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 template <typename OverflowOp> 209 bool inline_math_overflow(bool is_unary); 210 void inline_math_mathExact(Node* math, Node* test); 211 bool inline_math_addExactI(bool is_increment); 212 bool inline_math_addExactL(bool is_increment); 213 bool inline_math_multiplyExactI(); 214 bool inline_math_multiplyExactL(); 215 bool inline_math_negateExactI(); 216 bool inline_math_negateExactL(); 217 bool inline_math_subtractExactI(bool is_decrement); 218 bool inline_math_subtractExactL(bool is_decrement); 219 bool inline_exp(); 220 bool inline_pow(); 221 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 222 bool inline_min_max(vmIntrinsics::ID id); 223 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 224 // This returns Type::AnyPtr, RawPtr, or OopPtr. 225 int classify_unsafe_addr(Node* &base, Node* &offset); 226 Node* make_unsafe_address(Node* base, Node* offset); 227 // Helper for inline_unsafe_access. 228 // Generates the guards that check whether the result of 229 // Unsafe.getObject should be recorded in an SATB log buffer. 230 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 231 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 232 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 233 static bool klass_needs_init_guard(Node* kls); 234 bool inline_unsafe_allocate(); 235 bool inline_unsafe_copyMemory(); 236 bool inline_native_currentThread(); 237 #ifdef TRACE_HAVE_INTRINSICS 238 bool inline_native_classID(); 239 bool inline_native_threadID(); 240 #endif 241 bool inline_native_time_funcs(address method, const char* funcName); 242 bool inline_native_isInterrupted(); 243 bool inline_native_Class_query(vmIntrinsics::ID id); 244 bool inline_native_subtype_check(); 245 246 bool inline_native_newArray(); 247 bool inline_native_getLength(); 248 bool inline_array_copyOf(bool is_copyOfRange); 249 bool inline_array_equals(); 250 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 251 bool inline_native_clone(bool is_virtual); 252 bool inline_native_Reflection_getCallerClass(); 253 // Helper function for inlining native object hash method 254 bool inline_native_hashcode(bool is_virtual, bool is_static); 255 bool inline_native_getClass(); 256 257 // Helper functions for inlining arraycopy 258 bool inline_arraycopy(); 259 void generate_arraycopy(const TypePtr* adr_type, 260 BasicType basic_elem_type, 261 Node* src, Node* src_offset, 262 Node* dest, Node* dest_offset, 263 Node* copy_length, 264 bool disjoint_bases = false, 265 bool length_never_negative = false, 266 RegionNode* slow_region = NULL); 267 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 268 RegionNode* slow_region); 269 void generate_clear_array(const TypePtr* adr_type, 270 Node* dest, 271 BasicType basic_elem_type, 272 Node* slice_off, 273 Node* slice_len, 274 Node* slice_end); 275 bool generate_block_arraycopy(const TypePtr* adr_type, 276 BasicType basic_elem_type, 277 AllocateNode* alloc, 278 Node* src, Node* src_offset, 279 Node* dest, Node* dest_offset, 280 Node* dest_size, bool dest_uninitialized); 281 void generate_slow_arraycopy(const TypePtr* adr_type, 282 Node* src, Node* src_offset, 283 Node* dest, Node* dest_offset, 284 Node* copy_length, bool dest_uninitialized); 285 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 286 Node* dest_elem_klass, 287 Node* src, Node* src_offset, 288 Node* dest, Node* dest_offset, 289 Node* copy_length, bool dest_uninitialized); 290 Node* generate_generic_arraycopy(const TypePtr* adr_type, 291 Node* src, Node* src_offset, 292 Node* dest, Node* dest_offset, 293 Node* copy_length, bool dest_uninitialized); 294 void generate_unchecked_arraycopy(const TypePtr* adr_type, 295 BasicType basic_elem_type, 296 bool disjoint_bases, 297 Node* src, Node* src_offset, 298 Node* dest, Node* dest_offset, 299 Node* copy_length, bool dest_uninitialized); 300 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 301 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 302 bool inline_unsafe_ordered_store(BasicType type); 303 bool inline_unsafe_fence(vmIntrinsics::ID id); 304 bool inline_fp_conversions(vmIntrinsics::ID id); 305 bool inline_number_methods(vmIntrinsics::ID id); 306 bool inline_reference_get(); 307 bool inline_aescrypt_Block(vmIntrinsics::ID id); 308 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 309 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 310 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 311 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 312 bool inline_encodeISOArray(); 313 bool inline_updateCRC32(); 314 bool inline_updateBytesCRC32(); 315 bool inline_updateByteBufferCRC32(); 316 }; 317 318 319 //---------------------------make_vm_intrinsic---------------------------- 320 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 321 vmIntrinsics::ID id = m->intrinsic_id(); 322 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 323 324 if (DisableIntrinsic[0] != '\0' 325 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 326 // disabled by a user request on the command line: 327 // example: -XX:DisableIntrinsic=_hashCode,_getClass 328 return NULL; 329 } 330 331 if (!m->is_loaded()) { 332 // do not attempt to inline unloaded methods 333 return NULL; 334 } 335 336 // Only a few intrinsics implement a virtual dispatch. 337 // They are expensive calls which are also frequently overridden. 338 if (is_virtual) { 339 switch (id) { 340 case vmIntrinsics::_hashCode: 341 case vmIntrinsics::_clone: 342 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 343 break; 344 default: 345 return NULL; 346 } 347 } 348 349 // -XX:-InlineNatives disables nearly all intrinsics: 350 if (!InlineNatives) { 351 switch (id) { 352 case vmIntrinsics::_indexOf: 353 case vmIntrinsics::_compareTo: 354 case vmIntrinsics::_equals: 355 case vmIntrinsics::_equalsC: 356 case vmIntrinsics::_getAndAddInt: 357 case vmIntrinsics::_getAndAddLong: 358 case vmIntrinsics::_getAndSetInt: 359 case vmIntrinsics::_getAndSetLong: 360 case vmIntrinsics::_getAndSetObject: 361 case vmIntrinsics::_loadFence: 362 case vmIntrinsics::_storeFence: 363 case vmIntrinsics::_fullFence: 364 break; // InlineNatives does not control String.compareTo 365 case vmIntrinsics::_Reference_get: 366 break; // InlineNatives does not control Reference.get 367 default: 368 return NULL; 369 } 370 } 371 372 bool is_predicted = false; 373 bool does_virtual_dispatch = false; 374 375 switch (id) { 376 case vmIntrinsics::_compareTo: 377 if (!SpecialStringCompareTo) return NULL; 378 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 379 break; 380 case vmIntrinsics::_indexOf: 381 if (!SpecialStringIndexOf) return NULL; 382 break; 383 case vmIntrinsics::_equals: 384 if (!SpecialStringEquals) return NULL; 385 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 386 break; 387 case vmIntrinsics::_equalsC: 388 if (!SpecialArraysEquals) return NULL; 389 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 390 break; 391 case vmIntrinsics::_arraycopy: 392 if (!InlineArrayCopy) return NULL; 393 break; 394 case vmIntrinsics::_copyMemory: 395 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 396 if (!InlineArrayCopy) return NULL; 397 break; 398 case vmIntrinsics::_hashCode: 399 if (!InlineObjectHash) return NULL; 400 does_virtual_dispatch = true; 401 break; 402 case vmIntrinsics::_clone: 403 does_virtual_dispatch = true; 404 case vmIntrinsics::_copyOf: 405 case vmIntrinsics::_copyOfRange: 406 if (!InlineObjectCopy) return NULL; 407 // These also use the arraycopy intrinsic mechanism: 408 if (!InlineArrayCopy) return NULL; 409 break; 410 case vmIntrinsics::_encodeISOArray: 411 if (!SpecialEncodeISOArray) return NULL; 412 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 413 break; 414 case vmIntrinsics::_checkIndex: 415 // We do not intrinsify this. The optimizer does fine with it. 416 return NULL; 417 418 case vmIntrinsics::_getCallerClass: 419 if (!UseNewReflection) return NULL; 420 if (!InlineReflectionGetCallerClass) return NULL; 421 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 422 break; 423 424 case vmIntrinsics::_bitCount_i: 425 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 426 break; 427 428 case vmIntrinsics::_bitCount_l: 429 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 430 break; 431 432 case vmIntrinsics::_numberOfLeadingZeros_i: 433 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 434 break; 435 436 case vmIntrinsics::_numberOfLeadingZeros_l: 437 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 438 break; 439 440 case vmIntrinsics::_numberOfTrailingZeros_i: 441 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 442 break; 443 444 case vmIntrinsics::_numberOfTrailingZeros_l: 445 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 446 break; 447 448 case vmIntrinsics::_reverseBytes_c: 449 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 450 break; 451 case vmIntrinsics::_reverseBytes_s: 452 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 453 break; 454 case vmIntrinsics::_reverseBytes_i: 455 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 456 break; 457 case vmIntrinsics::_reverseBytes_l: 458 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 459 break; 460 461 case vmIntrinsics::_Reference_get: 462 // Use the intrinsic version of Reference.get() so that the value in 463 // the referent field can be registered by the G1 pre-barrier code. 464 // Also add memory barrier to prevent commoning reads from this field 465 // across safepoint since GC can change it value. 466 break; 467 468 case vmIntrinsics::_compareAndSwapObject: 469 #ifdef _LP64 470 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 471 #endif 472 break; 473 474 case vmIntrinsics::_compareAndSwapLong: 475 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 476 break; 477 478 case vmIntrinsics::_getAndAddInt: 479 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 480 break; 481 482 case vmIntrinsics::_getAndAddLong: 483 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 484 break; 485 486 case vmIntrinsics::_getAndSetInt: 487 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 488 break; 489 490 case vmIntrinsics::_getAndSetLong: 491 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 492 break; 493 494 case vmIntrinsics::_getAndSetObject: 495 #ifdef _LP64 496 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 497 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 498 break; 499 #else 500 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 501 break; 502 #endif 503 504 case vmIntrinsics::_aescrypt_encryptBlock: 505 case vmIntrinsics::_aescrypt_decryptBlock: 506 if (!UseAESIntrinsics) return NULL; 507 break; 508 509 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 510 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 511 if (!UseAESIntrinsics) return NULL; 512 // these two require the predicated logic 513 is_predicted = true; 514 break; 515 516 case vmIntrinsics::_updateCRC32: 517 case vmIntrinsics::_updateBytesCRC32: 518 case vmIntrinsics::_updateByteBufferCRC32: 519 if (!UseCRC32Intrinsics) return NULL; 520 break; 521 522 case vmIntrinsics::_incrementExactI: 523 case vmIntrinsics::_addExactI: 524 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 525 break; 526 case vmIntrinsics::_incrementExactL: 527 case vmIntrinsics::_addExactL: 528 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 529 break; 530 case vmIntrinsics::_decrementExactI: 531 case vmIntrinsics::_subtractExactI: 532 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 533 break; 534 case vmIntrinsics::_decrementExactL: 535 case vmIntrinsics::_subtractExactL: 536 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 537 break; 538 case vmIntrinsics::_negateExactI: 539 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 540 break; 541 case vmIntrinsics::_negateExactL: 542 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 543 break; 544 case vmIntrinsics::_multiplyExactI: 545 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 546 break; 547 case vmIntrinsics::_multiplyExactL: 548 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 549 break; 550 551 default: 552 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 553 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 554 break; 555 } 556 557 // -XX:-InlineClassNatives disables natives from the Class class. 558 // The flag applies to all reflective calls, notably Array.newArray 559 // (visible to Java programmers as Array.newInstance). 560 if (m->holder()->name() == ciSymbol::java_lang_Class() || 561 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 562 if (!InlineClassNatives) return NULL; 563 } 564 565 // -XX:-InlineThreadNatives disables natives from the Thread class. 566 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 567 if (!InlineThreadNatives) return NULL; 568 } 569 570 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 571 if (m->holder()->name() == ciSymbol::java_lang_Math() || 572 m->holder()->name() == ciSymbol::java_lang_Float() || 573 m->holder()->name() == ciSymbol::java_lang_Double()) { 574 if (!InlineMathNatives) return NULL; 575 } 576 577 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 578 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 579 if (!InlineUnsafeOps) return NULL; 580 } 581 582 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id); 583 } 584 585 //----------------------register_library_intrinsics----------------------- 586 // Initialize this file's data structures, for each Compile instance. 587 void Compile::register_library_intrinsics() { 588 // Nothing to do here. 589 } 590 591 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) { 592 LibraryCallKit kit(jvms, this); 593 Compile* C = kit.C; 594 int nodes = C->unique(); 595 #ifndef PRODUCT 596 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 597 char buf[1000]; 598 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 599 tty->print_cr("Intrinsic %s", str); 600 } 601 #endif 602 ciMethod* callee = kit.callee(); 603 const int bci = kit.bci(); 604 605 // Try to inline the intrinsic. 606 if (kit.try_to_inline()) { 607 if (C->print_intrinsics() || C->print_inlining()) { 608 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 609 } 610 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 611 if (C->log()) { 612 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 613 vmIntrinsics::name_at(intrinsic_id()), 614 (is_virtual() ? " virtual='1'" : ""), 615 C->unique() - nodes); 616 } 617 // Push the result from the inlined method onto the stack. 618 kit.push_result(); 619 return kit.transfer_exceptions_into_jvms(); 620 } 621 622 // The intrinsic bailed out 623 if (C->print_intrinsics() || C->print_inlining()) { 624 if (jvms->has_method()) { 625 // Not a root compile. 626 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 627 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 628 } else { 629 // Root compile 630 tty->print("Did not generate intrinsic %s%s at bci:%d in", 631 vmIntrinsics::name_at(intrinsic_id()), 632 (is_virtual() ? " (virtual)" : ""), bci); 633 } 634 } 635 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 636 return NULL; 637 } 638 639 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) { 640 LibraryCallKit kit(jvms, this); 641 Compile* C = kit.C; 642 int nodes = C->unique(); 643 #ifndef PRODUCT 644 assert(is_predicted(), "sanity"); 645 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 646 char buf[1000]; 647 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 648 tty->print_cr("Predicate for intrinsic %s", str); 649 } 650 #endif 651 ciMethod* callee = kit.callee(); 652 const int bci = kit.bci(); 653 654 Node* slow_ctl = kit.try_to_predicate(); 655 if (!kit.failing()) { 656 if (C->print_intrinsics() || C->print_inlining()) { 657 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 658 } 659 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 660 if (C->log()) { 661 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 662 vmIntrinsics::name_at(intrinsic_id()), 663 (is_virtual() ? " virtual='1'" : ""), 664 C->unique() - nodes); 665 } 666 return slow_ctl; // Could be NULL if the check folds. 667 } 668 669 // The intrinsic bailed out 670 if (C->print_intrinsics() || C->print_inlining()) { 671 if (jvms->has_method()) { 672 // Not a root compile. 673 const char* msg = "failed to generate predicate for intrinsic"; 674 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 675 } else { 676 // Root compile 677 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 678 vmIntrinsics::name_at(intrinsic_id()), 679 (is_virtual() ? " (virtual)" : ""), bci); 680 } 681 } 682 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 683 return NULL; 684 } 685 686 bool LibraryCallKit::try_to_inline() { 687 // Handle symbolic names for otherwise undistinguished boolean switches: 688 const bool is_store = true; 689 const bool is_native_ptr = true; 690 const bool is_static = true; 691 const bool is_volatile = true; 692 693 if (!jvms()->has_method()) { 694 // Root JVMState has a null method. 695 assert(map()->memory()->Opcode() == Op_Parm, ""); 696 // Insert the memory aliasing node 697 set_all_memory(reset_memory()); 698 } 699 assert(merged_memory(), ""); 700 701 702 switch (intrinsic_id()) { 703 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 704 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 705 case vmIntrinsics::_getClass: return inline_native_getClass(); 706 707 case vmIntrinsics::_dsin: 708 case vmIntrinsics::_dcos: 709 case vmIntrinsics::_dtan: 710 case vmIntrinsics::_dabs: 711 case vmIntrinsics::_datan2: 712 case vmIntrinsics::_dsqrt: 713 case vmIntrinsics::_dexp: 714 case vmIntrinsics::_dlog: 715 case vmIntrinsics::_dlog10: 716 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 717 718 case vmIntrinsics::_min: 719 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 720 721 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 722 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 723 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 724 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 725 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 726 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 727 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 728 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 729 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 730 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 731 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 732 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 733 734 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 735 736 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 737 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 738 case vmIntrinsics::_equals: return inline_string_equals(); 739 740 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 741 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 742 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 743 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 744 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 745 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 746 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 747 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 748 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 749 750 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 751 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 752 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 753 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 754 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 755 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 756 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 757 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 758 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 759 760 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 761 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 762 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 763 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 764 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 765 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 766 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 767 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 768 769 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 770 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 771 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 772 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 773 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 774 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 775 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 776 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 777 778 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 779 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 780 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 781 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 782 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 783 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 784 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 785 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 786 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 787 788 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 789 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 790 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 791 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 792 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 793 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 794 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 795 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 796 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 797 798 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 799 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 800 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 801 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 802 803 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 804 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 805 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 806 807 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 808 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 809 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 810 811 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 812 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 813 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 814 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 815 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 816 817 case vmIntrinsics::_loadFence: 818 case vmIntrinsics::_storeFence: 819 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 820 821 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 822 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 823 824 #ifdef TRACE_HAVE_INTRINSICS 825 case vmIntrinsics::_classID: return inline_native_classID(); 826 case vmIntrinsics::_threadID: return inline_native_threadID(); 827 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 828 #endif 829 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 830 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 831 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 832 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 833 case vmIntrinsics::_newArray: return inline_native_newArray(); 834 case vmIntrinsics::_getLength: return inline_native_getLength(); 835 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 836 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 837 case vmIntrinsics::_equalsC: return inline_array_equals(); 838 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 839 840 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 841 842 case vmIntrinsics::_isInstance: 843 case vmIntrinsics::_getModifiers: 844 case vmIntrinsics::_isInterface: 845 case vmIntrinsics::_isArray: 846 case vmIntrinsics::_isPrimitive: 847 case vmIntrinsics::_getSuperclass: 848 case vmIntrinsics::_getComponentType: 849 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 850 851 case vmIntrinsics::_floatToRawIntBits: 852 case vmIntrinsics::_floatToIntBits: 853 case vmIntrinsics::_intBitsToFloat: 854 case vmIntrinsics::_doubleToRawLongBits: 855 case vmIntrinsics::_doubleToLongBits: 856 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 857 858 case vmIntrinsics::_numberOfLeadingZeros_i: 859 case vmIntrinsics::_numberOfLeadingZeros_l: 860 case vmIntrinsics::_numberOfTrailingZeros_i: 861 case vmIntrinsics::_numberOfTrailingZeros_l: 862 case vmIntrinsics::_bitCount_i: 863 case vmIntrinsics::_bitCount_l: 864 case vmIntrinsics::_reverseBytes_i: 865 case vmIntrinsics::_reverseBytes_l: 866 case vmIntrinsics::_reverseBytes_s: 867 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 868 869 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 870 871 case vmIntrinsics::_Reference_get: return inline_reference_get(); 872 873 case vmIntrinsics::_aescrypt_encryptBlock: 874 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 875 876 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 877 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 878 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 879 880 case vmIntrinsics::_encodeISOArray: 881 return inline_encodeISOArray(); 882 883 case vmIntrinsics::_updateCRC32: 884 return inline_updateCRC32(); 885 case vmIntrinsics::_updateBytesCRC32: 886 return inline_updateBytesCRC32(); 887 case vmIntrinsics::_updateByteBufferCRC32: 888 return inline_updateByteBufferCRC32(); 889 890 default: 891 // If you get here, it may be that someone has added a new intrinsic 892 // to the list in vmSymbols.hpp without implementing it here. 893 #ifndef PRODUCT 894 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 895 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 896 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 897 } 898 #endif 899 return false; 900 } 901 } 902 903 Node* LibraryCallKit::try_to_predicate() { 904 if (!jvms()->has_method()) { 905 // Root JVMState has a null method. 906 assert(map()->memory()->Opcode() == Op_Parm, ""); 907 // Insert the memory aliasing node 908 set_all_memory(reset_memory()); 909 } 910 assert(merged_memory(), ""); 911 912 switch (intrinsic_id()) { 913 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 914 return inline_cipherBlockChaining_AESCrypt_predicate(false); 915 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 916 return inline_cipherBlockChaining_AESCrypt_predicate(true); 917 918 default: 919 // If you get here, it may be that someone has added a new intrinsic 920 // to the list in vmSymbols.hpp without implementing it here. 921 #ifndef PRODUCT 922 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 923 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 924 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 925 } 926 #endif 927 Node* slow_ctl = control(); 928 set_control(top()); // No fast path instrinsic 929 return slow_ctl; 930 } 931 } 932 933 //------------------------------set_result------------------------------- 934 // Helper function for finishing intrinsics. 935 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 936 record_for_igvn(region); 937 set_control(_gvn.transform(region)); 938 set_result( _gvn.transform(value)); 939 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 940 } 941 942 //------------------------------generate_guard--------------------------- 943 // Helper function for generating guarded fast-slow graph structures. 944 // The given 'test', if true, guards a slow path. If the test fails 945 // then a fast path can be taken. (We generally hope it fails.) 946 // In all cases, GraphKit::control() is updated to the fast path. 947 // The returned value represents the control for the slow path. 948 // The return value is never 'top'; it is either a valid control 949 // or NULL if it is obvious that the slow path can never be taken. 950 // Also, if region and the slow control are not NULL, the slow edge 951 // is appended to the region. 952 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 953 if (stopped()) { 954 // Already short circuited. 955 return NULL; 956 } 957 958 // Build an if node and its projections. 959 // If test is true we take the slow path, which we assume is uncommon. 960 if (_gvn.type(test) == TypeInt::ZERO) { 961 // The slow branch is never taken. No need to build this guard. 962 return NULL; 963 } 964 965 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 966 967 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff)); 968 if (if_slow == top()) { 969 // The slow branch is never taken. No need to build this guard. 970 return NULL; 971 } 972 973 if (region != NULL) 974 region->add_req(if_slow); 975 976 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff)); 977 set_control(if_fast); 978 979 return if_slow; 980 } 981 982 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 983 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 984 } 985 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 986 return generate_guard(test, region, PROB_FAIR); 987 } 988 989 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 990 Node* *pos_index) { 991 if (stopped()) 992 return NULL; // already stopped 993 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 994 return NULL; // index is already adequately typed 995 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0))); 996 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 997 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 998 if (is_neg != NULL && pos_index != NULL) { 999 // Emulate effect of Parse::adjust_map_after_if. 1000 Node* ccast = new (C) CastIINode(index, TypeInt::POS); 1001 ccast->set_req(0, control()); 1002 (*pos_index) = _gvn.transform(ccast); 1003 } 1004 return is_neg; 1005 } 1006 1007 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 1008 Node* *pos_index) { 1009 if (stopped()) 1010 return NULL; // already stopped 1011 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 1012 return NULL; // index is already adequately typed 1013 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0))); 1014 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 1015 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq)); 1016 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 1017 if (is_notp != NULL && pos_index != NULL) { 1018 // Emulate effect of Parse::adjust_map_after_if. 1019 Node* ccast = new (C) CastIINode(index, TypeInt::POS1); 1020 ccast->set_req(0, control()); 1021 (*pos_index) = _gvn.transform(ccast); 1022 } 1023 return is_notp; 1024 } 1025 1026 // Make sure that 'position' is a valid limit index, in [0..length]. 1027 // There are two equivalent plans for checking this: 1028 // A. (offset + copyLength) unsigned<= arrayLength 1029 // B. offset <= (arrayLength - copyLength) 1030 // We require that all of the values above, except for the sum and 1031 // difference, are already known to be non-negative. 1032 // Plan A is robust in the face of overflow, if offset and copyLength 1033 // are both hugely positive. 1034 // 1035 // Plan B is less direct and intuitive, but it does not overflow at 1036 // all, since the difference of two non-negatives is always 1037 // representable. Whenever Java methods must perform the equivalent 1038 // check they generally use Plan B instead of Plan A. 1039 // For the moment we use Plan A. 1040 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1041 Node* subseq_length, 1042 Node* array_length, 1043 RegionNode* region) { 1044 if (stopped()) 1045 return NULL; // already stopped 1046 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1047 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1048 return NULL; // common case of whole-array copy 1049 Node* last = subseq_length; 1050 if (!zero_offset) // last += offset 1051 last = _gvn.transform(new (C) AddINode(last, offset)); 1052 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last)); 1053 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 1054 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1055 return is_over; 1056 } 1057 1058 1059 //--------------------------generate_current_thread-------------------- 1060 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1061 ciKlass* thread_klass = env()->Thread_klass(); 1062 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1063 Node* thread = _gvn.transform(new (C) ThreadLocalNode()); 1064 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1065 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1066 tls_output = thread; 1067 return threadObj; 1068 } 1069 1070 1071 //------------------------------make_string_method_node------------------------ 1072 // Helper method for String intrinsic functions. This version is called 1073 // with str1 and str2 pointing to String object nodes. 1074 // 1075 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1076 Node* no_ctrl = NULL; 1077 1078 // Get start addr of string 1079 Node* str1_value = load_String_value(no_ctrl, str1); 1080 Node* str1_offset = load_String_offset(no_ctrl, str1); 1081 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1082 1083 // Get length of string 1 1084 Node* str1_len = load_String_length(no_ctrl, str1); 1085 1086 Node* str2_value = load_String_value(no_ctrl, str2); 1087 Node* str2_offset = load_String_offset(no_ctrl, str2); 1088 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1089 1090 Node* str2_len = NULL; 1091 Node* result = NULL; 1092 1093 switch (opcode) { 1094 case Op_StrIndexOf: 1095 // Get length of string 2 1096 str2_len = load_String_length(no_ctrl, str2); 1097 1098 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1099 str1_start, str1_len, str2_start, str2_len); 1100 break; 1101 case Op_StrComp: 1102 // Get length of string 2 1103 str2_len = load_String_length(no_ctrl, str2); 1104 1105 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1106 str1_start, str1_len, str2_start, str2_len); 1107 break; 1108 case Op_StrEquals: 1109 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1110 str1_start, str2_start, str1_len); 1111 break; 1112 default: 1113 ShouldNotReachHere(); 1114 return NULL; 1115 } 1116 1117 // All these intrinsics have checks. 1118 C->set_has_split_ifs(true); // Has chance for split-if optimization 1119 1120 return _gvn.transform(result); 1121 } 1122 1123 // Helper method for String intrinsic functions. This version is called 1124 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1125 // to Int nodes containing the lenghts of str1 and str2. 1126 // 1127 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1128 Node* result = NULL; 1129 switch (opcode) { 1130 case Op_StrIndexOf: 1131 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1132 str1_start, cnt1, str2_start, cnt2); 1133 break; 1134 case Op_StrComp: 1135 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1136 str1_start, cnt1, str2_start, cnt2); 1137 break; 1138 case Op_StrEquals: 1139 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1140 str1_start, str2_start, cnt1); 1141 break; 1142 default: 1143 ShouldNotReachHere(); 1144 return NULL; 1145 } 1146 1147 // All these intrinsics have checks. 1148 C->set_has_split_ifs(true); // Has chance for split-if optimization 1149 1150 return _gvn.transform(result); 1151 } 1152 1153 //------------------------------inline_string_compareTo------------------------ 1154 // public int java.lang.String.compareTo(String anotherString); 1155 bool LibraryCallKit::inline_string_compareTo() { 1156 Node* receiver = null_check(argument(0)); 1157 Node* arg = null_check(argument(1)); 1158 if (stopped()) { 1159 return true; 1160 } 1161 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1162 return true; 1163 } 1164 1165 //------------------------------inline_string_equals------------------------ 1166 bool LibraryCallKit::inline_string_equals() { 1167 Node* receiver = null_check_receiver(); 1168 // NOTE: Do not null check argument for String.equals() because spec 1169 // allows to specify NULL as argument. 1170 Node* argument = this->argument(1); 1171 if (stopped()) { 1172 return true; 1173 } 1174 1175 // paths (plus control) merge 1176 RegionNode* region = new (C) RegionNode(5); 1177 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 1178 1179 // does source == target string? 1180 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument)); 1181 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 1182 1183 Node* if_eq = generate_slow_guard(bol, NULL); 1184 if (if_eq != NULL) { 1185 // receiver == argument 1186 phi->init_req(2, intcon(1)); 1187 region->init_req(2, if_eq); 1188 } 1189 1190 // get String klass for instanceOf 1191 ciInstanceKlass* klass = env()->String_klass(); 1192 1193 if (!stopped()) { 1194 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1195 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1))); 1196 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 1197 1198 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1199 //instanceOf == true, fallthrough 1200 1201 if (inst_false != NULL) { 1202 phi->init_req(3, intcon(0)); 1203 region->init_req(3, inst_false); 1204 } 1205 } 1206 1207 if (!stopped()) { 1208 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1209 1210 // Properly cast the argument to String 1211 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type)); 1212 // This path is taken only when argument's type is String:NotNull. 1213 argument = cast_not_null(argument, false); 1214 1215 Node* no_ctrl = NULL; 1216 1217 // Get start addr of receiver 1218 Node* receiver_val = load_String_value(no_ctrl, receiver); 1219 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1220 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1221 1222 // Get length of receiver 1223 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1224 1225 // Get start addr of argument 1226 Node* argument_val = load_String_value(no_ctrl, argument); 1227 Node* argument_offset = load_String_offset(no_ctrl, argument); 1228 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1229 1230 // Get length of argument 1231 Node* argument_cnt = load_String_length(no_ctrl, argument); 1232 1233 // Check for receiver count != argument count 1234 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt)); 1235 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne)); 1236 Node* if_ne = generate_slow_guard(bol, NULL); 1237 if (if_ne != NULL) { 1238 phi->init_req(4, intcon(0)); 1239 region->init_req(4, if_ne); 1240 } 1241 1242 // Check for count == 0 is done by assembler code for StrEquals. 1243 1244 if (!stopped()) { 1245 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1246 phi->init_req(1, equals); 1247 region->init_req(1, control()); 1248 } 1249 } 1250 1251 // post merge 1252 set_control(_gvn.transform(region)); 1253 record_for_igvn(region); 1254 1255 set_result(_gvn.transform(phi)); 1256 return true; 1257 } 1258 1259 //------------------------------inline_array_equals---------------------------- 1260 bool LibraryCallKit::inline_array_equals() { 1261 Node* arg1 = argument(0); 1262 Node* arg2 = argument(1); 1263 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1264 return true; 1265 } 1266 1267 // Java version of String.indexOf(constant string) 1268 // class StringDecl { 1269 // StringDecl(char[] ca) { 1270 // offset = 0; 1271 // count = ca.length; 1272 // value = ca; 1273 // } 1274 // int offset; 1275 // int count; 1276 // char[] value; 1277 // } 1278 // 1279 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1280 // int targetOffset, int cache_i, int md2) { 1281 // int cache = cache_i; 1282 // int sourceOffset = string_object.offset; 1283 // int sourceCount = string_object.count; 1284 // int targetCount = target_object.length; 1285 // 1286 // int targetCountLess1 = targetCount - 1; 1287 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1288 // 1289 // char[] source = string_object.value; 1290 // char[] target = target_object; 1291 // int lastChar = target[targetCountLess1]; 1292 // 1293 // outer_loop: 1294 // for (int i = sourceOffset; i < sourceEnd; ) { 1295 // int src = source[i + targetCountLess1]; 1296 // if (src == lastChar) { 1297 // // With random strings and a 4-character alphabet, 1298 // // reverse matching at this point sets up 0.8% fewer 1299 // // frames, but (paradoxically) makes 0.3% more probes. 1300 // // Since those probes are nearer the lastChar probe, 1301 // // there is may be a net D$ win with reverse matching. 1302 // // But, reversing loop inhibits unroll of inner loop 1303 // // for unknown reason. So, does running outer loop from 1304 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1305 // for (int j = 0; j < targetCountLess1; j++) { 1306 // if (target[targetOffset + j] != source[i+j]) { 1307 // if ((cache & (1 << source[i+j])) == 0) { 1308 // if (md2 < j+1) { 1309 // i += j+1; 1310 // continue outer_loop; 1311 // } 1312 // } 1313 // i += md2; 1314 // continue outer_loop; 1315 // } 1316 // } 1317 // return i - sourceOffset; 1318 // } 1319 // if ((cache & (1 << src)) == 0) { 1320 // i += targetCountLess1; 1321 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1322 // i++; 1323 // } 1324 // return -1; 1325 // } 1326 1327 //------------------------------string_indexOf------------------------ 1328 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1329 jint cache_i, jint md2_i) { 1330 1331 Node* no_ctrl = NULL; 1332 float likely = PROB_LIKELY(0.9); 1333 float unlikely = PROB_UNLIKELY(0.9); 1334 1335 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1336 1337 Node* source = load_String_value(no_ctrl, string_object); 1338 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1339 Node* sourceCount = load_String_length(no_ctrl, string_object); 1340 1341 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1342 jint target_length = target_array->length(); 1343 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1344 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1345 1346 // String.value field is known to be @Stable. 1347 if (UseImplicitStableValues) { 1348 target = cast_array_to_stable(target, target_type); 1349 } 1350 1351 IdealKit kit(this, false, true); 1352 #define __ kit. 1353 Node* zero = __ ConI(0); 1354 Node* one = __ ConI(1); 1355 Node* cache = __ ConI(cache_i); 1356 Node* md2 = __ ConI(md2_i); 1357 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1358 Node* targetCount = __ ConI(target_length); 1359 Node* targetCountLess1 = __ ConI(target_length - 1); 1360 Node* targetOffset = __ ConI(targetOffset_i); 1361 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1362 1363 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1364 Node* outer_loop = __ make_label(2 /* goto */); 1365 Node* return_ = __ make_label(1); 1366 1367 __ set(rtn,__ ConI(-1)); 1368 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1369 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1370 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1371 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1372 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1373 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1374 Node* tpj = __ AddI(targetOffset, __ value(j)); 1375 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1376 Node* ipj = __ AddI(__ value(i), __ value(j)); 1377 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1378 __ if_then(targ, BoolTest::ne, src2); { 1379 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1380 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1381 __ increment(i, __ AddI(__ value(j), one)); 1382 __ goto_(outer_loop); 1383 } __ end_if(); __ dead(j); 1384 }__ end_if(); __ dead(j); 1385 __ increment(i, md2); 1386 __ goto_(outer_loop); 1387 }__ end_if(); 1388 __ increment(j, one); 1389 }__ end_loop(); __ dead(j); 1390 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1391 __ goto_(return_); 1392 }__ end_if(); 1393 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1394 __ increment(i, targetCountLess1); 1395 }__ end_if(); 1396 __ increment(i, one); 1397 __ bind(outer_loop); 1398 }__ end_loop(); __ dead(i); 1399 __ bind(return_); 1400 1401 // Final sync IdealKit and GraphKit. 1402 final_sync(kit); 1403 Node* result = __ value(rtn); 1404 #undef __ 1405 C->set_has_loops(true); 1406 return result; 1407 } 1408 1409 //------------------------------inline_string_indexOf------------------------ 1410 bool LibraryCallKit::inline_string_indexOf() { 1411 Node* receiver = argument(0); 1412 Node* arg = argument(1); 1413 1414 Node* result; 1415 // Disable the use of pcmpestri until it can be guaranteed that 1416 // the load doesn't cross into the uncommited space. 1417 if (Matcher::has_match_rule(Op_StrIndexOf) && 1418 UseSSE42Intrinsics) { 1419 // Generate SSE4.2 version of indexOf 1420 // We currently only have match rules that use SSE4.2 1421 1422 receiver = null_check(receiver); 1423 arg = null_check(arg); 1424 if (stopped()) { 1425 return true; 1426 } 1427 1428 ciInstanceKlass* str_klass = env()->String_klass(); 1429 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1430 1431 // Make the merge point 1432 RegionNode* result_rgn = new (C) RegionNode(4); 1433 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT); 1434 Node* no_ctrl = NULL; 1435 1436 // Get start addr of source string 1437 Node* source = load_String_value(no_ctrl, receiver); 1438 Node* source_offset = load_String_offset(no_ctrl, receiver); 1439 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1440 1441 // Get length of source string 1442 Node* source_cnt = load_String_length(no_ctrl, receiver); 1443 1444 // Get start addr of substring 1445 Node* substr = load_String_value(no_ctrl, arg); 1446 Node* substr_offset = load_String_offset(no_ctrl, arg); 1447 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1448 1449 // Get length of source string 1450 Node* substr_cnt = load_String_length(no_ctrl, arg); 1451 1452 // Check for substr count > string count 1453 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt)); 1454 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt)); 1455 Node* if_gt = generate_slow_guard(bol, NULL); 1456 if (if_gt != NULL) { 1457 result_phi->init_req(2, intcon(-1)); 1458 result_rgn->init_req(2, if_gt); 1459 } 1460 1461 if (!stopped()) { 1462 // Check for substr count == 0 1463 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0))); 1464 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 1465 Node* if_zero = generate_slow_guard(bol, NULL); 1466 if (if_zero != NULL) { 1467 result_phi->init_req(3, intcon(0)); 1468 result_rgn->init_req(3, if_zero); 1469 } 1470 } 1471 1472 if (!stopped()) { 1473 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1474 result_phi->init_req(1, result); 1475 result_rgn->init_req(1, control()); 1476 } 1477 set_control(_gvn.transform(result_rgn)); 1478 record_for_igvn(result_rgn); 1479 result = _gvn.transform(result_phi); 1480 1481 } else { // Use LibraryCallKit::string_indexOf 1482 // don't intrinsify if argument isn't a constant string. 1483 if (!arg->is_Con()) { 1484 return false; 1485 } 1486 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1487 if (str_type == NULL) { 1488 return false; 1489 } 1490 ciInstanceKlass* klass = env()->String_klass(); 1491 ciObject* str_const = str_type->const_oop(); 1492 if (str_const == NULL || str_const->klass() != klass) { 1493 return false; 1494 } 1495 ciInstance* str = str_const->as_instance(); 1496 assert(str != NULL, "must be instance"); 1497 1498 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1499 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1500 1501 int o; 1502 int c; 1503 if (java_lang_String::has_offset_field()) { 1504 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1505 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1506 } else { 1507 o = 0; 1508 c = pat->length(); 1509 } 1510 1511 // constant strings have no offset and count == length which 1512 // simplifies the resulting code somewhat so lets optimize for that. 1513 if (o != 0 || c != pat->length()) { 1514 return false; 1515 } 1516 1517 receiver = null_check(receiver, T_OBJECT); 1518 // NOTE: No null check on the argument is needed since it's a constant String oop. 1519 if (stopped()) { 1520 return true; 1521 } 1522 1523 // The null string as a pattern always returns 0 (match at beginning of string) 1524 if (c == 0) { 1525 set_result(intcon(0)); 1526 return true; 1527 } 1528 1529 // Generate default indexOf 1530 jchar lastChar = pat->char_at(o + (c - 1)); 1531 int cache = 0; 1532 int i; 1533 for (i = 0; i < c - 1; i++) { 1534 assert(i < pat->length(), "out of range"); 1535 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1536 } 1537 1538 int md2 = c; 1539 for (i = 0; i < c - 1; i++) { 1540 assert(i < pat->length(), "out of range"); 1541 if (pat->char_at(o + i) == lastChar) { 1542 md2 = (c - 1) - i; 1543 } 1544 } 1545 1546 result = string_indexOf(receiver, pat, o, cache, md2); 1547 } 1548 set_result(result); 1549 return true; 1550 } 1551 1552 //--------------------------round_double_node-------------------------------- 1553 // Round a double node if necessary. 1554 Node* LibraryCallKit::round_double_node(Node* n) { 1555 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1556 n = _gvn.transform(new (C) RoundDoubleNode(0, n)); 1557 return n; 1558 } 1559 1560 //------------------------------inline_math----------------------------------- 1561 // public static double Math.abs(double) 1562 // public static double Math.sqrt(double) 1563 // public static double Math.log(double) 1564 // public static double Math.log10(double) 1565 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1566 Node* arg = round_double_node(argument(0)); 1567 Node* n; 1568 switch (id) { 1569 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break; 1570 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break; 1571 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break; 1572 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break; 1573 default: fatal_unexpected_iid(id); break; 1574 } 1575 set_result(_gvn.transform(n)); 1576 return true; 1577 } 1578 1579 //------------------------------inline_trig---------------------------------- 1580 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1581 // argument reduction which will turn into a fast/slow diamond. 1582 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1583 Node* arg = round_double_node(argument(0)); 1584 Node* n = NULL; 1585 1586 switch (id) { 1587 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break; 1588 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break; 1589 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break; 1590 default: fatal_unexpected_iid(id); break; 1591 } 1592 n = _gvn.transform(n); 1593 1594 // Rounding required? Check for argument reduction! 1595 if (Matcher::strict_fp_requires_explicit_rounding) { 1596 static const double pi_4 = 0.7853981633974483; 1597 static const double neg_pi_4 = -0.7853981633974483; 1598 // pi/2 in 80-bit extended precision 1599 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1600 // -pi/2 in 80-bit extended precision 1601 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00}; 1602 // Cutoff value for using this argument reduction technique 1603 //static const double pi_2_minus_epsilon = 1.564660403643354; 1604 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1605 1606 // Pseudocode for sin: 1607 // if (x <= Math.PI / 4.0) { 1608 // if (x >= -Math.PI / 4.0) return fsin(x); 1609 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1610 // } else { 1611 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1612 // } 1613 // return StrictMath.sin(x); 1614 1615 // Pseudocode for cos: 1616 // if (x <= Math.PI / 4.0) { 1617 // if (x >= -Math.PI / 4.0) return fcos(x); 1618 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1619 // } else { 1620 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1621 // } 1622 // return StrictMath.cos(x); 1623 1624 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1625 // requires a special machine instruction to load it. Instead we'll try 1626 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1627 // probably do the math inside the SIN encoding. 1628 1629 // Make the merge point 1630 RegionNode* r = new (C) RegionNode(3); 1631 Node* phi = new (C) PhiNode(r, Type::DOUBLE); 1632 1633 // Flatten arg so we need only 1 test 1634 Node *abs = _gvn.transform(new (C) AbsDNode(arg)); 1635 // Node for PI/4 constant 1636 Node *pi4 = makecon(TypeD::make(pi_4)); 1637 // Check PI/4 : abs(arg) 1638 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs)); 1639 // Check: If PI/4 < abs(arg) then go slow 1640 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt )); 1641 // Branch either way 1642 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1643 set_control(opt_iff(r,iff)); 1644 1645 // Set fast path result 1646 phi->init_req(2, n); 1647 1648 // Slow path - non-blocking leaf call 1649 Node* call = NULL; 1650 switch (id) { 1651 case vmIntrinsics::_dsin: 1652 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1653 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1654 "Sin", NULL, arg, top()); 1655 break; 1656 case vmIntrinsics::_dcos: 1657 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1658 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1659 "Cos", NULL, arg, top()); 1660 break; 1661 case vmIntrinsics::_dtan: 1662 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1663 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1664 "Tan", NULL, arg, top()); 1665 break; 1666 } 1667 assert(control()->in(0) == call, ""); 1668 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1669 r->init_req(1, control()); 1670 phi->init_req(1, slow_result); 1671 1672 // Post-merge 1673 set_control(_gvn.transform(r)); 1674 record_for_igvn(r); 1675 n = _gvn.transform(phi); 1676 1677 C->set_has_split_ifs(true); // Has chance for split-if optimization 1678 } 1679 set_result(n); 1680 return true; 1681 } 1682 1683 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1684 //------------------- 1685 //result=(result.isNaN())? funcAddr():result; 1686 // Check: If isNaN() by checking result!=result? then either trap 1687 // or go to runtime 1688 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result)); 1689 // Build the boolean node 1690 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq)); 1691 1692 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1693 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1694 // The pow or exp intrinsic returned a NaN, which requires a call 1695 // to the runtime. Recompile with the runtime call. 1696 uncommon_trap(Deoptimization::Reason_intrinsic, 1697 Deoptimization::Action_make_not_entrant); 1698 } 1699 set_result(result); 1700 } else { 1701 // If this inlining ever returned NaN in the past, we compile a call 1702 // to the runtime to properly handle corner cases 1703 1704 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1705 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff)); 1706 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff)); 1707 1708 if (!if_slow->is_top()) { 1709 RegionNode* result_region = new (C) RegionNode(3); 1710 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE); 1711 1712 result_region->init_req(1, if_fast); 1713 result_val->init_req(1, result); 1714 1715 set_control(if_slow); 1716 1717 const TypePtr* no_memory_effects = NULL; 1718 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1719 no_memory_effects, 1720 x, top(), y, y ? top() : NULL); 1721 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0)); 1722 #ifdef ASSERT 1723 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1)); 1724 assert(value_top == top(), "second value must be top"); 1725 #endif 1726 1727 result_region->init_req(2, control()); 1728 result_val->init_req(2, value); 1729 set_result(result_region, result_val); 1730 } else { 1731 set_result(result); 1732 } 1733 } 1734 } 1735 1736 //------------------------------inline_exp------------------------------------- 1737 // Inline exp instructions, if possible. The Intel hardware only misses 1738 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1739 bool LibraryCallKit::inline_exp() { 1740 Node* arg = round_double_node(argument(0)); 1741 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg)); 1742 1743 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 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 (x <= 0.0) { 1754 // long longy = (long)y; 1755 // if ((double)longy == y) { // if y is long 1756 // if (y + 1 == y) longy = 0; // huge number: even 1757 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1758 // } else { 1759 // result = NaN; 1760 // } 1761 // } else { 1762 // result = DPow(x,y); 1763 // } 1764 // if (result != result)? { 1765 // result = uncommon_trap() or runtime_call(); 1766 // } 1767 // return result; 1768 1769 Node* x = round_double_node(argument(0)); 1770 Node* y = round_double_node(argument(2)); 1771 1772 Node* result = NULL; 1773 1774 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1775 // Short form: skip the fancy tests and just check for NaN result. 1776 result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1777 } else { 1778 // If this inlining ever returned NaN in the past, include all 1779 // checks + call to the runtime. 1780 1781 // Set the merge point for If node with condition of (x <= 0.0) 1782 // There are four possible paths to region node and phi node 1783 RegionNode *r = new (C) RegionNode(4); 1784 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1785 1786 // Build the first if node: if (x <= 0.0) 1787 // Node for 0 constant 1788 Node *zeronode = makecon(TypeD::ZERO); 1789 // Check x:0 1790 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode)); 1791 // Check: If (x<=0) then go complex path 1792 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le )); 1793 // Branch either way 1794 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1795 // Fast path taken; set region slot 3 1796 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1)); 1797 r->init_req(3,fast_taken); // Capture fast-control 1798 1799 // Fast path not-taken, i.e. slow path 1800 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1)); 1801 1802 // Set fast path result 1803 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1804 phi->init_req(3, fast_result); 1805 1806 // Complex path 1807 // Build the second if node (if y is long) 1808 // Node for (long)y 1809 Node *longy = _gvn.transform(new (C) ConvD2LNode(y)); 1810 // Node for (double)((long) y) 1811 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy)); 1812 // Check (double)((long) y) : y 1813 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y)); 1814 // Check if (y isn't long) then go to slow path 1815 1816 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne )); 1817 // Branch either way 1818 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1819 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2)); 1820 1821 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2)); 1822 1823 // Calculate DPow(abs(x), y)*(1 & (long)y) 1824 // Node for constant 1 1825 Node *conone = longcon(1); 1826 // 1& (long)y 1827 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy)); 1828 1829 // A huge number is always even. Detect a huge number by checking 1830 // if y + 1 == y and set integer to be tested for parity to 0. 1831 // Required for corner case: 1832 // (long)9.223372036854776E18 = max_jlong 1833 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1834 // max_jlong is odd but 9.223372036854776E18 is even 1835 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1)))); 1836 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y)); 1837 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq )); 1838 Node* correctedsign = NULL; 1839 if (ConditionalMoveLimit != 0) { 1840 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1841 } else { 1842 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1843 RegionNode *r = new (C) RegionNode(3); 1844 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 1845 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1))); 1846 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1))); 1847 phi->init_req(1, signnode); 1848 phi->init_req(2, longcon(0)); 1849 correctedsign = _gvn.transform(phi); 1850 ylong_path = _gvn.transform(r); 1851 record_for_igvn(r); 1852 } 1853 1854 // zero node 1855 Node *conzero = longcon(0); 1856 // Check (1&(long)y)==0? 1857 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero)); 1858 // Check if (1&(long)y)!=0?, if so the result is negative 1859 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne )); 1860 // abs(x) 1861 Node *absx=_gvn.transform(new (C) AbsDNode(x)); 1862 // abs(x)^y 1863 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y)); 1864 // -abs(x)^y 1865 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy)); 1866 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1867 Node *signresult = NULL; 1868 if (ConditionalMoveLimit != 0) { 1869 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1870 } else { 1871 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1872 RegionNode *r = new (C) RegionNode(3); 1873 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1874 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven))); 1875 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven))); 1876 phi->init_req(1, absxpowy); 1877 phi->init_req(2, negabsxpowy); 1878 signresult = _gvn.transform(phi); 1879 ylong_path = _gvn.transform(r); 1880 record_for_igvn(r); 1881 } 1882 // Set complex path fast result 1883 r->init_req(2, ylong_path); 1884 phi->init_req(2, signresult); 1885 1886 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1887 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1888 r->init_req(1,slow_path); 1889 phi->init_req(1,slow_result); 1890 1891 // Post merge 1892 set_control(_gvn.transform(r)); 1893 record_for_igvn(r); 1894 result = _gvn.transform(phi); 1895 } 1896 1897 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1898 1899 C->set_has_split_ifs(true); // Has chance for split-if optimization 1900 return true; 1901 } 1902 1903 //------------------------------runtime_math----------------------------- 1904 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1905 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1906 "must be (DD)D or (D)D type"); 1907 1908 // Inputs 1909 Node* a = round_double_node(argument(0)); 1910 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1911 1912 const TypePtr* no_memory_effects = NULL; 1913 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1914 no_memory_effects, 1915 a, top(), b, b ? top() : NULL); 1916 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0)); 1917 #ifdef ASSERT 1918 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1)); 1919 assert(value_top == top(), "second value must be top"); 1920 #endif 1921 1922 set_result(value); 1923 return true; 1924 } 1925 1926 //------------------------------inline_math_native----------------------------- 1927 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1928 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1929 switch (id) { 1930 // These intrinsics are not properly supported on all hardware 1931 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1932 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1933 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1934 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1935 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1936 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1937 1938 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1939 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1940 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1941 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1942 1943 // These intrinsics are supported on all hardware 1944 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false; 1945 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1946 1947 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1948 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1949 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1950 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1951 #undef FN_PTR 1952 1953 // These intrinsics are not yet correctly implemented 1954 case vmIntrinsics::_datan2: 1955 return false; 1956 1957 default: 1958 fatal_unexpected_iid(id); 1959 return false; 1960 } 1961 } 1962 1963 static bool is_simple_name(Node* n) { 1964 return (n->req() == 1 // constant 1965 || (n->is_Type() && n->as_Type()->type()->singleton()) 1966 || n->is_Proj() // parameter or return value 1967 || n->is_Phi() // local of some sort 1968 ); 1969 } 1970 1971 //----------------------------inline_min_max----------------------------------- 1972 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1973 set_result(generate_min_max(id, argument(0), argument(1))); 1974 return true; 1975 } 1976 1977 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1978 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) ); 1979 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1980 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check)); 1981 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) ); 1982 1983 { 1984 PreserveJVMState pjvms(this); 1985 PreserveReexecuteState preexecs(this); 1986 jvms()->set_should_reexecute(true); 1987 1988 set_control(slow_path); 1989 set_i_o(i_o()); 1990 1991 uncommon_trap(Deoptimization::Reason_intrinsic, 1992 Deoptimization::Action_none); 1993 } 1994 1995 set_control(fast_path); 1996 set_result(math); 1997 } 1998 1999 template <typename OverflowOp> 2000 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2001 typedef typename OverflowOp::MathOp MathOp; 2002 2003 MathOp* mathOp = new(C) MathOp(arg1, arg2); 2004 Node* operation = _gvn.transform( mathOp ); 2005 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) ); 2006 inline_math_mathExact(operation, ofcheck); 2007 return true; 2008 } 2009 2010 template <typename OverflowOp> 2011 bool LibraryCallKit::inline_math_overflow(bool is_unary) { 2012 Node* arg1 = argument(0); 2013 Node* arg2 = NULL; 2014 2015 if (OverflowOp::IsLong) { 2016 arg2 = (is_unary ? longcon(1) : argument(2)); 2017 } else { 2018 arg2 = (is_unary ? intcon(1) : argument(1)); 2019 } 2020 return inline_math_overflow<OverflowOp>(arg1, arg2); 2021 } 2022 2023 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2024 return inline_math_overflow<OverflowAddINode>(is_increment); 2025 } 2026 2027 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2028 return inline_math_overflow<OverflowAddLNode>(is_increment); 2029 } 2030 2031 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2032 return inline_math_overflow<OverflowSubINode>(is_decrement); 2033 } 2034 2035 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2036 return inline_math_overflow<OverflowSubLNode>(is_decrement); 2037 } 2038 2039 bool LibraryCallKit::inline_math_negateExactI() { 2040 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2041 } 2042 2043 bool LibraryCallKit::inline_math_negateExactL() { 2044 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2045 } 2046 2047 bool LibraryCallKit::inline_math_multiplyExactI() { 2048 return inline_math_overflow<OverflowMulINode>(false); 2049 } 2050 2051 bool LibraryCallKit::inline_math_multiplyExactL() { 2052 return inline_math_overflow<OverflowMulLNode>(false); 2053 } 2054 2055 Node* 2056 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2057 // These are the candidate return value: 2058 Node* xvalue = x0; 2059 Node* yvalue = y0; 2060 2061 if (xvalue == yvalue) { 2062 return xvalue; 2063 } 2064 2065 bool want_max = (id == vmIntrinsics::_max); 2066 2067 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2068 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2069 if (txvalue == NULL || tyvalue == NULL) return top(); 2070 // This is not really necessary, but it is consistent with a 2071 // hypothetical MaxINode::Value method: 2072 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2073 2074 // %%% This folding logic should (ideally) be in a different place. 2075 // Some should be inside IfNode, and there to be a more reliable 2076 // transformation of ?: style patterns into cmoves. We also want 2077 // more powerful optimizations around cmove and min/max. 2078 2079 // Try to find a dominating comparison of these guys. 2080 // It can simplify the index computation for Arrays.copyOf 2081 // and similar uses of System.arraycopy. 2082 // First, compute the normalized version of CmpI(x, y). 2083 int cmp_op = Op_CmpI; 2084 Node* xkey = xvalue; 2085 Node* ykey = yvalue; 2086 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey)); 2087 if (ideal_cmpxy->is_Cmp()) { 2088 // E.g., if we have CmpI(length - offset, count), 2089 // it might idealize to CmpI(length, count + offset) 2090 cmp_op = ideal_cmpxy->Opcode(); 2091 xkey = ideal_cmpxy->in(1); 2092 ykey = ideal_cmpxy->in(2); 2093 } 2094 2095 // Start by locating any relevant comparisons. 2096 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2097 Node* cmpxy = NULL; 2098 Node* cmpyx = NULL; 2099 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2100 Node* cmp = start_from->fast_out(k); 2101 if (cmp->outcnt() > 0 && // must have prior uses 2102 cmp->in(0) == NULL && // must be context-independent 2103 cmp->Opcode() == cmp_op) { // right kind of compare 2104 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2105 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2106 } 2107 } 2108 2109 const int NCMPS = 2; 2110 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2111 int cmpn; 2112 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2113 if (cmps[cmpn] != NULL) break; // find a result 2114 } 2115 if (cmpn < NCMPS) { 2116 // Look for a dominating test that tells us the min and max. 2117 int depth = 0; // Limit search depth for speed 2118 Node* dom = control(); 2119 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2120 if (++depth >= 100) break; 2121 Node* ifproj = dom; 2122 if (!ifproj->is_Proj()) continue; 2123 Node* iff = ifproj->in(0); 2124 if (!iff->is_If()) continue; 2125 Node* bol = iff->in(1); 2126 if (!bol->is_Bool()) continue; 2127 Node* cmp = bol->in(1); 2128 if (cmp == NULL) continue; 2129 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2130 if (cmps[cmpn] == cmp) break; 2131 if (cmpn == NCMPS) continue; 2132 BoolTest::mask btest = bol->as_Bool()->_test._test; 2133 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2134 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2135 // At this point, we know that 'x btest y' is true. 2136 switch (btest) { 2137 case BoolTest::eq: 2138 // They are proven equal, so we can collapse the min/max. 2139 // Either value is the answer. Choose the simpler. 2140 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2141 return yvalue; 2142 return xvalue; 2143 case BoolTest::lt: // x < y 2144 case BoolTest::le: // x <= y 2145 return (want_max ? yvalue : xvalue); 2146 case BoolTest::gt: // x > y 2147 case BoolTest::ge: // x >= y 2148 return (want_max ? xvalue : yvalue); 2149 } 2150 } 2151 } 2152 2153 // We failed to find a dominating test. 2154 // Let's pick a test that might GVN with prior tests. 2155 Node* best_bol = NULL; 2156 BoolTest::mask best_btest = BoolTest::illegal; 2157 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2158 Node* cmp = cmps[cmpn]; 2159 if (cmp == NULL) continue; 2160 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2161 Node* bol = cmp->fast_out(j); 2162 if (!bol->is_Bool()) continue; 2163 BoolTest::mask btest = bol->as_Bool()->_test._test; 2164 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2165 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2166 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2167 best_bol = bol->as_Bool(); 2168 best_btest = btest; 2169 } 2170 } 2171 } 2172 2173 Node* answer_if_true = NULL; 2174 Node* answer_if_false = NULL; 2175 switch (best_btest) { 2176 default: 2177 if (cmpxy == NULL) 2178 cmpxy = ideal_cmpxy; 2179 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt)); 2180 // and fall through: 2181 case BoolTest::lt: // x < y 2182 case BoolTest::le: // x <= y 2183 answer_if_true = (want_max ? yvalue : xvalue); 2184 answer_if_false = (want_max ? xvalue : yvalue); 2185 break; 2186 case BoolTest::gt: // x > y 2187 case BoolTest::ge: // x >= y 2188 answer_if_true = (want_max ? xvalue : yvalue); 2189 answer_if_false = (want_max ? yvalue : xvalue); 2190 break; 2191 } 2192 2193 jint hi, lo; 2194 if (want_max) { 2195 // We can sharpen the minimum. 2196 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2197 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2198 } else { 2199 // We can sharpen the maximum. 2200 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2201 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2202 } 2203 2204 // Use a flow-free graph structure, to avoid creating excess control edges 2205 // which could hinder other optimizations. 2206 // Since Math.min/max is often used with arraycopy, we want 2207 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2208 Node* cmov = CMoveNode::make(C, NULL, best_bol, 2209 answer_if_false, answer_if_true, 2210 TypeInt::make(lo, hi, widen)); 2211 2212 return _gvn.transform(cmov); 2213 2214 /* 2215 // This is not as desirable as it may seem, since Min and Max 2216 // nodes do not have a full set of optimizations. 2217 // And they would interfere, anyway, with 'if' optimizations 2218 // and with CMoveI canonical forms. 2219 switch (id) { 2220 case vmIntrinsics::_min: 2221 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2222 case vmIntrinsics::_max: 2223 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2224 default: 2225 ShouldNotReachHere(); 2226 } 2227 */ 2228 } 2229 2230 inline int 2231 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2232 const TypePtr* base_type = TypePtr::NULL_PTR; 2233 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2234 if (base_type == NULL) { 2235 // Unknown type. 2236 return Type::AnyPtr; 2237 } else if (base_type == TypePtr::NULL_PTR) { 2238 // Since this is a NULL+long form, we have to switch to a rawptr. 2239 base = _gvn.transform(new (C) CastX2PNode(offset)); 2240 offset = MakeConX(0); 2241 return Type::RawPtr; 2242 } else if (base_type->base() == Type::RawPtr) { 2243 return Type::RawPtr; 2244 } else if (base_type->isa_oopptr()) { 2245 // Base is never null => always a heap address. 2246 if (base_type->ptr() == TypePtr::NotNull) { 2247 return Type::OopPtr; 2248 } 2249 // Offset is small => always a heap address. 2250 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2251 if (offset_type != NULL && 2252 base_type->offset() == 0 && // (should always be?) 2253 offset_type->_lo >= 0 && 2254 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2255 return Type::OopPtr; 2256 } 2257 // Otherwise, it might either be oop+off or NULL+addr. 2258 return Type::AnyPtr; 2259 } else { 2260 // No information: 2261 return Type::AnyPtr; 2262 } 2263 } 2264 2265 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2266 int kind = classify_unsafe_addr(base, offset); 2267 if (kind == Type::RawPtr) { 2268 return basic_plus_adr(top(), base, offset); 2269 } else { 2270 return basic_plus_adr(base, offset); 2271 } 2272 } 2273 2274 //--------------------------inline_number_methods----------------------------- 2275 // inline int Integer.numberOfLeadingZeros(int) 2276 // inline int Long.numberOfLeadingZeros(long) 2277 // 2278 // inline int Integer.numberOfTrailingZeros(int) 2279 // inline int Long.numberOfTrailingZeros(long) 2280 // 2281 // inline int Integer.bitCount(int) 2282 // inline int Long.bitCount(long) 2283 // 2284 // inline char Character.reverseBytes(char) 2285 // inline short Short.reverseBytes(short) 2286 // inline int Integer.reverseBytes(int) 2287 // inline long Long.reverseBytes(long) 2288 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2289 Node* arg = argument(0); 2290 Node* n; 2291 switch (id) { 2292 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break; 2293 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break; 2294 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break; 2295 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break; 2296 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break; 2297 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break; 2298 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break; 2299 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break; 2300 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break; 2301 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break; 2302 default: fatal_unexpected_iid(id); break; 2303 } 2304 set_result(_gvn.transform(n)); 2305 return true; 2306 } 2307 2308 //----------------------------inline_unsafe_access---------------------------- 2309 2310 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2311 2312 // Helper that guards and inserts a pre-barrier. 2313 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2314 Node* pre_val, bool need_mem_bar) { 2315 // We could be accessing the referent field of a reference object. If so, when G1 2316 // is enabled, we need to log the value in the referent field in an SATB buffer. 2317 // This routine performs some compile time filters and generates suitable 2318 // runtime filters that guard the pre-barrier code. 2319 // Also add memory barrier for non volatile load from the referent field 2320 // to prevent commoning of loads across safepoint. 2321 if (!UseG1GC && !need_mem_bar) 2322 return; 2323 2324 // Some compile time checks. 2325 2326 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2327 const TypeX* otype = offset->find_intptr_t_type(); 2328 if (otype != NULL && otype->is_con() && 2329 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2330 // Constant offset but not the reference_offset so just return 2331 return; 2332 } 2333 2334 // We only need to generate the runtime guards for instances. 2335 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2336 if (btype != NULL) { 2337 if (btype->isa_aryptr()) { 2338 // Array type so nothing to do 2339 return; 2340 } 2341 2342 const TypeInstPtr* itype = btype->isa_instptr(); 2343 if (itype != NULL) { 2344 // Can the klass of base_oop be statically determined to be 2345 // _not_ a sub-class of Reference and _not_ Object? 2346 ciKlass* klass = itype->klass(); 2347 if ( klass->is_loaded() && 2348 !klass->is_subtype_of(env()->Reference_klass()) && 2349 !env()->Object_klass()->is_subtype_of(klass)) { 2350 return; 2351 } 2352 } 2353 } 2354 2355 // The compile time filters did not reject base_oop/offset so 2356 // we need to generate the following runtime filters 2357 // 2358 // if (offset == java_lang_ref_Reference::_reference_offset) { 2359 // if (instance_of(base, java.lang.ref.Reference)) { 2360 // pre_barrier(_, pre_val, ...); 2361 // } 2362 // } 2363 2364 float likely = PROB_LIKELY( 0.999); 2365 float unlikely = PROB_UNLIKELY(0.999); 2366 2367 IdealKit ideal(this); 2368 #define __ ideal. 2369 2370 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2371 2372 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2373 // Update graphKit memory and control from IdealKit. 2374 sync_kit(ideal); 2375 2376 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2377 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2378 2379 // Update IdealKit memory and control from graphKit. 2380 __ sync_kit(this); 2381 2382 Node* one = __ ConI(1); 2383 // is_instof == 0 if base_oop == NULL 2384 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2385 2386 // Update graphKit from IdeakKit. 2387 sync_kit(ideal); 2388 2389 // Use the pre-barrier to record the value in the referent field 2390 pre_barrier(false /* do_load */, 2391 __ ctrl(), 2392 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2393 pre_val /* pre_val */, 2394 T_OBJECT); 2395 if (need_mem_bar) { 2396 // Add memory barrier to prevent commoning reads from this field 2397 // across safepoint since GC can change its value. 2398 insert_mem_bar(Op_MemBarCPUOrder); 2399 } 2400 // Update IdealKit from graphKit. 2401 __ sync_kit(this); 2402 2403 } __ end_if(); // _ref_type != ref_none 2404 } __ end_if(); // offset == referent_offset 2405 2406 // Final sync IdealKit and GraphKit. 2407 final_sync(ideal); 2408 #undef __ 2409 } 2410 2411 2412 // Interpret Unsafe.fieldOffset cookies correctly: 2413 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2414 2415 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2416 // Attempt to infer a sharper value type from the offset and base type. 2417 ciKlass* sharpened_klass = NULL; 2418 2419 // See if it is an instance field, with an object type. 2420 if (alias_type->field() != NULL) { 2421 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2422 if (alias_type->field()->type()->is_klass()) { 2423 sharpened_klass = alias_type->field()->type()->as_klass(); 2424 } 2425 } 2426 2427 // See if it is a narrow oop array. 2428 if (adr_type->isa_aryptr()) { 2429 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2430 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2431 if (elem_type != NULL) { 2432 sharpened_klass = elem_type->klass(); 2433 } 2434 } 2435 } 2436 2437 // The sharpened class might be unloaded if there is no class loader 2438 // contraint in place. 2439 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2440 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2441 2442 #ifndef PRODUCT 2443 if (C->print_intrinsics() || C->print_inlining()) { 2444 tty->print(" from base type: "); adr_type->dump(); 2445 tty->print(" sharpened value: "); tjp->dump(); 2446 } 2447 #endif 2448 // Sharpen the value type. 2449 return tjp; 2450 } 2451 return NULL; 2452 } 2453 2454 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2455 if (callee()->is_static()) return false; // caller must have the capability! 2456 2457 #ifndef PRODUCT 2458 { 2459 ResourceMark rm; 2460 // Check the signatures. 2461 ciSignature* sig = callee()->signature(); 2462 #ifdef ASSERT 2463 if (!is_store) { 2464 // Object getObject(Object base, int/long offset), etc. 2465 BasicType rtype = sig->return_type()->basic_type(); 2466 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2467 rtype = T_ADDRESS; // it is really a C void* 2468 assert(rtype == type, "getter must return the expected value"); 2469 if (!is_native_ptr) { 2470 assert(sig->count() == 2, "oop getter has 2 arguments"); 2471 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2472 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2473 } else { 2474 assert(sig->count() == 1, "native getter has 1 argument"); 2475 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2476 } 2477 } else { 2478 // void putObject(Object base, int/long offset, Object x), etc. 2479 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2480 if (!is_native_ptr) { 2481 assert(sig->count() == 3, "oop putter has 3 arguments"); 2482 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2483 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2484 } else { 2485 assert(sig->count() == 2, "native putter has 2 arguments"); 2486 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2487 } 2488 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2489 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2490 vtype = T_ADDRESS; // it is really a C void* 2491 assert(vtype == type, "putter must accept the expected value"); 2492 } 2493 #endif // ASSERT 2494 } 2495 #endif //PRODUCT 2496 2497 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2498 2499 Node* receiver = argument(0); // type: oop 2500 2501 // Build address expression. See the code in inline_unsafe_prefetch. 2502 Node* adr; 2503 Node* heap_base_oop = top(); 2504 Node* offset = top(); 2505 Node* val; 2506 2507 if (!is_native_ptr) { 2508 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2509 Node* base = argument(1); // type: oop 2510 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2511 offset = argument(2); // type: long 2512 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2513 // to be plain byte offsets, which are also the same as those accepted 2514 // by oopDesc::field_base. 2515 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2516 "fieldOffset must be byte-scaled"); 2517 // 32-bit machines ignore the high half! 2518 offset = ConvL2X(offset); 2519 adr = make_unsafe_address(base, offset); 2520 heap_base_oop = base; 2521 val = is_store ? argument(4) : NULL; 2522 } else { 2523 Node* ptr = argument(1); // type: long 2524 ptr = ConvL2X(ptr); // adjust Java long to machine word 2525 adr = make_unsafe_address(NULL, ptr); 2526 val = is_store ? argument(3) : NULL; 2527 } 2528 2529 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2530 2531 // First guess at the value type. 2532 const Type *value_type = Type::get_const_basic_type(type); 2533 2534 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2535 // there was not enough information to nail it down. 2536 Compile::AliasType* alias_type = C->alias_type(adr_type); 2537 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2538 2539 // We will need memory barriers unless we can determine a unique 2540 // alias category for this reference. (Note: If for some reason 2541 // the barriers get omitted and the unsafe reference begins to "pollute" 2542 // the alias analysis of the rest of the graph, either Compile::can_alias 2543 // or Compile::must_alias will throw a diagnostic assert.) 2544 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2545 2546 // If we are reading the value of the referent field of a Reference 2547 // object (either by using Unsafe directly or through reflection) 2548 // then, if G1 is enabled, we need to record the referent in an 2549 // SATB log buffer using the pre-barrier mechanism. 2550 // Also we need to add memory barrier to prevent commoning reads 2551 // from this field across safepoint since GC can change its value. 2552 bool need_read_barrier = !is_native_ptr && !is_store && 2553 offset != top() && heap_base_oop != top(); 2554 2555 if (!is_store && type == T_OBJECT) { 2556 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2557 if (tjp != NULL) { 2558 value_type = tjp; 2559 } 2560 } 2561 2562 receiver = null_check(receiver); 2563 if (stopped()) { 2564 return true; 2565 } 2566 // Heap pointers get a null-check from the interpreter, 2567 // as a courtesy. However, this is not guaranteed by Unsafe, 2568 // and it is not possible to fully distinguish unintended nulls 2569 // from intended ones in this API. 2570 2571 if (is_volatile) { 2572 // We need to emit leading and trailing CPU membars (see below) in 2573 // addition to memory membars when is_volatile. This is a little 2574 // too strong, but avoids the need to insert per-alias-type 2575 // volatile membars (for stores; compare Parse::do_put_xxx), which 2576 // we cannot do effectively here because we probably only have a 2577 // rough approximation of type. 2578 need_mem_bar = true; 2579 // For Stores, place a memory ordering barrier now. 2580 if (is_store) { 2581 insert_mem_bar(Op_MemBarRelease); 2582 } else { 2583 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2584 insert_mem_bar(Op_MemBarVolatile); 2585 } 2586 } 2587 } 2588 2589 // Memory barrier to prevent normal and 'unsafe' accesses from 2590 // bypassing each other. Happens after null checks, so the 2591 // exception paths do not take memory state from the memory barrier, 2592 // so there's no problems making a strong assert about mixing users 2593 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2594 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2595 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2596 2597 if (!is_store) { 2598 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile); 2599 // load value 2600 switch (type) { 2601 case T_BOOLEAN: 2602 case T_CHAR: 2603 case T_BYTE: 2604 case T_SHORT: 2605 case T_INT: 2606 case T_LONG: 2607 case T_FLOAT: 2608 case T_DOUBLE: 2609 break; 2610 case T_OBJECT: 2611 if (need_read_barrier) { 2612 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2613 } 2614 break; 2615 case T_ADDRESS: 2616 // Cast to an int type. 2617 p = _gvn.transform(new (C) CastP2XNode(NULL, p)); 2618 p = ConvX2L(p); 2619 break; 2620 default: 2621 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2622 break; 2623 } 2624 // The load node has the control of the preceding MemBarCPUOrder. All 2625 // following nodes will have the control of the MemBarCPUOrder inserted at 2626 // the end of this method. So, pushing the load onto the stack at a later 2627 // point is fine. 2628 set_result(p); 2629 } else { 2630 // place effect of store into memory 2631 switch (type) { 2632 case T_DOUBLE: 2633 val = dstore_rounding(val); 2634 break; 2635 case T_ADDRESS: 2636 // Repackage the long as a pointer. 2637 val = ConvL2X(val); 2638 val = _gvn.transform(new (C) CastX2PNode(val)); 2639 break; 2640 } 2641 2642 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2643 if (type != T_OBJECT ) { 2644 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2645 } else { 2646 // Possibly an oop being stored to Java heap or native memory 2647 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2648 // oop to Java heap. 2649 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2650 } else { 2651 // We can't tell at compile time if we are storing in the Java heap or outside 2652 // of it. So we need to emit code to conditionally do the proper type of 2653 // store. 2654 2655 IdealKit ideal(this); 2656 #define __ ideal. 2657 // QQQ who knows what probability is here?? 2658 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2659 // Sync IdealKit and graphKit. 2660 sync_kit(ideal); 2661 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2662 // Update IdealKit memory. 2663 __ sync_kit(this); 2664 } __ else_(); { 2665 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2666 } __ end_if(); 2667 // Final sync IdealKit and GraphKit. 2668 final_sync(ideal); 2669 #undef __ 2670 } 2671 } 2672 } 2673 2674 if (is_volatile) { 2675 if (!is_store) { 2676 insert_mem_bar(Op_MemBarAcquire); 2677 } else { 2678 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2679 insert_mem_bar(Op_MemBarVolatile); 2680 } 2681 } 2682 } 2683 2684 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2685 2686 return true; 2687 } 2688 2689 //----------------------------inline_unsafe_prefetch---------------------------- 2690 2691 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2692 #ifndef PRODUCT 2693 { 2694 ResourceMark rm; 2695 // Check the signatures. 2696 ciSignature* sig = callee()->signature(); 2697 #ifdef ASSERT 2698 // Object getObject(Object base, int/long offset), etc. 2699 BasicType rtype = sig->return_type()->basic_type(); 2700 if (!is_native_ptr) { 2701 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2702 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2703 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2704 } else { 2705 assert(sig->count() == 1, "native prefetch has 1 argument"); 2706 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2707 } 2708 #endif // ASSERT 2709 } 2710 #endif // !PRODUCT 2711 2712 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2713 2714 const int idx = is_static ? 0 : 1; 2715 if (!is_static) { 2716 null_check_receiver(); 2717 if (stopped()) { 2718 return true; 2719 } 2720 } 2721 2722 // Build address expression. See the code in inline_unsafe_access. 2723 Node *adr; 2724 if (!is_native_ptr) { 2725 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2726 Node* base = argument(idx + 0); // type: oop 2727 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2728 Node* offset = argument(idx + 1); // type: long 2729 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2730 // to be plain byte offsets, which are also the same as those accepted 2731 // by oopDesc::field_base. 2732 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2733 "fieldOffset must be byte-scaled"); 2734 // 32-bit machines ignore the high half! 2735 offset = ConvL2X(offset); 2736 adr = make_unsafe_address(base, offset); 2737 } else { 2738 Node* ptr = argument(idx + 0); // type: long 2739 ptr = ConvL2X(ptr); // adjust Java long to machine word 2740 adr = make_unsafe_address(NULL, ptr); 2741 } 2742 2743 // Generate the read or write prefetch 2744 Node *prefetch; 2745 if (is_store) { 2746 prefetch = new (C) PrefetchWriteNode(i_o(), adr); 2747 } else { 2748 prefetch = new (C) PrefetchReadNode(i_o(), adr); 2749 } 2750 prefetch->init_req(0, control()); 2751 set_i_o(_gvn.transform(prefetch)); 2752 2753 return true; 2754 } 2755 2756 //----------------------------inline_unsafe_load_store---------------------------- 2757 // This method serves a couple of different customers (depending on LoadStoreKind): 2758 // 2759 // LS_cmpxchg: 2760 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2761 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2762 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2763 // 2764 // LS_xadd: 2765 // public int getAndAddInt( Object o, long offset, int delta) 2766 // public long getAndAddLong(Object o, long offset, long delta) 2767 // 2768 // LS_xchg: 2769 // int getAndSet(Object o, long offset, int newValue) 2770 // long getAndSet(Object o, long offset, long newValue) 2771 // Object getAndSet(Object o, long offset, Object newValue) 2772 // 2773 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2774 // This basic scheme here is the same as inline_unsafe_access, but 2775 // differs in enough details that combining them would make the code 2776 // overly confusing. (This is a true fact! I originally combined 2777 // them, but even I was confused by it!) As much code/comments as 2778 // possible are retained from inline_unsafe_access though to make 2779 // the correspondences clearer. - dl 2780 2781 if (callee()->is_static()) return false; // caller must have the capability! 2782 2783 #ifndef PRODUCT 2784 BasicType rtype; 2785 { 2786 ResourceMark rm; 2787 // Check the signatures. 2788 ciSignature* sig = callee()->signature(); 2789 rtype = sig->return_type()->basic_type(); 2790 if (kind == LS_xadd || kind == LS_xchg) { 2791 // Check the signatures. 2792 #ifdef ASSERT 2793 assert(rtype == type, "get and set must return the expected type"); 2794 assert(sig->count() == 3, "get and set has 3 arguments"); 2795 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2796 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2797 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2798 #endif // ASSERT 2799 } else if (kind == LS_cmpxchg) { 2800 // Check the signatures. 2801 #ifdef ASSERT 2802 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2803 assert(sig->count() == 4, "CAS has 4 arguments"); 2804 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2805 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2806 #endif // ASSERT 2807 } else { 2808 ShouldNotReachHere(); 2809 } 2810 } 2811 #endif //PRODUCT 2812 2813 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2814 2815 // Get arguments: 2816 Node* receiver = NULL; 2817 Node* base = NULL; 2818 Node* offset = NULL; 2819 Node* oldval = NULL; 2820 Node* newval = NULL; 2821 if (kind == LS_cmpxchg) { 2822 const bool two_slot_type = type2size[type] == 2; 2823 receiver = argument(0); // type: oop 2824 base = argument(1); // type: oop 2825 offset = argument(2); // type: long 2826 oldval = argument(4); // type: oop, int, or long 2827 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2828 } else if (kind == LS_xadd || kind == LS_xchg){ 2829 receiver = argument(0); // type: oop 2830 base = argument(1); // type: oop 2831 offset = argument(2); // type: long 2832 oldval = NULL; 2833 newval = argument(4); // type: oop, int, or long 2834 } 2835 2836 // Null check receiver. 2837 receiver = null_check(receiver); 2838 if (stopped()) { 2839 return true; 2840 } 2841 2842 // Build field offset expression. 2843 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2844 // to be plain byte offsets, which are also the same as those accepted 2845 // by oopDesc::field_base. 2846 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2847 // 32-bit machines ignore the high half of long offsets 2848 offset = ConvL2X(offset); 2849 Node* adr = make_unsafe_address(base, offset); 2850 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2851 2852 // For CAS, unlike inline_unsafe_access, there seems no point in 2853 // trying to refine types. Just use the coarse types here. 2854 const Type *value_type = Type::get_const_basic_type(type); 2855 Compile::AliasType* alias_type = C->alias_type(adr_type); 2856 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2857 2858 if (kind == LS_xchg && type == T_OBJECT) { 2859 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2860 if (tjp != NULL) { 2861 value_type = tjp; 2862 } 2863 } 2864 2865 int alias_idx = C->get_alias_index(adr_type); 2866 2867 // Memory-model-wise, a LoadStore acts like a little synchronized 2868 // block, so needs barriers on each side. These don't translate 2869 // into actual barriers on most machines, but we still need rest of 2870 // compiler to respect ordering. 2871 2872 insert_mem_bar(Op_MemBarRelease); 2873 insert_mem_bar(Op_MemBarCPUOrder); 2874 2875 // 4984716: MemBars must be inserted before this 2876 // memory node in order to avoid a false 2877 // dependency which will confuse the scheduler. 2878 Node *mem = memory(alias_idx); 2879 2880 // For now, we handle only those cases that actually exist: ints, 2881 // longs, and Object. Adding others should be straightforward. 2882 Node* load_store; 2883 switch(type) { 2884 case T_INT: 2885 if (kind == LS_xadd) { 2886 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type)); 2887 } else if (kind == LS_xchg) { 2888 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type)); 2889 } else if (kind == LS_cmpxchg) { 2890 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2891 } else { 2892 ShouldNotReachHere(); 2893 } 2894 break; 2895 case T_LONG: 2896 if (kind == LS_xadd) { 2897 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2898 } else if (kind == LS_xchg) { 2899 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2900 } else if (kind == LS_cmpxchg) { 2901 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2902 } else { 2903 ShouldNotReachHere(); 2904 } 2905 break; 2906 case T_OBJECT: 2907 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2908 // could be delayed during Parse (for example, in adjust_map_after_if()). 2909 // Execute transformation here to avoid barrier generation in such case. 2910 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2911 newval = _gvn.makecon(TypePtr::NULL_PTR); 2912 2913 // Reference stores need a store barrier. 2914 if (kind == LS_xchg) { 2915 // If pre-barrier must execute before the oop store, old value will require do_load here. 2916 if (!can_move_pre_barrier()) { 2917 pre_barrier(true /* do_load*/, 2918 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2919 NULL /* pre_val*/, 2920 T_OBJECT); 2921 } // Else move pre_barrier to use load_store value, see below. 2922 } else if (kind == LS_cmpxchg) { 2923 // Same as for newval above: 2924 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2925 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2926 } 2927 // The only known value which might get overwritten is oldval. 2928 pre_barrier(false /* do_load */, 2929 control(), NULL, NULL, max_juint, NULL, NULL, 2930 oldval /* pre_val */, 2931 T_OBJECT); 2932 } else { 2933 ShouldNotReachHere(); 2934 } 2935 2936 #ifdef _LP64 2937 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2938 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2939 if (kind == LS_xchg) { 2940 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr, 2941 newval_enc, adr_type, value_type->make_narrowoop())); 2942 } else { 2943 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2944 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2945 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr, 2946 newval_enc, oldval_enc)); 2947 } 2948 } else 2949 #endif 2950 { 2951 if (kind == LS_xchg) { 2952 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2953 } else { 2954 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2955 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2956 } 2957 } 2958 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2959 break; 2960 default: 2961 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2962 break; 2963 } 2964 2965 // SCMemProjNodes represent the memory state of a LoadStore. Their 2966 // main role is to prevent LoadStore nodes from being optimized away 2967 // when their results aren't used. 2968 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store)); 2969 set_memory(proj, alias_idx); 2970 2971 if (type == T_OBJECT && kind == LS_xchg) { 2972 #ifdef _LP64 2973 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2974 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type())); 2975 } 2976 #endif 2977 if (can_move_pre_barrier()) { 2978 // Don't need to load pre_val. The old value is returned by load_store. 2979 // The pre_barrier can execute after the xchg as long as no safepoint 2980 // gets inserted between them. 2981 pre_barrier(false /* do_load */, 2982 control(), NULL, NULL, max_juint, NULL, NULL, 2983 load_store /* pre_val */, 2984 T_OBJECT); 2985 } 2986 } 2987 2988 // Add the trailing membar surrounding the access 2989 insert_mem_bar(Op_MemBarCPUOrder); 2990 insert_mem_bar(Op_MemBarAcquire); 2991 2992 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2993 set_result(load_store); 2994 return true; 2995 } 2996 2997 //----------------------------inline_unsafe_ordered_store---------------------- 2998 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2999 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3000 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3001 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3002 // This is another variant of inline_unsafe_access, differing in 3003 // that it always issues store-store ("release") barrier and ensures 3004 // store-atomicity (which only matters for "long"). 3005 3006 if (callee()->is_static()) return false; // caller must have the capability! 3007 3008 #ifndef PRODUCT 3009 { 3010 ResourceMark rm; 3011 // Check the signatures. 3012 ciSignature* sig = callee()->signature(); 3013 #ifdef ASSERT 3014 BasicType rtype = sig->return_type()->basic_type(); 3015 assert(rtype == T_VOID, "must return void"); 3016 assert(sig->count() == 3, "has 3 arguments"); 3017 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3018 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3019 #endif // ASSERT 3020 } 3021 #endif //PRODUCT 3022 3023 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3024 3025 // Get arguments: 3026 Node* receiver = argument(0); // type: oop 3027 Node* base = argument(1); // type: oop 3028 Node* offset = argument(2); // type: long 3029 Node* val = argument(4); // type: oop, int, or long 3030 3031 // Null check receiver. 3032 receiver = null_check(receiver); 3033 if (stopped()) { 3034 return true; 3035 } 3036 3037 // Build field offset expression. 3038 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3039 // 32-bit machines ignore the high half of long offsets 3040 offset = ConvL2X(offset); 3041 Node* adr = make_unsafe_address(base, offset); 3042 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3043 const Type *value_type = Type::get_const_basic_type(type); 3044 Compile::AliasType* alias_type = C->alias_type(adr_type); 3045 3046 insert_mem_bar(Op_MemBarRelease); 3047 insert_mem_bar(Op_MemBarCPUOrder); 3048 // Ensure that the store is atomic for longs: 3049 const bool require_atomic_access = true; 3050 Node* store; 3051 if (type == T_OBJECT) // reference stores need a store barrier. 3052 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3053 else { 3054 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3055 } 3056 insert_mem_bar(Op_MemBarCPUOrder); 3057 return true; 3058 } 3059 3060 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3061 // Regardless of form, don't allow previous ld/st to move down, 3062 // then issue acquire, release, or volatile mem_bar. 3063 insert_mem_bar(Op_MemBarCPUOrder); 3064 switch(id) { 3065 case vmIntrinsics::_loadFence: 3066 insert_mem_bar(Op_LoadFence); 3067 return true; 3068 case vmIntrinsics::_storeFence: 3069 insert_mem_bar(Op_StoreFence); 3070 return true; 3071 case vmIntrinsics::_fullFence: 3072 insert_mem_bar(Op_MemBarVolatile); 3073 return true; 3074 default: 3075 fatal_unexpected_iid(id); 3076 return false; 3077 } 3078 } 3079 3080 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3081 if (!kls->is_Con()) { 3082 return true; 3083 } 3084 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3085 if (klsptr == NULL) { 3086 return true; 3087 } 3088 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3089 // don't need a guard for a klass that is already initialized 3090 return !ik->is_initialized(); 3091 } 3092 3093 //----------------------------inline_unsafe_allocate--------------------------- 3094 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3095 bool LibraryCallKit::inline_unsafe_allocate() { 3096 if (callee()->is_static()) return false; // caller must have the capability! 3097 3098 null_check_receiver(); // null-check, then ignore 3099 Node* cls = null_check(argument(1)); 3100 if (stopped()) return true; 3101 3102 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3103 kls = null_check(kls); 3104 if (stopped()) return true; // argument was like int.class 3105 3106 Node* test = NULL; 3107 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3108 // Note: The argument might still be an illegal value like 3109 // Serializable.class or Object[].class. The runtime will handle it. 3110 // But we must make an explicit check for initialization. 3111 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3112 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3113 // can generate code to load it as unsigned byte. 3114 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3115 Node* bits = intcon(InstanceKlass::fully_initialized); 3116 test = _gvn.transform(new (C) SubINode(inst, bits)); 3117 // The 'test' is non-zero if we need to take a slow path. 3118 } 3119 3120 Node* obj = new_instance(kls, test); 3121 set_result(obj); 3122 return true; 3123 } 3124 3125 #ifdef TRACE_HAVE_INTRINSICS 3126 /* 3127 * oop -> myklass 3128 * myklass->trace_id |= USED 3129 * return myklass->trace_id & ~0x3 3130 */ 3131 bool LibraryCallKit::inline_native_classID() { 3132 null_check_receiver(); // null-check, then ignore 3133 Node* cls = null_check(argument(1), T_OBJECT); 3134 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3135 kls = null_check(kls, T_OBJECT); 3136 ByteSize offset = TRACE_ID_OFFSET; 3137 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3138 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3139 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3140 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits)); 3141 Node* clsused = longcon(0x01l); // set the class bit 3142 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused)); 3143 3144 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3145 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3146 set_result(andl); 3147 return true; 3148 } 3149 3150 bool LibraryCallKit::inline_native_threadID() { 3151 Node* tls_ptr = NULL; 3152 Node* cur_thr = generate_current_thread(tls_ptr); 3153 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3154 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3155 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3156 3157 Node* threadid = NULL; 3158 size_t thread_id_size = OSThread::thread_id_size(); 3159 if (thread_id_size == (size_t) BytesPerLong) { 3160 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3161 } else if (thread_id_size == (size_t) BytesPerInt) { 3162 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3163 } else { 3164 ShouldNotReachHere(); 3165 } 3166 set_result(threadid); 3167 return true; 3168 } 3169 #endif 3170 3171 //------------------------inline_native_time_funcs-------------- 3172 // inline code for System.currentTimeMillis() and System.nanoTime() 3173 // these have the same type and signature 3174 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3175 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3176 const TypePtr* no_memory_effects = NULL; 3177 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3178 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0)); 3179 #ifdef ASSERT 3180 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1)); 3181 assert(value_top == top(), "second value must be top"); 3182 #endif 3183 set_result(value); 3184 return true; 3185 } 3186 3187 //------------------------inline_native_currentThread------------------ 3188 bool LibraryCallKit::inline_native_currentThread() { 3189 Node* junk = NULL; 3190 set_result(generate_current_thread(junk)); 3191 return true; 3192 } 3193 3194 //------------------------inline_native_isInterrupted------------------ 3195 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3196 bool LibraryCallKit::inline_native_isInterrupted() { 3197 // Add a fast path to t.isInterrupted(clear_int): 3198 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 3199 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3200 // So, in the common case that the interrupt bit is false, 3201 // we avoid making a call into the VM. Even if the interrupt bit 3202 // is true, if the clear_int argument is false, we avoid the VM call. 3203 // However, if the receiver is not currentThread, we must call the VM, 3204 // because there must be some locking done around the operation. 3205 3206 // We only go to the fast case code if we pass two guards. 3207 // Paths which do not pass are accumulated in the slow_region. 3208 3209 enum { 3210 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3211 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3212 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3213 PATH_LIMIT 3214 }; 3215 3216 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3217 // out of the function. 3218 insert_mem_bar(Op_MemBarCPUOrder); 3219 3220 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT); 3221 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL); 3222 3223 RegionNode* slow_region = new (C) RegionNode(1); 3224 record_for_igvn(slow_region); 3225 3226 // (a) Receiving thread must be the current thread. 3227 Node* rec_thr = argument(0); 3228 Node* tls_ptr = NULL; 3229 Node* cur_thr = generate_current_thread(tls_ptr); 3230 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr)); 3231 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne)); 3232 3233 generate_slow_guard(bol_thr, slow_region); 3234 3235 // (b) Interrupt bit on TLS must be false. 3236 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3237 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3238 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3239 3240 // Set the control input on the field _interrupted read to prevent it floating up. 3241 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3242 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0))); 3243 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne)); 3244 3245 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3246 3247 // First fast path: if (!TLS._interrupted) return false; 3248 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit)); 3249 result_rgn->init_req(no_int_result_path, false_bit); 3250 result_val->init_req(no_int_result_path, intcon(0)); 3251 3252 // drop through to next case 3253 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit))); 3254 3255 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3256 Node* clr_arg = argument(1); 3257 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0))); 3258 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne)); 3259 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3260 3261 // Second fast path: ... else if (!clear_int) return true; 3262 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg)); 3263 result_rgn->init_req(no_clear_result_path, false_arg); 3264 result_val->init_req(no_clear_result_path, intcon(1)); 3265 3266 // drop through to next case 3267 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg))); 3268 3269 // (d) Otherwise, go to the slow path. 3270 slow_region->add_req(control()); 3271 set_control( _gvn.transform(slow_region)); 3272 3273 if (stopped()) { 3274 // There is no slow path. 3275 result_rgn->init_req(slow_result_path, top()); 3276 result_val->init_req(slow_result_path, top()); 3277 } else { 3278 // non-virtual because it is a private non-static 3279 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3280 3281 Node* slow_val = set_results_for_java_call(slow_call); 3282 // this->control() comes from set_results_for_java_call 3283 3284 Node* fast_io = slow_call->in(TypeFunc::I_O); 3285 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3286 3287 // These two phis are pre-filled with copies of of the fast IO and Memory 3288 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3289 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3290 3291 result_rgn->init_req(slow_result_path, control()); 3292 result_io ->init_req(slow_result_path, i_o()); 3293 result_mem->init_req(slow_result_path, reset_memory()); 3294 result_val->init_req(slow_result_path, slow_val); 3295 3296 set_all_memory(_gvn.transform(result_mem)); 3297 set_i_o( _gvn.transform(result_io)); 3298 } 3299 3300 C->set_has_split_ifs(true); // Has chance for split-if optimization 3301 set_result(result_rgn, result_val); 3302 return true; 3303 } 3304 3305 //---------------------------load_mirror_from_klass---------------------------- 3306 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3307 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3308 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3309 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3310 } 3311 3312 //-----------------------load_klass_from_mirror_common------------------------- 3313 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3314 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3315 // and branch to the given path on the region. 3316 // If never_see_null, take an uncommon trap on null, so we can optimistically 3317 // compile for the non-null case. 3318 // If the region is NULL, force never_see_null = true. 3319 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3320 bool never_see_null, 3321 RegionNode* region, 3322 int null_path, 3323 int offset) { 3324 if (region == NULL) never_see_null = true; 3325 Node* p = basic_plus_adr(mirror, offset); 3326 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3327 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3328 Node* null_ctl = top(); 3329 kls = null_check_oop(kls, &null_ctl, never_see_null); 3330 if (region != NULL) { 3331 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3332 region->init_req(null_path, null_ctl); 3333 } else { 3334 assert(null_ctl == top(), "no loose ends"); 3335 } 3336 return kls; 3337 } 3338 3339 //--------------------(inline_native_Class_query helpers)--------------------- 3340 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3341 // Fall through if (mods & mask) == bits, take the guard otherwise. 3342 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3343 // Branch around if the given klass has the given modifier bit set. 3344 // Like generate_guard, adds a new path onto the region. 3345 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3346 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3347 Node* mask = intcon(modifier_mask); 3348 Node* bits = intcon(modifier_bits); 3349 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask)); 3350 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits)); 3351 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 3352 return generate_fair_guard(bol, region); 3353 } 3354 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3355 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3356 } 3357 3358 //-------------------------inline_native_Class_query------------------- 3359 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3360 const Type* return_type = TypeInt::BOOL; 3361 Node* prim_return_value = top(); // what happens if it's a primitive class? 3362 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3363 bool expect_prim = false; // most of these guys expect to work on refs 3364 3365 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3366 3367 Node* mirror = argument(0); 3368 Node* obj = top(); 3369 3370 switch (id) { 3371 case vmIntrinsics::_isInstance: 3372 // nothing is an instance of a primitive type 3373 prim_return_value = intcon(0); 3374 obj = argument(1); 3375 break; 3376 case vmIntrinsics::_getModifiers: 3377 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3378 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3379 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3380 break; 3381 case vmIntrinsics::_isInterface: 3382 prim_return_value = intcon(0); 3383 break; 3384 case vmIntrinsics::_isArray: 3385 prim_return_value = intcon(0); 3386 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3387 break; 3388 case vmIntrinsics::_isPrimitive: 3389 prim_return_value = intcon(1); 3390 expect_prim = true; // obviously 3391 break; 3392 case vmIntrinsics::_getSuperclass: 3393 prim_return_value = null(); 3394 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3395 break; 3396 case vmIntrinsics::_getComponentType: 3397 prim_return_value = null(); 3398 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3399 break; 3400 case vmIntrinsics::_getClassAccessFlags: 3401 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3402 return_type = TypeInt::INT; // not bool! 6297094 3403 break; 3404 default: 3405 fatal_unexpected_iid(id); 3406 break; 3407 } 3408 3409 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3410 if (mirror_con == NULL) return false; // cannot happen? 3411 3412 #ifndef PRODUCT 3413 if (C->print_intrinsics() || C->print_inlining()) { 3414 ciType* k = mirror_con->java_mirror_type(); 3415 if (k) { 3416 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3417 k->print_name(); 3418 tty->cr(); 3419 } 3420 } 3421 #endif 3422 3423 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3424 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3425 record_for_igvn(region); 3426 PhiNode* phi = new (C) PhiNode(region, return_type); 3427 3428 // The mirror will never be null of Reflection.getClassAccessFlags, however 3429 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3430 // if it is. See bug 4774291. 3431 3432 // For Reflection.getClassAccessFlags(), the null check occurs in 3433 // the wrong place; see inline_unsafe_access(), above, for a similar 3434 // situation. 3435 mirror = null_check(mirror); 3436 // If mirror or obj is dead, only null-path is taken. 3437 if (stopped()) return true; 3438 3439 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3440 3441 // Now load the mirror's klass metaobject, and null-check it. 3442 // Side-effects region with the control path if the klass is null. 3443 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3444 // If kls is null, we have a primitive mirror. 3445 phi->init_req(_prim_path, prim_return_value); 3446 if (stopped()) { set_result(region, phi); return true; } 3447 bool safe_for_replace = (region->in(_prim_path) == top()); 3448 3449 Node* p; // handy temp 3450 Node* null_ctl; 3451 3452 // Now that we have the non-null klass, we can perform the real query. 3453 // For constant classes, the query will constant-fold in LoadNode::Value. 3454 Node* query_value = top(); 3455 switch (id) { 3456 case vmIntrinsics::_isInstance: 3457 // nothing is an instance of a primitive type 3458 query_value = gen_instanceof(obj, kls, safe_for_replace); 3459 break; 3460 3461 case vmIntrinsics::_getModifiers: 3462 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3463 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3464 break; 3465 3466 case vmIntrinsics::_isInterface: 3467 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3468 if (generate_interface_guard(kls, region) != NULL) 3469 // A guard was added. If the guard is taken, it was an interface. 3470 phi->add_req(intcon(1)); 3471 // If we fall through, it's a plain class. 3472 query_value = intcon(0); 3473 break; 3474 3475 case vmIntrinsics::_isArray: 3476 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3477 if (generate_array_guard(kls, region) != NULL) 3478 // A guard was added. If the guard is taken, it was an array. 3479 phi->add_req(intcon(1)); 3480 // If we fall through, it's a plain class. 3481 query_value = intcon(0); 3482 break; 3483 3484 case vmIntrinsics::_isPrimitive: 3485 query_value = intcon(0); // "normal" path produces false 3486 break; 3487 3488 case vmIntrinsics::_getSuperclass: 3489 // The rules here are somewhat unfortunate, but we can still do better 3490 // with random logic than with a JNI call. 3491 // Interfaces store null or Object as _super, but must report null. 3492 // Arrays store an intermediate super as _super, but must report Object. 3493 // Other types can report the actual _super. 3494 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3495 if (generate_interface_guard(kls, region) != NULL) 3496 // A guard was added. If the guard is taken, it was an interface. 3497 phi->add_req(null()); 3498 if (generate_array_guard(kls, region) != NULL) 3499 // A guard was added. If the guard is taken, it was an array. 3500 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3501 // If we fall through, it's a plain class. Get its _super. 3502 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3503 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3504 null_ctl = top(); 3505 kls = null_check_oop(kls, &null_ctl); 3506 if (null_ctl != top()) { 3507 // If the guard is taken, Object.superClass is null (both klass and mirror). 3508 region->add_req(null_ctl); 3509 phi ->add_req(null()); 3510 } 3511 if (!stopped()) { 3512 query_value = load_mirror_from_klass(kls); 3513 } 3514 break; 3515 3516 case vmIntrinsics::_getComponentType: 3517 if (generate_array_guard(kls, region) != NULL) { 3518 // Be sure to pin the oop load to the guard edge just created: 3519 Node* is_array_ctrl = region->in(region->req()-1); 3520 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); 3521 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3522 phi->add_req(cmo); 3523 } 3524 query_value = null(); // non-array case is null 3525 break; 3526 3527 case vmIntrinsics::_getClassAccessFlags: 3528 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3529 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3530 break; 3531 3532 default: 3533 fatal_unexpected_iid(id); 3534 break; 3535 } 3536 3537 // Fall-through is the normal case of a query to a real class. 3538 phi->init_req(1, query_value); 3539 region->init_req(1, control()); 3540 3541 C->set_has_split_ifs(true); // Has chance for split-if optimization 3542 set_result(region, phi); 3543 return true; 3544 } 3545 3546 //--------------------------inline_native_subtype_check------------------------ 3547 // This intrinsic takes the JNI calls out of the heart of 3548 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3549 bool LibraryCallKit::inline_native_subtype_check() { 3550 // Pull both arguments off the stack. 3551 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3552 args[0] = argument(0); 3553 args[1] = argument(1); 3554 Node* klasses[2]; // corresponding Klasses: superk, subk 3555 klasses[0] = klasses[1] = top(); 3556 3557 enum { 3558 // A full decision tree on {superc is prim, subc is prim}: 3559 _prim_0_path = 1, // {P,N} => false 3560 // {P,P} & superc!=subc => false 3561 _prim_same_path, // {P,P} & superc==subc => true 3562 _prim_1_path, // {N,P} => false 3563 _ref_subtype_path, // {N,N} & subtype check wins => true 3564 _both_ref_path, // {N,N} & subtype check loses => false 3565 PATH_LIMIT 3566 }; 3567 3568 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3569 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 3570 record_for_igvn(region); 3571 3572 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3573 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3574 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3575 3576 // First null-check both mirrors and load each mirror's klass metaobject. 3577 int which_arg; 3578 for (which_arg = 0; which_arg <= 1; which_arg++) { 3579 Node* arg = args[which_arg]; 3580 arg = null_check(arg); 3581 if (stopped()) break; 3582 args[which_arg] = arg; 3583 3584 Node* p = basic_plus_adr(arg, class_klass_offset); 3585 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3586 klasses[which_arg] = _gvn.transform(kls); 3587 } 3588 3589 // Having loaded both klasses, test each for null. 3590 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3591 for (which_arg = 0; which_arg <= 1; which_arg++) { 3592 Node* kls = klasses[which_arg]; 3593 Node* null_ctl = top(); 3594 kls = null_check_oop(kls, &null_ctl, never_see_null); 3595 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3596 region->init_req(prim_path, null_ctl); 3597 if (stopped()) break; 3598 klasses[which_arg] = kls; 3599 } 3600 3601 if (!stopped()) { 3602 // now we have two reference types, in klasses[0..1] 3603 Node* subk = klasses[1]; // the argument to isAssignableFrom 3604 Node* superk = klasses[0]; // the receiver 3605 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3606 // now we have a successful reference subtype check 3607 region->set_req(_ref_subtype_path, control()); 3608 } 3609 3610 // If both operands are primitive (both klasses null), then 3611 // we must return true when they are identical primitives. 3612 // It is convenient to test this after the first null klass check. 3613 set_control(region->in(_prim_0_path)); // go back to first null check 3614 if (!stopped()) { 3615 // Since superc is primitive, make a guard for the superc==subc case. 3616 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1])); 3617 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq)); 3618 generate_guard(bol_eq, region, PROB_FAIR); 3619 if (region->req() == PATH_LIMIT+1) { 3620 // A guard was added. If the added guard is taken, superc==subc. 3621 region->swap_edges(PATH_LIMIT, _prim_same_path); 3622 region->del_req(PATH_LIMIT); 3623 } 3624 region->set_req(_prim_0_path, control()); // Not equal after all. 3625 } 3626 3627 // these are the only paths that produce 'true': 3628 phi->set_req(_prim_same_path, intcon(1)); 3629 phi->set_req(_ref_subtype_path, intcon(1)); 3630 3631 // pull together the cases: 3632 assert(region->req() == PATH_LIMIT, "sane region"); 3633 for (uint i = 1; i < region->req(); i++) { 3634 Node* ctl = region->in(i); 3635 if (ctl == NULL || ctl == top()) { 3636 region->set_req(i, top()); 3637 phi ->set_req(i, top()); 3638 } else if (phi->in(i) == NULL) { 3639 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3640 } 3641 } 3642 3643 set_control(_gvn.transform(region)); 3644 set_result(_gvn.transform(phi)); 3645 return true; 3646 } 3647 3648 //---------------------generate_array_guard_common------------------------ 3649 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3650 bool obj_array, bool not_array) { 3651 // If obj_array/non_array==false/false: 3652 // Branch around if the given klass is in fact an array (either obj or prim). 3653 // If obj_array/non_array==false/true: 3654 // Branch around if the given klass is not an array klass of any kind. 3655 // If obj_array/non_array==true/true: 3656 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3657 // If obj_array/non_array==true/false: 3658 // Branch around if the kls is an oop array (Object[] or subtype) 3659 // 3660 // Like generate_guard, adds a new path onto the region. 3661 jint layout_con = 0; 3662 Node* layout_val = get_layout_helper(kls, layout_con); 3663 if (layout_val == NULL) { 3664 bool query = (obj_array 3665 ? Klass::layout_helper_is_objArray(layout_con) 3666 : Klass::layout_helper_is_array(layout_con)); 3667 if (query == not_array) { 3668 return NULL; // never a branch 3669 } else { // always a branch 3670 Node* always_branch = control(); 3671 if (region != NULL) 3672 region->add_req(always_branch); 3673 set_control(top()); 3674 return always_branch; 3675 } 3676 } 3677 // Now test the correct condition. 3678 jint nval = (obj_array 3679 ? ((jint)Klass::_lh_array_tag_type_value 3680 << Klass::_lh_array_tag_shift) 3681 : Klass::_lh_neutral_value); 3682 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval))); 3683 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3684 // invert the test if we are looking for a non-array 3685 if (not_array) btest = BoolTest(btest).negate(); 3686 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest)); 3687 return generate_fair_guard(bol, region); 3688 } 3689 3690 3691 //-----------------------inline_native_newArray-------------------------- 3692 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3693 bool LibraryCallKit::inline_native_newArray() { 3694 Node* mirror = argument(0); 3695 Node* count_val = argument(1); 3696 3697 mirror = null_check(mirror); 3698 // If mirror or obj is dead, only null-path is taken. 3699 if (stopped()) return true; 3700 3701 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3702 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3703 PhiNode* result_val = new(C) PhiNode(result_reg, 3704 TypeInstPtr::NOTNULL); 3705 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3706 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3707 TypePtr::BOTTOM); 3708 3709 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3710 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3711 result_reg, _slow_path); 3712 Node* normal_ctl = control(); 3713 Node* no_array_ctl = result_reg->in(_slow_path); 3714 3715 // Generate code for the slow case. We make a call to newArray(). 3716 set_control(no_array_ctl); 3717 if (!stopped()) { 3718 // Either the input type is void.class, or else the 3719 // array klass has not yet been cached. Either the 3720 // ensuing call will throw an exception, or else it 3721 // will cache the array klass for next time. 3722 PreserveJVMState pjvms(this); 3723 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3724 Node* slow_result = set_results_for_java_call(slow_call); 3725 // this->control() comes from set_results_for_java_call 3726 result_reg->set_req(_slow_path, control()); 3727 result_val->set_req(_slow_path, slow_result); 3728 result_io ->set_req(_slow_path, i_o()); 3729 result_mem->set_req(_slow_path, reset_memory()); 3730 } 3731 3732 set_control(normal_ctl); 3733 if (!stopped()) { 3734 // Normal case: The array type has been cached in the java.lang.Class. 3735 // The following call works fine even if the array type is polymorphic. 3736 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3737 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3738 result_reg->init_req(_normal_path, control()); 3739 result_val->init_req(_normal_path, obj); 3740 result_io ->init_req(_normal_path, i_o()); 3741 result_mem->init_req(_normal_path, reset_memory()); 3742 } 3743 3744 // Return the combined state. 3745 set_i_o( _gvn.transform(result_io) ); 3746 set_all_memory( _gvn.transform(result_mem)); 3747 3748 C->set_has_split_ifs(true); // Has chance for split-if optimization 3749 set_result(result_reg, result_val); 3750 return true; 3751 } 3752 3753 //----------------------inline_native_getLength-------------------------- 3754 // public static native int java.lang.reflect.Array.getLength(Object array); 3755 bool LibraryCallKit::inline_native_getLength() { 3756 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3757 3758 Node* array = null_check(argument(0)); 3759 // If array is dead, only null-path is taken. 3760 if (stopped()) return true; 3761 3762 // Deoptimize if it is a non-array. 3763 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3764 3765 if (non_array != NULL) { 3766 PreserveJVMState pjvms(this); 3767 set_control(non_array); 3768 uncommon_trap(Deoptimization::Reason_intrinsic, 3769 Deoptimization::Action_maybe_recompile); 3770 } 3771 3772 // If control is dead, only non-array-path is taken. 3773 if (stopped()) return true; 3774 3775 // The works fine even if the array type is polymorphic. 3776 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3777 Node* result = load_array_length(array); 3778 3779 C->set_has_split_ifs(true); // Has chance for split-if optimization 3780 set_result(result); 3781 return true; 3782 } 3783 3784 //------------------------inline_array_copyOf---------------------------- 3785 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3786 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3787 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3788 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3789 3790 // Get the arguments. 3791 Node* original = argument(0); 3792 Node* start = is_copyOfRange? argument(1): intcon(0); 3793 Node* end = is_copyOfRange? argument(2): argument(1); 3794 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3795 3796 Node* newcopy; 3797 3798 // Set the original stack and the reexecute bit for the interpreter to reexecute 3799 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3800 { PreserveReexecuteState preexecs(this); 3801 jvms()->set_should_reexecute(true); 3802 3803 array_type_mirror = null_check(array_type_mirror); 3804 original = null_check(original); 3805 3806 // Check if a null path was taken unconditionally. 3807 if (stopped()) return true; 3808 3809 Node* orig_length = load_array_length(original); 3810 3811 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3812 klass_node = null_check(klass_node); 3813 3814 RegionNode* bailout = new (C) RegionNode(1); 3815 record_for_igvn(bailout); 3816 3817 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3818 // Bail out if that is so. 3819 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3820 if (not_objArray != NULL) { 3821 // Improve the klass node's type from the new optimistic assumption: 3822 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3823 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3824 Node* cast = new (C) CastPPNode(klass_node, akls); 3825 cast->init_req(0, control()); 3826 klass_node = _gvn.transform(cast); 3827 } 3828 3829 // Bail out if either start or end is negative. 3830 generate_negative_guard(start, bailout, &start); 3831 generate_negative_guard(end, bailout, &end); 3832 3833 Node* length = end; 3834 if (_gvn.type(start) != TypeInt::ZERO) { 3835 length = _gvn.transform(new (C) SubINode(end, start)); 3836 } 3837 3838 // Bail out if length is negative. 3839 // Without this the new_array would throw 3840 // NegativeArraySizeException but IllegalArgumentException is what 3841 // should be thrown 3842 generate_negative_guard(length, bailout, &length); 3843 3844 if (bailout->req() > 1) { 3845 PreserveJVMState pjvms(this); 3846 set_control(_gvn.transform(bailout)); 3847 uncommon_trap(Deoptimization::Reason_intrinsic, 3848 Deoptimization::Action_maybe_recompile); 3849 } 3850 3851 if (!stopped()) { 3852 // How many elements will we copy from the original? 3853 // The answer is MinI(orig_length - start, length). 3854 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start)); 3855 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3856 3857 newcopy = new_array(klass_node, length, 0); // no argments to push 3858 3859 // Generate a direct call to the right arraycopy function(s). 3860 // We know the copy is disjoint but we might not know if the 3861 // oop stores need checking. 3862 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3863 // This will fail a store-check if x contains any non-nulls. 3864 bool disjoint_bases = true; 3865 // if start > orig_length then the length of the copy may be 3866 // negative. 3867 bool length_never_negative = !is_copyOfRange; 3868 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3869 original, start, newcopy, intcon(0), moved, 3870 disjoint_bases, length_never_negative); 3871 } 3872 } // original reexecute is set back here 3873 3874 C->set_has_split_ifs(true); // Has chance for split-if optimization 3875 if (!stopped()) { 3876 set_result(newcopy); 3877 } 3878 return true; 3879 } 3880 3881 3882 //----------------------generate_virtual_guard--------------------------- 3883 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3884 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3885 RegionNode* slow_region) { 3886 ciMethod* method = callee(); 3887 int vtable_index = method->vtable_index(); 3888 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3889 err_msg_res("bad index %d", vtable_index)); 3890 // Get the Method* out of the appropriate vtable entry. 3891 int entry_offset = (InstanceKlass::vtable_start_offset() + 3892 vtable_index*vtableEntry::size()) * wordSize + 3893 vtableEntry::method_offset_in_bytes(); 3894 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3895 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3896 3897 // Compare the target method with the expected method (e.g., Object.hashCode). 3898 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3899 3900 Node* native_call = makecon(native_call_addr); 3901 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call)); 3902 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne)); 3903 3904 return generate_slow_guard(test_native, slow_region); 3905 } 3906 3907 //-----------------------generate_method_call---------------------------- 3908 // Use generate_method_call to make a slow-call to the real 3909 // method if the fast path fails. An alternative would be to 3910 // use a stub like OptoRuntime::slow_arraycopy_Java. 3911 // This only works for expanding the current library call, 3912 // not another intrinsic. (E.g., don't use this for making an 3913 // arraycopy call inside of the copyOf intrinsic.) 3914 CallJavaNode* 3915 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3916 // When compiling the intrinsic method itself, do not use this technique. 3917 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3918 3919 ciMethod* method = callee(); 3920 // ensure the JVMS we have will be correct for this call 3921 guarantee(method_id == method->intrinsic_id(), "must match"); 3922 3923 const TypeFunc* tf = TypeFunc::make(method); 3924 CallJavaNode* slow_call; 3925 if (is_static) { 3926 assert(!is_virtual, ""); 3927 slow_call = new(C) CallStaticJavaNode(C, tf, 3928 SharedRuntime::get_resolve_static_call_stub(), 3929 method, bci()); 3930 } else if (is_virtual) { 3931 null_check_receiver(); 3932 int vtable_index = Method::invalid_vtable_index; 3933 if (UseInlineCaches) { 3934 // Suppress the vtable call 3935 } else { 3936 // hashCode and clone are not a miranda methods, 3937 // so the vtable index is fixed. 3938 // No need to use the linkResolver to get it. 3939 vtable_index = method->vtable_index(); 3940 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3941 err_msg_res("bad index %d", vtable_index)); 3942 } 3943 slow_call = new(C) CallDynamicJavaNode(tf, 3944 SharedRuntime::get_resolve_virtual_call_stub(), 3945 method, vtable_index, bci()); 3946 } else { // neither virtual nor static: opt_virtual 3947 null_check_receiver(); 3948 slow_call = new(C) CallStaticJavaNode(C, tf, 3949 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3950 method, bci()); 3951 slow_call->set_optimized_virtual(true); 3952 } 3953 set_arguments_for_java_call(slow_call); 3954 set_edges_for_java_call(slow_call); 3955 return slow_call; 3956 } 3957 3958 3959 //------------------------------inline_native_hashcode-------------------- 3960 // Build special case code for calls to hashCode on an object. 3961 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3962 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3963 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3964 3965 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3966 3967 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3968 PhiNode* result_val = new(C) PhiNode(result_reg, 3969 TypeInt::INT); 3970 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3971 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3972 TypePtr::BOTTOM); 3973 Node* obj = NULL; 3974 if (!is_static) { 3975 // Check for hashing null object 3976 obj = null_check_receiver(); 3977 if (stopped()) return true; // unconditionally null 3978 result_reg->init_req(_null_path, top()); 3979 result_val->init_req(_null_path, top()); 3980 } else { 3981 // Do a null check, and return zero if null. 3982 // System.identityHashCode(null) == 0 3983 obj = argument(0); 3984 Node* null_ctl = top(); 3985 obj = null_check_oop(obj, &null_ctl); 3986 result_reg->init_req(_null_path, null_ctl); 3987 result_val->init_req(_null_path, _gvn.intcon(0)); 3988 } 3989 3990 // Unconditionally null? Then return right away. 3991 if (stopped()) { 3992 set_control( result_reg->in(_null_path)); 3993 if (!stopped()) 3994 set_result(result_val->in(_null_path)); 3995 return true; 3996 } 3997 3998 // After null check, get the object's klass. 3999 Node* obj_klass = load_object_klass(obj); 4000 4001 // This call may be virtual (invokevirtual) or bound (invokespecial). 4002 // For each case we generate slightly different code. 4003 4004 // We only go to the fast case code if we pass a number of guards. The 4005 // paths which do not pass are accumulated in the slow_region. 4006 RegionNode* slow_region = new (C) RegionNode(1); 4007 record_for_igvn(slow_region); 4008 4009 // If this is a virtual call, we generate a funny guard. We pull out 4010 // the vtable entry corresponding to hashCode() from the target object. 4011 // If the target method which we are calling happens to be the native 4012 // Object hashCode() method, we pass the guard. We do not need this 4013 // guard for non-virtual calls -- the caller is known to be the native 4014 // Object hashCode(). 4015 if (is_virtual) { 4016 generate_virtual_guard(obj_klass, slow_region); 4017 } 4018 4019 // Get the header out of the object, use LoadMarkNode when available 4020 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4021 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4022 4023 // Test the header to see if it is unlocked. 4024 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4025 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask)); 4026 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4027 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val)); 4028 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne)); 4029 4030 generate_slow_guard(test_unlocked, slow_region); 4031 4032 // Get the hash value and check to see that it has been properly assigned. 4033 // We depend on hash_mask being at most 32 bits and avoid the use of 4034 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4035 // vm: see markOop.hpp. 4036 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4037 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4038 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift)); 4039 // This hack lets the hash bits live anywhere in the mark object now, as long 4040 // as the shift drops the relevant bits into the low 32 bits. Note that 4041 // Java spec says that HashCode is an int so there's no point in capturing 4042 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4043 hshifted_header = ConvX2I(hshifted_header); 4044 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask)); 4045 4046 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4047 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val)); 4048 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq)); 4049 4050 generate_slow_guard(test_assigned, slow_region); 4051 4052 Node* init_mem = reset_memory(); 4053 // fill in the rest of the null path: 4054 result_io ->init_req(_null_path, i_o()); 4055 result_mem->init_req(_null_path, init_mem); 4056 4057 result_val->init_req(_fast_path, hash_val); 4058 result_reg->init_req(_fast_path, control()); 4059 result_io ->init_req(_fast_path, i_o()); 4060 result_mem->init_req(_fast_path, init_mem); 4061 4062 // Generate code for the slow case. We make a call to hashCode(). 4063 set_control(_gvn.transform(slow_region)); 4064 if (!stopped()) { 4065 // No need for PreserveJVMState, because we're using up the present state. 4066 set_all_memory(init_mem); 4067 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4068 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4069 Node* slow_result = set_results_for_java_call(slow_call); 4070 // this->control() comes from set_results_for_java_call 4071 result_reg->init_req(_slow_path, control()); 4072 result_val->init_req(_slow_path, slow_result); 4073 result_io ->set_req(_slow_path, i_o()); 4074 result_mem ->set_req(_slow_path, reset_memory()); 4075 } 4076 4077 // Return the combined state. 4078 set_i_o( _gvn.transform(result_io) ); 4079 set_all_memory( _gvn.transform(result_mem)); 4080 4081 set_result(result_reg, result_val); 4082 return true; 4083 } 4084 4085 //---------------------------inline_native_getClass---------------------------- 4086 // public final native Class<?> java.lang.Object.getClass(); 4087 // 4088 // Build special case code for calls to getClass on an object. 4089 bool LibraryCallKit::inline_native_getClass() { 4090 Node* obj = null_check_receiver(); 4091 if (stopped()) return true; 4092 set_result(load_mirror_from_klass(load_object_klass(obj))); 4093 return true; 4094 } 4095 4096 //-----------------inline_native_Reflection_getCallerClass--------------------- 4097 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4098 // 4099 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4100 // 4101 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4102 // in that it must skip particular security frames and checks for 4103 // caller sensitive methods. 4104 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4105 #ifndef PRODUCT 4106 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4107 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4108 } 4109 #endif 4110 4111 if (!jvms()->has_method()) { 4112 #ifndef PRODUCT 4113 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4114 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4115 } 4116 #endif 4117 return false; 4118 } 4119 4120 // Walk back up the JVM state to find the caller at the required 4121 // depth. 4122 JVMState* caller_jvms = jvms(); 4123 4124 // Cf. JVM_GetCallerClass 4125 // NOTE: Start the loop at depth 1 because the current JVM state does 4126 // not include the Reflection.getCallerClass() frame. 4127 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4128 ciMethod* m = caller_jvms->method(); 4129 switch (n) { 4130 case 0: 4131 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4132 break; 4133 case 1: 4134 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4135 if (!m->caller_sensitive()) { 4136 #ifndef PRODUCT 4137 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4138 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4139 } 4140 #endif 4141 return false; // bail-out; let JVM_GetCallerClass do the work 4142 } 4143 break; 4144 default: 4145 if (!m->is_ignored_by_security_stack_walk()) { 4146 // We have reached the desired frame; return the holder class. 4147 // Acquire method holder as java.lang.Class and push as constant. 4148 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4149 ciInstance* caller_mirror = caller_klass->java_mirror(); 4150 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4151 4152 #ifndef PRODUCT 4153 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4154 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()); 4155 tty->print_cr(" JVM state at this point:"); 4156 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4157 ciMethod* m = jvms()->of_depth(i)->method(); 4158 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4159 } 4160 } 4161 #endif 4162 return true; 4163 } 4164 break; 4165 } 4166 } 4167 4168 #ifndef PRODUCT 4169 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4170 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4171 tty->print_cr(" JVM state at this point:"); 4172 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4173 ciMethod* m = jvms()->of_depth(i)->method(); 4174 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4175 } 4176 } 4177 #endif 4178 4179 return false; // bail-out; let JVM_GetCallerClass do the work 4180 } 4181 4182 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4183 Node* arg = argument(0); 4184 Node* result; 4185 4186 switch (id) { 4187 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break; 4188 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break; 4189 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break; 4190 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break; 4191 4192 case vmIntrinsics::_doubleToLongBits: { 4193 // two paths (plus control) merge in a wood 4194 RegionNode *r = new (C) RegionNode(3); 4195 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 4196 4197 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg)); 4198 // Build the boolean node 4199 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4200 4201 // Branch either way. 4202 // NaN case is less traveled, which makes all the difference. 4203 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4204 Node *opt_isnan = _gvn.transform(ifisnan); 4205 assert( opt_isnan->is_If(), "Expect an IfNode"); 4206 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4207 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4208 4209 set_control(iftrue); 4210 4211 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4212 Node *slow_result = longcon(nan_bits); // return NaN 4213 phi->init_req(1, _gvn.transform( slow_result )); 4214 r->init_req(1, iftrue); 4215 4216 // Else fall through 4217 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4218 set_control(iffalse); 4219 4220 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg))); 4221 r->init_req(2, iffalse); 4222 4223 // Post merge 4224 set_control(_gvn.transform(r)); 4225 record_for_igvn(r); 4226 4227 C->set_has_split_ifs(true); // Has chance for split-if optimization 4228 result = phi; 4229 assert(result->bottom_type()->isa_long(), "must be"); 4230 break; 4231 } 4232 4233 case vmIntrinsics::_floatToIntBits: { 4234 // two paths (plus control) merge in a wood 4235 RegionNode *r = new (C) RegionNode(3); 4236 Node *phi = new (C) PhiNode(r, TypeInt::INT); 4237 4238 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg)); 4239 // Build the boolean node 4240 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4241 4242 // Branch either way. 4243 // NaN case is less traveled, which makes all the difference. 4244 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4245 Node *opt_isnan = _gvn.transform(ifisnan); 4246 assert( opt_isnan->is_If(), "Expect an IfNode"); 4247 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4248 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4249 4250 set_control(iftrue); 4251 4252 static const jint nan_bits = 0x7fc00000; 4253 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4254 phi->init_req(1, _gvn.transform( slow_result )); 4255 r->init_req(1, iftrue); 4256 4257 // Else fall through 4258 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4259 set_control(iffalse); 4260 4261 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg))); 4262 r->init_req(2, iffalse); 4263 4264 // Post merge 4265 set_control(_gvn.transform(r)); 4266 record_for_igvn(r); 4267 4268 C->set_has_split_ifs(true); // Has chance for split-if optimization 4269 result = phi; 4270 assert(result->bottom_type()->isa_int(), "must be"); 4271 break; 4272 } 4273 4274 default: 4275 fatal_unexpected_iid(id); 4276 break; 4277 } 4278 set_result(_gvn.transform(result)); 4279 return true; 4280 } 4281 4282 #ifdef _LP64 4283 #define XTOP ,top() /*additional argument*/ 4284 #else //_LP64 4285 #define XTOP /*no additional argument*/ 4286 #endif //_LP64 4287 4288 //----------------------inline_unsafe_copyMemory------------------------- 4289 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4290 bool LibraryCallKit::inline_unsafe_copyMemory() { 4291 if (callee()->is_static()) return false; // caller must have the capability! 4292 null_check_receiver(); // null-check receiver 4293 if (stopped()) return true; 4294 4295 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4296 4297 Node* src_ptr = argument(1); // type: oop 4298 Node* src_off = ConvL2X(argument(2)); // type: long 4299 Node* dst_ptr = argument(4); // type: oop 4300 Node* dst_off = ConvL2X(argument(5)); // type: long 4301 Node* size = ConvL2X(argument(7)); // type: long 4302 4303 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4304 "fieldOffset must be byte-scaled"); 4305 4306 Node* src = make_unsafe_address(src_ptr, src_off); 4307 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4308 4309 // Conservatively insert a memory barrier on all memory slices. 4310 // Do not let writes of the copy source or destination float below the copy. 4311 insert_mem_bar(Op_MemBarCPUOrder); 4312 4313 // Call it. Note that the length argument is not scaled. 4314 make_runtime_call(RC_LEAF|RC_NO_FP, 4315 OptoRuntime::fast_arraycopy_Type(), 4316 StubRoutines::unsafe_arraycopy(), 4317 "unsafe_arraycopy", 4318 TypeRawPtr::BOTTOM, 4319 src, dst, size XTOP); 4320 4321 // Do not let reads of the copy destination float above the copy. 4322 insert_mem_bar(Op_MemBarCPUOrder); 4323 4324 return true; 4325 } 4326 4327 //------------------------clone_coping----------------------------------- 4328 // Helper function for inline_native_clone. 4329 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4330 assert(obj_size != NULL, ""); 4331 Node* raw_obj = alloc_obj->in(1); 4332 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4333 4334 AllocateNode* alloc = NULL; 4335 if (ReduceBulkZeroing) { 4336 // We will be completely responsible for initializing this object - 4337 // mark Initialize node as complete. 4338 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4339 // The object was just allocated - there should be no any stores! 4340 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4341 // Mark as complete_with_arraycopy so that on AllocateNode 4342 // expansion, we know this AllocateNode is initialized by an array 4343 // copy and a StoreStore barrier exists after the array copy. 4344 alloc->initialization()->set_complete_with_arraycopy(); 4345 } 4346 4347 // Copy the fastest available way. 4348 // TODO: generate fields copies for small objects instead. 4349 Node* src = obj; 4350 Node* dest = alloc_obj; 4351 Node* size = _gvn.transform(obj_size); 4352 4353 // Exclude the header but include array length to copy by 8 bytes words. 4354 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4355 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4356 instanceOopDesc::base_offset_in_bytes(); 4357 // base_off: 4358 // 8 - 32-bit VM 4359 // 12 - 64-bit VM, compressed klass 4360 // 16 - 64-bit VM, normal klass 4361 if (base_off % BytesPerLong != 0) { 4362 assert(UseCompressedClassPointers, ""); 4363 if (is_array) { 4364 // Exclude length to copy by 8 bytes words. 4365 base_off += sizeof(int); 4366 } else { 4367 // Include klass to copy by 8 bytes words. 4368 base_off = instanceOopDesc::klass_offset_in_bytes(); 4369 } 4370 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4371 } 4372 src = basic_plus_adr(src, base_off); 4373 dest = basic_plus_adr(dest, base_off); 4374 4375 // Compute the length also, if needed: 4376 Node* countx = size; 4377 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off))); 4378 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4379 4380 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4381 bool disjoint_bases = true; 4382 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4383 src, NULL, dest, NULL, countx, 4384 /*dest_uninitialized*/true); 4385 4386 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4387 if (card_mark) { 4388 assert(!is_array, ""); 4389 // Put in store barrier for any and all oops we are sticking 4390 // into this object. (We could avoid this if we could prove 4391 // that the object type contains no oop fields at all.) 4392 Node* no_particular_value = NULL; 4393 Node* no_particular_field = NULL; 4394 int raw_adr_idx = Compile::AliasIdxRaw; 4395 post_barrier(control(), 4396 memory(raw_adr_type), 4397 alloc_obj, 4398 no_particular_field, 4399 raw_adr_idx, 4400 no_particular_value, 4401 T_OBJECT, 4402 false); 4403 } 4404 4405 // Do not let reads from the cloned object float above the arraycopy. 4406 if (alloc != NULL) { 4407 // Do not let stores that initialize this object be reordered with 4408 // a subsequent store that would make this object accessible by 4409 // other threads. 4410 // Record what AllocateNode this StoreStore protects so that 4411 // escape analysis can go from the MemBarStoreStoreNode to the 4412 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4413 // based on the escape status of the AllocateNode. 4414 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4415 } else { 4416 insert_mem_bar(Op_MemBarCPUOrder); 4417 } 4418 } 4419 4420 //------------------------inline_native_clone---------------------------- 4421 // protected native Object java.lang.Object.clone(); 4422 // 4423 // Here are the simple edge cases: 4424 // null receiver => normal trap 4425 // virtual and clone was overridden => slow path to out-of-line clone 4426 // not cloneable or finalizer => slow path to out-of-line Object.clone 4427 // 4428 // The general case has two steps, allocation and copying. 4429 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4430 // 4431 // Copying also has two cases, oop arrays and everything else. 4432 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4433 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4434 // 4435 // These steps fold up nicely if and when the cloned object's klass 4436 // can be sharply typed as an object array, a type array, or an instance. 4437 // 4438 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4439 PhiNode* result_val; 4440 4441 // Set the reexecute bit for the interpreter to reexecute 4442 // the bytecode that invokes Object.clone if deoptimization happens. 4443 { PreserveReexecuteState preexecs(this); 4444 jvms()->set_should_reexecute(true); 4445 4446 Node* obj = null_check_receiver(); 4447 if (stopped()) return true; 4448 4449 Node* obj_klass = load_object_klass(obj); 4450 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4451 const TypeOopPtr* toop = ((tklass != NULL) 4452 ? tklass->as_instance_type() 4453 : TypeInstPtr::NOTNULL); 4454 4455 // Conservatively insert a memory barrier on all memory slices. 4456 // Do not let writes into the original float below the clone. 4457 insert_mem_bar(Op_MemBarCPUOrder); 4458 4459 // paths into result_reg: 4460 enum { 4461 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4462 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4463 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4464 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4465 PATH_LIMIT 4466 }; 4467 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4468 result_val = new(C) PhiNode(result_reg, 4469 TypeInstPtr::NOTNULL); 4470 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO); 4471 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 4472 TypePtr::BOTTOM); 4473 record_for_igvn(result_reg); 4474 4475 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4476 int raw_adr_idx = Compile::AliasIdxRaw; 4477 4478 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4479 if (array_ctl != NULL) { 4480 // It's an array. 4481 PreserveJVMState pjvms(this); 4482 set_control(array_ctl); 4483 Node* obj_length = load_array_length(obj); 4484 Node* obj_size = NULL; 4485 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4486 4487 if (!use_ReduceInitialCardMarks()) { 4488 // If it is an oop array, it requires very special treatment, 4489 // because card marking is required on each card of the array. 4490 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4491 if (is_obja != NULL) { 4492 PreserveJVMState pjvms2(this); 4493 set_control(is_obja); 4494 // Generate a direct call to the right arraycopy function(s). 4495 bool disjoint_bases = true; 4496 bool length_never_negative = true; 4497 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4498 obj, intcon(0), alloc_obj, intcon(0), 4499 obj_length, 4500 disjoint_bases, length_never_negative); 4501 result_reg->init_req(_objArray_path, control()); 4502 result_val->init_req(_objArray_path, alloc_obj); 4503 result_i_o ->set_req(_objArray_path, i_o()); 4504 result_mem ->set_req(_objArray_path, reset_memory()); 4505 } 4506 } 4507 // Otherwise, there are no card marks to worry about. 4508 // (We can dispense with card marks if we know the allocation 4509 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4510 // causes the non-eden paths to take compensating steps to 4511 // simulate a fresh allocation, so that no further 4512 // card marks are required in compiled code to initialize 4513 // the object.) 4514 4515 if (!stopped()) { 4516 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4517 4518 // Present the results of the copy. 4519 result_reg->init_req(_array_path, control()); 4520 result_val->init_req(_array_path, alloc_obj); 4521 result_i_o ->set_req(_array_path, i_o()); 4522 result_mem ->set_req(_array_path, reset_memory()); 4523 } 4524 } 4525 4526 // We only go to the instance fast case code if we pass a number of guards. 4527 // The paths which do not pass are accumulated in the slow_region. 4528 RegionNode* slow_region = new (C) RegionNode(1); 4529 record_for_igvn(slow_region); 4530 if (!stopped()) { 4531 // It's an instance (we did array above). Make the slow-path tests. 4532 // If this is a virtual call, we generate a funny guard. We grab 4533 // the vtable entry corresponding to clone() from the target object. 4534 // If the target method which we are calling happens to be the 4535 // Object clone() method, we pass the guard. We do not need this 4536 // guard for non-virtual calls; the caller is known to be the native 4537 // Object clone(). 4538 if (is_virtual) { 4539 generate_virtual_guard(obj_klass, slow_region); 4540 } 4541 4542 // The object must be cloneable and must not have a finalizer. 4543 // Both of these conditions may be checked in a single test. 4544 // We could optimize the cloneable test further, but we don't care. 4545 generate_access_flags_guard(obj_klass, 4546 // Test both conditions: 4547 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4548 // Must be cloneable but not finalizer: 4549 JVM_ACC_IS_CLONEABLE, 4550 slow_region); 4551 } 4552 4553 if (!stopped()) { 4554 // It's an instance, and it passed the slow-path tests. 4555 PreserveJVMState pjvms(this); 4556 Node* obj_size = NULL; 4557 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size); 4558 4559 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4560 4561 // Present the results of the slow call. 4562 result_reg->init_req(_instance_path, control()); 4563 result_val->init_req(_instance_path, alloc_obj); 4564 result_i_o ->set_req(_instance_path, i_o()); 4565 result_mem ->set_req(_instance_path, reset_memory()); 4566 } 4567 4568 // Generate code for the slow case. We make a call to clone(). 4569 set_control(_gvn.transform(slow_region)); 4570 if (!stopped()) { 4571 PreserveJVMState pjvms(this); 4572 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4573 Node* slow_result = set_results_for_java_call(slow_call); 4574 // this->control() comes from set_results_for_java_call 4575 result_reg->init_req(_slow_path, control()); 4576 result_val->init_req(_slow_path, slow_result); 4577 result_i_o ->set_req(_slow_path, i_o()); 4578 result_mem ->set_req(_slow_path, reset_memory()); 4579 } 4580 4581 // Return the combined state. 4582 set_control( _gvn.transform(result_reg)); 4583 set_i_o( _gvn.transform(result_i_o)); 4584 set_all_memory( _gvn.transform(result_mem)); 4585 } // original reexecute is set back here 4586 4587 set_result(_gvn.transform(result_val)); 4588 return true; 4589 } 4590 4591 //------------------------------basictype2arraycopy---------------------------- 4592 address LibraryCallKit::basictype2arraycopy(BasicType t, 4593 Node* src_offset, 4594 Node* dest_offset, 4595 bool disjoint_bases, 4596 const char* &name, 4597 bool dest_uninitialized) { 4598 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4599 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4600 4601 bool aligned = false; 4602 bool disjoint = disjoint_bases; 4603 4604 // if the offsets are the same, we can treat the memory regions as 4605 // disjoint, because either the memory regions are in different arrays, 4606 // or they are identical (which we can treat as disjoint.) We can also 4607 // treat a copy with a destination index less that the source index 4608 // as disjoint since a low->high copy will work correctly in this case. 4609 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4610 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4611 // both indices are constants 4612 int s_offs = src_offset_inttype->get_con(); 4613 int d_offs = dest_offset_inttype->get_con(); 4614 int element_size = type2aelembytes(t); 4615 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4616 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4617 if (s_offs >= d_offs) disjoint = true; 4618 } else if (src_offset == dest_offset && src_offset != NULL) { 4619 // This can occur if the offsets are identical non-constants. 4620 disjoint = true; 4621 } 4622 4623 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4624 } 4625 4626 4627 //------------------------------inline_arraycopy----------------------- 4628 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4629 // Object dest, int destPos, 4630 // int length); 4631 bool LibraryCallKit::inline_arraycopy() { 4632 // Get the arguments. 4633 Node* src = argument(0); // type: oop 4634 Node* src_offset = argument(1); // type: int 4635 Node* dest = argument(2); // type: oop 4636 Node* dest_offset = argument(3); // type: int 4637 Node* length = argument(4); // type: int 4638 4639 // Compile time checks. If any of these checks cannot be verified at compile time, 4640 // we do not make a fast path for this call. Instead, we let the call remain as it 4641 // is. The checks we choose to mandate at compile time are: 4642 // 4643 // (1) src and dest are arrays. 4644 const Type* src_type = src->Value(&_gvn); 4645 const Type* dest_type = dest->Value(&_gvn); 4646 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4647 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4648 4649 // Do we have the type of src? 4650 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4651 // Do we have the type of dest? 4652 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4653 // Is the type for src from speculation? 4654 bool src_spec = false; 4655 // Is the type for dest from speculation? 4656 bool dest_spec = false; 4657 4658 if (!has_src || !has_dest) { 4659 // We don't have sufficient type information, let's see if 4660 // speculative types can help. We need to have types for both src 4661 // and dest so that it pays off. 4662 4663 // Do we already have or could we have type information for src 4664 bool could_have_src = has_src; 4665 // Do we already have or could we have type information for dest 4666 bool could_have_dest = has_dest; 4667 4668 ciKlass* src_k = NULL; 4669 if (!has_src) { 4670 src_k = src_type->speculative_type(); 4671 if (src_k != NULL && src_k->is_array_klass()) { 4672 could_have_src = true; 4673 } 4674 } 4675 4676 ciKlass* dest_k = NULL; 4677 if (!has_dest) { 4678 dest_k = dest_type->speculative_type(); 4679 if (dest_k != NULL && dest_k->is_array_klass()) { 4680 could_have_dest = true; 4681 } 4682 } 4683 4684 if (could_have_src && could_have_dest) { 4685 // This is going to pay off so emit the required guards 4686 if (!has_src) { 4687 src = maybe_cast_profiled_obj(src, src_k); 4688 src_type = _gvn.type(src); 4689 top_src = src_type->isa_aryptr(); 4690 has_src = (top_src != NULL && top_src->klass() != NULL); 4691 src_spec = true; 4692 } 4693 if (!has_dest) { 4694 dest = maybe_cast_profiled_obj(dest, dest_k); 4695 dest_type = _gvn.type(dest); 4696 top_dest = dest_type->isa_aryptr(); 4697 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4698 dest_spec = true; 4699 } 4700 } 4701 } 4702 4703 if (!has_src || !has_dest) { 4704 // Conservatively insert a memory barrier on all memory slices. 4705 // Do not let writes into the source float below the arraycopy. 4706 insert_mem_bar(Op_MemBarCPUOrder); 4707 4708 // Call StubRoutines::generic_arraycopy stub. 4709 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4710 src, src_offset, dest, dest_offset, length); 4711 4712 // Do not let reads from the destination float above the arraycopy. 4713 // Since we cannot type the arrays, we don't know which slices 4714 // might be affected. We could restrict this barrier only to those 4715 // memory slices which pertain to array elements--but don't bother. 4716 if (!InsertMemBarAfterArraycopy) 4717 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4718 insert_mem_bar(Op_MemBarCPUOrder); 4719 return true; 4720 } 4721 4722 // (2) src and dest arrays must have elements of the same BasicType 4723 // Figure out the size and type of the elements we will be copying. 4724 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4725 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4726 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4727 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4728 4729 if (src_elem != dest_elem || dest_elem == T_VOID) { 4730 // The component types are not the same or are not recognized. Punt. 4731 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4732 generate_slow_arraycopy(TypePtr::BOTTOM, 4733 src, src_offset, dest, dest_offset, length, 4734 /*dest_uninitialized*/false); 4735 return true; 4736 } 4737 4738 if (src_elem == T_OBJECT) { 4739 // If both arrays are object arrays then having the exact types 4740 // for both will remove the need for a subtype check at runtime 4741 // before the call and may make it possible to pick a faster copy 4742 // routine (without a subtype check on every element) 4743 // Do we have the exact type of src? 4744 bool could_have_src = src_spec; 4745 // Do we have the exact type of dest? 4746 bool could_have_dest = dest_spec; 4747 ciKlass* src_k = top_src->klass(); 4748 ciKlass* dest_k = top_dest->klass(); 4749 if (!src_spec) { 4750 src_k = src_type->speculative_type(); 4751 if (src_k != NULL && src_k->is_array_klass()) { 4752 could_have_src = true; 4753 } 4754 } 4755 if (!dest_spec) { 4756 dest_k = dest_type->speculative_type(); 4757 if (dest_k != NULL && dest_k->is_array_klass()) { 4758 could_have_dest = true; 4759 } 4760 } 4761 if (could_have_src && could_have_dest) { 4762 // If we can have both exact types, emit the missing guards 4763 if (could_have_src && !src_spec) { 4764 src = maybe_cast_profiled_obj(src, src_k); 4765 } 4766 if (could_have_dest && !dest_spec) { 4767 dest = maybe_cast_profiled_obj(dest, dest_k); 4768 } 4769 } 4770 } 4771 4772 //--------------------------------------------------------------------------- 4773 // We will make a fast path for this call to arraycopy. 4774 4775 // We have the following tests left to perform: 4776 // 4777 // (3) src and dest must not be null. 4778 // (4) src_offset must not be negative. 4779 // (5) dest_offset must not be negative. 4780 // (6) length must not be negative. 4781 // (7) src_offset + length must not exceed length of src. 4782 // (8) dest_offset + length must not exceed length of dest. 4783 // (9) each element of an oop array must be assignable 4784 4785 RegionNode* slow_region = new (C) RegionNode(1); 4786 record_for_igvn(slow_region); 4787 4788 // (3) operands must not be null 4789 // We currently perform our null checks with the null_check routine. 4790 // This means that the null exceptions will be reported in the caller 4791 // rather than (correctly) reported inside of the native arraycopy call. 4792 // This should be corrected, given time. We do our null check with the 4793 // stack pointer restored. 4794 src = null_check(src, T_ARRAY); 4795 dest = null_check(dest, T_ARRAY); 4796 4797 // (4) src_offset must not be negative. 4798 generate_negative_guard(src_offset, slow_region); 4799 4800 // (5) dest_offset must not be negative. 4801 generate_negative_guard(dest_offset, slow_region); 4802 4803 // (6) length must not be negative (moved to generate_arraycopy()). 4804 // generate_negative_guard(length, slow_region); 4805 4806 // (7) src_offset + length must not exceed length of src. 4807 generate_limit_guard(src_offset, length, 4808 load_array_length(src), 4809 slow_region); 4810 4811 // (8) dest_offset + length must not exceed length of dest. 4812 generate_limit_guard(dest_offset, length, 4813 load_array_length(dest), 4814 slow_region); 4815 4816 // (9) each element of an oop array must be assignable 4817 // The generate_arraycopy subroutine checks this. 4818 4819 // This is where the memory effects are placed: 4820 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4821 generate_arraycopy(adr_type, dest_elem, 4822 src, src_offset, dest, dest_offset, length, 4823 false, false, slow_region); 4824 4825 return true; 4826 } 4827 4828 //-----------------------------generate_arraycopy---------------------- 4829 // Generate an optimized call to arraycopy. 4830 // Caller must guard against non-arrays. 4831 // Caller must determine a common array basic-type for both arrays. 4832 // Caller must validate offsets against array bounds. 4833 // The slow_region has already collected guard failure paths 4834 // (such as out of bounds length or non-conformable array types). 4835 // The generated code has this shape, in general: 4836 // 4837 // if (length == 0) return // via zero_path 4838 // slowval = -1 4839 // if (types unknown) { 4840 // slowval = call generic copy loop 4841 // if (slowval == 0) return // via checked_path 4842 // } else if (indexes in bounds) { 4843 // if ((is object array) && !(array type check)) { 4844 // slowval = call checked copy loop 4845 // if (slowval == 0) return // via checked_path 4846 // } else { 4847 // call bulk copy loop 4848 // return // via fast_path 4849 // } 4850 // } 4851 // // adjust params for remaining work: 4852 // if (slowval != -1) { 4853 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4854 // } 4855 // slow_region: 4856 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4857 // return // via slow_call_path 4858 // 4859 // This routine is used from several intrinsics: System.arraycopy, 4860 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4861 // 4862 void 4863 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4864 BasicType basic_elem_type, 4865 Node* src, Node* src_offset, 4866 Node* dest, Node* dest_offset, 4867 Node* copy_length, 4868 bool disjoint_bases, 4869 bool length_never_negative, 4870 RegionNode* slow_region) { 4871 4872 if (slow_region == NULL) { 4873 slow_region = new(C) RegionNode(1); 4874 record_for_igvn(slow_region); 4875 } 4876 4877 Node* original_dest = dest; 4878 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4879 bool dest_uninitialized = false; 4880 4881 // See if this is the initialization of a newly-allocated array. 4882 // If so, we will take responsibility here for initializing it to zero. 4883 // (Note: Because tightly_coupled_allocation performs checks on the 4884 // out-edges of the dest, we need to avoid making derived pointers 4885 // from it until we have checked its uses.) 4886 if (ReduceBulkZeroing 4887 && !ZeroTLAB // pointless if already zeroed 4888 && basic_elem_type != T_CONFLICT // avoid corner case 4889 && !src->eqv_uncast(dest) 4890 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4891 != NULL) 4892 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4893 && alloc->maybe_set_complete(&_gvn)) { 4894 // "You break it, you buy it." 4895 InitializeNode* init = alloc->initialization(); 4896 assert(init->is_complete(), "we just did this"); 4897 init->set_complete_with_arraycopy(); 4898 assert(dest->is_CheckCastPP(), "sanity"); 4899 assert(dest->in(0)->in(0) == init, "dest pinned"); 4900 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4901 // From this point on, every exit path is responsible for 4902 // initializing any non-copied parts of the object to zero. 4903 // Also, if this flag is set we make sure that arraycopy interacts properly 4904 // with G1, eliding pre-barriers. See CR 6627983. 4905 dest_uninitialized = true; 4906 } else { 4907 // No zeroing elimination here. 4908 alloc = NULL; 4909 //original_dest = dest; 4910 //dest_uninitialized = false; 4911 } 4912 4913 // Results are placed here: 4914 enum { fast_path = 1, // normal void-returning assembly stub 4915 checked_path = 2, // special assembly stub with cleanup 4916 slow_call_path = 3, // something went wrong; call the VM 4917 zero_path = 4, // bypass when length of copy is zero 4918 bcopy_path = 5, // copy primitive array by 64-bit blocks 4919 PATH_LIMIT = 6 4920 }; 4921 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT); 4922 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO); 4923 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type); 4924 record_for_igvn(result_region); 4925 _gvn.set_type_bottom(result_i_o); 4926 _gvn.set_type_bottom(result_memory); 4927 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4928 4929 // The slow_control path: 4930 Node* slow_control; 4931 Node* slow_i_o = i_o(); 4932 Node* slow_mem = memory(adr_type); 4933 debug_only(slow_control = (Node*) badAddress); 4934 4935 // Checked control path: 4936 Node* checked_control = top(); 4937 Node* checked_mem = NULL; 4938 Node* checked_i_o = NULL; 4939 Node* checked_value = NULL; 4940 4941 if (basic_elem_type == T_CONFLICT) { 4942 assert(!dest_uninitialized, ""); 4943 Node* cv = generate_generic_arraycopy(adr_type, 4944 src, src_offset, dest, dest_offset, 4945 copy_length, dest_uninitialized); 4946 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4947 checked_control = control(); 4948 checked_i_o = i_o(); 4949 checked_mem = memory(adr_type); 4950 checked_value = cv; 4951 set_control(top()); // no fast path 4952 } 4953 4954 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4955 if (not_pos != NULL) { 4956 PreserveJVMState pjvms(this); 4957 set_control(not_pos); 4958 4959 // (6) length must not be negative. 4960 if (!length_never_negative) { 4961 generate_negative_guard(copy_length, slow_region); 4962 } 4963 4964 // copy_length is 0. 4965 if (!stopped() && dest_uninitialized) { 4966 Node* dest_length = alloc->in(AllocateNode::ALength); 4967 if (copy_length->eqv_uncast(dest_length) 4968 || _gvn.find_int_con(dest_length, 1) <= 0) { 4969 // There is no zeroing to do. No need for a secondary raw memory barrier. 4970 } else { 4971 // Clear the whole thing since there are no source elements to copy. 4972 generate_clear_array(adr_type, dest, basic_elem_type, 4973 intcon(0), NULL, 4974 alloc->in(AllocateNode::AllocSize)); 4975 // Use a secondary InitializeNode as raw memory barrier. 4976 // Currently it is needed only on this path since other 4977 // paths have stub or runtime calls as raw memory barriers. 4978 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4979 Compile::AliasIdxRaw, 4980 top())->as_Initialize(); 4981 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4982 } 4983 } 4984 4985 // Present the results of the fast call. 4986 result_region->init_req(zero_path, control()); 4987 result_i_o ->init_req(zero_path, i_o()); 4988 result_memory->init_req(zero_path, memory(adr_type)); 4989 } 4990 4991 if (!stopped() && dest_uninitialized) { 4992 // We have to initialize the *uncopied* part of the array to zero. 4993 // The copy destination is the slice dest[off..off+len]. The other slices 4994 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4995 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4996 Node* dest_length = alloc->in(AllocateNode::ALength); 4997 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset, 4998 copy_length)); 4999 5000 // If there is a head section that needs zeroing, do it now. 5001 if (find_int_con(dest_offset, -1) != 0) { 5002 generate_clear_array(adr_type, dest, basic_elem_type, 5003 intcon(0), dest_offset, 5004 NULL); 5005 } 5006 5007 // Next, perform a dynamic check on the tail length. 5008 // It is often zero, and we can win big if we prove this. 5009 // There are two wins: Avoid generating the ClearArray 5010 // with its attendant messy index arithmetic, and upgrade 5011 // the copy to a more hardware-friendly word size of 64 bits. 5012 Node* tail_ctl = NULL; 5013 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 5014 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length)); 5015 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt)); 5016 tail_ctl = generate_slow_guard(bol_lt, NULL); 5017 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 5018 } 5019 5020 // At this point, let's assume there is no tail. 5021 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 5022 // There is no tail. Try an upgrade to a 64-bit copy. 5023 bool didit = false; 5024 { PreserveJVMState pjvms(this); 5025 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 5026 src, src_offset, dest, dest_offset, 5027 dest_size, dest_uninitialized); 5028 if (didit) { 5029 // Present the results of the block-copying fast call. 5030 result_region->init_req(bcopy_path, control()); 5031 result_i_o ->init_req(bcopy_path, i_o()); 5032 result_memory->init_req(bcopy_path, memory(adr_type)); 5033 } 5034 } 5035 if (didit) 5036 set_control(top()); // no regular fast path 5037 } 5038 5039 // Clear the tail, if any. 5040 if (tail_ctl != NULL) { 5041 Node* notail_ctl = stopped() ? NULL : control(); 5042 set_control(tail_ctl); 5043 if (notail_ctl == NULL) { 5044 generate_clear_array(adr_type, dest, basic_elem_type, 5045 dest_tail, NULL, 5046 dest_size); 5047 } else { 5048 // Make a local merge. 5049 Node* done_ctl = new(C) RegionNode(3); 5050 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type); 5051 done_ctl->init_req(1, notail_ctl); 5052 done_mem->init_req(1, memory(adr_type)); 5053 generate_clear_array(adr_type, dest, basic_elem_type, 5054 dest_tail, NULL, 5055 dest_size); 5056 done_ctl->init_req(2, control()); 5057 done_mem->init_req(2, memory(adr_type)); 5058 set_control( _gvn.transform(done_ctl)); 5059 set_memory( _gvn.transform(done_mem), adr_type ); 5060 } 5061 } 5062 } 5063 5064 BasicType copy_type = basic_elem_type; 5065 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 5066 if (!stopped() && copy_type == T_OBJECT) { 5067 // If src and dest have compatible element types, we can copy bits. 5068 // Types S[] and D[] are compatible if D is a supertype of S. 5069 // 5070 // If they are not, we will use checked_oop_disjoint_arraycopy, 5071 // which performs a fast optimistic per-oop check, and backs off 5072 // further to JVM_ArrayCopy on the first per-oop check that fails. 5073 // (Actually, we don't move raw bits only; the GC requires card marks.) 5074 5075 // Get the Klass* for both src and dest 5076 Node* src_klass = load_object_klass(src); 5077 Node* dest_klass = load_object_klass(dest); 5078 5079 // Generate the subtype check. 5080 // This might fold up statically, or then again it might not. 5081 // 5082 // Non-static example: Copying List<String>.elements to a new String[]. 5083 // The backing store for a List<String> is always an Object[], 5084 // but its elements are always type String, if the generic types 5085 // are correct at the source level. 5086 // 5087 // Test S[] against D[], not S against D, because (probably) 5088 // the secondary supertype cache is less busy for S[] than S. 5089 // This usually only matters when D is an interface. 5090 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5091 // Plug failing path into checked_oop_disjoint_arraycopy 5092 if (not_subtype_ctrl != top()) { 5093 PreserveJVMState pjvms(this); 5094 set_control(not_subtype_ctrl); 5095 // (At this point we can assume disjoint_bases, since types differ.) 5096 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 5097 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 5098 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 5099 Node* dest_elem_klass = _gvn.transform(n1); 5100 Node* cv = generate_checkcast_arraycopy(adr_type, 5101 dest_elem_klass, 5102 src, src_offset, dest, dest_offset, 5103 ConvI2X(copy_length), dest_uninitialized); 5104 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 5105 checked_control = control(); 5106 checked_i_o = i_o(); 5107 checked_mem = memory(adr_type); 5108 checked_value = cv; 5109 } 5110 // At this point we know we do not need type checks on oop stores. 5111 5112 // Let's see if we need card marks: 5113 if (alloc != NULL && use_ReduceInitialCardMarks()) { 5114 // If we do not need card marks, copy using the jint or jlong stub. 5115 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 5116 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 5117 "sizes agree"); 5118 } 5119 } 5120 5121 if (!stopped()) { 5122 // Generate the fast path, if possible. 5123 PreserveJVMState pjvms(this); 5124 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 5125 src, src_offset, dest, dest_offset, 5126 ConvI2X(copy_length), dest_uninitialized); 5127 5128 // Present the results of the fast call. 5129 result_region->init_req(fast_path, control()); 5130 result_i_o ->init_req(fast_path, i_o()); 5131 result_memory->init_req(fast_path, memory(adr_type)); 5132 } 5133 5134 // Here are all the slow paths up to this point, in one bundle: 5135 slow_control = top(); 5136 if (slow_region != NULL) 5137 slow_control = _gvn.transform(slow_region); 5138 DEBUG_ONLY(slow_region = (RegionNode*)badAddress); 5139 5140 set_control(checked_control); 5141 if (!stopped()) { 5142 // Clean up after the checked call. 5143 // The returned value is either 0 or -1^K, 5144 // where K = number of partially transferred array elements. 5145 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0))); 5146 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 5147 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 5148 5149 // If it is 0, we are done, so transfer to the end. 5150 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff)); 5151 result_region->init_req(checked_path, checks_done); 5152 result_i_o ->init_req(checked_path, checked_i_o); 5153 result_memory->init_req(checked_path, checked_mem); 5154 5155 // If it is not zero, merge into the slow call. 5156 set_control( _gvn.transform(new(C) IfFalseNode(iff) )); 5157 RegionNode* slow_reg2 = new(C) RegionNode(3); 5158 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO); 5159 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type); 5160 record_for_igvn(slow_reg2); 5161 slow_reg2 ->init_req(1, slow_control); 5162 slow_i_o2 ->init_req(1, slow_i_o); 5163 slow_mem2 ->init_req(1, slow_mem); 5164 slow_reg2 ->init_req(2, control()); 5165 slow_i_o2 ->init_req(2, checked_i_o); 5166 slow_mem2 ->init_req(2, checked_mem); 5167 5168 slow_control = _gvn.transform(slow_reg2); 5169 slow_i_o = _gvn.transform(slow_i_o2); 5170 slow_mem = _gvn.transform(slow_mem2); 5171 5172 if (alloc != NULL) { 5173 // We'll restart from the very beginning, after zeroing the whole thing. 5174 // This can cause double writes, but that's OK since dest is brand new. 5175 // So we ignore the low 31 bits of the value returned from the stub. 5176 } else { 5177 // We must continue the copy exactly where it failed, or else 5178 // another thread might see the wrong number of writes to dest. 5179 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1))); 5180 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT); 5181 slow_offset->init_req(1, intcon(0)); 5182 slow_offset->init_req(2, checked_offset); 5183 slow_offset = _gvn.transform(slow_offset); 5184 5185 // Adjust the arguments by the conditionally incoming offset. 5186 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset)); 5187 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset)); 5188 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset)); 5189 5190 // Tweak the node variables to adjust the code produced below: 5191 src_offset = src_off_plus; 5192 dest_offset = dest_off_plus; 5193 copy_length = length_minus; 5194 } 5195 } 5196 5197 set_control(slow_control); 5198 if (!stopped()) { 5199 // Generate the slow path, if needed. 5200 PreserveJVMState pjvms(this); // replace_in_map may trash the map 5201 5202 set_memory(slow_mem, adr_type); 5203 set_i_o(slow_i_o); 5204 5205 if (dest_uninitialized) { 5206 generate_clear_array(adr_type, dest, basic_elem_type, 5207 intcon(0), NULL, 5208 alloc->in(AllocateNode::AllocSize)); 5209 } 5210 5211 generate_slow_arraycopy(adr_type, 5212 src, src_offset, dest, dest_offset, 5213 copy_length, /*dest_uninitialized*/false); 5214 5215 result_region->init_req(slow_call_path, control()); 5216 result_i_o ->init_req(slow_call_path, i_o()); 5217 result_memory->init_req(slow_call_path, memory(adr_type)); 5218 } 5219 5220 // Remove unused edges. 5221 for (uint i = 1; i < result_region->req(); i++) { 5222 if (result_region->in(i) == NULL) 5223 result_region->init_req(i, top()); 5224 } 5225 5226 // Finished; return the combined state. 5227 set_control( _gvn.transform(result_region)); 5228 set_i_o( _gvn.transform(result_i_o) ); 5229 set_memory( _gvn.transform(result_memory), adr_type ); 5230 5231 // The memory edges above are precise in order to model effects around 5232 // array copies accurately to allow value numbering of field loads around 5233 // arraycopy. Such field loads, both before and after, are common in Java 5234 // collections and similar classes involving header/array data structures. 5235 // 5236 // But with low number of register or when some registers are used or killed 5237 // by arraycopy calls it causes registers spilling on stack. See 6544710. 5238 // The next memory barrier is added to avoid it. If the arraycopy can be 5239 // optimized away (which it can, sometimes) then we can manually remove 5240 // the membar also. 5241 // 5242 // Do not let reads from the cloned object float above the arraycopy. 5243 if (alloc != NULL) { 5244 // Do not let stores that initialize this object be reordered with 5245 // a subsequent store that would make this object accessible by 5246 // other threads. 5247 // Record what AllocateNode this StoreStore protects so that 5248 // escape analysis can go from the MemBarStoreStoreNode to the 5249 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5250 // based on the escape status of the AllocateNode. 5251 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5252 } else if (InsertMemBarAfterArraycopy) 5253 insert_mem_bar(Op_MemBarCPUOrder); 5254 } 5255 5256 5257 // Helper function which determines if an arraycopy immediately follows 5258 // an allocation, with no intervening tests or other escapes for the object. 5259 AllocateArrayNode* 5260 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5261 RegionNode* slow_region) { 5262 if (stopped()) return NULL; // no fast path 5263 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5264 5265 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5266 if (alloc == NULL) return NULL; 5267 5268 Node* rawmem = memory(Compile::AliasIdxRaw); 5269 // Is the allocation's memory state untouched? 5270 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5271 // Bail out if there have been raw-memory effects since the allocation. 5272 // (Example: There might have been a call or safepoint.) 5273 return NULL; 5274 } 5275 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5276 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5277 return NULL; 5278 } 5279 5280 // There must be no unexpected observers of this allocation. 5281 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5282 Node* obs = ptr->fast_out(i); 5283 if (obs != this->map()) { 5284 return NULL; 5285 } 5286 } 5287 5288 // This arraycopy must unconditionally follow the allocation of the ptr. 5289 Node* alloc_ctl = ptr->in(0); 5290 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5291 5292 Node* ctl = control(); 5293 while (ctl != alloc_ctl) { 5294 // There may be guards which feed into the slow_region. 5295 // Any other control flow means that we might not get a chance 5296 // to finish initializing the allocated object. 5297 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5298 IfNode* iff = ctl->in(0)->as_If(); 5299 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5300 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5301 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5302 ctl = iff->in(0); // This test feeds the known slow_region. 5303 continue; 5304 } 5305 // One more try: Various low-level checks bottom out in 5306 // uncommon traps. If the debug-info of the trap omits 5307 // any reference to the allocation, as we've already 5308 // observed, then there can be no objection to the trap. 5309 bool found_trap = false; 5310 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5311 Node* obs = not_ctl->fast_out(j); 5312 if (obs->in(0) == not_ctl && obs->is_Call() && 5313 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5314 found_trap = true; break; 5315 } 5316 } 5317 if (found_trap) { 5318 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5319 continue; 5320 } 5321 } 5322 return NULL; 5323 } 5324 5325 // If we get this far, we have an allocation which immediately 5326 // precedes the arraycopy, and we can take over zeroing the new object. 5327 // The arraycopy will finish the initialization, and provide 5328 // a new control state to which we will anchor the destination pointer. 5329 5330 return alloc; 5331 } 5332 5333 // Helper for initialization of arrays, creating a ClearArray. 5334 // It writes zero bits in [start..end), within the body of an array object. 5335 // The memory effects are all chained onto the 'adr_type' alias category. 5336 // 5337 // Since the object is otherwise uninitialized, we are free 5338 // to put a little "slop" around the edges of the cleared area, 5339 // as long as it does not go back into the array's header, 5340 // or beyond the array end within the heap. 5341 // 5342 // The lower edge can be rounded down to the nearest jint and the 5343 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5344 // 5345 // Arguments: 5346 // adr_type memory slice where writes are generated 5347 // dest oop of the destination array 5348 // basic_elem_type element type of the destination 5349 // slice_idx array index of first element to store 5350 // slice_len number of elements to store (or NULL) 5351 // dest_size total size in bytes of the array object 5352 // 5353 // Exactly one of slice_len or dest_size must be non-NULL. 5354 // If dest_size is non-NULL, zeroing extends to the end of the object. 5355 // If slice_len is non-NULL, the slice_idx value must be a constant. 5356 void 5357 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5358 Node* dest, 5359 BasicType basic_elem_type, 5360 Node* slice_idx, 5361 Node* slice_len, 5362 Node* dest_size) { 5363 // one or the other but not both of slice_len and dest_size: 5364 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5365 if (slice_len == NULL) slice_len = top(); 5366 if (dest_size == NULL) dest_size = top(); 5367 5368 // operate on this memory slice: 5369 Node* mem = memory(adr_type); // memory slice to operate on 5370 5371 // scaling and rounding of indexes: 5372 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5373 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5374 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5375 int bump_bit = (-1 << scale) & BytesPerInt; 5376 5377 // determine constant starts and ends 5378 const intptr_t BIG_NEG = -128; 5379 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5380 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5381 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5382 if (slice_len_con == 0) { 5383 return; // nothing to do here 5384 } 5385 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5386 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5387 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5388 assert(end_con < 0, "not two cons"); 5389 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5390 BytesPerLong); 5391 } 5392 5393 if (start_con >= 0 && end_con >= 0) { 5394 // Constant start and end. Simple. 5395 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5396 start_con, end_con, &_gvn); 5397 } else if (start_con >= 0 && dest_size != top()) { 5398 // Constant start, pre-rounded end after the tail of the array. 5399 Node* end = dest_size; 5400 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5401 start_con, end, &_gvn); 5402 } else if (start_con >= 0 && slice_len != top()) { 5403 // Constant start, non-constant end. End needs rounding up. 5404 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5405 intptr_t end_base = abase + (slice_idx_con << scale); 5406 int end_round = (-1 << scale) & (BytesPerLong - 1); 5407 Node* end = ConvI2X(slice_len); 5408 if (scale != 0) 5409 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) )); 5410 end_base += end_round; 5411 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base))); 5412 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round))); 5413 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5414 start_con, end, &_gvn); 5415 } else if (start_con < 0 && dest_size != top()) { 5416 // Non-constant start, pre-rounded end after the tail of the array. 5417 // This is almost certainly a "round-to-end" operation. 5418 Node* start = slice_idx; 5419 start = ConvI2X(start); 5420 if (scale != 0) 5421 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) )); 5422 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase))); 5423 if ((bump_bit | clear_low) != 0) { 5424 int to_clear = (bump_bit | clear_low); 5425 // Align up mod 8, then store a jint zero unconditionally 5426 // just before the mod-8 boundary. 5427 if (((abase + bump_bit) & ~to_clear) - bump_bit 5428 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5429 bump_bit = 0; 5430 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5431 } else { 5432 // Bump 'start' up to (or past) the next jint boundary: 5433 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit))); 5434 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5435 } 5436 // Round bumped 'start' down to jlong boundary in body of array. 5437 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear))); 5438 if (bump_bit != 0) { 5439 // Store a zero to the immediately preceding jint: 5440 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit))); 5441 Node* p1 = basic_plus_adr(dest, x1); 5442 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered); 5443 mem = _gvn.transform(mem); 5444 } 5445 } 5446 Node* end = dest_size; // pre-rounded 5447 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5448 start, end, &_gvn); 5449 } else { 5450 // Non-constant start, unrounded non-constant end. 5451 // (Nobody zeroes a random midsection of an array using this routine.) 5452 ShouldNotReachHere(); // fix caller 5453 } 5454 5455 // Done. 5456 set_memory(mem, adr_type); 5457 } 5458 5459 5460 bool 5461 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5462 BasicType basic_elem_type, 5463 AllocateNode* alloc, 5464 Node* src, Node* src_offset, 5465 Node* dest, Node* dest_offset, 5466 Node* dest_size, bool dest_uninitialized) { 5467 // See if there is an advantage from block transfer. 5468 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5469 if (scale >= LogBytesPerLong) 5470 return false; // it is already a block transfer 5471 5472 // Look at the alignment of the starting offsets. 5473 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5474 5475 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5476 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5477 if (src_off_con < 0 || dest_off_con < 0) 5478 // At present, we can only understand constants. 5479 return false; 5480 5481 intptr_t src_off = abase + (src_off_con << scale); 5482 intptr_t dest_off = abase + (dest_off_con << scale); 5483 5484 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5485 // Non-aligned; too bad. 5486 // One more chance: Pick off an initial 32-bit word. 5487 // This is a common case, since abase can be odd mod 8. 5488 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5489 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5490 Node* sptr = basic_plus_adr(src, src_off); 5491 Node* dptr = basic_plus_adr(dest, dest_off); 5492 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 5493 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered); 5494 src_off += BytesPerInt; 5495 dest_off += BytesPerInt; 5496 } else { 5497 return false; 5498 } 5499 } 5500 assert(src_off % BytesPerLong == 0, ""); 5501 assert(dest_off % BytesPerLong == 0, ""); 5502 5503 // Do this copy by giant steps. 5504 Node* sptr = basic_plus_adr(src, src_off); 5505 Node* dptr = basic_plus_adr(dest, dest_off); 5506 Node* countx = dest_size; 5507 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off))); 5508 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong))); 5509 5510 bool disjoint_bases = true; // since alloc != NULL 5511 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5512 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5513 5514 return true; 5515 } 5516 5517 5518 // Helper function; generates code for the slow case. 5519 // We make a call to a runtime method which emulates the native method, 5520 // but without the native wrapper overhead. 5521 void 5522 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5523 Node* src, Node* src_offset, 5524 Node* dest, Node* dest_offset, 5525 Node* copy_length, bool dest_uninitialized) { 5526 assert(!dest_uninitialized, "Invariant"); 5527 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5528 OptoRuntime::slow_arraycopy_Type(), 5529 OptoRuntime::slow_arraycopy_Java(), 5530 "slow_arraycopy", adr_type, 5531 src, src_offset, dest, dest_offset, 5532 copy_length); 5533 5534 // Handle exceptions thrown by this fellow: 5535 make_slow_call_ex(call, env()->Throwable_klass(), false); 5536 } 5537 5538 // Helper function; generates code for cases requiring runtime checks. 5539 Node* 5540 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5541 Node* dest_elem_klass, 5542 Node* src, Node* src_offset, 5543 Node* dest, Node* dest_offset, 5544 Node* copy_length, bool dest_uninitialized) { 5545 if (stopped()) return NULL; 5546 5547 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5548 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5549 return NULL; 5550 } 5551 5552 // Pick out the parameters required to perform a store-check 5553 // for the target array. This is an optimistic check. It will 5554 // look in each non-null element's class, at the desired klass's 5555 // super_check_offset, for the desired klass. 5556 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5557 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5558 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered); 5559 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5560 Node* check_value = dest_elem_klass; 5561 5562 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5563 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5564 5565 // (We know the arrays are never conjoint, because their types differ.) 5566 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5567 OptoRuntime::checkcast_arraycopy_Type(), 5568 copyfunc_addr, "checkcast_arraycopy", adr_type, 5569 // five arguments, of which two are 5570 // intptr_t (jlong in LP64) 5571 src_start, dest_start, 5572 copy_length XTOP, 5573 check_offset XTOP, 5574 check_value); 5575 5576 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5577 } 5578 5579 5580 // Helper function; generates code for cases requiring runtime checks. 5581 Node* 5582 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5583 Node* src, Node* src_offset, 5584 Node* dest, Node* dest_offset, 5585 Node* copy_length, bool dest_uninitialized) { 5586 assert(!dest_uninitialized, "Invariant"); 5587 if (stopped()) return NULL; 5588 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5589 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5590 return NULL; 5591 } 5592 5593 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5594 OptoRuntime::generic_arraycopy_Type(), 5595 copyfunc_addr, "generic_arraycopy", adr_type, 5596 src, src_offset, dest, dest_offset, copy_length); 5597 5598 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5599 } 5600 5601 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5602 void 5603 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5604 BasicType basic_elem_type, 5605 bool disjoint_bases, 5606 Node* src, Node* src_offset, 5607 Node* dest, Node* dest_offset, 5608 Node* copy_length, bool dest_uninitialized) { 5609 if (stopped()) return; // nothing to do 5610 5611 Node* src_start = src; 5612 Node* dest_start = dest; 5613 if (src_offset != NULL || dest_offset != NULL) { 5614 assert(src_offset != NULL && dest_offset != NULL, ""); 5615 src_start = array_element_address(src, src_offset, basic_elem_type); 5616 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5617 } 5618 5619 // Figure out which arraycopy runtime method to call. 5620 const char* copyfunc_name = "arraycopy"; 5621 address copyfunc_addr = 5622 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5623 disjoint_bases, copyfunc_name, dest_uninitialized); 5624 5625 // Call it. Note that the count_ix value is not scaled to a byte-size. 5626 make_runtime_call(RC_LEAF|RC_NO_FP, 5627 OptoRuntime::fast_arraycopy_Type(), 5628 copyfunc_addr, copyfunc_name, adr_type, 5629 src_start, dest_start, copy_length XTOP); 5630 } 5631 5632 //-------------inline_encodeISOArray----------------------------------- 5633 // encode char[] to byte[] in ISO_8859_1 5634 bool LibraryCallKit::inline_encodeISOArray() { 5635 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5636 // no receiver since it is static method 5637 Node *src = argument(0); 5638 Node *src_offset = argument(1); 5639 Node *dst = argument(2); 5640 Node *dst_offset = argument(3); 5641 Node *length = argument(4); 5642 5643 const Type* src_type = src->Value(&_gvn); 5644 const Type* dst_type = dst->Value(&_gvn); 5645 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5646 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5647 if (top_src == NULL || top_src->klass() == NULL || 5648 top_dest == NULL || top_dest->klass() == NULL) { 5649 // failed array check 5650 return false; 5651 } 5652 5653 // Figure out the size and type of the elements we will be copying. 5654 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5655 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5656 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5657 return false; 5658 } 5659 Node* src_start = array_element_address(src, src_offset, src_elem); 5660 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5661 // 'src_start' points to src array + scaled offset 5662 // 'dst_start' points to dst array + scaled offset 5663 5664 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5665 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5666 enc = _gvn.transform(enc); 5667 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc)); 5668 set_memory(res_mem, mtype); 5669 set_result(enc); 5670 return true; 5671 } 5672 5673 /** 5674 * Calculate CRC32 for byte. 5675 * int java.util.zip.CRC32.update(int crc, int b) 5676 */ 5677 bool LibraryCallKit::inline_updateCRC32() { 5678 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5679 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5680 // no receiver since it is static method 5681 Node* crc = argument(0); // type: int 5682 Node* b = argument(1); // type: int 5683 5684 /* 5685 * int c = ~ crc; 5686 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5687 * b = b ^ (c >>> 8); 5688 * crc = ~b; 5689 */ 5690 5691 Node* M1 = intcon(-1); 5692 crc = _gvn.transform(new (C) XorINode(crc, M1)); 5693 Node* result = _gvn.transform(new (C) XorINode(crc, b)); 5694 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF))); 5695 5696 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5697 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2))); 5698 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5699 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5700 5701 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8))); 5702 result = _gvn.transform(new (C) XorINode(crc, result)); 5703 result = _gvn.transform(new (C) XorINode(result, M1)); 5704 set_result(result); 5705 return true; 5706 } 5707 5708 /** 5709 * Calculate CRC32 for byte[] array. 5710 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5711 */ 5712 bool LibraryCallKit::inline_updateBytesCRC32() { 5713 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5714 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5715 // no receiver since it is static method 5716 Node* crc = argument(0); // type: int 5717 Node* src = argument(1); // type: oop 5718 Node* offset = argument(2); // type: int 5719 Node* length = argument(3); // type: int 5720 5721 const Type* src_type = src->Value(&_gvn); 5722 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5723 if (top_src == NULL || top_src->klass() == NULL) { 5724 // failed array check 5725 return false; 5726 } 5727 5728 // Figure out the size and type of the elements we will be copying. 5729 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5730 if (src_elem != T_BYTE) { 5731 return false; 5732 } 5733 5734 // 'src_start' points to src array + scaled offset 5735 Node* src_start = array_element_address(src, offset, src_elem); 5736 5737 // We assume that range check is done by caller. 5738 // TODO: generate range check (offset+length < src.length) in debug VM. 5739 5740 // Call the stub. 5741 address stubAddr = StubRoutines::updateBytesCRC32(); 5742 const char *stubName = "updateBytesCRC32"; 5743 5744 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5745 stubAddr, stubName, TypePtr::BOTTOM, 5746 crc, src_start, length); 5747 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5748 set_result(result); 5749 return true; 5750 } 5751 5752 /** 5753 * Calculate CRC32 for ByteBuffer. 5754 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5755 */ 5756 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5757 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5758 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5759 // no receiver since it is static method 5760 Node* crc = argument(0); // type: int 5761 Node* src = argument(1); // type: long 5762 Node* offset = argument(3); // type: int 5763 Node* length = argument(4); // type: int 5764 5765 src = ConvL2X(src); // adjust Java long to machine word 5766 Node* base = _gvn.transform(new (C) CastX2PNode(src)); 5767 offset = ConvI2X(offset); 5768 5769 // 'src_start' points to src array + scaled offset 5770 Node* src_start = basic_plus_adr(top(), base, offset); 5771 5772 // Call the stub. 5773 address stubAddr = StubRoutines::updateBytesCRC32(); 5774 const char *stubName = "updateBytesCRC32"; 5775 5776 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5777 stubAddr, stubName, TypePtr::BOTTOM, 5778 crc, src_start, length); 5779 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5780 set_result(result); 5781 return true; 5782 } 5783 5784 //----------------------------inline_reference_get---------------------------- 5785 // public T java.lang.ref.Reference.get(); 5786 bool LibraryCallKit::inline_reference_get() { 5787 const int referent_offset = java_lang_ref_Reference::referent_offset; 5788 guarantee(referent_offset > 0, "should have already been set"); 5789 5790 // Get the argument: 5791 Node* reference_obj = null_check_receiver(); 5792 if (stopped()) return true; 5793 5794 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5795 5796 ciInstanceKlass* klass = env()->Object_klass(); 5797 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5798 5799 Node* no_ctrl = NULL; 5800 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5801 5802 // Use the pre-barrier to record the value in the referent field 5803 pre_barrier(false /* do_load */, 5804 control(), 5805 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5806 result /* pre_val */, 5807 T_OBJECT); 5808 5809 // Add memory barrier to prevent commoning reads from this field 5810 // across safepoint since GC can change its value. 5811 insert_mem_bar(Op_MemBarCPUOrder); 5812 5813 set_result(result); 5814 return true; 5815 } 5816 5817 5818 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5819 bool is_exact=true, bool is_static=false) { 5820 5821 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5822 assert(tinst != NULL, "obj is null"); 5823 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5824 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5825 5826 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5827 ciSymbol::make(fieldTypeString), 5828 is_static); 5829 if (field == NULL) return (Node *) NULL; 5830 assert (field != NULL, "undefined field"); 5831 5832 // Next code copied from Parse::do_get_xxx(): 5833 5834 // Compute address and memory type. 5835 int offset = field->offset_in_bytes(); 5836 bool is_vol = field->is_volatile(); 5837 ciType* field_klass = field->type(); 5838 assert(field_klass->is_loaded(), "should be loaded"); 5839 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5840 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5841 BasicType bt = field->layout_type(); 5842 5843 // Build the resultant type of the load 5844 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5845 5846 // Build the load. 5847 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol); 5848 return loadedField; 5849 } 5850 5851 5852 //------------------------------inline_aescrypt_Block----------------------- 5853 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5854 address stubAddr; 5855 const char *stubName; 5856 assert(UseAES, "need AES instruction support"); 5857 5858 switch(id) { 5859 case vmIntrinsics::_aescrypt_encryptBlock: 5860 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5861 stubName = "aescrypt_encryptBlock"; 5862 break; 5863 case vmIntrinsics::_aescrypt_decryptBlock: 5864 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5865 stubName = "aescrypt_decryptBlock"; 5866 break; 5867 } 5868 if (stubAddr == NULL) return false; 5869 5870 Node* aescrypt_object = argument(0); 5871 Node* src = argument(1); 5872 Node* src_offset = argument(2); 5873 Node* dest = argument(3); 5874 Node* dest_offset = argument(4); 5875 5876 // (1) src and dest are arrays. 5877 const Type* src_type = src->Value(&_gvn); 5878 const Type* dest_type = dest->Value(&_gvn); 5879 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5880 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5881 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5882 5883 // for the quick and dirty code we will skip all the checks. 5884 // we are just trying to get the call to be generated. 5885 Node* src_start = src; 5886 Node* dest_start = dest; 5887 if (src_offset != NULL || dest_offset != NULL) { 5888 assert(src_offset != NULL && dest_offset != NULL, ""); 5889 src_start = array_element_address(src, src_offset, T_BYTE); 5890 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5891 } 5892 5893 // now need to get the start of its expanded key array 5894 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5895 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5896 if (k_start == NULL) return false; 5897 5898 if (Matcher::pass_original_key_for_aes()) { 5899 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5900 // compatibility issues between Java key expansion and SPARC crypto instructions 5901 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5902 if (original_k_start == NULL) return false; 5903 5904 // Call the stub. 5905 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5906 stubAddr, stubName, TypePtr::BOTTOM, 5907 src_start, dest_start, k_start, original_k_start); 5908 } else { 5909 // Call the stub. 5910 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5911 stubAddr, stubName, TypePtr::BOTTOM, 5912 src_start, dest_start, k_start); 5913 } 5914 5915 return true; 5916 } 5917 5918 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5919 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5920 address stubAddr; 5921 const char *stubName; 5922 5923 assert(UseAES, "need AES instruction support"); 5924 5925 switch(id) { 5926 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5927 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5928 stubName = "cipherBlockChaining_encryptAESCrypt"; 5929 break; 5930 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5931 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5932 stubName = "cipherBlockChaining_decryptAESCrypt"; 5933 break; 5934 } 5935 if (stubAddr == NULL) return false; 5936 5937 Node* cipherBlockChaining_object = argument(0); 5938 Node* src = argument(1); 5939 Node* src_offset = argument(2); 5940 Node* len = argument(3); 5941 Node* dest = argument(4); 5942 Node* dest_offset = argument(5); 5943 5944 // (1) src and dest are arrays. 5945 const Type* src_type = src->Value(&_gvn); 5946 const Type* dest_type = dest->Value(&_gvn); 5947 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5948 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5949 assert (top_src != NULL && top_src->klass() != NULL 5950 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5951 5952 // checks are the responsibility of the caller 5953 Node* src_start = src; 5954 Node* dest_start = dest; 5955 if (src_offset != NULL || dest_offset != NULL) { 5956 assert(src_offset != NULL && dest_offset != NULL, ""); 5957 src_start = array_element_address(src, src_offset, T_BYTE); 5958 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5959 } 5960 5961 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5962 // (because of the predicated logic executed earlier). 5963 // so we cast it here safely. 5964 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5965 5966 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5967 if (embeddedCipherObj == NULL) return false; 5968 5969 // cast it to what we know it will be at runtime 5970 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5971 assert(tinst != NULL, "CBC obj is null"); 5972 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5973 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5974 if (!klass_AESCrypt->is_loaded()) return false; 5975 5976 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5977 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5978 const TypeOopPtr* xtype = aklass->as_instance_type(); 5979 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype); 5980 aescrypt_object = _gvn.transform(aescrypt_object); 5981 5982 // we need to get the start of the aescrypt_object's expanded key array 5983 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5984 if (k_start == NULL) return false; 5985 5986 // similarly, get the start address of the r vector 5987 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5988 if (objRvec == NULL) return false; 5989 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5990 5991 Node* cbcCrypt; 5992 if (Matcher::pass_original_key_for_aes()) { 5993 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5994 // compatibility issues between Java key expansion and SPARC crypto instructions 5995 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5996 if (original_k_start == NULL) return false; 5997 5998 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5999 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6000 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6001 stubAddr, stubName, TypePtr::BOTTOM, 6002 src_start, dest_start, k_start, r_start, len, original_k_start); 6003 } else { 6004 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6005 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6006 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6007 stubAddr, stubName, TypePtr::BOTTOM, 6008 src_start, dest_start, k_start, r_start, len); 6009 } 6010 6011 // return cipher length (int) 6012 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms)); 6013 set_result(retvalue); 6014 return true; 6015 } 6016 6017 //------------------------------get_key_start_from_aescrypt_object----------------------- 6018 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6019 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6020 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6021 if (objAESCryptKey == NULL) return (Node *) NULL; 6022 6023 // now have the array, need to get the start address of the K array 6024 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6025 return k_start; 6026 } 6027 6028 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6029 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6030 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6031 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6032 if (objAESCryptKey == NULL) return (Node *) NULL; 6033 6034 // now have the array, need to get the start address of the lastKey array 6035 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6036 return original_k_start; 6037 } 6038 6039 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6040 // Return node representing slow path of predicate check. 6041 // the pseudo code we want to emulate with this predicate is: 6042 // for encryption: 6043 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6044 // for decryption: 6045 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6046 // note cipher==plain is more conservative than the original java code but that's OK 6047 // 6048 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6049 // First, check receiver for NULL since it is virtual method. 6050 Node* objCBC = argument(0); 6051 objCBC = null_check(objCBC); 6052 6053 if (stopped()) return NULL; // Always NULL 6054 6055 // Load embeddedCipher field of CipherBlockChaining object. 6056 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6057 6058 // get AESCrypt klass for instanceOf check 6059 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6060 // will have same classloader as CipherBlockChaining object 6061 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6062 assert(tinst != NULL, "CBCobj is null"); 6063 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6064 6065 // we want to do an instanceof comparison against the AESCrypt class 6066 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6067 if (!klass_AESCrypt->is_loaded()) { 6068 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6069 Node* ctrl = control(); 6070 set_control(top()); // no regular fast path 6071 return ctrl; 6072 } 6073 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6074 6075 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6076 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1))); 6077 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); 6078 6079 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6080 6081 // for encryption, we are done 6082 if (!decrypting) 6083 return instof_false; // even if it is NULL 6084 6085 // for decryption, we need to add a further check to avoid 6086 // taking the intrinsic path when cipher and plain are the same 6087 // see the original java code for why. 6088 RegionNode* region = new(C) RegionNode(3); 6089 region->init_req(1, instof_false); 6090 Node* src = argument(1); 6091 Node* dest = argument(4); 6092 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest)); 6093 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq)); 6094 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6095 region->init_req(2, src_dest_conjoint); 6096 6097 record_for_igvn(region); 6098 return _gvn.transform(region); 6099 }