 2023 2024 set_control(fast_path); 2025 set_result(math); 2026 } 2027 2028 template <typename OverflowOp> 2029 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2030 typedef typename OverflowOp::MathOp MathOp; 2031 2032 MathOp* mathOp = new(C) MathOp(arg1, arg2); 2033 Node* operation = _gvn.transform( mathOp ); 2034 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) ); 2035 inline_math_mathExact(operation, ofcheck); 2036 return true; 2037 } 2038 2039 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2040 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2041 } 2042 2043 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2044 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2045 } 2046 2047 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2048 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2049 } 2050 2051 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2052 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2053 } 2054 2055 bool LibraryCallKit::inline_math_negateExactI() { 2056 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2057 } 2058 2059 bool LibraryCallKit::inline_math_negateExactL() { 2060 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2061 } 2062 2063 bool LibraryCallKit::inline_math_multiplyExactI() { 2064 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2065 } 2066 2067 bool LibraryCallKit::inline_math_multiplyExactL() { 2068 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2069 } 2070 2071 Node* 2072 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2073 // These are the candidate return value: 2074 Node* xvalue = x0; 2075 Node* yvalue = y0; 2076 2077 if (xvalue == yvalue) { 2078 return xvalue; 2079 } 2080 2081 bool want_max = (id == vmIntrinsics::_max); 2082 2083 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2084 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2085 if (txvalue == NULL || tyvalue == NULL) return top(); 2086 // This is not really necessary, but it is consistent with a 2087 // hypothetical MaxINode::Value method: 2088 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2089 2090 // %%% This folding logic should (ideally) be in a different place. 2091 // Some should be inside IfNode, and there to be a more reliable 2092 // transformation of ?: style patterns into cmoves. We also want 2093 // more powerful optimizations around cmove and min/max. 2094 2095 // Try to find a dominating comparison of these guys. 2096 // It can simplify the index computation for Arrays.copyOf 2097 // and similar uses of System.arraycopy. 2098 // First, compute the normalized version of CmpI(x, y). 2099 int cmp_op = Op_CmpI; 2100 Node* xkey = xvalue; 2101 Node* ykey = yvalue; 2102 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey)); 2103 if (ideal_cmpxy->is_Cmp()) { 2104 // E.g., if we have CmpI(length - offset, count), 2105 // it might idealize to CmpI(length, count + offset) 2106 cmp_op = ideal_cmpxy->Opcode(); 2107 xkey = ideal_cmpxy->in(1); 2108 ykey = ideal_cmpxy->in(2); 2109 } 2110 2111 // Start by locating any relevant comparisons. 2112 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2113 Node* cmpxy = NULL; 2114 Node* cmpyx = NULL; 2115 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2116 Node* cmp = start_from->fast_out(k); 2117 if (cmp->outcnt() > 0 && // must have prior uses 2118 cmp->in(0) == NULL && // must be context-independent 2119 cmp->Opcode() == cmp_op) { // right kind of compare 2120 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2121 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2122 } 2123 } 2124 2125 const int NCMPS = 2; 2126 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2127 int cmpn; 2128 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2129 if (cmps[cmpn] != NULL) break; // find a result 2130 } 2131 if (cmpn < NCMPS) { 2132 // Look for a dominating test that tells us the min and max. 2133 int depth = 0; // Limit search depth for speed 2134 Node* dom = control(); 2135 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2136 if (++depth >= 100) break; 2137 Node* ifproj = dom; 2138 if (!ifproj->is_Proj()) continue; 2139 Node* iff = ifproj->in(0); 2140 if (!iff->is_If()) continue; 2141 Node* bol = iff->in(1); 2142 if (!bol->is_Bool()) continue; 2143 Node* cmp = bol->in(1); 2144 if (cmp == NULL) continue; 2145 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2146 if (cmps[cmpn] == cmp) break; 2147 if (cmpn == NCMPS) continue; 2148 BoolTest::mask btest = bol->as_Bool()->_test._test; 2149 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2150 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2151 // At this point, we know that 'x btest y' is true. 2152 switch (btest) { 2153 case BoolTest::eq: 2154 // They are proven equal, so we can collapse the min/max. 2155 // Either value is the answer. Choose the simpler. 2156 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2157 return yvalue; 2158 return xvalue; 2159 case BoolTest::lt: // x < y 2160 case BoolTest::le: // x <= y 2161 return (want_max ? yvalue : xvalue); 2162 case BoolTest::gt: // x > y 2163 case BoolTest::ge: // x >= y 2164 return (want_max ? xvalue : yvalue); 2165 } 2166 } 2167 } 2168 2169 // We failed to find a dominating test. 2170 // Let's pick a test that might GVN with prior tests. 2171 Node* best_bol = NULL; 2172 BoolTest::mask best_btest = BoolTest::illegal; 2173 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2174 Node* cmp = cmps[cmpn]; 2175 if (cmp == NULL) continue; 2176 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2177 Node* bol = cmp->fast_out(j); 2178 if (!bol->is_Bool()) continue; 2179 BoolTest::mask btest = bol->as_Bool()->_test._test; 2180 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2181 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2182 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2183 best_bol = bol->as_Bool(); 2184 best_btest = btest; 2185 } 2186 } 2187 } 2188 2189 Node* answer_if_true = NULL; 2190 Node* answer_if_false = NULL; 2191 switch (best_btest) { 2192 default: 2193 if (cmpxy == NULL) 2194 cmpxy = ideal_cmpxy; 2195 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt)); 2196 // and fall through: 2197 case BoolTest::lt: // x < y 2198 case BoolTest::le: // x <= y 2199 answer_if_true = (want_max ? yvalue : xvalue); 2200 answer_if_false = (want_max ? xvalue : yvalue); 2201 break; 2202 case BoolTest::gt: // x > y 2203 case BoolTest::ge: // x >= y 2204 answer_if_true = (want_max ? xvalue : yvalue); 2205 answer_if_false = (want_max ? yvalue : xvalue); 2206 break; 2207 } 2208 2209 jint hi, lo; 2210 if (want_max) { 2211 // We can sharpen the minimum. 2212 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2213 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2214 } else { 2215 // We can sharpen the maximum. 2216 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2217 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2218 } 2219 2220 // Use a flow-free graph structure, to avoid creating excess control edges 2221 // which could hinder other optimizations. 2222 // Since Math.min/max is often used with arraycopy, we want 2223 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2224 Node* cmov = CMoveNode::make(C, NULL, best_bol, 2225 answer_if_false, answer_if_true, 2226 TypeInt::make(lo, hi, widen)); 2227 2228 return _gvn.transform(cmov); 2229 2230 /* 2231 // This is not as desirable as it may seem, since Min and Max 2232 // nodes do not have a full set of optimizations. 2233 // And they would interfere, anyway, with 'if' optimizations 2234 // and with CMoveI canonical forms. 2235 switch (id) { 2236 case vmIntrinsics::_min: 2237 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2238 case vmIntrinsics::_max: 2239 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2240 default: 2241 ShouldNotReachHere(); 2242 } 2243 */ 2244 } 2245 2246 inline int 2247 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2248 const TypePtr* base_type = TypePtr::NULL_PTR; 2249 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2250 if (base_type == NULL) { 2251 // Unknown type. 2252 return Type::AnyPtr; 2253 } else if (base_type == TypePtr::NULL_PTR) { 2254 // Since this is a NULL+long form, we have to switch to a rawptr. 2255 base = _gvn.transform(new (C) CastX2PNode(offset)); 2256 offset = MakeConX(0); 2257 return Type::RawPtr; 2258 } else if (base_type->base() == Type::RawPtr) { 2259 return Type::RawPtr; 2260 } else if (base_type->isa_oopptr()) { 2261 // Base is never null => always a heap address. 2262 if (base_type->ptr() == TypePtr::NotNull) { 2263 return Type::OopPtr; 2264 } 2265 // Offset is small => always a heap address. 2266 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2267 if (offset_type != NULL && 2268 base_type->offset() == 0 && // (should always be?) 2269 offset_type->_lo >= 0 && 2270 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2271 return Type::OopPtr; 2272 } 2273 // Otherwise, it might either be oop+off or NULL+addr. 2274 return Type::AnyPtr; 2275 } else { 2276 // No information: 2277 return Type::AnyPtr; 2278 } 2279 } 2280 2281 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2282 int kind = classify_unsafe_addr(base, offset); 2283 if (kind == Type::RawPtr) { 2284 return basic_plus_adr(top(), base, offset); 2285 } else { 2286 return basic_plus_adr(base, offset); 2287 } 2288 } 2289 2290 //--------------------------inline_number_methods----------------------------- 2291 // inline int Integer.numberOfLeadingZeros(int) 2292 // inline int Long.numberOfLeadingZeros(long) 2293 // 2294 // inline int Integer.numberOfTrailingZeros(int) 2295 // inline int Long.numberOfTrailingZeros(long) 2296 // 2297 // inline int Integer.bitCount(int) 2298 // inline int Long.bitCount(long) 2299 // 2300 // inline char Character.reverseBytes(char) 2301 // inline short Short.reverseBytes(short) 2302 // inline int Integer.reverseBytes(int) 2303 // inline long Long.reverseBytes(long) 2304 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2305 Node* arg = argument(0); 2306 Node* n; 2307 switch (id) { 2308 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break; 2309 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break; 2310 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break; 2311 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break; 2312 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break; 2313 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break; 2314 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break; 2315 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break; 2316 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break; 2317 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break; 2318 default: fatal_unexpected_iid(id); break; 2319 } 2320 set_result(_gvn.transform(n)); 2321 return true; 2322 } 2323 2324 //----------------------------inline_unsafe_access---------------------------- 2325 2326 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2327 2328 // Helper that guards and inserts a pre-barrier. 2329 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2330 Node* pre_val, bool need_mem_bar) { 2331 // We could be accessing the referent field of a reference object. If so, when G1 2332 // is enabled, we need to log the value in the referent field in an SATB buffer. 2333 // This routine performs some compile time filters and generates suitable 2334 // runtime filters that guard the pre-barrier code. 2335 // Also add memory barrier for non volatile load from the referent field 2336 // to prevent commoning of loads across safepoint. 2337 if (!UseG1GC && !need_mem_bar) 2338 return; 2339 2340 // Some compile time checks. 2341 2342 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2343 const TypeX* otype = offset->find_intptr_t_type(); 2344 if (otype != NULL && otype->is_con() && 2345 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2346 // Constant offset but not the reference_offset so just return 2347 return; 2348 } 2349 2350 // We only need to generate the runtime guards for instances. 2351 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2352 if (btype != NULL) { 2353 if (btype->isa_aryptr()) { 2354 // Array type so nothing to do 2355 return; 2356 } 2357 2358 const TypeInstPtr* itype = btype->isa_instptr(); 2359 if (itype != NULL) { 2360 // Can the klass of base_oop be statically determined to be 2361 // _not_ a sub-class of Reference and _not_ Object? 2362 ciKlass* klass = itype->klass(); 2363 if ( klass->is_loaded() && 2364 !klass->is_subtype_of(env()->Reference_klass()) && 2365 !env()->Object_klass()->is_subtype_of(klass)) { 2366 return; 2367 } 2368 } 2369 } 2370 2371 // The compile time filters did not reject base_oop/offset so 2372 // we need to generate the following runtime filters 2373 // 2374 // if (offset == java_lang_ref_Reference::_reference_offset) { 2375 // if (instance_of(base, java.lang.ref.Reference)) { 2376 // pre_barrier(_, pre_val, ...); 2377 // } 2378 // } 2379 2380 float likely = PROB_LIKELY( 0.999); 2381 float unlikely = PROB_UNLIKELY(0.999); 2382 2383 IdealKit ideal(this); 2384 #define __ ideal. 2385 2386 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2387 2388 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2389 // Update graphKit memory and control from IdealKit. 2390 sync_kit(ideal); 2391 2392 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2393 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2394 2395 // Update IdealKit memory and control from graphKit. 2396 __ sync_kit(this); 2397 2398 Node* one = __ ConI(1); 2399 // is_instof == 0 if base_oop == NULL 2400 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2401 2402 // Update graphKit from IdeakKit. 2403 sync_kit(ideal); 2404 2405 // Use the pre-barrier to record the value in the referent field 2406 pre_barrier(false /* do_load */, 2407 __ ctrl(), 2408 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2409 pre_val /* pre_val */, 2410 T_OBJECT); 2411 if (need_mem_bar) { 2412 // Add memory barrier to prevent commoning reads from this field 2413 // across safepoint since GC can change its value. 2414 insert_mem_bar(Op_MemBarCPUOrder); 2415 } 2416 // Update IdealKit from graphKit. 2417 __ sync_kit(this); 2418 2419 } __ end_if(); // _ref_type != ref_none 2420 } __ end_if(); // offset == referent_offset 2421 2422 // Final sync IdealKit and GraphKit. 2423 final_sync(ideal); 2424 #undef __ 2425 } 2426 2427 2428 // Interpret Unsafe.fieldOffset cookies correctly: 2429 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2430 2431 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2432 // Attempt to infer a sharper value type from the offset and base type. 2433 ciKlass* sharpened_klass = NULL; 2434 2435 // See if it is an instance field, with an object type. 2436 if (alias_type->field() != NULL) { 2437 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2438 if (alias_type->field()->type()->is_klass()) { 2439 sharpened_klass = alias_type->field()->type()->as_klass(); 2440 } 2441 } 2442 2443 // See if it is a narrow oop array. 2444 if (adr_type->isa_aryptr()) { 2445 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2446 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2447 if (elem_type != NULL) { 2448 sharpened_klass = elem_type->klass(); 2449 } 2450 } 2451 } 2452 2453 // The sharpened class might be unloaded if there is no class loader 2454 // contraint in place. 2455 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2456 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2457 2458 #ifndef PRODUCT 2459 if (C->print_intrinsics() || C->print_inlining()) { 2460 tty->print(" from base type: "); adr_type->dump(); 2461 tty->print(" sharpened value: "); tjp->dump(); 2462 } 2463 #endif 2464 // Sharpen the value type. 2465 return tjp; 2466 } 2467 return NULL; 2468 } 2469 2470 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2471 if (callee()->is_static()) return false; // caller must have the capability! 2472 2473 #ifndef PRODUCT 2474 { 2475 ResourceMark rm; 2476 // Check the signatures. 2477 ciSignature* sig = callee()->signature(); 2478 #ifdef ASSERT 2479 if (!is_store) { 2480 // Object getObject(Object base, int/long offset), etc. 2481 BasicType rtype = sig->return_type()->basic_type(); 2482 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2483 rtype = T_ADDRESS; // it is really a C void* 2484 assert(rtype == type, "getter must return the expected value"); 2485 if (!is_native_ptr) { 2486 assert(sig->count() == 2, "oop getter has 2 arguments"); 2487 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2488 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2489 } else { 2490 assert(sig->count() == 1, "native getter has 1 argument"); 2491 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2492 } 2493 } else { 2494 // void putObject(Object base, int/long offset, Object x), etc. 2495 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2496 if (!is_native_ptr) { 2497 assert(sig->count() == 3, "oop putter has 3 arguments"); 2498 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2499 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2500 } else { 2501 assert(sig->count() == 2, "native putter has 2 arguments"); 2502 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2503 } 2504 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2505 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2506 vtype = T_ADDRESS; // it is really a C void* 2507 assert(vtype == type, "putter must accept the expected value"); 2508 } 2509 #endif // ASSERT 2510 } 2511 #endif //PRODUCT 2512 2513 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2514 2515 Node* receiver = argument(0); // type: oop 2516 2517 // Build address expression. See the code in inline_unsafe_prefetch. 2518 Node* adr; 2519 Node* heap_base_oop = top(); 2520 Node* offset = top(); 2521 Node* val; 2522 2523 if (!is_native_ptr) { 2524 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2525 Node* base = argument(1); // type: oop 2526 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2527 offset = argument(2); // type: long 2528 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2529 // to be plain byte offsets, which are also the same as those accepted 2530 // by oopDesc::field_base. 2531 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2532 "fieldOffset must be byte-scaled"); 2533 // 32-bit machines ignore the high half! 2534 offset = ConvL2X(offset); 2535 adr = make_unsafe_address(base, offset); 2536 heap_base_oop = base; 2537 val = is_store ? argument(4) : NULL; 2538 } else { 2539 Node* ptr = argument(1); // type: long 2540 ptr = ConvL2X(ptr); // adjust Java long to machine word 2541 adr = make_unsafe_address(NULL, ptr); 2542 val = is_store ? argument(3) : NULL; 2543 } 2544 2545 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2546 2547 // First guess at the value type. 2548 const Type *value_type = Type::get_const_basic_type(type); 2549 2550 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2551 // there was not enough information to nail it down. 2552 Compile::AliasType* alias_type = C->alias_type(adr_type); 2553 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2554 2555 // We will need memory barriers unless we can determine a unique 2556 // alias category for this reference. (Note: If for some reason 2557 // the barriers get omitted and the unsafe reference begins to "pollute" 2558 // the alias analysis of the rest of the graph, either Compile::can_alias 2559 // or Compile::must_alias will throw a diagnostic assert.) 2560 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2561 2562 // If we are reading the value of the referent field of a Reference 2563 // object (either by using Unsafe directly or through reflection) 2564 // then, if G1 is enabled, we need to record the referent in an 2565 // SATB log buffer using the pre-barrier mechanism. 2566 // Also we need to add memory barrier to prevent commoning reads 2567 // from this field across safepoint since GC can change its value. 2568 bool need_read_barrier = !is_native_ptr && !is_store && 2569 offset != top() && heap_base_oop != top(); 2570 2571 if (!is_store && type == T_OBJECT) { 2572 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2573 if (tjp != NULL) { 2574 value_type = tjp; 2575 } 2576 } 2577 2578 receiver = null_check(receiver); 2579 if (stopped()) { 2580 return true; 2581 } 2582 // Heap pointers get a null-check from the interpreter, 2583 // as a courtesy. However, this is not guaranteed by Unsafe, 2584 // and it is not possible to fully distinguish unintended nulls 2585 // from intended ones in this API. 2586 2587 if (is_volatile) { 2588 // We need to emit leading and trailing CPU membars (see below) in 2589 // addition to memory membars when is_volatile. This is a little 2590 // too strong, but avoids the need to insert per-alias-type 2591 // volatile membars (for stores; compare Parse::do_put_xxx), which 2592 // we cannot do effectively here because we probably only have a 2593 // rough approximation of type. 2594 need_mem_bar = true; 2595 // For Stores, place a memory ordering barrier now. 2596 if (is_store) { 2597 insert_mem_bar(Op_MemBarRelease); 2598 } else { 2599 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2600 insert_mem_bar(Op_MemBarVolatile); 2601 } 2602 } 2603 } 2604 2605 // Memory barrier to prevent normal and 'unsafe' accesses from 2606 // bypassing each other. Happens after null checks, so the 2607 // exception paths do not take memory state from the memory barrier, 2608 // so there's no problems making a strong assert about mixing users 2609 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2610 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2611 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2612 2613 if (!is_store) { 2614 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile); 2615 // load value 2616 switch (type) { 2617 case T_BOOLEAN: 2618 case T_CHAR: 2619 case T_BYTE: 2620 case T_SHORT: 2621 case T_INT: 2622 case T_LONG: 2623 case T_FLOAT: 2624 case T_DOUBLE: 2625 break; 2626 case T_OBJECT: 2627 if (need_read_barrier) { 2628 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2629 } 2630 break; 2631 case T_ADDRESS: 2632 // Cast to an int type. 2633 p = _gvn.transform(new (C) CastP2XNode(NULL, p)); 2634 p = ConvX2UL(p); 2635 break; 2636 default: 2637 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2638 break; 2639 } 2640 // The load node has the control of the preceding MemBarCPUOrder. All 2641 // following nodes will have the control of the MemBarCPUOrder inserted at 2642 // the end of this method. So, pushing the load onto the stack at a later 2643 // point is fine. 2644 set_result(p); 2645 } else { 2646 // place effect of store into memory 2647 switch (type) { 2648 case T_DOUBLE: 2649 val = dstore_rounding(val); 2650 break; 2651 case T_ADDRESS: 2652 // Repackage the long as a pointer. 2653 val = ConvL2X(val); 2654 val = _gvn.transform(new (C) CastX2PNode(val)); 2655 break; 2656 } 2657 2658 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2659 if (type != T_OBJECT ) { 2660 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2661 } else { 2662 // Possibly an oop being stored to Java heap or native memory 2663 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2664 // oop to Java heap. 2665 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2666 } else { 2667 // We can't tell at compile time if we are storing in the Java heap or outside 2668 // of it. So we need to emit code to conditionally do the proper type of 2669 // store. 2670 2671 IdealKit ideal(this); 2672 #define __ ideal. 2673 // QQQ who knows what probability is here?? 2674 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2675 // Sync IdealKit and graphKit. 2676 sync_kit(ideal); 2677 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2678 // Update IdealKit memory. 2679 __ sync_kit(this); 2680 } __ else_(); { 2681 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2682 } __ end_if(); 2683 // Final sync IdealKit and GraphKit. 2684 final_sync(ideal); 2685 #undef __ 2686 } 2687 } 2688 } 2689 2690 if (is_volatile) { 2691 if (!is_store) { 2692 insert_mem_bar(Op_MemBarAcquire); 2693 } else { 2694 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2695 insert_mem_bar(Op_MemBarVolatile); 2696 } 2697 } 2698 } 2699 2700 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2701 2702 return true; 2703 } 2704 2705 //----------------------------inline_unsafe_prefetch---------------------------- 2706 2707 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2708 #ifndef PRODUCT 2709 { 2710 ResourceMark rm; 2711 // Check the signatures. 2712 ciSignature* sig = callee()->signature(); 2713 #ifdef ASSERT 2714 // Object getObject(Object base, int/long offset), etc. 2715 BasicType rtype = sig->return_type()->basic_type(); 2716 if (!is_native_ptr) { 2717 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2718 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2719 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2720 } else { 2721 assert(sig->count() == 1, "native prefetch has 1 argument"); 2722 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2723 } 2724 #endif // ASSERT 2725 } 2726 #endif // !PRODUCT 2727 2728 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2729 2730 const int idx = is_static ? 0 : 1; 2731 if (!is_static) { 2732 null_check_receiver(); 2733 if (stopped()) { 2734 return true; 2735 } 2736 } 2737 2738 // Build address expression. See the code in inline_unsafe_access. 2739 Node *adr; 2740 if (!is_native_ptr) { 2741 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2742 Node* base = argument(idx + 0); // type: oop 2743 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2744 Node* offset = argument(idx + 1); // type: long 2745 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2746 // to be plain byte offsets, which are also the same as those accepted 2747 // by oopDesc::field_base. 2748 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2749 "fieldOffset must be byte-scaled"); 2750 // 32-bit machines ignore the high half! 2751 offset = ConvL2X(offset); 2752 adr = make_unsafe_address(base, offset); 2753 } else { 2754 Node* ptr = argument(idx + 0); // type: long 2755 ptr = ConvL2X(ptr); // adjust Java long to machine word 2756 adr = make_unsafe_address(NULL, ptr); 2757 } 2758 2759 // Generate the read or write prefetch 2760 Node *prefetch; 2761 if (is_store) { 2762 prefetch = new (C) PrefetchWriteNode(i_o(), adr); 2763 } else { 2764 prefetch = new (C) PrefetchReadNode(i_o(), adr); 2765 } 2766 prefetch->init_req(0, control()); 2767 set_i_o(_gvn.transform(prefetch)); 2768 2769 return true; 2770 } 2771 2772 //----------------------------inline_unsafe_load_store---------------------------- 2773 // This method serves a couple of different customers (depending on LoadStoreKind): 2774 // 2775 // LS_cmpxchg: 2776 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2777 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2778 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2779 // 2780 // LS_xadd: 2781 // public int getAndAddInt( Object o, long offset, int delta) 2782 // public long getAndAddLong(Object o, long offset, long delta) 2783 // 2784 // LS_xchg: 2785 // int getAndSet(Object o, long offset, int newValue) 2786 // long getAndSet(Object o, long offset, long newValue) 2787 // Object getAndSet(Object o, long offset, Object newValue) 2788 // 2789 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2790 // This basic scheme here is the same as inline_unsafe_access, but 2791 // differs in enough details that combining them would make the code 2792 // overly confusing. (This is a true fact! I originally combined 2793 // them, but even I was confused by it!) As much code/comments as 2794 // possible are retained from inline_unsafe_access though to make 2795 // the correspondences clearer. - dl 2796 2797 if (callee()->is_static()) return false; // caller must have the capability! 2798 2799 #ifndef PRODUCT 2800 BasicType rtype; 2801 { 2802 ResourceMark rm; 2803 // Check the signatures. 2804 ciSignature* sig = callee()->signature(); 2805 rtype = sig->return_type()->basic_type(); 2806 if (kind == LS_xadd || kind == LS_xchg) { 2807 // Check the signatures. 2808 #ifdef ASSERT 2809 assert(rtype == type, "get and set must return the expected type"); 2810 assert(sig->count() == 3, "get and set has 3 arguments"); 2811 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2812 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2813 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2814 #endif // ASSERT 2815 } else if (kind == LS_cmpxchg) { 2816 // Check the signatures. 2817 #ifdef ASSERT 2818 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2819 assert(sig->count() == 4, "CAS has 4 arguments"); 2820 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2821 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2822 #endif // ASSERT 2823 } else { 2824 ShouldNotReachHere(); 2825 } 2826 } 2827 #endif //PRODUCT 2828 2829 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2830 2831 // Get arguments: 2832 Node* receiver = NULL; 2833 Node* base = NULL; 2834 Node* offset = NULL; 2835 Node* oldval = NULL; 2836 Node* newval = NULL; 2837 if (kind == LS_cmpxchg) { 2838 const bool two_slot_type = type2size[type] == 2; 2839 receiver = argument(0); // type: oop 2840 base = argument(1); // type: oop 2841 offset = argument(2); // type: long 2842 oldval = argument(4); // type: oop, int, or long 2843 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2844 } else if (kind == LS_xadd || kind == LS_xchg){ 2845 receiver = argument(0); // type: oop 2846 base = argument(1); // type: oop 2847 offset = argument(2); // type: long 2848 oldval = NULL; 2849 newval = argument(4); // type: oop, int, or long 2850 } 2851 2852 // Null check receiver. 2853 receiver = null_check(receiver); 2854 if (stopped()) { 2855 return true; 2856 } 2857 2858 // Build field offset expression. 2859 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2860 // to be plain byte offsets, which are also the same as those accepted 2861 // by oopDesc::field_base. 2862 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2863 // 32-bit machines ignore the high half of long offsets 2864 offset = ConvL2X(offset); 2865 Node* adr = make_unsafe_address(base, offset); 2866 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2867 2868 // For CAS, unlike inline_unsafe_access, there seems no point in 2869 // trying to refine types. Just use the coarse types here. 2870 const Type *value_type = Type::get_const_basic_type(type); 2871 Compile::AliasType* alias_type = C->alias_type(adr_type); 2872 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2873 2874 if (kind == LS_xchg && type == T_OBJECT) { 2875 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2876 if (tjp != NULL) { 2877 value_type = tjp; 2878 } 2879 } 2880 2881 int alias_idx = C->get_alias_index(adr_type); 2882 2883 // Memory-model-wise, a LoadStore acts like a little synchronized 2884 // block, so needs barriers on each side. These don't translate 2885 // into actual barriers on most machines, but we still need rest of 2886 // compiler to respect ordering. 2887 2888 insert_mem_bar(Op_MemBarRelease); 2889 insert_mem_bar(Op_MemBarCPUOrder); 2890 2891 // 4984716: MemBars must be inserted before this 2892 // memory node in order to avoid a false 2893 // dependency which will confuse the scheduler. 2894 Node *mem = memory(alias_idx); 2895 2896 // For now, we handle only those cases that actually exist: ints, 2897 // longs, and Object. Adding others should be straightforward. 2898 Node* load_store; 2899 switch(type) { 2900 case T_INT: 2901 if (kind == LS_xadd) { 2902 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type)); 2903 } else if (kind == LS_xchg) { 2904 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type)); 2905 } else if (kind == LS_cmpxchg) { 2906 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2907 } else { 2908 ShouldNotReachHere(); 2909 } 2910 break; 2911 case T_LONG: 2912 if (kind == LS_xadd) { 2913 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2914 } else if (kind == LS_xchg) { 2915 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2916 } else if (kind == LS_cmpxchg) { 2917 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2918 } else { 2919 ShouldNotReachHere(); 2920 } 2921 break; 2922 case T_OBJECT: 2923 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2924 // could be delayed during Parse (for example, in adjust_map_after_if()). 2925 // Execute transformation here to avoid barrier generation in such case. 2926 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2927 newval = _gvn.makecon(TypePtr::NULL_PTR); 2928 2929 // Reference stores need a store barrier. 2930 if (kind == LS_xchg) { 2931 // If pre-barrier must execute before the oop store, old value will require do_load here. 2932 if (!can_move_pre_barrier()) { 2933 pre_barrier(true /* do_load*/, 2934 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2935 NULL /* pre_val*/, 2936 T_OBJECT); 2937 } // Else move pre_barrier to use load_store value, see below. 2938 } else if (kind == LS_cmpxchg) { 2939 // Same as for newval above: 2940 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2941 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2942 } 2943 // The only known value which might get overwritten is oldval. 2944 pre_barrier(false /* do_load */, 2945 control(), NULL, NULL, max_juint, NULL, NULL, 2946 oldval /* pre_val */, 2947 T_OBJECT); 2948 } else { 2949 ShouldNotReachHere(); 2950 } 2951 2952 #ifdef _LP64 2953 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2954 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2955 if (kind == LS_xchg) { 2956 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr, 2957 newval_enc, adr_type, value_type->make_narrowoop())); 2958 } else { 2959 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2960 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2961 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr, 2962 newval_enc, oldval_enc)); 2963 } 2964 } else 2965 #endif 2966 { 2967 if (kind == LS_xchg) { 2968 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2969 } else { 2970 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2971 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2972 } 2973 } 2974 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2975 break; 2976 default: 2977 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2978 break; 2979 } 2980 2981 // SCMemProjNodes represent the memory state of a LoadStore. Their 2982 // main role is to prevent LoadStore nodes from being optimized away 2983 // when their results aren't used. 2984 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store)); 2985 set_memory(proj, alias_idx); 2986 2987 if (type == T_OBJECT && kind == LS_xchg) { 2988 #ifdef _LP64 2989 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2990 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type())); 2991 } 2992 #endif 2993 if (can_move_pre_barrier()) { 2994 // Don't need to load pre_val. The old value is returned by load_store. 2995 // The pre_barrier can execute after the xchg as long as no safepoint 2996 // gets inserted between them. 2997 pre_barrier(false /* do_load */, 2998 control(), NULL, NULL, max_juint, NULL, NULL, 2999 load_store /* pre_val */, 3000 T_OBJECT); 3001 } 3002 } 3003 3004 // Add the trailing membar surrounding the access 3005 insert_mem_bar(Op_MemBarCPUOrder); 3006 insert_mem_bar(Op_MemBarAcquire); 3007 3008 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3009 set_result(load_store); 3010 return true; 3011 } 3012 3013 //----------------------------inline_unsafe_ordered_store---------------------- 3014 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3015 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3016 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3017 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3018 // This is another variant of inline_unsafe_access, differing in 3019 // that it always issues store-store ("release") barrier and ensures 3020 // store-atomicity (which only matters for "long"). 3021 3022 if (callee()->is_static()) return false; // caller must have the capability! 3023 3024 #ifndef PRODUCT 3025 { 3026 ResourceMark rm; 3027 // Check the signatures. 3028 ciSignature* sig = callee()->signature(); 3029 #ifdef ASSERT 3030 BasicType rtype = sig->return_type()->basic_type(); 3031 assert(rtype == T_VOID, "must return void"); 3032 assert(sig->count() == 3, "has 3 arguments"); 3033 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3034 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3035 #endif // ASSERT 3036 } 3037 #endif //PRODUCT 3038 3039 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3040 3041 // Get arguments: 3042 Node* receiver = argument(0); // type: oop 3043 Node* base = argument(1); // type: oop 3044 Node* offset = argument(2); // type: long 3045 Node* val = argument(4); // type: oop, int, or long 3046 3047 // Null check receiver. 3048 receiver = null_check(receiver); 3049 if (stopped()) { 3050 return true; 3051 } 3052 3053 // Build field offset expression. 3054 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3055 // 32-bit machines ignore the high half of long offsets 3056 offset = ConvL2X(offset); 3057 Node* adr = make_unsafe_address(base, offset); 3058 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3059 const Type *value_type = Type::get_const_basic_type(type); 3060 Compile::AliasType* alias_type = C->alias_type(adr_type); 3061 3062 insert_mem_bar(Op_MemBarRelease); 3063 insert_mem_bar(Op_MemBarCPUOrder); 3064 // Ensure that the store is atomic for longs: 3065 const bool require_atomic_access = true; 3066 Node* store; 3067 if (type == T_OBJECT) // reference stores need a store barrier. 3068 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3069 else { 3070 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3071 } 3072 insert_mem_bar(Op_MemBarCPUOrder); 3073 return true; 3074 } 3075 3076 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3077 // Regardless of form, don't allow previous ld/st to move down, 3078 // then issue acquire, release, or volatile mem_bar. 3079 insert_mem_bar(Op_MemBarCPUOrder); 3080 switch(id) { 3081 case vmIntrinsics::_loadFence: 3082 insert_mem_bar(Op_LoadFence); 3083 return true; 3084 case vmIntrinsics::_storeFence: 3085 insert_mem_bar(Op_StoreFence); 3086 return true; 3087 case vmIntrinsics::_fullFence: 3088 insert_mem_bar(Op_MemBarVolatile); 3089 return true; 3090 default: 3091 fatal_unexpected_iid(id); 3092 return false; 3093 } 3094 } 3095 3096 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3097 if (!kls->is_Con()) { 3098 return true; 3099 } 3100 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3101 if (klsptr == NULL) { 3102 return true; 3103 } 3104 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3105 // don't need a guard for a klass that is already initialized 3106 return !ik->is_initialized(); 3107 } 3108 3109 //----------------------------inline_unsafe_allocate--------------------------- 3110 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3111 bool LibraryCallKit::inline_unsafe_allocate() { 3112 if (callee()->is_static()) return false; // caller must have the capability! 3113 3114 null_check_receiver(); // null-check, then ignore 3115 Node* cls = null_check(argument(1)); 3116 if (stopped()) return true; 3117 3118 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3119 kls = null_check(kls); 3120 if (stopped()) return true; // argument was like int.class 3121 3122 Node* test = NULL; 3123 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3124 // Note: The argument might still be an illegal value like 3125 // Serializable.class or Object[].class. The runtime will handle it. 3126 // But we must make an explicit check for initialization. 3127 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3128 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3129 // can generate code to load it as unsigned byte. 3130 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3131 Node* bits = intcon(InstanceKlass::fully_initialized); 3132 test = _gvn.transform(new (C) SubINode(inst, bits)); 3133 // The 'test' is non-zero if we need to take a slow path. 3134 } 3135 3136 Node* obj = new_instance(kls, test); 3137 set_result(obj); 3138 return true; 3139 } 3140 3141 #ifdef TRACE_HAVE_INTRINSICS 3142 /* 3143 * oop -> myklass 3144 * myklass->trace_id |= USED 3145 * return myklass->trace_id & ~0x3 3146 */ 3147 bool LibraryCallKit::inline_native_classID() { 3148 null_check_receiver(); // null-check, then ignore 3149 Node* cls = null_check(argument(1), T_OBJECT); 3150 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3151 kls = null_check(kls, T_OBJECT); 3152 ByteSize offset = TRACE_ID_OFFSET; 3153 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3154 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3155 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3156 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits)); 3157 Node* clsused = longcon(0x01l); // set the class bit 3158 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused)); 3159 3160 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3161 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3162 set_result(andl); 3163 return true; 3164 } 3165 3166 bool LibraryCallKit::inline_native_threadID() { 3167 Node* tls_ptr = NULL; 3168 Node* cur_thr = generate_current_thread(tls_ptr); 3169 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3170 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3171 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3172 3173 Node* threadid = NULL; 3174 size_t thread_id_size = OSThread::thread_id_size(); 3175 if (thread_id_size == (size_t) BytesPerLong) { 3176 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3177 } else if (thread_id_size == (size_t) BytesPerInt) { 3178 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3179 } else { 3180 ShouldNotReachHere(); 3181 } 3182 set_result(threadid); 3183 return true; 3184 } 3185 #endif 3186 3187 //------------------------inline_native_time_funcs-------------- 3188 // inline code for System.currentTimeMillis() and System.nanoTime() 3189 // these have the same type and signature 3190 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3191 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3192 const TypePtr* no_memory_effects = NULL; 3193 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3194 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0)); 3195 #ifdef ASSERT 3196 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1)); 3197 assert(value_top == top(), "second value must be top"); 3198 #endif 3199 set_result(value); 3200 return true; 3201 } 3202 3203 //------------------------inline_native_currentThread------------------ 3204 bool LibraryCallKit::inline_native_currentThread() { 3205 Node* junk = NULL; 3206 set_result(generate_current_thread(junk)); 3207 return true; 3208 } 3209 3210 //------------------------inline_native_isInterrupted------------------ 3211 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3212 bool LibraryCallKit::inline_native_isInterrupted() { 3213 // Add a fast path to t.isInterrupted(clear_int): 3214 // (t == Thread.current() && 3215 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3216 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3217 // So, in the common case that the interrupt bit is false, 3218 // we avoid making a call into the VM. Even if the interrupt bit 3219 // is true, if the clear_int argument is false, we avoid the VM call. 3220 // However, if the receiver is not currentThread, we must call the VM, 3221 // because there must be some locking done around the operation. 3222 3223 // We only go to the fast case code if we pass two guards. 3224 // Paths which do not pass are accumulated in the slow_region. 3225 3226 enum { 3227 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3228 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3229 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3230 PATH_LIMIT 3231 }; 3232 3233 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3234 // out of the function. 3235 insert_mem_bar(Op_MemBarCPUOrder); 3236 3237 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT); 3238 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL); 3239 3240 RegionNode* slow_region = new (C) RegionNode(1); 3241 record_for_igvn(slow_region); 3242 3243 // (a) Receiving thread must be the current thread. 3244 Node* rec_thr = argument(0); 3245 Node* tls_ptr = NULL; 3246 Node* cur_thr = generate_current_thread(tls_ptr); 3247 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr)); 3248 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne)); 3249 3250 generate_slow_guard(bol_thr, slow_region); 3251 3252 // (b) Interrupt bit on TLS must be false. 3253 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3254 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3255 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3256 3257 // Set the control input on the field _interrupted read to prevent it floating up. 3258 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3259 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0))); 3260 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne)); 3261 3262 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3263 3264 // First fast path: if (!TLS._interrupted) return false; 3265 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit)); 3266 result_rgn->init_req(no_int_result_path, false_bit); 3267 result_val->init_req(no_int_result_path, intcon(0)); 3268 3269 // drop through to next case 3270 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit))); 3271 3272 #ifndef TARGET_OS_FAMILY_windows 3273 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3274 Node* clr_arg = argument(1); 3275 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0))); 3276 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne)); 3277 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3278 3279 // Second fast path: ... else if (!clear_int) return true; 3280 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg)); 3281 result_rgn->init_req(no_clear_result_path, false_arg); 3282 result_val->init_req(no_clear_result_path, intcon(1)); 3283 3284 // drop through to next case 3285 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg))); 3286 #else 3287 // To return true on Windows you must read the _interrupted field 3288 // and check the the event state i.e. take the slow path. 3289 #endif // TARGET_OS_FAMILY_windows 3290 3291 // (d) Otherwise, go to the slow path. 3292 slow_region->add_req(control()); 3293 set_control( _gvn.transform(slow_region)); 3294 3295 if (stopped()) { 3296 // There is no slow path. 3297 result_rgn->init_req(slow_result_path, top()); 3298 result_val->init_req(slow_result_path, top()); 3299 } else { 3300 // non-virtual because it is a private non-static 3301 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3302 3303 Node* slow_val = set_results_for_java_call(slow_call); 3304 // this->control() comes from set_results_for_java_call 3305 3306 Node* fast_io = slow_call->in(TypeFunc::I_O); 3307 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3308 3309 // These two phis are pre-filled with copies of of the fast IO and Memory 3310 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3311 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3312 3313 result_rgn->init_req(slow_result_path, control()); 3314 result_io ->init_req(slow_result_path, i_o()); 3315 result_mem->init_req(slow_result_path, reset_memory()); 3316 result_val->init_req(slow_result_path, slow_val); 3317 3318 set_all_memory(_gvn.transform(result_mem)); 3319 set_i_o( _gvn.transform(result_io)); 3320 } 3321 3322 C->set_has_split_ifs(true); // Has chance for split-if optimization 3323 set_result(result_rgn, result_val); 3324 return true; 3325 } 3326 3327 //---------------------------load_mirror_from_klass---------------------------- 3328 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3329 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3330 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3331 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3332 } 3333 3334 //-----------------------load_klass_from_mirror_common------------------------- 3335 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3336 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3337 // and branch to the given path on the region. 3338 // If never_see_null, take an uncommon trap on null, so we can optimistically 3339 // compile for the non-null case. 3340 // If the region is NULL, force never_see_null = true. 3341 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3342 bool never_see_null, 3343 RegionNode* region, 3344 int null_path, 3345 int offset) { 3346 if (region == NULL) never_see_null = true; 3347 Node* p = basic_plus_adr(mirror, offset); 3348 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3349 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3350 Node* null_ctl = top(); 3351 kls = null_check_oop(kls, &null_ctl, never_see_null); 3352 if (region != NULL) { 3353 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3354 region->init_req(null_path, null_ctl); 3355 } else { 3356 assert(null_ctl == top(), "no loose ends"); 3357 } 3358 return kls; 3359 } 3360 3361 //--------------------(inline_native_Class_query helpers)--------------------- 3362 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3363 // Fall through if (mods & mask) == bits, take the guard otherwise. 3364 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3365 // Branch around if the given klass has the given modifier bit set. 3366 // Like generate_guard, adds a new path onto the region. 3367 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3368 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3369 Node* mask = intcon(modifier_mask); 3370 Node* bits = intcon(modifier_bits); 3371 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask)); 3372 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits)); 3373 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 3374 return generate_fair_guard(bol, region); 3375 } 3376 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3377 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3378 } 3379 3380 //-------------------------inline_native_Class_query------------------- 3381 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3382 const Type* return_type = TypeInt::BOOL; 3383 Node* prim_return_value = top(); // what happens if it's a primitive class? 3384 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3385 bool expect_prim = false; // most of these guys expect to work on refs 3386 3387 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3388 3389 Node* mirror = argument(0); 3390 Node* obj = top(); 3391 3392 switch (id) { 3393 case vmIntrinsics::_isInstance: 3394 // nothing is an instance of a primitive type 3395 prim_return_value = intcon(0); 3396 obj = argument(1); 3397 break; 3398 case vmIntrinsics::_getModifiers: 3399 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3400 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3401 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3402 break; 3403 case vmIntrinsics::_isInterface: 3404 prim_return_value = intcon(0); 3405 break; 3406 case vmIntrinsics::_isArray: 3407 prim_return_value = intcon(0); 3408 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3409 break; 3410 case vmIntrinsics::_isPrimitive: 3411 prim_return_value = intcon(1); 3412 expect_prim = true; // obviously 3413 break; 3414 case vmIntrinsics::_getSuperclass: 3415 prim_return_value = null(); 3416 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3417 break; 3418 case vmIntrinsics::_getComponentType: 3419 prim_return_value = null(); 3420 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3421 break; 3422 case vmIntrinsics::_getClassAccessFlags: 3423 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3424 return_type = TypeInt::INT; // not bool! 6297094 3425 break; 3426 default: 3427 fatal_unexpected_iid(id); 3428 break; 3429 } 3430 3431 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3432 if (mirror_con == NULL) return false; // cannot happen? 3433 3434 #ifndef PRODUCT 3435 if (C->print_intrinsics() || C->print_inlining()) { 3436 ciType* k = mirror_con->java_mirror_type(); 3437 if (k) { 3438 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3439 k->print_name(); 3440 tty->cr(); 3441 } 3442 } 3443 #endif 3444 3445 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3446 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3447 record_for_igvn(region); 3448 PhiNode* phi = new (C) PhiNode(region, return_type); 3449 3450 // The mirror will never be null of Reflection.getClassAccessFlags, however 3451 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3452 // if it is. See bug 4774291. 3453 3454 // For Reflection.getClassAccessFlags(), the null check occurs in 3455 // the wrong place; see inline_unsafe_access(), above, for a similar 3456 // situation. 3457 mirror = null_check(mirror); 3458 // If mirror or obj is dead, only null-path is taken. 3459 if (stopped()) return true; 3460 3461 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3462 3463 // Now load the mirror's klass metaobject, and null-check it. 3464 // Side-effects region with the control path if the klass is null. 3465 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3466 // If kls is null, we have a primitive mirror. 3467 phi->init_req(_prim_path, prim_return_value); 3468 if (stopped()) { set_result(region, phi); return true; } 3469 bool safe_for_replace = (region->in(_prim_path) == top()); 3470 3471 Node* p; // handy temp 3472 Node* null_ctl; 3473 3474 // Now that we have the non-null klass, we can perform the real query. 3475 // For constant classes, the query will constant-fold in LoadNode::Value. 3476 Node* query_value = top(); 3477 switch (id) { 3478 case vmIntrinsics::_isInstance: 3479 // nothing is an instance of a primitive type 3480 query_value = gen_instanceof(obj, kls, safe_for_replace); 3481 break; 3482 3483 case vmIntrinsics::_getModifiers: 3484 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3485 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3486 break; 3487 3488 case vmIntrinsics::_isInterface: 3489 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3490 if (generate_interface_guard(kls, region) != NULL) 3491 // A guard was added. If the guard is taken, it was an interface. 3492 phi->add_req(intcon(1)); 3493 // If we fall through, it's a plain class. 3494 query_value = intcon(0); 3495 break; 3496 3497 case vmIntrinsics::_isArray: 3498 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3499 if (generate_array_guard(kls, region) != NULL) 3500 // A guard was added. If the guard is taken, it was an array. 3501 phi->add_req(intcon(1)); 3502 // If we fall through, it's a plain class. 3503 query_value = intcon(0); 3504 break; 3505 3506 case vmIntrinsics::_isPrimitive: 3507 query_value = intcon(0); // "normal" path produces false 3508 break; 3509 3510 case vmIntrinsics::_getSuperclass: 3511 // The rules here are somewhat unfortunate, but we can still do better 3512 // with random logic than with a JNI call. 3513 // Interfaces store null or Object as _super, but must report null. 3514 // Arrays store an intermediate super as _super, but must report Object. 3515 // Other types can report the actual _super. 3516 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3517 if (generate_interface_guard(kls, region) != NULL) 3518 // A guard was added. If the guard is taken, it was an interface. 3519 phi->add_req(null()); 3520 if (generate_array_guard(kls, region) != NULL) 3521 // A guard was added. If the guard is taken, it was an array. 3522 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3523 // If we fall through, it's a plain class. Get its _super. 3524 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3525 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3526 null_ctl = top(); 3527 kls = null_check_oop(kls, &null_ctl); 3528 if (null_ctl != top()) { 3529 // If the guard is taken, Object.superClass is null (both klass and mirror). 3530 region->add_req(null_ctl); 3531 phi ->add_req(null()); 3532 } 3533 if (!stopped()) { 3534 query_value = load_mirror_from_klass(kls); 3535 } 3536 break; 3537 3538 case vmIntrinsics::_getComponentType: 3539 if (generate_array_guard(kls, region) != NULL) { 3540 // Be sure to pin the oop load to the guard edge just created: 3541 Node* is_array_ctrl = region->in(region->req()-1); 3542 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); 3543 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3544 phi->add_req(cmo); 3545 } 3546 query_value = null(); // non-array case is null 3547 break; 3548 3549 case vmIntrinsics::_getClassAccessFlags: 3550 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3551 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3552 break; 3553 3554 default: 3555 fatal_unexpected_iid(id); 3556 break; 3557 } 3558 3559 // Fall-through is the normal case of a query to a real class. 3560 phi->init_req(1, query_value); 3561 region->init_req(1, control()); 3562 3563 C->set_has_split_ifs(true); // Has chance for split-if optimization 3564 set_result(region, phi); 3565 return true; 3566 } 3567 3568 //--------------------------inline_native_subtype_check------------------------ 3569 // This intrinsic takes the JNI calls out of the heart of 3570 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3571 bool LibraryCallKit::inline_native_subtype_check() { 3572 // Pull both arguments off the stack. 3573 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3574 args[0] = argument(0); 3575 args[1] = argument(1); 3576 Node* klasses[2]; // corresponding Klasses: superk, subk 3577 klasses[0] = klasses[1] = top(); 3578 3579 enum { 3580 // A full decision tree on {superc is prim, subc is prim}: 3581 _prim_0_path = 1, // {P,N} => false 3582 // {P,P} & superc!=subc => false 3583 _prim_same_path, // {P,P} & superc==subc => true 3584 _prim_1_path, // {N,P} => false 3585 _ref_subtype_path, // {N,N} & subtype check wins => true 3586 _both_ref_path, // {N,N} & subtype check loses => false 3587 PATH_LIMIT 3588 }; 3589 3590 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3591 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 3592 record_for_igvn(region); 3593 3594 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3595 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3596 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3597 3598 // First null-check both mirrors and load each mirror's klass metaobject. 3599 int which_arg; 3600 for (which_arg = 0; which_arg <= 1; which_arg++) { 3601 Node* arg = args[which_arg]; 3602 arg = null_check(arg); 3603 if (stopped()) break; 3604 args[which_arg] = arg; 3605 3606 Node* p = basic_plus_adr(arg, class_klass_offset); 3607 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3608 klasses[which_arg] = _gvn.transform(kls); 3609 } 3610 3611 // Having loaded both klasses, test each for null. 3612 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3613 for (which_arg = 0; which_arg <= 1; which_arg++) { 3614 Node* kls = klasses[which_arg]; 3615 Node* null_ctl = top(); 3616 kls = null_check_oop(kls, &null_ctl, never_see_null); 3617 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3618 region->init_req(prim_path, null_ctl); 3619 if (stopped()) break; 3620 klasses[which_arg] = kls; 3621 } 3622 3623 if (!stopped()) { 3624 // now we have two reference types, in klasses[0..1] 3625 Node* subk = klasses[1]; // the argument to isAssignableFrom 3626 Node* superk = klasses[0]; // the receiver 3627 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3628 // now we have a successful reference subtype check 3629 region->set_req(_ref_subtype_path, control()); 3630 } 3631 3632 // If both operands are primitive (both klasses null), then 3633 // we must return true when they are identical primitives. 3634 // It is convenient to test this after the first null klass check. 3635 set_control(region->in(_prim_0_path)); // go back to first null check 3636 if (!stopped()) { 3637 // Since superc is primitive, make a guard for the superc==subc case. 3638 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1])); 3639 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq)); 3640 generate_guard(bol_eq, region, PROB_FAIR); 3641 if (region->req() == PATH_LIMIT+1) { 3642 // A guard was added. If the added guard is taken, superc==subc. 3643 region->swap_edges(PATH_LIMIT, _prim_same_path); 3644 region->del_req(PATH_LIMIT); 3645 } 3646 region->set_req(_prim_0_path, control()); // Not equal after all. 3647 } 3648 3649 // these are the only paths that produce 'true': 3650 phi->set_req(_prim_same_path, intcon(1)); 3651 phi->set_req(_ref_subtype_path, intcon(1)); 3652 3653 // pull together the cases: 3654 assert(region->req() == PATH_LIMIT, "sane region"); 3655 for (uint i = 1; i < region->req(); i++) { 3656 Node* ctl = region->in(i); 3657 if (ctl == NULL || ctl == top()) { 3658 region->set_req(i, top()); 3659 phi ->set_req(i, top()); 3660 } else if (phi->in(i) == NULL) { 3661 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3662 } 3663 } 3664 3665 set_control(_gvn.transform(region)); 3666 set_result(_gvn.transform(phi)); 3667 return true; 3668 } 3669 3670 //---------------------generate_array_guard_common------------------------ 3671 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3672 bool obj_array, bool not_array) { 3673 // If obj_array/non_array==false/false: 3674 // Branch around if the given klass is in fact an array (either obj or prim). 3675 // If obj_array/non_array==false/true: 3676 // Branch around if the given klass is not an array klass of any kind. 3677 // If obj_array/non_array==true/true: 3678 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3679 // If obj_array/non_array==true/false: 3680 // Branch around if the kls is an oop array (Object[] or subtype) 3681 // 3682 // Like generate_guard, adds a new path onto the region. 3683 jint layout_con = 0; 3684 Node* layout_val = get_layout_helper(kls, layout_con); 3685 if (layout_val == NULL) { 3686 bool query = (obj_array 3687 ? Klass::layout_helper_is_objArray(layout_con) 3688 : Klass::layout_helper_is_array(layout_con)); 3689 if (query == not_array) { 3690 return NULL; // never a branch 3691 } else { // always a branch 3692 Node* always_branch = control(); 3693 if (region != NULL) 3694 region->add_req(always_branch); 3695 set_control(top()); 3696 return always_branch; 3697 } 3698 } 3699 // Now test the correct condition. 3700 jint nval = (obj_array 3701 ? ((jint)Klass::_lh_array_tag_type_value 3702 << Klass::_lh_array_tag_shift) 3703 : Klass::_lh_neutral_value); 3704 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval))); 3705 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3706 // invert the test if we are looking for a non-array 3707 if (not_array) btest = BoolTest(btest).negate(); 3708 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest)); 3709 return generate_fair_guard(bol, region); 3710 } 3711 3712 3713 //-----------------------inline_native_newArray-------------------------- 3714 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3715 bool LibraryCallKit::inline_native_newArray() { 3716 Node* mirror = argument(0); 3717 Node* count_val = argument(1); 3718 3719 mirror = null_check(mirror); 3720 // If mirror or obj is dead, only null-path is taken. 3721 if (stopped()) return true; 3722 3723 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3724 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3725 PhiNode* result_val = new(C) PhiNode(result_reg, 3726 TypeInstPtr::NOTNULL); 3727 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3728 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3729 TypePtr::BOTTOM); 3730 3731 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3732 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3733 result_reg, _slow_path); 3734 Node* normal_ctl = control(); 3735 Node* no_array_ctl = result_reg->in(_slow_path); 3736 3737 // Generate code for the slow case. We make a call to newArray(). 3738 set_control(no_array_ctl); 3739 if (!stopped()) { 3740 // Either the input type is void.class, or else the 3741 // array klass has not yet been cached. Either the 3742 // ensuing call will throw an exception, or else it 3743 // will cache the array klass for next time. 3744 PreserveJVMState pjvms(this); 3745 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3746 Node* slow_result = set_results_for_java_call(slow_call); 3747 // this->control() comes from set_results_for_java_call 3748 result_reg->set_req(_slow_path, control()); 3749 result_val->set_req(_slow_path, slow_result); 3750 result_io ->set_req(_slow_path, i_o()); 3751 result_mem->set_req(_slow_path, reset_memory()); 3752 } 3753 3754 set_control(normal_ctl); 3755 if (!stopped()) { 3756 // Normal case: The array type has been cached in the java.lang.Class. 3757 // The following call works fine even if the array type is polymorphic. 3758 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3759 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3760 result_reg->init_req(_normal_path, control()); 3761 result_val->init_req(_normal_path, obj); 3762 result_io ->init_req(_normal_path, i_o()); 3763 result_mem->init_req(_normal_path, reset_memory()); 3764 } 3765 3766 // Return the combined state. 3767 set_i_o( _gvn.transform(result_io) ); 3768 set_all_memory( _gvn.transform(result_mem)); 3769 3770 C->set_has_split_ifs(true); // Has chance for split-if optimization 3771 set_result(result_reg, result_val); 3772 return true; 3773 } 3774 3775 //----------------------inline_native_getLength-------------------------- 3776 // public static native int java.lang.reflect.Array.getLength(Object array); 3777 bool LibraryCallKit::inline_native_getLength() { 3778 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3779 3780 Node* array = null_check(argument(0)); 3781 // If array is dead, only null-path is taken. 3782 if (stopped()) return true; 3783 3784 // Deoptimize if it is a non-array. 3785 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3786 3787 if (non_array != NULL) { 3788 PreserveJVMState pjvms(this); 3789 set_control(non_array); 3790 uncommon_trap(Deoptimization::Reason_intrinsic, 3791 Deoptimization::Action_maybe_recompile); 3792 } 3793 3794 // If control is dead, only non-array-path is taken. 3795 if (stopped()) return true; 3796 3797 // The works fine even if the array type is polymorphic. 3798 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3799 Node* result = load_array_length(array); 3800 3801 C->set_has_split_ifs(true); // Has chance for split-if optimization 3802 set_result(result); 3803 return true; 3804 } 3805 3806 //------------------------inline_array_copyOf---------------------------- 3807 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3808 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3809 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3810 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3811 3812 // Get the arguments. 3813 Node* original = argument(0); 3814 Node* start = is_copyOfRange? argument(1): intcon(0); 3815 Node* end = is_copyOfRange? argument(2): argument(1); 3816 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3817 3818 Node* newcopy; 3819 3820 // Set the original stack and the reexecute bit for the interpreter to reexecute 3821 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3822 { PreserveReexecuteState preexecs(this); 3823 jvms()->set_should_reexecute(true); 3824 3825 array_type_mirror = null_check(array_type_mirror); 3826 original = null_check(original); 3827 3828 // Check if a null path was taken unconditionally. 3829 if (stopped()) return true; 3830 3831 Node* orig_length = load_array_length(original); 3832 3833 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3834 klass_node = null_check(klass_node); 3835 3836 RegionNode* bailout = new (C) RegionNode(1); 3837 record_for_igvn(bailout); 3838 3839 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3840 // Bail out if that is so. 3841 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3842 if (not_objArray != NULL) { 3843 // Improve the klass node's type from the new optimistic assumption: 3844 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3845 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3846 Node* cast = new (C) CastPPNode(klass_node, akls); 3847 cast->init_req(0, control()); 3848 klass_node = _gvn.transform(cast); 3849 } 3850 3851 // Bail out if either start or end is negative. 3852 generate_negative_guard(start, bailout, &start); 3853 generate_negative_guard(end, bailout, &end); 3854 3855 Node* length = end; 3856 if (_gvn.type(start) != TypeInt::ZERO) { 3857 length = _gvn.transform(new (C) SubINode(end, start)); 3858 } 3859 3860 // Bail out if length is negative. 3861 // Without this the new_array would throw 3862 // NegativeArraySizeException but IllegalArgumentException is what 3863 // should be thrown 3864 generate_negative_guard(length, bailout, &length); 3865 3866 if (bailout->req() > 1) { 3867 PreserveJVMState pjvms(this); 3868 set_control(_gvn.transform(bailout)); 3869 uncommon_trap(Deoptimization::Reason_intrinsic, 3870 Deoptimization::Action_maybe_recompile); 3871 } 3872 3873 if (!stopped()) { 3874 // How many elements will we copy from the original? 3875 // The answer is MinI(orig_length - start, length). 3876 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start)); 3877 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3878 3879 newcopy = new_array(klass_node, length, 0); // no argments to push 3880 3881 // Generate a direct call to the right arraycopy function(s). 3882 // We know the copy is disjoint but we might not know if the 3883 // oop stores need checking. 3884 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3885 // This will fail a store-check if x contains any non-nulls. 3886 bool disjoint_bases = true; 3887 // if start > orig_length then the length of the copy may be 3888 // negative. 3889 bool length_never_negative = !is_copyOfRange; 3890 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3891 original, start, newcopy, intcon(0), moved, 3892 disjoint_bases, length_never_negative); 3893 } 3894 } // original reexecute is set back here 3895 3896 C->set_has_split_ifs(true); // Has chance for split-if optimization 3897 if (!stopped()) { 3898 set_result(newcopy); 3899 } 3900 return true; 3901 } 3902 3903 3904 //----------------------generate_virtual_guard--------------------------- 3905 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3906 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3907 RegionNode* slow_region) { 3908 ciMethod* method = callee(); 3909 int vtable_index = method->vtable_index(); 3910 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3911 err_msg_res("bad index %d", vtable_index)); 3912 // Get the Method* out of the appropriate vtable entry. 3913 int entry_offset = (InstanceKlass::vtable_start_offset() + 3914 vtable_index*vtableEntry::size()) * wordSize + 3915 vtableEntry::method_offset_in_bytes(); 3916 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3917 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3918 3919 // Compare the target method with the expected method (e.g., Object.hashCode). 3920 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3921 3922 Node* native_call = makecon(native_call_addr); 3923 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call)); 3924 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne)); 3925 3926 return generate_slow_guard(test_native, slow_region); 3927 } 3928 3929 //-----------------------generate_method_call---------------------------- 3930 // Use generate_method_call to make a slow-call to the real 3931 // method if the fast path fails. An alternative would be to 3932 // use a stub like OptoRuntime::slow_arraycopy_Java. 3933 // This only works for expanding the current library call, 3934 // not another intrinsic. (E.g., don't use this for making an 3935 // arraycopy call inside of the copyOf intrinsic.) 3936 CallJavaNode* 3937 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3938 // When compiling the intrinsic method itself, do not use this technique. 3939 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3940 3941 ciMethod* method = callee(); 3942 // ensure the JVMS we have will be correct for this call 3943 guarantee(method_id == method->intrinsic_id(), "must match"); 3944 3945 const TypeFunc* tf = TypeFunc::make(method); 3946 CallJavaNode* slow_call; 3947 if (is_static) { 3948 assert(!is_virtual, ""); 3949 slow_call = new(C) CallStaticJavaNode(C, tf, 3950 SharedRuntime::get_resolve_static_call_stub(), 3951 method, bci()); 3952 } else if (is_virtual) { 3953 null_check_receiver(); 3954 int vtable_index = Method::invalid_vtable_index; 3955 if (UseInlineCaches) { 3956 // Suppress the vtable call 3957 } else { 3958 // hashCode and clone are not a miranda methods, 3959 // so the vtable index is fixed. 3960 // No need to use the linkResolver to get it. 3961 vtable_index = method->vtable_index(); 3962 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3963 err_msg_res("bad index %d", vtable_index)); 3964 } 3965 slow_call = new(C) CallDynamicJavaNode(tf, 3966 SharedRuntime::get_resolve_virtual_call_stub(), 3967 method, vtable_index, bci()); 3968 } else { // neither virtual nor static: opt_virtual 3969 null_check_receiver(); 3970 slow_call = new(C) CallStaticJavaNode(C, tf, 3971 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3972 method, bci()); 3973 slow_call->set_optimized_virtual(true); 3974 } 3975 set_arguments_for_java_call(slow_call); 3976 set_edges_for_java_call(slow_call); 3977 return slow_call; 3978 } 3979 3980 3981 //------------------------------inline_native_hashcode-------------------- 3982 // Build special case code for calls to hashCode on an object. 3983 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3984 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3985 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3986 3987 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3988 3989 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3990 PhiNode* result_val = new(C) PhiNode(result_reg, 3991 TypeInt::INT); 3992 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3993 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3994 TypePtr::BOTTOM); 3995 Node* obj = NULL; 3996 if (!is_static) { 3997 // Check for hashing null object 3998 obj = null_check_receiver(); 3999 if (stopped()) return true; // unconditionally null 4000 result_reg->init_req(_null_path, top()); 4001 result_val->init_req(_null_path, top()); 4002 } else { 4003 // Do a null check, and return zero if null. 4004 // System.identityHashCode(null) == 0 4005 obj = argument(0); 4006 Node* null_ctl = top(); 4007 obj = null_check_oop(obj, &null_ctl); 4008 result_reg->init_req(_null_path, null_ctl); 4009 result_val->init_req(_null_path, _gvn.intcon(0)); 4010 } 4011 4012 // Unconditionally null? Then return right away. 4013 if (stopped()) { 4014 set_control( result_reg->in(_null_path)); 4015 if (!stopped()) 4016 set_result(result_val->in(_null_path)); 4017 return true; 4018 } 4019 4020 // After null check, get the object's klass. 4021 Node* obj_klass = load_object_klass(obj); 4022 4023 // This call may be virtual (invokevirtual) or bound (invokespecial). 4024 // For each case we generate slightly different code. 4025 4026 // We only go to the fast case code if we pass a number of guards. The 4027 // paths which do not pass are accumulated in the slow_region. 4028 RegionNode* slow_region = new (C) RegionNode(1); 4029 record_for_igvn(slow_region); 4030 4031 // If this is a virtual call, we generate a funny guard. We pull out 4032 // the vtable entry corresponding to hashCode() from the target object. 4033 // If the target method which we are calling happens to be the native 4034 // Object hashCode() method, we pass the guard. We do not need this 4035 // guard for non-virtual calls -- the caller is known to be the native 4036 // Object hashCode(). 4037 if (is_virtual) { 4038 generate_virtual_guard(obj_klass, slow_region); 4039 } 4040 4041 // Get the header out of the object, use LoadMarkNode when available 4042 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4043 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4044 4045 // Test the header to see if it is unlocked. 4046 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4047 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask)); 4048 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4049 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val)); 4050 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne)); 4051 4052 generate_slow_guard(test_unlocked, slow_region); 4053 4054 // Get the hash value and check to see that it has been properly assigned. 4055 // We depend on hash_mask being at most 32 bits and avoid the use of 4056 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4057 // vm: see markOop.hpp. 4058 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4059 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4060 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift)); 4061 // This hack lets the hash bits live anywhere in the mark object now, as long 4062 // as the shift drops the relevant bits into the low 32 bits. Note that 4063 // Java spec says that HashCode is an int so there's no point in capturing 4064 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4065 hshifted_header = ConvX2I(hshifted_header); 4066 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask)); 4067 4068 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4069 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val)); 4070 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq)); 4071 4072 generate_slow_guard(test_assigned, slow_region); 4073 4074 Node* init_mem = reset_memory(); 4075 // fill in the rest of the null path: 4076 result_io ->init_req(_null_path, i_o()); 4077 result_mem->init_req(_null_path, init_mem); 4078 4079 result_val->init_req(_fast_path, hash_val); 4080 result_reg->init_req(_fast_path, control()); 4081 result_io ->init_req(_fast_path, i_o()); 4082 result_mem->init_req(_fast_path, init_mem); 4083 4084 // Generate code for the slow case. We make a call to hashCode(). 4085 set_control(_gvn.transform(slow_region)); 4086 if (!stopped()) { 4087 // No need for PreserveJVMState, because we're using up the present state. 4088 set_all_memory(init_mem); 4089 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4090 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4091 Node* slow_result = set_results_for_java_call(slow_call); 4092 // this->control() comes from set_results_for_java_call 4093 result_reg->init_req(_slow_path, control()); 4094 result_val->init_req(_slow_path, slow_result); 4095 result_io ->set_req(_slow_path, i_o()); 4096 result_mem ->set_req(_slow_path, reset_memory()); 4097 } 4098 4099 // Return the combined state. 4100 set_i_o( _gvn.transform(result_io) ); 4101 set_all_memory( _gvn.transform(result_mem)); 4102 4103 set_result(result_reg, result_val); 4104 return true; 4105 } 4106 4107 //---------------------------inline_native_getClass---------------------------- 4108 // public final native Class<?> java.lang.Object.getClass(); 4109 // 4110 // Build special case code for calls to getClass on an object. 4111 bool LibraryCallKit::inline_native_getClass() { 4112 Node* obj = null_check_receiver(); 4113 if (stopped()) return true; 4114 set_result(load_mirror_from_klass(load_object_klass(obj))); 4115 return true; 4116 } 4117 4118 //-----------------inline_native_Reflection_getCallerClass--------------------- 4119 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4120 // 4121 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4122 // 4123 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4124 // in that it must skip particular security frames and checks for 4125 // caller sensitive methods. 4126 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4127 #ifndef PRODUCT 4128 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4129 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4130 } 4131 #endif 4132 4133 if (!jvms()->has_method()) { 4134 #ifndef PRODUCT 4135 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4136 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4137 } 4138 #endif 4139 return false; 4140 } 4141 4142 // Walk back up the JVM state to find the caller at the required 4143 // depth. 4144 JVMState* caller_jvms = jvms(); 4145 4146 // Cf. JVM_GetCallerClass 4147 // NOTE: Start the loop at depth 1 because the current JVM state does 4148 // not include the Reflection.getCallerClass() frame. 4149 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4150 ciMethod* m = caller_jvms->method(); 4151 switch (n) { 4152 case 0: 4153 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4154 break; 4155 case 1: 4156 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4157 if (!m->caller_sensitive()) { 4158 #ifndef PRODUCT 4159 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4160 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4161 } 4162 #endif 4163 return false; // bail-out; let JVM_GetCallerClass do the work 4164 } 4165 break; 4166 default: 4167 if (!m->is_ignored_by_security_stack_walk()) { 4168 // We have reached the desired frame; return the holder class. 4169 // Acquire method holder as java.lang.Class and push as constant. 4170 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4171 ciInstance* caller_mirror = caller_klass->java_mirror(); 4172 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4173 4174 #ifndef PRODUCT 4175 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4176 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()); 4177 tty->print_cr(" JVM state at this point:"); 4178 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4179 ciMethod* m = jvms()->of_depth(i)->method(); 4180 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4181 } 4182 } 4183 #endif 4184 return true; 4185 } 4186 break; 4187 } 4188 } 4189 4190 #ifndef PRODUCT 4191 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4192 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4193 tty->print_cr(" JVM state at this point:"); 4194 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4195 ciMethod* m = jvms()->of_depth(i)->method(); 4196 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4197 } 4198 } 4199 #endif 4200 4201 return false; // bail-out; let JVM_GetCallerClass do the work 4202 } 4203 4204 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4205 Node* arg = argument(0); 4206 Node* result; 4207 4208 switch (id) { 4209 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break; 4210 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break; 4211 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break; 4212 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break; 4213 4214 case vmIntrinsics::_doubleToLongBits: { 4215 // two paths (plus control) merge in a wood 4216 RegionNode *r = new (C) RegionNode(3); 4217 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 4218 4219 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg)); 4220 // Build the boolean node 4221 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4222 4223 // Branch either way. 4224 // NaN case is less traveled, which makes all the difference. 4225 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4226 Node *opt_isnan = _gvn.transform(ifisnan); 4227 assert( opt_isnan->is_If(), "Expect an IfNode"); 4228 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4229 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4230 4231 set_control(iftrue); 4232 4233 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4234 Node *slow_result = longcon(nan_bits); // return NaN 4235 phi->init_req(1, _gvn.transform( slow_result )); 4236 r->init_req(1, iftrue); 4237 4238 // Else fall through 4239 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4240 set_control(iffalse); 4241 4242 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg))); 4243 r->init_req(2, iffalse); 4244 4245 // Post merge 4246 set_control(_gvn.transform(r)); 4247 record_for_igvn(r); 4248 4249 C->set_has_split_ifs(true); // Has chance for split-if optimization 4250 result = phi; 4251 assert(result->bottom_type()->isa_long(), "must be"); 4252 break; 4253 } 4254 4255 case vmIntrinsics::_floatToIntBits: { 4256 // two paths (plus control) merge in a wood 4257 RegionNode *r = new (C) RegionNode(3); 4258 Node *phi = new (C) PhiNode(r, TypeInt::INT); 4259 4260 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg)); 4261 // Build the boolean node 4262 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4263 4264 // Branch either way. 4265 // NaN case is less traveled, which makes all the difference. 4266 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4267 Node *opt_isnan = _gvn.transform(ifisnan); 4268 assert( opt_isnan->is_If(), "Expect an IfNode"); 4269 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4270 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4271 4272 set_control(iftrue); 4273 4274 static const jint nan_bits = 0x7fc00000; 4275 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4276 phi->init_req(1, _gvn.transform( slow_result )); 4277 r->init_req(1, iftrue); 4278 4279 // Else fall through 4280 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4281 set_control(iffalse); 4282 4283 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg))); 4284 r->init_req(2, iffalse); 4285 4286 // Post merge 4287 set_control(_gvn.transform(r)); 4288 record_for_igvn(r); 4289 4290 C->set_has_split_ifs(true); // Has chance for split-if optimization 4291 result = phi; 4292 assert(result->bottom_type()->isa_int(), "must be"); 4293 break; 4294 } 4295 4296 default: 4297 fatal_unexpected_iid(id); 4298 break; 4299 } 4300 set_result(_gvn.transform(result)); 4301 return true; 4302 } 4303 4304 #ifdef _LP64 4305 #define XTOP ,top() /*additional argument*/ 4306 #else //_LP64 4307 #define XTOP /*no additional argument*/ 4308 #endif //_LP64 4309 4310 //----------------------inline_unsafe_copyMemory------------------------- 4311 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4312 bool LibraryCallKit::inline_unsafe_copyMemory() { 4313 if (callee()->is_static()) return false; // caller must have the capability! 4314 null_check_receiver(); // null-check receiver 4315 if (stopped()) return true; 4316 4317 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4318 4319 Node* src_ptr = argument(1); // type: oop 4320 Node* src_off = ConvL2X(argument(2)); // type: long 4321 Node* dst_ptr = argument(4); // type: oop 4322 Node* dst_off = ConvL2X(argument(5)); // type: long 4323 Node* size = ConvL2X(argument(7)); // type: long 4324 4325 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4326 "fieldOffset must be byte-scaled"); 4327 4328 Node* src = make_unsafe_address(src_ptr, src_off); 4329 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4330 4331 // Conservatively insert a memory barrier on all memory slices. 4332 // Do not let writes of the copy source or destination float below the copy. 4333 insert_mem_bar(Op_MemBarCPUOrder); 4334 4335 // Call it. Note that the length argument is not scaled. 4336 make_runtime_call(RC_LEAF|RC_NO_FP, 4337 OptoRuntime::fast_arraycopy_Type(), 4338 StubRoutines::unsafe_arraycopy(), 4339 "unsafe_arraycopy", 4340 TypeRawPtr::BOTTOM, 4341 src, dst, size XTOP); 4342 4343 // Do not let reads of the copy destination float above the copy. 4344 insert_mem_bar(Op_MemBarCPUOrder); 4345 4346 return true; 4347 } 4348 4349 //------------------------clone_coping----------------------------------- 4350 // Helper function for inline_native_clone. 4351 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4352 assert(obj_size != NULL, ""); 4353 Node* raw_obj = alloc_obj->in(1); 4354 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4355 4356 AllocateNode* alloc = NULL; 4357 if (ReduceBulkZeroing) { 4358 // We will be completely responsible for initializing this object - 4359 // mark Initialize node as complete. 4360 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4361 // The object was just allocated - there should be no any stores! 4362 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4363 // Mark as complete_with_arraycopy so that on AllocateNode 4364 // expansion, we know this AllocateNode is initialized by an array 4365 // copy and a StoreStore barrier exists after the array copy. 4366 alloc->initialization()->set_complete_with_arraycopy(); 4367 } 4368 4369 // Copy the fastest available way. 4370 // TODO: generate fields copies for small objects instead. 4371 Node* src = obj; 4372 Node* dest = alloc_obj; 4373 Node* size = _gvn.transform(obj_size); 4374 4375 // Exclude the header but include array length to copy by 8 bytes words. 4376 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4377 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4378 instanceOopDesc::base_offset_in_bytes(); 4379 // base_off: 4380 // 8 - 32-bit VM 4381 // 12 - 64-bit VM, compressed klass 4382 // 16 - 64-bit VM, normal klass 4383 if (base_off % BytesPerLong != 0) { 4384 assert(UseCompressedClassPointers, ""); 4385 if (is_array) { 4386 // Exclude length to copy by 8 bytes words. 4387 base_off += sizeof(int); 4388 } else { 4389 // Include klass to copy by 8 bytes words. 4390 base_off = instanceOopDesc::klass_offset_in_bytes(); 4391 } 4392 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4393 } 4394 src = basic_plus_adr(src, base_off); 4395 dest = basic_plus_adr(dest, base_off); 4396 4397 // Compute the length also, if needed: 4398 Node* countx = size; 4399 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off))); 4400 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4401 4402 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4403 bool disjoint_bases = true; 4404 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4405 src, NULL, dest, NULL, countx, 4406 /*dest_uninitialized*/true); 4407 4408 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4409 if (card_mark) { 4410 assert(!is_array, ""); 4411 // Put in store barrier for any and all oops we are sticking 4412 // into this object. (We could avoid this if we could prove 4413 // that the object type contains no oop fields at all.) 4414 Node* no_particular_value = NULL; 4415 Node* no_particular_field = NULL; 4416 int raw_adr_idx = Compile::AliasIdxRaw; 4417 post_barrier(control(), 4418 memory(raw_adr_type), 4419 alloc_obj, 4420 no_particular_field, 4421 raw_adr_idx, 4422 no_particular_value, 4423 T_OBJECT, 4424 false); 4425 } 4426 4427 // Do not let reads from the cloned object float above the arraycopy. 4428 if (alloc != NULL) { 4429 // Do not let stores that initialize this object be reordered with 4430 // a subsequent store that would make this object accessible by 4431 // other threads. 4432 // Record what AllocateNode this StoreStore protects so that 4433 // escape analysis can go from the MemBarStoreStoreNode to the 4434 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4435 // based on the escape status of the AllocateNode. 4436 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4437 } else { 4438 insert_mem_bar(Op_MemBarCPUOrder); 4439 } 4440 } 4441 4442 //------------------------inline_native_clone---------------------------- 4443 // protected native Object java.lang.Object.clone(); 4444 // 4445 // Here are the simple edge cases: 4446 // null receiver => normal trap 4447 // virtual and clone was overridden => slow path to out-of-line clone 4448 // not cloneable or finalizer => slow path to out-of-line Object.clone 4449 // 4450 // The general case has two steps, allocation and copying. 4451 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4452 // 4453 // Copying also has two cases, oop arrays and everything else. 4454 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4455 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4456 // 4457 // These steps fold up nicely if and when the cloned object's klass 4458 // can be sharply typed as an object array, a type array, or an instance. 4459 // 4460 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4461 PhiNode* result_val; 4462 4463 // Set the reexecute bit for the interpreter to reexecute 4464 // the bytecode that invokes Object.clone if deoptimization happens. 4465 { PreserveReexecuteState preexecs(this); 4466 jvms()->set_should_reexecute(true); 4467 4468 Node* obj = null_check_receiver(); 4469 if (stopped()) return true; 4470 4471 Node* obj_klass = load_object_klass(obj); 4472 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4473 const TypeOopPtr* toop = ((tklass != NULL) 4474 ? tklass->as_instance_type() 4475 : TypeInstPtr::NOTNULL); 4476 4477 // Conservatively insert a memory barrier on all memory slices. 4478 // Do not let writes into the original float below the clone. 4479 insert_mem_bar(Op_MemBarCPUOrder); 4480 4481 // paths into result_reg: 4482 enum { 4483 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4484 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4485 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4486 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4487 PATH_LIMIT 4488 }; 4489 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4490 result_val = new(C) PhiNode(result_reg, 4491 TypeInstPtr::NOTNULL); 4492 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO); 4493 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 4494 TypePtr::BOTTOM); 4495 record_for_igvn(result_reg); 4496 4497 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4498 int raw_adr_idx = Compile::AliasIdxRaw; 4499 4500 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4501 if (array_ctl != NULL) { 4502 // It's an array. 4503 PreserveJVMState pjvms(this); 4504 set_control(array_ctl); 4505 Node* obj_length = load_array_length(obj); 4506 Node* obj_size = NULL; 4507 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4508 4509 if (!use_ReduceInitialCardMarks()) { 4510 // If it is an oop array, it requires very special treatment, 4511 // because card marking is required on each card of the array. 4512 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4513 if (is_obja != NULL) { 4514 PreserveJVMState pjvms2(this); 4515 set_control(is_obja); 4516 // Generate a direct call to the right arraycopy function(s). 4517 bool disjoint_bases = true; 4518 bool length_never_negative = true; 4519 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4520 obj, intcon(0), alloc_obj, intcon(0), 4521 obj_length, 4522 disjoint_bases, length_never_negative); 4523 result_reg->init_req(_objArray_path, control()); 4524 result_val->init_req(_objArray_path, alloc_obj); 4525 result_i_o ->set_req(_objArray_path, i_o()); 4526 result_mem ->set_req(_objArray_path, reset_memory()); 4527 } 4528 } 4529 // Otherwise, there are no card marks to worry about. 4530 // (We can dispense with card marks if we know the allocation 4531 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4532 // causes the non-eden paths to take compensating steps to 4533 // simulate a fresh allocation, so that no further 4534 // card marks are required in compiled code to initialize 4535 // the object.) 4536 4537 if (!stopped()) { 4538 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4539 4540 // Present the results of the copy. 4541 result_reg->init_req(_array_path, control()); 4542 result_val->init_req(_array_path, alloc_obj); 4543 result_i_o ->set_req(_array_path, i_o()); 4544 result_mem ->set_req(_array_path, reset_memory()); 4545 } 4546 } 4547 4548 // We only go to the instance fast case code if we pass a number of guards. 4549 // The paths which do not pass are accumulated in the slow_region. 4550 RegionNode* slow_region = new (C) RegionNode(1); 4551 record_for_igvn(slow_region); 4552 if (!stopped()) { 4553 // It's an instance (we did array above). Make the slow-path tests. 4554 // If this is a virtual call, we generate a funny guard. We grab 4555 // the vtable entry corresponding to clone() from the target object. 4556 // If the target method which we are calling happens to be the 4557 // Object clone() method, we pass the guard. We do not need this 4558 // guard for non-virtual calls; the caller is known to be the native 4559 // Object clone(). 4560 if (is_virtual) { 4561 generate_virtual_guard(obj_klass, slow_region); 4562 } 4563 4564 // The object must be cloneable and must not have a finalizer. 4565 // Both of these conditions may be checked in a single test. 4566 // We could optimize the cloneable test further, but we don't care. 4567 generate_access_flags_guard(obj_klass, 4568 // Test both conditions: 4569 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4570 // Must be cloneable but not finalizer: 4571 JVM_ACC_IS_CLONEABLE, 4572 slow_region); 4573 } 4574 4575 if (!stopped()) { 4576 // It's an instance, and it passed the slow-path tests. 4577 PreserveJVMState pjvms(this); 4578 Node* obj_size = NULL; 4579 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size); 4580 4581 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4582 4583 // Present the results of the slow call. 4584 result_reg->init_req(_instance_path, control()); 4585 result_val->init_req(_instance_path, alloc_obj); 4586 result_i_o ->set_req(_instance_path, i_o()); 4587 result_mem ->set_req(_instance_path, reset_memory()); 4588 } 4589 4590 // Generate code for the slow case. We make a call to clone(). 4591 set_control(_gvn.transform(slow_region)); 4592 if (!stopped()) { 4593 PreserveJVMState pjvms(this); 4594 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4595 Node* slow_result = set_results_for_java_call(slow_call); 4596 // this->control() comes from set_results_for_java_call 4597 result_reg->init_req(_slow_path, control()); 4598 result_val->init_req(_slow_path, slow_result); 4599 result_i_o ->set_req(_slow_path, i_o()); 4600 result_mem ->set_req(_slow_path, reset_memory()); 4601 } 4602 4603 // Return the combined state. 4604 set_control( _gvn.transform(result_reg)); 4605 set_i_o( _gvn.transform(result_i_o)); 4606 set_all_memory( _gvn.transform(result_mem)); 4607 } // original reexecute is set back here 4608 4609 set_result(_gvn.transform(result_val)); 4610 return true; 4611 } 4612 4613 //------------------------------basictype2arraycopy---------------------------- 4614 address LibraryCallKit::basictype2arraycopy(BasicType t, 4615 Node* src_offset, 4616 Node* dest_offset, 4617 bool disjoint_bases, 4618 const char* &name, 4619 bool dest_uninitialized) { 4620 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4621 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4622 4623 bool aligned = false; 4624 bool disjoint = disjoint_bases; 4625 4626 // if the offsets are the same, we can treat the memory regions as 4627 // disjoint, because either the memory regions are in different arrays, 4628 // or they are identical (which we can treat as disjoint.) We can also 4629 // treat a copy with a destination index less that the source index 4630 // as disjoint since a low->high copy will work correctly in this case. 4631 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4632 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4633 // both indices are constants 4634 int s_offs = src_offset_inttype->get_con(); 4635 int d_offs = dest_offset_inttype->get_con(); 4636 int element_size = type2aelembytes(t); 4637 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4638 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4639 if (s_offs >= d_offs) disjoint = true; 4640 } else if (src_offset == dest_offset && src_offset != NULL) { 4641 // This can occur if the offsets are identical non-constants. 4642 disjoint = true; 4643 } 4644 4645 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4646 } 4647 4648 4649 //------------------------------inline_arraycopy----------------------- 4650 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4651 // Object dest, int destPos, 4652 // int length); 4653 bool LibraryCallKit::inline_arraycopy() { 4654 // Get the arguments. 4655 Node* src = argument(0); // type: oop 4656 Node* src_offset = argument(1); // type: int 4657 Node* dest = argument(2); // type: oop 4658 Node* dest_offset = argument(3); // type: int 4659 Node* length = argument(4); // type: int 4660 4661 // Compile time checks. If any of these checks cannot be verified at compile time, 4662 // we do not make a fast path for this call. Instead, we let the call remain as it 4663 // is. The checks we choose to mandate at compile time are: 4664 // 4665 // (1) src and dest are arrays. 4666 const Type* src_type = src->Value(&_gvn); 4667 const Type* dest_type = dest->Value(&_gvn); 4668 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4669 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4670 4671 // Do we have the type of src? 4672 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4673 // Do we have the type of dest? 4674 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4675 // Is the type for src from speculation? 4676 bool src_spec = false; 4677 // Is the type for dest from speculation? 4678 bool dest_spec = false; 4679 4680 if (!has_src || !has_dest) { 4681 // We don't have sufficient type information, let's see if 4682 // speculative types can help. We need to have types for both src 4683 // and dest so that it pays off. 4684 4685 // Do we already have or could we have type information for src 4686 bool could_have_src = has_src; 4687 // Do we already have or could we have type information for dest 4688 bool could_have_dest = has_dest; 4689 4690 ciKlass* src_k = NULL; 4691 if (!has_src) { 4692 src_k = src_type->speculative_type(); 4693 if (src_k != NULL && src_k->is_array_klass()) { 4694 could_have_src = true; 4695 } 4696 } 4697 4698 ciKlass* dest_k = NULL; 4699 if (!has_dest) { 4700 dest_k = dest_type->speculative_type(); 4701 if (dest_k != NULL && dest_k->is_array_klass()) { 4702 could_have_dest = true; 4703 } 4704 } 4705 4706 if (could_have_src && could_have_dest) { 4707 // This is going to pay off so emit the required guards 4708 if (!has_src) { 4709 src = maybe_cast_profiled_obj(src, src_k); 4710 src_type = _gvn.type(src); 4711 top_src = src_type->isa_aryptr(); 4712 has_src = (top_src != NULL && top_src->klass() != NULL); 4713 src_spec = true; 4714 } 4715 if (!has_dest) { 4716 dest = maybe_cast_profiled_obj(dest, dest_k); 4717 dest_type = _gvn.type(dest); 4718 top_dest = dest_type->isa_aryptr(); 4719 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4720 dest_spec = true; 4721 } 4722 } 4723 } 4724 4725 if (!has_src || !has_dest) { 4726 // Conservatively insert a memory barrier on all memory slices. 4727 // Do not let writes into the source float below the arraycopy. 4728 insert_mem_bar(Op_MemBarCPUOrder); 4729 4730 // Call StubRoutines::generic_arraycopy stub. 4731 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4732 src, src_offset, dest, dest_offset, length); 4733 4734 // Do not let reads from the destination float above the arraycopy. 4735 // Since we cannot type the arrays, we don't know which slices 4736 // might be affected. We could restrict this barrier only to those 4737 // memory slices which pertain to array elements--but don't bother. 4738 if (!InsertMemBarAfterArraycopy) 4739 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4740 insert_mem_bar(Op_MemBarCPUOrder); 4741 return true; 4742 } 4743 4744 // (2) src and dest arrays must have elements of the same BasicType 4745 // Figure out the size and type of the elements we will be copying. 4746 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4747 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4748 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4749 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4750 4751 if (src_elem != dest_elem || dest_elem == T_VOID) { 4752 // The component types are not the same or are not recognized. Punt. 4753 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4754 generate_slow_arraycopy(TypePtr::BOTTOM, 4755 src, src_offset, dest, dest_offset, length, 4756 /*dest_uninitialized*/false); 4757 return true; 4758 } 4759 4760 if (src_elem == T_OBJECT) { 4761 // If both arrays are object arrays then having the exact types 4762 // for both will remove the need for a subtype check at runtime 4763 // before the call and may make it possible to pick a faster copy 4764 // routine (without a subtype check on every element) 4765 // Do we have the exact type of src? 4766 bool could_have_src = src_spec; 4767 // Do we have the exact type of dest? 4768 bool could_have_dest = dest_spec; 4769 ciKlass* src_k = top_src->klass(); 4770 ciKlass* dest_k = top_dest->klass(); 4771 if (!src_spec) { 4772 src_k = src_type->speculative_type(); 4773 if (src_k != NULL && src_k->is_array_klass()) { 4774 could_have_src = true; 4775 } 4776 } 4777 if (!dest_spec) { 4778 dest_k = dest_type->speculative_type(); 4779 if (dest_k != NULL && dest_k->is_array_klass()) { 4780 could_have_dest = true; 4781 } 4782 } 4783 if (could_have_src && could_have_dest) { 4784 // If we can have both exact types, emit the missing guards 4785 if (could_have_src && !src_spec) { 4786 src = maybe_cast_profiled_obj(src, src_k); 4787 } 4788 if (could_have_dest && !dest_spec) { 4789 dest = maybe_cast_profiled_obj(dest, dest_k); 4790 } 4791 } 4792 } 4793 4794 //--------------------------------------------------------------------------- 4795 // We will make a fast path for this call to arraycopy. 4796 4797 // We have the following tests left to perform: 4798 // 4799 // (3) src and dest must not be null. 4800 // (4) src_offset must not be negative. 4801 // (5) dest_offset must not be negative. 4802 // (6) length must not be negative. 4803 // (7) src_offset + length must not exceed length of src. 4804 // (8) dest_offset + length must not exceed length of dest. 4805 // (9) each element of an oop array must be assignable 4806 4807 RegionNode* slow_region = new (C) RegionNode(1); 4808 record_for_igvn(slow_region); 4809 4810 // (3) operands must not be null 4811 // We currently perform our null checks with the null_check routine. 4812 // This means that the null exceptions will be reported in the caller 4813 // rather than (correctly) reported inside of the native arraycopy call. 4814 // This should be corrected, given time. We do our null check with the 4815 // stack pointer restored. 4816 src = null_check(src, T_ARRAY); 4817 dest = null_check(dest, T_ARRAY); 4818 4819 // (4) src_offset must not be negative. 4820 generate_negative_guard(src_offset, slow_region); 4821 4822 // (5) dest_offset must not be negative. 4823 generate_negative_guard(dest_offset, slow_region); 4824 4825 // (6) length must not be negative (moved to generate_arraycopy()). 4826 // generate_negative_guard(length, slow_region); 4827 4828 // (7) src_offset + length must not exceed length of src. 4829 generate_limit_guard(src_offset, length, 4830 load_array_length(src), 4831 slow_region); 4832 4833 // (8) dest_offset + length must not exceed length of dest. 4834 generate_limit_guard(dest_offset, length, 4835 load_array_length(dest), 4836 slow_region); 4837 4838 // (9) each element of an oop array must be assignable 4839 // The generate_arraycopy subroutine checks this. 4840 4841 // This is where the memory effects are placed: 4842 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4843 generate_arraycopy(adr_type, dest_elem, 4844 src, src_offset, dest, dest_offset, length, 4845 false, false, slow_region); 4846 4847 return true; 4848 } 4849 4850 //-----------------------------generate_arraycopy---------------------- 4851 // Generate an optimized call to arraycopy. 4852 // Caller must guard against non-arrays. 4853 // Caller must determine a common array basic-type for both arrays. 4854 // Caller must validate offsets against array bounds. 4855 // The slow_region has already collected guard failure paths 4856 // (such as out of bounds length or non-conformable array types). 4857 // The generated code has this shape, in general: 4858 // 4859 // if (length == 0) return // via zero_path 4860 // slowval = -1 4861 // if (types unknown) { 4862 // slowval = call generic copy loop 4863 // if (slowval == 0) return // via checked_path 4864 // } else if (indexes in bounds) { 4865 // if ((is object array) && !(array type check)) { 4866 // slowval = call checked copy loop 4867 // if (slowval == 0) return // via checked_path 4868 // } else { 4869 // call bulk copy loop 4870 // return // via fast_path 4871 // } 4872 // } 4873 // // adjust params for remaining work: 4874 // if (slowval != -1) { 4875 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4876 // } 4877 // slow_region: 4878 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4879 // return // via slow_call_path 4880 // 4881 // This routine is used from several intrinsics: System.arraycopy, 4882 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4883 // 4884 void 4885 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4886 BasicType basic_elem_type, 4887 Node* src, Node* src_offset, 4888 Node* dest, Node* dest_offset, 4889 Node* copy_length, 4890 bool disjoint_bases, 4891 bool length_never_negative, 4892 RegionNode* slow_region) { 4893 4894 if (slow_region == NULL) { 4895 slow_region = new(C) RegionNode(1); 4896 record_for_igvn(slow_region); 4897 } 4898 4899 Node* original_dest = dest; 4900 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4901 bool dest_uninitialized = false; 4902 4903 // See if this is the initialization of a newly-allocated array. 4904 // If so, we will take responsibility here for initializing it to zero. 4905 // (Note: Because tightly_coupled_allocation performs checks on the 4906 // out-edges of the dest, we need to avoid making derived pointers 4907 // from it until we have checked its uses.) 4908 if (ReduceBulkZeroing 4909 && !ZeroTLAB // pointless if already zeroed 4910 && basic_elem_type != T_CONFLICT // avoid corner case 4911 && !src->eqv_uncast(dest) 4912 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4913 != NULL) 4914 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4915 && alloc->maybe_set_complete(&_gvn)) { 4916 // "You break it, you buy it." 4917 InitializeNode* init = alloc->initialization(); 4918 assert(init->is_complete(), "we just did this"); 4919 init->set_complete_with_arraycopy(); 4920 assert(dest->is_CheckCastPP(), "sanity"); 4921 assert(dest->in(0)->in(0) == init, "dest pinned"); 4922 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4923 // From this point on, every exit path is responsible for 4924 // initializing any non-copied parts of the object to zero. 4925 // Also, if this flag is set we make sure that arraycopy interacts properly 4926 // with G1, eliding pre-barriers. See CR 6627983. 4927 dest_uninitialized = true; 4928 } else { 4929 // No zeroing elimination here. 4930 alloc = NULL; 4931 //original_dest = dest; 4932 //dest_uninitialized = false; 4933 } 4934 4935 // Results are placed here: 4936 enum { fast_path = 1, // normal void-returning assembly stub 4937 checked_path = 2, // special assembly stub with cleanup 4938 slow_call_path = 3, // something went wrong; call the VM 4939 zero_path = 4, // bypass when length of copy is zero 4940 bcopy_path = 5, // copy primitive array by 64-bit blocks 4941 PATH_LIMIT = 6 4942 }; 4943 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT); 4944 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO); 4945 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type); 4946 record_for_igvn(result_region); 4947 _gvn.set_type_bottom(result_i_o); 4948 _gvn.set_type_bottom(result_memory); 4949 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4950 4951 // The slow_control path: 4952 Node* slow_control; 4953 Node* slow_i_o = i_o(); 4954 Node* slow_mem = memory(adr_type); 4955 debug_only(slow_control = (Node*) badAddress); 4956 4957 // Checked control path: 4958 Node* checked_control = top(); 4959 Node* checked_mem = NULL; 4960 Node* checked_i_o = NULL; 4961 Node* checked_value = NULL; 4962 4963 if (basic_elem_type == T_CONFLICT) { 4964 assert(!dest_uninitialized, ""); 4965 Node* cv = generate_generic_arraycopy(adr_type, 4966 src, src_offset, dest, dest_offset, 4967 copy_length, dest_uninitialized); 4968 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4969 checked_control = control(); 4970 checked_i_o = i_o(); 4971 checked_mem = memory(adr_type); 4972 checked_value = cv; 4973 set_control(top()); // no fast path 4974 } 4975 4976 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4977 if (not_pos != NULL) { 4978 PreserveJVMState pjvms(this); 4979 set_control(not_pos); 4980 4981 // (6) length must not be negative. 4982 if (!length_never_negative) { 4983 generate_negative_guard(copy_length, slow_region); 4984 } 4985 4986 // copy_length is 0. 4987 if (!stopped() && dest_uninitialized) { 4988 Node* dest_length = alloc->in(AllocateNode::ALength); 4989 if (copy_length->eqv_uncast(dest_length) 4990 || _gvn.find_int_con(dest_length, 1) <= 0) { 4991 // There is no zeroing to do. No need for a secondary raw memory barrier. 4992 } else { 4993 // Clear the whole thing since there are no source elements to copy. 4994 generate_clear_array(adr_type, dest, basic_elem_type, 4995 intcon(0), NULL, 4996 alloc->in(AllocateNode::AllocSize)); 4997 // Use a secondary InitializeNode as raw memory barrier. 4998 // Currently it is needed only on this path since other 4999 // paths have stub or runtime calls as raw memory barriers. 5000 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 5001 Compile::AliasIdxRaw, 5002 top())->as_Initialize(); 5003 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 5004 } 5005 } 5006 5007 // Present the results of the fast call. 5008 result_region->init_req(zero_path, control()); 5009 result_i_o ->init_req(zero_path, i_o()); 5010 result_memory->init_req(zero_path, memory(adr_type)); 5011 } 5012 5013 if (!stopped() && dest_uninitialized) { 5014 // We have to initialize the *uncopied* part of the array to zero. 5015 // The copy destination is the slice dest[off..off+len]. The other slices 5016 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 5017 Node* dest_size = alloc->in(AllocateNode::AllocSize); 5018 Node* dest_length = alloc->in(AllocateNode::ALength); 5019 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset, 5020 copy_length)); 5021 5022 // If there is a head section that needs zeroing, do it now. 5023 if (find_int_con(dest_offset, -1) != 0) { 5024 generate_clear_array(adr_type, dest, basic_elem_type, 5025 intcon(0), dest_offset, 5026 NULL); 5027 } 5028 5029 // Next, perform a dynamic check on the tail length. 5030 // It is often zero, and we can win big if we prove this. 5031 // There are two wins: Avoid generating the ClearArray 5032 // with its attendant messy index arithmetic, and upgrade 5033 // the copy to a more hardware-friendly word size of 64 bits. 5034 Node* tail_ctl = NULL; 5035 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 5036 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length)); 5037 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt)); 5038 tail_ctl = generate_slow_guard(bol_lt, NULL); 5039 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 5040 } 5041 5042 // At this point, let's assume there is no tail. 5043 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 5044 // There is no tail. Try an upgrade to a 64-bit copy. 5045 bool didit = false; 5046 { PreserveJVMState pjvms(this); 5047 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 5048 src, src_offset, dest, dest_offset, 5049 dest_size, dest_uninitialized); 5050 if (didit) { 5051 // Present the results of the block-copying fast call. 5052 result_region->init_req(bcopy_path, control()); 5053 result_i_o ->init_req(bcopy_path, i_o()); 5054 result_memory->init_req(bcopy_path, memory(adr_type)); 5055 } 5056 } 5057 if (didit) 5058 set_control(top()); // no regular fast path 5059 } 5060 5061 // Clear the tail, if any. 5062 if (tail_ctl != NULL) { 5063 Node* notail_ctl = stopped() ? NULL : control(); 5064 set_control(tail_ctl); 5065 if (notail_ctl == NULL) { 5066 generate_clear_array(adr_type, dest, basic_elem_type, 5067 dest_tail, NULL, 5068 dest_size); 5069 } else { 5070 // Make a local merge. 5071 Node* done_ctl = new(C) RegionNode(3); 5072 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type); 5073 done_ctl->init_req(1, notail_ctl); 5074 done_mem->init_req(1, memory(adr_type)); 5075 generate_clear_array(adr_type, dest, basic_elem_type, 5076 dest_tail, NULL, 5077 dest_size); 5078 done_ctl->init_req(2, control()); 5079 done_mem->init_req(2, memory(adr_type)); 5080 set_control( _gvn.transform(done_ctl)); 5081 set_memory( _gvn.transform(done_mem), adr_type ); 5082 } 5083 } 5084 } 5085 5086 BasicType copy_type = basic_elem_type; 5087 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 5088 if (!stopped() && copy_type == T_OBJECT) { 5089 // If src and dest have compatible element types, we can copy bits. 5090 // Types S[] and D[] are compatible if D is a supertype of S. 5091 // 5092 // If they are not, we will use checked_oop_disjoint_arraycopy, 5093 // which performs a fast optimistic per-oop check, and backs off 5094 // further to JVM_ArrayCopy on the first per-oop check that fails. 5095 // (Actually, we don't move raw bits only; the GC requires card marks.) 5096 5097 // Get the Klass* for both src and dest 5098 Node* src_klass = load_object_klass(src); 5099 Node* dest_klass = load_object_klass(dest); 5100 5101 // Generate the subtype check. 5102 // This might fold up statically, or then again it might not. 5103 // 5104 // Non-static example: Copying List<String>.elements to a new String[]. 5105 // The backing store for a List<String> is always an Object[], 5106 // but its elements are always type String, if the generic types 5107 // are correct at the source level. 5108 // 5109 // Test S[] against D[], not S against D, because (probably) 5110 // the secondary supertype cache is less busy for S[] than S. 5111 // This usually only matters when D is an interface. 5112 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5113 // Plug failing path into checked_oop_disjoint_arraycopy 5114 if (not_subtype_ctrl != top()) { 5115 PreserveJVMState pjvms(this); 5116 set_control(not_subtype_ctrl); 5117 // (At this point we can assume disjoint_bases, since types differ.) 5118 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 5119 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 5120 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 5121 Node* dest_elem_klass = _gvn.transform(n1); 5122 Node* cv = generate_checkcast_arraycopy(adr_type, 5123 dest_elem_klass, 5124 src, src_offset, dest, dest_offset, 5125 ConvI2X(copy_length), dest_uninitialized); 5126 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 5127 checked_control = control(); 5128 checked_i_o = i_o(); 5129 checked_mem = memory(adr_type); 5130 checked_value = cv; 5131 } 5132 // At this point we know we do not need type checks on oop stores. 5133 5134 // Let's see if we need card marks: 5135 if (alloc != NULL && use_ReduceInitialCardMarks()) { 5136 // If we do not need card marks, copy using the jint or jlong stub. 5137 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 5138 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 5139 "sizes agree"); 5140 } 5141 } 5142 5143 if (!stopped()) { 5144 // Generate the fast path, if possible. 5145 PreserveJVMState pjvms(this); 5146 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 5147 src, src_offset, dest, dest_offset, 5148 ConvI2X(copy_length), dest_uninitialized); 5149 5150 // Present the results of the fast call. 5151 result_region->init_req(fast_path, control()); 5152 result_i_o ->init_req(fast_path, i_o()); 5153 result_memory->init_req(fast_path, memory(adr_type)); 5154 } 5155 5156 // Here are all the slow paths up to this point, in one bundle: 5157 slow_control = top(); 5158 if (slow_region != NULL) 5159 slow_control = _gvn.transform(slow_region); 5160 DEBUG_ONLY(slow_region = (RegionNode*)badAddress); 5161 5162 set_control(checked_control); 5163 if (!stopped()) { 5164 // Clean up after the checked call. 5165 // The returned value is either 0 or -1^K, 5166 // where K = number of partially transferred array elements. 5167 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0))); 5168 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 5169 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 5170 5171 // If it is 0, we are done, so transfer to the end. 5172 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff)); 5173 result_region->init_req(checked_path, checks_done); 5174 result_i_o ->init_req(checked_path, checked_i_o); 5175 result_memory->init_req(checked_path, checked_mem); 5176 5177 // If it is not zero, merge into the slow call. 5178 set_control( _gvn.transform(new(C) IfFalseNode(iff) )); 5179 RegionNode* slow_reg2 = new(C) RegionNode(3); 5180 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO); 5181 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type); 5182 record_for_igvn(slow_reg2); 5183 slow_reg2 ->init_req(1, slow_control); 5184 slow_i_o2 ->init_req(1, slow_i_o); 5185 slow_mem2 ->init_req(1, slow_mem); 5186 slow_reg2 ->init_req(2, control()); 5187 slow_i_o2 ->init_req(2, checked_i_o); 5188 slow_mem2 ->init_req(2, checked_mem); 5189 5190 slow_control = _gvn.transform(slow_reg2); 5191 slow_i_o = _gvn.transform(slow_i_o2); 5192 slow_mem = _gvn.transform(slow_mem2); 5193 5194 if (alloc != NULL) { 5195 // We'll restart from the very beginning, after zeroing the whole thing. 5196 // This can cause double writes, but that's OK since dest is brand new. 5197 // So we ignore the low 31 bits of the value returned from the stub. 5198 } else { 5199 // We must continue the copy exactly where it failed, or else 5200 // another thread might see the wrong number of writes to dest. 5201 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1))); 5202 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT); 5203 slow_offset->init_req(1, intcon(0)); 5204 slow_offset->init_req(2, checked_offset); 5205 slow_offset = _gvn.transform(slow_offset); 5206 5207 // Adjust the arguments by the conditionally incoming offset. 5208 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset)); 5209 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset)); 5210 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset)); 5211 5212 // Tweak the node variables to adjust the code produced below: 5213 src_offset = src_off_plus; 5214 dest_offset = dest_off_plus; 5215 copy_length = length_minus; 5216 } 5217 } 5218 5219 set_control(slow_control); 5220 if (!stopped()) { 5221 // Generate the slow path, if needed. 5222 PreserveJVMState pjvms(this); // replace_in_map may trash the map 5223 5224 set_memory(slow_mem, adr_type); 5225 set_i_o(slow_i_o); 5226 5227 if (dest_uninitialized) { 5228 generate_clear_array(adr_type, dest, basic_elem_type, 5229 intcon(0), NULL, 5230 alloc->in(AllocateNode::AllocSize)); 5231 } 5232 5233 generate_slow_arraycopy(adr_type, 5234 src, src_offset, dest, dest_offset, 5235 copy_length, /*dest_uninitialized*/false); 5236 5237 result_region->init_req(slow_call_path, control()); 5238 result_i_o ->init_req(slow_call_path, i_o()); 5239 result_memory->init_req(slow_call_path, memory(adr_type)); 5240 } 5241 5242 // Remove unused edges. 5243 for (uint i = 1; i < result_region->req(); i++) { 5244 if (result_region->in(i) == NULL) 5245 result_region->init_req(i, top()); 5246 } 5247 5248 // Finished; return the combined state. 5249 set_control( _gvn.transform(result_region)); 5250 set_i_o( _gvn.transform(result_i_o) ); 5251 set_memory( _gvn.transform(result_memory), adr_type ); 5252 5253 // The memory edges above are precise in order to model effects around 5254 // array copies accurately to allow value numbering of field loads around 5255 // arraycopy. Such field loads, both before and after, are common in Java 5256 // collections and similar classes involving header/array data structures. 5257 // 5258 // But with low number of register or when some registers are used or killed 5259 // by arraycopy calls it causes registers spilling on stack. See 6544710. 5260 // The next memory barrier is added to avoid it. If the arraycopy can be 5261 // optimized away (which it can, sometimes) then we can manually remove 5262 // the membar also. 5263 // 5264 // Do not let reads from the cloned object float above the arraycopy. 5265 if (alloc != NULL) { 5266 // Do not let stores that initialize this object be reordered with 5267 // a subsequent store that would make this object accessible by 5268 // other threads. 5269 // Record what AllocateNode this StoreStore protects so that 5270 // escape analysis can go from the MemBarStoreStoreNode to the 5271 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5272 // based on the escape status of the AllocateNode. 5273 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5274 } else if (InsertMemBarAfterArraycopy) 5275 insert_mem_bar(Op_MemBarCPUOrder); 5276 } 5277 5278 5279 // Helper function which determines if an arraycopy immediately follows 5280 // an allocation, with no intervening tests or other escapes for the object. 5281 AllocateArrayNode* 5282 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5283 RegionNode* slow_region) { 5284 if (stopped()) return NULL; // no fast path 5285 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5286 5287 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5288 if (alloc == NULL) return NULL; 5289 5290 Node* rawmem = memory(Compile::AliasIdxRaw); 5291 // Is the allocation's memory state untouched? 5292 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5293 // Bail out if there have been raw-memory effects since the allocation. 5294 // (Example: There might have been a call or safepoint.) 5295 return NULL; 5296 } 5297 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5298 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5299 return NULL; 5300 } 5301 5302 // There must be no unexpected observers of this allocation. 5303 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5304 Node* obs = ptr->fast_out(i); 5305 if (obs != this->map()) { 5306 return NULL; 5307 } 5308 } 5309 5310 // This arraycopy must unconditionally follow the allocation of the ptr. 5311 Node* alloc_ctl = ptr->in(0); 5312 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5313 5314 Node* ctl = control(); 5315 while (ctl != alloc_ctl) { 5316 // There may be guards which feed into the slow_region. 5317 // Any other control flow means that we might not get a chance 5318 // to finish initializing the allocated object. 5319 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5320 IfNode* iff = ctl->in(0)->as_If(); 5321 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5322 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5323 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5324 ctl = iff->in(0); // This test feeds the known slow_region. 5325 continue; 5326 } 5327 // One more try: Various low-level checks bottom out in 5328 // uncommon traps. If the debug-info of the trap omits 5329 // any reference to the allocation, as we've already 5330 // observed, then there can be no objection to the trap. 5331 bool found_trap = false; 5332 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5333 Node* obs = not_ctl->fast_out(j); 5334 if (obs->in(0) == not_ctl && obs->is_Call() && 5335 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5336 found_trap = true; break; 5337 } 5338 } 5339 if (found_trap) { 5340 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5341 continue; 5342 } 5343 } 5344 return NULL; 5345 } 5346 5347 // If we get this far, we have an allocation which immediately 5348 // precedes the arraycopy, and we can take over zeroing the new object. 5349 // The arraycopy will finish the initialization, and provide 5350 // a new control state to which we will anchor the destination pointer. 5351 5352 return alloc; 5353 } 5354 5355 // Helper for initialization of arrays, creating a ClearArray. 5356 // It writes zero bits in [start..end), within the body of an array object. 5357 // The memory effects are all chained onto the 'adr_type' alias category. 5358 // 5359 // Since the object is otherwise uninitialized, we are free 5360 // to put a little "slop" around the edges of the cleared area, 5361 // as long as it does not go back into the array's header, 5362 // or beyond the array end within the heap. 5363 // 5364 // The lower edge can be rounded down to the nearest jint and the 5365 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5366 // 5367 // Arguments: 5368 // adr_type memory slice where writes are generated 5369 // dest oop of the destination array 5370 // basic_elem_type element type of the destination 5371 // slice_idx array index of first element to store 5372 // slice_len number of elements to store (or NULL) 5373 // dest_size total size in bytes of the array object 5374 // 5375 // Exactly one of slice_len or dest_size must be non-NULL. 5376 // If dest_size is non-NULL, zeroing extends to the end of the object. 5377 // If slice_len is non-NULL, the slice_idx value must be a constant. 5378 void 5379 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5380 Node* dest, 5381 BasicType basic_elem_type, 5382 Node* slice_idx, 5383 Node* slice_len, 5384 Node* dest_size) { 5385 // one or the other but not both of slice_len and dest_size: 5386 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5387 if (slice_len == NULL) slice_len = top(); 5388 if (dest_size == NULL) dest_size = top(); 5389 5390 // operate on this memory slice: 5391 Node* mem = memory(adr_type); // memory slice to operate on 5392 5393 // scaling and rounding of indexes: 5394 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5395 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5396 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5397 int bump_bit = (-1 << scale) & BytesPerInt; 5398 5399 // determine constant starts and ends 5400 const intptr_t BIG_NEG = -128; 5401 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5402 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5403 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5404 if (slice_len_con == 0) { 5405 return; // nothing to do here 5406 } 5407 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5408 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5409 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5410 assert(end_con < 0, "not two cons"); 5411 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5412 BytesPerLong); 5413 } 5414 5415 if (start_con >= 0 && end_con >= 0) { 5416 // Constant start and end. Simple. 5417 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5418 start_con, end_con, &_gvn); 5419 } else if (start_con >= 0 && dest_size != top()) { 5420 // Constant start, pre-rounded end after the tail of the array. 5421 Node* end = dest_size; 5422 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5423 start_con, end, &_gvn); 5424 } else if (start_con >= 0 && slice_len != top()) { 5425 // Constant start, non-constant end. End needs rounding up. 5426 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5427 intptr_t end_base = abase + (slice_idx_con << scale); 5428 int end_round = (-1 << scale) & (BytesPerLong - 1); 5429 Node* end = ConvI2X(slice_len); 5430 if (scale != 0) 5431 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) )); 5432 end_base += end_round; 5433 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base))); 5434 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round))); 5435 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5436 start_con, end, &_gvn); 5437 } else if (start_con < 0 && dest_size != top()) { 5438 // Non-constant start, pre-rounded end after the tail of the array. 5439 // This is almost certainly a "round-to-end" operation. 5440 Node* start = slice_idx; 5441 start = ConvI2X(start); 5442 if (scale != 0) 5443 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) )); 5444 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase))); 5445 if ((bump_bit | clear_low) != 0) { 5446 int to_clear = (bump_bit | clear_low); 5447 // Align up mod 8, then store a jint zero unconditionally 5448 // just before the mod-8 boundary. 5449 if (((abase + bump_bit) & ~to_clear) - bump_bit 5450 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5451 bump_bit = 0; 5452 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5453 } else { 5454 // Bump 'start' up to (or past) the next jint boundary: 5455 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit))); 5456 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5457 } 5458 // Round bumped 'start' down to jlong boundary in body of array. 5459 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear))); 5460 if (bump_bit != 0) { 5461 // Store a zero to the immediately preceding jint: 5462 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit))); 5463 Node* p1 = basic_plus_adr(dest, x1); 5464 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered); 5465 mem = _gvn.transform(mem); 5466 } 5467 } 5468 Node* end = dest_size; // pre-rounded 5469 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5470 start, end, &_gvn); 5471 } else { 5472 // Non-constant start, unrounded non-constant end. 5473 // (Nobody zeroes a random midsection of an array using this routine.) 5474 ShouldNotReachHere(); // fix caller 5475 } 5476 5477 // Done. 5478 set_memory(mem, adr_type); 5479 } 5480 5481 5482 bool 5483 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5484 BasicType basic_elem_type, 5485 AllocateNode* alloc, 5486 Node* src, Node* src_offset, 5487 Node* dest, Node* dest_offset, 5488 Node* dest_size, bool dest_uninitialized) { 5489 // See if there is an advantage from block transfer. 5490 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5491 if (scale >= LogBytesPerLong) 5492 return false; // it is already a block transfer 5493 5494 // Look at the alignment of the starting offsets. 5495 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5496 5497 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5498 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5499 if (src_off_con < 0 || dest_off_con < 0) 5500 // At present, we can only understand constants. 5501 return false; 5502 5503 intptr_t src_off = abase + (src_off_con << scale); 5504 intptr_t dest_off = abase + (dest_off_con << scale); 5505 5506 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5507 // Non-aligned; too bad. 5508 // One more chance: Pick off an initial 32-bit word. 5509 // This is a common case, since abase can be odd mod 8. 5510 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5511 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5512 Node* sptr = basic_plus_adr(src, src_off); 5513 Node* dptr = basic_plus_adr(dest, dest_off); 5514 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 5515 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered); 5516 src_off += BytesPerInt; 5517 dest_off += BytesPerInt; 5518 } else { 5519 return false; 5520 } 5521 } 5522 assert(src_off % BytesPerLong == 0, ""); 5523 assert(dest_off % BytesPerLong == 0, ""); 5524 5525 // Do this copy by giant steps. 5526 Node* sptr = basic_plus_adr(src, src_off); 5527 Node* dptr = basic_plus_adr(dest, dest_off); 5528 Node* countx = dest_size; 5529 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off))); 5530 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong))); 5531 5532 bool disjoint_bases = true; // since alloc != NULL 5533 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5534 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5535 5536 return true; 5537 } 5538 5539 5540 // Helper function; generates code for the slow case. 5541 // We make a call to a runtime method which emulates the native method, 5542 // but without the native wrapper overhead. 5543 void 5544 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5545 Node* src, Node* src_offset, 5546 Node* dest, Node* dest_offset, 5547 Node* copy_length, bool dest_uninitialized) { 5548 assert(!dest_uninitialized, "Invariant"); 5549 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5550 OptoRuntime::slow_arraycopy_Type(), 5551 OptoRuntime::slow_arraycopy_Java(), 5552 "slow_arraycopy", adr_type, 5553 src, src_offset, dest, dest_offset, 5554 copy_length); 5555 5556 // Handle exceptions thrown by this fellow: 5557 make_slow_call_ex(call, env()->Throwable_klass(), false); 5558 } 5559 5560 // Helper function; generates code for cases requiring runtime checks. 5561 Node* 5562 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5563 Node* dest_elem_klass, 5564 Node* src, Node* src_offset, 5565 Node* dest, Node* dest_offset, 5566 Node* copy_length, bool dest_uninitialized) { 5567 if (stopped()) return NULL; 5568 5569 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5570 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5571 return NULL; 5572 } 5573 5574 // Pick out the parameters required to perform a store-check 5575 // for the target array. This is an optimistic check. It will 5576 // look in each non-null element's class, at the desired klass's 5577 // super_check_offset, for the desired klass. 5578 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5579 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5580 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered); 5581 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5582 Node* check_value = dest_elem_klass; 5583 5584 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5585 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5586 5587 // (We know the arrays are never conjoint, because their types differ.) 5588 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5589 OptoRuntime::checkcast_arraycopy_Type(), 5590 copyfunc_addr, "checkcast_arraycopy", adr_type, 5591 // five arguments, of which two are 5592 // intptr_t (jlong in LP64) 5593 src_start, dest_start, 5594 copy_length XTOP, 5595 check_offset XTOP, 5596 check_value); 5597 5598 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5599 } 5600 5601 5602 // Helper function; generates code for cases requiring runtime checks. 5603 Node* 5604 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5605 Node* src, Node* src_offset, 5606 Node* dest, Node* dest_offset, 5607 Node* copy_length, bool dest_uninitialized) { 5608 assert(!dest_uninitialized, "Invariant"); 5609 if (stopped()) return NULL; 5610 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5611 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5612 return NULL; 5613 } 5614 5615 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5616 OptoRuntime::generic_arraycopy_Type(), 5617 copyfunc_addr, "generic_arraycopy", adr_type, 5618 src, src_offset, dest, dest_offset, copy_length); 5619 5620 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5621 } 5622 5623 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5624 void 5625 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5626 BasicType basic_elem_type, 5627 bool disjoint_bases, 5628 Node* src, Node* src_offset, 5629 Node* dest, Node* dest_offset, 5630 Node* copy_length, bool dest_uninitialized) { 5631 if (stopped()) return; // nothing to do 5632 5633 Node* src_start = src; 5634 Node* dest_start = dest; 5635 if (src_offset != NULL || dest_offset != NULL) { 5636 assert(src_offset != NULL && dest_offset != NULL, ""); 5637 src_start = array_element_address(src, src_offset, basic_elem_type); 5638 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5639 } 5640 5641 // Figure out which arraycopy runtime method to call. 5642 const char* copyfunc_name = "arraycopy"; 5643 address copyfunc_addr = 5644 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5645 disjoint_bases, copyfunc_name, dest_uninitialized); 5646 5647 // Call it. Note that the count_ix value is not scaled to a byte-size. 5648 make_runtime_call(RC_LEAF|RC_NO_FP, 5649 OptoRuntime::fast_arraycopy_Type(), 5650 copyfunc_addr, copyfunc_name, adr_type, 5651 src_start, dest_start, copy_length XTOP); 5652 } 5653 5654 //-------------inline_encodeISOArray----------------------------------- 5655 // encode char[] to byte[] in ISO_8859_1 5656 bool LibraryCallKit::inline_encodeISOArray() { 5657 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5658 // no receiver since it is static method 5659 Node *src = argument(0); 5660 Node *src_offset = argument(1); 5661 Node *dst = argument(2); 5662 Node *dst_offset = argument(3); 5663 Node *length = argument(4); 5664 5665 const Type* src_type = src->Value(&_gvn); 5666 const Type* dst_type = dst->Value(&_gvn); 5667 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5668 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5669 if (top_src == NULL || top_src->klass() == NULL || 5670 top_dest == NULL || top_dest->klass() == NULL) { 5671 // failed array check 5672 return false; 5673 } 5674 5675 // Figure out the size and type of the elements we will be copying. 5676 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5677 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5678 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5679 return false; 5680 } 5681 Node* src_start = array_element_address(src, src_offset, src_elem); 5682 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5683 // 'src_start' points to src array + scaled offset 5684 // 'dst_start' points to dst array + scaled offset 5685 5686 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5687 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5688 enc = _gvn.transform(enc); 5689 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc)); 5690 set_memory(res_mem, mtype); 5691 set_result(enc); 5692 return true; 5693 } 5694 5695 /** 5696 * Calculate CRC32 for byte. 5697 * int java.util.zip.CRC32.update(int crc, int b) 5698 */ 5699 bool LibraryCallKit::inline_updateCRC32() { 5700 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5701 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5702 // no receiver since it is static method 5703 Node* crc = argument(0); // type: int 5704 Node* b = argument(1); // type: int 5705 5706 /* 5707 * int c = ~ crc; 5708 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5709 * b = b ^ (c >>> 8); 5710 * crc = ~b; 5711 */ 5712 5713 Node* M1 = intcon(-1); 5714 crc = _gvn.transform(new (C) XorINode(crc, M1)); 5715 Node* result = _gvn.transform(new (C) XorINode(crc, b)); 5716 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF))); 5717 5718 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5719 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2))); 5720 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5721 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5722 5723 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8))); 5724 result = _gvn.transform(new (C) XorINode(crc, result)); 5725 result = _gvn.transform(new (C) XorINode(result, M1)); 5726 set_result(result); 5727 return true; 5728 } 5729 5730 /** 5731 * Calculate CRC32 for byte[] array. 5732 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5733 */ 5734 bool LibraryCallKit::inline_updateBytesCRC32() { 5735 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5736 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5737 // no receiver since it is static method 5738 Node* crc = argument(0); // type: int 5739 Node* src = argument(1); // type: oop 5740 Node* offset = argument(2); // type: int 5741 Node* length = argument(3); // type: int 5742 5743 const Type* src_type = src->Value(&_gvn); 5744 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5745 if (top_src == NULL || top_src->klass() == NULL) { 5746 // failed array check 5747 return false; 5748 } 5749 5750 // Figure out the size and type of the elements we will be copying. 5751 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5752 if (src_elem != T_BYTE) { 5753 return false; 5754 } 5755 5756 // 'src_start' points to src array + scaled offset 5757 Node* src_start = array_element_address(src, offset, src_elem); 5758 5759 // We assume that range check is done by caller. 5760 // TODO: generate range check (offset+length < src.length) in debug VM. 5761 5762 // Call the stub. 5763 address stubAddr = StubRoutines::updateBytesCRC32(); 5764 const char *stubName = "updateBytesCRC32"; 5765 5766 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5767 stubAddr, stubName, TypePtr::BOTTOM, 5768 crc, src_start, length); 5769 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5770 set_result(result); 5771 return true; 5772 } 5773 5774 /** 5775 * Calculate CRC32 for ByteBuffer. 5776 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5777 */ 5778 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5779 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5780 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5781 // no receiver since it is static method 5782 Node* crc = argument(0); // type: int 5783 Node* src = argument(1); // type: long 5784 Node* offset = argument(3); // type: int 5785 Node* length = argument(4); // type: int 5786 5787 src = ConvL2X(src); // adjust Java long to machine word 5788 Node* base = _gvn.transform(new (C) CastX2PNode(src)); 5789 offset = ConvI2X(offset); 5790 5791 // 'src_start' points to src array + scaled offset 5792 Node* src_start = basic_plus_adr(top(), base, offset); 5793 5794 // Call the stub. 5795 address stubAddr = StubRoutines::updateBytesCRC32(); 5796 const char *stubName = "updateBytesCRC32"; 5797 5798 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5799 stubAddr, stubName, TypePtr::BOTTOM, 5800 crc, src_start, length); 5801 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5802 set_result(result); 5803 return true; 5804 } 5805 5806 //----------------------------inline_reference_get---------------------------- 5807 // public T java.lang.ref.Reference.get(); 5808 bool LibraryCallKit::inline_reference_get() { 5809 const int referent_offset = java_lang_ref_Reference::referent_offset; 5810 guarantee(referent_offset > 0, "should have already been set"); 5811 5812 // Get the argument: 5813 Node* reference_obj = null_check_receiver(); 5814 if (stopped()) return true; 5815 5816 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5817 5818 ciInstanceKlass* klass = env()->Object_klass(); 5819 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5820 5821 Node* no_ctrl = NULL; 5822 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5823 5824 // Use the pre-barrier to record the value in the referent field 5825 pre_barrier(false /* do_load */, 5826 control(), 5827 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5828 result /* pre_val */, 5829 T_OBJECT); 5830 5831 // Add memory barrier to prevent commoning reads from this field 5832 // across safepoint since GC can change its value. 5833 insert_mem_bar(Op_MemBarCPUOrder); 5834 5835 set_result(result); 5836 return true; 5837 } 5838 5839 5840 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5841 bool is_exact=true, bool is_static=false) { 5842 5843 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5844 assert(tinst != NULL, "obj is null"); 5845 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5846 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5847 5848 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5849 ciSymbol::make(fieldTypeString), 5850 is_static); 5851 if (field == NULL) return (Node *) NULL; 5852 assert (field != NULL, "undefined field"); 5853 5854 // Next code copied from Parse::do_get_xxx(): 5855 5856 // Compute address and memory type. 5857 int offset = field->offset_in_bytes(); 5858 bool is_vol = field->is_volatile(); 5859 ciType* field_klass = field->type(); 5860 assert(field_klass->is_loaded(), "should be loaded"); 5861 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5862 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5863 BasicType bt = field->layout_type(); 5864 5865 // Build the resultant type of the load 5866 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5867 5868 // Build the load. 5869 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol); 5870 return loadedField; 5871 } 5872 5873 5874 //------------------------------inline_aescrypt_Block----------------------- 5875 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5876 address stubAddr; 5877 const char *stubName; 5878 assert(UseAES, "need AES instruction support"); 5879 5880 switch(id) { 5881 case vmIntrinsics::_aescrypt_encryptBlock: 5882 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5883 stubName = "aescrypt_encryptBlock"; 5884 break; 5885 case vmIntrinsics::_aescrypt_decryptBlock: 5886 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5887 stubName = "aescrypt_decryptBlock"; 5888 break; 5889 } 5890 if (stubAddr == NULL) return false; 5891 5892 Node* aescrypt_object = argument(0); 5893 Node* src = argument(1); 5894 Node* src_offset = argument(2); 5895 Node* dest = argument(3); 5896 Node* dest_offset = argument(4); 5897 5898 // (1) src and dest are arrays. 5899 const Type* src_type = src->Value(&_gvn); 5900 const Type* dest_type = dest->Value(&_gvn); 5901 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5902 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5903 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5904 5905 // for the quick and dirty code we will skip all the checks. 5906 // we are just trying to get the call to be generated. 5907 Node* src_start = src; 5908 Node* dest_start = dest; 5909 if (src_offset != NULL || dest_offset != NULL) { 5910 assert(src_offset != NULL && dest_offset != NULL, ""); 5911 src_start = array_element_address(src, src_offset, T_BYTE); 5912 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5913 } 5914 5915 // now need to get the start of its expanded key array 5916 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5917 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5918 if (k_start == NULL) return false; 5919 5920 if (Matcher::pass_original_key_for_aes()) { 5921 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5922 // compatibility issues between Java key expansion and SPARC crypto instructions 5923 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5924 if (original_k_start == NULL) return false; 5925 5926 // Call the stub. 5927 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5928 stubAddr, stubName, TypePtr::BOTTOM, 5929 src_start, dest_start, k_start, original_k_start); 5930 } else { 5931 // Call the stub. 5932 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5933 stubAddr, stubName, TypePtr::BOTTOM, 5934 src_start, dest_start, k_start); 5935 } 5936 5937 return true; 5938 } 5939 5940 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5941 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5942 address stubAddr; 5943 const char *stubName; 5944 5945 assert(UseAES, "need AES instruction support"); 5946 5947 switch(id) { 5948 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5949 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5950 stubName = "cipherBlockChaining_encryptAESCrypt"; 5951 break; 5952 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5953 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5954 stubName = "cipherBlockChaining_decryptAESCrypt"; 5955 break; 5956 } 5957 if (stubAddr == NULL) return false; 5958 5959 Node* cipherBlockChaining_object = argument(0); 5960 Node* src = argument(1); 5961 Node* src_offset = argument(2); 5962 Node* len = argument(3); 5963 Node* dest = argument(4); 5964 Node* dest_offset = argument(5); 5965 5966 // (1) src and dest are arrays. 5967 const Type* src_type = src->Value(&_gvn); 5968 const Type* dest_type = dest->Value(&_gvn); 5969 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5970 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5971 assert (top_src != NULL && top_src->klass() != NULL 5972 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5973 5974 // checks are the responsibility of the caller 5975 Node* src_start = src; 5976 Node* dest_start = dest; 5977 if (src_offset != NULL || dest_offset != NULL) { 5978 assert(src_offset != NULL && dest_offset != NULL, ""); 5979 src_start = array_element_address(src, src_offset, T_BYTE); 5980 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5981 } 5982 5983 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5984 // (because of the predicated logic executed earlier). 5985 // so we cast it here safely. 5986 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5987 5988 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5989 if (embeddedCipherObj == NULL) return false; 5990 5991 // cast it to what we know it will be at runtime 5992 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5993 assert(tinst != NULL, "CBC obj is null"); 5994 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5995 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5996 if (!klass_AESCrypt->is_loaded()) return false; 5997 5998 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5999 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6000 const TypeOopPtr* xtype = aklass->as_instance_type(); 6001 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype); 6002 aescrypt_object = _gvn.transform(aescrypt_object); 6003 6004 // we need to get the start of the aescrypt_object's expanded key array 6005 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6006 if (k_start == NULL) return false; 6007 6008 // similarly, get the start address of the r vector 6009 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6010 if (objRvec == NULL) return false; 6011 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6012 6013 Node* cbcCrypt; 6014 if (Matcher::pass_original_key_for_aes()) { 6015 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6016 // compatibility issues between Java key expansion and SPARC crypto instructions 6017 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6018 if (original_k_start == NULL) return false; 6019 6020 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6021 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6022 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6023 stubAddr, stubName, TypePtr::BOTTOM, 6024 src_start, dest_start, k_start, r_start, len, original_k_start); 6025 } else { 6026 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6027 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6028 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6029 stubAddr, stubName, TypePtr::BOTTOM, 6030 src_start, dest_start, k_start, r_start, len); 6031 } 6032 6033 // return cipher length (int) 6034 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms)); 6035 set_result(retvalue); 6036 return true; 6037 } 6038 6039 //------------------------------get_key_start_from_aescrypt_object----------------------- 6040 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6041 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6042 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6043 if (objAESCryptKey == NULL) return (Node *) NULL; 6044 6045 // now have the array, need to get the start address of the K array 6046 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6047 return k_start; 6048 } 6049 6050 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6051 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6052 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6053 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6054 if (objAESCryptKey == NULL) return (Node *) NULL; 6055 6056 // now have the array, need to get the start address of the lastKey array 6057 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6058 return original_k_start; 6059 } 6060 6061 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6062 // Return node representing slow path of predicate check. 6063 // the pseudo code we want to emulate with this predicate is: 6064 // for encryption: 6065 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6066 // for decryption: 6067 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6068 // note cipher==plain is more conservative than the original java code but that's OK 6069 // 6070 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6071 // First, check receiver for NULL since it is virtual method. 6072 Node* objCBC = argument(0); 6073 objCBC = null_check(objCBC); 6074 6075 if (stopped()) return NULL; // Always NULL 6076 6077 // Load embeddedCipher field of CipherBlockChaining object. 6078 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6079 6080 // get AESCrypt klass for instanceOf check 6081 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6082 // will have same classloader as CipherBlockChaining object 6083 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6084 assert(tinst != NULL, "CBCobj is null"); 6085 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6086 6087 // we want to do an instanceof comparison against the AESCrypt class 6088 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6089 if (!klass_AESCrypt->is_loaded()) { 6090 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6091 Node* ctrl = control(); 6092 set_control(top()); // no regular fast path 6093 return ctrl; 6094 } 6095 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6096 6097 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6098 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1))); 6099 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); 6100 6101 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6102 6103 // for encryption, we are done 6104 if (!decrypting) 6105 return instof_false; // even if it is NULL 6106 6107 // for decryption, we need to add a further check to avoid 6108 // taking the intrinsic path when cipher and plain are the same 6109 // see the original java code for why. 6110 RegionNode* region = new(C) RegionNode(3); 6111 region->init_req(1, instof_false); 6112 Node* src = argument(1); 6113 Node* dest = argument(4); 6114 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest)); 6115 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq)); 6116 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6117 region->init_req(2, src_dest_conjoint); 6118 6119 record_for_igvn(region); 6120 return _gvn.transform(region); 6121 }