1 /* 2 * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "classfile/vmSymbols.hpp" 28 #include "compiler/compileBroker.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "oops/objArrayKlass.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/callGenerator.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/idealKit.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/parse.hpp" 37 #include "opto/runtime.hpp" 38 #include "opto/subnode.hpp" 39 #include "prims/nativeLookup.hpp" 40 #include "runtime/sharedRuntime.hpp" 41 #include "trace/traceMacros.hpp" 42 43 class LibraryIntrinsic : public InlineCallGenerator { 44 // Extend the set of intrinsics known to the runtime: 45 public: 46 private: 47 bool _is_virtual; 48 bool _is_predicted; 49 vmIntrinsics::ID _intrinsic_id; 50 51 public: 52 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, vmIntrinsics::ID id) 53 : InlineCallGenerator(m), 54 _is_virtual(is_virtual), 55 _is_predicted(is_predicted), 56 _intrinsic_id(id) 57 { 58 } 59 virtual bool is_intrinsic() const { return true; } 60 virtual bool is_virtual() const { return _is_virtual; } 61 virtual bool is_predicted() const { return _is_predicted; } 62 virtual JVMState* generate(JVMState* jvms); 63 virtual Node* generate_predicate(JVMState* jvms); 64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 65 }; 66 67 68 // Local helper class for LibraryIntrinsic: 69 class LibraryCallKit : public GraphKit { 70 private: 71 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 72 Node* _result; // the result node, if any 73 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 74 75 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 76 77 public: 78 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 79 : GraphKit(jvms), 80 _intrinsic(intrinsic), 81 _result(NULL) 82 { 83 // Check if this is a root compile. In that case we don't have a caller. 84 if (!jvms->has_method()) { 85 _reexecute_sp = sp(); 86 } else { 87 // Find out how many arguments the interpreter needs when deoptimizing 88 // and save the stack pointer value so it can used by uncommon_trap. 89 // We find the argument count by looking at the declared signature. 90 bool ignored_will_link; 91 ciSignature* declared_signature = NULL; 92 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 93 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 94 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 95 } 96 } 97 98 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 99 100 ciMethod* caller() const { return jvms()->method(); } 101 int bci() const { return jvms()->bci(); } 102 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 103 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 104 ciMethod* callee() const { return _intrinsic->method(); } 105 106 bool try_to_inline(); 107 Node* try_to_predicate(); 108 109 void push_result() { 110 // Push the result onto the stack. 111 if (!stopped() && result() != NULL) { 112 BasicType bt = result()->bottom_type()->basic_type(); 113 push_node(bt, result()); 114 } 115 } 116 117 private: 118 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 119 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 120 } 121 122 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 123 void set_result(RegionNode* region, PhiNode* value); 124 Node* result() { return _result; } 125 126 virtual int reexecute_sp() { return _reexecute_sp; } 127 128 // Helper functions to inline natives 129 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 130 Node* generate_slow_guard(Node* test, RegionNode* region); 131 Node* generate_fair_guard(Node* test, RegionNode* region); 132 Node* generate_negative_guard(Node* index, RegionNode* region, 133 // resulting CastII of index: 134 Node* *pos_index = NULL); 135 Node* generate_nonpositive_guard(Node* index, bool never_negative, 136 // resulting CastII of index: 137 Node* *pos_index = NULL); 138 Node* generate_limit_guard(Node* offset, Node* subseq_length, 139 Node* array_length, 140 RegionNode* region); 141 Node* generate_current_thread(Node* &tls_output); 142 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 143 bool disjoint_bases, const char* &name, bool dest_uninitialized); 144 Node* load_mirror_from_klass(Node* klass); 145 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 146 RegionNode* region, int null_path, 147 int offset); 148 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 149 RegionNode* region, int null_path) { 150 int offset = java_lang_Class::klass_offset_in_bytes(); 151 return load_klass_from_mirror_common(mirror, never_see_null, 152 region, null_path, 153 offset); 154 } 155 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 156 RegionNode* region, int null_path) { 157 int offset = java_lang_Class::array_klass_offset_in_bytes(); 158 return load_klass_from_mirror_common(mirror, never_see_null, 159 region, null_path, 160 offset); 161 } 162 Node* generate_access_flags_guard(Node* kls, 163 int modifier_mask, int modifier_bits, 164 RegionNode* region); 165 Node* generate_interface_guard(Node* kls, RegionNode* region); 166 Node* generate_array_guard(Node* kls, RegionNode* region) { 167 return generate_array_guard_common(kls, region, false, false); 168 } 169 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 170 return generate_array_guard_common(kls, region, false, true); 171 } 172 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 173 return generate_array_guard_common(kls, region, true, false); 174 } 175 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 176 return generate_array_guard_common(kls, region, true, true); 177 } 178 Node* generate_array_guard_common(Node* kls, RegionNode* region, 179 bool obj_array, bool not_array); 180 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 181 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 182 bool is_virtual = false, bool is_static = false); 183 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 184 return generate_method_call(method_id, false, true); 185 } 186 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 187 return generate_method_call(method_id, true, false); 188 } 189 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 190 191 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 192 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 193 bool inline_string_compareTo(); 194 bool inline_string_indexOf(); 195 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 196 bool inline_string_equals(); 197 Node* round_double_node(Node* n); 198 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 199 bool inline_math_native(vmIntrinsics::ID id); 200 bool inline_trig(vmIntrinsics::ID id); 201 bool inline_math(vmIntrinsics::ID id); 202 bool inline_exp(); 203 bool inline_pow(); 204 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 205 bool inline_min_max(vmIntrinsics::ID id); 206 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 207 // This returns Type::AnyPtr, RawPtr, or OopPtr. 208 int classify_unsafe_addr(Node* &base, Node* &offset); 209 Node* make_unsafe_address(Node* base, Node* offset); 210 // Helper for inline_unsafe_access. 211 // Generates the guards that check whether the result of 212 // Unsafe.getObject should be recorded in an SATB log buffer. 213 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 214 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 215 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 216 bool inline_unsafe_allocate(); 217 bool inline_unsafe_copyMemory(); 218 bool inline_native_currentThread(); 219 #ifdef TRACE_HAVE_INTRINSICS 220 bool inline_native_classID(); 221 bool inline_native_threadID(); 222 #endif 223 bool inline_native_time_funcs(address method, const char* funcName); 224 bool inline_native_isInterrupted(); 225 bool inline_native_Class_query(vmIntrinsics::ID id); 226 bool inline_native_subtype_check(); 227 228 bool inline_native_newArray(); 229 bool inline_native_getLength(); 230 bool inline_array_copyOf(bool is_copyOfRange); 231 bool inline_array_equals(); 232 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 233 bool inline_native_clone(bool is_virtual); 234 bool inline_native_Reflection_getCallerClass(); 235 // Helper function for inlining native object hash method 236 bool inline_native_hashcode(bool is_virtual, bool is_static); 237 bool inline_native_getClass(); 238 239 // Helper functions for inlining arraycopy 240 bool inline_arraycopy(); 241 void generate_arraycopy(const TypePtr* adr_type, 242 BasicType basic_elem_type, 243 Node* src, Node* src_offset, 244 Node* dest, Node* dest_offset, 245 Node* copy_length, 246 bool disjoint_bases = false, 247 bool length_never_negative = false, 248 RegionNode* slow_region = NULL); 249 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 250 RegionNode* slow_region); 251 void generate_clear_array(const TypePtr* adr_type, 252 Node* dest, 253 BasicType basic_elem_type, 254 Node* slice_off, 255 Node* slice_len, 256 Node* slice_end); 257 bool generate_block_arraycopy(const TypePtr* adr_type, 258 BasicType basic_elem_type, 259 AllocateNode* alloc, 260 Node* src, Node* src_offset, 261 Node* dest, Node* dest_offset, 262 Node* dest_size, bool dest_uninitialized); 263 void generate_slow_arraycopy(const TypePtr* adr_type, 264 Node* src, Node* src_offset, 265 Node* dest, Node* dest_offset, 266 Node* copy_length, bool dest_uninitialized); 267 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 268 Node* dest_elem_klass, 269 Node* src, Node* src_offset, 270 Node* dest, Node* dest_offset, 271 Node* copy_length, bool dest_uninitialized); 272 Node* generate_generic_arraycopy(const TypePtr* adr_type, 273 Node* src, Node* src_offset, 274 Node* dest, Node* dest_offset, 275 Node* copy_length, bool dest_uninitialized); 276 void generate_unchecked_arraycopy(const TypePtr* adr_type, 277 BasicType basic_elem_type, 278 bool disjoint_bases, 279 Node* src, Node* src_offset, 280 Node* dest, Node* dest_offset, 281 Node* copy_length, bool dest_uninitialized); 282 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 283 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 284 bool inline_unsafe_ordered_store(BasicType type); 285 bool inline_unsafe_fence(vmIntrinsics::ID id); 286 bool inline_fp_conversions(vmIntrinsics::ID id); 287 bool inline_number_methods(vmIntrinsics::ID id); 288 bool inline_reference_get(); 289 bool inline_aescrypt_Block(vmIntrinsics::ID id); 290 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 291 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 292 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 293 bool inline_encodeISOArray(); 294 bool inline_updateCRC32(); 295 bool inline_updateBytesCRC32(); 296 bool inline_updateByteBufferCRC32(); 297 }; 298 299 300 //---------------------------make_vm_intrinsic---------------------------- 301 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 302 vmIntrinsics::ID id = m->intrinsic_id(); 303 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 304 305 if (DisableIntrinsic[0] != '\0' 306 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 307 // disabled by a user request on the command line: 308 // example: -XX:DisableIntrinsic=_hashCode,_getClass 309 return NULL; 310 } 311 312 if (!m->is_loaded()) { 313 // do not attempt to inline unloaded methods 314 return NULL; 315 } 316 317 // Only a few intrinsics implement a virtual dispatch. 318 // They are expensive calls which are also frequently overridden. 319 if (is_virtual) { 320 switch (id) { 321 case vmIntrinsics::_hashCode: 322 case vmIntrinsics::_clone: 323 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 324 break; 325 default: 326 return NULL; 327 } 328 } 329 330 // -XX:-InlineNatives disables nearly all intrinsics: 331 if (!InlineNatives) { 332 switch (id) { 333 case vmIntrinsics::_indexOf: 334 case vmIntrinsics::_compareTo: 335 case vmIntrinsics::_equals: 336 case vmIntrinsics::_equalsC: 337 case vmIntrinsics::_getAndAddInt: 338 case vmIntrinsics::_getAndAddLong: 339 case vmIntrinsics::_getAndSetInt: 340 case vmIntrinsics::_getAndSetLong: 341 case vmIntrinsics::_getAndSetObject: 342 case vmIntrinsics::_loadFence: 343 case vmIntrinsics::_storeFence: 344 case vmIntrinsics::_fullFence: 345 break; // InlineNatives does not control String.compareTo 346 case vmIntrinsics::_Reference_get: 347 break; // InlineNatives does not control Reference.get 348 default: 349 return NULL; 350 } 351 } 352 353 bool is_predicted = false; 354 355 switch (id) { 356 case vmIntrinsics::_compareTo: 357 if (!SpecialStringCompareTo) return NULL; 358 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 359 break; 360 case vmIntrinsics::_indexOf: 361 if (!SpecialStringIndexOf) return NULL; 362 break; 363 case vmIntrinsics::_equals: 364 if (!SpecialStringEquals) return NULL; 365 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 366 break; 367 case vmIntrinsics::_equalsC: 368 if (!SpecialArraysEquals) return NULL; 369 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 370 break; 371 case vmIntrinsics::_arraycopy: 372 if (!InlineArrayCopy) return NULL; 373 break; 374 case vmIntrinsics::_copyMemory: 375 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 376 if (!InlineArrayCopy) return NULL; 377 break; 378 case vmIntrinsics::_hashCode: 379 if (!InlineObjectHash) return NULL; 380 break; 381 case vmIntrinsics::_clone: 382 case vmIntrinsics::_copyOf: 383 case vmIntrinsics::_copyOfRange: 384 if (!InlineObjectCopy) return NULL; 385 // These also use the arraycopy intrinsic mechanism: 386 if (!InlineArrayCopy) return NULL; 387 break; 388 case vmIntrinsics::_encodeISOArray: 389 if (!SpecialEncodeISOArray) return NULL; 390 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 391 break; 392 case vmIntrinsics::_checkIndex: 393 // We do not intrinsify this. The optimizer does fine with it. 394 return NULL; 395 396 case vmIntrinsics::_getCallerClass: 397 if (!UseNewReflection) return NULL; 398 if (!InlineReflectionGetCallerClass) return NULL; 399 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 400 break; 401 402 case vmIntrinsics::_bitCount_i: 403 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 404 break; 405 406 case vmIntrinsics::_bitCount_l: 407 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 408 break; 409 410 case vmIntrinsics::_numberOfLeadingZeros_i: 411 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 412 break; 413 414 case vmIntrinsics::_numberOfLeadingZeros_l: 415 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 416 break; 417 418 case vmIntrinsics::_numberOfTrailingZeros_i: 419 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 420 break; 421 422 case vmIntrinsics::_numberOfTrailingZeros_l: 423 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 424 break; 425 426 case vmIntrinsics::_reverseBytes_c: 427 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 428 break; 429 case vmIntrinsics::_reverseBytes_s: 430 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 431 break; 432 case vmIntrinsics::_reverseBytes_i: 433 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 434 break; 435 case vmIntrinsics::_reverseBytes_l: 436 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 437 break; 438 439 case vmIntrinsics::_Reference_get: 440 // Use the intrinsic version of Reference.get() so that the value in 441 // the referent field can be registered by the G1 pre-barrier code. 442 // Also add memory barrier to prevent commoning reads from this field 443 // across safepoint since GC can change it value. 444 break; 445 446 case vmIntrinsics::_compareAndSwapObject: 447 #ifdef _LP64 448 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 449 #endif 450 break; 451 452 case vmIntrinsics::_compareAndSwapLong: 453 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 454 break; 455 456 case vmIntrinsics::_getAndAddInt: 457 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 458 break; 459 460 case vmIntrinsics::_getAndAddLong: 461 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 462 break; 463 464 case vmIntrinsics::_getAndSetInt: 465 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 466 break; 467 468 case vmIntrinsics::_getAndSetLong: 469 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 470 break; 471 472 case vmIntrinsics::_getAndSetObject: 473 #ifdef _LP64 474 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 475 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 476 break; 477 #else 478 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 479 break; 480 #endif 481 482 case vmIntrinsics::_aescrypt_encryptBlock: 483 case vmIntrinsics::_aescrypt_decryptBlock: 484 if (!UseAESIntrinsics) return NULL; 485 break; 486 487 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 488 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 489 if (!UseAESIntrinsics) return NULL; 490 // these two require the predicated logic 491 is_predicted = true; 492 break; 493 494 case vmIntrinsics::_updateCRC32: 495 case vmIntrinsics::_updateBytesCRC32: 496 case vmIntrinsics::_updateByteBufferCRC32: 497 if (!UseCRC32Intrinsics) return NULL; 498 break; 499 500 default: 501 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 502 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 503 break; 504 } 505 506 // -XX:-InlineClassNatives disables natives from the Class class. 507 // The flag applies to all reflective calls, notably Array.newArray 508 // (visible to Java programmers as Array.newInstance). 509 if (m->holder()->name() == ciSymbol::java_lang_Class() || 510 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 511 if (!InlineClassNatives) return NULL; 512 } 513 514 // -XX:-InlineThreadNatives disables natives from the Thread class. 515 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 516 if (!InlineThreadNatives) return NULL; 517 } 518 519 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 520 if (m->holder()->name() == ciSymbol::java_lang_Math() || 521 m->holder()->name() == ciSymbol::java_lang_Float() || 522 m->holder()->name() == ciSymbol::java_lang_Double()) { 523 if (!InlineMathNatives) return NULL; 524 } 525 526 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 527 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 528 if (!InlineUnsafeOps) return NULL; 529 } 530 531 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id); 532 } 533 534 //----------------------register_library_intrinsics----------------------- 535 // Initialize this file's data structures, for each Compile instance. 536 void Compile::register_library_intrinsics() { 537 // Nothing to do here. 538 } 539 540 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 541 LibraryCallKit kit(jvms, this); 542 Compile* C = kit.C; 543 int nodes = C->unique(); 544 #ifndef PRODUCT 545 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 546 char buf[1000]; 547 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 548 tty->print_cr("Intrinsic %s", str); 549 } 550 #endif 551 ciMethod* callee = kit.callee(); 552 const int bci = kit.bci(); 553 554 // Try to inline the intrinsic. 555 if (kit.try_to_inline()) { 556 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 557 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 558 } 559 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 560 if (C->log()) { 561 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 562 vmIntrinsics::name_at(intrinsic_id()), 563 (is_virtual() ? " virtual='1'" : ""), 564 C->unique() - nodes); 565 } 566 // Push the result from the inlined method onto the stack. 567 kit.push_result(); 568 return kit.transfer_exceptions_into_jvms(); 569 } 570 571 // The intrinsic bailed out 572 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 573 if (jvms->has_method()) { 574 // Not a root compile. 575 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 576 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 577 } else { 578 // Root compile 579 tty->print("Did not generate intrinsic %s%s at bci:%d in", 580 vmIntrinsics::name_at(intrinsic_id()), 581 (is_virtual() ? " (virtual)" : ""), bci); 582 } 583 } 584 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 585 return NULL; 586 } 587 588 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) { 589 LibraryCallKit kit(jvms, this); 590 Compile* C = kit.C; 591 int nodes = C->unique(); 592 #ifndef PRODUCT 593 assert(is_predicted(), "sanity"); 594 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 595 char buf[1000]; 596 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 597 tty->print_cr("Predicate for intrinsic %s", str); 598 } 599 #endif 600 ciMethod* callee = kit.callee(); 601 const int bci = kit.bci(); 602 603 Node* slow_ctl = kit.try_to_predicate(); 604 if (!kit.failing()) { 605 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 606 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 607 } 608 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 609 if (C->log()) { 610 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 611 vmIntrinsics::name_at(intrinsic_id()), 612 (is_virtual() ? " virtual='1'" : ""), 613 C->unique() - nodes); 614 } 615 return slow_ctl; // Could be NULL if the check folds. 616 } 617 618 // The intrinsic bailed out 619 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 620 if (jvms->has_method()) { 621 // Not a root compile. 622 const char* msg = "failed to generate predicate for intrinsic"; 623 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 624 } else { 625 // Root compile 626 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 627 vmIntrinsics::name_at(intrinsic_id()), 628 (is_virtual() ? " (virtual)" : ""), bci); 629 } 630 } 631 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 632 return NULL; 633 } 634 635 bool LibraryCallKit::try_to_inline() { 636 // Handle symbolic names for otherwise undistinguished boolean switches: 637 const bool is_store = true; 638 const bool is_native_ptr = true; 639 const bool is_static = true; 640 const bool is_volatile = true; 641 642 if (!jvms()->has_method()) { 643 // Root JVMState has a null method. 644 assert(map()->memory()->Opcode() == Op_Parm, ""); 645 // Insert the memory aliasing node 646 set_all_memory(reset_memory()); 647 } 648 assert(merged_memory(), ""); 649 650 651 switch (intrinsic_id()) { 652 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 653 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 654 case vmIntrinsics::_getClass: return inline_native_getClass(); 655 656 case vmIntrinsics::_dsin: 657 case vmIntrinsics::_dcos: 658 case vmIntrinsics::_dtan: 659 case vmIntrinsics::_dabs: 660 case vmIntrinsics::_datan2: 661 case vmIntrinsics::_dsqrt: 662 case vmIntrinsics::_dexp: 663 case vmIntrinsics::_dlog: 664 case vmIntrinsics::_dlog10: 665 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 666 667 case vmIntrinsics::_min: 668 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 669 670 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 671 672 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 673 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 674 case vmIntrinsics::_equals: return inline_string_equals(); 675 676 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 677 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 678 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 679 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 680 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 681 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 682 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 683 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 684 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 685 686 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 687 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 688 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 689 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 690 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 691 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 692 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 693 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 694 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 695 696 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 697 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 698 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 699 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 700 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 701 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 702 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 703 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 704 705 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 706 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 707 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 708 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 709 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 710 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 711 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 712 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 713 714 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 715 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 716 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 717 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 718 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 719 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 720 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 721 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 722 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 723 724 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 725 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 726 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 727 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 728 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 729 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 730 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 731 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 732 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 733 734 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 735 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 736 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 737 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 738 739 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 740 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 741 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 742 743 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 744 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 745 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 746 747 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 748 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 749 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 750 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 751 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 752 753 case vmIntrinsics::_loadFence: 754 case vmIntrinsics::_storeFence: 755 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 756 757 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 758 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 759 760 #ifdef TRACE_HAVE_INTRINSICS 761 case vmIntrinsics::_classID: return inline_native_classID(); 762 case vmIntrinsics::_threadID: return inline_native_threadID(); 763 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 764 #endif 765 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 766 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 767 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 768 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 769 case vmIntrinsics::_newArray: return inline_native_newArray(); 770 case vmIntrinsics::_getLength: return inline_native_getLength(); 771 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 772 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 773 case vmIntrinsics::_equalsC: return inline_array_equals(); 774 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 775 776 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 777 778 case vmIntrinsics::_isInstance: 779 case vmIntrinsics::_getModifiers: 780 case vmIntrinsics::_isInterface: 781 case vmIntrinsics::_isArray: 782 case vmIntrinsics::_isPrimitive: 783 case vmIntrinsics::_getSuperclass: 784 case vmIntrinsics::_getComponentType: 785 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 786 787 case vmIntrinsics::_floatToRawIntBits: 788 case vmIntrinsics::_floatToIntBits: 789 case vmIntrinsics::_intBitsToFloat: 790 case vmIntrinsics::_doubleToRawLongBits: 791 case vmIntrinsics::_doubleToLongBits: 792 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 793 794 case vmIntrinsics::_numberOfLeadingZeros_i: 795 case vmIntrinsics::_numberOfLeadingZeros_l: 796 case vmIntrinsics::_numberOfTrailingZeros_i: 797 case vmIntrinsics::_numberOfTrailingZeros_l: 798 case vmIntrinsics::_bitCount_i: 799 case vmIntrinsics::_bitCount_l: 800 case vmIntrinsics::_reverseBytes_i: 801 case vmIntrinsics::_reverseBytes_l: 802 case vmIntrinsics::_reverseBytes_s: 803 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 804 805 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 806 807 case vmIntrinsics::_Reference_get: return inline_reference_get(); 808 809 case vmIntrinsics::_aescrypt_encryptBlock: 810 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 811 812 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 813 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 814 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 815 816 case vmIntrinsics::_encodeISOArray: 817 return inline_encodeISOArray(); 818 819 case vmIntrinsics::_updateCRC32: 820 return inline_updateCRC32(); 821 case vmIntrinsics::_updateBytesCRC32: 822 return inline_updateBytesCRC32(); 823 case vmIntrinsics::_updateByteBufferCRC32: 824 return inline_updateByteBufferCRC32(); 825 826 default: 827 // If you get here, it may be that someone has added a new intrinsic 828 // to the list in vmSymbols.hpp without implementing it here. 829 #ifndef PRODUCT 830 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 831 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 832 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 833 } 834 #endif 835 return false; 836 } 837 } 838 839 Node* LibraryCallKit::try_to_predicate() { 840 if (!jvms()->has_method()) { 841 // Root JVMState has a null method. 842 assert(map()->memory()->Opcode() == Op_Parm, ""); 843 // Insert the memory aliasing node 844 set_all_memory(reset_memory()); 845 } 846 assert(merged_memory(), ""); 847 848 switch (intrinsic_id()) { 849 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 850 return inline_cipherBlockChaining_AESCrypt_predicate(false); 851 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 852 return inline_cipherBlockChaining_AESCrypt_predicate(true); 853 854 default: 855 // If you get here, it may be that someone has added a new intrinsic 856 // to the list in vmSymbols.hpp without implementing it here. 857 #ifndef PRODUCT 858 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 859 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 860 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 861 } 862 #endif 863 Node* slow_ctl = control(); 864 set_control(top()); // No fast path instrinsic 865 return slow_ctl; 866 } 867 } 868 869 //------------------------------set_result------------------------------- 870 // Helper function for finishing intrinsics. 871 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 872 record_for_igvn(region); 873 set_control(_gvn.transform(region)); 874 set_result( _gvn.transform(value)); 875 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 876 } 877 878 //------------------------------generate_guard--------------------------- 879 // Helper function for generating guarded fast-slow graph structures. 880 // The given 'test', if true, guards a slow path. If the test fails 881 // then a fast path can be taken. (We generally hope it fails.) 882 // In all cases, GraphKit::control() is updated to the fast path. 883 // The returned value represents the control for the slow path. 884 // The return value is never 'top'; it is either a valid control 885 // or NULL if it is obvious that the slow path can never be taken. 886 // Also, if region and the slow control are not NULL, the slow edge 887 // is appended to the region. 888 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 889 if (stopped()) { 890 // Already short circuited. 891 return NULL; 892 } 893 894 // Build an if node and its projections. 895 // If test is true we take the slow path, which we assume is uncommon. 896 if (_gvn.type(test) == TypeInt::ZERO) { 897 // The slow branch is never taken. No need to build this guard. 898 return NULL; 899 } 900 901 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 902 903 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff)); 904 if (if_slow == top()) { 905 // The slow branch is never taken. No need to build this guard. 906 return NULL; 907 } 908 909 if (region != NULL) 910 region->add_req(if_slow); 911 912 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff)); 913 set_control(if_fast); 914 915 return if_slow; 916 } 917 918 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 919 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 920 } 921 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 922 return generate_guard(test, region, PROB_FAIR); 923 } 924 925 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 926 Node* *pos_index) { 927 if (stopped()) 928 return NULL; // already stopped 929 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 930 return NULL; // index is already adequately typed 931 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0))); 932 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 933 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 934 if (is_neg != NULL && pos_index != NULL) { 935 // Emulate effect of Parse::adjust_map_after_if. 936 Node* ccast = new (C) CastIINode(index, TypeInt::POS); 937 ccast->set_req(0, control()); 938 (*pos_index) = _gvn.transform(ccast); 939 } 940 return is_neg; 941 } 942 943 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 944 Node* *pos_index) { 945 if (stopped()) 946 return NULL; // already stopped 947 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 948 return NULL; // index is already adequately typed 949 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0))); 950 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 951 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq)); 952 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 953 if (is_notp != NULL && pos_index != NULL) { 954 // Emulate effect of Parse::adjust_map_after_if. 955 Node* ccast = new (C) CastIINode(index, TypeInt::POS1); 956 ccast->set_req(0, control()); 957 (*pos_index) = _gvn.transform(ccast); 958 } 959 return is_notp; 960 } 961 962 // Make sure that 'position' is a valid limit index, in [0..length]. 963 // There are two equivalent plans for checking this: 964 // A. (offset + copyLength) unsigned<= arrayLength 965 // B. offset <= (arrayLength - copyLength) 966 // We require that all of the values above, except for the sum and 967 // difference, are already known to be non-negative. 968 // Plan A is robust in the face of overflow, if offset and copyLength 969 // are both hugely positive. 970 // 971 // Plan B is less direct and intuitive, but it does not overflow at 972 // all, since the difference of two non-negatives is always 973 // representable. Whenever Java methods must perform the equivalent 974 // check they generally use Plan B instead of Plan A. 975 // For the moment we use Plan A. 976 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 977 Node* subseq_length, 978 Node* array_length, 979 RegionNode* region) { 980 if (stopped()) 981 return NULL; // already stopped 982 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 983 if (zero_offset && subseq_length->eqv_uncast(array_length)) 984 return NULL; // common case of whole-array copy 985 Node* last = subseq_length; 986 if (!zero_offset) // last += offset 987 last = _gvn.transform(new (C) AddINode(last, offset)); 988 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last)); 989 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 990 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 991 return is_over; 992 } 993 994 995 //--------------------------generate_current_thread-------------------- 996 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 997 ciKlass* thread_klass = env()->Thread_klass(); 998 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 999 Node* thread = _gvn.transform(new (C) ThreadLocalNode()); 1000 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1001 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 1002 tls_output = thread; 1003 return threadObj; 1004 } 1005 1006 1007 //------------------------------make_string_method_node------------------------ 1008 // Helper method for String intrinsic functions. This version is called 1009 // with str1 and str2 pointing to String object nodes. 1010 // 1011 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1012 Node* no_ctrl = NULL; 1013 1014 // Get start addr of string 1015 Node* str1_value = load_String_value(no_ctrl, str1); 1016 Node* str1_offset = load_String_offset(no_ctrl, str1); 1017 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1018 1019 // Get length of string 1 1020 Node* str1_len = load_String_length(no_ctrl, str1); 1021 1022 Node* str2_value = load_String_value(no_ctrl, str2); 1023 Node* str2_offset = load_String_offset(no_ctrl, str2); 1024 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1025 1026 Node* str2_len = NULL; 1027 Node* result = NULL; 1028 1029 switch (opcode) { 1030 case Op_StrIndexOf: 1031 // Get length of string 2 1032 str2_len = load_String_length(no_ctrl, str2); 1033 1034 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1035 str1_start, str1_len, str2_start, str2_len); 1036 break; 1037 case Op_StrComp: 1038 // Get length of string 2 1039 str2_len = load_String_length(no_ctrl, str2); 1040 1041 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1042 str1_start, str1_len, str2_start, str2_len); 1043 break; 1044 case Op_StrEquals: 1045 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1046 str1_start, str2_start, str1_len); 1047 break; 1048 default: 1049 ShouldNotReachHere(); 1050 return NULL; 1051 } 1052 1053 // All these intrinsics have checks. 1054 C->set_has_split_ifs(true); // Has chance for split-if optimization 1055 1056 return _gvn.transform(result); 1057 } 1058 1059 // Helper method for String intrinsic functions. This version is called 1060 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1061 // to Int nodes containing the lenghts of str1 and str2. 1062 // 1063 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1064 Node* result = NULL; 1065 switch (opcode) { 1066 case Op_StrIndexOf: 1067 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1068 str1_start, cnt1, str2_start, cnt2); 1069 break; 1070 case Op_StrComp: 1071 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1072 str1_start, cnt1, str2_start, cnt2); 1073 break; 1074 case Op_StrEquals: 1075 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1076 str1_start, str2_start, cnt1); 1077 break; 1078 default: 1079 ShouldNotReachHere(); 1080 return NULL; 1081 } 1082 1083 // All these intrinsics have checks. 1084 C->set_has_split_ifs(true); // Has chance for split-if optimization 1085 1086 return _gvn.transform(result); 1087 } 1088 1089 //------------------------------inline_string_compareTo------------------------ 1090 // public int java.lang.String.compareTo(String anotherString); 1091 bool LibraryCallKit::inline_string_compareTo() { 1092 Node* receiver = null_check(argument(0)); 1093 Node* arg = null_check(argument(1)); 1094 if (stopped()) { 1095 return true; 1096 } 1097 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1098 return true; 1099 } 1100 1101 //------------------------------inline_string_equals------------------------ 1102 bool LibraryCallKit::inline_string_equals() { 1103 Node* receiver = null_check_receiver(); 1104 // NOTE: Do not null check argument for String.equals() because spec 1105 // allows to specify NULL as argument. 1106 Node* argument = this->argument(1); 1107 if (stopped()) { 1108 return true; 1109 } 1110 1111 // paths (plus control) merge 1112 RegionNode* region = new (C) RegionNode(5); 1113 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 1114 1115 // does source == target string? 1116 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument)); 1117 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 1118 1119 Node* if_eq = generate_slow_guard(bol, NULL); 1120 if (if_eq != NULL) { 1121 // receiver == argument 1122 phi->init_req(2, intcon(1)); 1123 region->init_req(2, if_eq); 1124 } 1125 1126 // get String klass for instanceOf 1127 ciInstanceKlass* klass = env()->String_klass(); 1128 1129 if (!stopped()) { 1130 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1131 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1))); 1132 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 1133 1134 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1135 //instanceOf == true, fallthrough 1136 1137 if (inst_false != NULL) { 1138 phi->init_req(3, intcon(0)); 1139 region->init_req(3, inst_false); 1140 } 1141 } 1142 1143 if (!stopped()) { 1144 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1145 1146 // Properly cast the argument to String 1147 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type)); 1148 // This path is taken only when argument's type is String:NotNull. 1149 argument = cast_not_null(argument, false); 1150 1151 Node* no_ctrl = NULL; 1152 1153 // Get start addr of receiver 1154 Node* receiver_val = load_String_value(no_ctrl, receiver); 1155 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1156 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1157 1158 // Get length of receiver 1159 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1160 1161 // Get start addr of argument 1162 Node* argument_val = load_String_value(no_ctrl, argument); 1163 Node* argument_offset = load_String_offset(no_ctrl, argument); 1164 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1165 1166 // Get length of argument 1167 Node* argument_cnt = load_String_length(no_ctrl, argument); 1168 1169 // Check for receiver count != argument count 1170 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt)); 1171 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne)); 1172 Node* if_ne = generate_slow_guard(bol, NULL); 1173 if (if_ne != NULL) { 1174 phi->init_req(4, intcon(0)); 1175 region->init_req(4, if_ne); 1176 } 1177 1178 // Check for count == 0 is done by assembler code for StrEquals. 1179 1180 if (!stopped()) { 1181 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1182 phi->init_req(1, equals); 1183 region->init_req(1, control()); 1184 } 1185 } 1186 1187 // post merge 1188 set_control(_gvn.transform(region)); 1189 record_for_igvn(region); 1190 1191 set_result(_gvn.transform(phi)); 1192 return true; 1193 } 1194 1195 //------------------------------inline_array_equals---------------------------- 1196 bool LibraryCallKit::inline_array_equals() { 1197 Node* arg1 = argument(0); 1198 Node* arg2 = argument(1); 1199 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1200 return true; 1201 } 1202 1203 // Java version of String.indexOf(constant string) 1204 // class StringDecl { 1205 // StringDecl(char[] ca) { 1206 // offset = 0; 1207 // count = ca.length; 1208 // value = ca; 1209 // } 1210 // int offset; 1211 // int count; 1212 // char[] value; 1213 // } 1214 // 1215 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1216 // int targetOffset, int cache_i, int md2) { 1217 // int cache = cache_i; 1218 // int sourceOffset = string_object.offset; 1219 // int sourceCount = string_object.count; 1220 // int targetCount = target_object.length; 1221 // 1222 // int targetCountLess1 = targetCount - 1; 1223 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1224 // 1225 // char[] source = string_object.value; 1226 // char[] target = target_object; 1227 // int lastChar = target[targetCountLess1]; 1228 // 1229 // outer_loop: 1230 // for (int i = sourceOffset; i < sourceEnd; ) { 1231 // int src = source[i + targetCountLess1]; 1232 // if (src == lastChar) { 1233 // // With random strings and a 4-character alphabet, 1234 // // reverse matching at this point sets up 0.8% fewer 1235 // // frames, but (paradoxically) makes 0.3% more probes. 1236 // // Since those probes are nearer the lastChar probe, 1237 // // there is may be a net D$ win with reverse matching. 1238 // // But, reversing loop inhibits unroll of inner loop 1239 // // for unknown reason. So, does running outer loop from 1240 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1241 // for (int j = 0; j < targetCountLess1; j++) { 1242 // if (target[targetOffset + j] != source[i+j]) { 1243 // if ((cache & (1 << source[i+j])) == 0) { 1244 // if (md2 < j+1) { 1245 // i += j+1; 1246 // continue outer_loop; 1247 // } 1248 // } 1249 // i += md2; 1250 // continue outer_loop; 1251 // } 1252 // } 1253 // return i - sourceOffset; 1254 // } 1255 // if ((cache & (1 << src)) == 0) { 1256 // i += targetCountLess1; 1257 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1258 // i++; 1259 // } 1260 // return -1; 1261 // } 1262 1263 //------------------------------string_indexOf------------------------ 1264 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1265 jint cache_i, jint md2_i) { 1266 1267 Node* no_ctrl = NULL; 1268 float likely = PROB_LIKELY(0.9); 1269 float unlikely = PROB_UNLIKELY(0.9); 1270 1271 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1272 1273 Node* source = load_String_value(no_ctrl, string_object); 1274 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1275 Node* sourceCount = load_String_length(no_ctrl, string_object); 1276 1277 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1278 jint target_length = target_array->length(); 1279 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1280 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1281 1282 IdealKit kit(this, false, true); 1283 #define __ kit. 1284 Node* zero = __ ConI(0); 1285 Node* one = __ ConI(1); 1286 Node* cache = __ ConI(cache_i); 1287 Node* md2 = __ ConI(md2_i); 1288 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1289 Node* targetCount = __ ConI(target_length); 1290 Node* targetCountLess1 = __ ConI(target_length - 1); 1291 Node* targetOffset = __ ConI(targetOffset_i); 1292 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1293 1294 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1295 Node* outer_loop = __ make_label(2 /* goto */); 1296 Node* return_ = __ make_label(1); 1297 1298 __ set(rtn,__ ConI(-1)); 1299 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1300 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1301 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1302 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1303 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1304 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1305 Node* tpj = __ AddI(targetOffset, __ value(j)); 1306 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1307 Node* ipj = __ AddI(__ value(i), __ value(j)); 1308 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1309 __ if_then(targ, BoolTest::ne, src2); { 1310 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1311 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1312 __ increment(i, __ AddI(__ value(j), one)); 1313 __ goto_(outer_loop); 1314 } __ end_if(); __ dead(j); 1315 }__ end_if(); __ dead(j); 1316 __ increment(i, md2); 1317 __ goto_(outer_loop); 1318 }__ end_if(); 1319 __ increment(j, one); 1320 }__ end_loop(); __ dead(j); 1321 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1322 __ goto_(return_); 1323 }__ end_if(); 1324 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1325 __ increment(i, targetCountLess1); 1326 }__ end_if(); 1327 __ increment(i, one); 1328 __ bind(outer_loop); 1329 }__ end_loop(); __ dead(i); 1330 __ bind(return_); 1331 1332 // Final sync IdealKit and GraphKit. 1333 final_sync(kit); 1334 Node* result = __ value(rtn); 1335 #undef __ 1336 C->set_has_loops(true); 1337 return result; 1338 } 1339 1340 //------------------------------inline_string_indexOf------------------------ 1341 bool LibraryCallKit::inline_string_indexOf() { 1342 Node* receiver = argument(0); 1343 Node* arg = argument(1); 1344 1345 Node* result; 1346 // Disable the use of pcmpestri until it can be guaranteed that 1347 // the load doesn't cross into the uncommited space. 1348 if (Matcher::has_match_rule(Op_StrIndexOf) && 1349 UseSSE42Intrinsics) { 1350 // Generate SSE4.2 version of indexOf 1351 // We currently only have match rules that use SSE4.2 1352 1353 receiver = null_check(receiver); 1354 arg = null_check(arg); 1355 if (stopped()) { 1356 return true; 1357 } 1358 1359 ciInstanceKlass* str_klass = env()->String_klass(); 1360 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1361 1362 // Make the merge point 1363 RegionNode* result_rgn = new (C) RegionNode(4); 1364 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT); 1365 Node* no_ctrl = NULL; 1366 1367 // Get start addr of source string 1368 Node* source = load_String_value(no_ctrl, receiver); 1369 Node* source_offset = load_String_offset(no_ctrl, receiver); 1370 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1371 1372 // Get length of source string 1373 Node* source_cnt = load_String_length(no_ctrl, receiver); 1374 1375 // Get start addr of substring 1376 Node* substr = load_String_value(no_ctrl, arg); 1377 Node* substr_offset = load_String_offset(no_ctrl, arg); 1378 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1379 1380 // Get length of source string 1381 Node* substr_cnt = load_String_length(no_ctrl, arg); 1382 1383 // Check for substr count > string count 1384 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt)); 1385 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt)); 1386 Node* if_gt = generate_slow_guard(bol, NULL); 1387 if (if_gt != NULL) { 1388 result_phi->init_req(2, intcon(-1)); 1389 result_rgn->init_req(2, if_gt); 1390 } 1391 1392 if (!stopped()) { 1393 // Check for substr count == 0 1394 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0))); 1395 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 1396 Node* if_zero = generate_slow_guard(bol, NULL); 1397 if (if_zero != NULL) { 1398 result_phi->init_req(3, intcon(0)); 1399 result_rgn->init_req(3, if_zero); 1400 } 1401 } 1402 1403 if (!stopped()) { 1404 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1405 result_phi->init_req(1, result); 1406 result_rgn->init_req(1, control()); 1407 } 1408 set_control(_gvn.transform(result_rgn)); 1409 record_for_igvn(result_rgn); 1410 result = _gvn.transform(result_phi); 1411 1412 } else { // Use LibraryCallKit::string_indexOf 1413 // don't intrinsify if argument isn't a constant string. 1414 if (!arg->is_Con()) { 1415 return false; 1416 } 1417 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1418 if (str_type == NULL) { 1419 return false; 1420 } 1421 ciInstanceKlass* klass = env()->String_klass(); 1422 ciObject* str_const = str_type->const_oop(); 1423 if (str_const == NULL || str_const->klass() != klass) { 1424 return false; 1425 } 1426 ciInstance* str = str_const->as_instance(); 1427 assert(str != NULL, "must be instance"); 1428 1429 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1430 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1431 1432 int o; 1433 int c; 1434 if (java_lang_String::has_offset_field()) { 1435 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1436 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1437 } else { 1438 o = 0; 1439 c = pat->length(); 1440 } 1441 1442 // constant strings have no offset and count == length which 1443 // simplifies the resulting code somewhat so lets optimize for that. 1444 if (o != 0 || c != pat->length()) { 1445 return false; 1446 } 1447 1448 receiver = null_check(receiver, T_OBJECT); 1449 // NOTE: No null check on the argument is needed since it's a constant String oop. 1450 if (stopped()) { 1451 return true; 1452 } 1453 1454 // The null string as a pattern always returns 0 (match at beginning of string) 1455 if (c == 0) { 1456 set_result(intcon(0)); 1457 return true; 1458 } 1459 1460 // Generate default indexOf 1461 jchar lastChar = pat->char_at(o + (c - 1)); 1462 int cache = 0; 1463 int i; 1464 for (i = 0; i < c - 1; i++) { 1465 assert(i < pat->length(), "out of range"); 1466 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1467 } 1468 1469 int md2 = c; 1470 for (i = 0; i < c - 1; i++) { 1471 assert(i < pat->length(), "out of range"); 1472 if (pat->char_at(o + i) == lastChar) { 1473 md2 = (c - 1) - i; 1474 } 1475 } 1476 1477 result = string_indexOf(receiver, pat, o, cache, md2); 1478 } 1479 set_result(result); 1480 return true; 1481 } 1482 1483 //--------------------------round_double_node-------------------------------- 1484 // Round a double node if necessary. 1485 Node* LibraryCallKit::round_double_node(Node* n) { 1486 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1487 n = _gvn.transform(new (C) RoundDoubleNode(0, n)); 1488 return n; 1489 } 1490 1491 //------------------------------inline_math----------------------------------- 1492 // public static double Math.abs(double) 1493 // public static double Math.sqrt(double) 1494 // public static double Math.log(double) 1495 // public static double Math.log10(double) 1496 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1497 Node* arg = round_double_node(argument(0)); 1498 Node* n; 1499 switch (id) { 1500 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break; 1501 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break; 1502 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break; 1503 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break; 1504 default: fatal_unexpected_iid(id); break; 1505 } 1506 set_result(_gvn.transform(n)); 1507 return true; 1508 } 1509 1510 //------------------------------inline_trig---------------------------------- 1511 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1512 // argument reduction which will turn into a fast/slow diamond. 1513 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1514 Node* arg = round_double_node(argument(0)); 1515 Node* n = NULL; 1516 1517 switch (id) { 1518 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break; 1519 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break; 1520 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break; 1521 default: fatal_unexpected_iid(id); break; 1522 } 1523 n = _gvn.transform(n); 1524 1525 // Rounding required? Check for argument reduction! 1526 if (Matcher::strict_fp_requires_explicit_rounding) { 1527 static const double pi_4 = 0.7853981633974483; 1528 static const double neg_pi_4 = -0.7853981633974483; 1529 // pi/2 in 80-bit extended precision 1530 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1531 // -pi/2 in 80-bit extended precision 1532 // 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}; 1533 // Cutoff value for using this argument reduction technique 1534 //static const double pi_2_minus_epsilon = 1.564660403643354; 1535 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1536 1537 // Pseudocode for sin: 1538 // if (x <= Math.PI / 4.0) { 1539 // if (x >= -Math.PI / 4.0) return fsin(x); 1540 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1541 // } else { 1542 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1543 // } 1544 // return StrictMath.sin(x); 1545 1546 // Pseudocode for cos: 1547 // if (x <= Math.PI / 4.0) { 1548 // if (x >= -Math.PI / 4.0) return fcos(x); 1549 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1550 // } else { 1551 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1552 // } 1553 // return StrictMath.cos(x); 1554 1555 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1556 // requires a special machine instruction to load it. Instead we'll try 1557 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1558 // probably do the math inside the SIN encoding. 1559 1560 // Make the merge point 1561 RegionNode* r = new (C) RegionNode(3); 1562 Node* phi = new (C) PhiNode(r, Type::DOUBLE); 1563 1564 // Flatten arg so we need only 1 test 1565 Node *abs = _gvn.transform(new (C) AbsDNode(arg)); 1566 // Node for PI/4 constant 1567 Node *pi4 = makecon(TypeD::make(pi_4)); 1568 // Check PI/4 : abs(arg) 1569 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs)); 1570 // Check: If PI/4 < abs(arg) then go slow 1571 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt )); 1572 // Branch either way 1573 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1574 set_control(opt_iff(r,iff)); 1575 1576 // Set fast path result 1577 phi->init_req(2, n); 1578 1579 // Slow path - non-blocking leaf call 1580 Node* call = NULL; 1581 switch (id) { 1582 case vmIntrinsics::_dsin: 1583 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1584 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1585 "Sin", NULL, arg, top()); 1586 break; 1587 case vmIntrinsics::_dcos: 1588 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1589 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1590 "Cos", NULL, arg, top()); 1591 break; 1592 case vmIntrinsics::_dtan: 1593 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1594 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1595 "Tan", NULL, arg, top()); 1596 break; 1597 } 1598 assert(control()->in(0) == call, ""); 1599 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1600 r->init_req(1, control()); 1601 phi->init_req(1, slow_result); 1602 1603 // Post-merge 1604 set_control(_gvn.transform(r)); 1605 record_for_igvn(r); 1606 n = _gvn.transform(phi); 1607 1608 C->set_has_split_ifs(true); // Has chance for split-if optimization 1609 } 1610 set_result(n); 1611 return true; 1612 } 1613 1614 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1615 //------------------- 1616 //result=(result.isNaN())? funcAddr():result; 1617 // Check: If isNaN() by checking result!=result? then either trap 1618 // or go to runtime 1619 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result)); 1620 // Build the boolean node 1621 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq)); 1622 1623 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1624 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1625 // The pow or exp intrinsic returned a NaN, which requires a call 1626 // to the runtime. Recompile with the runtime call. 1627 uncommon_trap(Deoptimization::Reason_intrinsic, 1628 Deoptimization::Action_make_not_entrant); 1629 } 1630 set_result(result); 1631 } else { 1632 // If this inlining ever returned NaN in the past, we compile a call 1633 // to the runtime to properly handle corner cases 1634 1635 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1636 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff)); 1637 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff)); 1638 1639 if (!if_slow->is_top()) { 1640 RegionNode* result_region = new (C) RegionNode(3); 1641 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE); 1642 1643 result_region->init_req(1, if_fast); 1644 result_val->init_req(1, result); 1645 1646 set_control(if_slow); 1647 1648 const TypePtr* no_memory_effects = NULL; 1649 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1650 no_memory_effects, 1651 x, top(), y, y ? top() : NULL); 1652 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0)); 1653 #ifdef ASSERT 1654 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1)); 1655 assert(value_top == top(), "second value must be top"); 1656 #endif 1657 1658 result_region->init_req(2, control()); 1659 result_val->init_req(2, value); 1660 set_result(result_region, result_val); 1661 } else { 1662 set_result(result); 1663 } 1664 } 1665 } 1666 1667 //------------------------------inline_exp------------------------------------- 1668 // Inline exp instructions, if possible. The Intel hardware only misses 1669 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1670 bool LibraryCallKit::inline_exp() { 1671 Node* arg = round_double_node(argument(0)); 1672 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg)); 1673 1674 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1675 1676 C->set_has_split_ifs(true); // Has chance for split-if optimization 1677 return true; 1678 } 1679 1680 //------------------------------inline_pow------------------------------------- 1681 // Inline power instructions, if possible. 1682 bool LibraryCallKit::inline_pow() { 1683 // Pseudocode for pow 1684 // if (x <= 0.0) { 1685 // long longy = (long)y; 1686 // if ((double)longy == y) { // if y is long 1687 // if (y + 1 == y) longy = 0; // huge number: even 1688 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1689 // } else { 1690 // result = NaN; 1691 // } 1692 // } else { 1693 // result = DPow(x,y); 1694 // } 1695 // if (result != result)? { 1696 // result = uncommon_trap() or runtime_call(); 1697 // } 1698 // return result; 1699 1700 Node* x = round_double_node(argument(0)); 1701 Node* y = round_double_node(argument(2)); 1702 1703 Node* result = NULL; 1704 1705 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1706 // Short form: skip the fancy tests and just check for NaN result. 1707 result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1708 } else { 1709 // If this inlining ever returned NaN in the past, include all 1710 // checks + call to the runtime. 1711 1712 // Set the merge point for If node with condition of (x <= 0.0) 1713 // There are four possible paths to region node and phi node 1714 RegionNode *r = new (C) RegionNode(4); 1715 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1716 1717 // Build the first if node: if (x <= 0.0) 1718 // Node for 0 constant 1719 Node *zeronode = makecon(TypeD::ZERO); 1720 // Check x:0 1721 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode)); 1722 // Check: If (x<=0) then go complex path 1723 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le )); 1724 // Branch either way 1725 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1726 // Fast path taken; set region slot 3 1727 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1)); 1728 r->init_req(3,fast_taken); // Capture fast-control 1729 1730 // Fast path not-taken, i.e. slow path 1731 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1)); 1732 1733 // Set fast path result 1734 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1735 phi->init_req(3, fast_result); 1736 1737 // Complex path 1738 // Build the second if node (if y is long) 1739 // Node for (long)y 1740 Node *longy = _gvn.transform(new (C) ConvD2LNode(y)); 1741 // Node for (double)((long) y) 1742 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy)); 1743 // Check (double)((long) y) : y 1744 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y)); 1745 // Check if (y isn't long) then go to slow path 1746 1747 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne )); 1748 // Branch either way 1749 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1750 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2)); 1751 1752 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2)); 1753 1754 // Calculate DPow(abs(x), y)*(1 & (long)y) 1755 // Node for constant 1 1756 Node *conone = longcon(1); 1757 // 1& (long)y 1758 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy)); 1759 1760 // A huge number is always even. Detect a huge number by checking 1761 // if y + 1 == y and set integer to be tested for parity to 0. 1762 // Required for corner case: 1763 // (long)9.223372036854776E18 = max_jlong 1764 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1765 // max_jlong is odd but 9.223372036854776E18 is even 1766 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1)))); 1767 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y)); 1768 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq )); 1769 Node* correctedsign = NULL; 1770 if (ConditionalMoveLimit != 0) { 1771 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1772 } else { 1773 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1774 RegionNode *r = new (C) RegionNode(3); 1775 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 1776 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1))); 1777 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1))); 1778 phi->init_req(1, signnode); 1779 phi->init_req(2, longcon(0)); 1780 correctedsign = _gvn.transform(phi); 1781 ylong_path = _gvn.transform(r); 1782 record_for_igvn(r); 1783 } 1784 1785 // zero node 1786 Node *conzero = longcon(0); 1787 // Check (1&(long)y)==0? 1788 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero)); 1789 // Check if (1&(long)y)!=0?, if so the result is negative 1790 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne )); 1791 // abs(x) 1792 Node *absx=_gvn.transform(new (C) AbsDNode(x)); 1793 // abs(x)^y 1794 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y)); 1795 // -abs(x)^y 1796 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy)); 1797 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1798 Node *signresult = NULL; 1799 if (ConditionalMoveLimit != 0) { 1800 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1801 } else { 1802 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1803 RegionNode *r = new (C) RegionNode(3); 1804 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1805 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven))); 1806 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven))); 1807 phi->init_req(1, absxpowy); 1808 phi->init_req(2, negabsxpowy); 1809 signresult = _gvn.transform(phi); 1810 ylong_path = _gvn.transform(r); 1811 record_for_igvn(r); 1812 } 1813 // Set complex path fast result 1814 r->init_req(2, ylong_path); 1815 phi->init_req(2, signresult); 1816 1817 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1818 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1819 r->init_req(1,slow_path); 1820 phi->init_req(1,slow_result); 1821 1822 // Post merge 1823 set_control(_gvn.transform(r)); 1824 record_for_igvn(r); 1825 result = _gvn.transform(phi); 1826 } 1827 1828 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1829 1830 C->set_has_split_ifs(true); // Has chance for split-if optimization 1831 return true; 1832 } 1833 1834 //------------------------------runtime_math----------------------------- 1835 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1836 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1837 "must be (DD)D or (D)D type"); 1838 1839 // Inputs 1840 Node* a = round_double_node(argument(0)); 1841 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1842 1843 const TypePtr* no_memory_effects = NULL; 1844 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1845 no_memory_effects, 1846 a, top(), b, b ? top() : NULL); 1847 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0)); 1848 #ifdef ASSERT 1849 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1)); 1850 assert(value_top == top(), "second value must be top"); 1851 #endif 1852 1853 set_result(value); 1854 return true; 1855 } 1856 1857 //------------------------------inline_math_native----------------------------- 1858 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1859 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1860 switch (id) { 1861 // These intrinsics are not properly supported on all hardware 1862 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1863 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1864 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1865 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1866 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1867 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1868 1869 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1870 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1871 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1872 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1873 1874 // These intrinsics are supported on all hardware 1875 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false; 1876 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1877 1878 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1879 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1880 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1881 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1882 #undef FN_PTR 1883 1884 // These intrinsics are not yet correctly implemented 1885 case vmIntrinsics::_datan2: 1886 return false; 1887 1888 default: 1889 fatal_unexpected_iid(id); 1890 return false; 1891 } 1892 } 1893 1894 static bool is_simple_name(Node* n) { 1895 return (n->req() == 1 // constant 1896 || (n->is_Type() && n->as_Type()->type()->singleton()) 1897 || n->is_Proj() // parameter or return value 1898 || n->is_Phi() // local of some sort 1899 ); 1900 } 1901 1902 //----------------------------inline_min_max----------------------------------- 1903 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1904 set_result(generate_min_max(id, argument(0), argument(1))); 1905 return true; 1906 } 1907 1908 Node* 1909 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1910 // These are the candidate return value: 1911 Node* xvalue = x0; 1912 Node* yvalue = y0; 1913 1914 if (xvalue == yvalue) { 1915 return xvalue; 1916 } 1917 1918 bool want_max = (id == vmIntrinsics::_max); 1919 1920 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1921 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1922 if (txvalue == NULL || tyvalue == NULL) return top(); 1923 // This is not really necessary, but it is consistent with a 1924 // hypothetical MaxINode::Value method: 1925 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1926 1927 // %%% This folding logic should (ideally) be in a different place. 1928 // Some should be inside IfNode, and there to be a more reliable 1929 // transformation of ?: style patterns into cmoves. We also want 1930 // more powerful optimizations around cmove and min/max. 1931 1932 // Try to find a dominating comparison of these guys. 1933 // It can simplify the index computation for Arrays.copyOf 1934 // and similar uses of System.arraycopy. 1935 // First, compute the normalized version of CmpI(x, y). 1936 int cmp_op = Op_CmpI; 1937 Node* xkey = xvalue; 1938 Node* ykey = yvalue; 1939 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey)); 1940 if (ideal_cmpxy->is_Cmp()) { 1941 // E.g., if we have CmpI(length - offset, count), 1942 // it might idealize to CmpI(length, count + offset) 1943 cmp_op = ideal_cmpxy->Opcode(); 1944 xkey = ideal_cmpxy->in(1); 1945 ykey = ideal_cmpxy->in(2); 1946 } 1947 1948 // Start by locating any relevant comparisons. 1949 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1950 Node* cmpxy = NULL; 1951 Node* cmpyx = NULL; 1952 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1953 Node* cmp = start_from->fast_out(k); 1954 if (cmp->outcnt() > 0 && // must have prior uses 1955 cmp->in(0) == NULL && // must be context-independent 1956 cmp->Opcode() == cmp_op) { // right kind of compare 1957 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1958 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1959 } 1960 } 1961 1962 const int NCMPS = 2; 1963 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1964 int cmpn; 1965 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1966 if (cmps[cmpn] != NULL) break; // find a result 1967 } 1968 if (cmpn < NCMPS) { 1969 // Look for a dominating test that tells us the min and max. 1970 int depth = 0; // Limit search depth for speed 1971 Node* dom = control(); 1972 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1973 if (++depth >= 100) break; 1974 Node* ifproj = dom; 1975 if (!ifproj->is_Proj()) continue; 1976 Node* iff = ifproj->in(0); 1977 if (!iff->is_If()) continue; 1978 Node* bol = iff->in(1); 1979 if (!bol->is_Bool()) continue; 1980 Node* cmp = bol->in(1); 1981 if (cmp == NULL) continue; 1982 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1983 if (cmps[cmpn] == cmp) break; 1984 if (cmpn == NCMPS) continue; 1985 BoolTest::mask btest = bol->as_Bool()->_test._test; 1986 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1987 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1988 // At this point, we know that 'x btest y' is true. 1989 switch (btest) { 1990 case BoolTest::eq: 1991 // They are proven equal, so we can collapse the min/max. 1992 // Either value is the answer. Choose the simpler. 1993 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1994 return yvalue; 1995 return xvalue; 1996 case BoolTest::lt: // x < y 1997 case BoolTest::le: // x <= y 1998 return (want_max ? yvalue : xvalue); 1999 case BoolTest::gt: // x > y 2000 case BoolTest::ge: // x >= y 2001 return (want_max ? xvalue : yvalue); 2002 } 2003 } 2004 } 2005 2006 // We failed to find a dominating test. 2007 // Let's pick a test that might GVN with prior tests. 2008 Node* best_bol = NULL; 2009 BoolTest::mask best_btest = BoolTest::illegal; 2010 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2011 Node* cmp = cmps[cmpn]; 2012 if (cmp == NULL) continue; 2013 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2014 Node* bol = cmp->fast_out(j); 2015 if (!bol->is_Bool()) continue; 2016 BoolTest::mask btest = bol->as_Bool()->_test._test; 2017 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2018 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2019 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2020 best_bol = bol->as_Bool(); 2021 best_btest = btest; 2022 } 2023 } 2024 } 2025 2026 Node* answer_if_true = NULL; 2027 Node* answer_if_false = NULL; 2028 switch (best_btest) { 2029 default: 2030 if (cmpxy == NULL) 2031 cmpxy = ideal_cmpxy; 2032 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt)); 2033 // and fall through: 2034 case BoolTest::lt: // x < y 2035 case BoolTest::le: // x <= y 2036 answer_if_true = (want_max ? yvalue : xvalue); 2037 answer_if_false = (want_max ? xvalue : yvalue); 2038 break; 2039 case BoolTest::gt: // x > y 2040 case BoolTest::ge: // x >= y 2041 answer_if_true = (want_max ? xvalue : yvalue); 2042 answer_if_false = (want_max ? yvalue : xvalue); 2043 break; 2044 } 2045 2046 jint hi, lo; 2047 if (want_max) { 2048 // We can sharpen the minimum. 2049 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2050 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2051 } else { 2052 // We can sharpen the maximum. 2053 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2054 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2055 } 2056 2057 // Use a flow-free graph structure, to avoid creating excess control edges 2058 // which could hinder other optimizations. 2059 // Since Math.min/max is often used with arraycopy, we want 2060 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2061 Node* cmov = CMoveNode::make(C, NULL, best_bol, 2062 answer_if_false, answer_if_true, 2063 TypeInt::make(lo, hi, widen)); 2064 2065 return _gvn.transform(cmov); 2066 2067 /* 2068 // This is not as desirable as it may seem, since Min and Max 2069 // nodes do not have a full set of optimizations. 2070 // And they would interfere, anyway, with 'if' optimizations 2071 // and with CMoveI canonical forms. 2072 switch (id) { 2073 case vmIntrinsics::_min: 2074 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2075 case vmIntrinsics::_max: 2076 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2077 default: 2078 ShouldNotReachHere(); 2079 } 2080 */ 2081 } 2082 2083 inline int 2084 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2085 const TypePtr* base_type = TypePtr::NULL_PTR; 2086 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2087 if (base_type == NULL) { 2088 // Unknown type. 2089 return Type::AnyPtr; 2090 } else if (base_type == TypePtr::NULL_PTR) { 2091 // Since this is a NULL+long form, we have to switch to a rawptr. 2092 base = _gvn.transform(new (C) CastX2PNode(offset)); 2093 offset = MakeConX(0); 2094 return Type::RawPtr; 2095 } else if (base_type->base() == Type::RawPtr) { 2096 return Type::RawPtr; 2097 } else if (base_type->isa_oopptr()) { 2098 // Base is never null => always a heap address. 2099 if (base_type->ptr() == TypePtr::NotNull) { 2100 return Type::OopPtr; 2101 } 2102 // Offset is small => always a heap address. 2103 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2104 if (offset_type != NULL && 2105 base_type->offset() == 0 && // (should always be?) 2106 offset_type->_lo >= 0 && 2107 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2108 return Type::OopPtr; 2109 } 2110 // Otherwise, it might either be oop+off or NULL+addr. 2111 return Type::AnyPtr; 2112 } else { 2113 // No information: 2114 return Type::AnyPtr; 2115 } 2116 } 2117 2118 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2119 int kind = classify_unsafe_addr(base, offset); 2120 if (kind == Type::RawPtr) { 2121 return basic_plus_adr(top(), base, offset); 2122 } else { 2123 return basic_plus_adr(base, offset); 2124 } 2125 } 2126 2127 //--------------------------inline_number_methods----------------------------- 2128 // inline int Integer.numberOfLeadingZeros(int) 2129 // inline int Long.numberOfLeadingZeros(long) 2130 // 2131 // inline int Integer.numberOfTrailingZeros(int) 2132 // inline int Long.numberOfTrailingZeros(long) 2133 // 2134 // inline int Integer.bitCount(int) 2135 // inline int Long.bitCount(long) 2136 // 2137 // inline char Character.reverseBytes(char) 2138 // inline short Short.reverseBytes(short) 2139 // inline int Integer.reverseBytes(int) 2140 // inline long Long.reverseBytes(long) 2141 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2142 Node* arg = argument(0); 2143 Node* n; 2144 switch (id) { 2145 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break; 2146 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break; 2147 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break; 2148 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break; 2149 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break; 2150 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break; 2151 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break; 2152 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break; 2153 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break; 2154 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break; 2155 default: fatal_unexpected_iid(id); break; 2156 } 2157 set_result(_gvn.transform(n)); 2158 return true; 2159 } 2160 2161 //----------------------------inline_unsafe_access---------------------------- 2162 2163 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2164 2165 // Helper that guards and inserts a pre-barrier. 2166 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2167 Node* pre_val, bool need_mem_bar) { 2168 // We could be accessing the referent field of a reference object. If so, when G1 2169 // is enabled, we need to log the value in the referent field in an SATB buffer. 2170 // This routine performs some compile time filters and generates suitable 2171 // runtime filters that guard the pre-barrier code. 2172 // Also add memory barrier for non volatile load from the referent field 2173 // to prevent commoning of loads across safepoint. 2174 if (!UseG1GC && !need_mem_bar) 2175 return; 2176 2177 // Some compile time checks. 2178 2179 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2180 const TypeX* otype = offset->find_intptr_t_type(); 2181 if (otype != NULL && otype->is_con() && 2182 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2183 // Constant offset but not the reference_offset so just return 2184 return; 2185 } 2186 2187 // We only need to generate the runtime guards for instances. 2188 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2189 if (btype != NULL) { 2190 if (btype->isa_aryptr()) { 2191 // Array type so nothing to do 2192 return; 2193 } 2194 2195 const TypeInstPtr* itype = btype->isa_instptr(); 2196 if (itype != NULL) { 2197 // Can the klass of base_oop be statically determined to be 2198 // _not_ a sub-class of Reference and _not_ Object? 2199 ciKlass* klass = itype->klass(); 2200 if ( klass->is_loaded() && 2201 !klass->is_subtype_of(env()->Reference_klass()) && 2202 !env()->Object_klass()->is_subtype_of(klass)) { 2203 return; 2204 } 2205 } 2206 } 2207 2208 // The compile time filters did not reject base_oop/offset so 2209 // we need to generate the following runtime filters 2210 // 2211 // if (offset == java_lang_ref_Reference::_reference_offset) { 2212 // if (instance_of(base, java.lang.ref.Reference)) { 2213 // pre_barrier(_, pre_val, ...); 2214 // } 2215 // } 2216 2217 float likely = PROB_LIKELY( 0.999); 2218 float unlikely = PROB_UNLIKELY(0.999); 2219 2220 IdealKit ideal(this); 2221 #define __ ideal. 2222 2223 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2224 2225 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2226 // Update graphKit memory and control from IdealKit. 2227 sync_kit(ideal); 2228 2229 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2230 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2231 2232 // Update IdealKit memory and control from graphKit. 2233 __ sync_kit(this); 2234 2235 Node* one = __ ConI(1); 2236 // is_instof == 0 if base_oop == NULL 2237 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2238 2239 // Update graphKit from IdeakKit. 2240 sync_kit(ideal); 2241 2242 // Use the pre-barrier to record the value in the referent field 2243 pre_barrier(false /* do_load */, 2244 __ ctrl(), 2245 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2246 pre_val /* pre_val */, 2247 T_OBJECT); 2248 if (need_mem_bar) { 2249 // Add memory barrier to prevent commoning reads from this field 2250 // across safepoint since GC can change its value. 2251 insert_mem_bar(Op_MemBarCPUOrder); 2252 } 2253 // Update IdealKit from graphKit. 2254 __ sync_kit(this); 2255 2256 } __ end_if(); // _ref_type != ref_none 2257 } __ end_if(); // offset == referent_offset 2258 2259 // Final sync IdealKit and GraphKit. 2260 final_sync(ideal); 2261 #undef __ 2262 } 2263 2264 2265 // Interpret Unsafe.fieldOffset cookies correctly: 2266 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2267 2268 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2269 // Attempt to infer a sharper value type from the offset and base type. 2270 ciKlass* sharpened_klass = NULL; 2271 2272 // See if it is an instance field, with an object type. 2273 if (alias_type->field() != NULL) { 2274 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2275 if (alias_type->field()->type()->is_klass()) { 2276 sharpened_klass = alias_type->field()->type()->as_klass(); 2277 } 2278 } 2279 2280 // See if it is a narrow oop array. 2281 if (adr_type->isa_aryptr()) { 2282 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2283 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2284 if (elem_type != NULL) { 2285 sharpened_klass = elem_type->klass(); 2286 } 2287 } 2288 } 2289 2290 // The sharpened class might be unloaded if there is no class loader 2291 // contraint in place. 2292 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2293 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2294 2295 #ifndef PRODUCT 2296 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2297 tty->print(" from base type: "); adr_type->dump(); 2298 tty->print(" sharpened value: "); tjp->dump(); 2299 } 2300 #endif 2301 // Sharpen the value type. 2302 return tjp; 2303 } 2304 return NULL; 2305 } 2306 2307 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2308 if (callee()->is_static()) return false; // caller must have the capability! 2309 2310 #ifndef PRODUCT 2311 { 2312 ResourceMark rm; 2313 // Check the signatures. 2314 ciSignature* sig = callee()->signature(); 2315 #ifdef ASSERT 2316 if (!is_store) { 2317 // Object getObject(Object base, int/long offset), etc. 2318 BasicType rtype = sig->return_type()->basic_type(); 2319 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2320 rtype = T_ADDRESS; // it is really a C void* 2321 assert(rtype == type, "getter must return the expected value"); 2322 if (!is_native_ptr) { 2323 assert(sig->count() == 2, "oop getter has 2 arguments"); 2324 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2325 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2326 } else { 2327 assert(sig->count() == 1, "native getter has 1 argument"); 2328 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2329 } 2330 } else { 2331 // void putObject(Object base, int/long offset, Object x), etc. 2332 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2333 if (!is_native_ptr) { 2334 assert(sig->count() == 3, "oop putter has 3 arguments"); 2335 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2336 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2337 } else { 2338 assert(sig->count() == 2, "native putter has 2 arguments"); 2339 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2340 } 2341 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2342 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2343 vtype = T_ADDRESS; // it is really a C void* 2344 assert(vtype == type, "putter must accept the expected value"); 2345 } 2346 #endif // ASSERT 2347 } 2348 #endif //PRODUCT 2349 2350 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2351 2352 Node* receiver = argument(0); // type: oop 2353 2354 // Build address expression. See the code in inline_unsafe_prefetch. 2355 Node* adr; 2356 Node* heap_base_oop = top(); 2357 Node* offset = top(); 2358 Node* val; 2359 2360 if (!is_native_ptr) { 2361 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2362 Node* base = argument(1); // type: oop 2363 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2364 offset = argument(2); // type: long 2365 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2366 // to be plain byte offsets, which are also the same as those accepted 2367 // by oopDesc::field_base. 2368 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2369 "fieldOffset must be byte-scaled"); 2370 // 32-bit machines ignore the high half! 2371 offset = ConvL2X(offset); 2372 adr = make_unsafe_address(base, offset); 2373 heap_base_oop = base; 2374 val = is_store ? argument(4) : NULL; 2375 } else { 2376 Node* ptr = argument(1); // type: long 2377 ptr = ConvL2X(ptr); // adjust Java long to machine word 2378 adr = make_unsafe_address(NULL, ptr); 2379 val = is_store ? argument(3) : NULL; 2380 } 2381 2382 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2383 2384 // First guess at the value type. 2385 const Type *value_type = Type::get_const_basic_type(type); 2386 2387 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2388 // there was not enough information to nail it down. 2389 Compile::AliasType* alias_type = C->alias_type(adr_type); 2390 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2391 2392 // We will need memory barriers unless we can determine a unique 2393 // alias category for this reference. (Note: If for some reason 2394 // the barriers get omitted and the unsafe reference begins to "pollute" 2395 // the alias analysis of the rest of the graph, either Compile::can_alias 2396 // or Compile::must_alias will throw a diagnostic assert.) 2397 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2398 2399 // If we are reading the value of the referent field of a Reference 2400 // object (either by using Unsafe directly or through reflection) 2401 // then, if G1 is enabled, we need to record the referent in an 2402 // SATB log buffer using the pre-barrier mechanism. 2403 // Also we need to add memory barrier to prevent commoning reads 2404 // from this field across safepoint since GC can change its value. 2405 bool need_read_barrier = !is_native_ptr && !is_store && 2406 offset != top() && heap_base_oop != top(); 2407 2408 if (!is_store && type == T_OBJECT) { 2409 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2410 if (tjp != NULL) { 2411 value_type = tjp; 2412 } 2413 } 2414 2415 receiver = null_check(receiver); 2416 if (stopped()) { 2417 return true; 2418 } 2419 // Heap pointers get a null-check from the interpreter, 2420 // as a courtesy. However, this is not guaranteed by Unsafe, 2421 // and it is not possible to fully distinguish unintended nulls 2422 // from intended ones in this API. 2423 2424 if (is_volatile) { 2425 // We need to emit leading and trailing CPU membars (see below) in 2426 // addition to memory membars when is_volatile. This is a little 2427 // too strong, but avoids the need to insert per-alias-type 2428 // volatile membars (for stores; compare Parse::do_put_xxx), which 2429 // we cannot do effectively here because we probably only have a 2430 // rough approximation of type. 2431 need_mem_bar = true; 2432 // For Stores, place a memory ordering barrier now. 2433 if (is_store) 2434 insert_mem_bar(Op_MemBarRelease); 2435 } 2436 2437 // Memory barrier to prevent normal and 'unsafe' accesses from 2438 // bypassing each other. Happens after null checks, so the 2439 // exception paths do not take memory state from the memory barrier, 2440 // so there's no problems making a strong assert about mixing users 2441 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2442 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2443 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2444 2445 if (!is_store) { 2446 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2447 // load value 2448 switch (type) { 2449 case T_BOOLEAN: 2450 case T_CHAR: 2451 case T_BYTE: 2452 case T_SHORT: 2453 case T_INT: 2454 case T_LONG: 2455 case T_FLOAT: 2456 case T_DOUBLE: 2457 break; 2458 case T_OBJECT: 2459 if (need_read_barrier) { 2460 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2461 } 2462 break; 2463 case T_ADDRESS: 2464 // Cast to an int type. 2465 p = _gvn.transform(new (C) CastP2XNode(NULL, p)); 2466 p = ConvX2L(p); 2467 break; 2468 default: 2469 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2470 break; 2471 } 2472 // The load node has the control of the preceding MemBarCPUOrder. All 2473 // following nodes will have the control of the MemBarCPUOrder inserted at 2474 // the end of this method. So, pushing the load onto the stack at a later 2475 // point is fine. 2476 set_result(p); 2477 } else { 2478 // place effect of store into memory 2479 switch (type) { 2480 case T_DOUBLE: 2481 val = dstore_rounding(val); 2482 break; 2483 case T_ADDRESS: 2484 // Repackage the long as a pointer. 2485 val = ConvL2X(val); 2486 val = _gvn.transform(new (C) CastX2PNode(val)); 2487 break; 2488 } 2489 2490 if (type != T_OBJECT ) { 2491 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2492 } else { 2493 // Possibly an oop being stored to Java heap or native memory 2494 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2495 // oop to Java heap. 2496 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2497 } else { 2498 // We can't tell at compile time if we are storing in the Java heap or outside 2499 // of it. So we need to emit code to conditionally do the proper type of 2500 // store. 2501 2502 IdealKit ideal(this); 2503 #define __ ideal. 2504 // QQQ who knows what probability is here?? 2505 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2506 // Sync IdealKit and graphKit. 2507 sync_kit(ideal); 2508 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2509 // Update IdealKit memory. 2510 __ sync_kit(this); 2511 } __ else_(); { 2512 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile); 2513 } __ end_if(); 2514 // Final sync IdealKit and GraphKit. 2515 final_sync(ideal); 2516 #undef __ 2517 } 2518 } 2519 } 2520 2521 if (is_volatile) { 2522 if (!is_store) 2523 insert_mem_bar(Op_MemBarAcquire); 2524 else 2525 insert_mem_bar(Op_MemBarVolatile); 2526 } 2527 2528 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2529 2530 return true; 2531 } 2532 2533 //----------------------------inline_unsafe_prefetch---------------------------- 2534 2535 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2536 #ifndef PRODUCT 2537 { 2538 ResourceMark rm; 2539 // Check the signatures. 2540 ciSignature* sig = callee()->signature(); 2541 #ifdef ASSERT 2542 // Object getObject(Object base, int/long offset), etc. 2543 BasicType rtype = sig->return_type()->basic_type(); 2544 if (!is_native_ptr) { 2545 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2546 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2547 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2548 } else { 2549 assert(sig->count() == 1, "native prefetch has 1 argument"); 2550 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2551 } 2552 #endif // ASSERT 2553 } 2554 #endif // !PRODUCT 2555 2556 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2557 2558 const int idx = is_static ? 0 : 1; 2559 if (!is_static) { 2560 null_check_receiver(); 2561 if (stopped()) { 2562 return true; 2563 } 2564 } 2565 2566 // Build address expression. See the code in inline_unsafe_access. 2567 Node *adr; 2568 if (!is_native_ptr) { 2569 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2570 Node* base = argument(idx + 0); // type: oop 2571 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2572 Node* offset = argument(idx + 1); // type: long 2573 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2574 // to be plain byte offsets, which are also the same as those accepted 2575 // by oopDesc::field_base. 2576 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2577 "fieldOffset must be byte-scaled"); 2578 // 32-bit machines ignore the high half! 2579 offset = ConvL2X(offset); 2580 adr = make_unsafe_address(base, offset); 2581 } else { 2582 Node* ptr = argument(idx + 0); // type: long 2583 ptr = ConvL2X(ptr); // adjust Java long to machine word 2584 adr = make_unsafe_address(NULL, ptr); 2585 } 2586 2587 // Generate the read or write prefetch 2588 Node *prefetch; 2589 if (is_store) { 2590 prefetch = new (C) PrefetchWriteNode(i_o(), adr); 2591 } else { 2592 prefetch = new (C) PrefetchReadNode(i_o(), adr); 2593 } 2594 prefetch->init_req(0, control()); 2595 set_i_o(_gvn.transform(prefetch)); 2596 2597 return true; 2598 } 2599 2600 //----------------------------inline_unsafe_load_store---------------------------- 2601 // This method serves a couple of different customers (depending on LoadStoreKind): 2602 // 2603 // LS_cmpxchg: 2604 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2605 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2606 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2607 // 2608 // LS_xadd: 2609 // public int getAndAddInt( Object o, long offset, int delta) 2610 // public long getAndAddLong(Object o, long offset, long delta) 2611 // 2612 // LS_xchg: 2613 // int getAndSet(Object o, long offset, int newValue) 2614 // long getAndSet(Object o, long offset, long newValue) 2615 // Object getAndSet(Object o, long offset, Object newValue) 2616 // 2617 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2618 // This basic scheme here is the same as inline_unsafe_access, but 2619 // differs in enough details that combining them would make the code 2620 // overly confusing. (This is a true fact! I originally combined 2621 // them, but even I was confused by it!) As much code/comments as 2622 // possible are retained from inline_unsafe_access though to make 2623 // the correspondences clearer. - dl 2624 2625 if (callee()->is_static()) return false; // caller must have the capability! 2626 2627 #ifndef PRODUCT 2628 BasicType rtype; 2629 { 2630 ResourceMark rm; 2631 // Check the signatures. 2632 ciSignature* sig = callee()->signature(); 2633 rtype = sig->return_type()->basic_type(); 2634 if (kind == LS_xadd || kind == LS_xchg) { 2635 // Check the signatures. 2636 #ifdef ASSERT 2637 assert(rtype == type, "get and set must return the expected type"); 2638 assert(sig->count() == 3, "get and set has 3 arguments"); 2639 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2640 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2641 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2642 #endif // ASSERT 2643 } else if (kind == LS_cmpxchg) { 2644 // Check the signatures. 2645 #ifdef ASSERT 2646 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2647 assert(sig->count() == 4, "CAS has 4 arguments"); 2648 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2649 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2650 #endif // ASSERT 2651 } else { 2652 ShouldNotReachHere(); 2653 } 2654 } 2655 #endif //PRODUCT 2656 2657 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2658 2659 // Get arguments: 2660 Node* receiver = NULL; 2661 Node* base = NULL; 2662 Node* offset = NULL; 2663 Node* oldval = NULL; 2664 Node* newval = NULL; 2665 if (kind == LS_cmpxchg) { 2666 const bool two_slot_type = type2size[type] == 2; 2667 receiver = argument(0); // type: oop 2668 base = argument(1); // type: oop 2669 offset = argument(2); // type: long 2670 oldval = argument(4); // type: oop, int, or long 2671 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2672 } else if (kind == LS_xadd || kind == LS_xchg){ 2673 receiver = argument(0); // type: oop 2674 base = argument(1); // type: oop 2675 offset = argument(2); // type: long 2676 oldval = NULL; 2677 newval = argument(4); // type: oop, int, or long 2678 } 2679 2680 // Null check receiver. 2681 receiver = null_check(receiver); 2682 if (stopped()) { 2683 return true; 2684 } 2685 2686 // Build field offset expression. 2687 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2688 // to be plain byte offsets, which are also the same as those accepted 2689 // by oopDesc::field_base. 2690 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2691 // 32-bit machines ignore the high half of long offsets 2692 offset = ConvL2X(offset); 2693 Node* adr = make_unsafe_address(base, offset); 2694 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2695 2696 // For CAS, unlike inline_unsafe_access, there seems no point in 2697 // trying to refine types. Just use the coarse types here. 2698 const Type *value_type = Type::get_const_basic_type(type); 2699 Compile::AliasType* alias_type = C->alias_type(adr_type); 2700 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2701 2702 if (kind == LS_xchg && type == T_OBJECT) { 2703 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2704 if (tjp != NULL) { 2705 value_type = tjp; 2706 } 2707 } 2708 2709 int alias_idx = C->get_alias_index(adr_type); 2710 2711 // Memory-model-wise, a LoadStore acts like a little synchronized 2712 // block, so needs barriers on each side. These don't translate 2713 // into actual barriers on most machines, but we still need rest of 2714 // compiler to respect ordering. 2715 2716 insert_mem_bar(Op_MemBarRelease); 2717 insert_mem_bar(Op_MemBarCPUOrder); 2718 2719 // 4984716: MemBars must be inserted before this 2720 // memory node in order to avoid a false 2721 // dependency which will confuse the scheduler. 2722 Node *mem = memory(alias_idx); 2723 2724 // For now, we handle only those cases that actually exist: ints, 2725 // longs, and Object. Adding others should be straightforward. 2726 Node* load_store; 2727 switch(type) { 2728 case T_INT: 2729 if (kind == LS_xadd) { 2730 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type)); 2731 } else if (kind == LS_xchg) { 2732 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type)); 2733 } else if (kind == LS_cmpxchg) { 2734 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2735 } else { 2736 ShouldNotReachHere(); 2737 } 2738 break; 2739 case T_LONG: 2740 if (kind == LS_xadd) { 2741 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2742 } else if (kind == LS_xchg) { 2743 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2744 } else if (kind == LS_cmpxchg) { 2745 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2746 } else { 2747 ShouldNotReachHere(); 2748 } 2749 break; 2750 case T_OBJECT: 2751 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2752 // could be delayed during Parse (for example, in adjust_map_after_if()). 2753 // Execute transformation here to avoid barrier generation in such case. 2754 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2755 newval = _gvn.makecon(TypePtr::NULL_PTR); 2756 2757 // Reference stores need a store barrier. 2758 pre_barrier(true /* do_load*/, 2759 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2760 NULL /* pre_val*/, 2761 T_OBJECT); 2762 #ifdef _LP64 2763 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2764 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2765 if (kind == LS_xchg) { 2766 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr, 2767 newval_enc, adr_type, value_type->make_narrowoop())); 2768 } else { 2769 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2770 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2771 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr, 2772 newval_enc, oldval_enc)); 2773 } 2774 } else 2775 #endif 2776 { 2777 if (kind == LS_xchg) { 2778 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2779 } else { 2780 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2781 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2782 } 2783 } 2784 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2785 break; 2786 default: 2787 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2788 break; 2789 } 2790 2791 // SCMemProjNodes represent the memory state of a LoadStore. Their 2792 // main role is to prevent LoadStore nodes from being optimized away 2793 // when their results aren't used. 2794 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store)); 2795 set_memory(proj, alias_idx); 2796 2797 // Add the trailing membar surrounding the access 2798 insert_mem_bar(Op_MemBarCPUOrder); 2799 insert_mem_bar(Op_MemBarAcquire); 2800 2801 #ifdef _LP64 2802 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) { 2803 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type())); 2804 } 2805 #endif 2806 2807 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2808 set_result(load_store); 2809 return true; 2810 } 2811 2812 //----------------------------inline_unsafe_ordered_store---------------------- 2813 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2814 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 2815 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 2816 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2817 // This is another variant of inline_unsafe_access, differing in 2818 // that it always issues store-store ("release") barrier and ensures 2819 // store-atomicity (which only matters for "long"). 2820 2821 if (callee()->is_static()) return false; // caller must have the capability! 2822 2823 #ifndef PRODUCT 2824 { 2825 ResourceMark rm; 2826 // Check the signatures. 2827 ciSignature* sig = callee()->signature(); 2828 #ifdef ASSERT 2829 BasicType rtype = sig->return_type()->basic_type(); 2830 assert(rtype == T_VOID, "must return void"); 2831 assert(sig->count() == 3, "has 3 arguments"); 2832 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2833 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2834 #endif // ASSERT 2835 } 2836 #endif //PRODUCT 2837 2838 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2839 2840 // Get arguments: 2841 Node* receiver = argument(0); // type: oop 2842 Node* base = argument(1); // type: oop 2843 Node* offset = argument(2); // type: long 2844 Node* val = argument(4); // type: oop, int, or long 2845 2846 // Null check receiver. 2847 receiver = null_check(receiver); 2848 if (stopped()) { 2849 return true; 2850 } 2851 2852 // Build field offset expression. 2853 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2854 // 32-bit machines ignore the high half of long offsets 2855 offset = ConvL2X(offset); 2856 Node* adr = make_unsafe_address(base, offset); 2857 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2858 const Type *value_type = Type::get_const_basic_type(type); 2859 Compile::AliasType* alias_type = C->alias_type(adr_type); 2860 2861 insert_mem_bar(Op_MemBarRelease); 2862 insert_mem_bar(Op_MemBarCPUOrder); 2863 // Ensure that the store is atomic for longs: 2864 const bool require_atomic_access = true; 2865 Node* store; 2866 if (type == T_OBJECT) // reference stores need a store barrier. 2867 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2868 else { 2869 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2870 } 2871 insert_mem_bar(Op_MemBarCPUOrder); 2872 return true; 2873 } 2874 2875 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2876 // Regardless of form, don't allow previous ld/st to move down, 2877 // then issue acquire, release, or volatile mem_bar. 2878 insert_mem_bar(Op_MemBarCPUOrder); 2879 switch(id) { 2880 case vmIntrinsics::_loadFence: 2881 insert_mem_bar(Op_MemBarAcquire); 2882 return true; 2883 case vmIntrinsics::_storeFence: 2884 insert_mem_bar(Op_MemBarRelease); 2885 return true; 2886 case vmIntrinsics::_fullFence: 2887 insert_mem_bar(Op_MemBarVolatile); 2888 return true; 2889 default: 2890 fatal_unexpected_iid(id); 2891 return false; 2892 } 2893 } 2894 2895 //----------------------------inline_unsafe_allocate--------------------------- 2896 // public native Object sun.mics.Unsafe.allocateInstance(Class<?> cls); 2897 bool LibraryCallKit::inline_unsafe_allocate() { 2898 if (callee()->is_static()) return false; // caller must have the capability! 2899 2900 null_check_receiver(); // null-check, then ignore 2901 Node* cls = null_check(argument(1)); 2902 if (stopped()) return true; 2903 2904 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2905 kls = null_check(kls); 2906 if (stopped()) return true; // argument was like int.class 2907 2908 // Note: The argument might still be an illegal value like 2909 // Serializable.class or Object[].class. The runtime will handle it. 2910 // But we must make an explicit check for initialization. 2911 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 2912 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 2913 // can generate code to load it as unsigned byte. 2914 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN); 2915 Node* bits = intcon(InstanceKlass::fully_initialized); 2916 Node* test = _gvn.transform(new (C) SubINode(inst, bits)); 2917 // The 'test' is non-zero if we need to take a slow path. 2918 2919 Node* obj = new_instance(kls, test); 2920 set_result(obj); 2921 return true; 2922 } 2923 2924 #ifdef TRACE_HAVE_INTRINSICS 2925 /* 2926 * oop -> myklass 2927 * myklass->trace_id |= USED 2928 * return myklass->trace_id & ~0x3 2929 */ 2930 bool LibraryCallKit::inline_native_classID() { 2931 null_check_receiver(); // null-check, then ignore 2932 Node* cls = null_check(argument(1), T_OBJECT); 2933 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2934 kls = null_check(kls, T_OBJECT); 2935 ByteSize offset = TRACE_ID_OFFSET; 2936 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 2937 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG); 2938 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 2939 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits)); 2940 Node* clsused = longcon(0x01l); // set the class bit 2941 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused)); 2942 2943 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 2944 store_to_memory(control(), insp, orl, T_LONG, adr_type); 2945 set_result(andl); 2946 return true; 2947 } 2948 2949 bool LibraryCallKit::inline_native_threadID() { 2950 Node* tls_ptr = NULL; 2951 Node* cur_thr = generate_current_thread(tls_ptr); 2952 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2953 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 2954 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 2955 2956 Node* threadid = NULL; 2957 size_t thread_id_size = OSThread::thread_id_size(); 2958 if (thread_id_size == (size_t) BytesPerLong) { 2959 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG)); 2960 } else if (thread_id_size == (size_t) BytesPerInt) { 2961 threadid = make_load(control(), p, TypeInt::INT, T_INT); 2962 } else { 2963 ShouldNotReachHere(); 2964 } 2965 set_result(threadid); 2966 return true; 2967 } 2968 #endif 2969 2970 //------------------------inline_native_time_funcs-------------- 2971 // inline code for System.currentTimeMillis() and System.nanoTime() 2972 // these have the same type and signature 2973 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 2974 const TypeFunc* tf = OptoRuntime::void_long_Type(); 2975 const TypePtr* no_memory_effects = NULL; 2976 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2977 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0)); 2978 #ifdef ASSERT 2979 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1)); 2980 assert(value_top == top(), "second value must be top"); 2981 #endif 2982 set_result(value); 2983 return true; 2984 } 2985 2986 //------------------------inline_native_currentThread------------------ 2987 bool LibraryCallKit::inline_native_currentThread() { 2988 Node* junk = NULL; 2989 set_result(generate_current_thread(junk)); 2990 return true; 2991 } 2992 2993 //------------------------inline_native_isInterrupted------------------ 2994 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 2995 bool LibraryCallKit::inline_native_isInterrupted() { 2996 // Add a fast path to t.isInterrupted(clear_int): 2997 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 2998 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 2999 // So, in the common case that the interrupt bit is false, 3000 // we avoid making a call into the VM. Even if the interrupt bit 3001 // is true, if the clear_int argument is false, we avoid the VM call. 3002 // However, if the receiver is not currentThread, we must call the VM, 3003 // because there must be some locking done around the operation. 3004 3005 // We only go to the fast case code if we pass two guards. 3006 // Paths which do not pass are accumulated in the slow_region. 3007 3008 enum { 3009 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3010 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3011 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3012 PATH_LIMIT 3013 }; 3014 3015 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3016 // out of the function. 3017 insert_mem_bar(Op_MemBarCPUOrder); 3018 3019 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT); 3020 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL); 3021 3022 RegionNode* slow_region = new (C) RegionNode(1); 3023 record_for_igvn(slow_region); 3024 3025 // (a) Receiving thread must be the current thread. 3026 Node* rec_thr = argument(0); 3027 Node* tls_ptr = NULL; 3028 Node* cur_thr = generate_current_thread(tls_ptr); 3029 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr)); 3030 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne)); 3031 3032 generate_slow_guard(bol_thr, slow_region); 3033 3034 // (b) Interrupt bit on TLS must be false. 3035 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3036 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 3037 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3038 3039 // Set the control input on the field _interrupted read to prevent it floating up. 3040 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 3041 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0))); 3042 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne)); 3043 3044 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3045 3046 // First fast path: if (!TLS._interrupted) return false; 3047 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit)); 3048 result_rgn->init_req(no_int_result_path, false_bit); 3049 result_val->init_req(no_int_result_path, intcon(0)); 3050 3051 // drop through to next case 3052 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit))); 3053 3054 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3055 Node* clr_arg = argument(1); 3056 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0))); 3057 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne)); 3058 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3059 3060 // Second fast path: ... else if (!clear_int) return true; 3061 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg)); 3062 result_rgn->init_req(no_clear_result_path, false_arg); 3063 result_val->init_req(no_clear_result_path, intcon(1)); 3064 3065 // drop through to next case 3066 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg))); 3067 3068 // (d) Otherwise, go to the slow path. 3069 slow_region->add_req(control()); 3070 set_control( _gvn.transform(slow_region)); 3071 3072 if (stopped()) { 3073 // There is no slow path. 3074 result_rgn->init_req(slow_result_path, top()); 3075 result_val->init_req(slow_result_path, top()); 3076 } else { 3077 // non-virtual because it is a private non-static 3078 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3079 3080 Node* slow_val = set_results_for_java_call(slow_call); 3081 // this->control() comes from set_results_for_java_call 3082 3083 Node* fast_io = slow_call->in(TypeFunc::I_O); 3084 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3085 3086 // These two phis are pre-filled with copies of of the fast IO and Memory 3087 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3088 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3089 3090 result_rgn->init_req(slow_result_path, control()); 3091 result_io ->init_req(slow_result_path, i_o()); 3092 result_mem->init_req(slow_result_path, reset_memory()); 3093 result_val->init_req(slow_result_path, slow_val); 3094 3095 set_all_memory(_gvn.transform(result_mem)); 3096 set_i_o( _gvn.transform(result_io)); 3097 } 3098 3099 C->set_has_split_ifs(true); // Has chance for split-if optimization 3100 set_result(result_rgn, result_val); 3101 return true; 3102 } 3103 3104 //---------------------------load_mirror_from_klass---------------------------- 3105 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3106 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3107 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3108 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 3109 } 3110 3111 //-----------------------load_klass_from_mirror_common------------------------- 3112 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3113 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3114 // and branch to the given path on the region. 3115 // If never_see_null, take an uncommon trap on null, so we can optimistically 3116 // compile for the non-null case. 3117 // If the region is NULL, force never_see_null = true. 3118 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3119 bool never_see_null, 3120 RegionNode* region, 3121 int null_path, 3122 int offset) { 3123 if (region == NULL) never_see_null = true; 3124 Node* p = basic_plus_adr(mirror, offset); 3125 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3126 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3127 Node* null_ctl = top(); 3128 kls = null_check_oop(kls, &null_ctl, never_see_null); 3129 if (region != NULL) { 3130 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3131 region->init_req(null_path, null_ctl); 3132 } else { 3133 assert(null_ctl == top(), "no loose ends"); 3134 } 3135 return kls; 3136 } 3137 3138 //--------------------(inline_native_Class_query helpers)--------------------- 3139 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3140 // Fall through if (mods & mask) == bits, take the guard otherwise. 3141 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3142 // Branch around if the given klass has the given modifier bit set. 3143 // Like generate_guard, adds a new path onto the region. 3144 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3145 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 3146 Node* mask = intcon(modifier_mask); 3147 Node* bits = intcon(modifier_bits); 3148 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask)); 3149 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits)); 3150 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 3151 return generate_fair_guard(bol, region); 3152 } 3153 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3154 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3155 } 3156 3157 //-------------------------inline_native_Class_query------------------- 3158 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3159 const Type* return_type = TypeInt::BOOL; 3160 Node* prim_return_value = top(); // what happens if it's a primitive class? 3161 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3162 bool expect_prim = false; // most of these guys expect to work on refs 3163 3164 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3165 3166 Node* mirror = argument(0); 3167 Node* obj = top(); 3168 3169 switch (id) { 3170 case vmIntrinsics::_isInstance: 3171 // nothing is an instance of a primitive type 3172 prim_return_value = intcon(0); 3173 obj = argument(1); 3174 break; 3175 case vmIntrinsics::_getModifiers: 3176 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3177 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3178 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3179 break; 3180 case vmIntrinsics::_isInterface: 3181 prim_return_value = intcon(0); 3182 break; 3183 case vmIntrinsics::_isArray: 3184 prim_return_value = intcon(0); 3185 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3186 break; 3187 case vmIntrinsics::_isPrimitive: 3188 prim_return_value = intcon(1); 3189 expect_prim = true; // obviously 3190 break; 3191 case vmIntrinsics::_getSuperclass: 3192 prim_return_value = null(); 3193 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3194 break; 3195 case vmIntrinsics::_getComponentType: 3196 prim_return_value = null(); 3197 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3198 break; 3199 case vmIntrinsics::_getClassAccessFlags: 3200 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3201 return_type = TypeInt::INT; // not bool! 6297094 3202 break; 3203 default: 3204 fatal_unexpected_iid(id); 3205 break; 3206 } 3207 3208 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3209 if (mirror_con == NULL) return false; // cannot happen? 3210 3211 #ifndef PRODUCT 3212 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 3213 ciType* k = mirror_con->java_mirror_type(); 3214 if (k) { 3215 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3216 k->print_name(); 3217 tty->cr(); 3218 } 3219 } 3220 #endif 3221 3222 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3223 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3224 record_for_igvn(region); 3225 PhiNode* phi = new (C) PhiNode(region, return_type); 3226 3227 // The mirror will never be null of Reflection.getClassAccessFlags, however 3228 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3229 // if it is. See bug 4774291. 3230 3231 // For Reflection.getClassAccessFlags(), the null check occurs in 3232 // the wrong place; see inline_unsafe_access(), above, for a similar 3233 // situation. 3234 mirror = null_check(mirror); 3235 // If mirror or obj is dead, only null-path is taken. 3236 if (stopped()) return true; 3237 3238 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3239 3240 // Now load the mirror's klass metaobject, and null-check it. 3241 // Side-effects region with the control path if the klass is null. 3242 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3243 // If kls is null, we have a primitive mirror. 3244 phi->init_req(_prim_path, prim_return_value); 3245 if (stopped()) { set_result(region, phi); return true; } 3246 3247 Node* p; // handy temp 3248 Node* null_ctl; 3249 3250 // Now that we have the non-null klass, we can perform the real query. 3251 // For constant classes, the query will constant-fold in LoadNode::Value. 3252 Node* query_value = top(); 3253 switch (id) { 3254 case vmIntrinsics::_isInstance: 3255 // nothing is an instance of a primitive type 3256 query_value = gen_instanceof(obj, kls); 3257 break; 3258 3259 case vmIntrinsics::_getModifiers: 3260 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3261 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3262 break; 3263 3264 case vmIntrinsics::_isInterface: 3265 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3266 if (generate_interface_guard(kls, region) != NULL) 3267 // A guard was added. If the guard is taken, it was an interface. 3268 phi->add_req(intcon(1)); 3269 // If we fall through, it's a plain class. 3270 query_value = intcon(0); 3271 break; 3272 3273 case vmIntrinsics::_isArray: 3274 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3275 if (generate_array_guard(kls, region) != NULL) 3276 // A guard was added. If the guard is taken, it was an array. 3277 phi->add_req(intcon(1)); 3278 // If we fall through, it's a plain class. 3279 query_value = intcon(0); 3280 break; 3281 3282 case vmIntrinsics::_isPrimitive: 3283 query_value = intcon(0); // "normal" path produces false 3284 break; 3285 3286 case vmIntrinsics::_getSuperclass: 3287 // The rules here are somewhat unfortunate, but we can still do better 3288 // with random logic than with a JNI call. 3289 // Interfaces store null or Object as _super, but must report null. 3290 // Arrays store an intermediate super as _super, but must report Object. 3291 // Other types can report the actual _super. 3292 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3293 if (generate_interface_guard(kls, region) != NULL) 3294 // A guard was added. If the guard is taken, it was an interface. 3295 phi->add_req(null()); 3296 if (generate_array_guard(kls, region) != NULL) 3297 // A guard was added. If the guard is taken, it was an array. 3298 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3299 // If we fall through, it's a plain class. Get its _super. 3300 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3301 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3302 null_ctl = top(); 3303 kls = null_check_oop(kls, &null_ctl); 3304 if (null_ctl != top()) { 3305 // If the guard is taken, Object.superClass is null (both klass and mirror). 3306 region->add_req(null_ctl); 3307 phi ->add_req(null()); 3308 } 3309 if (!stopped()) { 3310 query_value = load_mirror_from_klass(kls); 3311 } 3312 break; 3313 3314 case vmIntrinsics::_getComponentType: 3315 if (generate_array_guard(kls, region) != NULL) { 3316 // Be sure to pin the oop load to the guard edge just created: 3317 Node* is_array_ctrl = region->in(region->req()-1); 3318 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); 3319 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 3320 phi->add_req(cmo); 3321 } 3322 query_value = null(); // non-array case is null 3323 break; 3324 3325 case vmIntrinsics::_getClassAccessFlags: 3326 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3327 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3328 break; 3329 3330 default: 3331 fatal_unexpected_iid(id); 3332 break; 3333 } 3334 3335 // Fall-through is the normal case of a query to a real class. 3336 phi->init_req(1, query_value); 3337 region->init_req(1, control()); 3338 3339 C->set_has_split_ifs(true); // Has chance for split-if optimization 3340 set_result(region, phi); 3341 return true; 3342 } 3343 3344 //--------------------------inline_native_subtype_check------------------------ 3345 // This intrinsic takes the JNI calls out of the heart of 3346 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3347 bool LibraryCallKit::inline_native_subtype_check() { 3348 // Pull both arguments off the stack. 3349 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3350 args[0] = argument(0); 3351 args[1] = argument(1); 3352 Node* klasses[2]; // corresponding Klasses: superk, subk 3353 klasses[0] = klasses[1] = top(); 3354 3355 enum { 3356 // A full decision tree on {superc is prim, subc is prim}: 3357 _prim_0_path = 1, // {P,N} => false 3358 // {P,P} & superc!=subc => false 3359 _prim_same_path, // {P,P} & superc==subc => true 3360 _prim_1_path, // {N,P} => false 3361 _ref_subtype_path, // {N,N} & subtype check wins => true 3362 _both_ref_path, // {N,N} & subtype check loses => false 3363 PATH_LIMIT 3364 }; 3365 3366 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3367 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 3368 record_for_igvn(region); 3369 3370 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3371 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3372 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3373 3374 // First null-check both mirrors and load each mirror's klass metaobject. 3375 int which_arg; 3376 for (which_arg = 0; which_arg <= 1; which_arg++) { 3377 Node* arg = args[which_arg]; 3378 arg = null_check(arg); 3379 if (stopped()) break; 3380 args[which_arg] = arg; 3381 3382 Node* p = basic_plus_adr(arg, class_klass_offset); 3383 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3384 klasses[which_arg] = _gvn.transform(kls); 3385 } 3386 3387 // Having loaded both klasses, test each for null. 3388 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3389 for (which_arg = 0; which_arg <= 1; which_arg++) { 3390 Node* kls = klasses[which_arg]; 3391 Node* null_ctl = top(); 3392 kls = null_check_oop(kls, &null_ctl, never_see_null); 3393 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3394 region->init_req(prim_path, null_ctl); 3395 if (stopped()) break; 3396 klasses[which_arg] = kls; 3397 } 3398 3399 if (!stopped()) { 3400 // now we have two reference types, in klasses[0..1] 3401 Node* subk = klasses[1]; // the argument to isAssignableFrom 3402 Node* superk = klasses[0]; // the receiver 3403 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3404 // now we have a successful reference subtype check 3405 region->set_req(_ref_subtype_path, control()); 3406 } 3407 3408 // If both operands are primitive (both klasses null), then 3409 // we must return true when they are identical primitives. 3410 // It is convenient to test this after the first null klass check. 3411 set_control(region->in(_prim_0_path)); // go back to first null check 3412 if (!stopped()) { 3413 // Since superc is primitive, make a guard for the superc==subc case. 3414 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1])); 3415 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq)); 3416 generate_guard(bol_eq, region, PROB_FAIR); 3417 if (region->req() == PATH_LIMIT+1) { 3418 // A guard was added. If the added guard is taken, superc==subc. 3419 region->swap_edges(PATH_LIMIT, _prim_same_path); 3420 region->del_req(PATH_LIMIT); 3421 } 3422 region->set_req(_prim_0_path, control()); // Not equal after all. 3423 } 3424 3425 // these are the only paths that produce 'true': 3426 phi->set_req(_prim_same_path, intcon(1)); 3427 phi->set_req(_ref_subtype_path, intcon(1)); 3428 3429 // pull together the cases: 3430 assert(region->req() == PATH_LIMIT, "sane region"); 3431 for (uint i = 1; i < region->req(); i++) { 3432 Node* ctl = region->in(i); 3433 if (ctl == NULL || ctl == top()) { 3434 region->set_req(i, top()); 3435 phi ->set_req(i, top()); 3436 } else if (phi->in(i) == NULL) { 3437 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3438 } 3439 } 3440 3441 set_control(_gvn.transform(region)); 3442 set_result(_gvn.transform(phi)); 3443 return true; 3444 } 3445 3446 //---------------------generate_array_guard_common------------------------ 3447 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3448 bool obj_array, bool not_array) { 3449 // If obj_array/non_array==false/false: 3450 // Branch around if the given klass is in fact an array (either obj or prim). 3451 // If obj_array/non_array==false/true: 3452 // Branch around if the given klass is not an array klass of any kind. 3453 // If obj_array/non_array==true/true: 3454 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3455 // If obj_array/non_array==true/false: 3456 // Branch around if the kls is an oop array (Object[] or subtype) 3457 // 3458 // Like generate_guard, adds a new path onto the region. 3459 jint layout_con = 0; 3460 Node* layout_val = get_layout_helper(kls, layout_con); 3461 if (layout_val == NULL) { 3462 bool query = (obj_array 3463 ? Klass::layout_helper_is_objArray(layout_con) 3464 : Klass::layout_helper_is_array(layout_con)); 3465 if (query == not_array) { 3466 return NULL; // never a branch 3467 } else { // always a branch 3468 Node* always_branch = control(); 3469 if (region != NULL) 3470 region->add_req(always_branch); 3471 set_control(top()); 3472 return always_branch; 3473 } 3474 } 3475 // Now test the correct condition. 3476 jint nval = (obj_array 3477 ? ((jint)Klass::_lh_array_tag_type_value 3478 << Klass::_lh_array_tag_shift) 3479 : Klass::_lh_neutral_value); 3480 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval))); 3481 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3482 // invert the test if we are looking for a non-array 3483 if (not_array) btest = BoolTest(btest).negate(); 3484 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest)); 3485 return generate_fair_guard(bol, region); 3486 } 3487 3488 3489 //-----------------------inline_native_newArray-------------------------- 3490 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3491 bool LibraryCallKit::inline_native_newArray() { 3492 Node* mirror = argument(0); 3493 Node* count_val = argument(1); 3494 3495 mirror = null_check(mirror); 3496 // If mirror or obj is dead, only null-path is taken. 3497 if (stopped()) return true; 3498 3499 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3500 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3501 PhiNode* result_val = new(C) PhiNode(result_reg, 3502 TypeInstPtr::NOTNULL); 3503 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3504 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3505 TypePtr::BOTTOM); 3506 3507 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3508 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3509 result_reg, _slow_path); 3510 Node* normal_ctl = control(); 3511 Node* no_array_ctl = result_reg->in(_slow_path); 3512 3513 // Generate code for the slow case. We make a call to newArray(). 3514 set_control(no_array_ctl); 3515 if (!stopped()) { 3516 // Either the input type is void.class, or else the 3517 // array klass has not yet been cached. Either the 3518 // ensuing call will throw an exception, or else it 3519 // will cache the array klass for next time. 3520 PreserveJVMState pjvms(this); 3521 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3522 Node* slow_result = set_results_for_java_call(slow_call); 3523 // this->control() comes from set_results_for_java_call 3524 result_reg->set_req(_slow_path, control()); 3525 result_val->set_req(_slow_path, slow_result); 3526 result_io ->set_req(_slow_path, i_o()); 3527 result_mem->set_req(_slow_path, reset_memory()); 3528 } 3529 3530 set_control(normal_ctl); 3531 if (!stopped()) { 3532 // Normal case: The array type has been cached in the java.lang.Class. 3533 // The following call works fine even if the array type is polymorphic. 3534 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3535 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3536 result_reg->init_req(_normal_path, control()); 3537 result_val->init_req(_normal_path, obj); 3538 result_io ->init_req(_normal_path, i_o()); 3539 result_mem->init_req(_normal_path, reset_memory()); 3540 } 3541 3542 // Return the combined state. 3543 set_i_o( _gvn.transform(result_io) ); 3544 set_all_memory( _gvn.transform(result_mem)); 3545 3546 C->set_has_split_ifs(true); // Has chance for split-if optimization 3547 set_result(result_reg, result_val); 3548 return true; 3549 } 3550 3551 //----------------------inline_native_getLength-------------------------- 3552 // public static native int java.lang.reflect.Array.getLength(Object array); 3553 bool LibraryCallKit::inline_native_getLength() { 3554 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3555 3556 Node* array = null_check(argument(0)); 3557 // If array is dead, only null-path is taken. 3558 if (stopped()) return true; 3559 3560 // Deoptimize if it is a non-array. 3561 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3562 3563 if (non_array != NULL) { 3564 PreserveJVMState pjvms(this); 3565 set_control(non_array); 3566 uncommon_trap(Deoptimization::Reason_intrinsic, 3567 Deoptimization::Action_maybe_recompile); 3568 } 3569 3570 // If control is dead, only non-array-path is taken. 3571 if (stopped()) return true; 3572 3573 // The works fine even if the array type is polymorphic. 3574 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3575 Node* result = load_array_length(array); 3576 3577 C->set_has_split_ifs(true); // Has chance for split-if optimization 3578 set_result(result); 3579 return true; 3580 } 3581 3582 //------------------------inline_array_copyOf---------------------------- 3583 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3584 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3585 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3586 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3587 3588 // Get the arguments. 3589 Node* original = argument(0); 3590 Node* start = is_copyOfRange? argument(1): intcon(0); 3591 Node* end = is_copyOfRange? argument(2): argument(1); 3592 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3593 3594 Node* newcopy; 3595 3596 // Set the original stack and the reexecute bit for the interpreter to reexecute 3597 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3598 { PreserveReexecuteState preexecs(this); 3599 jvms()->set_should_reexecute(true); 3600 3601 array_type_mirror = null_check(array_type_mirror); 3602 original = null_check(original); 3603 3604 // Check if a null path was taken unconditionally. 3605 if (stopped()) return true; 3606 3607 Node* orig_length = load_array_length(original); 3608 3609 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3610 klass_node = null_check(klass_node); 3611 3612 RegionNode* bailout = new (C) RegionNode(1); 3613 record_for_igvn(bailout); 3614 3615 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3616 // Bail out if that is so. 3617 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3618 if (not_objArray != NULL) { 3619 // Improve the klass node's type from the new optimistic assumption: 3620 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3621 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3622 Node* cast = new (C) CastPPNode(klass_node, akls); 3623 cast->init_req(0, control()); 3624 klass_node = _gvn.transform(cast); 3625 } 3626 3627 // Bail out if either start or end is negative. 3628 generate_negative_guard(start, bailout, &start); 3629 generate_negative_guard(end, bailout, &end); 3630 3631 Node* length = end; 3632 if (_gvn.type(start) != TypeInt::ZERO) { 3633 length = _gvn.transform(new (C) SubINode(end, start)); 3634 } 3635 3636 // Bail out if length is negative. 3637 // Without this the new_array would throw 3638 // NegativeArraySizeException but IllegalArgumentException is what 3639 // should be thrown 3640 generate_negative_guard(length, bailout, &length); 3641 3642 if (bailout->req() > 1) { 3643 PreserveJVMState pjvms(this); 3644 set_control(_gvn.transform(bailout)); 3645 uncommon_trap(Deoptimization::Reason_intrinsic, 3646 Deoptimization::Action_maybe_recompile); 3647 } 3648 3649 if (!stopped()) { 3650 // How many elements will we copy from the original? 3651 // The answer is MinI(orig_length - start, length). 3652 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start)); 3653 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3654 3655 newcopy = new_array(klass_node, length, 0); // no argments to push 3656 3657 // Generate a direct call to the right arraycopy function(s). 3658 // We know the copy is disjoint but we might not know if the 3659 // oop stores need checking. 3660 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3661 // This will fail a store-check if x contains any non-nulls. 3662 bool disjoint_bases = true; 3663 // if start > orig_length then the length of the copy may be 3664 // negative. 3665 bool length_never_negative = !is_copyOfRange; 3666 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3667 original, start, newcopy, intcon(0), moved, 3668 disjoint_bases, length_never_negative); 3669 } 3670 } // original reexecute is set back here 3671 3672 C->set_has_split_ifs(true); // Has chance for split-if optimization 3673 if (!stopped()) { 3674 set_result(newcopy); 3675 } 3676 return true; 3677 } 3678 3679 3680 //----------------------generate_virtual_guard--------------------------- 3681 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3682 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3683 RegionNode* slow_region) { 3684 ciMethod* method = callee(); 3685 int vtable_index = method->vtable_index(); 3686 // Get the Method* out of the appropriate vtable entry. 3687 int entry_offset = (InstanceKlass::vtable_start_offset() + 3688 vtable_index*vtableEntry::size()) * wordSize + 3689 vtableEntry::method_offset_in_bytes(); 3690 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3691 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS); 3692 3693 // Compare the target method with the expected method (e.g., Object.hashCode). 3694 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3695 3696 Node* native_call = makecon(native_call_addr); 3697 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call)); 3698 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne)); 3699 3700 return generate_slow_guard(test_native, slow_region); 3701 } 3702 3703 //-----------------------generate_method_call---------------------------- 3704 // Use generate_method_call to make a slow-call to the real 3705 // method if the fast path fails. An alternative would be to 3706 // use a stub like OptoRuntime::slow_arraycopy_Java. 3707 // This only works for expanding the current library call, 3708 // not another intrinsic. (E.g., don't use this for making an 3709 // arraycopy call inside of the copyOf intrinsic.) 3710 CallJavaNode* 3711 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3712 // When compiling the intrinsic method itself, do not use this technique. 3713 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3714 3715 ciMethod* method = callee(); 3716 // ensure the JVMS we have will be correct for this call 3717 guarantee(method_id == method->intrinsic_id(), "must match"); 3718 3719 const TypeFunc* tf = TypeFunc::make(method); 3720 CallJavaNode* slow_call; 3721 if (is_static) { 3722 assert(!is_virtual, ""); 3723 slow_call = new(C) CallStaticJavaNode(C, tf, 3724 SharedRuntime::get_resolve_static_call_stub(), 3725 method, bci()); 3726 } else if (is_virtual) { 3727 null_check_receiver(); 3728 int vtable_index = Method::invalid_vtable_index; 3729 if (UseInlineCaches) { 3730 // Suppress the vtable call 3731 } else { 3732 // hashCode and clone are not a miranda methods, 3733 // so the vtable index is fixed. 3734 // No need to use the linkResolver to get it. 3735 vtable_index = method->vtable_index(); 3736 } 3737 slow_call = new(C) CallDynamicJavaNode(tf, 3738 SharedRuntime::get_resolve_virtual_call_stub(), 3739 method, vtable_index, bci()); 3740 } else { // neither virtual nor static: opt_virtual 3741 null_check_receiver(); 3742 slow_call = new(C) CallStaticJavaNode(C, tf, 3743 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3744 method, bci()); 3745 slow_call->set_optimized_virtual(true); 3746 } 3747 set_arguments_for_java_call(slow_call); 3748 set_edges_for_java_call(slow_call); 3749 return slow_call; 3750 } 3751 3752 3753 //------------------------------inline_native_hashcode-------------------- 3754 // Build special case code for calls to hashCode on an object. 3755 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3756 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3757 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3758 3759 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3760 3761 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3762 PhiNode* result_val = new(C) PhiNode(result_reg, 3763 TypeInt::INT); 3764 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3765 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3766 TypePtr::BOTTOM); 3767 Node* obj = NULL; 3768 if (!is_static) { 3769 // Check for hashing null object 3770 obj = null_check_receiver(); 3771 if (stopped()) return true; // unconditionally null 3772 result_reg->init_req(_null_path, top()); 3773 result_val->init_req(_null_path, top()); 3774 } else { 3775 // Do a null check, and return zero if null. 3776 // System.identityHashCode(null) == 0 3777 obj = argument(0); 3778 Node* null_ctl = top(); 3779 obj = null_check_oop(obj, &null_ctl); 3780 result_reg->init_req(_null_path, null_ctl); 3781 result_val->init_req(_null_path, _gvn.intcon(0)); 3782 } 3783 3784 // Unconditionally null? Then return right away. 3785 if (stopped()) { 3786 set_control( result_reg->in(_null_path)); 3787 if (!stopped()) 3788 set_result(result_val->in(_null_path)); 3789 return true; 3790 } 3791 3792 // After null check, get the object's klass. 3793 Node* obj_klass = load_object_klass(obj); 3794 3795 // This call may be virtual (invokevirtual) or bound (invokespecial). 3796 // For each case we generate slightly different code. 3797 3798 // We only go to the fast case code if we pass a number of guards. The 3799 // paths which do not pass are accumulated in the slow_region. 3800 RegionNode* slow_region = new (C) RegionNode(1); 3801 record_for_igvn(slow_region); 3802 3803 // If this is a virtual call, we generate a funny guard. We pull out 3804 // the vtable entry corresponding to hashCode() from the target object. 3805 // If the target method which we are calling happens to be the native 3806 // Object hashCode() method, we pass the guard. We do not need this 3807 // guard for non-virtual calls -- the caller is known to be the native 3808 // Object hashCode(). 3809 if (is_virtual) { 3810 generate_virtual_guard(obj_klass, slow_region); 3811 } 3812 3813 // Get the header out of the object, use LoadMarkNode when available 3814 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3815 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type()); 3816 3817 // Test the header to see if it is unlocked. 3818 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3819 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask)); 3820 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3821 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val)); 3822 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne)); 3823 3824 generate_slow_guard(test_unlocked, slow_region); 3825 3826 // Get the hash value and check to see that it has been properly assigned. 3827 // We depend on hash_mask being at most 32 bits and avoid the use of 3828 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3829 // vm: see markOop.hpp. 3830 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3831 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3832 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift)); 3833 // This hack lets the hash bits live anywhere in the mark object now, as long 3834 // as the shift drops the relevant bits into the low 32 bits. Note that 3835 // Java spec says that HashCode is an int so there's no point in capturing 3836 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3837 hshifted_header = ConvX2I(hshifted_header); 3838 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask)); 3839 3840 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3841 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val)); 3842 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq)); 3843 3844 generate_slow_guard(test_assigned, slow_region); 3845 3846 Node* init_mem = reset_memory(); 3847 // fill in the rest of the null path: 3848 result_io ->init_req(_null_path, i_o()); 3849 result_mem->init_req(_null_path, init_mem); 3850 3851 result_val->init_req(_fast_path, hash_val); 3852 result_reg->init_req(_fast_path, control()); 3853 result_io ->init_req(_fast_path, i_o()); 3854 result_mem->init_req(_fast_path, init_mem); 3855 3856 // Generate code for the slow case. We make a call to hashCode(). 3857 set_control(_gvn.transform(slow_region)); 3858 if (!stopped()) { 3859 // No need for PreserveJVMState, because we're using up the present state. 3860 set_all_memory(init_mem); 3861 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 3862 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3863 Node* slow_result = set_results_for_java_call(slow_call); 3864 // this->control() comes from set_results_for_java_call 3865 result_reg->init_req(_slow_path, control()); 3866 result_val->init_req(_slow_path, slow_result); 3867 result_io ->set_req(_slow_path, i_o()); 3868 result_mem ->set_req(_slow_path, reset_memory()); 3869 } 3870 3871 // Return the combined state. 3872 set_i_o( _gvn.transform(result_io) ); 3873 set_all_memory( _gvn.transform(result_mem)); 3874 3875 set_result(result_reg, result_val); 3876 return true; 3877 } 3878 3879 //---------------------------inline_native_getClass---------------------------- 3880 // public final native Class<?> java.lang.Object.getClass(); 3881 // 3882 // Build special case code for calls to getClass on an object. 3883 bool LibraryCallKit::inline_native_getClass() { 3884 Node* obj = null_check_receiver(); 3885 if (stopped()) return true; 3886 set_result(load_mirror_from_klass(load_object_klass(obj))); 3887 return true; 3888 } 3889 3890 //-----------------inline_native_Reflection_getCallerClass--------------------- 3891 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 3892 // 3893 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3894 // 3895 // NOTE: This code must perform the same logic as JVM_GetCallerClass 3896 // in that it must skip particular security frames and checks for 3897 // caller sensitive methods. 3898 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3899 #ifndef PRODUCT 3900 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3901 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3902 } 3903 #endif 3904 3905 if (!jvms()->has_method()) { 3906 #ifndef PRODUCT 3907 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3908 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3909 } 3910 #endif 3911 return false; 3912 } 3913 3914 // Walk back up the JVM state to find the caller at the required 3915 // depth. 3916 JVMState* caller_jvms = jvms(); 3917 3918 // Cf. JVM_GetCallerClass 3919 // NOTE: Start the loop at depth 1 because the current JVM state does 3920 // not include the Reflection.getCallerClass() frame. 3921 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 3922 ciMethod* m = caller_jvms->method(); 3923 switch (n) { 3924 case 0: 3925 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 3926 break; 3927 case 1: 3928 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 3929 if (!m->caller_sensitive()) { 3930 #ifndef PRODUCT 3931 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3932 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 3933 } 3934 #endif 3935 return false; // bail-out; let JVM_GetCallerClass do the work 3936 } 3937 break; 3938 default: 3939 if (!m->is_ignored_by_security_stack_walk()) { 3940 // We have reached the desired frame; return the holder class. 3941 // Acquire method holder as java.lang.Class and push as constant. 3942 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 3943 ciInstance* caller_mirror = caller_klass->java_mirror(); 3944 set_result(makecon(TypeInstPtr::make(caller_mirror))); 3945 3946 #ifndef PRODUCT 3947 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3948 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()); 3949 tty->print_cr(" JVM state at this point:"); 3950 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 3951 ciMethod* m = jvms()->of_depth(i)->method(); 3952 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 3953 } 3954 } 3955 #endif 3956 return true; 3957 } 3958 break; 3959 } 3960 } 3961 3962 #ifndef PRODUCT 3963 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3964 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 3965 tty->print_cr(" JVM state at this point:"); 3966 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 3967 ciMethod* m = jvms()->of_depth(i)->method(); 3968 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 3969 } 3970 } 3971 #endif 3972 3973 return false; // bail-out; let JVM_GetCallerClass do the work 3974 } 3975 3976 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 3977 Node* arg = argument(0); 3978 Node* result; 3979 3980 switch (id) { 3981 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break; 3982 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break; 3983 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break; 3984 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break; 3985 3986 case vmIntrinsics::_doubleToLongBits: { 3987 // two paths (plus control) merge in a wood 3988 RegionNode *r = new (C) RegionNode(3); 3989 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 3990 3991 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg)); 3992 // Build the boolean node 3993 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 3994 3995 // Branch either way. 3996 // NaN case is less traveled, which makes all the difference. 3997 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3998 Node *opt_isnan = _gvn.transform(ifisnan); 3999 assert( opt_isnan->is_If(), "Expect an IfNode"); 4000 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4001 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4002 4003 set_control(iftrue); 4004 4005 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4006 Node *slow_result = longcon(nan_bits); // return NaN 4007 phi->init_req(1, _gvn.transform( slow_result )); 4008 r->init_req(1, iftrue); 4009 4010 // Else fall through 4011 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4012 set_control(iffalse); 4013 4014 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg))); 4015 r->init_req(2, iffalse); 4016 4017 // Post merge 4018 set_control(_gvn.transform(r)); 4019 record_for_igvn(r); 4020 4021 C->set_has_split_ifs(true); // Has chance for split-if optimization 4022 result = phi; 4023 assert(result->bottom_type()->isa_long(), "must be"); 4024 break; 4025 } 4026 4027 case vmIntrinsics::_floatToIntBits: { 4028 // two paths (plus control) merge in a wood 4029 RegionNode *r = new (C) RegionNode(3); 4030 Node *phi = new (C) PhiNode(r, TypeInt::INT); 4031 4032 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg)); 4033 // Build the boolean node 4034 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4035 4036 // Branch either way. 4037 // NaN case is less traveled, which makes all the difference. 4038 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4039 Node *opt_isnan = _gvn.transform(ifisnan); 4040 assert( opt_isnan->is_If(), "Expect an IfNode"); 4041 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4042 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4043 4044 set_control(iftrue); 4045 4046 static const jint nan_bits = 0x7fc00000; 4047 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4048 phi->init_req(1, _gvn.transform( slow_result )); 4049 r->init_req(1, iftrue); 4050 4051 // Else fall through 4052 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4053 set_control(iffalse); 4054 4055 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg))); 4056 r->init_req(2, iffalse); 4057 4058 // Post merge 4059 set_control(_gvn.transform(r)); 4060 record_for_igvn(r); 4061 4062 C->set_has_split_ifs(true); // Has chance for split-if optimization 4063 result = phi; 4064 assert(result->bottom_type()->isa_int(), "must be"); 4065 break; 4066 } 4067 4068 default: 4069 fatal_unexpected_iid(id); 4070 break; 4071 } 4072 set_result(_gvn.transform(result)); 4073 return true; 4074 } 4075 4076 #ifdef _LP64 4077 #define XTOP ,top() /*additional argument*/ 4078 #else //_LP64 4079 #define XTOP /*no additional argument*/ 4080 #endif //_LP64 4081 4082 //----------------------inline_unsafe_copyMemory------------------------- 4083 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4084 bool LibraryCallKit::inline_unsafe_copyMemory() { 4085 if (callee()->is_static()) return false; // caller must have the capability! 4086 null_check_receiver(); // null-check receiver 4087 if (stopped()) return true; 4088 4089 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4090 4091 Node* src_ptr = argument(1); // type: oop 4092 Node* src_off = ConvL2X(argument(2)); // type: long 4093 Node* dst_ptr = argument(4); // type: oop 4094 Node* dst_off = ConvL2X(argument(5)); // type: long 4095 Node* size = ConvL2X(argument(7)); // type: long 4096 4097 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4098 "fieldOffset must be byte-scaled"); 4099 4100 Node* src = make_unsafe_address(src_ptr, src_off); 4101 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4102 4103 // Conservatively insert a memory barrier on all memory slices. 4104 // Do not let writes of the copy source or destination float below the copy. 4105 insert_mem_bar(Op_MemBarCPUOrder); 4106 4107 // Call it. Note that the length argument is not scaled. 4108 make_runtime_call(RC_LEAF|RC_NO_FP, 4109 OptoRuntime::fast_arraycopy_Type(), 4110 StubRoutines::unsafe_arraycopy(), 4111 "unsafe_arraycopy", 4112 TypeRawPtr::BOTTOM, 4113 src, dst, size XTOP); 4114 4115 // Do not let reads of the copy destination float above the copy. 4116 insert_mem_bar(Op_MemBarCPUOrder); 4117 4118 return true; 4119 } 4120 4121 //------------------------clone_coping----------------------------------- 4122 // Helper function for inline_native_clone. 4123 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4124 assert(obj_size != NULL, ""); 4125 Node* raw_obj = alloc_obj->in(1); 4126 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4127 4128 AllocateNode* alloc = NULL; 4129 if (ReduceBulkZeroing) { 4130 // We will be completely responsible for initializing this object - 4131 // mark Initialize node as complete. 4132 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4133 // The object was just allocated - there should be no any stores! 4134 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4135 // Mark as complete_with_arraycopy so that on AllocateNode 4136 // expansion, we know this AllocateNode is initialized by an array 4137 // copy and a StoreStore barrier exists after the array copy. 4138 alloc->initialization()->set_complete_with_arraycopy(); 4139 } 4140 4141 // Copy the fastest available way. 4142 // TODO: generate fields copies for small objects instead. 4143 Node* src = obj; 4144 Node* dest = alloc_obj; 4145 Node* size = _gvn.transform(obj_size); 4146 4147 // Exclude the header but include array length to copy by 8 bytes words. 4148 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4149 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4150 instanceOopDesc::base_offset_in_bytes(); 4151 // base_off: 4152 // 8 - 32-bit VM 4153 // 12 - 64-bit VM, compressed klass 4154 // 16 - 64-bit VM, normal klass 4155 if (base_off % BytesPerLong != 0) { 4156 assert(UseCompressedKlassPointers, ""); 4157 if (is_array) { 4158 // Exclude length to copy by 8 bytes words. 4159 base_off += sizeof(int); 4160 } else { 4161 // Include klass to copy by 8 bytes words. 4162 base_off = instanceOopDesc::klass_offset_in_bytes(); 4163 } 4164 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4165 } 4166 src = basic_plus_adr(src, base_off); 4167 dest = basic_plus_adr(dest, base_off); 4168 4169 // Compute the length also, if needed: 4170 Node* countx = size; 4171 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off))); 4172 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4173 4174 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4175 bool disjoint_bases = true; 4176 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4177 src, NULL, dest, NULL, countx, 4178 /*dest_uninitialized*/true); 4179 4180 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4181 if (card_mark) { 4182 assert(!is_array, ""); 4183 // Put in store barrier for any and all oops we are sticking 4184 // into this object. (We could avoid this if we could prove 4185 // that the object type contains no oop fields at all.) 4186 Node* no_particular_value = NULL; 4187 Node* no_particular_field = NULL; 4188 int raw_adr_idx = Compile::AliasIdxRaw; 4189 post_barrier(control(), 4190 memory(raw_adr_type), 4191 alloc_obj, 4192 no_particular_field, 4193 raw_adr_idx, 4194 no_particular_value, 4195 T_OBJECT, 4196 false); 4197 } 4198 4199 // Do not let reads from the cloned object float above the arraycopy. 4200 if (alloc != NULL) { 4201 // Do not let stores that initialize this object be reordered with 4202 // a subsequent store that would make this object accessible by 4203 // other threads. 4204 // Record what AllocateNode this StoreStore protects so that 4205 // escape analysis can go from the MemBarStoreStoreNode to the 4206 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4207 // based on the escape status of the AllocateNode. 4208 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4209 } else { 4210 insert_mem_bar(Op_MemBarCPUOrder); 4211 } 4212 } 4213 4214 //------------------------inline_native_clone---------------------------- 4215 // protected native Object java.lang.Object.clone(); 4216 // 4217 // Here are the simple edge cases: 4218 // null receiver => normal trap 4219 // virtual and clone was overridden => slow path to out-of-line clone 4220 // not cloneable or finalizer => slow path to out-of-line Object.clone 4221 // 4222 // The general case has two steps, allocation and copying. 4223 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4224 // 4225 // Copying also has two cases, oop arrays and everything else. 4226 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4227 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4228 // 4229 // These steps fold up nicely if and when the cloned object's klass 4230 // can be sharply typed as an object array, a type array, or an instance. 4231 // 4232 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4233 PhiNode* result_val; 4234 4235 // Set the reexecute bit for the interpreter to reexecute 4236 // the bytecode that invokes Object.clone if deoptimization happens. 4237 { PreserveReexecuteState preexecs(this); 4238 jvms()->set_should_reexecute(true); 4239 4240 Node* obj = null_check_receiver(); 4241 if (stopped()) return true; 4242 4243 Node* obj_klass = load_object_klass(obj); 4244 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4245 const TypeOopPtr* toop = ((tklass != NULL) 4246 ? tklass->as_instance_type() 4247 : TypeInstPtr::NOTNULL); 4248 4249 // Conservatively insert a memory barrier on all memory slices. 4250 // Do not let writes into the original float below the clone. 4251 insert_mem_bar(Op_MemBarCPUOrder); 4252 4253 // paths into result_reg: 4254 enum { 4255 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4256 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4257 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4258 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4259 PATH_LIMIT 4260 }; 4261 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4262 result_val = new(C) PhiNode(result_reg, 4263 TypeInstPtr::NOTNULL); 4264 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO); 4265 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 4266 TypePtr::BOTTOM); 4267 record_for_igvn(result_reg); 4268 4269 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4270 int raw_adr_idx = Compile::AliasIdxRaw; 4271 4272 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4273 if (array_ctl != NULL) { 4274 // It's an array. 4275 PreserveJVMState pjvms(this); 4276 set_control(array_ctl); 4277 Node* obj_length = load_array_length(obj); 4278 Node* obj_size = NULL; 4279 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4280 4281 if (!use_ReduceInitialCardMarks()) { 4282 // If it is an oop array, it requires very special treatment, 4283 // because card marking is required on each card of the array. 4284 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4285 if (is_obja != NULL) { 4286 PreserveJVMState pjvms2(this); 4287 set_control(is_obja); 4288 // Generate a direct call to the right arraycopy function(s). 4289 bool disjoint_bases = true; 4290 bool length_never_negative = true; 4291 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4292 obj, intcon(0), alloc_obj, intcon(0), 4293 obj_length, 4294 disjoint_bases, length_never_negative); 4295 result_reg->init_req(_objArray_path, control()); 4296 result_val->init_req(_objArray_path, alloc_obj); 4297 result_i_o ->set_req(_objArray_path, i_o()); 4298 result_mem ->set_req(_objArray_path, reset_memory()); 4299 } 4300 } 4301 // Otherwise, there are no card marks to worry about. 4302 // (We can dispense with card marks if we know the allocation 4303 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4304 // causes the non-eden paths to take compensating steps to 4305 // simulate a fresh allocation, so that no further 4306 // card marks are required in compiled code to initialize 4307 // the object.) 4308 4309 if (!stopped()) { 4310 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4311 4312 // Present the results of the copy. 4313 result_reg->init_req(_array_path, control()); 4314 result_val->init_req(_array_path, alloc_obj); 4315 result_i_o ->set_req(_array_path, i_o()); 4316 result_mem ->set_req(_array_path, reset_memory()); 4317 } 4318 } 4319 4320 // We only go to the instance fast case code if we pass a number of guards. 4321 // The paths which do not pass are accumulated in the slow_region. 4322 RegionNode* slow_region = new (C) RegionNode(1); 4323 record_for_igvn(slow_region); 4324 if (!stopped()) { 4325 // It's an instance (we did array above). Make the slow-path tests. 4326 // If this is a virtual call, we generate a funny guard. We grab 4327 // the vtable entry corresponding to clone() from the target object. 4328 // If the target method which we are calling happens to be the 4329 // Object clone() method, we pass the guard. We do not need this 4330 // guard for non-virtual calls; the caller is known to be the native 4331 // Object clone(). 4332 if (is_virtual) { 4333 generate_virtual_guard(obj_klass, slow_region); 4334 } 4335 4336 // The object must be cloneable and must not have a finalizer. 4337 // Both of these conditions may be checked in a single test. 4338 // We could optimize the cloneable test further, but we don't care. 4339 generate_access_flags_guard(obj_klass, 4340 // Test both conditions: 4341 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4342 // Must be cloneable but not finalizer: 4343 JVM_ACC_IS_CLONEABLE, 4344 slow_region); 4345 } 4346 4347 if (!stopped()) { 4348 // It's an instance, and it passed the slow-path tests. 4349 PreserveJVMState pjvms(this); 4350 Node* obj_size = NULL; 4351 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size); 4352 4353 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4354 4355 // Present the results of the slow call. 4356 result_reg->init_req(_instance_path, control()); 4357 result_val->init_req(_instance_path, alloc_obj); 4358 result_i_o ->set_req(_instance_path, i_o()); 4359 result_mem ->set_req(_instance_path, reset_memory()); 4360 } 4361 4362 // Generate code for the slow case. We make a call to clone(). 4363 set_control(_gvn.transform(slow_region)); 4364 if (!stopped()) { 4365 PreserveJVMState pjvms(this); 4366 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4367 Node* slow_result = set_results_for_java_call(slow_call); 4368 // this->control() comes from set_results_for_java_call 4369 result_reg->init_req(_slow_path, control()); 4370 result_val->init_req(_slow_path, slow_result); 4371 result_i_o ->set_req(_slow_path, i_o()); 4372 result_mem ->set_req(_slow_path, reset_memory()); 4373 } 4374 4375 // Return the combined state. 4376 set_control( _gvn.transform(result_reg)); 4377 set_i_o( _gvn.transform(result_i_o)); 4378 set_all_memory( _gvn.transform(result_mem)); 4379 } // original reexecute is set back here 4380 4381 set_result(_gvn.transform(result_val)); 4382 return true; 4383 } 4384 4385 //------------------------------basictype2arraycopy---------------------------- 4386 address LibraryCallKit::basictype2arraycopy(BasicType t, 4387 Node* src_offset, 4388 Node* dest_offset, 4389 bool disjoint_bases, 4390 const char* &name, 4391 bool dest_uninitialized) { 4392 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4393 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4394 4395 bool aligned = false; 4396 bool disjoint = disjoint_bases; 4397 4398 // if the offsets are the same, we can treat the memory regions as 4399 // disjoint, because either the memory regions are in different arrays, 4400 // or they are identical (which we can treat as disjoint.) We can also 4401 // treat a copy with a destination index less that the source index 4402 // as disjoint since a low->high copy will work correctly in this case. 4403 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4404 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4405 // both indices are constants 4406 int s_offs = src_offset_inttype->get_con(); 4407 int d_offs = dest_offset_inttype->get_con(); 4408 int element_size = type2aelembytes(t); 4409 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4410 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4411 if (s_offs >= d_offs) disjoint = true; 4412 } else if (src_offset == dest_offset && src_offset != NULL) { 4413 // This can occur if the offsets are identical non-constants. 4414 disjoint = true; 4415 } 4416 4417 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4418 } 4419 4420 4421 //------------------------------inline_arraycopy----------------------- 4422 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4423 // Object dest, int destPos, 4424 // int length); 4425 bool LibraryCallKit::inline_arraycopy() { 4426 // Get the arguments. 4427 Node* src = argument(0); // type: oop 4428 Node* src_offset = argument(1); // type: int 4429 Node* dest = argument(2); // type: oop 4430 Node* dest_offset = argument(3); // type: int 4431 Node* length = argument(4); // type: int 4432 4433 // Compile time checks. If any of these checks cannot be verified at compile time, 4434 // we do not make a fast path for this call. Instead, we let the call remain as it 4435 // is. The checks we choose to mandate at compile time are: 4436 // 4437 // (1) src and dest are arrays. 4438 const Type* src_type = src->Value(&_gvn); 4439 const Type* dest_type = dest->Value(&_gvn); 4440 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4441 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4442 if (top_src == NULL || top_src->klass() == NULL || 4443 top_dest == NULL || top_dest->klass() == NULL) { 4444 // Conservatively insert a memory barrier on all memory slices. 4445 // Do not let writes into the source float below the arraycopy. 4446 insert_mem_bar(Op_MemBarCPUOrder); 4447 4448 // Call StubRoutines::generic_arraycopy stub. 4449 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4450 src, src_offset, dest, dest_offset, length); 4451 4452 // Do not let reads from the destination float above the arraycopy. 4453 // Since we cannot type the arrays, we don't know which slices 4454 // might be affected. We could restrict this barrier only to those 4455 // memory slices which pertain to array elements--but don't bother. 4456 if (!InsertMemBarAfterArraycopy) 4457 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4458 insert_mem_bar(Op_MemBarCPUOrder); 4459 return true; 4460 } 4461 4462 // (2) src and dest arrays must have elements of the same BasicType 4463 // Figure out the size and type of the elements we will be copying. 4464 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4465 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4466 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4467 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4468 4469 if (src_elem != dest_elem || dest_elem == T_VOID) { 4470 // The component types are not the same or are not recognized. Punt. 4471 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4472 generate_slow_arraycopy(TypePtr::BOTTOM, 4473 src, src_offset, dest, dest_offset, length, 4474 /*dest_uninitialized*/false); 4475 return true; 4476 } 4477 4478 //--------------------------------------------------------------------------- 4479 // We will make a fast path for this call to arraycopy. 4480 4481 // We have the following tests left to perform: 4482 // 4483 // (3) src and dest must not be null. 4484 // (4) src_offset must not be negative. 4485 // (5) dest_offset must not be negative. 4486 // (6) length must not be negative. 4487 // (7) src_offset + length must not exceed length of src. 4488 // (8) dest_offset + length must not exceed length of dest. 4489 // (9) each element of an oop array must be assignable 4490 4491 RegionNode* slow_region = new (C) RegionNode(1); 4492 record_for_igvn(slow_region); 4493 4494 // (3) operands must not be null 4495 // We currently perform our null checks with the null_check routine. 4496 // This means that the null exceptions will be reported in the caller 4497 // rather than (correctly) reported inside of the native arraycopy call. 4498 // This should be corrected, given time. We do our null check with the 4499 // stack pointer restored. 4500 src = null_check(src, T_ARRAY); 4501 dest = null_check(dest, T_ARRAY); 4502 4503 // (4) src_offset must not be negative. 4504 generate_negative_guard(src_offset, slow_region); 4505 4506 // (5) dest_offset must not be negative. 4507 generate_negative_guard(dest_offset, slow_region); 4508 4509 // (6) length must not be negative (moved to generate_arraycopy()). 4510 // generate_negative_guard(length, slow_region); 4511 4512 // (7) src_offset + length must not exceed length of src. 4513 generate_limit_guard(src_offset, length, 4514 load_array_length(src), 4515 slow_region); 4516 4517 // (8) dest_offset + length must not exceed length of dest. 4518 generate_limit_guard(dest_offset, length, 4519 load_array_length(dest), 4520 slow_region); 4521 4522 // (9) each element of an oop array must be assignable 4523 // The generate_arraycopy subroutine checks this. 4524 4525 // This is where the memory effects are placed: 4526 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4527 generate_arraycopy(adr_type, dest_elem, 4528 src, src_offset, dest, dest_offset, length, 4529 false, false, slow_region); 4530 4531 return true; 4532 } 4533 4534 //-----------------------------generate_arraycopy---------------------- 4535 // Generate an optimized call to arraycopy. 4536 // Caller must guard against non-arrays. 4537 // Caller must determine a common array basic-type for both arrays. 4538 // Caller must validate offsets against array bounds. 4539 // The slow_region has already collected guard failure paths 4540 // (such as out of bounds length or non-conformable array types). 4541 // The generated code has this shape, in general: 4542 // 4543 // if (length == 0) return // via zero_path 4544 // slowval = -1 4545 // if (types unknown) { 4546 // slowval = call generic copy loop 4547 // if (slowval == 0) return // via checked_path 4548 // } else if (indexes in bounds) { 4549 // if ((is object array) && !(array type check)) { 4550 // slowval = call checked copy loop 4551 // if (slowval == 0) return // via checked_path 4552 // } else { 4553 // call bulk copy loop 4554 // return // via fast_path 4555 // } 4556 // } 4557 // // adjust params for remaining work: 4558 // if (slowval != -1) { 4559 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4560 // } 4561 // slow_region: 4562 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4563 // return // via slow_call_path 4564 // 4565 // This routine is used from several intrinsics: System.arraycopy, 4566 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4567 // 4568 void 4569 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4570 BasicType basic_elem_type, 4571 Node* src, Node* src_offset, 4572 Node* dest, Node* dest_offset, 4573 Node* copy_length, 4574 bool disjoint_bases, 4575 bool length_never_negative, 4576 RegionNode* slow_region) { 4577 4578 if (slow_region == NULL) { 4579 slow_region = new(C) RegionNode(1); 4580 record_for_igvn(slow_region); 4581 } 4582 4583 Node* original_dest = dest; 4584 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4585 bool dest_uninitialized = false; 4586 4587 // See if this is the initialization of a newly-allocated array. 4588 // If so, we will take responsibility here for initializing it to zero. 4589 // (Note: Because tightly_coupled_allocation performs checks on the 4590 // out-edges of the dest, we need to avoid making derived pointers 4591 // from it until we have checked its uses.) 4592 if (ReduceBulkZeroing 4593 && !ZeroTLAB // pointless if already zeroed 4594 && basic_elem_type != T_CONFLICT // avoid corner case 4595 && !src->eqv_uncast(dest) 4596 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4597 != NULL) 4598 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4599 && alloc->maybe_set_complete(&_gvn)) { 4600 // "You break it, you buy it." 4601 InitializeNode* init = alloc->initialization(); 4602 assert(init->is_complete(), "we just did this"); 4603 init->set_complete_with_arraycopy(); 4604 assert(dest->is_CheckCastPP(), "sanity"); 4605 assert(dest->in(0)->in(0) == init, "dest pinned"); 4606 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4607 // From this point on, every exit path is responsible for 4608 // initializing any non-copied parts of the object to zero. 4609 // Also, if this flag is set we make sure that arraycopy interacts properly 4610 // with G1, eliding pre-barriers. See CR 6627983. 4611 dest_uninitialized = true; 4612 } else { 4613 // No zeroing elimination here. 4614 alloc = NULL; 4615 //original_dest = dest; 4616 //dest_uninitialized = false; 4617 } 4618 4619 // Results are placed here: 4620 enum { fast_path = 1, // normal void-returning assembly stub 4621 checked_path = 2, // special assembly stub with cleanup 4622 slow_call_path = 3, // something went wrong; call the VM 4623 zero_path = 4, // bypass when length of copy is zero 4624 bcopy_path = 5, // copy primitive array by 64-bit blocks 4625 PATH_LIMIT = 6 4626 }; 4627 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT); 4628 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO); 4629 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type); 4630 record_for_igvn(result_region); 4631 _gvn.set_type_bottom(result_i_o); 4632 _gvn.set_type_bottom(result_memory); 4633 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4634 4635 // The slow_control path: 4636 Node* slow_control; 4637 Node* slow_i_o = i_o(); 4638 Node* slow_mem = memory(adr_type); 4639 debug_only(slow_control = (Node*) badAddress); 4640 4641 // Checked control path: 4642 Node* checked_control = top(); 4643 Node* checked_mem = NULL; 4644 Node* checked_i_o = NULL; 4645 Node* checked_value = NULL; 4646 4647 if (basic_elem_type == T_CONFLICT) { 4648 assert(!dest_uninitialized, ""); 4649 Node* cv = generate_generic_arraycopy(adr_type, 4650 src, src_offset, dest, dest_offset, 4651 copy_length, dest_uninitialized); 4652 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4653 checked_control = control(); 4654 checked_i_o = i_o(); 4655 checked_mem = memory(adr_type); 4656 checked_value = cv; 4657 set_control(top()); // no fast path 4658 } 4659 4660 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4661 if (not_pos != NULL) { 4662 PreserveJVMState pjvms(this); 4663 set_control(not_pos); 4664 4665 // (6) length must not be negative. 4666 if (!length_never_negative) { 4667 generate_negative_guard(copy_length, slow_region); 4668 } 4669 4670 // copy_length is 0. 4671 if (!stopped() && dest_uninitialized) { 4672 Node* dest_length = alloc->in(AllocateNode::ALength); 4673 if (copy_length->eqv_uncast(dest_length) 4674 || _gvn.find_int_con(dest_length, 1) <= 0) { 4675 // There is no zeroing to do. No need for a secondary raw memory barrier. 4676 } else { 4677 // Clear the whole thing since there are no source elements to copy. 4678 generate_clear_array(adr_type, dest, basic_elem_type, 4679 intcon(0), NULL, 4680 alloc->in(AllocateNode::AllocSize)); 4681 // Use a secondary InitializeNode as raw memory barrier. 4682 // Currently it is needed only on this path since other 4683 // paths have stub or runtime calls as raw memory barriers. 4684 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4685 Compile::AliasIdxRaw, 4686 top())->as_Initialize(); 4687 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4688 } 4689 } 4690 4691 // Present the results of the fast call. 4692 result_region->init_req(zero_path, control()); 4693 result_i_o ->init_req(zero_path, i_o()); 4694 result_memory->init_req(zero_path, memory(adr_type)); 4695 } 4696 4697 if (!stopped() && dest_uninitialized) { 4698 // We have to initialize the *uncopied* part of the array to zero. 4699 // The copy destination is the slice dest[off..off+len]. The other slices 4700 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4701 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4702 Node* dest_length = alloc->in(AllocateNode::ALength); 4703 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset, 4704 copy_length)); 4705 4706 // If there is a head section that needs zeroing, do it now. 4707 if (find_int_con(dest_offset, -1) != 0) { 4708 generate_clear_array(adr_type, dest, basic_elem_type, 4709 intcon(0), dest_offset, 4710 NULL); 4711 } 4712 4713 // Next, perform a dynamic check on the tail length. 4714 // It is often zero, and we can win big if we prove this. 4715 // There are two wins: Avoid generating the ClearArray 4716 // with its attendant messy index arithmetic, and upgrade 4717 // the copy to a more hardware-friendly word size of 64 bits. 4718 Node* tail_ctl = NULL; 4719 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 4720 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length)); 4721 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt)); 4722 tail_ctl = generate_slow_guard(bol_lt, NULL); 4723 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4724 } 4725 4726 // At this point, let's assume there is no tail. 4727 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4728 // There is no tail. Try an upgrade to a 64-bit copy. 4729 bool didit = false; 4730 { PreserveJVMState pjvms(this); 4731 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4732 src, src_offset, dest, dest_offset, 4733 dest_size, dest_uninitialized); 4734 if (didit) { 4735 // Present the results of the block-copying fast call. 4736 result_region->init_req(bcopy_path, control()); 4737 result_i_o ->init_req(bcopy_path, i_o()); 4738 result_memory->init_req(bcopy_path, memory(adr_type)); 4739 } 4740 } 4741 if (didit) 4742 set_control(top()); // no regular fast path 4743 } 4744 4745 // Clear the tail, if any. 4746 if (tail_ctl != NULL) { 4747 Node* notail_ctl = stopped() ? NULL : control(); 4748 set_control(tail_ctl); 4749 if (notail_ctl == NULL) { 4750 generate_clear_array(adr_type, dest, basic_elem_type, 4751 dest_tail, NULL, 4752 dest_size); 4753 } else { 4754 // Make a local merge. 4755 Node* done_ctl = new(C) RegionNode(3); 4756 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type); 4757 done_ctl->init_req(1, notail_ctl); 4758 done_mem->init_req(1, memory(adr_type)); 4759 generate_clear_array(adr_type, dest, basic_elem_type, 4760 dest_tail, NULL, 4761 dest_size); 4762 done_ctl->init_req(2, control()); 4763 done_mem->init_req(2, memory(adr_type)); 4764 set_control( _gvn.transform(done_ctl)); 4765 set_memory( _gvn.transform(done_mem), adr_type ); 4766 } 4767 } 4768 } 4769 4770 BasicType copy_type = basic_elem_type; 4771 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 4772 if (!stopped() && copy_type == T_OBJECT) { 4773 // If src and dest have compatible element types, we can copy bits. 4774 // Types S[] and D[] are compatible if D is a supertype of S. 4775 // 4776 // If they are not, we will use checked_oop_disjoint_arraycopy, 4777 // which performs a fast optimistic per-oop check, and backs off 4778 // further to JVM_ArrayCopy on the first per-oop check that fails. 4779 // (Actually, we don't move raw bits only; the GC requires card marks.) 4780 4781 // Get the Klass* for both src and dest 4782 Node* src_klass = load_object_klass(src); 4783 Node* dest_klass = load_object_klass(dest); 4784 4785 // Generate the subtype check. 4786 // This might fold up statically, or then again it might not. 4787 // 4788 // Non-static example: Copying List<String>.elements to a new String[]. 4789 // The backing store for a List<String> is always an Object[], 4790 // but its elements are always type String, if the generic types 4791 // are correct at the source level. 4792 // 4793 // Test S[] against D[], not S against D, because (probably) 4794 // the secondary supertype cache is less busy for S[] than S. 4795 // This usually only matters when D is an interface. 4796 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4797 // Plug failing path into checked_oop_disjoint_arraycopy 4798 if (not_subtype_ctrl != top()) { 4799 PreserveJVMState pjvms(this); 4800 set_control(not_subtype_ctrl); 4801 // (At this point we can assume disjoint_bases, since types differ.) 4802 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 4803 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 4804 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 4805 Node* dest_elem_klass = _gvn.transform(n1); 4806 Node* cv = generate_checkcast_arraycopy(adr_type, 4807 dest_elem_klass, 4808 src, src_offset, dest, dest_offset, 4809 ConvI2X(copy_length), dest_uninitialized); 4810 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4811 checked_control = control(); 4812 checked_i_o = i_o(); 4813 checked_mem = memory(adr_type); 4814 checked_value = cv; 4815 } 4816 // At this point we know we do not need type checks on oop stores. 4817 4818 // Let's see if we need card marks: 4819 if (alloc != NULL && use_ReduceInitialCardMarks()) { 4820 // If we do not need card marks, copy using the jint or jlong stub. 4821 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 4822 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 4823 "sizes agree"); 4824 } 4825 } 4826 4827 if (!stopped()) { 4828 // Generate the fast path, if possible. 4829 PreserveJVMState pjvms(this); 4830 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 4831 src, src_offset, dest, dest_offset, 4832 ConvI2X(copy_length), dest_uninitialized); 4833 4834 // Present the results of the fast call. 4835 result_region->init_req(fast_path, control()); 4836 result_i_o ->init_req(fast_path, i_o()); 4837 result_memory->init_req(fast_path, memory(adr_type)); 4838 } 4839 4840 // Here are all the slow paths up to this point, in one bundle: 4841 slow_control = top(); 4842 if (slow_region != NULL) 4843 slow_control = _gvn.transform(slow_region); 4844 DEBUG_ONLY(slow_region = (RegionNode*)badAddress); 4845 4846 set_control(checked_control); 4847 if (!stopped()) { 4848 // Clean up after the checked call. 4849 // The returned value is either 0 or -1^K, 4850 // where K = number of partially transferred array elements. 4851 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0))); 4852 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 4853 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4854 4855 // If it is 0, we are done, so transfer to the end. 4856 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff)); 4857 result_region->init_req(checked_path, checks_done); 4858 result_i_o ->init_req(checked_path, checked_i_o); 4859 result_memory->init_req(checked_path, checked_mem); 4860 4861 // If it is not zero, merge into the slow call. 4862 set_control( _gvn.transform(new(C) IfFalseNode(iff) )); 4863 RegionNode* slow_reg2 = new(C) RegionNode(3); 4864 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO); 4865 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type); 4866 record_for_igvn(slow_reg2); 4867 slow_reg2 ->init_req(1, slow_control); 4868 slow_i_o2 ->init_req(1, slow_i_o); 4869 slow_mem2 ->init_req(1, slow_mem); 4870 slow_reg2 ->init_req(2, control()); 4871 slow_i_o2 ->init_req(2, checked_i_o); 4872 slow_mem2 ->init_req(2, checked_mem); 4873 4874 slow_control = _gvn.transform(slow_reg2); 4875 slow_i_o = _gvn.transform(slow_i_o2); 4876 slow_mem = _gvn.transform(slow_mem2); 4877 4878 if (alloc != NULL) { 4879 // We'll restart from the very beginning, after zeroing the whole thing. 4880 // This can cause double writes, but that's OK since dest is brand new. 4881 // So we ignore the low 31 bits of the value returned from the stub. 4882 } else { 4883 // We must continue the copy exactly where it failed, or else 4884 // another thread might see the wrong number of writes to dest. 4885 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1))); 4886 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT); 4887 slow_offset->init_req(1, intcon(0)); 4888 slow_offset->init_req(2, checked_offset); 4889 slow_offset = _gvn.transform(slow_offset); 4890 4891 // Adjust the arguments by the conditionally incoming offset. 4892 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset)); 4893 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset)); 4894 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset)); 4895 4896 // Tweak the node variables to adjust the code produced below: 4897 src_offset = src_off_plus; 4898 dest_offset = dest_off_plus; 4899 copy_length = length_minus; 4900 } 4901 } 4902 4903 set_control(slow_control); 4904 if (!stopped()) { 4905 // Generate the slow path, if needed. 4906 PreserveJVMState pjvms(this); // replace_in_map may trash the map 4907 4908 set_memory(slow_mem, adr_type); 4909 set_i_o(slow_i_o); 4910 4911 if (dest_uninitialized) { 4912 generate_clear_array(adr_type, dest, basic_elem_type, 4913 intcon(0), NULL, 4914 alloc->in(AllocateNode::AllocSize)); 4915 } 4916 4917 generate_slow_arraycopy(adr_type, 4918 src, src_offset, dest, dest_offset, 4919 copy_length, /*dest_uninitialized*/false); 4920 4921 result_region->init_req(slow_call_path, control()); 4922 result_i_o ->init_req(slow_call_path, i_o()); 4923 result_memory->init_req(slow_call_path, memory(adr_type)); 4924 } 4925 4926 // Remove unused edges. 4927 for (uint i = 1; i < result_region->req(); i++) { 4928 if (result_region->in(i) == NULL) 4929 result_region->init_req(i, top()); 4930 } 4931 4932 // Finished; return the combined state. 4933 set_control( _gvn.transform(result_region)); 4934 set_i_o( _gvn.transform(result_i_o) ); 4935 set_memory( _gvn.transform(result_memory), adr_type ); 4936 4937 // The memory edges above are precise in order to model effects around 4938 // array copies accurately to allow value numbering of field loads around 4939 // arraycopy. Such field loads, both before and after, are common in Java 4940 // collections and similar classes involving header/array data structures. 4941 // 4942 // But with low number of register or when some registers are used or killed 4943 // by arraycopy calls it causes registers spilling on stack. See 6544710. 4944 // The next memory barrier is added to avoid it. If the arraycopy can be 4945 // optimized away (which it can, sometimes) then we can manually remove 4946 // the membar also. 4947 // 4948 // Do not let reads from the cloned object float above the arraycopy. 4949 if (alloc != NULL) { 4950 // Do not let stores that initialize this object be reordered with 4951 // a subsequent store that would make this object accessible by 4952 // other threads. 4953 // Record what AllocateNode this StoreStore protects so that 4954 // escape analysis can go from the MemBarStoreStoreNode to the 4955 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4956 // based on the escape status of the AllocateNode. 4957 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4958 } else if (InsertMemBarAfterArraycopy) 4959 insert_mem_bar(Op_MemBarCPUOrder); 4960 } 4961 4962 4963 // Helper function which determines if an arraycopy immediately follows 4964 // an allocation, with no intervening tests or other escapes for the object. 4965 AllocateArrayNode* 4966 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4967 RegionNode* slow_region) { 4968 if (stopped()) return NULL; // no fast path 4969 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4970 4971 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4972 if (alloc == NULL) return NULL; 4973 4974 Node* rawmem = memory(Compile::AliasIdxRaw); 4975 // Is the allocation's memory state untouched? 4976 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4977 // Bail out if there have been raw-memory effects since the allocation. 4978 // (Example: There might have been a call or safepoint.) 4979 return NULL; 4980 } 4981 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4982 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4983 return NULL; 4984 } 4985 4986 // There must be no unexpected observers of this allocation. 4987 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4988 Node* obs = ptr->fast_out(i); 4989 if (obs != this->map()) { 4990 return NULL; 4991 } 4992 } 4993 4994 // This arraycopy must unconditionally follow the allocation of the ptr. 4995 Node* alloc_ctl = ptr->in(0); 4996 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4997 4998 Node* ctl = control(); 4999 while (ctl != alloc_ctl) { 5000 // There may be guards which feed into the slow_region. 5001 // Any other control flow means that we might not get a chance 5002 // to finish initializing the allocated object. 5003 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5004 IfNode* iff = ctl->in(0)->as_If(); 5005 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5006 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5007 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5008 ctl = iff->in(0); // This test feeds the known slow_region. 5009 continue; 5010 } 5011 // One more try: Various low-level checks bottom out in 5012 // uncommon traps. If the debug-info of the trap omits 5013 // any reference to the allocation, as we've already 5014 // observed, then there can be no objection to the trap. 5015 bool found_trap = false; 5016 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5017 Node* obs = not_ctl->fast_out(j); 5018 if (obs->in(0) == not_ctl && obs->is_Call() && 5019 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5020 found_trap = true; break; 5021 } 5022 } 5023 if (found_trap) { 5024 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5025 continue; 5026 } 5027 } 5028 return NULL; 5029 } 5030 5031 // If we get this far, we have an allocation which immediately 5032 // precedes the arraycopy, and we can take over zeroing the new object. 5033 // The arraycopy will finish the initialization, and provide 5034 // a new control state to which we will anchor the destination pointer. 5035 5036 return alloc; 5037 } 5038 5039 // Helper for initialization of arrays, creating a ClearArray. 5040 // It writes zero bits in [start..end), within the body of an array object. 5041 // The memory effects are all chained onto the 'adr_type' alias category. 5042 // 5043 // Since the object is otherwise uninitialized, we are free 5044 // to put a little "slop" around the edges of the cleared area, 5045 // as long as it does not go back into the array's header, 5046 // or beyond the array end within the heap. 5047 // 5048 // The lower edge can be rounded down to the nearest jint and the 5049 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5050 // 5051 // Arguments: 5052 // adr_type memory slice where writes are generated 5053 // dest oop of the destination array 5054 // basic_elem_type element type of the destination 5055 // slice_idx array index of first element to store 5056 // slice_len number of elements to store (or NULL) 5057 // dest_size total size in bytes of the array object 5058 // 5059 // Exactly one of slice_len or dest_size must be non-NULL. 5060 // If dest_size is non-NULL, zeroing extends to the end of the object. 5061 // If slice_len is non-NULL, the slice_idx value must be a constant. 5062 void 5063 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5064 Node* dest, 5065 BasicType basic_elem_type, 5066 Node* slice_idx, 5067 Node* slice_len, 5068 Node* dest_size) { 5069 // one or the other but not both of slice_len and dest_size: 5070 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5071 if (slice_len == NULL) slice_len = top(); 5072 if (dest_size == NULL) dest_size = top(); 5073 5074 // operate on this memory slice: 5075 Node* mem = memory(adr_type); // memory slice to operate on 5076 5077 // scaling and rounding of indexes: 5078 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5079 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5080 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5081 int bump_bit = (-1 << scale) & BytesPerInt; 5082 5083 // determine constant starts and ends 5084 const intptr_t BIG_NEG = -128; 5085 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5086 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5087 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5088 if (slice_len_con == 0) { 5089 return; // nothing to do here 5090 } 5091 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5092 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5093 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5094 assert(end_con < 0, "not two cons"); 5095 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5096 BytesPerLong); 5097 } 5098 5099 if (start_con >= 0 && end_con >= 0) { 5100 // Constant start and end. Simple. 5101 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5102 start_con, end_con, &_gvn); 5103 } else if (start_con >= 0 && dest_size != top()) { 5104 // Constant start, pre-rounded end after the tail of the array. 5105 Node* end = dest_size; 5106 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5107 start_con, end, &_gvn); 5108 } else if (start_con >= 0 && slice_len != top()) { 5109 // Constant start, non-constant end. End needs rounding up. 5110 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5111 intptr_t end_base = abase + (slice_idx_con << scale); 5112 int end_round = (-1 << scale) & (BytesPerLong - 1); 5113 Node* end = ConvI2X(slice_len); 5114 if (scale != 0) 5115 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) )); 5116 end_base += end_round; 5117 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base))); 5118 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round))); 5119 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5120 start_con, end, &_gvn); 5121 } else if (start_con < 0 && dest_size != top()) { 5122 // Non-constant start, pre-rounded end after the tail of the array. 5123 // This is almost certainly a "round-to-end" operation. 5124 Node* start = slice_idx; 5125 start = ConvI2X(start); 5126 if (scale != 0) 5127 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) )); 5128 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase))); 5129 if ((bump_bit | clear_low) != 0) { 5130 int to_clear = (bump_bit | clear_low); 5131 // Align up mod 8, then store a jint zero unconditionally 5132 // just before the mod-8 boundary. 5133 if (((abase + bump_bit) & ~to_clear) - bump_bit 5134 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5135 bump_bit = 0; 5136 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5137 } else { 5138 // Bump 'start' up to (or past) the next jint boundary: 5139 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit))); 5140 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5141 } 5142 // Round bumped 'start' down to jlong boundary in body of array. 5143 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear))); 5144 if (bump_bit != 0) { 5145 // Store a zero to the immediately preceding jint: 5146 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit))); 5147 Node* p1 = basic_plus_adr(dest, x1); 5148 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 5149 mem = _gvn.transform(mem); 5150 } 5151 } 5152 Node* end = dest_size; // pre-rounded 5153 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5154 start, end, &_gvn); 5155 } else { 5156 // Non-constant start, unrounded non-constant end. 5157 // (Nobody zeroes a random midsection of an array using this routine.) 5158 ShouldNotReachHere(); // fix caller 5159 } 5160 5161 // Done. 5162 set_memory(mem, adr_type); 5163 } 5164 5165 5166 bool 5167 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5168 BasicType basic_elem_type, 5169 AllocateNode* alloc, 5170 Node* src, Node* src_offset, 5171 Node* dest, Node* dest_offset, 5172 Node* dest_size, bool dest_uninitialized) { 5173 // See if there is an advantage from block transfer. 5174 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5175 if (scale >= LogBytesPerLong) 5176 return false; // it is already a block transfer 5177 5178 // Look at the alignment of the starting offsets. 5179 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5180 5181 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5182 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5183 if (src_off_con < 0 || dest_off_con < 0) 5184 // At present, we can only understand constants. 5185 return false; 5186 5187 intptr_t src_off = abase + (src_off_con << scale); 5188 intptr_t dest_off = abase + (dest_off_con << scale); 5189 5190 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5191 // Non-aligned; too bad. 5192 // One more chance: Pick off an initial 32-bit word. 5193 // This is a common case, since abase can be odd mod 8. 5194 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5195 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5196 Node* sptr = basic_plus_adr(src, src_off); 5197 Node* dptr = basic_plus_adr(dest, dest_off); 5198 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 5199 store_to_memory(control(), dptr, sval, T_INT, adr_type); 5200 src_off += BytesPerInt; 5201 dest_off += BytesPerInt; 5202 } else { 5203 return false; 5204 } 5205 } 5206 assert(src_off % BytesPerLong == 0, ""); 5207 assert(dest_off % BytesPerLong == 0, ""); 5208 5209 // Do this copy by giant steps. 5210 Node* sptr = basic_plus_adr(src, src_off); 5211 Node* dptr = basic_plus_adr(dest, dest_off); 5212 Node* countx = dest_size; 5213 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off))); 5214 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong))); 5215 5216 bool disjoint_bases = true; // since alloc != NULL 5217 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5218 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5219 5220 return true; 5221 } 5222 5223 5224 // Helper function; generates code for the slow case. 5225 // We make a call to a runtime method which emulates the native method, 5226 // but without the native wrapper overhead. 5227 void 5228 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5229 Node* src, Node* src_offset, 5230 Node* dest, Node* dest_offset, 5231 Node* copy_length, bool dest_uninitialized) { 5232 assert(!dest_uninitialized, "Invariant"); 5233 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5234 OptoRuntime::slow_arraycopy_Type(), 5235 OptoRuntime::slow_arraycopy_Java(), 5236 "slow_arraycopy", adr_type, 5237 src, src_offset, dest, dest_offset, 5238 copy_length); 5239 5240 // Handle exceptions thrown by this fellow: 5241 make_slow_call_ex(call, env()->Throwable_klass(), false); 5242 } 5243 5244 // Helper function; generates code for cases requiring runtime checks. 5245 Node* 5246 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5247 Node* dest_elem_klass, 5248 Node* src, Node* src_offset, 5249 Node* dest, Node* dest_offset, 5250 Node* copy_length, bool dest_uninitialized) { 5251 if (stopped()) return NULL; 5252 5253 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5254 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5255 return NULL; 5256 } 5257 5258 // Pick out the parameters required to perform a store-check 5259 // for the target array. This is an optimistic check. It will 5260 // look in each non-null element's class, at the desired klass's 5261 // super_check_offset, for the desired klass. 5262 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5263 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5264 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr()); 5265 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5266 Node* check_value = dest_elem_klass; 5267 5268 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5269 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5270 5271 // (We know the arrays are never conjoint, because their types differ.) 5272 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5273 OptoRuntime::checkcast_arraycopy_Type(), 5274 copyfunc_addr, "checkcast_arraycopy", adr_type, 5275 // five arguments, of which two are 5276 // intptr_t (jlong in LP64) 5277 src_start, dest_start, 5278 copy_length XTOP, 5279 check_offset XTOP, 5280 check_value); 5281 5282 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5283 } 5284 5285 5286 // Helper function; generates code for cases requiring runtime checks. 5287 Node* 5288 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5289 Node* src, Node* src_offset, 5290 Node* dest, Node* dest_offset, 5291 Node* copy_length, bool dest_uninitialized) { 5292 assert(!dest_uninitialized, "Invariant"); 5293 if (stopped()) return NULL; 5294 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5295 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5296 return NULL; 5297 } 5298 5299 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5300 OptoRuntime::generic_arraycopy_Type(), 5301 copyfunc_addr, "generic_arraycopy", adr_type, 5302 src, src_offset, dest, dest_offset, copy_length); 5303 5304 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5305 } 5306 5307 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5308 void 5309 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5310 BasicType basic_elem_type, 5311 bool disjoint_bases, 5312 Node* src, Node* src_offset, 5313 Node* dest, Node* dest_offset, 5314 Node* copy_length, bool dest_uninitialized) { 5315 if (stopped()) return; // nothing to do 5316 5317 Node* src_start = src; 5318 Node* dest_start = dest; 5319 if (src_offset != NULL || dest_offset != NULL) { 5320 assert(src_offset != NULL && dest_offset != NULL, ""); 5321 src_start = array_element_address(src, src_offset, basic_elem_type); 5322 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5323 } 5324 5325 // Figure out which arraycopy runtime method to call. 5326 const char* copyfunc_name = "arraycopy"; 5327 address copyfunc_addr = 5328 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5329 disjoint_bases, copyfunc_name, dest_uninitialized); 5330 5331 // Call it. Note that the count_ix value is not scaled to a byte-size. 5332 make_runtime_call(RC_LEAF|RC_NO_FP, 5333 OptoRuntime::fast_arraycopy_Type(), 5334 copyfunc_addr, copyfunc_name, adr_type, 5335 src_start, dest_start, copy_length XTOP); 5336 } 5337 5338 //-------------inline_encodeISOArray----------------------------------- 5339 // encode char[] to byte[] in ISO_8859_1 5340 bool LibraryCallKit::inline_encodeISOArray() { 5341 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5342 // no receiver since it is static method 5343 Node *src = argument(0); 5344 Node *src_offset = argument(1); 5345 Node *dst = argument(2); 5346 Node *dst_offset = argument(3); 5347 Node *length = argument(4); 5348 5349 const Type* src_type = src->Value(&_gvn); 5350 const Type* dst_type = dst->Value(&_gvn); 5351 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5352 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5353 if (top_src == NULL || top_src->klass() == NULL || 5354 top_dest == NULL || top_dest->klass() == NULL) { 5355 // failed array check 5356 return false; 5357 } 5358 5359 // Figure out the size and type of the elements we will be copying. 5360 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5361 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5362 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5363 return false; 5364 } 5365 Node* src_start = array_element_address(src, src_offset, src_elem); 5366 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5367 // 'src_start' points to src array + scaled offset 5368 // 'dst_start' points to dst array + scaled offset 5369 5370 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5371 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5372 enc = _gvn.transform(enc); 5373 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc)); 5374 set_memory(res_mem, mtype); 5375 set_result(enc); 5376 return true; 5377 } 5378 5379 /** 5380 * Calculate CRC32 for byte. 5381 * int java.util.zip.CRC32.update(int crc, int b) 5382 */ 5383 bool LibraryCallKit::inline_updateCRC32() { 5384 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5385 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5386 // no receiver since it is static method 5387 Node* crc = argument(0); // type: int 5388 Node* b = argument(1); // type: int 5389 5390 /* 5391 * int c = ~ crc; 5392 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5393 * b = b ^ (c >>> 8); 5394 * crc = ~b; 5395 */ 5396 5397 Node* M1 = intcon(-1); 5398 crc = _gvn.transform(new (C) XorINode(crc, M1)); 5399 Node* result = _gvn.transform(new (C) XorINode(crc, b)); 5400 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF))); 5401 5402 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5403 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2))); 5404 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5405 result = make_load(control(), adr, TypeInt::INT, T_INT); 5406 5407 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8))); 5408 result = _gvn.transform(new (C) XorINode(crc, result)); 5409 result = _gvn.transform(new (C) XorINode(result, M1)); 5410 set_result(result); 5411 return true; 5412 } 5413 5414 /** 5415 * Calculate CRC32 for byte[] array. 5416 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5417 */ 5418 bool LibraryCallKit::inline_updateBytesCRC32() { 5419 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5420 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5421 // no receiver since it is static method 5422 Node* crc = argument(0); // type: int 5423 Node* src = argument(1); // type: oop 5424 Node* offset = argument(2); // type: int 5425 Node* length = argument(3); // type: int 5426 5427 const Type* src_type = src->Value(&_gvn); 5428 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5429 if (top_src == NULL || top_src->klass() == NULL) { 5430 // failed array check 5431 return false; 5432 } 5433 5434 // Figure out the size and type of the elements we will be copying. 5435 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5436 if (src_elem != T_BYTE) { 5437 return false; 5438 } 5439 5440 // 'src_start' points to src array + scaled offset 5441 Node* src_start = array_element_address(src, offset, src_elem); 5442 5443 // We assume that range check is done by caller. 5444 // TODO: generate range check (offset+length < src.length) in debug VM. 5445 5446 // Call the stub. 5447 address stubAddr = StubRoutines::updateBytesCRC32(); 5448 const char *stubName = "updateBytesCRC32"; 5449 5450 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5451 stubAddr, stubName, TypePtr::BOTTOM, 5452 crc, src_start, length); 5453 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5454 set_result(result); 5455 return true; 5456 } 5457 5458 /** 5459 * Calculate CRC32 for ByteBuffer. 5460 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5461 */ 5462 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5463 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5464 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5465 // no receiver since it is static method 5466 Node* crc = argument(0); // type: int 5467 Node* src = argument(1); // type: long 5468 Node* offset = argument(3); // type: int 5469 Node* length = argument(4); // type: int 5470 5471 src = ConvL2X(src); // adjust Java long to machine word 5472 Node* base = _gvn.transform(new (C) CastX2PNode(src)); 5473 offset = ConvI2X(offset); 5474 5475 // 'src_start' points to src array + scaled offset 5476 Node* src_start = basic_plus_adr(top(), base, offset); 5477 5478 // Call the stub. 5479 address stubAddr = StubRoutines::updateBytesCRC32(); 5480 const char *stubName = "updateBytesCRC32"; 5481 5482 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5483 stubAddr, stubName, TypePtr::BOTTOM, 5484 crc, src_start, length); 5485 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5486 set_result(result); 5487 return true; 5488 } 5489 5490 //----------------------------inline_reference_get---------------------------- 5491 // public T java.lang.ref.Reference.get(); 5492 bool LibraryCallKit::inline_reference_get() { 5493 const int referent_offset = java_lang_ref_Reference::referent_offset; 5494 guarantee(referent_offset > 0, "should have already been set"); 5495 5496 // Get the argument: 5497 Node* reference_obj = null_check_receiver(); 5498 if (stopped()) return true; 5499 5500 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5501 5502 ciInstanceKlass* klass = env()->Object_klass(); 5503 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5504 5505 Node* no_ctrl = NULL; 5506 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT); 5507 5508 // Use the pre-barrier to record the value in the referent field 5509 pre_barrier(false /* do_load */, 5510 control(), 5511 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5512 result /* pre_val */, 5513 T_OBJECT); 5514 5515 // Add memory barrier to prevent commoning reads from this field 5516 // across safepoint since GC can change its value. 5517 insert_mem_bar(Op_MemBarCPUOrder); 5518 5519 set_result(result); 5520 return true; 5521 } 5522 5523 5524 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5525 bool is_exact=true, bool is_static=false) { 5526 5527 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5528 assert(tinst != NULL, "obj is null"); 5529 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5530 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5531 5532 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5533 ciSymbol::make(fieldTypeString), 5534 is_static); 5535 if (field == NULL) return (Node *) NULL; 5536 assert (field != NULL, "undefined field"); 5537 5538 // Next code copied from Parse::do_get_xxx(): 5539 5540 // Compute address and memory type. 5541 int offset = field->offset_in_bytes(); 5542 bool is_vol = field->is_volatile(); 5543 ciType* field_klass = field->type(); 5544 assert(field_klass->is_loaded(), "should be loaded"); 5545 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5546 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5547 BasicType bt = field->layout_type(); 5548 5549 // Build the resultant type of the load 5550 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5551 5552 // Build the load. 5553 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol); 5554 return loadedField; 5555 } 5556 5557 5558 //------------------------------inline_aescrypt_Block----------------------- 5559 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5560 address stubAddr; 5561 const char *stubName; 5562 assert(UseAES, "need AES instruction support"); 5563 5564 switch(id) { 5565 case vmIntrinsics::_aescrypt_encryptBlock: 5566 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5567 stubName = "aescrypt_encryptBlock"; 5568 break; 5569 case vmIntrinsics::_aescrypt_decryptBlock: 5570 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5571 stubName = "aescrypt_decryptBlock"; 5572 break; 5573 } 5574 if (stubAddr == NULL) return false; 5575 5576 Node* aescrypt_object = argument(0); 5577 Node* src = argument(1); 5578 Node* src_offset = argument(2); 5579 Node* dest = argument(3); 5580 Node* dest_offset = argument(4); 5581 5582 // (1) src and dest are arrays. 5583 const Type* src_type = src->Value(&_gvn); 5584 const Type* dest_type = dest->Value(&_gvn); 5585 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5586 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5587 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5588 5589 // for the quick and dirty code we will skip all the checks. 5590 // we are just trying to get the call to be generated. 5591 Node* src_start = src; 5592 Node* dest_start = dest; 5593 if (src_offset != NULL || dest_offset != NULL) { 5594 assert(src_offset != NULL && dest_offset != NULL, ""); 5595 src_start = array_element_address(src, src_offset, T_BYTE); 5596 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5597 } 5598 5599 // now need to get the start of its expanded key array 5600 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5601 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5602 if (k_start == NULL) return false; 5603 5604 // Call the stub. 5605 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5606 stubAddr, stubName, TypePtr::BOTTOM, 5607 src_start, dest_start, k_start); 5608 5609 return true; 5610 } 5611 5612 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5613 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5614 address stubAddr; 5615 const char *stubName; 5616 5617 assert(UseAES, "need AES instruction support"); 5618 5619 switch(id) { 5620 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5621 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5622 stubName = "cipherBlockChaining_encryptAESCrypt"; 5623 break; 5624 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5625 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5626 stubName = "cipherBlockChaining_decryptAESCrypt"; 5627 break; 5628 } 5629 if (stubAddr == NULL) return false; 5630 5631 Node* cipherBlockChaining_object = argument(0); 5632 Node* src = argument(1); 5633 Node* src_offset = argument(2); 5634 Node* len = argument(3); 5635 Node* dest = argument(4); 5636 Node* dest_offset = argument(5); 5637 5638 // (1) src and dest are arrays. 5639 const Type* src_type = src->Value(&_gvn); 5640 const Type* dest_type = dest->Value(&_gvn); 5641 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5642 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5643 assert (top_src != NULL && top_src->klass() != NULL 5644 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5645 5646 // checks are the responsibility of the caller 5647 Node* src_start = src; 5648 Node* dest_start = dest; 5649 if (src_offset != NULL || dest_offset != NULL) { 5650 assert(src_offset != NULL && dest_offset != NULL, ""); 5651 src_start = array_element_address(src, src_offset, T_BYTE); 5652 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5653 } 5654 5655 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5656 // (because of the predicated logic executed earlier). 5657 // so we cast it here safely. 5658 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5659 5660 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5661 if (embeddedCipherObj == NULL) return false; 5662 5663 // cast it to what we know it will be at runtime 5664 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5665 assert(tinst != NULL, "CBC obj is null"); 5666 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5667 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5668 if (!klass_AESCrypt->is_loaded()) return false; 5669 5670 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5671 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5672 const TypeOopPtr* xtype = aklass->as_instance_type(); 5673 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype); 5674 aescrypt_object = _gvn.transform(aescrypt_object); 5675 5676 // we need to get the start of the aescrypt_object's expanded key array 5677 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5678 if (k_start == NULL) return false; 5679 5680 // similarly, get the start address of the r vector 5681 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5682 if (objRvec == NULL) return false; 5683 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5684 5685 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5686 make_runtime_call(RC_LEAF|RC_NO_FP, 5687 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5688 stubAddr, stubName, TypePtr::BOTTOM, 5689 src_start, dest_start, k_start, r_start, len); 5690 5691 // return is void so no result needs to be pushed 5692 5693 return true; 5694 } 5695 5696 //------------------------------get_key_start_from_aescrypt_object----------------------- 5697 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5698 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5699 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5700 if (objAESCryptKey == NULL) return (Node *) NULL; 5701 5702 // now have the array, need to get the start address of the K array 5703 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5704 return k_start; 5705 } 5706 5707 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5708 // Return node representing slow path of predicate check. 5709 // the pseudo code we want to emulate with this predicate is: 5710 // for encryption: 5711 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5712 // for decryption: 5713 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5714 // note cipher==plain is more conservative than the original java code but that's OK 5715 // 5716 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5717 // First, check receiver for NULL since it is virtual method. 5718 Node* objCBC = argument(0); 5719 objCBC = null_check(objCBC); 5720 5721 if (stopped()) return NULL; // Always NULL 5722 5723 // Load embeddedCipher field of CipherBlockChaining object. 5724 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5725 5726 // get AESCrypt klass for instanceOf check 5727 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5728 // will have same classloader as CipherBlockChaining object 5729 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5730 assert(tinst != NULL, "CBCobj is null"); 5731 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5732 5733 // we want to do an instanceof comparison against the AESCrypt class 5734 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5735 if (!klass_AESCrypt->is_loaded()) { 5736 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5737 Node* ctrl = control(); 5738 set_control(top()); // no regular fast path 5739 return ctrl; 5740 } 5741 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5742 5743 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5744 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1))); 5745 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); 5746 5747 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5748 5749 // for encryption, we are done 5750 if (!decrypting) 5751 return instof_false; // even if it is NULL 5752 5753 // for decryption, we need to add a further check to avoid 5754 // taking the intrinsic path when cipher and plain are the same 5755 // see the original java code for why. 5756 RegionNode* region = new(C) RegionNode(3); 5757 region->init_req(1, instof_false); 5758 Node* src = argument(1); 5759 Node* dest = argument(4); 5760 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest)); 5761 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq)); 5762 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5763 region->init_req(2, src_dest_conjoint); 5764 5765 record_for_igvn(region); 5766 return _gvn.transform(region); 5767 }