library_call.cpp revision 8722:564b61ae7dc8
1/* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25#include "precompiled.hpp" 26#include "asm/macroAssembler.hpp" 27#include "classfile/systemDictionary.hpp" 28#include "classfile/vmSymbols.hpp" 29#include "compiler/compileBroker.hpp" 30#include "compiler/compileLog.hpp" 31#include "oops/objArrayKlass.hpp" 32#include "opto/addnode.hpp" 33#include "opto/arraycopynode.hpp" 34#include "opto/callGenerator.hpp" 35#include "opto/castnode.hpp" 36#include "opto/cfgnode.hpp" 37#include "opto/convertnode.hpp" 38#include "opto/countbitsnode.hpp" 39#include "opto/intrinsicnode.hpp" 40#include "opto/idealKit.hpp" 41#include "opto/mathexactnode.hpp" 42#include "opto/movenode.hpp" 43#include "opto/mulnode.hpp" 44#include "opto/narrowptrnode.hpp" 45#include "opto/opaquenode.hpp" 46#include "opto/parse.hpp" 47#include "opto/runtime.hpp" 48#include "opto/subnode.hpp" 49#include "prims/nativeLookup.hpp" 50#include "runtime/sharedRuntime.hpp" 51#include "trace/traceMacros.hpp" 52 53class LibraryIntrinsic : public InlineCallGenerator { 54 // Extend the set of intrinsics known to the runtime: 55 public: 56 private: 57 bool _is_virtual; 58 bool _does_virtual_dispatch; 59 int8_t _predicates_count; // Intrinsic is predicated by several conditions 60 int8_t _last_predicate; // Last generated predicate 61 vmIntrinsics::ID _intrinsic_id; 62 63 public: 64 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 65 : InlineCallGenerator(m), 66 _is_virtual(is_virtual), 67 _does_virtual_dispatch(does_virtual_dispatch), 68 _predicates_count((int8_t)predicates_count), 69 _last_predicate((int8_t)-1), 70 _intrinsic_id(id) 71 { 72 } 73 virtual bool is_intrinsic() const { return true; } 74 virtual bool is_virtual() const { return _is_virtual; } 75 virtual bool is_predicated() const { return _predicates_count > 0; } 76 virtual int predicates_count() const { return _predicates_count; } 77 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 78 virtual JVMState* generate(JVMState* jvms); 79 virtual Node* generate_predicate(JVMState* jvms, int predicate); 80 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 81}; 82 83 84// Local helper class for LibraryIntrinsic: 85class LibraryCallKit : public GraphKit { 86 private: 87 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 88 Node* _result; // the result node, if any 89 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 90 91 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 92 93 public: 94 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 95 : GraphKit(jvms), 96 _intrinsic(intrinsic), 97 _result(NULL) 98 { 99 // Check if this is a root compile. In that case we don't have a caller. 100 if (!jvms->has_method()) { 101 _reexecute_sp = sp(); 102 } else { 103 // Find out how many arguments the interpreter needs when deoptimizing 104 // and save the stack pointer value so it can used by uncommon_trap. 105 // We find the argument count by looking at the declared signature. 106 bool ignored_will_link; 107 ciSignature* declared_signature = NULL; 108 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 109 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 110 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 111 } 112 } 113 114 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 115 116 ciMethod* caller() const { return jvms()->method(); } 117 int bci() const { return jvms()->bci(); } 118 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 119 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 120 ciMethod* callee() const { return _intrinsic->method(); } 121 122 bool try_to_inline(int predicate); 123 Node* try_to_predicate(int predicate); 124 125 void push_result() { 126 // Push the result onto the stack. 127 if (!stopped() && result() != NULL) { 128 BasicType bt = result()->bottom_type()->basic_type(); 129 push_node(bt, result()); 130 } 131 } 132 133 private: 134 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 135 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 136 } 137 138 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 139 void set_result(RegionNode* region, PhiNode* value); 140 Node* result() { return _result; } 141 142 virtual int reexecute_sp() { return _reexecute_sp; } 143 144 // Helper functions to inline natives 145 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 146 Node* generate_slow_guard(Node* test, RegionNode* region); 147 Node* generate_fair_guard(Node* test, RegionNode* region); 148 Node* generate_negative_guard(Node* index, RegionNode* region, 149 // resulting CastII of index: 150 Node* *pos_index = NULL); 151 Node* generate_limit_guard(Node* offset, Node* subseq_length, 152 Node* array_length, 153 RegionNode* region); 154 Node* generate_current_thread(Node* &tls_output); 155 Node* load_mirror_from_klass(Node* klass); 156 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 157 RegionNode* region, int null_path, 158 int offset); 159 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path) { 161 int offset = java_lang_Class::klass_offset_in_bytes(); 162 return load_klass_from_mirror_common(mirror, never_see_null, 163 region, null_path, 164 offset); 165 } 166 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 167 RegionNode* region, int null_path) { 168 int offset = java_lang_Class::array_klass_offset_in_bytes(); 169 return load_klass_from_mirror_common(mirror, never_see_null, 170 region, null_path, 171 offset); 172 } 173 Node* generate_access_flags_guard(Node* kls, 174 int modifier_mask, int modifier_bits, 175 RegionNode* region); 176 Node* generate_interface_guard(Node* kls, RegionNode* region); 177 Node* generate_array_guard(Node* kls, RegionNode* region) { 178 return generate_array_guard_common(kls, region, false, false); 179 } 180 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 181 return generate_array_guard_common(kls, region, false, true); 182 } 183 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 184 return generate_array_guard_common(kls, region, true, false); 185 } 186 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 187 return generate_array_guard_common(kls, region, true, true); 188 } 189 Node* generate_array_guard_common(Node* kls, RegionNode* region, 190 bool obj_array, bool not_array); 191 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 192 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 193 bool is_virtual = false, bool is_static = false); 194 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 195 return generate_method_call(method_id, false, true); 196 } 197 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 198 return generate_method_call(method_id, true, false); 199 } 200 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 201 202 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 203 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 204 bool inline_string_compareTo(); 205 bool inline_string_indexOf(); 206 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 207 bool inline_string_equals(); 208 Node* round_double_node(Node* n); 209 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 210 bool inline_math_native(vmIntrinsics::ID id); 211 bool inline_trig(vmIntrinsics::ID id); 212 bool inline_math(vmIntrinsics::ID id); 213 template <typename OverflowOp> 214 bool inline_math_overflow(Node* arg1, Node* arg2); 215 void inline_math_mathExact(Node* math, Node* test); 216 bool inline_math_addExactI(bool is_increment); 217 bool inline_math_addExactL(bool is_increment); 218 bool inline_math_multiplyExactI(); 219 bool inline_math_multiplyExactL(); 220 bool inline_math_negateExactI(); 221 bool inline_math_negateExactL(); 222 bool inline_math_subtractExactI(bool is_decrement); 223 bool inline_math_subtractExactL(bool is_decrement); 224 bool inline_exp(); 225 bool inline_pow(); 226 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 227 bool inline_min_max(vmIntrinsics::ID id); 228 bool inline_notify(vmIntrinsics::ID id); 229 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 230 // This returns Type::AnyPtr, RawPtr, or OopPtr. 231 int classify_unsafe_addr(Node* &base, Node* &offset); 232 Node* make_unsafe_address(Node* base, Node* offset); 233 // Helper for inline_unsafe_access. 234 // Generates the guards that check whether the result of 235 // Unsafe.getObject should be recorded in an SATB log buffer. 236 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 237 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 238 static bool klass_needs_init_guard(Node* kls); 239 bool inline_unsafe_allocate(); 240 bool inline_unsafe_copyMemory(); 241 bool inline_native_currentThread(); 242#ifdef TRACE_HAVE_INTRINSICS 243 bool inline_native_classID(); 244 bool inline_native_threadID(); 245#endif 246 bool inline_native_time_funcs(address method, const char* funcName); 247 bool inline_native_isInterrupted(); 248 bool inline_native_Class_query(vmIntrinsics::ID id); 249 bool inline_native_subtype_check(); 250 251 bool inline_native_newArray(); 252 bool inline_native_getLength(); 253 bool inline_array_copyOf(bool is_copyOfRange); 254 bool inline_array_equals(); 255 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 256 bool inline_native_clone(bool is_virtual); 257 bool inline_native_Reflection_getCallerClass(); 258 // Helper function for inlining native object hash method 259 bool inline_native_hashcode(bool is_virtual, bool is_static); 260 bool inline_native_getClass(); 261 262 // Helper functions for inlining arraycopy 263 bool inline_arraycopy(); 264 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 265 RegionNode* slow_region); 266 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 267 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 268 269 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 270 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 271 bool inline_unsafe_ordered_store(BasicType type); 272 bool inline_unsafe_fence(vmIntrinsics::ID id); 273 bool inline_fp_conversions(vmIntrinsics::ID id); 274 bool inline_number_methods(vmIntrinsics::ID id); 275 bool inline_reference_get(); 276 bool inline_Class_cast(); 277 bool inline_aescrypt_Block(vmIntrinsics::ID id); 278 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 279 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 280 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 281 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 282 bool inline_ghash_processBlocks(); 283 bool inline_sha_implCompress(vmIntrinsics::ID id); 284 bool inline_digestBase_implCompressMB(int predicate); 285 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 286 bool long_state, address stubAddr, const char *stubName, 287 Node* src_start, Node* ofs, Node* limit); 288 Node* get_state_from_sha_object(Node *sha_object); 289 Node* get_state_from_sha5_object(Node *sha_object); 290 Node* inline_digestBase_implCompressMB_predicate(int predicate); 291 bool inline_encodeISOArray(); 292 bool inline_updateCRC32(); 293 bool inline_updateBytesCRC32(); 294 bool inline_updateByteBufferCRC32(); 295 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 296 bool inline_updateBytesCRC32C(); 297 bool inline_updateDirectByteBufferCRC32C(); 298 bool inline_multiplyToLen(); 299 bool inline_squareToLen(); 300 bool inline_mulAdd(); 301 bool inline_montgomeryMultiply(); 302 bool inline_montgomerySquare(); 303 304 bool inline_profileBoolean(); 305 bool inline_isCompileConstant(); 306}; 307 308 309//---------------------------make_vm_intrinsic---------------------------- 310CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 311 vmIntrinsics::ID id = m->intrinsic_id(); 312 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 313 314 ccstr disable_intr = NULL; 315 316 if ((DisableIntrinsic[0] != '\0' 317 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 318 (method_has_option_value("DisableIntrinsic", disable_intr) 319 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 320 // disabled by a user request on the command line: 321 // example: -XX:DisableIntrinsic=_hashCode,_getClass 322 return NULL; 323 } 324 325 if (!m->is_loaded()) { 326 // do not attempt to inline unloaded methods 327 return NULL; 328 } 329 330 // Only a few intrinsics implement a virtual dispatch. 331 // They are expensive calls which are also frequently overridden. 332 if (is_virtual) { 333 switch (id) { 334 case vmIntrinsics::_hashCode: 335 case vmIntrinsics::_clone: 336 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 337 break; 338 default: 339 return NULL; 340 } 341 } 342 343 // -XX:-InlineNatives disables nearly all intrinsics: 344 if (!InlineNatives) { 345 switch (id) { 346 case vmIntrinsics::_indexOf: 347 case vmIntrinsics::_compareTo: 348 case vmIntrinsics::_equals: 349 case vmIntrinsics::_equalsC: 350 case vmIntrinsics::_getAndAddInt: 351 case vmIntrinsics::_getAndAddLong: 352 case vmIntrinsics::_getAndSetInt: 353 case vmIntrinsics::_getAndSetLong: 354 case vmIntrinsics::_getAndSetObject: 355 case vmIntrinsics::_loadFence: 356 case vmIntrinsics::_storeFence: 357 case vmIntrinsics::_fullFence: 358 break; // InlineNatives does not control String.compareTo 359 case vmIntrinsics::_Reference_get: 360 break; // InlineNatives does not control Reference.get 361 default: 362 return NULL; 363 } 364 } 365 366 int predicates = 0; 367 bool does_virtual_dispatch = false; 368 369 switch (id) { 370 case vmIntrinsics::_compareTo: 371 if (!SpecialStringCompareTo) return NULL; 372 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 373 break; 374 case vmIntrinsics::_indexOf: 375 if (!SpecialStringIndexOf) return NULL; 376 break; 377 case vmIntrinsics::_equals: 378 if (!SpecialStringEquals) return NULL; 379 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 380 break; 381 case vmIntrinsics::_equalsC: 382 if (!SpecialArraysEquals) return NULL; 383 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 384 break; 385 case vmIntrinsics::_arraycopy: 386 if (!InlineArrayCopy) return NULL; 387 break; 388 case vmIntrinsics::_copyMemory: 389 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 390 if (!InlineArrayCopy) return NULL; 391 break; 392 case vmIntrinsics::_hashCode: 393 if (!InlineObjectHash) return NULL; 394 does_virtual_dispatch = true; 395 break; 396 case vmIntrinsics::_clone: 397 does_virtual_dispatch = true; 398 case vmIntrinsics::_copyOf: 399 case vmIntrinsics::_copyOfRange: 400 if (!InlineObjectCopy) return NULL; 401 // These also use the arraycopy intrinsic mechanism: 402 if (!InlineArrayCopy) return NULL; 403 break; 404 case vmIntrinsics::_encodeISOArray: 405 if (!SpecialEncodeISOArray) return NULL; 406 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 407 break; 408 case vmIntrinsics::_checkIndex: 409 // We do not intrinsify this. The optimizer does fine with it. 410 return NULL; 411 412 case vmIntrinsics::_getCallerClass: 413 if (!InlineReflectionGetCallerClass) return NULL; 414 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 415 break; 416 417 case vmIntrinsics::_bitCount_i: 418 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 419 break; 420 421 case vmIntrinsics::_bitCount_l: 422 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 423 break; 424 425 case vmIntrinsics::_numberOfLeadingZeros_i: 426 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 427 break; 428 429 case vmIntrinsics::_numberOfLeadingZeros_l: 430 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 431 break; 432 433 case vmIntrinsics::_numberOfTrailingZeros_i: 434 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 435 break; 436 437 case vmIntrinsics::_numberOfTrailingZeros_l: 438 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 439 break; 440 441 case vmIntrinsics::_reverseBytes_c: 442 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 443 break; 444 case vmIntrinsics::_reverseBytes_s: 445 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 446 break; 447 case vmIntrinsics::_reverseBytes_i: 448 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 449 break; 450 case vmIntrinsics::_reverseBytes_l: 451 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 452 break; 453 454 case vmIntrinsics::_Reference_get: 455 // Use the intrinsic version of Reference.get() so that the value in 456 // the referent field can be registered by the G1 pre-barrier code. 457 // Also add memory barrier to prevent commoning reads from this field 458 // across safepoint since GC can change it value. 459 break; 460 461 case vmIntrinsics::_compareAndSwapObject: 462#ifdef _LP64 463 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 464#endif 465 break; 466 467 case vmIntrinsics::_compareAndSwapLong: 468 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 469 break; 470 471 case vmIntrinsics::_getAndAddInt: 472 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 473 break; 474 475 case vmIntrinsics::_getAndAddLong: 476 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 477 break; 478 479 case vmIntrinsics::_getAndSetInt: 480 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 481 break; 482 483 case vmIntrinsics::_getAndSetLong: 484 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 485 break; 486 487 case vmIntrinsics::_getAndSetObject: 488#ifdef _LP64 489 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 490 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 491 break; 492#else 493 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 494 break; 495#endif 496 497 case vmIntrinsics::_aescrypt_encryptBlock: 498 case vmIntrinsics::_aescrypt_decryptBlock: 499 if (!UseAESIntrinsics) return NULL; 500 break; 501 502 case vmIntrinsics::_multiplyToLen: 503 if (!UseMultiplyToLenIntrinsic) return NULL; 504 break; 505 506 case vmIntrinsics::_squareToLen: 507 if (!UseSquareToLenIntrinsic) return NULL; 508 break; 509 510 case vmIntrinsics::_mulAdd: 511 if (!UseMulAddIntrinsic) return NULL; 512 break; 513 514 case vmIntrinsics::_montgomeryMultiply: 515 if (!UseMontgomeryMultiplyIntrinsic) return NULL; 516 break; 517 case vmIntrinsics::_montgomerySquare: 518 if (!UseMontgomerySquareIntrinsic) return NULL; 519 break; 520 521 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 522 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 523 if (!UseAESIntrinsics) return NULL; 524 // these two require the predicated logic 525 predicates = 1; 526 break; 527 528 case vmIntrinsics::_sha_implCompress: 529 if (!UseSHA1Intrinsics) return NULL; 530 break; 531 532 case vmIntrinsics::_sha2_implCompress: 533 if (!UseSHA256Intrinsics) return NULL; 534 break; 535 536 case vmIntrinsics::_sha5_implCompress: 537 if (!UseSHA512Intrinsics) return NULL; 538 break; 539 540 case vmIntrinsics::_digestBase_implCompressMB: 541 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 542 predicates = 3; 543 break; 544 545 case vmIntrinsics::_ghash_processBlocks: 546 if (!UseGHASHIntrinsics) return NULL; 547 break; 548 549 case vmIntrinsics::_updateCRC32: 550 case vmIntrinsics::_updateBytesCRC32: 551 case vmIntrinsics::_updateByteBufferCRC32: 552 if (!UseCRC32Intrinsics) return NULL; 553 break; 554 555 case vmIntrinsics::_updateBytesCRC32C: 556 case vmIntrinsics::_updateDirectByteBufferCRC32C: 557 if (!UseCRC32CIntrinsics) return NULL; 558 break; 559 560 case vmIntrinsics::_incrementExactI: 561 case vmIntrinsics::_addExactI: 562 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 563 break; 564 case vmIntrinsics::_incrementExactL: 565 case vmIntrinsics::_addExactL: 566 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 567 break; 568 case vmIntrinsics::_decrementExactI: 569 case vmIntrinsics::_subtractExactI: 570 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 571 break; 572 case vmIntrinsics::_decrementExactL: 573 case vmIntrinsics::_subtractExactL: 574 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 575 break; 576 case vmIntrinsics::_negateExactI: 577 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 578 break; 579 case vmIntrinsics::_negateExactL: 580 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 581 break; 582 case vmIntrinsics::_multiplyExactI: 583 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 584 break; 585 case vmIntrinsics::_multiplyExactL: 586 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 587 break; 588 589 case vmIntrinsics::_getShortUnaligned: 590 case vmIntrinsics::_getCharUnaligned: 591 case vmIntrinsics::_getIntUnaligned: 592 case vmIntrinsics::_getLongUnaligned: 593 case vmIntrinsics::_putShortUnaligned: 594 case vmIntrinsics::_putCharUnaligned: 595 case vmIntrinsics::_putIntUnaligned: 596 case vmIntrinsics::_putLongUnaligned: 597 if (!UseUnalignedAccesses) return NULL; 598 break; 599 600 default: 601 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 602 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 603 break; 604 } 605 606 // -XX:-InlineClassNatives disables natives from the Class class. 607 // The flag applies to all reflective calls, notably Array.newArray 608 // (visible to Java programmers as Array.newInstance). 609 if (m->holder()->name() == ciSymbol::java_lang_Class() || 610 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 611 if (!InlineClassNatives) return NULL; 612 } 613 614 // -XX:-InlineThreadNatives disables natives from the Thread class. 615 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 616 if (!InlineThreadNatives) return NULL; 617 } 618 619 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 620 if (m->holder()->name() == ciSymbol::java_lang_Math() || 621 m->holder()->name() == ciSymbol::java_lang_Float() || 622 m->holder()->name() == ciSymbol::java_lang_Double()) { 623 if (!InlineMathNatives) return NULL; 624 } 625 626 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 627 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 628 if (!InlineUnsafeOps) return NULL; 629 } 630 631 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 632} 633 634//----------------------register_library_intrinsics----------------------- 635// Initialize this file's data structures, for each Compile instance. 636void Compile::register_library_intrinsics() { 637 // Nothing to do here. 638} 639 640JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 641 LibraryCallKit kit(jvms, this); 642 Compile* C = kit.C; 643 int nodes = C->unique(); 644#ifndef PRODUCT 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("Intrinsic %s", str); 649 } 650#endif 651 ciMethod* callee = kit.callee(); 652 const int bci = kit.bci(); 653 654 // Try to inline the intrinsic. 655 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 656 kit.try_to_inline(_last_predicate)) { 657 if (C->print_intrinsics() || C->print_inlining()) { 658 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 659 } 660 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 661 if (C->log()) { 662 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 663 vmIntrinsics::name_at(intrinsic_id()), 664 (is_virtual() ? " virtual='1'" : ""), 665 C->unique() - nodes); 666 } 667 // Push the result from the inlined method onto the stack. 668 kit.push_result(); 669 C->print_inlining_update(this); 670 return kit.transfer_exceptions_into_jvms(); 671 } 672 673 // The intrinsic bailed out 674 if (C->print_intrinsics() || C->print_inlining()) { 675 if (jvms->has_method()) { 676 // Not a root compile. 677 const char* msg; 678 if (callee->intrinsic_candidate()) { 679 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 680 } else { 681 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 682 : "failed to inline (intrinsic), method not annotated"; 683 } 684 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 685 } else { 686 // Root compile 687 tty->print("Did not generate intrinsic %s%s at bci:%d in", 688 vmIntrinsics::name_at(intrinsic_id()), 689 (is_virtual() ? " (virtual)" : ""), bci); 690 } 691 } 692 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 693 C->print_inlining_update(this); 694 return NULL; 695} 696 697Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 698 LibraryCallKit kit(jvms, this); 699 Compile* C = kit.C; 700 int nodes = C->unique(); 701 _last_predicate = predicate; 702#ifndef PRODUCT 703 assert(is_predicated() && predicate < predicates_count(), "sanity"); 704 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 705 char buf[1000]; 706 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 707 tty->print_cr("Predicate for intrinsic %s", str); 708 } 709#endif 710 ciMethod* callee = kit.callee(); 711 const int bci = kit.bci(); 712 713 Node* slow_ctl = kit.try_to_predicate(predicate); 714 if (!kit.failing()) { 715 if (C->print_intrinsics() || C->print_inlining()) { 716 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 717 } 718 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 719 if (C->log()) { 720 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 721 vmIntrinsics::name_at(intrinsic_id()), 722 (is_virtual() ? " virtual='1'" : ""), 723 C->unique() - nodes); 724 } 725 return slow_ctl; // Could be NULL if the check folds. 726 } 727 728 // The intrinsic bailed out 729 if (C->print_intrinsics() || C->print_inlining()) { 730 if (jvms->has_method()) { 731 // Not a root compile. 732 const char* msg = "failed to generate predicate for intrinsic"; 733 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 734 } else { 735 // Root compile 736 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 737 vmIntrinsics::name_at(intrinsic_id()), 738 (is_virtual() ? " (virtual)" : ""), bci); 739 } 740 } 741 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 742 return NULL; 743} 744 745bool LibraryCallKit::try_to_inline(int predicate) { 746 // Handle symbolic names for otherwise undistinguished boolean switches: 747 const bool is_store = true; 748 const bool is_native_ptr = true; 749 const bool is_static = true; 750 const bool is_volatile = true; 751 752 if (!jvms()->has_method()) { 753 // Root JVMState has a null method. 754 assert(map()->memory()->Opcode() == Op_Parm, ""); 755 // Insert the memory aliasing node 756 set_all_memory(reset_memory()); 757 } 758 assert(merged_memory(), ""); 759 760 761 switch (intrinsic_id()) { 762 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 763 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 764 case vmIntrinsics::_getClass: return inline_native_getClass(); 765 766 case vmIntrinsics::_dsin: 767 case vmIntrinsics::_dcos: 768 case vmIntrinsics::_dtan: 769 case vmIntrinsics::_dabs: 770 case vmIntrinsics::_datan2: 771 case vmIntrinsics::_dsqrt: 772 case vmIntrinsics::_dexp: 773 case vmIntrinsics::_dlog: 774 case vmIntrinsics::_dlog10: 775 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 776 777 case vmIntrinsics::_min: 778 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 779 780 case vmIntrinsics::_notify: 781 case vmIntrinsics::_notifyAll: 782 if (InlineNotify) { 783 return inline_notify(intrinsic_id()); 784 } 785 return false; 786 787 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 788 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 789 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 790 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 791 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 792 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 793 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 794 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 795 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 796 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 797 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 798 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 799 800 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 801 802 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 803 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 804 case vmIntrinsics::_equals: return inline_string_equals(); 805 806 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 807 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 808 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 809 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 810 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 811 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 812 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 813 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 814 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 815 816 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 817 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 818 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 819 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 820 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 821 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 822 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 823 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 824 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 825 826 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 827 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 828 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 829 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 830 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 831 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 832 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 833 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 834 835 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 836 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 837 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 838 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 839 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 840 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 841 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 842 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 843 844 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 845 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 846 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 847 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 848 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 849 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 850 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 851 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 852 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 853 854 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 855 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 856 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 857 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 858 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 859 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 860 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 861 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 862 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 863 864 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 865 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 866 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 867 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 868 869 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 870 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 871 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 872 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 873 874 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 875 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 876 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 877 878 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 879 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 880 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 881 882 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 883 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 884 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 885 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 886 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 887 888 case vmIntrinsics::_loadFence: 889 case vmIntrinsics::_storeFence: 890 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 891 892 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 893 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 894 895#ifdef TRACE_HAVE_INTRINSICS 896 case vmIntrinsics::_classID: return inline_native_classID(); 897 case vmIntrinsics::_threadID: return inline_native_threadID(); 898 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 899#endif 900 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 901 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 902 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 903 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 904 case vmIntrinsics::_newArray: return inline_native_newArray(); 905 case vmIntrinsics::_getLength: return inline_native_getLength(); 906 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 907 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 908 case vmIntrinsics::_equalsC: return inline_array_equals(); 909 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 910 911 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 912 913 case vmIntrinsics::_isInstance: 914 case vmIntrinsics::_getModifiers: 915 case vmIntrinsics::_isInterface: 916 case vmIntrinsics::_isArray: 917 case vmIntrinsics::_isPrimitive: 918 case vmIntrinsics::_getSuperclass: 919 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 920 921 case vmIntrinsics::_floatToRawIntBits: 922 case vmIntrinsics::_floatToIntBits: 923 case vmIntrinsics::_intBitsToFloat: 924 case vmIntrinsics::_doubleToRawLongBits: 925 case vmIntrinsics::_doubleToLongBits: 926 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 927 928 case vmIntrinsics::_numberOfLeadingZeros_i: 929 case vmIntrinsics::_numberOfLeadingZeros_l: 930 case vmIntrinsics::_numberOfTrailingZeros_i: 931 case vmIntrinsics::_numberOfTrailingZeros_l: 932 case vmIntrinsics::_bitCount_i: 933 case vmIntrinsics::_bitCount_l: 934 case vmIntrinsics::_reverseBytes_i: 935 case vmIntrinsics::_reverseBytes_l: 936 case vmIntrinsics::_reverseBytes_s: 937 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 938 939 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 940 941 case vmIntrinsics::_Reference_get: return inline_reference_get(); 942 943 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 944 945 case vmIntrinsics::_aescrypt_encryptBlock: 946 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 947 948 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 949 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 950 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 951 952 case vmIntrinsics::_sha_implCompress: 953 case vmIntrinsics::_sha2_implCompress: 954 case vmIntrinsics::_sha5_implCompress: 955 return inline_sha_implCompress(intrinsic_id()); 956 957 case vmIntrinsics::_digestBase_implCompressMB: 958 return inline_digestBase_implCompressMB(predicate); 959 960 case vmIntrinsics::_multiplyToLen: 961 return inline_multiplyToLen(); 962 963 case vmIntrinsics::_squareToLen: 964 return inline_squareToLen(); 965 966 case vmIntrinsics::_mulAdd: 967 return inline_mulAdd(); 968 969 case vmIntrinsics::_montgomeryMultiply: 970 return inline_montgomeryMultiply(); 971 case vmIntrinsics::_montgomerySquare: 972 return inline_montgomerySquare(); 973 974 case vmIntrinsics::_ghash_processBlocks: 975 return inline_ghash_processBlocks(); 976 977 case vmIntrinsics::_encodeISOArray: 978 return inline_encodeISOArray(); 979 980 case vmIntrinsics::_updateCRC32: 981 return inline_updateCRC32(); 982 case vmIntrinsics::_updateBytesCRC32: 983 return inline_updateBytesCRC32(); 984 case vmIntrinsics::_updateByteBufferCRC32: 985 return inline_updateByteBufferCRC32(); 986 987 case vmIntrinsics::_updateBytesCRC32C: 988 return inline_updateBytesCRC32C(); 989 case vmIntrinsics::_updateDirectByteBufferCRC32C: 990 return inline_updateDirectByteBufferCRC32C(); 991 992 case vmIntrinsics::_profileBoolean: 993 return inline_profileBoolean(); 994 case vmIntrinsics::_isCompileConstant: 995 return inline_isCompileConstant(); 996 997 default: 998 // If you get here, it may be that someone has added a new intrinsic 999 // to the list in vmSymbols.hpp without implementing it here. 1000#ifndef PRODUCT 1001 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 1002 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 1003 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 1004 } 1005#endif 1006 return false; 1007 } 1008} 1009 1010Node* LibraryCallKit::try_to_predicate(int predicate) { 1011 if (!jvms()->has_method()) { 1012 // Root JVMState has a null method. 1013 assert(map()->memory()->Opcode() == Op_Parm, ""); 1014 // Insert the memory aliasing node 1015 set_all_memory(reset_memory()); 1016 } 1017 assert(merged_memory(), ""); 1018 1019 switch (intrinsic_id()) { 1020 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 1021 return inline_cipherBlockChaining_AESCrypt_predicate(false); 1022 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 1023 return inline_cipherBlockChaining_AESCrypt_predicate(true); 1024 case vmIntrinsics::_digestBase_implCompressMB: 1025 return inline_digestBase_implCompressMB_predicate(predicate); 1026 1027 default: 1028 // If you get here, it may be that someone has added a new intrinsic 1029 // to the list in vmSymbols.hpp without implementing it here. 1030#ifndef PRODUCT 1031 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 1032 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 1033 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 1034 } 1035#endif 1036 Node* slow_ctl = control(); 1037 set_control(top()); // No fast path instrinsic 1038 return slow_ctl; 1039 } 1040} 1041 1042//------------------------------set_result------------------------------- 1043// Helper function for finishing intrinsics. 1044void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 1045 record_for_igvn(region); 1046 set_control(_gvn.transform(region)); 1047 set_result( _gvn.transform(value)); 1048 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 1049} 1050 1051//------------------------------generate_guard--------------------------- 1052// Helper function for generating guarded fast-slow graph structures. 1053// The given 'test', if true, guards a slow path. If the test fails 1054// then a fast path can be taken. (We generally hope it fails.) 1055// In all cases, GraphKit::control() is updated to the fast path. 1056// The returned value represents the control for the slow path. 1057// The return value is never 'top'; it is either a valid control 1058// or NULL if it is obvious that the slow path can never be taken. 1059// Also, if region and the slow control are not NULL, the slow edge 1060// is appended to the region. 1061Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 1062 if (stopped()) { 1063 // Already short circuited. 1064 return NULL; 1065 } 1066 1067 // Build an if node and its projections. 1068 // If test is true we take the slow path, which we assume is uncommon. 1069 if (_gvn.type(test) == TypeInt::ZERO) { 1070 // The slow branch is never taken. No need to build this guard. 1071 return NULL; 1072 } 1073 1074 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1075 1076 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1077 if (if_slow == top()) { 1078 // The slow branch is never taken. No need to build this guard. 1079 return NULL; 1080 } 1081 1082 if (region != NULL) 1083 region->add_req(if_slow); 1084 1085 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1086 set_control(if_fast); 1087 1088 return if_slow; 1089} 1090 1091inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1092 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1093} 1094inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1095 return generate_guard(test, region, PROB_FAIR); 1096} 1097 1098inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1099 Node* *pos_index) { 1100 if (stopped()) 1101 return NULL; // already stopped 1102 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1103 return NULL; // index is already adequately typed 1104 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1105 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1106 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1107 if (is_neg != NULL && pos_index != NULL) { 1108 // Emulate effect of Parse::adjust_map_after_if. 1109 Node* ccast = new CastIINode(index, TypeInt::POS); 1110 ccast->set_req(0, control()); 1111 (*pos_index) = _gvn.transform(ccast); 1112 } 1113 return is_neg; 1114} 1115 1116// Make sure that 'position' is a valid limit index, in [0..length]. 1117// There are two equivalent plans for checking this: 1118// A. (offset + copyLength) unsigned<= arrayLength 1119// B. offset <= (arrayLength - copyLength) 1120// We require that all of the values above, except for the sum and 1121// difference, are already known to be non-negative. 1122// Plan A is robust in the face of overflow, if offset and copyLength 1123// are both hugely positive. 1124// 1125// Plan B is less direct and intuitive, but it does not overflow at 1126// all, since the difference of two non-negatives is always 1127// representable. Whenever Java methods must perform the equivalent 1128// check they generally use Plan B instead of Plan A. 1129// For the moment we use Plan A. 1130inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1131 Node* subseq_length, 1132 Node* array_length, 1133 RegionNode* region) { 1134 if (stopped()) 1135 return NULL; // already stopped 1136 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1137 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1138 return NULL; // common case of whole-array copy 1139 Node* last = subseq_length; 1140 if (!zero_offset) // last += offset 1141 last = _gvn.transform(new AddINode(last, offset)); 1142 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1143 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1144 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1145 return is_over; 1146} 1147 1148 1149//--------------------------generate_current_thread-------------------- 1150Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1151 ciKlass* thread_klass = env()->Thread_klass(); 1152 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1153 Node* thread = _gvn.transform(new ThreadLocalNode()); 1154 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1155 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1156 tls_output = thread; 1157 return threadObj; 1158} 1159 1160 1161//------------------------------make_string_method_node------------------------ 1162// Helper method for String intrinsic functions. This version is called 1163// with str1 and str2 pointing to String object nodes. 1164// 1165Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1166 Node* no_ctrl = NULL; 1167 1168 // Get start addr of string 1169 Node* str1_value = load_String_value(no_ctrl, str1); 1170 Node* str1_offset = load_String_offset(no_ctrl, str1); 1171 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1172 1173 // Get length of string 1 1174 Node* str1_len = load_String_length(no_ctrl, str1); 1175 1176 Node* str2_value = load_String_value(no_ctrl, str2); 1177 Node* str2_offset = load_String_offset(no_ctrl, str2); 1178 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1179 1180 Node* str2_len = NULL; 1181 Node* result = NULL; 1182 1183 switch (opcode) { 1184 case Op_StrIndexOf: 1185 // Get length of string 2 1186 str2_len = load_String_length(no_ctrl, str2); 1187 1188 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1189 str1_start, str1_len, str2_start, str2_len); 1190 break; 1191 case Op_StrComp: 1192 // Get length of string 2 1193 str2_len = load_String_length(no_ctrl, str2); 1194 1195 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1196 str1_start, str1_len, str2_start, str2_len); 1197 break; 1198 case Op_StrEquals: 1199 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1200 str1_start, str2_start, str1_len); 1201 break; 1202 default: 1203 ShouldNotReachHere(); 1204 return NULL; 1205 } 1206 1207 // All these intrinsics have checks. 1208 C->set_has_split_ifs(true); // Has chance for split-if optimization 1209 1210 return _gvn.transform(result); 1211} 1212 1213// Helper method for String intrinsic functions. This version is called 1214// with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1215// to Int nodes containing the lenghts of str1 and str2. 1216// 1217Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1218 Node* result = NULL; 1219 switch (opcode) { 1220 case Op_StrIndexOf: 1221 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1222 str1_start, cnt1, str2_start, cnt2); 1223 break; 1224 case Op_StrComp: 1225 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1226 str1_start, cnt1, str2_start, cnt2); 1227 break; 1228 case Op_StrEquals: 1229 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1230 str1_start, str2_start, cnt1); 1231 break; 1232 default: 1233 ShouldNotReachHere(); 1234 return NULL; 1235 } 1236 1237 // All these intrinsics have checks. 1238 C->set_has_split_ifs(true); // Has chance for split-if optimization 1239 1240 return _gvn.transform(result); 1241} 1242 1243//------------------------------inline_string_compareTo------------------------ 1244// public int java.lang.String.compareTo(String anotherString); 1245bool LibraryCallKit::inline_string_compareTo() { 1246 Node* receiver = null_check(argument(0)); 1247 Node* arg = null_check(argument(1)); 1248 if (stopped()) { 1249 return true; 1250 } 1251 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1252 return true; 1253} 1254 1255//------------------------------inline_string_equals------------------------ 1256bool LibraryCallKit::inline_string_equals() { 1257 Node* receiver = null_check_receiver(); 1258 // NOTE: Do not null check argument for String.equals() because spec 1259 // allows to specify NULL as argument. 1260 Node* argument = this->argument(1); 1261 if (stopped()) { 1262 return true; 1263 } 1264 1265 // paths (plus control) merge 1266 RegionNode* region = new RegionNode(5); 1267 Node* phi = new PhiNode(region, TypeInt::BOOL); 1268 1269 // does source == target string? 1270 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1271 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1272 1273 Node* if_eq = generate_slow_guard(bol, NULL); 1274 if (if_eq != NULL) { 1275 // receiver == argument 1276 phi->init_req(2, intcon(1)); 1277 region->init_req(2, if_eq); 1278 } 1279 1280 // get String klass for instanceOf 1281 ciInstanceKlass* klass = env()->String_klass(); 1282 1283 if (!stopped()) { 1284 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1285 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1286 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1287 1288 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1289 //instanceOf == true, fallthrough 1290 1291 if (inst_false != NULL) { 1292 phi->init_req(3, intcon(0)); 1293 region->init_req(3, inst_false); 1294 } 1295 } 1296 1297 if (!stopped()) { 1298 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1299 1300 // Properly cast the argument to String 1301 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1302 // This path is taken only when argument's type is String:NotNull. 1303 argument = cast_not_null(argument, false); 1304 1305 Node* no_ctrl = NULL; 1306 1307 // Get start addr of receiver 1308 Node* receiver_val = load_String_value(no_ctrl, receiver); 1309 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1310 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1311 1312 // Get length of receiver 1313 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1314 1315 // Get start addr of argument 1316 Node* argument_val = load_String_value(no_ctrl, argument); 1317 Node* argument_offset = load_String_offset(no_ctrl, argument); 1318 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1319 1320 // Get length of argument 1321 Node* argument_cnt = load_String_length(no_ctrl, argument); 1322 1323 // Check for receiver count != argument count 1324 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1325 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1326 Node* if_ne = generate_slow_guard(bol, NULL); 1327 if (if_ne != NULL) { 1328 phi->init_req(4, intcon(0)); 1329 region->init_req(4, if_ne); 1330 } 1331 1332 // Check for count == 0 is done by assembler code for StrEquals. 1333 1334 if (!stopped()) { 1335 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1336 phi->init_req(1, equals); 1337 region->init_req(1, control()); 1338 } 1339 } 1340 1341 // post merge 1342 set_control(_gvn.transform(region)); 1343 record_for_igvn(region); 1344 1345 set_result(_gvn.transform(phi)); 1346 return true; 1347} 1348 1349//------------------------------inline_array_equals---------------------------- 1350bool LibraryCallKit::inline_array_equals() { 1351 Node* arg1 = argument(0); 1352 Node* arg2 = argument(1); 1353 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1354 return true; 1355} 1356 1357// Java version of String.indexOf(constant string) 1358// class StringDecl { 1359// StringDecl(char[] ca) { 1360// offset = 0; 1361// count = ca.length; 1362// value = ca; 1363// } 1364// int offset; 1365// int count; 1366// char[] value; 1367// } 1368// 1369// static int string_indexOf_J(StringDecl string_object, char[] target_object, 1370// int targetOffset, int cache_i, int md2) { 1371// int cache = cache_i; 1372// int sourceOffset = string_object.offset; 1373// int sourceCount = string_object.count; 1374// int targetCount = target_object.length; 1375// 1376// int targetCountLess1 = targetCount - 1; 1377// int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1378// 1379// char[] source = string_object.value; 1380// char[] target = target_object; 1381// int lastChar = target[targetCountLess1]; 1382// 1383// outer_loop: 1384// for (int i = sourceOffset; i < sourceEnd; ) { 1385// int src = source[i + targetCountLess1]; 1386// if (src == lastChar) { 1387// // With random strings and a 4-character alphabet, 1388// // reverse matching at this point sets up 0.8% fewer 1389// // frames, but (paradoxically) makes 0.3% more probes. 1390// // Since those probes are nearer the lastChar probe, 1391// // there is may be a net D$ win with reverse matching. 1392// // But, reversing loop inhibits unroll of inner loop 1393// // for unknown reason. So, does running outer loop from 1394// // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1395// for (int j = 0; j < targetCountLess1; j++) { 1396// if (target[targetOffset + j] != source[i+j]) { 1397// if ((cache & (1 << source[i+j])) == 0) { 1398// if (md2 < j+1) { 1399// i += j+1; 1400// continue outer_loop; 1401// } 1402// } 1403// i += md2; 1404// continue outer_loop; 1405// } 1406// } 1407// return i - sourceOffset; 1408// } 1409// if ((cache & (1 << src)) == 0) { 1410// i += targetCountLess1; 1411// } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1412// i++; 1413// } 1414// return -1; 1415// } 1416 1417//------------------------------string_indexOf------------------------ 1418Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1419 jint cache_i, jint md2_i) { 1420 1421 Node* no_ctrl = NULL; 1422 float likely = PROB_LIKELY(0.9); 1423 float unlikely = PROB_UNLIKELY(0.9); 1424 1425 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1426 1427 Node* source = load_String_value(no_ctrl, string_object); 1428 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1429 Node* sourceCount = load_String_length(no_ctrl, string_object); 1430 1431 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1432 jint target_length = target_array->length(); 1433 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1434 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1435 1436 // String.value field is known to be @Stable. 1437 if (UseImplicitStableValues) { 1438 target = cast_array_to_stable(target, target_type); 1439 } 1440 1441 IdealKit kit(this, false, true); 1442#define __ kit. 1443 Node* zero = __ ConI(0); 1444 Node* one = __ ConI(1); 1445 Node* cache = __ ConI(cache_i); 1446 Node* md2 = __ ConI(md2_i); 1447 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1448 Node* targetCountLess1 = __ ConI(target_length - 1); 1449 Node* targetOffset = __ ConI(targetOffset_i); 1450 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1451 1452 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1453 Node* outer_loop = __ make_label(2 /* goto */); 1454 Node* return_ = __ make_label(1); 1455 1456 __ set(rtn,__ ConI(-1)); 1457 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1458 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1459 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1460 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1461 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1462 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1463 Node* tpj = __ AddI(targetOffset, __ value(j)); 1464 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1465 Node* ipj = __ AddI(__ value(i), __ value(j)); 1466 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1467 __ if_then(targ, BoolTest::ne, src2); { 1468 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1469 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1470 __ increment(i, __ AddI(__ value(j), one)); 1471 __ goto_(outer_loop); 1472 } __ end_if(); __ dead(j); 1473 }__ end_if(); __ dead(j); 1474 __ increment(i, md2); 1475 __ goto_(outer_loop); 1476 }__ end_if(); 1477 __ increment(j, one); 1478 }__ end_loop(); __ dead(j); 1479 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1480 __ goto_(return_); 1481 }__ end_if(); 1482 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1483 __ increment(i, targetCountLess1); 1484 }__ end_if(); 1485 __ increment(i, one); 1486 __ bind(outer_loop); 1487 }__ end_loop(); __ dead(i); 1488 __ bind(return_); 1489 1490 // Final sync IdealKit and GraphKit. 1491 final_sync(kit); 1492 Node* result = __ value(rtn); 1493#undef __ 1494 C->set_has_loops(true); 1495 return result; 1496} 1497 1498//------------------------------inline_string_indexOf------------------------ 1499bool LibraryCallKit::inline_string_indexOf() { 1500 Node* receiver = argument(0); 1501 Node* arg = argument(1); 1502 1503 Node* result; 1504 if (Matcher::has_match_rule(Op_StrIndexOf) && 1505 UseSSE42Intrinsics) { 1506 // Generate SSE4.2 version of indexOf 1507 // We currently only have match rules that use SSE4.2 1508 1509 receiver = null_check(receiver); 1510 arg = null_check(arg); 1511 if (stopped()) { 1512 return true; 1513 } 1514 1515 // Make the merge point 1516 RegionNode* result_rgn = new RegionNode(4); 1517 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1518 Node* no_ctrl = NULL; 1519 1520 // Get start addr of source string 1521 Node* source = load_String_value(no_ctrl, receiver); 1522 Node* source_offset = load_String_offset(no_ctrl, receiver); 1523 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1524 1525 // Get length of source string 1526 Node* source_cnt = load_String_length(no_ctrl, receiver); 1527 1528 // Get start addr of substring 1529 Node* substr = load_String_value(no_ctrl, arg); 1530 Node* substr_offset = load_String_offset(no_ctrl, arg); 1531 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1532 1533 // Get length of source string 1534 Node* substr_cnt = load_String_length(no_ctrl, arg); 1535 1536 // Check for substr count > string count 1537 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1538 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1539 Node* if_gt = generate_slow_guard(bol, NULL); 1540 if (if_gt != NULL) { 1541 result_phi->init_req(2, intcon(-1)); 1542 result_rgn->init_req(2, if_gt); 1543 } 1544 1545 if (!stopped()) { 1546 // Check for substr count == 0 1547 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1548 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1549 Node* if_zero = generate_slow_guard(bol, NULL); 1550 if (if_zero != NULL) { 1551 result_phi->init_req(3, intcon(0)); 1552 result_rgn->init_req(3, if_zero); 1553 } 1554 } 1555 1556 if (!stopped()) { 1557 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1558 result_phi->init_req(1, result); 1559 result_rgn->init_req(1, control()); 1560 } 1561 set_control(_gvn.transform(result_rgn)); 1562 record_for_igvn(result_rgn); 1563 result = _gvn.transform(result_phi); 1564 1565 } else { // Use LibraryCallKit::string_indexOf 1566 // don't intrinsify if argument isn't a constant string. 1567 if (!arg->is_Con()) { 1568 return false; 1569 } 1570 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1571 if (str_type == NULL) { 1572 return false; 1573 } 1574 ciInstanceKlass* klass = env()->String_klass(); 1575 ciObject* str_const = str_type->const_oop(); 1576 if (str_const == NULL || str_const->klass() != klass) { 1577 return false; 1578 } 1579 ciInstance* str = str_const->as_instance(); 1580 assert(str != NULL, "must be instance"); 1581 1582 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1583 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1584 1585 int o; 1586 int c; 1587 if (java_lang_String::has_offset_field()) { 1588 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1589 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1590 } else { 1591 o = 0; 1592 c = pat->length(); 1593 } 1594 1595 // constant strings have no offset and count == length which 1596 // simplifies the resulting code somewhat so lets optimize for that. 1597 if (o != 0 || c != pat->length()) { 1598 return false; 1599 } 1600 1601 receiver = null_check(receiver, T_OBJECT); 1602 // NOTE: No null check on the argument is needed since it's a constant String oop. 1603 if (stopped()) { 1604 return true; 1605 } 1606 1607 // The null string as a pattern always returns 0 (match at beginning of string) 1608 if (c == 0) { 1609 set_result(intcon(0)); 1610 return true; 1611 } 1612 1613 // Generate default indexOf 1614 jchar lastChar = pat->char_at(o + (c - 1)); 1615 int cache = 0; 1616 int i; 1617 for (i = 0; i < c - 1; i++) { 1618 assert(i < pat->length(), "out of range"); 1619 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1620 } 1621 1622 int md2 = c; 1623 for (i = 0; i < c - 1; i++) { 1624 assert(i < pat->length(), "out of range"); 1625 if (pat->char_at(o + i) == lastChar) { 1626 md2 = (c - 1) - i; 1627 } 1628 } 1629 1630 result = string_indexOf(receiver, pat, o, cache, md2); 1631 } 1632 set_result(result); 1633 return true; 1634} 1635 1636//--------------------------round_double_node-------------------------------- 1637// Round a double node if necessary. 1638Node* LibraryCallKit::round_double_node(Node* n) { 1639 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1640 n = _gvn.transform(new RoundDoubleNode(0, n)); 1641 return n; 1642} 1643 1644//------------------------------inline_math----------------------------------- 1645// public static double Math.abs(double) 1646// public static double Math.sqrt(double) 1647// public static double Math.log(double) 1648// public static double Math.log10(double) 1649bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1650 Node* arg = round_double_node(argument(0)); 1651 Node* n; 1652 switch (id) { 1653 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1654 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1655 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1656 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1657 default: fatal_unexpected_iid(id); break; 1658 } 1659 set_result(_gvn.transform(n)); 1660 return true; 1661} 1662 1663//------------------------------inline_trig---------------------------------- 1664// Inline sin/cos/tan instructions, if possible. If rounding is required, do 1665// argument reduction which will turn into a fast/slow diamond. 1666bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1667 Node* arg = round_double_node(argument(0)); 1668 Node* n = NULL; 1669 1670 switch (id) { 1671 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1672 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1673 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1674 default: fatal_unexpected_iid(id); break; 1675 } 1676 n = _gvn.transform(n); 1677 1678 // Rounding required? Check for argument reduction! 1679 if (Matcher::strict_fp_requires_explicit_rounding) { 1680 static const double pi_4 = 0.7853981633974483; 1681 static const double neg_pi_4 = -0.7853981633974483; 1682 // pi/2 in 80-bit extended precision 1683 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1684 // -pi/2 in 80-bit extended precision 1685 // 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}; 1686 // Cutoff value for using this argument reduction technique 1687 //static const double pi_2_minus_epsilon = 1.564660403643354; 1688 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1689 1690 // Pseudocode for sin: 1691 // if (x <= Math.PI / 4.0) { 1692 // if (x >= -Math.PI / 4.0) return fsin(x); 1693 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1694 // } else { 1695 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1696 // } 1697 // return StrictMath.sin(x); 1698 1699 // Pseudocode for cos: 1700 // if (x <= Math.PI / 4.0) { 1701 // if (x >= -Math.PI / 4.0) return fcos(x); 1702 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1703 // } else { 1704 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1705 // } 1706 // return StrictMath.cos(x); 1707 1708 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1709 // requires a special machine instruction to load it. Instead we'll try 1710 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1711 // probably do the math inside the SIN encoding. 1712 1713 // Make the merge point 1714 RegionNode* r = new RegionNode(3); 1715 Node* phi = new PhiNode(r, Type::DOUBLE); 1716 1717 // Flatten arg so we need only 1 test 1718 Node *abs = _gvn.transform(new AbsDNode(arg)); 1719 // Node for PI/4 constant 1720 Node *pi4 = makecon(TypeD::make(pi_4)); 1721 // Check PI/4 : abs(arg) 1722 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1723 // Check: If PI/4 < abs(arg) then go slow 1724 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1725 // Branch either way 1726 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1727 set_control(opt_iff(r,iff)); 1728 1729 // Set fast path result 1730 phi->init_req(2, n); 1731 1732 // Slow path - non-blocking leaf call 1733 Node* call = NULL; 1734 switch (id) { 1735 case vmIntrinsics::_dsin: 1736 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1737 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1738 "Sin", NULL, arg, top()); 1739 break; 1740 case vmIntrinsics::_dcos: 1741 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1742 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1743 "Cos", NULL, arg, top()); 1744 break; 1745 case vmIntrinsics::_dtan: 1746 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1747 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1748 "Tan", NULL, arg, top()); 1749 break; 1750 } 1751 assert(control()->in(0) == call, ""); 1752 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1753 r->init_req(1, control()); 1754 phi->init_req(1, slow_result); 1755 1756 // Post-merge 1757 set_control(_gvn.transform(r)); 1758 record_for_igvn(r); 1759 n = _gvn.transform(phi); 1760 1761 C->set_has_split_ifs(true); // Has chance for split-if optimization 1762 } 1763 set_result(n); 1764 return true; 1765} 1766 1767Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1768 //------------------- 1769 //result=(result.isNaN())? funcAddr():result; 1770 // Check: If isNaN() by checking result!=result? then either trap 1771 // or go to runtime 1772 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1773 // Build the boolean node 1774 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1775 1776 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1777 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1778 // The pow or exp intrinsic returned a NaN, which requires a call 1779 // to the runtime. Recompile with the runtime call. 1780 uncommon_trap(Deoptimization::Reason_intrinsic, 1781 Deoptimization::Action_make_not_entrant); 1782 } 1783 return result; 1784 } else { 1785 // If this inlining ever returned NaN in the past, we compile a call 1786 // to the runtime to properly handle corner cases 1787 1788 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1789 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1790 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1791 1792 if (!if_slow->is_top()) { 1793 RegionNode* result_region = new RegionNode(3); 1794 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1795 1796 result_region->init_req(1, if_fast); 1797 result_val->init_req(1, result); 1798 1799 set_control(if_slow); 1800 1801 const TypePtr* no_memory_effects = NULL; 1802 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1803 no_memory_effects, 1804 x, top(), y, y ? top() : NULL); 1805 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1806#ifdef ASSERT 1807 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1808 assert(value_top == top(), "second value must be top"); 1809#endif 1810 1811 result_region->init_req(2, control()); 1812 result_val->init_req(2, value); 1813 set_control(_gvn.transform(result_region)); 1814 return _gvn.transform(result_val); 1815 } else { 1816 return result; 1817 } 1818 } 1819} 1820 1821//------------------------------inline_exp------------------------------------- 1822// Inline exp instructions, if possible. The Intel hardware only misses 1823// really odd corner cases (+/- Infinity). Just uncommon-trap them. 1824bool LibraryCallKit::inline_exp() { 1825 Node* arg = round_double_node(argument(0)); 1826 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1827 1828 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1829 set_result(n); 1830 1831 C->set_has_split_ifs(true); // Has chance for split-if optimization 1832 return true; 1833} 1834 1835//------------------------------inline_pow------------------------------------- 1836// Inline power instructions, if possible. 1837bool LibraryCallKit::inline_pow() { 1838 // Pseudocode for pow 1839 // if (y == 2) { 1840 // return x * x; 1841 // } else { 1842 // if (x <= 0.0) { 1843 // long longy = (long)y; 1844 // if ((double)longy == y) { // if y is long 1845 // if (y + 1 == y) longy = 0; // huge number: even 1846 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1847 // } else { 1848 // result = NaN; 1849 // } 1850 // } else { 1851 // result = DPow(x,y); 1852 // } 1853 // if (result != result)? { 1854 // result = uncommon_trap() or runtime_call(); 1855 // } 1856 // return result; 1857 // } 1858 1859 Node* x = round_double_node(argument(0)); 1860 Node* y = round_double_node(argument(2)); 1861 1862 Node* result = NULL; 1863 1864 Node* const_two_node = makecon(TypeD::make(2.0)); 1865 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1866 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1867 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1868 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1869 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1870 1871 RegionNode* region_node = new RegionNode(3); 1872 region_node->init_req(1, if_true); 1873 1874 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1875 // special case for x^y where y == 2, we can convert it to x * x 1876 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1877 1878 // set control to if_false since we will now process the false branch 1879 set_control(if_false); 1880 1881 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1882 // Short form: skip the fancy tests and just check for NaN result. 1883 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1884 } else { 1885 // If this inlining ever returned NaN in the past, include all 1886 // checks + call to the runtime. 1887 1888 // Set the merge point for If node with condition of (x <= 0.0) 1889 // There are four possible paths to region node and phi node 1890 RegionNode *r = new RegionNode(4); 1891 Node *phi = new PhiNode(r, Type::DOUBLE); 1892 1893 // Build the first if node: if (x <= 0.0) 1894 // Node for 0 constant 1895 Node *zeronode = makecon(TypeD::ZERO); 1896 // Check x:0 1897 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1898 // Check: If (x<=0) then go complex path 1899 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1900 // Branch either way 1901 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1902 // Fast path taken; set region slot 3 1903 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1904 r->init_req(3,fast_taken); // Capture fast-control 1905 1906 // Fast path not-taken, i.e. slow path 1907 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1908 1909 // Set fast path result 1910 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1911 phi->init_req(3, fast_result); 1912 1913 // Complex path 1914 // Build the second if node (if y is long) 1915 // Node for (long)y 1916 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1917 // Node for (double)((long) y) 1918 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1919 // Check (double)((long) y) : y 1920 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1921 // Check if (y isn't long) then go to slow path 1922 1923 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1924 // Branch either way 1925 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1926 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1927 1928 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1929 1930 // Calculate DPow(abs(x), y)*(1 & (long)y) 1931 // Node for constant 1 1932 Node *conone = longcon(1); 1933 // 1& (long)y 1934 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1935 1936 // A huge number is always even. Detect a huge number by checking 1937 // if y + 1 == y and set integer to be tested for parity to 0. 1938 // Required for corner case: 1939 // (long)9.223372036854776E18 = max_jlong 1940 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1941 // max_jlong is odd but 9.223372036854776E18 is even 1942 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1943 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1944 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1945 Node* correctedsign = NULL; 1946 if (ConditionalMoveLimit != 0) { 1947 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1948 } else { 1949 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1950 RegionNode *r = new RegionNode(3); 1951 Node *phi = new PhiNode(r, TypeLong::LONG); 1952 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1953 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1954 phi->init_req(1, signnode); 1955 phi->init_req(2, longcon(0)); 1956 correctedsign = _gvn.transform(phi); 1957 ylong_path = _gvn.transform(r); 1958 record_for_igvn(r); 1959 } 1960 1961 // zero node 1962 Node *conzero = longcon(0); 1963 // Check (1&(long)y)==0? 1964 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1965 // Check if (1&(long)y)!=0?, if so the result is negative 1966 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1967 // abs(x) 1968 Node *absx=_gvn.transform(new AbsDNode(x)); 1969 // abs(x)^y 1970 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1971 // -abs(x)^y 1972 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1973 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1974 Node *signresult = NULL; 1975 if (ConditionalMoveLimit != 0) { 1976 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1977 } else { 1978 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1979 RegionNode *r = new RegionNode(3); 1980 Node *phi = new PhiNode(r, Type::DOUBLE); 1981 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1982 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1983 phi->init_req(1, absxpowy); 1984 phi->init_req(2, negabsxpowy); 1985 signresult = _gvn.transform(phi); 1986 ylong_path = _gvn.transform(r); 1987 record_for_igvn(r); 1988 } 1989 // Set complex path fast result 1990 r->init_req(2, ylong_path); 1991 phi->init_req(2, signresult); 1992 1993 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1994 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1995 r->init_req(1,slow_path); 1996 phi->init_req(1,slow_result); 1997 1998 // Post merge 1999 set_control(_gvn.transform(r)); 2000 record_for_igvn(r); 2001 result = _gvn.transform(phi); 2002 } 2003 2004 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 2005 2006 // control from finish_pow_exp is now input to the region node 2007 region_node->set_req(2, control()); 2008 // the result from finish_pow_exp is now input to the phi node 2009 phi_node->init_req(2, result); 2010 set_control(_gvn.transform(region_node)); 2011 record_for_igvn(region_node); 2012 set_result(_gvn.transform(phi_node)); 2013 2014 C->set_has_split_ifs(true); // Has chance for split-if optimization 2015 return true; 2016} 2017 2018//------------------------------runtime_math----------------------------- 2019bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 2020 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 2021 "must be (DD)D or (D)D type"); 2022 2023 // Inputs 2024 Node* a = round_double_node(argument(0)); 2025 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 2026 2027 const TypePtr* no_memory_effects = NULL; 2028 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 2029 no_memory_effects, 2030 a, top(), b, b ? top() : NULL); 2031 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 2032#ifdef ASSERT 2033 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 2034 assert(value_top == top(), "second value must be top"); 2035#endif 2036 2037 set_result(value); 2038 return true; 2039} 2040 2041//------------------------------inline_math_native----------------------------- 2042bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 2043#define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 2044 switch (id) { 2045 // These intrinsics are not properly supported on all hardware 2046 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 2047 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 2048 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 2049 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2050 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2051 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2052 2053 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2054 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2055 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2056 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2057 2058 // These intrinsics are supported on all hardware 2059 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2060 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2061 2062 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2063 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2064 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2065 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2066#undef FN_PTR 2067 2068 // These intrinsics are not yet correctly implemented 2069 case vmIntrinsics::_datan2: 2070 return false; 2071 2072 default: 2073 fatal_unexpected_iid(id); 2074 return false; 2075 } 2076} 2077 2078static bool is_simple_name(Node* n) { 2079 return (n->req() == 1 // constant 2080 || (n->is_Type() && n->as_Type()->type()->singleton()) 2081 || n->is_Proj() // parameter or return value 2082 || n->is_Phi() // local of some sort 2083 ); 2084} 2085 2086//----------------------------inline_notify-----------------------------------* 2087bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 2088 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 2089 address func; 2090 if (id == vmIntrinsics::_notify) { 2091 func = OptoRuntime::monitor_notify_Java(); 2092 } else { 2093 func = OptoRuntime::monitor_notifyAll_Java(); 2094 } 2095 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 2096 make_slow_call_ex(call, env()->Throwable_klass(), false); 2097 return true; 2098} 2099 2100 2101//----------------------------inline_min_max----------------------------------- 2102bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2103 set_result(generate_min_max(id, argument(0), argument(1))); 2104 return true; 2105} 2106 2107void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2108 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2109 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2110 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2111 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2112 2113 { 2114 PreserveJVMState pjvms(this); 2115 PreserveReexecuteState preexecs(this); 2116 jvms()->set_should_reexecute(true); 2117 2118 set_control(slow_path); 2119 set_i_o(i_o()); 2120 2121 uncommon_trap(Deoptimization::Reason_intrinsic, 2122 Deoptimization::Action_none); 2123 } 2124 2125 set_control(fast_path); 2126 set_result(math); 2127} 2128 2129template <typename OverflowOp> 2130bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2131 typedef typename OverflowOp::MathOp MathOp; 2132 2133 MathOp* mathOp = new MathOp(arg1, arg2); 2134 Node* operation = _gvn.transform( mathOp ); 2135 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2136 inline_math_mathExact(operation, ofcheck); 2137 return true; 2138} 2139 2140bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2141 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2142} 2143 2144bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2145 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2146} 2147 2148bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2149 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2150} 2151 2152bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2153 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2154} 2155 2156bool LibraryCallKit::inline_math_negateExactI() { 2157 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2158} 2159 2160bool LibraryCallKit::inline_math_negateExactL() { 2161 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2162} 2163 2164bool LibraryCallKit::inline_math_multiplyExactI() { 2165 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2166} 2167 2168bool LibraryCallKit::inline_math_multiplyExactL() { 2169 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2170} 2171 2172Node* 2173LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2174 // These are the candidate return value: 2175 Node* xvalue = x0; 2176 Node* yvalue = y0; 2177 2178 if (xvalue == yvalue) { 2179 return xvalue; 2180 } 2181 2182 bool want_max = (id == vmIntrinsics::_max); 2183 2184 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2185 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2186 if (txvalue == NULL || tyvalue == NULL) return top(); 2187 // This is not really necessary, but it is consistent with a 2188 // hypothetical MaxINode::Value method: 2189 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2190 2191 // %%% This folding logic should (ideally) be in a different place. 2192 // Some should be inside IfNode, and there to be a more reliable 2193 // transformation of ?: style patterns into cmoves. We also want 2194 // more powerful optimizations around cmove and min/max. 2195 2196 // Try to find a dominating comparison of these guys. 2197 // It can simplify the index computation for Arrays.copyOf 2198 // and similar uses of System.arraycopy. 2199 // First, compute the normalized version of CmpI(x, y). 2200 int cmp_op = Op_CmpI; 2201 Node* xkey = xvalue; 2202 Node* ykey = yvalue; 2203 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2204 if (ideal_cmpxy->is_Cmp()) { 2205 // E.g., if we have CmpI(length - offset, count), 2206 // it might idealize to CmpI(length, count + offset) 2207 cmp_op = ideal_cmpxy->Opcode(); 2208 xkey = ideal_cmpxy->in(1); 2209 ykey = ideal_cmpxy->in(2); 2210 } 2211 2212 // Start by locating any relevant comparisons. 2213 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2214 Node* cmpxy = NULL; 2215 Node* cmpyx = NULL; 2216 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2217 Node* cmp = start_from->fast_out(k); 2218 if (cmp->outcnt() > 0 && // must have prior uses 2219 cmp->in(0) == NULL && // must be context-independent 2220 cmp->Opcode() == cmp_op) { // right kind of compare 2221 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2222 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2223 } 2224 } 2225 2226 const int NCMPS = 2; 2227 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2228 int cmpn; 2229 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2230 if (cmps[cmpn] != NULL) break; // find a result 2231 } 2232 if (cmpn < NCMPS) { 2233 // Look for a dominating test that tells us the min and max. 2234 int depth = 0; // Limit search depth for speed 2235 Node* dom = control(); 2236 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2237 if (++depth >= 100) break; 2238 Node* ifproj = dom; 2239 if (!ifproj->is_Proj()) continue; 2240 Node* iff = ifproj->in(0); 2241 if (!iff->is_If()) continue; 2242 Node* bol = iff->in(1); 2243 if (!bol->is_Bool()) continue; 2244 Node* cmp = bol->in(1); 2245 if (cmp == NULL) continue; 2246 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2247 if (cmps[cmpn] == cmp) break; 2248 if (cmpn == NCMPS) continue; 2249 BoolTest::mask btest = bol->as_Bool()->_test._test; 2250 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2251 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2252 // At this point, we know that 'x btest y' is true. 2253 switch (btest) { 2254 case BoolTest::eq: 2255 // They are proven equal, so we can collapse the min/max. 2256 // Either value is the answer. Choose the simpler. 2257 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2258 return yvalue; 2259 return xvalue; 2260 case BoolTest::lt: // x < y 2261 case BoolTest::le: // x <= y 2262 return (want_max ? yvalue : xvalue); 2263 case BoolTest::gt: // x > y 2264 case BoolTest::ge: // x >= y 2265 return (want_max ? xvalue : yvalue); 2266 } 2267 } 2268 } 2269 2270 // We failed to find a dominating test. 2271 // Let's pick a test that might GVN with prior tests. 2272 Node* best_bol = NULL; 2273 BoolTest::mask best_btest = BoolTest::illegal; 2274 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2275 Node* cmp = cmps[cmpn]; 2276 if (cmp == NULL) continue; 2277 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2278 Node* bol = cmp->fast_out(j); 2279 if (!bol->is_Bool()) continue; 2280 BoolTest::mask btest = bol->as_Bool()->_test._test; 2281 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2282 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2283 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2284 best_bol = bol->as_Bool(); 2285 best_btest = btest; 2286 } 2287 } 2288 } 2289 2290 Node* answer_if_true = NULL; 2291 Node* answer_if_false = NULL; 2292 switch (best_btest) { 2293 default: 2294 if (cmpxy == NULL) 2295 cmpxy = ideal_cmpxy; 2296 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2297 // and fall through: 2298 case BoolTest::lt: // x < y 2299 case BoolTest::le: // x <= y 2300 answer_if_true = (want_max ? yvalue : xvalue); 2301 answer_if_false = (want_max ? xvalue : yvalue); 2302 break; 2303 case BoolTest::gt: // x > y 2304 case BoolTest::ge: // x >= y 2305 answer_if_true = (want_max ? xvalue : yvalue); 2306 answer_if_false = (want_max ? yvalue : xvalue); 2307 break; 2308 } 2309 2310 jint hi, lo; 2311 if (want_max) { 2312 // We can sharpen the minimum. 2313 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2314 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2315 } else { 2316 // We can sharpen the maximum. 2317 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2318 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2319 } 2320 2321 // Use a flow-free graph structure, to avoid creating excess control edges 2322 // which could hinder other optimizations. 2323 // Since Math.min/max is often used with arraycopy, we want 2324 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2325 Node* cmov = CMoveNode::make(NULL, best_bol, 2326 answer_if_false, answer_if_true, 2327 TypeInt::make(lo, hi, widen)); 2328 2329 return _gvn.transform(cmov); 2330 2331 /* 2332 // This is not as desirable as it may seem, since Min and Max 2333 // nodes do not have a full set of optimizations. 2334 // And they would interfere, anyway, with 'if' optimizations 2335 // and with CMoveI canonical forms. 2336 switch (id) { 2337 case vmIntrinsics::_min: 2338 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2339 case vmIntrinsics::_max: 2340 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2341 default: 2342 ShouldNotReachHere(); 2343 } 2344 */ 2345} 2346 2347inline int 2348LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2349 const TypePtr* base_type = TypePtr::NULL_PTR; 2350 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2351 if (base_type == NULL) { 2352 // Unknown type. 2353 return Type::AnyPtr; 2354 } else if (base_type == TypePtr::NULL_PTR) { 2355 // Since this is a NULL+long form, we have to switch to a rawptr. 2356 base = _gvn.transform(new CastX2PNode(offset)); 2357 offset = MakeConX(0); 2358 return Type::RawPtr; 2359 } else if (base_type->base() == Type::RawPtr) { 2360 return Type::RawPtr; 2361 } else if (base_type->isa_oopptr()) { 2362 // Base is never null => always a heap address. 2363 if (base_type->ptr() == TypePtr::NotNull) { 2364 return Type::OopPtr; 2365 } 2366 // Offset is small => always a heap address. 2367 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2368 if (offset_type != NULL && 2369 base_type->offset() == 0 && // (should always be?) 2370 offset_type->_lo >= 0 && 2371 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2372 return Type::OopPtr; 2373 } 2374 // Otherwise, it might either be oop+off or NULL+addr. 2375 return Type::AnyPtr; 2376 } else { 2377 // No information: 2378 return Type::AnyPtr; 2379 } 2380} 2381 2382inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2383 int kind = classify_unsafe_addr(base, offset); 2384 if (kind == Type::RawPtr) { 2385 return basic_plus_adr(top(), base, offset); 2386 } else { 2387 return basic_plus_adr(base, offset); 2388 } 2389} 2390 2391//--------------------------inline_number_methods----------------------------- 2392// inline int Integer.numberOfLeadingZeros(int) 2393// inline int Long.numberOfLeadingZeros(long) 2394// 2395// inline int Integer.numberOfTrailingZeros(int) 2396// inline int Long.numberOfTrailingZeros(long) 2397// 2398// inline int Integer.bitCount(int) 2399// inline int Long.bitCount(long) 2400// 2401// inline char Character.reverseBytes(char) 2402// inline short Short.reverseBytes(short) 2403// inline int Integer.reverseBytes(int) 2404// inline long Long.reverseBytes(long) 2405bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2406 Node* arg = argument(0); 2407 Node* n; 2408 switch (id) { 2409 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2410 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2411 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2412 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2413 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2414 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2415 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2416 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2417 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2418 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2419 default: fatal_unexpected_iid(id); break; 2420 } 2421 set_result(_gvn.transform(n)); 2422 return true; 2423} 2424 2425//----------------------------inline_unsafe_access---------------------------- 2426 2427const static BasicType T_ADDRESS_HOLDER = T_LONG; 2428 2429// Helper that guards and inserts a pre-barrier. 2430void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2431 Node* pre_val, bool need_mem_bar) { 2432 // We could be accessing the referent field of a reference object. If so, when G1 2433 // is enabled, we need to log the value in the referent field in an SATB buffer. 2434 // This routine performs some compile time filters and generates suitable 2435 // runtime filters that guard the pre-barrier code. 2436 // Also add memory barrier for non volatile load from the referent field 2437 // to prevent commoning of loads across safepoint. 2438 if (!UseG1GC && !need_mem_bar) 2439 return; 2440 2441 // Some compile time checks. 2442 2443 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2444 const TypeX* otype = offset->find_intptr_t_type(); 2445 if (otype != NULL && otype->is_con() && 2446 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2447 // Constant offset but not the reference_offset so just return 2448 return; 2449 } 2450 2451 // We only need to generate the runtime guards for instances. 2452 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2453 if (btype != NULL) { 2454 if (btype->isa_aryptr()) { 2455 // Array type so nothing to do 2456 return; 2457 } 2458 2459 const TypeInstPtr* itype = btype->isa_instptr(); 2460 if (itype != NULL) { 2461 // Can the klass of base_oop be statically determined to be 2462 // _not_ a sub-class of Reference and _not_ Object? 2463 ciKlass* klass = itype->klass(); 2464 if ( klass->is_loaded() && 2465 !klass->is_subtype_of(env()->Reference_klass()) && 2466 !env()->Object_klass()->is_subtype_of(klass)) { 2467 return; 2468 } 2469 } 2470 } 2471 2472 // The compile time filters did not reject base_oop/offset so 2473 // we need to generate the following runtime filters 2474 // 2475 // if (offset == java_lang_ref_Reference::_reference_offset) { 2476 // if (instance_of(base, java.lang.ref.Reference)) { 2477 // pre_barrier(_, pre_val, ...); 2478 // } 2479 // } 2480 2481 float likely = PROB_LIKELY( 0.999); 2482 float unlikely = PROB_UNLIKELY(0.999); 2483 2484 IdealKit ideal(this); 2485#define __ ideal. 2486 2487 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2488 2489 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2490 // Update graphKit memory and control from IdealKit. 2491 sync_kit(ideal); 2492 2493 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2494 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2495 2496 // Update IdealKit memory and control from graphKit. 2497 __ sync_kit(this); 2498 2499 Node* one = __ ConI(1); 2500 // is_instof == 0 if base_oop == NULL 2501 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2502 2503 // Update graphKit from IdeakKit. 2504 sync_kit(ideal); 2505 2506 // Use the pre-barrier to record the value in the referent field 2507 pre_barrier(false /* do_load */, 2508 __ ctrl(), 2509 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2510 pre_val /* pre_val */, 2511 T_OBJECT); 2512 if (need_mem_bar) { 2513 // Add memory barrier to prevent commoning reads from this field 2514 // across safepoint since GC can change its value. 2515 insert_mem_bar(Op_MemBarCPUOrder); 2516 } 2517 // Update IdealKit from graphKit. 2518 __ sync_kit(this); 2519 2520 } __ end_if(); // _ref_type != ref_none 2521 } __ end_if(); // offset == referent_offset 2522 2523 // Final sync IdealKit and GraphKit. 2524 final_sync(ideal); 2525#undef __ 2526} 2527 2528 2529// Interpret Unsafe.fieldOffset cookies correctly: 2530extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2531 2532const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2533 // Attempt to infer a sharper value type from the offset and base type. 2534 ciKlass* sharpened_klass = NULL; 2535 2536 // See if it is an instance field, with an object type. 2537 if (alias_type->field() != NULL) { 2538 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2539 if (alias_type->field()->type()->is_klass()) { 2540 sharpened_klass = alias_type->field()->type()->as_klass(); 2541 } 2542 } 2543 2544 // See if it is a narrow oop array. 2545 if (adr_type->isa_aryptr()) { 2546 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2547 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2548 if (elem_type != NULL) { 2549 sharpened_klass = elem_type->klass(); 2550 } 2551 } 2552 } 2553 2554 // The sharpened class might be unloaded if there is no class loader 2555 // contraint in place. 2556 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2557 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2558 2559#ifndef PRODUCT 2560 if (C->print_intrinsics() || C->print_inlining()) { 2561 tty->print(" from base type: "); adr_type->dump(); 2562 tty->print(" sharpened value: "); tjp->dump(); 2563 } 2564#endif 2565 // Sharpen the value type. 2566 return tjp; 2567 } 2568 return NULL; 2569} 2570 2571bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2572 if (callee()->is_static()) return false; // caller must have the capability! 2573 2574#ifndef PRODUCT 2575 { 2576 ResourceMark rm; 2577 // Check the signatures. 2578 ciSignature* sig = callee()->signature(); 2579#ifdef ASSERT 2580 if (!is_store) { 2581 // Object getObject(Object base, int/long offset), etc. 2582 BasicType rtype = sig->return_type()->basic_type(); 2583 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2584 rtype = T_ADDRESS; // it is really a C void* 2585 assert(rtype == type, "getter must return the expected value"); 2586 if (!is_native_ptr) { 2587 assert(sig->count() == 2, "oop getter has 2 arguments"); 2588 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2589 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2590 } else { 2591 assert(sig->count() == 1, "native getter has 1 argument"); 2592 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2593 } 2594 } else { 2595 // void putObject(Object base, int/long offset, Object x), etc. 2596 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2597 if (!is_native_ptr) { 2598 assert(sig->count() == 3, "oop putter has 3 arguments"); 2599 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2600 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2601 } else { 2602 assert(sig->count() == 2, "native putter has 2 arguments"); 2603 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2604 } 2605 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2606 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2607 vtype = T_ADDRESS; // it is really a C void* 2608 assert(vtype == type, "putter must accept the expected value"); 2609 } 2610#endif // ASSERT 2611 } 2612#endif //PRODUCT 2613 2614 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2615 2616 Node* receiver = argument(0); // type: oop 2617 2618 // Build address expression. 2619 Node* adr; 2620 Node* heap_base_oop = top(); 2621 Node* offset = top(); 2622 Node* val; 2623 2624 if (!is_native_ptr) { 2625 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2626 Node* base = argument(1); // type: oop 2627 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2628 offset = argument(2); // type: long 2629 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2630 // to be plain byte offsets, which are also the same as those accepted 2631 // by oopDesc::field_base. 2632 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2633 "fieldOffset must be byte-scaled"); 2634 // 32-bit machines ignore the high half! 2635 offset = ConvL2X(offset); 2636 adr = make_unsafe_address(base, offset); 2637 heap_base_oop = base; 2638 val = is_store ? argument(4) : NULL; 2639 } else { 2640 Node* ptr = argument(1); // type: long 2641 ptr = ConvL2X(ptr); // adjust Java long to machine word 2642 adr = make_unsafe_address(NULL, ptr); 2643 val = is_store ? argument(3) : NULL; 2644 } 2645 2646 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2647 2648 // First guess at the value type. 2649 const Type *value_type = Type::get_const_basic_type(type); 2650 2651 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2652 // there was not enough information to nail it down. 2653 Compile::AliasType* alias_type = C->alias_type(adr_type); 2654 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2655 2656 // We will need memory barriers unless we can determine a unique 2657 // alias category for this reference. (Note: If for some reason 2658 // the barriers get omitted and the unsafe reference begins to "pollute" 2659 // the alias analysis of the rest of the graph, either Compile::can_alias 2660 // or Compile::must_alias will throw a diagnostic assert.) 2661 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2662 2663 // If we are reading the value of the referent field of a Reference 2664 // object (either by using Unsafe directly or through reflection) 2665 // then, if G1 is enabled, we need to record the referent in an 2666 // SATB log buffer using the pre-barrier mechanism. 2667 // Also we need to add memory barrier to prevent commoning reads 2668 // from this field across safepoint since GC can change its value. 2669 bool need_read_barrier = !is_native_ptr && !is_store && 2670 offset != top() && heap_base_oop != top(); 2671 2672 if (!is_store && type == T_OBJECT) { 2673 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2674 if (tjp != NULL) { 2675 value_type = tjp; 2676 } 2677 } 2678 2679 receiver = null_check(receiver); 2680 if (stopped()) { 2681 return true; 2682 } 2683 // Heap pointers get a null-check from the interpreter, 2684 // as a courtesy. However, this is not guaranteed by Unsafe, 2685 // and it is not possible to fully distinguish unintended nulls 2686 // from intended ones in this API. 2687 2688 if (is_volatile) { 2689 // We need to emit leading and trailing CPU membars (see below) in 2690 // addition to memory membars when is_volatile. This is a little 2691 // too strong, but avoids the need to insert per-alias-type 2692 // volatile membars (for stores; compare Parse::do_put_xxx), which 2693 // we cannot do effectively here because we probably only have a 2694 // rough approximation of type. 2695 need_mem_bar = true; 2696 // For Stores, place a memory ordering barrier now. 2697 if (is_store) { 2698 insert_mem_bar(Op_MemBarRelease); 2699 } else { 2700 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2701 insert_mem_bar(Op_MemBarVolatile); 2702 } 2703 } 2704 } 2705 2706 // Memory barrier to prevent normal and 'unsafe' accesses from 2707 // bypassing each other. Happens after null checks, so the 2708 // exception paths do not take memory state from the memory barrier, 2709 // so there's no problems making a strong assert about mixing users 2710 // of safe & unsafe memory. 2711 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2712 2713 if (!is_store) { 2714 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2715 // To be valid, unsafe loads may depend on other conditions than 2716 // the one that guards them: pin the Load node 2717 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile); 2718 // load value 2719 switch (type) { 2720 case T_BOOLEAN: 2721 case T_CHAR: 2722 case T_BYTE: 2723 case T_SHORT: 2724 case T_INT: 2725 case T_LONG: 2726 case T_FLOAT: 2727 case T_DOUBLE: 2728 break; 2729 case T_OBJECT: 2730 if (need_read_barrier) { 2731 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2732 } 2733 break; 2734 case T_ADDRESS: 2735 // Cast to an int type. 2736 p = _gvn.transform(new CastP2XNode(NULL, p)); 2737 p = ConvX2UL(p); 2738 break; 2739 default: 2740 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2741 break; 2742 } 2743 // The load node has the control of the preceding MemBarCPUOrder. All 2744 // following nodes will have the control of the MemBarCPUOrder inserted at 2745 // the end of this method. So, pushing the load onto the stack at a later 2746 // point is fine. 2747 set_result(p); 2748 } else { 2749 // place effect of store into memory 2750 switch (type) { 2751 case T_DOUBLE: 2752 val = dstore_rounding(val); 2753 break; 2754 case T_ADDRESS: 2755 // Repackage the long as a pointer. 2756 val = ConvL2X(val); 2757 val = _gvn.transform(new CastX2PNode(val)); 2758 break; 2759 } 2760 2761 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2762 if (type != T_OBJECT ) { 2763 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2764 } else { 2765 // Possibly an oop being stored to Java heap or native memory 2766 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2767 // oop to Java heap. 2768 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2769 } else { 2770 // We can't tell at compile time if we are storing in the Java heap or outside 2771 // of it. So we need to emit code to conditionally do the proper type of 2772 // store. 2773 2774 IdealKit ideal(this); 2775#define __ ideal. 2776 // QQQ who knows what probability is here?? 2777 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2778 // Sync IdealKit and graphKit. 2779 sync_kit(ideal); 2780 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2781 // Update IdealKit memory. 2782 __ sync_kit(this); 2783 } __ else_(); { 2784 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2785 } __ end_if(); 2786 // Final sync IdealKit and GraphKit. 2787 final_sync(ideal); 2788#undef __ 2789 } 2790 } 2791 } 2792 2793 if (is_volatile) { 2794 if (!is_store) { 2795 insert_mem_bar(Op_MemBarAcquire); 2796 } else { 2797 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2798 insert_mem_bar(Op_MemBarVolatile); 2799 } 2800 } 2801 } 2802 2803 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2804 2805 return true; 2806} 2807 2808//----------------------------inline_unsafe_load_store---------------------------- 2809// This method serves a couple of different customers (depending on LoadStoreKind): 2810// 2811// LS_cmpxchg: 2812// public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2813// public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2814// public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2815// 2816// LS_xadd: 2817// public int getAndAddInt( Object o, long offset, int delta) 2818// public long getAndAddLong(Object o, long offset, long delta) 2819// 2820// LS_xchg: 2821// int getAndSet(Object o, long offset, int newValue) 2822// long getAndSet(Object o, long offset, long newValue) 2823// Object getAndSet(Object o, long offset, Object newValue) 2824// 2825bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2826 // This basic scheme here is the same as inline_unsafe_access, but 2827 // differs in enough details that combining them would make the code 2828 // overly confusing. (This is a true fact! I originally combined 2829 // them, but even I was confused by it!) As much code/comments as 2830 // possible are retained from inline_unsafe_access though to make 2831 // the correspondences clearer. - dl 2832 2833 if (callee()->is_static()) return false; // caller must have the capability! 2834 2835#ifndef PRODUCT 2836 BasicType rtype; 2837 { 2838 ResourceMark rm; 2839 // Check the signatures. 2840 ciSignature* sig = callee()->signature(); 2841 rtype = sig->return_type()->basic_type(); 2842 if (kind == LS_xadd || kind == LS_xchg) { 2843 // Check the signatures. 2844#ifdef ASSERT 2845 assert(rtype == type, "get and set must return the expected type"); 2846 assert(sig->count() == 3, "get and set has 3 arguments"); 2847 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2848 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2849 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2850#endif // ASSERT 2851 } else if (kind == LS_cmpxchg) { 2852 // Check the signatures. 2853#ifdef ASSERT 2854 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2855 assert(sig->count() == 4, "CAS has 4 arguments"); 2856 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2857 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2858#endif // ASSERT 2859 } else { 2860 ShouldNotReachHere(); 2861 } 2862 } 2863#endif //PRODUCT 2864 2865 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2866 2867 // Get arguments: 2868 Node* receiver = NULL; 2869 Node* base = NULL; 2870 Node* offset = NULL; 2871 Node* oldval = NULL; 2872 Node* newval = NULL; 2873 if (kind == LS_cmpxchg) { 2874 const bool two_slot_type = type2size[type] == 2; 2875 receiver = argument(0); // type: oop 2876 base = argument(1); // type: oop 2877 offset = argument(2); // type: long 2878 oldval = argument(4); // type: oop, int, or long 2879 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2880 } else if (kind == LS_xadd || kind == LS_xchg){ 2881 receiver = argument(0); // type: oop 2882 base = argument(1); // type: oop 2883 offset = argument(2); // type: long 2884 oldval = NULL; 2885 newval = argument(4); // type: oop, int, or long 2886 } 2887 2888 // Null check receiver. 2889 receiver = null_check(receiver); 2890 if (stopped()) { 2891 return true; 2892 } 2893 2894 // Build field offset expression. 2895 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2896 // to be plain byte offsets, which are also the same as those accepted 2897 // by oopDesc::field_base. 2898 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2899 // 32-bit machines ignore the high half of long offsets 2900 offset = ConvL2X(offset); 2901 Node* adr = make_unsafe_address(base, offset); 2902 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2903 2904 // For CAS, unlike inline_unsafe_access, there seems no point in 2905 // trying to refine types. Just use the coarse types here. 2906 const Type *value_type = Type::get_const_basic_type(type); 2907 Compile::AliasType* alias_type = C->alias_type(adr_type); 2908 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2909 2910 if (kind == LS_xchg && type == T_OBJECT) { 2911 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2912 if (tjp != NULL) { 2913 value_type = tjp; 2914 } 2915 } 2916 2917 int alias_idx = C->get_alias_index(adr_type); 2918 2919 // Memory-model-wise, a LoadStore acts like a little synchronized 2920 // block, so needs barriers on each side. These don't translate 2921 // into actual barriers on most machines, but we still need rest of 2922 // compiler to respect ordering. 2923 2924 insert_mem_bar(Op_MemBarRelease); 2925 insert_mem_bar(Op_MemBarCPUOrder); 2926 2927 // 4984716: MemBars must be inserted before this 2928 // memory node in order to avoid a false 2929 // dependency which will confuse the scheduler. 2930 Node *mem = memory(alias_idx); 2931 2932 // For now, we handle only those cases that actually exist: ints, 2933 // longs, and Object. Adding others should be straightforward. 2934 Node* load_store; 2935 switch(type) { 2936 case T_INT: 2937 if (kind == LS_xadd) { 2938 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2939 } else if (kind == LS_xchg) { 2940 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2941 } else if (kind == LS_cmpxchg) { 2942 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2943 } else { 2944 ShouldNotReachHere(); 2945 } 2946 break; 2947 case T_LONG: 2948 if (kind == LS_xadd) { 2949 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2950 } else if (kind == LS_xchg) { 2951 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2952 } else if (kind == LS_cmpxchg) { 2953 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2954 } else { 2955 ShouldNotReachHere(); 2956 } 2957 break; 2958 case T_OBJECT: 2959 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2960 // could be delayed during Parse (for example, in adjust_map_after_if()). 2961 // Execute transformation here to avoid barrier generation in such case. 2962 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2963 newval = _gvn.makecon(TypePtr::NULL_PTR); 2964 2965 // Reference stores need a store barrier. 2966 if (kind == LS_xchg) { 2967 // If pre-barrier must execute before the oop store, old value will require do_load here. 2968 if (!can_move_pre_barrier()) { 2969 pre_barrier(true /* do_load*/, 2970 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2971 NULL /* pre_val*/, 2972 T_OBJECT); 2973 } // Else move pre_barrier to use load_store value, see below. 2974 } else if (kind == LS_cmpxchg) { 2975 // Same as for newval above: 2976 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2977 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2978 } 2979 // The only known value which might get overwritten is oldval. 2980 pre_barrier(false /* do_load */, 2981 control(), NULL, NULL, max_juint, NULL, NULL, 2982 oldval /* pre_val */, 2983 T_OBJECT); 2984 } else { 2985 ShouldNotReachHere(); 2986 } 2987 2988#ifdef _LP64 2989 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2990 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2991 if (kind == LS_xchg) { 2992 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2993 newval_enc, adr_type, value_type->make_narrowoop())); 2994 } else { 2995 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2996 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2997 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2998 newval_enc, oldval_enc)); 2999 } 3000 } else 3001#endif 3002 { 3003 if (kind == LS_xchg) { 3004 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 3005 } else { 3006 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 3007 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 3008 } 3009 } 3010 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 3011 break; 3012 default: 3013 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 3014 break; 3015 } 3016 3017 // SCMemProjNodes represent the memory state of a LoadStore. Their 3018 // main role is to prevent LoadStore nodes from being optimized away 3019 // when their results aren't used. 3020 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 3021 set_memory(proj, alias_idx); 3022 3023 if (type == T_OBJECT && kind == LS_xchg) { 3024#ifdef _LP64 3025 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3026 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 3027 } 3028#endif 3029 if (can_move_pre_barrier()) { 3030 // Don't need to load pre_val. The old value is returned by load_store. 3031 // The pre_barrier can execute after the xchg as long as no safepoint 3032 // gets inserted between them. 3033 pre_barrier(false /* do_load */, 3034 control(), NULL, NULL, max_juint, NULL, NULL, 3035 load_store /* pre_val */, 3036 T_OBJECT); 3037 } 3038 } 3039 3040 // Add the trailing membar surrounding the access 3041 insert_mem_bar(Op_MemBarCPUOrder); 3042 insert_mem_bar(Op_MemBarAcquire); 3043 3044 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3045 set_result(load_store); 3046 return true; 3047} 3048 3049//----------------------------inline_unsafe_ordered_store---------------------- 3050// public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3051// public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3052// public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3053bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3054 // This is another variant of inline_unsafe_access, differing in 3055 // that it always issues store-store ("release") barrier and ensures 3056 // store-atomicity (which only matters for "long"). 3057 3058 if (callee()->is_static()) return false; // caller must have the capability! 3059 3060#ifndef PRODUCT 3061 { 3062 ResourceMark rm; 3063 // Check the signatures. 3064 ciSignature* sig = callee()->signature(); 3065#ifdef ASSERT 3066 BasicType rtype = sig->return_type()->basic_type(); 3067 assert(rtype == T_VOID, "must return void"); 3068 assert(sig->count() == 3, "has 3 arguments"); 3069 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3070 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3071#endif // ASSERT 3072 } 3073#endif //PRODUCT 3074 3075 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3076 3077 // Get arguments: 3078 Node* receiver = argument(0); // type: oop 3079 Node* base = argument(1); // type: oop 3080 Node* offset = argument(2); // type: long 3081 Node* val = argument(4); // type: oop, int, or long 3082 3083 // Null check receiver. 3084 receiver = null_check(receiver); 3085 if (stopped()) { 3086 return true; 3087 } 3088 3089 // Build field offset expression. 3090 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3091 // 32-bit machines ignore the high half of long offsets 3092 offset = ConvL2X(offset); 3093 Node* adr = make_unsafe_address(base, offset); 3094 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3095 const Type *value_type = Type::get_const_basic_type(type); 3096 Compile::AliasType* alias_type = C->alias_type(adr_type); 3097 3098 insert_mem_bar(Op_MemBarRelease); 3099 insert_mem_bar(Op_MemBarCPUOrder); 3100 // Ensure that the store is atomic for longs: 3101 const bool require_atomic_access = true; 3102 Node* store; 3103 if (type == T_OBJECT) // reference stores need a store barrier. 3104 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3105 else { 3106 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3107 } 3108 insert_mem_bar(Op_MemBarCPUOrder); 3109 return true; 3110} 3111 3112bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3113 // Regardless of form, don't allow previous ld/st to move down, 3114 // then issue acquire, release, or volatile mem_bar. 3115 insert_mem_bar(Op_MemBarCPUOrder); 3116 switch(id) { 3117 case vmIntrinsics::_loadFence: 3118 insert_mem_bar(Op_LoadFence); 3119 return true; 3120 case vmIntrinsics::_storeFence: 3121 insert_mem_bar(Op_StoreFence); 3122 return true; 3123 case vmIntrinsics::_fullFence: 3124 insert_mem_bar(Op_MemBarVolatile); 3125 return true; 3126 default: 3127 fatal_unexpected_iid(id); 3128 return false; 3129 } 3130} 3131 3132bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3133 if (!kls->is_Con()) { 3134 return true; 3135 } 3136 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3137 if (klsptr == NULL) { 3138 return true; 3139 } 3140 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3141 // don't need a guard for a klass that is already initialized 3142 return !ik->is_initialized(); 3143} 3144 3145//----------------------------inline_unsafe_allocate--------------------------- 3146// public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3147bool LibraryCallKit::inline_unsafe_allocate() { 3148 if (callee()->is_static()) return false; // caller must have the capability! 3149 3150 null_check_receiver(); // null-check, then ignore 3151 Node* cls = null_check(argument(1)); 3152 if (stopped()) return true; 3153 3154 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3155 kls = null_check(kls); 3156 if (stopped()) return true; // argument was like int.class 3157 3158 Node* test = NULL; 3159 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3160 // Note: The argument might still be an illegal value like 3161 // Serializable.class or Object[].class. The runtime will handle it. 3162 // But we must make an explicit check for initialization. 3163 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3164 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3165 // can generate code to load it as unsigned byte. 3166 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3167 Node* bits = intcon(InstanceKlass::fully_initialized); 3168 test = _gvn.transform(new SubINode(inst, bits)); 3169 // The 'test' is non-zero if we need to take a slow path. 3170 } 3171 3172 Node* obj = new_instance(kls, test); 3173 set_result(obj); 3174 return true; 3175} 3176 3177#ifdef TRACE_HAVE_INTRINSICS 3178/* 3179 * oop -> myklass 3180 * myklass->trace_id |= USED 3181 * return myklass->trace_id & ~0x3 3182 */ 3183bool LibraryCallKit::inline_native_classID() { 3184 null_check_receiver(); // null-check, then ignore 3185 Node* cls = null_check(argument(1), T_OBJECT); 3186 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3187 kls = null_check(kls, T_OBJECT); 3188 ByteSize offset = TRACE_ID_OFFSET; 3189 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3190 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3191 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3192 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3193 Node* clsused = longcon(0x01l); // set the class bit 3194 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3195 3196 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3197 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3198 set_result(andl); 3199 return true; 3200} 3201 3202bool LibraryCallKit::inline_native_threadID() { 3203 Node* tls_ptr = NULL; 3204 Node* cur_thr = generate_current_thread(tls_ptr); 3205 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3206 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3207 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3208 3209 Node* threadid = NULL; 3210 size_t thread_id_size = OSThread::thread_id_size(); 3211 if (thread_id_size == (size_t) BytesPerLong) { 3212 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3213 } else if (thread_id_size == (size_t) BytesPerInt) { 3214 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3215 } else { 3216 ShouldNotReachHere(); 3217 } 3218 set_result(threadid); 3219 return true; 3220} 3221#endif 3222 3223//------------------------inline_native_time_funcs-------------- 3224// inline code for System.currentTimeMillis() and System.nanoTime() 3225// these have the same type and signature 3226bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3227 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3228 const TypePtr* no_memory_effects = NULL; 3229 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3230 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3231#ifdef ASSERT 3232 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3233 assert(value_top == top(), "second value must be top"); 3234#endif 3235 set_result(value); 3236 return true; 3237} 3238 3239//------------------------inline_native_currentThread------------------ 3240bool LibraryCallKit::inline_native_currentThread() { 3241 Node* junk = NULL; 3242 set_result(generate_current_thread(junk)); 3243 return true; 3244} 3245 3246//------------------------inline_native_isInterrupted------------------ 3247// private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3248bool LibraryCallKit::inline_native_isInterrupted() { 3249 // Add a fast path to t.isInterrupted(clear_int): 3250 // (t == Thread.current() && 3251 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3252 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3253 // So, in the common case that the interrupt bit is false, 3254 // we avoid making a call into the VM. Even if the interrupt bit 3255 // is true, if the clear_int argument is false, we avoid the VM call. 3256 // However, if the receiver is not currentThread, we must call the VM, 3257 // because there must be some locking done around the operation. 3258 3259 // We only go to the fast case code if we pass two guards. 3260 // Paths which do not pass are accumulated in the slow_region. 3261 3262 enum { 3263 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3264 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3265 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3266 PATH_LIMIT 3267 }; 3268 3269 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3270 // out of the function. 3271 insert_mem_bar(Op_MemBarCPUOrder); 3272 3273 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3274 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3275 3276 RegionNode* slow_region = new RegionNode(1); 3277 record_for_igvn(slow_region); 3278 3279 // (a) Receiving thread must be the current thread. 3280 Node* rec_thr = argument(0); 3281 Node* tls_ptr = NULL; 3282 Node* cur_thr = generate_current_thread(tls_ptr); 3283 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3284 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3285 3286 generate_slow_guard(bol_thr, slow_region); 3287 3288 // (b) Interrupt bit on TLS must be false. 3289 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3290 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3291 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3292 3293 // Set the control input on the field _interrupted read to prevent it floating up. 3294 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3295 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3296 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3297 3298 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3299 3300 // First fast path: if (!TLS._interrupted) return false; 3301 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3302 result_rgn->init_req(no_int_result_path, false_bit); 3303 result_val->init_req(no_int_result_path, intcon(0)); 3304 3305 // drop through to next case 3306 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3307 3308#ifndef TARGET_OS_FAMILY_windows 3309 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3310 Node* clr_arg = argument(1); 3311 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3312 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3313 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3314 3315 // Second fast path: ... else if (!clear_int) return true; 3316 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3317 result_rgn->init_req(no_clear_result_path, false_arg); 3318 result_val->init_req(no_clear_result_path, intcon(1)); 3319 3320 // drop through to next case 3321 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3322#else 3323 // To return true on Windows you must read the _interrupted field 3324 // and check the the event state i.e. take the slow path. 3325#endif // TARGET_OS_FAMILY_windows 3326 3327 // (d) Otherwise, go to the slow path. 3328 slow_region->add_req(control()); 3329 set_control( _gvn.transform(slow_region)); 3330 3331 if (stopped()) { 3332 // There is no slow path. 3333 result_rgn->init_req(slow_result_path, top()); 3334 result_val->init_req(slow_result_path, top()); 3335 } else { 3336 // non-virtual because it is a private non-static 3337 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3338 3339 Node* slow_val = set_results_for_java_call(slow_call); 3340 // this->control() comes from set_results_for_java_call 3341 3342 Node* fast_io = slow_call->in(TypeFunc::I_O); 3343 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3344 3345 // These two phis are pre-filled with copies of of the fast IO and Memory 3346 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3347 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3348 3349 result_rgn->init_req(slow_result_path, control()); 3350 result_io ->init_req(slow_result_path, i_o()); 3351 result_mem->init_req(slow_result_path, reset_memory()); 3352 result_val->init_req(slow_result_path, slow_val); 3353 3354 set_all_memory(_gvn.transform(result_mem)); 3355 set_i_o( _gvn.transform(result_io)); 3356 } 3357 3358 C->set_has_split_ifs(true); // Has chance for split-if optimization 3359 set_result(result_rgn, result_val); 3360 return true; 3361} 3362 3363//---------------------------load_mirror_from_klass---------------------------- 3364// Given a klass oop, load its java mirror (a java.lang.Class oop). 3365Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3366 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3367 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3368} 3369 3370//-----------------------load_klass_from_mirror_common------------------------- 3371// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3372// Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3373// and branch to the given path on the region. 3374// If never_see_null, take an uncommon trap on null, so we can optimistically 3375// compile for the non-null case. 3376// If the region is NULL, force never_see_null = true. 3377Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3378 bool never_see_null, 3379 RegionNode* region, 3380 int null_path, 3381 int offset) { 3382 if (region == NULL) never_see_null = true; 3383 Node* p = basic_plus_adr(mirror, offset); 3384 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3385 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3386 Node* null_ctl = top(); 3387 kls = null_check_oop(kls, &null_ctl, never_see_null); 3388 if (region != NULL) { 3389 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3390 region->init_req(null_path, null_ctl); 3391 } else { 3392 assert(null_ctl == top(), "no loose ends"); 3393 } 3394 return kls; 3395} 3396 3397//--------------------(inline_native_Class_query helpers)--------------------- 3398// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3399// Fall through if (mods & mask) == bits, take the guard otherwise. 3400Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3401 // Branch around if the given klass has the given modifier bit set. 3402 // Like generate_guard, adds a new path onto the region. 3403 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3404 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3405 Node* mask = intcon(modifier_mask); 3406 Node* bits = intcon(modifier_bits); 3407 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3408 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3409 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3410 return generate_fair_guard(bol, region); 3411} 3412Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3413 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3414} 3415 3416//-------------------------inline_native_Class_query------------------- 3417bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3418 const Type* return_type = TypeInt::BOOL; 3419 Node* prim_return_value = top(); // what happens if it's a primitive class? 3420 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3421 bool expect_prim = false; // most of these guys expect to work on refs 3422 3423 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3424 3425 Node* mirror = argument(0); 3426 Node* obj = top(); 3427 3428 switch (id) { 3429 case vmIntrinsics::_isInstance: 3430 // nothing is an instance of a primitive type 3431 prim_return_value = intcon(0); 3432 obj = argument(1); 3433 break; 3434 case vmIntrinsics::_getModifiers: 3435 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3436 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3437 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3438 break; 3439 case vmIntrinsics::_isInterface: 3440 prim_return_value = intcon(0); 3441 break; 3442 case vmIntrinsics::_isArray: 3443 prim_return_value = intcon(0); 3444 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3445 break; 3446 case vmIntrinsics::_isPrimitive: 3447 prim_return_value = intcon(1); 3448 expect_prim = true; // obviously 3449 break; 3450 case vmIntrinsics::_getSuperclass: 3451 prim_return_value = null(); 3452 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3453 break; 3454 case vmIntrinsics::_getClassAccessFlags: 3455 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3456 return_type = TypeInt::INT; // not bool! 6297094 3457 break; 3458 default: 3459 fatal_unexpected_iid(id); 3460 break; 3461 } 3462 3463 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3464 if (mirror_con == NULL) return false; // cannot happen? 3465 3466#ifndef PRODUCT 3467 if (C->print_intrinsics() || C->print_inlining()) { 3468 ciType* k = mirror_con->java_mirror_type(); 3469 if (k) { 3470 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3471 k->print_name(); 3472 tty->cr(); 3473 } 3474 } 3475#endif 3476 3477 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3478 RegionNode* region = new RegionNode(PATH_LIMIT); 3479 record_for_igvn(region); 3480 PhiNode* phi = new PhiNode(region, return_type); 3481 3482 // The mirror will never be null of Reflection.getClassAccessFlags, however 3483 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3484 // if it is. See bug 4774291. 3485 3486 // For Reflection.getClassAccessFlags(), the null check occurs in 3487 // the wrong place; see inline_unsafe_access(), above, for a similar 3488 // situation. 3489 mirror = null_check(mirror); 3490 // If mirror or obj is dead, only null-path is taken. 3491 if (stopped()) return true; 3492 3493 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3494 3495 // Now load the mirror's klass metaobject, and null-check it. 3496 // Side-effects region with the control path if the klass is null. 3497 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3498 // If kls is null, we have a primitive mirror. 3499 phi->init_req(_prim_path, prim_return_value); 3500 if (stopped()) { set_result(region, phi); return true; } 3501 bool safe_for_replace = (region->in(_prim_path) == top()); 3502 3503 Node* p; // handy temp 3504 Node* null_ctl; 3505 3506 // Now that we have the non-null klass, we can perform the real query. 3507 // For constant classes, the query will constant-fold in LoadNode::Value. 3508 Node* query_value = top(); 3509 switch (id) { 3510 case vmIntrinsics::_isInstance: 3511 // nothing is an instance of a primitive type 3512 query_value = gen_instanceof(obj, kls, safe_for_replace); 3513 break; 3514 3515 case vmIntrinsics::_getModifiers: 3516 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3517 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3518 break; 3519 3520 case vmIntrinsics::_isInterface: 3521 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3522 if (generate_interface_guard(kls, region) != NULL) 3523 // A guard was added. If the guard is taken, it was an interface. 3524 phi->add_req(intcon(1)); 3525 // If we fall through, it's a plain class. 3526 query_value = intcon(0); 3527 break; 3528 3529 case vmIntrinsics::_isArray: 3530 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3531 if (generate_array_guard(kls, region) != NULL) 3532 // A guard was added. If the guard is taken, it was an array. 3533 phi->add_req(intcon(1)); 3534 // If we fall through, it's a plain class. 3535 query_value = intcon(0); 3536 break; 3537 3538 case vmIntrinsics::_isPrimitive: 3539 query_value = intcon(0); // "normal" path produces false 3540 break; 3541 3542 case vmIntrinsics::_getSuperclass: 3543 // The rules here are somewhat unfortunate, but we can still do better 3544 // with random logic than with a JNI call. 3545 // Interfaces store null or Object as _super, but must report null. 3546 // Arrays store an intermediate super as _super, but must report Object. 3547 // Other types can report the actual _super. 3548 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3549 if (generate_interface_guard(kls, region) != NULL) 3550 // A guard was added. If the guard is taken, it was an interface. 3551 phi->add_req(null()); 3552 if (generate_array_guard(kls, region) != NULL) 3553 // A guard was added. If the guard is taken, it was an array. 3554 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3555 // If we fall through, it's a plain class. Get its _super. 3556 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3557 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3558 null_ctl = top(); 3559 kls = null_check_oop(kls, &null_ctl); 3560 if (null_ctl != top()) { 3561 // If the guard is taken, Object.superClass is null (both klass and mirror). 3562 region->add_req(null_ctl); 3563 phi ->add_req(null()); 3564 } 3565 if (!stopped()) { 3566 query_value = load_mirror_from_klass(kls); 3567 } 3568 break; 3569 3570 case vmIntrinsics::_getClassAccessFlags: 3571 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3572 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3573 break; 3574 3575 default: 3576 fatal_unexpected_iid(id); 3577 break; 3578 } 3579 3580 // Fall-through is the normal case of a query to a real class. 3581 phi->init_req(1, query_value); 3582 region->init_req(1, control()); 3583 3584 C->set_has_split_ifs(true); // Has chance for split-if optimization 3585 set_result(region, phi); 3586 return true; 3587} 3588 3589//-------------------------inline_Class_cast------------------- 3590bool LibraryCallKit::inline_Class_cast() { 3591 Node* mirror = argument(0); // Class 3592 Node* obj = argument(1); 3593 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3594 if (mirror_con == NULL) { 3595 return false; // dead path (mirror->is_top()). 3596 } 3597 if (obj == NULL || obj->is_top()) { 3598 return false; // dead path 3599 } 3600 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3601 3602 // First, see if Class.cast() can be folded statically. 3603 // java_mirror_type() returns non-null for compile-time Class constants. 3604 ciType* tm = mirror_con->java_mirror_type(); 3605 if (tm != NULL && tm->is_klass() && 3606 tp != NULL && tp->klass() != NULL) { 3607 if (!tp->klass()->is_loaded()) { 3608 // Don't use intrinsic when class is not loaded. 3609 return false; 3610 } else { 3611 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3612 if (static_res == Compile::SSC_always_true) { 3613 // isInstance() is true - fold the code. 3614 set_result(obj); 3615 return true; 3616 } else if (static_res == Compile::SSC_always_false) { 3617 // Don't use intrinsic, have to throw ClassCastException. 3618 // If the reference is null, the non-intrinsic bytecode will 3619 // be optimized appropriately. 3620 return false; 3621 } 3622 } 3623 } 3624 3625 // Bailout intrinsic and do normal inlining if exception path is frequent. 3626 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3627 return false; 3628 } 3629 3630 // Generate dynamic checks. 3631 // Class.cast() is java implementation of _checkcast bytecode. 3632 // Do checkcast (Parse::do_checkcast()) optimizations here. 3633 3634 mirror = null_check(mirror); 3635 // If mirror is dead, only null-path is taken. 3636 if (stopped()) { 3637 return true; 3638 } 3639 3640 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3641 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3642 RegionNode* region = new RegionNode(PATH_LIMIT); 3643 record_for_igvn(region); 3644 3645 // Now load the mirror's klass metaobject, and null-check it. 3646 // If kls is null, we have a primitive mirror and 3647 // nothing is an instance of a primitive type. 3648 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3649 3650 Node* res = top(); 3651 if (!stopped()) { 3652 Node* bad_type_ctrl = top(); 3653 // Do checkcast optimizations. 3654 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3655 region->init_req(_bad_type_path, bad_type_ctrl); 3656 } 3657 if (region->in(_prim_path) != top() || 3658 region->in(_bad_type_path) != top()) { 3659 // Let Interpreter throw ClassCastException. 3660 PreserveJVMState pjvms(this); 3661 set_control(_gvn.transform(region)); 3662 uncommon_trap(Deoptimization::Reason_intrinsic, 3663 Deoptimization::Action_maybe_recompile); 3664 } 3665 if (!stopped()) { 3666 set_result(res); 3667 } 3668 return true; 3669} 3670 3671 3672//--------------------------inline_native_subtype_check------------------------ 3673// This intrinsic takes the JNI calls out of the heart of 3674// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3675bool LibraryCallKit::inline_native_subtype_check() { 3676 // Pull both arguments off the stack. 3677 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3678 args[0] = argument(0); 3679 args[1] = argument(1); 3680 Node* klasses[2]; // corresponding Klasses: superk, subk 3681 klasses[0] = klasses[1] = top(); 3682 3683 enum { 3684 // A full decision tree on {superc is prim, subc is prim}: 3685 _prim_0_path = 1, // {P,N} => false 3686 // {P,P} & superc!=subc => false 3687 _prim_same_path, // {P,P} & superc==subc => true 3688 _prim_1_path, // {N,P} => false 3689 _ref_subtype_path, // {N,N} & subtype check wins => true 3690 _both_ref_path, // {N,N} & subtype check loses => false 3691 PATH_LIMIT 3692 }; 3693 3694 RegionNode* region = new RegionNode(PATH_LIMIT); 3695 Node* phi = new PhiNode(region, TypeInt::BOOL); 3696 record_for_igvn(region); 3697 3698 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3699 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3700 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3701 3702 // First null-check both mirrors and load each mirror's klass metaobject. 3703 int which_arg; 3704 for (which_arg = 0; which_arg <= 1; which_arg++) { 3705 Node* arg = args[which_arg]; 3706 arg = null_check(arg); 3707 if (stopped()) break; 3708 args[which_arg] = arg; 3709 3710 Node* p = basic_plus_adr(arg, class_klass_offset); 3711 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3712 klasses[which_arg] = _gvn.transform(kls); 3713 } 3714 3715 // Having loaded both klasses, test each for null. 3716 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3717 for (which_arg = 0; which_arg <= 1; which_arg++) { 3718 Node* kls = klasses[which_arg]; 3719 Node* null_ctl = top(); 3720 kls = null_check_oop(kls, &null_ctl, never_see_null); 3721 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3722 region->init_req(prim_path, null_ctl); 3723 if (stopped()) break; 3724 klasses[which_arg] = kls; 3725 } 3726 3727 if (!stopped()) { 3728 // now we have two reference types, in klasses[0..1] 3729 Node* subk = klasses[1]; // the argument to isAssignableFrom 3730 Node* superk = klasses[0]; // the receiver 3731 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3732 // now we have a successful reference subtype check 3733 region->set_req(_ref_subtype_path, control()); 3734 } 3735 3736 // If both operands are primitive (both klasses null), then 3737 // we must return true when they are identical primitives. 3738 // It is convenient to test this after the first null klass check. 3739 set_control(region->in(_prim_0_path)); // go back to first null check 3740 if (!stopped()) { 3741 // Since superc is primitive, make a guard for the superc==subc case. 3742 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3743 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3744 generate_guard(bol_eq, region, PROB_FAIR); 3745 if (region->req() == PATH_LIMIT+1) { 3746 // A guard was added. If the added guard is taken, superc==subc. 3747 region->swap_edges(PATH_LIMIT, _prim_same_path); 3748 region->del_req(PATH_LIMIT); 3749 } 3750 region->set_req(_prim_0_path, control()); // Not equal after all. 3751 } 3752 3753 // these are the only paths that produce 'true': 3754 phi->set_req(_prim_same_path, intcon(1)); 3755 phi->set_req(_ref_subtype_path, intcon(1)); 3756 3757 // pull together the cases: 3758 assert(region->req() == PATH_LIMIT, "sane region"); 3759 for (uint i = 1; i < region->req(); i++) { 3760 Node* ctl = region->in(i); 3761 if (ctl == NULL || ctl == top()) { 3762 region->set_req(i, top()); 3763 phi ->set_req(i, top()); 3764 } else if (phi->in(i) == NULL) { 3765 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3766 } 3767 } 3768 3769 set_control(_gvn.transform(region)); 3770 set_result(_gvn.transform(phi)); 3771 return true; 3772} 3773 3774//---------------------generate_array_guard_common------------------------ 3775Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3776 bool obj_array, bool not_array) { 3777 3778 if (stopped()) { 3779 return NULL; 3780 } 3781 3782 // If obj_array/non_array==false/false: 3783 // Branch around if the given klass is in fact an array (either obj or prim). 3784 // If obj_array/non_array==false/true: 3785 // Branch around if the given klass is not an array klass of any kind. 3786 // If obj_array/non_array==true/true: 3787 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3788 // If obj_array/non_array==true/false: 3789 // Branch around if the kls is an oop array (Object[] or subtype) 3790 // 3791 // Like generate_guard, adds a new path onto the region. 3792 jint layout_con = 0; 3793 Node* layout_val = get_layout_helper(kls, layout_con); 3794 if (layout_val == NULL) { 3795 bool query = (obj_array 3796 ? Klass::layout_helper_is_objArray(layout_con) 3797 : Klass::layout_helper_is_array(layout_con)); 3798 if (query == not_array) { 3799 return NULL; // never a branch 3800 } else { // always a branch 3801 Node* always_branch = control(); 3802 if (region != NULL) 3803 region->add_req(always_branch); 3804 set_control(top()); 3805 return always_branch; 3806 } 3807 } 3808 // Now test the correct condition. 3809 jint nval = (obj_array 3810 ? ((jint)Klass::_lh_array_tag_type_value 3811 << Klass::_lh_array_tag_shift) 3812 : Klass::_lh_neutral_value); 3813 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3814 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3815 // invert the test if we are looking for a non-array 3816 if (not_array) btest = BoolTest(btest).negate(); 3817 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3818 return generate_fair_guard(bol, region); 3819} 3820 3821 3822//-----------------------inline_native_newArray-------------------------- 3823// private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3824bool LibraryCallKit::inline_native_newArray() { 3825 Node* mirror = argument(0); 3826 Node* count_val = argument(1); 3827 3828 mirror = null_check(mirror); 3829 // If mirror or obj is dead, only null-path is taken. 3830 if (stopped()) return true; 3831 3832 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3833 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3834 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3835 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3836 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3837 3838 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3839 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3840 result_reg, _slow_path); 3841 Node* normal_ctl = control(); 3842 Node* no_array_ctl = result_reg->in(_slow_path); 3843 3844 // Generate code for the slow case. We make a call to newArray(). 3845 set_control(no_array_ctl); 3846 if (!stopped()) { 3847 // Either the input type is void.class, or else the 3848 // array klass has not yet been cached. Either the 3849 // ensuing call will throw an exception, or else it 3850 // will cache the array klass for next time. 3851 PreserveJVMState pjvms(this); 3852 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3853 Node* slow_result = set_results_for_java_call(slow_call); 3854 // this->control() comes from set_results_for_java_call 3855 result_reg->set_req(_slow_path, control()); 3856 result_val->set_req(_slow_path, slow_result); 3857 result_io ->set_req(_slow_path, i_o()); 3858 result_mem->set_req(_slow_path, reset_memory()); 3859 } 3860 3861 set_control(normal_ctl); 3862 if (!stopped()) { 3863 // Normal case: The array type has been cached in the java.lang.Class. 3864 // The following call works fine even if the array type is polymorphic. 3865 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3866 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3867 result_reg->init_req(_normal_path, control()); 3868 result_val->init_req(_normal_path, obj); 3869 result_io ->init_req(_normal_path, i_o()); 3870 result_mem->init_req(_normal_path, reset_memory()); 3871 } 3872 3873 // Return the combined state. 3874 set_i_o( _gvn.transform(result_io) ); 3875 set_all_memory( _gvn.transform(result_mem)); 3876 3877 C->set_has_split_ifs(true); // Has chance for split-if optimization 3878 set_result(result_reg, result_val); 3879 return true; 3880} 3881 3882//----------------------inline_native_getLength-------------------------- 3883// public static native int java.lang.reflect.Array.getLength(Object array); 3884bool LibraryCallKit::inline_native_getLength() { 3885 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3886 3887 Node* array = null_check(argument(0)); 3888 // If array is dead, only null-path is taken. 3889 if (stopped()) return true; 3890 3891 // Deoptimize if it is a non-array. 3892 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3893 3894 if (non_array != NULL) { 3895 PreserveJVMState pjvms(this); 3896 set_control(non_array); 3897 uncommon_trap(Deoptimization::Reason_intrinsic, 3898 Deoptimization::Action_maybe_recompile); 3899 } 3900 3901 // If control is dead, only non-array-path is taken. 3902 if (stopped()) return true; 3903 3904 // The works fine even if the array type is polymorphic. 3905 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3906 Node* result = load_array_length(array); 3907 3908 C->set_has_split_ifs(true); // Has chance for split-if optimization 3909 set_result(result); 3910 return true; 3911} 3912 3913//------------------------inline_array_copyOf---------------------------- 3914// public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3915// public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3916bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3917 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3918 3919 // Get the arguments. 3920 Node* original = argument(0); 3921 Node* start = is_copyOfRange? argument(1): intcon(0); 3922 Node* end = is_copyOfRange? argument(2): argument(1); 3923 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3924 3925 Node* newcopy; 3926 3927 // Set the original stack and the reexecute bit for the interpreter to reexecute 3928 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3929 { PreserveReexecuteState preexecs(this); 3930 jvms()->set_should_reexecute(true); 3931 3932 array_type_mirror = null_check(array_type_mirror); 3933 original = null_check(original); 3934 3935 // Check if a null path was taken unconditionally. 3936 if (stopped()) return true; 3937 3938 Node* orig_length = load_array_length(original); 3939 3940 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3941 klass_node = null_check(klass_node); 3942 3943 RegionNode* bailout = new RegionNode(1); 3944 record_for_igvn(bailout); 3945 3946 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3947 // Bail out if that is so. 3948 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3949 if (not_objArray != NULL) { 3950 // Improve the klass node's type from the new optimistic assumption: 3951 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3952 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3953 Node* cast = new CastPPNode(klass_node, akls); 3954 cast->init_req(0, control()); 3955 klass_node = _gvn.transform(cast); 3956 } 3957 3958 // Bail out if either start or end is negative. 3959 generate_negative_guard(start, bailout, &start); 3960 generate_negative_guard(end, bailout, &end); 3961 3962 Node* length = end; 3963 if (_gvn.type(start) != TypeInt::ZERO) { 3964 length = _gvn.transform(new SubINode(end, start)); 3965 } 3966 3967 // Bail out if length is negative. 3968 // Without this the new_array would throw 3969 // NegativeArraySizeException but IllegalArgumentException is what 3970 // should be thrown 3971 generate_negative_guard(length, bailout, &length); 3972 3973 if (bailout->req() > 1) { 3974 PreserveJVMState pjvms(this); 3975 set_control(_gvn.transform(bailout)); 3976 uncommon_trap(Deoptimization::Reason_intrinsic, 3977 Deoptimization::Action_maybe_recompile); 3978 } 3979 3980 if (!stopped()) { 3981 // How many elements will we copy from the original? 3982 // The answer is MinI(orig_length - start, length). 3983 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3984 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3985 3986 // Generate a direct call to the right arraycopy function(s). 3987 // We know the copy is disjoint but we might not know if the 3988 // oop stores need checking. 3989 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3990 // This will fail a store-check if x contains any non-nulls. 3991 3992 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3993 // loads/stores but it is legal only if we're sure the 3994 // Arrays.copyOf would succeed. So we need all input arguments 3995 // to the copyOf to be validated, including that the copy to the 3996 // new array won't trigger an ArrayStoreException. That subtype 3997 // check can be optimized if we know something on the type of 3998 // the input array from type speculation. 3999 if (_gvn.type(klass_node)->singleton()) { 4000 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 4001 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 4002 4003 int test = C->static_subtype_check(superk, subk); 4004 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 4005 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 4006 if (t_original->speculative_type() != NULL) { 4007 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 4008 } 4009 } 4010 } 4011 4012 bool validated = false; 4013 // Reason_class_check rather than Reason_intrinsic because we 4014 // want to intrinsify even if this traps. 4015 if (!too_many_traps(Deoptimization::Reason_class_check)) { 4016 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 4017 klass_node); 4018 4019 if (not_subtype_ctrl != top()) { 4020 PreserveJVMState pjvms(this); 4021 set_control(not_subtype_ctrl); 4022 uncommon_trap(Deoptimization::Reason_class_check, 4023 Deoptimization::Action_make_not_entrant); 4024 assert(stopped(), "Should be stopped"); 4025 } 4026 validated = true; 4027 } 4028 4029 if (!stopped()) { 4030 newcopy = new_array(klass_node, length, 0); // no arguments to push 4031 4032 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 4033 load_object_klass(original), klass_node); 4034 if (!is_copyOfRange) { 4035 ac->set_copyof(validated); 4036 } else { 4037 ac->set_copyofrange(validated); 4038 } 4039 Node* n = _gvn.transform(ac); 4040 if (n == ac) { 4041 ac->connect_outputs(this); 4042 } else { 4043 assert(validated, "shouldn't transform if all arguments not validated"); 4044 set_all_memory(n); 4045 } 4046 } 4047 } 4048 } // original reexecute is set back here 4049 4050 C->set_has_split_ifs(true); // Has chance for split-if optimization 4051 if (!stopped()) { 4052 set_result(newcopy); 4053 } 4054 return true; 4055} 4056 4057 4058//----------------------generate_virtual_guard--------------------------- 4059// Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4060Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4061 RegionNode* slow_region) { 4062 ciMethod* method = callee(); 4063 int vtable_index = method->vtable_index(); 4064 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4065 err_msg_res("bad index %d", vtable_index)); 4066 // Get the Method* out of the appropriate vtable entry. 4067 int entry_offset = (InstanceKlass::vtable_start_offset() + 4068 vtable_index*vtableEntry::size()) * wordSize + 4069 vtableEntry::method_offset_in_bytes(); 4070 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4071 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4072 4073 // Compare the target method with the expected method (e.g., Object.hashCode). 4074 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4075 4076 Node* native_call = makecon(native_call_addr); 4077 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4078 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4079 4080 return generate_slow_guard(test_native, slow_region); 4081} 4082 4083//-----------------------generate_method_call---------------------------- 4084// Use generate_method_call to make a slow-call to the real 4085// method if the fast path fails. An alternative would be to 4086// use a stub like OptoRuntime::slow_arraycopy_Java. 4087// This only works for expanding the current library call, 4088// not another intrinsic. (E.g., don't use this for making an 4089// arraycopy call inside of the copyOf intrinsic.) 4090CallJavaNode* 4091LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4092 // When compiling the intrinsic method itself, do not use this technique. 4093 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4094 4095 ciMethod* method = callee(); 4096 // ensure the JVMS we have will be correct for this call 4097 guarantee(method_id == method->intrinsic_id(), "must match"); 4098 4099 const TypeFunc* tf = TypeFunc::make(method); 4100 CallJavaNode* slow_call; 4101 if (is_static) { 4102 assert(!is_virtual, ""); 4103 slow_call = new CallStaticJavaNode(C, tf, 4104 SharedRuntime::get_resolve_static_call_stub(), 4105 method, bci()); 4106 } else if (is_virtual) { 4107 null_check_receiver(); 4108 int vtable_index = Method::invalid_vtable_index; 4109 if (UseInlineCaches) { 4110 // Suppress the vtable call 4111 } else { 4112 // hashCode and clone are not a miranda methods, 4113 // so the vtable index is fixed. 4114 // No need to use the linkResolver to get it. 4115 vtable_index = method->vtable_index(); 4116 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4117 err_msg_res("bad index %d", vtable_index)); 4118 } 4119 slow_call = new CallDynamicJavaNode(tf, 4120 SharedRuntime::get_resolve_virtual_call_stub(), 4121 method, vtable_index, bci()); 4122 } else { // neither virtual nor static: opt_virtual 4123 null_check_receiver(); 4124 slow_call = new CallStaticJavaNode(C, tf, 4125 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4126 method, bci()); 4127 slow_call->set_optimized_virtual(true); 4128 } 4129 set_arguments_for_java_call(slow_call); 4130 set_edges_for_java_call(slow_call); 4131 return slow_call; 4132} 4133 4134 4135/** 4136 * Build special case code for calls to hashCode on an object. This call may 4137 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4138 * slightly different code. 4139 */ 4140bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4141 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4142 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4143 4144 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4145 4146 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4147 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4148 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4149 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4150 Node* obj = NULL; 4151 if (!is_static) { 4152 // Check for hashing null object 4153 obj = null_check_receiver(); 4154 if (stopped()) return true; // unconditionally null 4155 result_reg->init_req(_null_path, top()); 4156 result_val->init_req(_null_path, top()); 4157 } else { 4158 // Do a null check, and return zero if null. 4159 // System.identityHashCode(null) == 0 4160 obj = argument(0); 4161 Node* null_ctl = top(); 4162 obj = null_check_oop(obj, &null_ctl); 4163 result_reg->init_req(_null_path, null_ctl); 4164 result_val->init_req(_null_path, _gvn.intcon(0)); 4165 } 4166 4167 // Unconditionally null? Then return right away. 4168 if (stopped()) { 4169 set_control( result_reg->in(_null_path)); 4170 if (!stopped()) 4171 set_result(result_val->in(_null_path)); 4172 return true; 4173 } 4174 4175 // We only go to the fast case code if we pass a number of guards. The 4176 // paths which do not pass are accumulated in the slow_region. 4177 RegionNode* slow_region = new RegionNode(1); 4178 record_for_igvn(slow_region); 4179 4180 // If this is a virtual call, we generate a funny guard. We pull out 4181 // the vtable entry corresponding to hashCode() from the target object. 4182 // If the target method which we are calling happens to be the native 4183 // Object hashCode() method, we pass the guard. We do not need this 4184 // guard for non-virtual calls -- the caller is known to be the native 4185 // Object hashCode(). 4186 if (is_virtual) { 4187 // After null check, get the object's klass. 4188 Node* obj_klass = load_object_klass(obj); 4189 generate_virtual_guard(obj_klass, slow_region); 4190 } 4191 4192 // Get the header out of the object, use LoadMarkNode when available 4193 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4194 // The control of the load must be NULL. Otherwise, the load can move before 4195 // the null check after castPP removal. 4196 Node* no_ctrl = NULL; 4197 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4198 4199 // Test the header to see if it is unlocked. 4200 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4201 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4202 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4203 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4204 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4205 4206 generate_slow_guard(test_unlocked, slow_region); 4207 4208 // Get the hash value and check to see that it has been properly assigned. 4209 // We depend on hash_mask being at most 32 bits and avoid the use of 4210 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4211 // vm: see markOop.hpp. 4212 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4213 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4214 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4215 // This hack lets the hash bits live anywhere in the mark object now, as long 4216 // as the shift drops the relevant bits into the low 32 bits. Note that 4217 // Java spec says that HashCode is an int so there's no point in capturing 4218 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4219 hshifted_header = ConvX2I(hshifted_header); 4220 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4221 4222 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4223 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4224 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4225 4226 generate_slow_guard(test_assigned, slow_region); 4227 4228 Node* init_mem = reset_memory(); 4229 // fill in the rest of the null path: 4230 result_io ->init_req(_null_path, i_o()); 4231 result_mem->init_req(_null_path, init_mem); 4232 4233 result_val->init_req(_fast_path, hash_val); 4234 result_reg->init_req(_fast_path, control()); 4235 result_io ->init_req(_fast_path, i_o()); 4236 result_mem->init_req(_fast_path, init_mem); 4237 4238 // Generate code for the slow case. We make a call to hashCode(). 4239 set_control(_gvn.transform(slow_region)); 4240 if (!stopped()) { 4241 // No need for PreserveJVMState, because we're using up the present state. 4242 set_all_memory(init_mem); 4243 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4244 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4245 Node* slow_result = set_results_for_java_call(slow_call); 4246 // this->control() comes from set_results_for_java_call 4247 result_reg->init_req(_slow_path, control()); 4248 result_val->init_req(_slow_path, slow_result); 4249 result_io ->set_req(_slow_path, i_o()); 4250 result_mem ->set_req(_slow_path, reset_memory()); 4251 } 4252 4253 // Return the combined state. 4254 set_i_o( _gvn.transform(result_io) ); 4255 set_all_memory( _gvn.transform(result_mem)); 4256 4257 set_result(result_reg, result_val); 4258 return true; 4259} 4260 4261//---------------------------inline_native_getClass---------------------------- 4262// public final native Class<?> java.lang.Object.getClass(); 4263// 4264// Build special case code for calls to getClass on an object. 4265bool LibraryCallKit::inline_native_getClass() { 4266 Node* obj = null_check_receiver(); 4267 if (stopped()) return true; 4268 set_result(load_mirror_from_klass(load_object_klass(obj))); 4269 return true; 4270} 4271 4272//-----------------inline_native_Reflection_getCallerClass--------------------- 4273// public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4274// 4275// In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4276// 4277// NOTE: This code must perform the same logic as JVM_GetCallerClass 4278// in that it must skip particular security frames and checks for 4279// caller sensitive methods. 4280bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4281#ifndef PRODUCT 4282 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4283 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4284 } 4285#endif 4286 4287 if (!jvms()->has_method()) { 4288#ifndef PRODUCT 4289 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4290 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4291 } 4292#endif 4293 return false; 4294 } 4295 4296 // Walk back up the JVM state to find the caller at the required 4297 // depth. 4298 JVMState* caller_jvms = jvms(); 4299 4300 // Cf. JVM_GetCallerClass 4301 // NOTE: Start the loop at depth 1 because the current JVM state does 4302 // not include the Reflection.getCallerClass() frame. 4303 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4304 ciMethod* m = caller_jvms->method(); 4305 switch (n) { 4306 case 0: 4307 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4308 break; 4309 case 1: 4310 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4311 if (!m->caller_sensitive()) { 4312#ifndef PRODUCT 4313 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4314 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4315 } 4316#endif 4317 return false; // bail-out; let JVM_GetCallerClass do the work 4318 } 4319 break; 4320 default: 4321 if (!m->is_ignored_by_security_stack_walk()) { 4322 // We have reached the desired frame; return the holder class. 4323 // Acquire method holder as java.lang.Class and push as constant. 4324 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4325 ciInstance* caller_mirror = caller_klass->java_mirror(); 4326 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4327 4328#ifndef PRODUCT 4329 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4330 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()); 4331 tty->print_cr(" JVM state at this point:"); 4332 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4333 ciMethod* m = jvms()->of_depth(i)->method(); 4334 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4335 } 4336 } 4337#endif 4338 return true; 4339 } 4340 break; 4341 } 4342 } 4343 4344#ifndef PRODUCT 4345 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4346 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4347 tty->print_cr(" JVM state at this point:"); 4348 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4349 ciMethod* m = jvms()->of_depth(i)->method(); 4350 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4351 } 4352 } 4353#endif 4354 4355 return false; // bail-out; let JVM_GetCallerClass do the work 4356} 4357 4358bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4359 Node* arg = argument(0); 4360 Node* result; 4361 4362 switch (id) { 4363 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4364 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4365 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4366 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4367 4368 case vmIntrinsics::_doubleToLongBits: { 4369 // two paths (plus control) merge in a wood 4370 RegionNode *r = new RegionNode(3); 4371 Node *phi = new PhiNode(r, TypeLong::LONG); 4372 4373 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4374 // Build the boolean node 4375 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4376 4377 // Branch either way. 4378 // NaN case is less traveled, which makes all the difference. 4379 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4380 Node *opt_isnan = _gvn.transform(ifisnan); 4381 assert( opt_isnan->is_If(), "Expect an IfNode"); 4382 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4383 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4384 4385 set_control(iftrue); 4386 4387 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4388 Node *slow_result = longcon(nan_bits); // return NaN 4389 phi->init_req(1, _gvn.transform( slow_result )); 4390 r->init_req(1, iftrue); 4391 4392 // Else fall through 4393 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4394 set_control(iffalse); 4395 4396 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4397 r->init_req(2, iffalse); 4398 4399 // Post merge 4400 set_control(_gvn.transform(r)); 4401 record_for_igvn(r); 4402 4403 C->set_has_split_ifs(true); // Has chance for split-if optimization 4404 result = phi; 4405 assert(result->bottom_type()->isa_long(), "must be"); 4406 break; 4407 } 4408 4409 case vmIntrinsics::_floatToIntBits: { 4410 // two paths (plus control) merge in a wood 4411 RegionNode *r = new RegionNode(3); 4412 Node *phi = new PhiNode(r, TypeInt::INT); 4413 4414 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4415 // Build the boolean node 4416 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4417 4418 // Branch either way. 4419 // NaN case is less traveled, which makes all the difference. 4420 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4421 Node *opt_isnan = _gvn.transform(ifisnan); 4422 assert( opt_isnan->is_If(), "Expect an IfNode"); 4423 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4424 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4425 4426 set_control(iftrue); 4427 4428 static const jint nan_bits = 0x7fc00000; 4429 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4430 phi->init_req(1, _gvn.transform( slow_result )); 4431 r->init_req(1, iftrue); 4432 4433 // Else fall through 4434 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4435 set_control(iffalse); 4436 4437 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4438 r->init_req(2, iffalse); 4439 4440 // Post merge 4441 set_control(_gvn.transform(r)); 4442 record_for_igvn(r); 4443 4444 C->set_has_split_ifs(true); // Has chance for split-if optimization 4445 result = phi; 4446 assert(result->bottom_type()->isa_int(), "must be"); 4447 break; 4448 } 4449 4450 default: 4451 fatal_unexpected_iid(id); 4452 break; 4453 } 4454 set_result(_gvn.transform(result)); 4455 return true; 4456} 4457 4458#ifdef _LP64 4459#define XTOP ,top() /*additional argument*/ 4460#else //_LP64 4461#define XTOP /*no additional argument*/ 4462#endif //_LP64 4463 4464//----------------------inline_unsafe_copyMemory------------------------- 4465// public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4466bool LibraryCallKit::inline_unsafe_copyMemory() { 4467 if (callee()->is_static()) return false; // caller must have the capability! 4468 null_check_receiver(); // null-check receiver 4469 if (stopped()) return true; 4470 4471 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4472 4473 Node* src_ptr = argument(1); // type: oop 4474 Node* src_off = ConvL2X(argument(2)); // type: long 4475 Node* dst_ptr = argument(4); // type: oop 4476 Node* dst_off = ConvL2X(argument(5)); // type: long 4477 Node* size = ConvL2X(argument(7)); // type: long 4478 4479 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4480 "fieldOffset must be byte-scaled"); 4481 4482 Node* src = make_unsafe_address(src_ptr, src_off); 4483 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4484 4485 // Conservatively insert a memory barrier on all memory slices. 4486 // Do not let writes of the copy source or destination float below the copy. 4487 insert_mem_bar(Op_MemBarCPUOrder); 4488 4489 // Call it. Note that the length argument is not scaled. 4490 make_runtime_call(RC_LEAF|RC_NO_FP, 4491 OptoRuntime::fast_arraycopy_Type(), 4492 StubRoutines::unsafe_arraycopy(), 4493 "unsafe_arraycopy", 4494 TypeRawPtr::BOTTOM, 4495 src, dst, size XTOP); 4496 4497 // Do not let reads of the copy destination float above the copy. 4498 insert_mem_bar(Op_MemBarCPUOrder); 4499 4500 return true; 4501} 4502 4503//------------------------clone_coping----------------------------------- 4504// Helper function for inline_native_clone. 4505void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4506 assert(obj_size != NULL, ""); 4507 Node* raw_obj = alloc_obj->in(1); 4508 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4509 4510 AllocateNode* alloc = NULL; 4511 if (ReduceBulkZeroing) { 4512 // We will be completely responsible for initializing this object - 4513 // mark Initialize node as complete. 4514 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4515 // The object was just allocated - there should be no any stores! 4516 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4517 // Mark as complete_with_arraycopy so that on AllocateNode 4518 // expansion, we know this AllocateNode is initialized by an array 4519 // copy and a StoreStore barrier exists after the array copy. 4520 alloc->initialization()->set_complete_with_arraycopy(); 4521 } 4522 4523 // Copy the fastest available way. 4524 // TODO: generate fields copies for small objects instead. 4525 Node* src = obj; 4526 Node* dest = alloc_obj; 4527 Node* size = _gvn.transform(obj_size); 4528 4529 // Exclude the header but include array length to copy by 8 bytes words. 4530 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4531 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4532 instanceOopDesc::base_offset_in_bytes(); 4533 // base_off: 4534 // 8 - 32-bit VM 4535 // 12 - 64-bit VM, compressed klass 4536 // 16 - 64-bit VM, normal klass 4537 if (base_off % BytesPerLong != 0) { 4538 assert(UseCompressedClassPointers, ""); 4539 if (is_array) { 4540 // Exclude length to copy by 8 bytes words. 4541 base_off += sizeof(int); 4542 } else { 4543 // Include klass to copy by 8 bytes words. 4544 base_off = instanceOopDesc::klass_offset_in_bytes(); 4545 } 4546 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4547 } 4548 src = basic_plus_adr(src, base_off); 4549 dest = basic_plus_adr(dest, base_off); 4550 4551 // Compute the length also, if needed: 4552 Node* countx = size; 4553 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4554 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4555 4556 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4557 4558 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4559 ac->set_clonebasic(); 4560 Node* n = _gvn.transform(ac); 4561 if (n == ac) { 4562 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4563 } else { 4564 set_all_memory(n); 4565 } 4566 4567 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4568 if (card_mark) { 4569 assert(!is_array, ""); 4570 // Put in store barrier for any and all oops we are sticking 4571 // into this object. (We could avoid this if we could prove 4572 // that the object type contains no oop fields at all.) 4573 Node* no_particular_value = NULL; 4574 Node* no_particular_field = NULL; 4575 int raw_adr_idx = Compile::AliasIdxRaw; 4576 post_barrier(control(), 4577 memory(raw_adr_type), 4578 alloc_obj, 4579 no_particular_field, 4580 raw_adr_idx, 4581 no_particular_value, 4582 T_OBJECT, 4583 false); 4584 } 4585 4586 // Do not let reads from the cloned object float above the arraycopy. 4587 if (alloc != NULL) { 4588 // Do not let stores that initialize this object be reordered with 4589 // a subsequent store that would make this object accessible by 4590 // other threads. 4591 // Record what AllocateNode this StoreStore protects so that 4592 // escape analysis can go from the MemBarStoreStoreNode to the 4593 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4594 // based on the escape status of the AllocateNode. 4595 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4596 } else { 4597 insert_mem_bar(Op_MemBarCPUOrder); 4598 } 4599} 4600 4601//------------------------inline_native_clone---------------------------- 4602// protected native Object java.lang.Object.clone(); 4603// 4604// Here are the simple edge cases: 4605// null receiver => normal trap 4606// virtual and clone was overridden => slow path to out-of-line clone 4607// not cloneable or finalizer => slow path to out-of-line Object.clone 4608// 4609// The general case has two steps, allocation and copying. 4610// Allocation has two cases, and uses GraphKit::new_instance or new_array. 4611// 4612// Copying also has two cases, oop arrays and everything else. 4613// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4614// Everything else uses the tight inline loop supplied by CopyArrayNode. 4615// 4616// These steps fold up nicely if and when the cloned object's klass 4617// can be sharply typed as an object array, a type array, or an instance. 4618// 4619bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4620 PhiNode* result_val; 4621 4622 // Set the reexecute bit for the interpreter to reexecute 4623 // the bytecode that invokes Object.clone if deoptimization happens. 4624 { PreserveReexecuteState preexecs(this); 4625 jvms()->set_should_reexecute(true); 4626 4627 Node* obj = null_check_receiver(); 4628 if (stopped()) return true; 4629 4630 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4631 4632 // If we are going to clone an instance, we need its exact type to 4633 // know the number and types of fields to convert the clone to 4634 // loads/stores. Maybe a speculative type can help us. 4635 if (!obj_type->klass_is_exact() && 4636 obj_type->speculative_type() != NULL && 4637 obj_type->speculative_type()->is_instance_klass()) { 4638 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4639 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4640 !spec_ik->has_injected_fields()) { 4641 ciKlass* k = obj_type->klass(); 4642 if (!k->is_instance_klass() || 4643 k->as_instance_klass()->is_interface() || 4644 k->as_instance_klass()->has_subklass()) { 4645 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4646 } 4647 } 4648 } 4649 4650 Node* obj_klass = load_object_klass(obj); 4651 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4652 const TypeOopPtr* toop = ((tklass != NULL) 4653 ? tklass->as_instance_type() 4654 : TypeInstPtr::NOTNULL); 4655 4656 // Conservatively insert a memory barrier on all memory slices. 4657 // Do not let writes into the original float below the clone. 4658 insert_mem_bar(Op_MemBarCPUOrder); 4659 4660 // paths into result_reg: 4661 enum { 4662 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4663 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4664 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4665 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4666 PATH_LIMIT 4667 }; 4668 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4669 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4670 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4671 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4672 record_for_igvn(result_reg); 4673 4674 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4675 int raw_adr_idx = Compile::AliasIdxRaw; 4676 4677 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4678 if (array_ctl != NULL) { 4679 // It's an array. 4680 PreserveJVMState pjvms(this); 4681 set_control(array_ctl); 4682 Node* obj_length = load_array_length(obj); 4683 Node* obj_size = NULL; 4684 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4685 4686 if (!use_ReduceInitialCardMarks()) { 4687 // If it is an oop array, it requires very special treatment, 4688 // because card marking is required on each card of the array. 4689 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4690 if (is_obja != NULL) { 4691 PreserveJVMState pjvms2(this); 4692 set_control(is_obja); 4693 // Generate a direct call to the right arraycopy function(s). 4694 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4695 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4696 ac->set_cloneoop(); 4697 Node* n = _gvn.transform(ac); 4698 assert(n == ac, "cannot disappear"); 4699 ac->connect_outputs(this); 4700 4701 result_reg->init_req(_objArray_path, control()); 4702 result_val->init_req(_objArray_path, alloc_obj); 4703 result_i_o ->set_req(_objArray_path, i_o()); 4704 result_mem ->set_req(_objArray_path, reset_memory()); 4705 } 4706 } 4707 // Otherwise, there are no card marks to worry about. 4708 // (We can dispense with card marks if we know the allocation 4709 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4710 // causes the non-eden paths to take compensating steps to 4711 // simulate a fresh allocation, so that no further 4712 // card marks are required in compiled code to initialize 4713 // the object.) 4714 4715 if (!stopped()) { 4716 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4717 4718 // Present the results of the copy. 4719 result_reg->init_req(_array_path, control()); 4720 result_val->init_req(_array_path, alloc_obj); 4721 result_i_o ->set_req(_array_path, i_o()); 4722 result_mem ->set_req(_array_path, reset_memory()); 4723 } 4724 } 4725 4726 // We only go to the instance fast case code if we pass a number of guards. 4727 // The paths which do not pass are accumulated in the slow_region. 4728 RegionNode* slow_region = new RegionNode(1); 4729 record_for_igvn(slow_region); 4730 if (!stopped()) { 4731 // It's an instance (we did array above). Make the slow-path tests. 4732 // If this is a virtual call, we generate a funny guard. We grab 4733 // the vtable entry corresponding to clone() from the target object. 4734 // If the target method which we are calling happens to be the 4735 // Object clone() method, we pass the guard. We do not need this 4736 // guard for non-virtual calls; the caller is known to be the native 4737 // Object clone(). 4738 if (is_virtual) { 4739 generate_virtual_guard(obj_klass, slow_region); 4740 } 4741 4742 // The object must be cloneable and must not have a finalizer. 4743 // Both of these conditions may be checked in a single test. 4744 // We could optimize the cloneable test further, but we don't care. 4745 generate_access_flags_guard(obj_klass, 4746 // Test both conditions: 4747 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4748 // Must be cloneable but not finalizer: 4749 JVM_ACC_IS_CLONEABLE, 4750 slow_region); 4751 } 4752 4753 if (!stopped()) { 4754 // It's an instance, and it passed the slow-path tests. 4755 PreserveJVMState pjvms(this); 4756 Node* obj_size = NULL; 4757 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4758 // is reexecuted if deoptimization occurs and there could be problems when merging 4759 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4760 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4761 4762 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4763 4764 // Present the results of the slow call. 4765 result_reg->init_req(_instance_path, control()); 4766 result_val->init_req(_instance_path, alloc_obj); 4767 result_i_o ->set_req(_instance_path, i_o()); 4768 result_mem ->set_req(_instance_path, reset_memory()); 4769 } 4770 4771 // Generate code for the slow case. We make a call to clone(). 4772 set_control(_gvn.transform(slow_region)); 4773 if (!stopped()) { 4774 PreserveJVMState pjvms(this); 4775 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4776 Node* slow_result = set_results_for_java_call(slow_call); 4777 // this->control() comes from set_results_for_java_call 4778 result_reg->init_req(_slow_path, control()); 4779 result_val->init_req(_slow_path, slow_result); 4780 result_i_o ->set_req(_slow_path, i_o()); 4781 result_mem ->set_req(_slow_path, reset_memory()); 4782 } 4783 4784 // Return the combined state. 4785 set_control( _gvn.transform(result_reg)); 4786 set_i_o( _gvn.transform(result_i_o)); 4787 set_all_memory( _gvn.transform(result_mem)); 4788 } // original reexecute is set back here 4789 4790 set_result(_gvn.transform(result_val)); 4791 return true; 4792} 4793 4794// If we have a tighly coupled allocation, the arraycopy may take care 4795// of the array initialization. If one of the guards we insert between 4796// the allocation and the arraycopy causes a deoptimization, an 4797// unitialized array will escape the compiled method. To prevent that 4798// we set the JVM state for uncommon traps between the allocation and 4799// the arraycopy to the state before the allocation so, in case of 4800// deoptimization, we'll reexecute the allocation and the 4801// initialization. 4802JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4803 if (alloc != NULL) { 4804 ciMethod* trap_method = alloc->jvms()->method(); 4805 int trap_bci = alloc->jvms()->bci(); 4806 4807 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4808 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4809 // Make sure there's no store between the allocation and the 4810 // arraycopy otherwise visible side effects could be rexecuted 4811 // in case of deoptimization and cause incorrect execution. 4812 bool no_interfering_store = true; 4813 Node* mem = alloc->in(TypeFunc::Memory); 4814 if (mem->is_MergeMem()) { 4815 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4816 Node* n = mms.memory(); 4817 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4818 assert(n->is_Store(), "what else?"); 4819 no_interfering_store = false; 4820 break; 4821 } 4822 } 4823 } else { 4824 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4825 Node* n = mms.memory(); 4826 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4827 assert(n->is_Store(), "what else?"); 4828 no_interfering_store = false; 4829 break; 4830 } 4831 } 4832 } 4833 4834 if (no_interfering_store) { 4835 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4836 uint size = alloc->req(); 4837 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4838 old_jvms->set_map(sfpt); 4839 for (uint i = 0; i < size; i++) { 4840 sfpt->init_req(i, alloc->in(i)); 4841 } 4842 // re-push array length for deoptimization 4843 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4844 old_jvms->set_sp(old_jvms->sp()+1); 4845 old_jvms->set_monoff(old_jvms->monoff()+1); 4846 old_jvms->set_scloff(old_jvms->scloff()+1); 4847 old_jvms->set_endoff(old_jvms->endoff()+1); 4848 old_jvms->set_should_reexecute(true); 4849 4850 sfpt->set_i_o(map()->i_o()); 4851 sfpt->set_memory(map()->memory()); 4852 sfpt->set_control(map()->control()); 4853 4854 JVMState* saved_jvms = jvms(); 4855 saved_reexecute_sp = _reexecute_sp; 4856 4857 set_jvms(sfpt->jvms()); 4858 _reexecute_sp = jvms()->sp(); 4859 4860 return saved_jvms; 4861 } 4862 } 4863 } 4864 return NULL; 4865} 4866 4867// In case of a deoptimization, we restart execution at the 4868// allocation, allocating a new array. We would leave an uninitialized 4869// array in the heap that GCs wouldn't expect. Move the allocation 4870// after the traps so we don't allocate the array if we 4871// deoptimize. This is possible because tightly_coupled_allocation() 4872// guarantees there's no observer of the allocated array at this point 4873// and the control flow is simple enough. 4874void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4875 if (saved_jvms != NULL && !stopped()) { 4876 assert(alloc != NULL, "only with a tightly coupled allocation"); 4877 // restore JVM state to the state at the arraycopy 4878 saved_jvms->map()->set_control(map()->control()); 4879 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4880 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4881 // If we've improved the types of some nodes (null check) while 4882 // emitting the guards, propagate them to the current state 4883 map()->replaced_nodes().apply(saved_jvms->map()); 4884 set_jvms(saved_jvms); 4885 _reexecute_sp = saved_reexecute_sp; 4886 4887 // Remove the allocation from above the guards 4888 CallProjections callprojs; 4889 alloc->extract_projections(&callprojs, true); 4890 InitializeNode* init = alloc->initialization(); 4891 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4892 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4893 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4894 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4895 4896 // move the allocation here (after the guards) 4897 _gvn.hash_delete(alloc); 4898 alloc->set_req(TypeFunc::Control, control()); 4899 alloc->set_req(TypeFunc::I_O, i_o()); 4900 Node *mem = reset_memory(); 4901 set_all_memory(mem); 4902 alloc->set_req(TypeFunc::Memory, mem); 4903 set_control(init->proj_out(TypeFunc::Control)); 4904 set_i_o(callprojs.fallthrough_ioproj); 4905 4906 // Update memory as done in GraphKit::set_output_for_allocation() 4907 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4908 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4909 if (ary_type->isa_aryptr() && length_type != NULL) { 4910 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4911 } 4912 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4913 int elemidx = C->get_alias_index(telemref); 4914 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4915 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4916 4917 Node* allocx = _gvn.transform(alloc); 4918 assert(allocx == alloc, "where has the allocation gone?"); 4919 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4920 4921 _gvn.hash_delete(dest); 4922 dest->set_req(0, control()); 4923 Node* destx = _gvn.transform(dest); 4924 assert(destx == dest, "where has the allocation result gone?"); 4925 } 4926} 4927 4928 4929//------------------------------inline_arraycopy----------------------- 4930// public static native void java.lang.System.arraycopy(Object src, int srcPos, 4931// Object dest, int destPos, 4932// int length); 4933bool LibraryCallKit::inline_arraycopy() { 4934 // Get the arguments. 4935 Node* src = argument(0); // type: oop 4936 Node* src_offset = argument(1); // type: int 4937 Node* dest = argument(2); // type: oop 4938 Node* dest_offset = argument(3); // type: int 4939 Node* length = argument(4); // type: int 4940 4941 4942 // Check for allocation before we add nodes that would confuse 4943 // tightly_coupled_allocation() 4944 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4945 4946 int saved_reexecute_sp = -1; 4947 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4948 // See arraycopy_restore_alloc_state() comment 4949 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4950 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4951 // if saved_jvms == NULL and alloc != NULL, we can���t emit any guards 4952 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4953 4954 // The following tests must be performed 4955 // (1) src and dest are arrays. 4956 // (2) src and dest arrays must have elements of the same BasicType 4957 // (3) src and dest must not be null. 4958 // (4) src_offset must not be negative. 4959 // (5) dest_offset must not be negative. 4960 // (6) length must not be negative. 4961 // (7) src_offset + length must not exceed length of src. 4962 // (8) dest_offset + length must not exceed length of dest. 4963 // (9) each element of an oop array must be assignable 4964 4965 // (3) src and dest must not be null. 4966 // always do this here because we need the JVM state for uncommon traps 4967 Node* null_ctl = top(); 4968 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4969 assert(null_ctl->is_top(), "no null control here"); 4970 dest = null_check(dest, T_ARRAY); 4971 4972 if (!can_emit_guards) { 4973 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4974 // guards but the arraycopy node could still take advantage of a 4975 // tightly allocated allocation. tightly_coupled_allocation() is 4976 // called again to make sure it takes the null check above into 4977 // account: the null check is mandatory and if it caused an 4978 // uncommon trap to be emitted then the allocation can't be 4979 // considered tightly coupled in this context. 4980 alloc = tightly_coupled_allocation(dest, NULL); 4981 } 4982 4983 bool validated = false; 4984 4985 const Type* src_type = _gvn.type(src); 4986 const Type* dest_type = _gvn.type(dest); 4987 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4988 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4989 4990 // Do we have the type of src? 4991 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4992 // Do we have the type of dest? 4993 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4994 // Is the type for src from speculation? 4995 bool src_spec = false; 4996 // Is the type for dest from speculation? 4997 bool dest_spec = false; 4998 4999 if ((!has_src || !has_dest) && can_emit_guards) { 5000 // We don't have sufficient type information, let's see if 5001 // speculative types can help. We need to have types for both src 5002 // and dest so that it pays off. 5003 5004 // Do we already have or could we have type information for src 5005 bool could_have_src = has_src; 5006 // Do we already have or could we have type information for dest 5007 bool could_have_dest = has_dest; 5008 5009 ciKlass* src_k = NULL; 5010 if (!has_src) { 5011 src_k = src_type->speculative_type_not_null(); 5012 if (src_k != NULL && src_k->is_array_klass()) { 5013 could_have_src = true; 5014 } 5015 } 5016 5017 ciKlass* dest_k = NULL; 5018 if (!has_dest) { 5019 dest_k = dest_type->speculative_type_not_null(); 5020 if (dest_k != NULL && dest_k->is_array_klass()) { 5021 could_have_dest = true; 5022 } 5023 } 5024 5025 if (could_have_src && could_have_dest) { 5026 // This is going to pay off so emit the required guards 5027 if (!has_src) { 5028 src = maybe_cast_profiled_obj(src, src_k, true); 5029 src_type = _gvn.type(src); 5030 top_src = src_type->isa_aryptr(); 5031 has_src = (top_src != NULL && top_src->klass() != NULL); 5032 src_spec = true; 5033 } 5034 if (!has_dest) { 5035 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5036 dest_type = _gvn.type(dest); 5037 top_dest = dest_type->isa_aryptr(); 5038 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 5039 dest_spec = true; 5040 } 5041 } 5042 } 5043 5044 if (has_src && has_dest && can_emit_guards) { 5045 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 5046 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 5047 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 5048 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 5049 5050 if (src_elem == dest_elem && src_elem == T_OBJECT) { 5051 // If both arrays are object arrays then having the exact types 5052 // for both will remove the need for a subtype check at runtime 5053 // before the call and may make it possible to pick a faster copy 5054 // routine (without a subtype check on every element) 5055 // Do we have the exact type of src? 5056 bool could_have_src = src_spec; 5057 // Do we have the exact type of dest? 5058 bool could_have_dest = dest_spec; 5059 ciKlass* src_k = top_src->klass(); 5060 ciKlass* dest_k = top_dest->klass(); 5061 if (!src_spec) { 5062 src_k = src_type->speculative_type_not_null(); 5063 if (src_k != NULL && src_k->is_array_klass()) { 5064 could_have_src = true; 5065 } 5066 } 5067 if (!dest_spec) { 5068 dest_k = dest_type->speculative_type_not_null(); 5069 if (dest_k != NULL && dest_k->is_array_klass()) { 5070 could_have_dest = true; 5071 } 5072 } 5073 if (could_have_src && could_have_dest) { 5074 // If we can have both exact types, emit the missing guards 5075 if (could_have_src && !src_spec) { 5076 src = maybe_cast_profiled_obj(src, src_k, true); 5077 } 5078 if (could_have_dest && !dest_spec) { 5079 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5080 } 5081 } 5082 } 5083 } 5084 5085 ciMethod* trap_method = method(); 5086 int trap_bci = bci(); 5087 if (saved_jvms != NULL) { 5088 trap_method = alloc->jvms()->method(); 5089 trap_bci = alloc->jvms()->bci(); 5090 } 5091 5092 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5093 can_emit_guards && 5094 !src->is_top() && !dest->is_top()) { 5095 // validate arguments: enables transformation the ArrayCopyNode 5096 validated = true; 5097 5098 RegionNode* slow_region = new RegionNode(1); 5099 record_for_igvn(slow_region); 5100 5101 // (1) src and dest are arrays. 5102 generate_non_array_guard(load_object_klass(src), slow_region); 5103 generate_non_array_guard(load_object_klass(dest), slow_region); 5104 5105 // (2) src and dest arrays must have elements of the same BasicType 5106 // done at macro expansion or at Ideal transformation time 5107 5108 // (4) src_offset must not be negative. 5109 generate_negative_guard(src_offset, slow_region); 5110 5111 // (5) dest_offset must not be negative. 5112 generate_negative_guard(dest_offset, slow_region); 5113 5114 // (7) src_offset + length must not exceed length of src. 5115 generate_limit_guard(src_offset, length, 5116 load_array_length(src), 5117 slow_region); 5118 5119 // (8) dest_offset + length must not exceed length of dest. 5120 generate_limit_guard(dest_offset, length, 5121 load_array_length(dest), 5122 slow_region); 5123 5124 // (9) each element of an oop array must be assignable 5125 Node* src_klass = load_object_klass(src); 5126 Node* dest_klass = load_object_klass(dest); 5127 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5128 5129 if (not_subtype_ctrl != top()) { 5130 PreserveJVMState pjvms(this); 5131 set_control(not_subtype_ctrl); 5132 uncommon_trap(Deoptimization::Reason_intrinsic, 5133 Deoptimization::Action_make_not_entrant); 5134 assert(stopped(), "Should be stopped"); 5135 } 5136 { 5137 PreserveJVMState pjvms(this); 5138 set_control(_gvn.transform(slow_region)); 5139 uncommon_trap(Deoptimization::Reason_intrinsic, 5140 Deoptimization::Action_make_not_entrant); 5141 assert(stopped(), "Should be stopped"); 5142 } 5143 } 5144 5145 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5146 5147 if (stopped()) { 5148 return true; 5149 } 5150 5151 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5152 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5153 // so the compiler has a chance to eliminate them: during macro expansion, 5154 // we have to set their control (CastPP nodes are eliminated). 5155 load_object_klass(src), load_object_klass(dest), 5156 load_array_length(src), load_array_length(dest)); 5157 5158 ac->set_arraycopy(validated); 5159 5160 Node* n = _gvn.transform(ac); 5161 if (n == ac) { 5162 ac->connect_outputs(this); 5163 } else { 5164 assert(validated, "shouldn't transform if all arguments not validated"); 5165 set_all_memory(n); 5166 } 5167 5168 return true; 5169} 5170 5171 5172// Helper function which determines if an arraycopy immediately follows 5173// an allocation, with no intervening tests or other escapes for the object. 5174AllocateArrayNode* 5175LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5176 RegionNode* slow_region) { 5177 if (stopped()) return NULL; // no fast path 5178 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5179 5180 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5181 if (alloc == NULL) return NULL; 5182 5183 Node* rawmem = memory(Compile::AliasIdxRaw); 5184 // Is the allocation's memory state untouched? 5185 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5186 // Bail out if there have been raw-memory effects since the allocation. 5187 // (Example: There might have been a call or safepoint.) 5188 return NULL; 5189 } 5190 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5191 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5192 return NULL; 5193 } 5194 5195 // There must be no unexpected observers of this allocation. 5196 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5197 Node* obs = ptr->fast_out(i); 5198 if (obs != this->map()) { 5199 return NULL; 5200 } 5201 } 5202 5203 // This arraycopy must unconditionally follow the allocation of the ptr. 5204 Node* alloc_ctl = ptr->in(0); 5205 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5206 5207 Node* ctl = control(); 5208 while (ctl != alloc_ctl) { 5209 // There may be guards which feed into the slow_region. 5210 // Any other control flow means that we might not get a chance 5211 // to finish initializing the allocated object. 5212 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5213 IfNode* iff = ctl->in(0)->as_If(); 5214 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5215 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5216 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5217 ctl = iff->in(0); // This test feeds the known slow_region. 5218 continue; 5219 } 5220 // One more try: Various low-level checks bottom out in 5221 // uncommon traps. If the debug-info of the trap omits 5222 // any reference to the allocation, as we've already 5223 // observed, then there can be no objection to the trap. 5224 bool found_trap = false; 5225 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5226 Node* obs = not_ctl->fast_out(j); 5227 if (obs->in(0) == not_ctl && obs->is_Call() && 5228 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5229 found_trap = true; break; 5230 } 5231 } 5232 if (found_trap) { 5233 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5234 continue; 5235 } 5236 } 5237 return NULL; 5238 } 5239 5240 // If we get this far, we have an allocation which immediately 5241 // precedes the arraycopy, and we can take over zeroing the new object. 5242 // The arraycopy will finish the initialization, and provide 5243 // a new control state to which we will anchor the destination pointer. 5244 5245 return alloc; 5246} 5247 5248//-------------inline_encodeISOArray----------------------------------- 5249// encode char[] to byte[] in ISO_8859_1 5250bool LibraryCallKit::inline_encodeISOArray() { 5251 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5252 // no receiver since it is static method 5253 Node *src = argument(0); 5254 Node *src_offset = argument(1); 5255 Node *dst = argument(2); 5256 Node *dst_offset = argument(3); 5257 Node *length = argument(4); 5258 5259 const Type* src_type = src->Value(&_gvn); 5260 const Type* dst_type = dst->Value(&_gvn); 5261 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5262 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5263 if (top_src == NULL || top_src->klass() == NULL || 5264 top_dest == NULL || top_dest->klass() == NULL) { 5265 // failed array check 5266 return false; 5267 } 5268 5269 // Figure out the size and type of the elements we will be copying. 5270 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5271 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5272 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5273 return false; 5274 } 5275 Node* src_start = array_element_address(src, src_offset, src_elem); 5276 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5277 // 'src_start' points to src array + scaled offset 5278 // 'dst_start' points to dst array + scaled offset 5279 5280 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5281 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5282 enc = _gvn.transform(enc); 5283 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5284 set_memory(res_mem, mtype); 5285 set_result(enc); 5286 return true; 5287} 5288 5289//-------------inline_multiplyToLen----------------------------------- 5290bool LibraryCallKit::inline_multiplyToLen() { 5291 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5292 5293 address stubAddr = StubRoutines::multiplyToLen(); 5294 if (stubAddr == NULL) { 5295 return false; // Intrinsic's stub is not implemented on this platform 5296 } 5297 const char* stubName = "multiplyToLen"; 5298 5299 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5300 5301 // no receiver because it is a static method 5302 Node* x = argument(0); 5303 Node* xlen = argument(1); 5304 Node* y = argument(2); 5305 Node* ylen = argument(3); 5306 Node* z = argument(4); 5307 5308 const Type* x_type = x->Value(&_gvn); 5309 const Type* y_type = y->Value(&_gvn); 5310 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5311 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5312 if (top_x == NULL || top_x->klass() == NULL || 5313 top_y == NULL || top_y->klass() == NULL) { 5314 // failed array check 5315 return false; 5316 } 5317 5318 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5319 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5320 if (x_elem != T_INT || y_elem != T_INT) { 5321 return false; 5322 } 5323 5324 // Set the original stack and the reexecute bit for the interpreter to reexecute 5325 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5326 // on the return from z array allocation in runtime. 5327 { PreserveReexecuteState preexecs(this); 5328 jvms()->set_should_reexecute(true); 5329 5330 Node* x_start = array_element_address(x, intcon(0), x_elem); 5331 Node* y_start = array_element_address(y, intcon(0), y_elem); 5332 // 'x_start' points to x array + scaled xlen 5333 // 'y_start' points to y array + scaled ylen 5334 5335 // Allocate the result array 5336 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5337 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5338 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5339 5340 IdealKit ideal(this); 5341 5342#define __ ideal. 5343 Node* one = __ ConI(1); 5344 Node* zero = __ ConI(0); 5345 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5346 __ set(need_alloc, zero); 5347 __ set(z_alloc, z); 5348 __ if_then(z, BoolTest::eq, null()); { 5349 __ increment (need_alloc, one); 5350 } __ else_(); { 5351 // Update graphKit memory and control from IdealKit. 5352 sync_kit(ideal); 5353 Node* zlen_arg = load_array_length(z); 5354 // Update IdealKit memory and control from graphKit. 5355 __ sync_kit(this); 5356 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5357 __ increment (need_alloc, one); 5358 } __ end_if(); 5359 } __ end_if(); 5360 5361 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5362 // Update graphKit memory and control from IdealKit. 5363 sync_kit(ideal); 5364 Node * narr = new_array(klass_node, zlen, 1); 5365 // Update IdealKit memory and control from graphKit. 5366 __ sync_kit(this); 5367 __ set(z_alloc, narr); 5368 } __ end_if(); 5369 5370 sync_kit(ideal); 5371 z = __ value(z_alloc); 5372 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5373 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5374 // Final sync IdealKit and GraphKit. 5375 final_sync(ideal); 5376#undef __ 5377 5378 Node* z_start = array_element_address(z, intcon(0), T_INT); 5379 5380 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5381 OptoRuntime::multiplyToLen_Type(), 5382 stubAddr, stubName, TypePtr::BOTTOM, 5383 x_start, xlen, y_start, ylen, z_start, zlen); 5384 } // original reexecute is set back here 5385 5386 C->set_has_split_ifs(true); // Has chance for split-if optimization 5387 set_result(z); 5388 return true; 5389} 5390 5391//-------------inline_squareToLen------------------------------------ 5392bool LibraryCallKit::inline_squareToLen() { 5393 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5394 5395 address stubAddr = StubRoutines::squareToLen(); 5396 if (stubAddr == NULL) { 5397 return false; // Intrinsic's stub is not implemented on this platform 5398 } 5399 const char* stubName = "squareToLen"; 5400 5401 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5402 5403 Node* x = argument(0); 5404 Node* len = argument(1); 5405 Node* z = argument(2); 5406 Node* zlen = argument(3); 5407 5408 const Type* x_type = x->Value(&_gvn); 5409 const Type* z_type = z->Value(&_gvn); 5410 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5411 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5412 if (top_x == NULL || top_x->klass() == NULL || 5413 top_z == NULL || top_z->klass() == NULL) { 5414 // failed array check 5415 return false; 5416 } 5417 5418 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5419 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5420 if (x_elem != T_INT || z_elem != T_INT) { 5421 return false; 5422 } 5423 5424 5425 Node* x_start = array_element_address(x, intcon(0), x_elem); 5426 Node* z_start = array_element_address(z, intcon(0), z_elem); 5427 5428 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5429 OptoRuntime::squareToLen_Type(), 5430 stubAddr, stubName, TypePtr::BOTTOM, 5431 x_start, len, z_start, zlen); 5432 5433 set_result(z); 5434 return true; 5435} 5436 5437//-------------inline_mulAdd------------------------------------------ 5438bool LibraryCallKit::inline_mulAdd() { 5439 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5440 5441 address stubAddr = StubRoutines::mulAdd(); 5442 if (stubAddr == NULL) { 5443 return false; // Intrinsic's stub is not implemented on this platform 5444 } 5445 const char* stubName = "mulAdd"; 5446 5447 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5448 5449 Node* out = argument(0); 5450 Node* in = argument(1); 5451 Node* offset = argument(2); 5452 Node* len = argument(3); 5453 Node* k = argument(4); 5454 5455 const Type* out_type = out->Value(&_gvn); 5456 const Type* in_type = in->Value(&_gvn); 5457 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5458 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5459 if (top_out == NULL || top_out->klass() == NULL || 5460 top_in == NULL || top_in->klass() == NULL) { 5461 // failed array check 5462 return false; 5463 } 5464 5465 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5466 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5467 if (out_elem != T_INT || in_elem != T_INT) { 5468 return false; 5469 } 5470 5471 Node* outlen = load_array_length(out); 5472 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5473 Node* out_start = array_element_address(out, intcon(0), out_elem); 5474 Node* in_start = array_element_address(in, intcon(0), in_elem); 5475 5476 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5477 OptoRuntime::mulAdd_Type(), 5478 stubAddr, stubName, TypePtr::BOTTOM, 5479 out_start,in_start, new_offset, len, k); 5480 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5481 set_result(result); 5482 return true; 5483} 5484 5485//-------------inline_montgomeryMultiply----------------------------------- 5486bool LibraryCallKit::inline_montgomeryMultiply() { 5487 address stubAddr = StubRoutines::montgomeryMultiply(); 5488 if (stubAddr == NULL) { 5489 return false; // Intrinsic's stub is not implemented on this platform 5490 } 5491 5492 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5493 const char* stubName = "montgomery_square"; 5494 5495 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5496 5497 Node* a = argument(0); 5498 Node* b = argument(1); 5499 Node* n = argument(2); 5500 Node* len = argument(3); 5501 Node* inv = argument(4); 5502 Node* m = argument(6); 5503 5504 const Type* a_type = a->Value(&_gvn); 5505 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5506 const Type* b_type = b->Value(&_gvn); 5507 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5508 const Type* n_type = a->Value(&_gvn); 5509 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5510 const Type* m_type = a->Value(&_gvn); 5511 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5512 if (top_a == NULL || top_a->klass() == NULL || 5513 top_b == NULL || top_b->klass() == NULL || 5514 top_n == NULL || top_n->klass() == NULL || 5515 top_m == NULL || top_m->klass() == NULL) { 5516 // failed array check 5517 return false; 5518 } 5519 5520 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5521 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5522 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5523 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5524 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5525 return false; 5526 } 5527 5528 // Make the call 5529 { 5530 Node* a_start = array_element_address(a, intcon(0), a_elem); 5531 Node* b_start = array_element_address(b, intcon(0), b_elem); 5532 Node* n_start = array_element_address(n, intcon(0), n_elem); 5533 Node* m_start = array_element_address(m, intcon(0), m_elem); 5534 5535 Node* call = make_runtime_call(RC_LEAF, 5536 OptoRuntime::montgomeryMultiply_Type(), 5537 stubAddr, stubName, TypePtr::BOTTOM, 5538 a_start, b_start, n_start, len, inv, top(), 5539 m_start); 5540 set_result(m); 5541 } 5542 5543 return true; 5544} 5545 5546bool LibraryCallKit::inline_montgomerySquare() { 5547 address stubAddr = StubRoutines::montgomerySquare(); 5548 if (stubAddr == NULL) { 5549 return false; // Intrinsic's stub is not implemented on this platform 5550 } 5551 5552 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5553 const char* stubName = "montgomery_square"; 5554 5555 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5556 5557 Node* a = argument(0); 5558 Node* n = argument(1); 5559 Node* len = argument(2); 5560 Node* inv = argument(3); 5561 Node* m = argument(5); 5562 5563 const Type* a_type = a->Value(&_gvn); 5564 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5565 const Type* n_type = a->Value(&_gvn); 5566 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5567 const Type* m_type = a->Value(&_gvn); 5568 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5569 if (top_a == NULL || top_a->klass() == NULL || 5570 top_n == NULL || top_n->klass() == NULL || 5571 top_m == NULL || top_m->klass() == NULL) { 5572 // failed array check 5573 return false; 5574 } 5575 5576 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5577 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5578 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5579 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5580 return false; 5581 } 5582 5583 // Make the call 5584 { 5585 Node* a_start = array_element_address(a, intcon(0), a_elem); 5586 Node* n_start = array_element_address(n, intcon(0), n_elem); 5587 Node* m_start = array_element_address(m, intcon(0), m_elem); 5588 5589 Node* call = make_runtime_call(RC_LEAF, 5590 OptoRuntime::montgomerySquare_Type(), 5591 stubAddr, stubName, TypePtr::BOTTOM, 5592 a_start, n_start, len, inv, top(), 5593 m_start); 5594 set_result(m); 5595 } 5596 5597 return true; 5598} 5599 5600 5601/** 5602 * Calculate CRC32 for byte. 5603 * int java.util.zip.CRC32.update(int crc, int b) 5604 */ 5605bool LibraryCallKit::inline_updateCRC32() { 5606 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5607 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5608 // no receiver since it is static method 5609 Node* crc = argument(0); // type: int 5610 Node* b = argument(1); // type: int 5611 5612 /* 5613 * int c = ~ crc; 5614 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5615 * b = b ^ (c >>> 8); 5616 * crc = ~b; 5617 */ 5618 5619 Node* M1 = intcon(-1); 5620 crc = _gvn.transform(new XorINode(crc, M1)); 5621 Node* result = _gvn.transform(new XorINode(crc, b)); 5622 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5623 5624 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5625 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5626 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5627 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5628 5629 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5630 result = _gvn.transform(new XorINode(crc, result)); 5631 result = _gvn.transform(new XorINode(result, M1)); 5632 set_result(result); 5633 return true; 5634} 5635 5636/** 5637 * Calculate CRC32 for byte[] array. 5638 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5639 */ 5640bool LibraryCallKit::inline_updateBytesCRC32() { 5641 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5642 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5643 // no receiver since it is static method 5644 Node* crc = argument(0); // type: int 5645 Node* src = argument(1); // type: oop 5646 Node* offset = argument(2); // type: int 5647 Node* length = argument(3); // type: int 5648 5649 const Type* src_type = src->Value(&_gvn); 5650 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5651 if (top_src == NULL || top_src->klass() == NULL) { 5652 // failed array check 5653 return false; 5654 } 5655 5656 // Figure out the size and type of the elements we will be copying. 5657 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5658 if (src_elem != T_BYTE) { 5659 return false; 5660 } 5661 5662 // 'src_start' points to src array + scaled offset 5663 Node* src_start = array_element_address(src, offset, src_elem); 5664 5665 // We assume that range check is done by caller. 5666 // TODO: generate range check (offset+length < src.length) in debug VM. 5667 5668 // Call the stub. 5669 address stubAddr = StubRoutines::updateBytesCRC32(); 5670 const char *stubName = "updateBytesCRC32"; 5671 5672 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5673 stubAddr, stubName, TypePtr::BOTTOM, 5674 crc, src_start, length); 5675 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5676 set_result(result); 5677 return true; 5678} 5679 5680/** 5681 * Calculate CRC32 for ByteBuffer. 5682 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5683 */ 5684bool LibraryCallKit::inline_updateByteBufferCRC32() { 5685 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5686 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5687 // no receiver since it is static method 5688 Node* crc = argument(0); // type: int 5689 Node* src = argument(1); // type: long 5690 Node* offset = argument(3); // type: int 5691 Node* length = argument(4); // type: int 5692 5693 src = ConvL2X(src); // adjust Java long to machine word 5694 Node* base = _gvn.transform(new CastX2PNode(src)); 5695 offset = ConvI2X(offset); 5696 5697 // 'src_start' points to src array + scaled offset 5698 Node* src_start = basic_plus_adr(top(), base, offset); 5699 5700 // Call the stub. 5701 address stubAddr = StubRoutines::updateBytesCRC32(); 5702 const char *stubName = "updateBytesCRC32"; 5703 5704 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5705 stubAddr, stubName, TypePtr::BOTTOM, 5706 crc, src_start, length); 5707 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5708 set_result(result); 5709 return true; 5710} 5711 5712//------------------------------get_table_from_crc32c_class----------------------- 5713Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5714 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5715 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5716 5717 return table; 5718} 5719 5720//------------------------------inline_updateBytesCRC32C----------------------- 5721// 5722// Calculate CRC32C for byte[] array. 5723// int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5724// 5725bool LibraryCallKit::inline_updateBytesCRC32C() { 5726 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5727 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5728 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5729 // no receiver since it is a static method 5730 Node* crc = argument(0); // type: int 5731 Node* src = argument(1); // type: oop 5732 Node* offset = argument(2); // type: int 5733 Node* end = argument(3); // type: int 5734 5735 Node* length = _gvn.transform(new SubINode(end, offset)); 5736 5737 const Type* src_type = src->Value(&_gvn); 5738 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5739 if (top_src == NULL || top_src->klass() == NULL) { 5740 // failed array check 5741 return false; 5742 } 5743 5744 // Figure out the size and type of the elements we will be copying. 5745 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5746 if (src_elem != T_BYTE) { 5747 return false; 5748 } 5749 5750 // 'src_start' points to src array + scaled offset 5751 Node* src_start = array_element_address(src, offset, src_elem); 5752 5753 // static final int[] byteTable in class CRC32C 5754 Node* table = get_table_from_crc32c_class(callee()->holder()); 5755 Node* table_start = array_element_address(table, intcon(0), T_INT); 5756 5757 // We assume that range check is done by caller. 5758 // TODO: generate range check (offset+length < src.length) in debug VM. 5759 5760 // Call the stub. 5761 address stubAddr = StubRoutines::updateBytesCRC32C(); 5762 const char *stubName = "updateBytesCRC32C"; 5763 5764 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5765 stubAddr, stubName, TypePtr::BOTTOM, 5766 crc, src_start, length, table_start); 5767 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5768 set_result(result); 5769 return true; 5770} 5771 5772//------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5773// 5774// Calculate CRC32C for DirectByteBuffer. 5775// int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5776// 5777bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5778 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5779 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5780 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5781 // no receiver since it is a static method 5782 Node* crc = argument(0); // type: int 5783 Node* src = argument(1); // type: long 5784 Node* offset = argument(3); // type: int 5785 Node* end = argument(4); // type: int 5786 5787 Node* length = _gvn.transform(new SubINode(end, offset)); 5788 5789 src = ConvL2X(src); // adjust Java long to machine word 5790 Node* base = _gvn.transform(new CastX2PNode(src)); 5791 offset = ConvI2X(offset); 5792 5793 // 'src_start' points to src array + scaled offset 5794 Node* src_start = basic_plus_adr(top(), base, offset); 5795 5796 // static final int[] byteTable in class CRC32C 5797 Node* table = get_table_from_crc32c_class(callee()->holder()); 5798 Node* table_start = array_element_address(table, intcon(0), T_INT); 5799 5800 // Call the stub. 5801 address stubAddr = StubRoutines::updateBytesCRC32C(); 5802 const char *stubName = "updateBytesCRC32C"; 5803 5804 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5805 stubAddr, stubName, TypePtr::BOTTOM, 5806 crc, src_start, length, table_start); 5807 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5808 set_result(result); 5809 return true; 5810} 5811 5812//----------------------------inline_reference_get---------------------------- 5813// public T java.lang.ref.Reference.get(); 5814bool LibraryCallKit::inline_reference_get() { 5815 const int referent_offset = java_lang_ref_Reference::referent_offset; 5816 guarantee(referent_offset > 0, "should have already been set"); 5817 5818 // Get the argument: 5819 Node* reference_obj = null_check_receiver(); 5820 if (stopped()) return true; 5821 5822 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5823 5824 ciInstanceKlass* klass = env()->Object_klass(); 5825 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5826 5827 Node* no_ctrl = NULL; 5828 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5829 5830 // Use the pre-barrier to record the value in the referent field 5831 pre_barrier(false /* do_load */, 5832 control(), 5833 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5834 result /* pre_val */, 5835 T_OBJECT); 5836 5837 // Add memory barrier to prevent commoning reads from this field 5838 // across safepoint since GC can change its value. 5839 insert_mem_bar(Op_MemBarCPUOrder); 5840 5841 set_result(result); 5842 return true; 5843} 5844 5845 5846Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5847 bool is_exact=true, bool is_static=false, 5848 ciInstanceKlass * fromKls=NULL) { 5849 if (fromKls == NULL) { 5850 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5851 assert(tinst != NULL, "obj is null"); 5852 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5853 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5854 fromKls = tinst->klass()->as_instance_klass(); 5855 } else { 5856 assert(is_static, "only for static field access"); 5857 } 5858 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5859 ciSymbol::make(fieldTypeString), 5860 is_static); 5861 5862 assert (field != NULL, "undefined field"); 5863 if (field == NULL) return (Node *) NULL; 5864 5865 if (is_static) { 5866 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5867 fromObj = makecon(tip); 5868 } 5869 5870 // Next code copied from Parse::do_get_xxx(): 5871 5872 // Compute address and memory type. 5873 int offset = field->offset_in_bytes(); 5874 bool is_vol = field->is_volatile(); 5875 ciType* field_klass = field->type(); 5876 assert(field_klass->is_loaded(), "should be loaded"); 5877 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5878 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5879 BasicType bt = field->layout_type(); 5880 5881 // Build the resultant type of the load 5882 const Type *type; 5883 if (bt == T_OBJECT) { 5884 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5885 } else { 5886 type = Type::get_const_basic_type(bt); 5887 } 5888 5889 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5890 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5891 } 5892 // Build the load. 5893 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5894 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5895 // If reference is volatile, prevent following memory ops from 5896 // floating up past the volatile read. Also prevents commoning 5897 // another volatile read. 5898 if (is_vol) { 5899 // Memory barrier includes bogus read of value to force load BEFORE membar 5900 insert_mem_bar(Op_MemBarAcquire, loadedField); 5901 } 5902 return loadedField; 5903} 5904 5905 5906//------------------------------inline_aescrypt_Block----------------------- 5907bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5908 address stubAddr; 5909 const char *stubName; 5910 assert(UseAES, "need AES instruction support"); 5911 5912 switch(id) { 5913 case vmIntrinsics::_aescrypt_encryptBlock: 5914 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5915 stubName = "aescrypt_encryptBlock"; 5916 break; 5917 case vmIntrinsics::_aescrypt_decryptBlock: 5918 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5919 stubName = "aescrypt_decryptBlock"; 5920 break; 5921 } 5922 if (stubAddr == NULL) return false; 5923 5924 Node* aescrypt_object = argument(0); 5925 Node* src = argument(1); 5926 Node* src_offset = argument(2); 5927 Node* dest = argument(3); 5928 Node* dest_offset = argument(4); 5929 5930 // (1) src and dest are arrays. 5931 const Type* src_type = src->Value(&_gvn); 5932 const Type* dest_type = dest->Value(&_gvn); 5933 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5934 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5935 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5936 5937 // for the quick and dirty code we will skip all the checks. 5938 // we are just trying to get the call to be generated. 5939 Node* src_start = src; 5940 Node* dest_start = dest; 5941 if (src_offset != NULL || dest_offset != NULL) { 5942 assert(src_offset != NULL && dest_offset != NULL, ""); 5943 src_start = array_element_address(src, src_offset, T_BYTE); 5944 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5945 } 5946 5947 // now need to get the start of its expanded key array 5948 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5949 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5950 if (k_start == NULL) return false; 5951 5952 if (Matcher::pass_original_key_for_aes()) { 5953 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5954 // compatibility issues between Java key expansion and SPARC crypto instructions 5955 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5956 if (original_k_start == NULL) return false; 5957 5958 // Call the stub. 5959 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5960 stubAddr, stubName, TypePtr::BOTTOM, 5961 src_start, dest_start, k_start, original_k_start); 5962 } else { 5963 // Call the stub. 5964 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5965 stubAddr, stubName, TypePtr::BOTTOM, 5966 src_start, dest_start, k_start); 5967 } 5968 5969 return true; 5970} 5971 5972//------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5973bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5974 address stubAddr; 5975 const char *stubName; 5976 5977 assert(UseAES, "need AES instruction support"); 5978 5979 switch(id) { 5980 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5981 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5982 stubName = "cipherBlockChaining_encryptAESCrypt"; 5983 break; 5984 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5985 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5986 stubName = "cipherBlockChaining_decryptAESCrypt"; 5987 break; 5988 } 5989 if (stubAddr == NULL) return false; 5990 5991 Node* cipherBlockChaining_object = argument(0); 5992 Node* src = argument(1); 5993 Node* src_offset = argument(2); 5994 Node* len = argument(3); 5995 Node* dest = argument(4); 5996 Node* dest_offset = argument(5); 5997 5998 // (1) src and dest are arrays. 5999 const Type* src_type = src->Value(&_gvn); 6000 const Type* dest_type = dest->Value(&_gvn); 6001 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6002 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6003 assert (top_src != NULL && top_src->klass() != NULL 6004 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6005 6006 // checks are the responsibility of the caller 6007 Node* src_start = src; 6008 Node* dest_start = dest; 6009 if (src_offset != NULL || dest_offset != NULL) { 6010 assert(src_offset != NULL && dest_offset != NULL, ""); 6011 src_start = array_element_address(src, src_offset, T_BYTE); 6012 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6013 } 6014 6015 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6016 // (because of the predicated logic executed earlier). 6017 // so we cast it here safely. 6018 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6019 6020 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6021 if (embeddedCipherObj == NULL) return false; 6022 6023 // cast it to what we know it will be at runtime 6024 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6025 assert(tinst != NULL, "CBC obj is null"); 6026 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6027 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6028 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6029 6030 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6031 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6032 const TypeOopPtr* xtype = aklass->as_instance_type(); 6033 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6034 aescrypt_object = _gvn.transform(aescrypt_object); 6035 6036 // we need to get the start of the aescrypt_object's expanded key array 6037 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6038 if (k_start == NULL) return false; 6039 6040 // similarly, get the start address of the r vector 6041 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6042 if (objRvec == NULL) return false; 6043 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6044 6045 Node* cbcCrypt; 6046 if (Matcher::pass_original_key_for_aes()) { 6047 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6048 // compatibility issues between Java key expansion and SPARC crypto instructions 6049 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6050 if (original_k_start == NULL) return false; 6051 6052 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6053 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6054 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6055 stubAddr, stubName, TypePtr::BOTTOM, 6056 src_start, dest_start, k_start, r_start, len, original_k_start); 6057 } else { 6058 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6059 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6060 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6061 stubAddr, stubName, TypePtr::BOTTOM, 6062 src_start, dest_start, k_start, r_start, len); 6063 } 6064 6065 // return cipher length (int) 6066 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6067 set_result(retvalue); 6068 return true; 6069} 6070 6071//------------------------------get_key_start_from_aescrypt_object----------------------- 6072Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6073 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6074 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6075 if (objAESCryptKey == NULL) return (Node *) NULL; 6076 6077 // now have the array, need to get the start address of the K array 6078 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6079 return k_start; 6080} 6081 6082//------------------------------get_original_key_start_from_aescrypt_object----------------------- 6083Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6084 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6085 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6086 if (objAESCryptKey == NULL) return (Node *) NULL; 6087 6088 // now have the array, need to get the start address of the lastKey array 6089 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6090 return original_k_start; 6091} 6092 6093//----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6094// Return node representing slow path of predicate check. 6095// the pseudo code we want to emulate with this predicate is: 6096// for encryption: 6097// if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6098// for decryption: 6099// if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6100// note cipher==plain is more conservative than the original java code but that's OK 6101// 6102Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6103 // The receiver was checked for NULL already. 6104 Node* objCBC = argument(0); 6105 6106 // Load embeddedCipher field of CipherBlockChaining object. 6107 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6108 6109 // get AESCrypt klass for instanceOf check 6110 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6111 // will have same classloader as CipherBlockChaining object 6112 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6113 assert(tinst != NULL, "CBCobj is null"); 6114 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6115 6116 // we want to do an instanceof comparison against the AESCrypt class 6117 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6118 if (!klass_AESCrypt->is_loaded()) { 6119 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6120 Node* ctrl = control(); 6121 set_control(top()); // no regular fast path 6122 return ctrl; 6123 } 6124 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6125 6126 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6127 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6128 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6129 6130 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6131 6132 // for encryption, we are done 6133 if (!decrypting) 6134 return instof_false; // even if it is NULL 6135 6136 // for decryption, we need to add a further check to avoid 6137 // taking the intrinsic path when cipher and plain are the same 6138 // see the original java code for why. 6139 RegionNode* region = new RegionNode(3); 6140 region->init_req(1, instof_false); 6141 Node* src = argument(1); 6142 Node* dest = argument(4); 6143 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6144 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6145 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6146 region->init_req(2, src_dest_conjoint); 6147 6148 record_for_igvn(region); 6149 return _gvn.transform(region); 6150} 6151 6152//------------------------------inline_ghash_processBlocks 6153bool LibraryCallKit::inline_ghash_processBlocks() { 6154 address stubAddr; 6155 const char *stubName; 6156 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6157 6158 stubAddr = StubRoutines::ghash_processBlocks(); 6159 stubName = "ghash_processBlocks"; 6160 6161 Node* data = argument(0); 6162 Node* offset = argument(1); 6163 Node* len = argument(2); 6164 Node* state = argument(3); 6165 Node* subkeyH = argument(4); 6166 6167 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6168 assert(state_start, "state is NULL"); 6169 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6170 assert(subkeyH_start, "subkeyH is NULL"); 6171 Node* data_start = array_element_address(data, offset, T_BYTE); 6172 assert(data_start, "data is NULL"); 6173 6174 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6175 OptoRuntime::ghash_processBlocks_Type(), 6176 stubAddr, stubName, TypePtr::BOTTOM, 6177 state_start, subkeyH_start, data_start, len); 6178 return true; 6179} 6180 6181//------------------------------inline_sha_implCompress----------------------- 6182// 6183// Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6184// void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6185// 6186// Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6187// void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6188// 6189// Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6190// void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6191// 6192bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6193 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6194 6195 Node* sha_obj = argument(0); 6196 Node* src = argument(1); // type oop 6197 Node* ofs = argument(2); // type int 6198 6199 const Type* src_type = src->Value(&_gvn); 6200 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6201 if (top_src == NULL || top_src->klass() == NULL) { 6202 // failed array check 6203 return false; 6204 } 6205 // Figure out the size and type of the elements we will be copying. 6206 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6207 if (src_elem != T_BYTE) { 6208 return false; 6209 } 6210 // 'src_start' points to src array + offset 6211 Node* src_start = array_element_address(src, ofs, src_elem); 6212 Node* state = NULL; 6213 address stubAddr; 6214 const char *stubName; 6215 6216 switch(id) { 6217 case vmIntrinsics::_sha_implCompress: 6218 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6219 state = get_state_from_sha_object(sha_obj); 6220 stubAddr = StubRoutines::sha1_implCompress(); 6221 stubName = "sha1_implCompress"; 6222 break; 6223 case vmIntrinsics::_sha2_implCompress: 6224 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6225 state = get_state_from_sha_object(sha_obj); 6226 stubAddr = StubRoutines::sha256_implCompress(); 6227 stubName = "sha256_implCompress"; 6228 break; 6229 case vmIntrinsics::_sha5_implCompress: 6230 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6231 state = get_state_from_sha5_object(sha_obj); 6232 stubAddr = StubRoutines::sha512_implCompress(); 6233 stubName = "sha512_implCompress"; 6234 break; 6235 default: 6236 fatal_unexpected_iid(id); 6237 return false; 6238 } 6239 if (state == NULL) return false; 6240 6241 // Call the stub. 6242 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6243 stubAddr, stubName, TypePtr::BOTTOM, 6244 src_start, state); 6245 6246 return true; 6247} 6248 6249//------------------------------inline_digestBase_implCompressMB----------------------- 6250// 6251// Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6252// int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6253// 6254bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6255 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6256 "need SHA1/SHA256/SHA512 instruction support"); 6257 assert((uint)predicate < 3, "sanity"); 6258 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6259 6260 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6261 Node* src = argument(1); // byte[] array 6262 Node* ofs = argument(2); // type int 6263 Node* limit = argument(3); // type int 6264 6265 const Type* src_type = src->Value(&_gvn); 6266 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6267 if (top_src == NULL || top_src->klass() == NULL) { 6268 // failed array check 6269 return false; 6270 } 6271 // Figure out the size and type of the elements we will be copying. 6272 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6273 if (src_elem != T_BYTE) { 6274 return false; 6275 } 6276 // 'src_start' points to src array + offset 6277 Node* src_start = array_element_address(src, ofs, src_elem); 6278 6279 const char* klass_SHA_name = NULL; 6280 const char* stub_name = NULL; 6281 address stub_addr = NULL; 6282 bool long_state = false; 6283 6284 switch (predicate) { 6285 case 0: 6286 if (UseSHA1Intrinsics) { 6287 klass_SHA_name = "sun/security/provider/SHA"; 6288 stub_name = "sha1_implCompressMB"; 6289 stub_addr = StubRoutines::sha1_implCompressMB(); 6290 } 6291 break; 6292 case 1: 6293 if (UseSHA256Intrinsics) { 6294 klass_SHA_name = "sun/security/provider/SHA2"; 6295 stub_name = "sha256_implCompressMB"; 6296 stub_addr = StubRoutines::sha256_implCompressMB(); 6297 } 6298 break; 6299 case 2: 6300 if (UseSHA512Intrinsics) { 6301 klass_SHA_name = "sun/security/provider/SHA5"; 6302 stub_name = "sha512_implCompressMB"; 6303 stub_addr = StubRoutines::sha512_implCompressMB(); 6304 long_state = true; 6305 } 6306 break; 6307 default: 6308 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6309 } 6310 if (klass_SHA_name != NULL) { 6311 // get DigestBase klass to lookup for SHA klass 6312 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6313 assert(tinst != NULL, "digestBase_obj is not instance???"); 6314 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6315 6316 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6317 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6318 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6319 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6320 } 6321 return false; 6322} 6323//------------------------------inline_sha_implCompressMB----------------------- 6324bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6325 bool long_state, address stubAddr, const char *stubName, 6326 Node* src_start, Node* ofs, Node* limit) { 6327 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6328 const TypeOopPtr* xtype = aklass->as_instance_type(); 6329 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6330 sha_obj = _gvn.transform(sha_obj); 6331 6332 Node* state; 6333 if (long_state) { 6334 state = get_state_from_sha5_object(sha_obj); 6335 } else { 6336 state = get_state_from_sha_object(sha_obj); 6337 } 6338 if (state == NULL) return false; 6339 6340 // Call the stub. 6341 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6342 OptoRuntime::digestBase_implCompressMB_Type(), 6343 stubAddr, stubName, TypePtr::BOTTOM, 6344 src_start, state, ofs, limit); 6345 // return ofs (int) 6346 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6347 set_result(result); 6348 6349 return true; 6350} 6351 6352//------------------------------get_state_from_sha_object----------------------- 6353Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6354 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6355 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6356 if (sha_state == NULL) return (Node *) NULL; 6357 6358 // now have the array, need to get the start address of the state array 6359 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6360 return state; 6361} 6362 6363//------------------------------get_state_from_sha5_object----------------------- 6364Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6365 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6366 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6367 if (sha_state == NULL) return (Node *) NULL; 6368 6369 // now have the array, need to get the start address of the state array 6370 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6371 return state; 6372} 6373 6374//----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6375// Return node representing slow path of predicate check. 6376// the pseudo code we want to emulate with this predicate is: 6377// if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6378// 6379Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6380 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6381 "need SHA1/SHA256/SHA512 instruction support"); 6382 assert((uint)predicate < 3, "sanity"); 6383 6384 // The receiver was checked for NULL already. 6385 Node* digestBaseObj = argument(0); 6386 6387 // get DigestBase klass for instanceOf check 6388 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6389 assert(tinst != NULL, "digestBaseObj is null"); 6390 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6391 6392 const char* klass_SHA_name = NULL; 6393 switch (predicate) { 6394 case 0: 6395 if (UseSHA1Intrinsics) { 6396 // we want to do an instanceof comparison against the SHA class 6397 klass_SHA_name = "sun/security/provider/SHA"; 6398 } 6399 break; 6400 case 1: 6401 if (UseSHA256Intrinsics) { 6402 // we want to do an instanceof comparison against the SHA2 class 6403 klass_SHA_name = "sun/security/provider/SHA2"; 6404 } 6405 break; 6406 case 2: 6407 if (UseSHA512Intrinsics) { 6408 // we want to do an instanceof comparison against the SHA5 class 6409 klass_SHA_name = "sun/security/provider/SHA5"; 6410 } 6411 break; 6412 default: 6413 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6414 } 6415 6416 ciKlass* klass_SHA = NULL; 6417 if (klass_SHA_name != NULL) { 6418 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6419 } 6420 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6421 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6422 Node* ctrl = control(); 6423 set_control(top()); // no intrinsic path 6424 return ctrl; 6425 } 6426 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6427 6428 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6429 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6430 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6431 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6432 6433 return instof_false; // even if it is NULL 6434} 6435 6436bool LibraryCallKit::inline_profileBoolean() { 6437 Node* counts = argument(1); 6438 const TypeAryPtr* ary = NULL; 6439 ciArray* aobj = NULL; 6440 if (counts->is_Con() 6441 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6442 && (aobj = ary->const_oop()->as_array()) != NULL 6443 && (aobj->length() == 2)) { 6444 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6445 jint false_cnt = aobj->element_value(0).as_int(); 6446 jint true_cnt = aobj->element_value(1).as_int(); 6447 6448 if (C->log() != NULL) { 6449 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6450 false_cnt, true_cnt); 6451 } 6452 6453 if (false_cnt + true_cnt == 0) { 6454 // According to profile, never executed. 6455 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6456 Deoptimization::Action_reinterpret); 6457 return true; 6458 } 6459 6460 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6461 // is a number of each value occurrences. 6462 Node* result = argument(0); 6463 if (false_cnt == 0 || true_cnt == 0) { 6464 // According to profile, one value has been never seen. 6465 int expected_val = (false_cnt == 0) ? 1 : 0; 6466 6467 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6468 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6469 6470 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6471 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6472 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6473 6474 { // Slow path: uncommon trap for never seen value and then reexecute 6475 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6476 // the value has been seen at least once. 6477 PreserveJVMState pjvms(this); 6478 PreserveReexecuteState preexecs(this); 6479 jvms()->set_should_reexecute(true); 6480 6481 set_control(slow_path); 6482 set_i_o(i_o()); 6483 6484 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6485 Deoptimization::Action_reinterpret); 6486 } 6487 // The guard for never seen value enables sharpening of the result and 6488 // returning a constant. It allows to eliminate branches on the same value 6489 // later on. 6490 set_control(fast_path); 6491 result = intcon(expected_val); 6492 } 6493 // Stop profiling. 6494 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6495 // By replacing method body with profile data (represented as ProfileBooleanNode 6496 // on IR level) we effectively disable profiling. 6497 // It enables full speed execution once optimized code is generated. 6498 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6499 C->record_for_igvn(profile); 6500 set_result(profile); 6501 return true; 6502 } else { 6503 // Continue profiling. 6504 // Profile data isn't available at the moment. So, execute method's bytecode version. 6505 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6506 // is compiled and counters aren't available since corresponding MethodHandle 6507 // isn't a compile-time constant. 6508 return false; 6509 } 6510} 6511 6512bool LibraryCallKit::inline_isCompileConstant() { 6513 Node* n = argument(0); 6514 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6515 return true; 6516} 6517