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 static bool klass_needs_init_guard(Node* kls); 217 bool inline_unsafe_allocate(); 218 bool inline_unsafe_copyMemory(); 219 bool inline_native_currentThread(); 220 #ifdef TRACE_HAVE_INTRINSICS 221 bool inline_native_classID(); 222 bool inline_native_threadID(); 223 #endif 224 bool inline_native_time_funcs(address method, const char* funcName); 225 bool inline_native_isInterrupted(); 226 bool inline_native_Class_query(vmIntrinsics::ID id); 227 bool inline_native_subtype_check(); 228 229 bool inline_native_newArray(); 230 bool inline_native_getLength(); 231 bool inline_array_copyOf(bool is_copyOfRange); 232 bool inline_array_equals(); 233 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 234 bool inline_native_clone(bool is_virtual); 235 bool inline_native_Reflection_getCallerClass(); 236 // Helper function for inlining native object hash method 237 bool inline_native_hashcode(bool is_virtual, bool is_static); 238 bool inline_native_getClass(); 239 240 // Helper functions for inlining arraycopy 241 bool inline_arraycopy(); 242 void generate_arraycopy(const TypePtr* adr_type, 243 BasicType basic_elem_type, 244 Node* src, Node* src_offset, 245 Node* dest, Node* dest_offset, 246 Node* copy_length, 247 bool disjoint_bases = false, 248 bool length_never_negative = false, 249 RegionNode* slow_region = NULL); 250 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 251 RegionNode* slow_region); 252 void generate_clear_array(const TypePtr* adr_type, 253 Node* dest, 254 BasicType basic_elem_type, 255 Node* slice_off, 256 Node* slice_len, 257 Node* slice_end); 258 bool generate_block_arraycopy(const TypePtr* adr_type, 259 BasicType basic_elem_type, 260 AllocateNode* alloc, 261 Node* src, Node* src_offset, 262 Node* dest, Node* dest_offset, 263 Node* dest_size, bool dest_uninitialized); 264 void generate_slow_arraycopy(const TypePtr* adr_type, 265 Node* src, Node* src_offset, 266 Node* dest, Node* dest_offset, 267 Node* copy_length, bool dest_uninitialized); 268 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 269 Node* dest_elem_klass, 270 Node* src, Node* src_offset, 271 Node* dest, Node* dest_offset, 272 Node* copy_length, bool dest_uninitialized); 273 Node* generate_generic_arraycopy(const TypePtr* adr_type, 274 Node* src, Node* src_offset, 275 Node* dest, Node* dest_offset, 276 Node* copy_length, bool dest_uninitialized); 277 void generate_unchecked_arraycopy(const TypePtr* adr_type, 278 BasicType basic_elem_type, 279 bool disjoint_bases, 280 Node* src, Node* src_offset, 281 Node* dest, Node* dest_offset, 282 Node* copy_length, bool dest_uninitialized); 283 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 284 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 285 bool inline_unsafe_ordered_store(BasicType type); 286 bool inline_unsafe_fence(vmIntrinsics::ID id); 287 bool inline_fp_conversions(vmIntrinsics::ID id); 288 bool inline_number_methods(vmIntrinsics::ID id); 289 bool inline_reference_get(); 290 bool inline_aescrypt_Block(vmIntrinsics::ID id); 291 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 292 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 293 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 294 bool inline_encodeISOArray(); 295 bool inline_updateCRC32(); 296 bool inline_updateBytesCRC32(); 297 bool inline_updateByteBufferCRC32(); 298 }; 299 300 301 //---------------------------make_vm_intrinsic---------------------------- 302 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 303 vmIntrinsics::ID id = m->intrinsic_id(); 304 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 305 306 if (DisableIntrinsic[0] != '\0' 307 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 308 // disabled by a user request on the command line: 309 // example: -XX:DisableIntrinsic=_hashCode,_getClass 310 return NULL; 311 } 312 313 if (!m->is_loaded()) { 314 // do not attempt to inline unloaded methods 315 return NULL; 316 } 317 318 // Only a few intrinsics implement a virtual dispatch. 319 // They are expensive calls which are also frequently overridden. 320 if (is_virtual) { 321 switch (id) { 322 case vmIntrinsics::_hashCode: 323 case vmIntrinsics::_clone: 324 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 325 break; 326 default: 327 return NULL; 328 } 329 } 330 331 // -XX:-InlineNatives disables nearly all intrinsics: 332 if (!InlineNatives) { 333 switch (id) { 334 case vmIntrinsics::_indexOf: 335 case vmIntrinsics::_compareTo: 336 case vmIntrinsics::_equals: 337 case vmIntrinsics::_equalsC: 338 case vmIntrinsics::_getAndAddInt: 339 case vmIntrinsics::_getAndAddLong: 340 case vmIntrinsics::_getAndSetInt: 341 case vmIntrinsics::_getAndSetLong: 342 case vmIntrinsics::_getAndSetObject: 343 case vmIntrinsics::_loadFence: 344 case vmIntrinsics::_storeFence: 345 case vmIntrinsics::_fullFence: 346 break; // InlineNatives does not control String.compareTo 347 case vmIntrinsics::_Reference_get: 348 break; // InlineNatives does not control Reference.get 349 default: 350 return NULL; 351 } 352 } 353 354 bool is_predicted = false; 355 356 switch (id) { 357 case vmIntrinsics::_compareTo: 358 if (!SpecialStringCompareTo) return NULL; 359 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 360 break; 361 case vmIntrinsics::_indexOf: 362 if (!SpecialStringIndexOf) return NULL; 363 break; 364 case vmIntrinsics::_equals: 365 if (!SpecialStringEquals) return NULL; 366 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 367 break; 368 case vmIntrinsics::_equalsC: 369 if (!SpecialArraysEquals) return NULL; 370 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 371 break; 372 case vmIntrinsics::_arraycopy: 373 if (!InlineArrayCopy) return NULL; 374 break; 375 case vmIntrinsics::_copyMemory: 376 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 377 if (!InlineArrayCopy) return NULL; 378 break; 379 case vmIntrinsics::_hashCode: 380 if (!InlineObjectHash) return NULL; 381 break; 382 case vmIntrinsics::_clone: 383 case vmIntrinsics::_copyOf: 384 case vmIntrinsics::_copyOfRange: 385 if (!InlineObjectCopy) return NULL; 386 // These also use the arraycopy intrinsic mechanism: 387 if (!InlineArrayCopy) return NULL; 388 break; 389 case vmIntrinsics::_encodeISOArray: 390 if (!SpecialEncodeISOArray) return NULL; 391 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 392 break; 393 case vmIntrinsics::_checkIndex: 394 // We do not intrinsify this. The optimizer does fine with it. 395 return NULL; 396 397 case vmIntrinsics::_getCallerClass: 398 if (!UseNewReflection) return NULL; 399 if (!InlineReflectionGetCallerClass) return NULL; 400 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 401 break; 402 403 case vmIntrinsics::_bitCount_i: 404 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 405 break; 406 407 case vmIntrinsics::_bitCount_l: 408 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 409 break; 410 411 case vmIntrinsics::_numberOfLeadingZeros_i: 412 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 413 break; 414 415 case vmIntrinsics::_numberOfLeadingZeros_l: 416 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 417 break; 418 419 case vmIntrinsics::_numberOfTrailingZeros_i: 420 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 421 break; 422 423 case vmIntrinsics::_numberOfTrailingZeros_l: 424 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 425 break; 426 427 case vmIntrinsics::_reverseBytes_c: 428 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 429 break; 430 case vmIntrinsics::_reverseBytes_s: 431 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 432 break; 433 case vmIntrinsics::_reverseBytes_i: 434 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 435 break; 436 case vmIntrinsics::_reverseBytes_l: 437 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 438 break; 439 440 case vmIntrinsics::_Reference_get: 441 // Use the intrinsic version of Reference.get() so that the value in 442 // the referent field can be registered by the G1 pre-barrier code. 443 // Also add memory barrier to prevent commoning reads from this field 444 // across safepoint since GC can change it value. 445 break; 446 447 case vmIntrinsics::_compareAndSwapObject: 448 #ifdef _LP64 449 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 450 #endif 451 break; 452 453 case vmIntrinsics::_compareAndSwapLong: 454 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 455 break; 456 457 case vmIntrinsics::_getAndAddInt: 458 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 459 break; 460 461 case vmIntrinsics::_getAndAddLong: 462 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 463 break; 464 465 case vmIntrinsics::_getAndSetInt: 466 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 467 break; 468 469 case vmIntrinsics::_getAndSetLong: 470 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 471 break; 472 473 case vmIntrinsics::_getAndSetObject: 474 #ifdef _LP64 475 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 476 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 477 break; 478 #else 479 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 480 break; 481 #endif 482 483 case vmIntrinsics::_aescrypt_encryptBlock: 484 case vmIntrinsics::_aescrypt_decryptBlock: 485 if (!UseAESIntrinsics) return NULL; 486 break; 487 488 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 489 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 490 if (!UseAESIntrinsics) return NULL; 491 // these two require the predicated logic 492 is_predicted = true; 493 break; 494 495 case vmIntrinsics::_updateCRC32: 496 case vmIntrinsics::_updateBytesCRC32: 497 case vmIntrinsics::_updateByteBufferCRC32: 498 if (!UseCRC32Intrinsics) return NULL; 499 break; 500 501 default: 502 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 503 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 504 break; 505 } 506 507 // -XX:-InlineClassNatives disables natives from the Class class. 508 // The flag applies to all reflective calls, notably Array.newArray 509 // (visible to Java programmers as Array.newInstance). 510 if (m->holder()->name() == ciSymbol::java_lang_Class() || 511 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 512 if (!InlineClassNatives) return NULL; 513 } 514 515 // -XX:-InlineThreadNatives disables natives from the Thread class. 516 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 517 if (!InlineThreadNatives) return NULL; 518 } 519 520 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 521 if (m->holder()->name() == ciSymbol::java_lang_Math() || 522 m->holder()->name() == ciSymbol::java_lang_Float() || 523 m->holder()->name() == ciSymbol::java_lang_Double()) { 524 if (!InlineMathNatives) return NULL; 525 } 526 527 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 528 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 529 if (!InlineUnsafeOps) return NULL; 530 } 531 532 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id); 533 } 534 535 //----------------------register_library_intrinsics----------------------- 536 // Initialize this file's data structures, for each Compile instance. 537 void Compile::register_library_intrinsics() { 538 // Nothing to do here. 539 } 540 541 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 542 LibraryCallKit kit(jvms, this); 543 Compile* C = kit.C; 544 int nodes = C->unique(); 545 #ifndef PRODUCT 546 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 547 char buf[1000]; 548 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 549 tty->print_cr("Intrinsic %s", str); 550 } 551 #endif 552 ciMethod* callee = kit.callee(); 553 const int bci = kit.bci(); 554 555 // Try to inline the intrinsic. 556 if (kit.try_to_inline()) { 557 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 558 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 559 } 560 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 561 if (C->log()) { 562 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 563 vmIntrinsics::name_at(intrinsic_id()), 564 (is_virtual() ? " virtual='1'" : ""), 565 C->unique() - nodes); 566 } 567 // Push the result from the inlined method onto the stack. 568 kit.push_result(); 569 return kit.transfer_exceptions_into_jvms(); 570 } 571 572 // The intrinsic bailed out 573 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 574 if (jvms->has_method()) { 575 // Not a root compile. 576 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 577 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 578 } else { 579 // Root compile 580 tty->print("Did not generate intrinsic %s%s at bci:%d in", 581 vmIntrinsics::name_at(intrinsic_id()), 582 (is_virtual() ? " (virtual)" : ""), bci); 583 } 584 } 585 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 586 return NULL; 587 } 588 589 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) { 590 LibraryCallKit kit(jvms, this); 591 Compile* C = kit.C; 592 int nodes = C->unique(); 593 #ifndef PRODUCT 594 assert(is_predicted(), "sanity"); 595 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 596 char buf[1000]; 597 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 598 tty->print_cr("Predicate for intrinsic %s", str); 599 } 600 #endif 601 ciMethod* callee = kit.callee(); 602 const int bci = kit.bci(); 603 604 Node* slow_ctl = kit.try_to_predicate(); 605 if (!kit.failing()) { 606 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 607 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 608 } 609 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 610 if (C->log()) { 611 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 612 vmIntrinsics::name_at(intrinsic_id()), 613 (is_virtual() ? " virtual='1'" : ""), 614 C->unique() - nodes); 615 } 616 return slow_ctl; // Could be NULL if the check folds. 617 } 618 619 // The intrinsic bailed out 620 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 621 if (jvms->has_method()) { 622 // Not a root compile. 623 const char* msg = "failed to generate predicate for intrinsic"; 624 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 625 } else { 626 // Root compile 627 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 628 vmIntrinsics::name_at(intrinsic_id()), 629 (is_virtual() ? " (virtual)" : ""), bci); 630 } 631 } 632 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 633 return NULL; 634 } 635 636 bool LibraryCallKit::try_to_inline() { 637 // Handle symbolic names for otherwise undistinguished boolean switches: 638 const bool is_store = true; 639 const bool is_native_ptr = true; 640 const bool is_static = true; 641 const bool is_volatile = true; 642 643 if (!jvms()->has_method()) { 644 // Root JVMState has a null method. 645 assert(map()->memory()->Opcode() == Op_Parm, ""); 646 // Insert the memory aliasing node 647 set_all_memory(reset_memory()); 648 } 649 assert(merged_memory(), ""); 650 651 652 switch (intrinsic_id()) { 653 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 654 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 655 case vmIntrinsics::_getClass: return inline_native_getClass(); 656 657 case vmIntrinsics::_dsin: 658 case vmIntrinsics::_dcos: 659 case vmIntrinsics::_dtan: 660 case vmIntrinsics::_dabs: 661 case vmIntrinsics::_datan2: 662 case vmIntrinsics::_dsqrt: 663 case vmIntrinsics::_dexp: 664 case vmIntrinsics::_dlog: 665 case vmIntrinsics::_dlog10: 666 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 667 668 case vmIntrinsics::_min: 669 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 670 671 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 672 673 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 674 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 675 case vmIntrinsics::_equals: return inline_string_equals(); 676 677 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 678 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 679 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 680 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 681 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 682 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 683 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 684 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 685 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 686 687 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 688 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 689 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 690 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 691 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 692 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 693 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 694 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 695 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 696 697 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 698 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 699 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 700 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 701 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 702 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 703 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 704 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 705 706 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 707 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 708 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 709 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 710 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 711 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 712 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 713 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 714 715 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 716 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 717 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 718 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 719 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 720 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 721 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 722 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 723 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 724 725 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 726 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 727 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 728 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 729 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 730 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 731 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 732 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 733 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 734 735 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 736 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 737 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 738 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 739 740 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 741 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 742 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 743 744 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 745 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 746 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 747 748 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 749 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 750 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 751 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 752 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 753 754 case vmIntrinsics::_loadFence: 755 case vmIntrinsics::_storeFence: 756 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 757 758 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 759 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 760 761 #ifdef TRACE_HAVE_INTRINSICS 762 case vmIntrinsics::_classID: return inline_native_classID(); 763 case vmIntrinsics::_threadID: return inline_native_threadID(); 764 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 765 #endif 766 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 767 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 768 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 769 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 770 case vmIntrinsics::_newArray: return inline_native_newArray(); 771 case vmIntrinsics::_getLength: return inline_native_getLength(); 772 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 773 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 774 case vmIntrinsics::_equalsC: return inline_array_equals(); 775 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 776 777 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 778 779 case vmIntrinsics::_isInstance: 780 case vmIntrinsics::_getModifiers: 781 case vmIntrinsics::_isInterface: 782 case vmIntrinsics::_isArray: 783 case vmIntrinsics::_isPrimitive: 784 case vmIntrinsics::_getSuperclass: 785 case vmIntrinsics::_getComponentType: 786 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 787 788 case vmIntrinsics::_floatToRawIntBits: 789 case vmIntrinsics::_floatToIntBits: 790 case vmIntrinsics::_intBitsToFloat: 791 case vmIntrinsics::_doubleToRawLongBits: 792 case vmIntrinsics::_doubleToLongBits: 793 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 794 795 case vmIntrinsics::_numberOfLeadingZeros_i: 796 case vmIntrinsics::_numberOfLeadingZeros_l: 797 case vmIntrinsics::_numberOfTrailingZeros_i: 798 case vmIntrinsics::_numberOfTrailingZeros_l: 799 case vmIntrinsics::_bitCount_i: 800 case vmIntrinsics::_bitCount_l: 801 case vmIntrinsics::_reverseBytes_i: 802 case vmIntrinsics::_reverseBytes_l: 803 case vmIntrinsics::_reverseBytes_s: 804 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 805 806 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 807 808 case vmIntrinsics::_Reference_get: return inline_reference_get(); 809 810 case vmIntrinsics::_aescrypt_encryptBlock: 811 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 812 813 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 814 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 815 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 816 817 case vmIntrinsics::_encodeISOArray: 818 return inline_encodeISOArray(); 819 820 case vmIntrinsics::_updateCRC32: 821 return inline_updateCRC32(); 822 case vmIntrinsics::_updateBytesCRC32: 823 return inline_updateBytesCRC32(); 824 case vmIntrinsics::_updateByteBufferCRC32: 825 return inline_updateByteBufferCRC32(); 826 827 default: 828 // If you get here, it may be that someone has added a new intrinsic 829 // to the list in vmSymbols.hpp without implementing it here. 830 #ifndef PRODUCT 831 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 832 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 833 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 834 } 835 #endif 836 return false; 837 } 838 } 839 840 Node* LibraryCallKit::try_to_predicate() { 841 if (!jvms()->has_method()) { 842 // Root JVMState has a null method. 843 assert(map()->memory()->Opcode() == Op_Parm, ""); 844 // Insert the memory aliasing node 845 set_all_memory(reset_memory()); 846 } 847 assert(merged_memory(), ""); 848 849 switch (intrinsic_id()) { 850 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 851 return inline_cipherBlockChaining_AESCrypt_predicate(false); 852 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 853 return inline_cipherBlockChaining_AESCrypt_predicate(true); 854 855 default: 856 // If you get here, it may be that someone has added a new intrinsic 857 // to the list in vmSymbols.hpp without implementing it here. 858 #ifndef PRODUCT 859 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 860 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 861 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 862 } 863 #endif 864 Node* slow_ctl = control(); 865 set_control(top()); // No fast path instrinsic 866 return slow_ctl; 867 } 868 } 869 870 //------------------------------set_result------------------------------- 871 // Helper function for finishing intrinsics. 872 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 873 record_for_igvn(region); 874 set_control(_gvn.transform(region)); 875 set_result( _gvn.transform(value)); 876 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 877 } 878 879 //------------------------------generate_guard--------------------------- 880 // Helper function for generating guarded fast-slow graph structures. 881 // The given 'test', if true, guards a slow path. If the test fails 882 // then a fast path can be taken. (We generally hope it fails.) 883 // In all cases, GraphKit::control() is updated to the fast path. 884 // The returned value represents the control for the slow path. 885 // The return value is never 'top'; it is either a valid control 886 // or NULL if it is obvious that the slow path can never be taken. 887 // Also, if region and the slow control are not NULL, the slow edge 888 // is appended to the region. 889 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 890 if (stopped()) { 891 // Already short circuited. 892 return NULL; 893 } 894 895 // Build an if node and its projections. 896 // If test is true we take the slow path, which we assume is uncommon. 897 if (_gvn.type(test) == TypeInt::ZERO) { 898 // The slow branch is never taken. No need to build this guard. 899 return NULL; 900 } 901 902 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 903 904 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff)); 905 if (if_slow == top()) { 906 // The slow branch is never taken. No need to build this guard. 907 return NULL; 908 } 909 910 if (region != NULL) 911 region->add_req(if_slow); 912 913 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff)); 914 set_control(if_fast); 915 916 return if_slow; 917 } 918 919 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 920 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 921 } 922 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 923 return generate_guard(test, region, PROB_FAIR); 924 } 925 926 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 927 Node* *pos_index) { 928 if (stopped()) 929 return NULL; // already stopped 930 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 931 return NULL; // index is already adequately typed 932 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0))); 933 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 934 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 935 if (is_neg != NULL && pos_index != NULL) { 936 // Emulate effect of Parse::adjust_map_after_if. 937 Node* ccast = new (C) CastIINode(index, TypeInt::POS); 938 ccast->set_req(0, control()); 939 (*pos_index) = _gvn.transform(ccast); 940 } 941 return is_neg; 942 } 943 944 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 945 Node* *pos_index) { 946 if (stopped()) 947 return NULL; // already stopped 948 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 949 return NULL; // index is already adequately typed 950 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0))); 951 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 952 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq)); 953 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 954 if (is_notp != NULL && pos_index != NULL) { 955 // Emulate effect of Parse::adjust_map_after_if. 956 Node* ccast = new (C) CastIINode(index, TypeInt::POS1); 957 ccast->set_req(0, control()); 958 (*pos_index) = _gvn.transform(ccast); 959 } 960 return is_notp; 961 } 962 963 // Make sure that 'position' is a valid limit index, in [0..length]. 964 // There are two equivalent plans for checking this: 965 // A. (offset + copyLength) unsigned<= arrayLength 966 // B. offset <= (arrayLength - copyLength) 967 // We require that all of the values above, except for the sum and 968 // difference, are already known to be non-negative. 969 // Plan A is robust in the face of overflow, if offset and copyLength 970 // are both hugely positive. 971 // 972 // Plan B is less direct and intuitive, but it does not overflow at 973 // all, since the difference of two non-negatives is always 974 // representable. Whenever Java methods must perform the equivalent 975 // check they generally use Plan B instead of Plan A. 976 // For the moment we use Plan A. 977 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 978 Node* subseq_length, 979 Node* array_length, 980 RegionNode* region) { 981 if (stopped()) 982 return NULL; // already stopped 983 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 984 if (zero_offset && subseq_length->eqv_uncast(array_length)) 985 return NULL; // common case of whole-array copy 986 Node* last = subseq_length; 987 if (!zero_offset) // last += offset 988 last = _gvn.transform(new (C) AddINode(last, offset)); 989 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last)); 990 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 991 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 992 return is_over; 993 } 994 995 996 //--------------------------generate_current_thread-------------------- 997 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 998 ciKlass* thread_klass = env()->Thread_klass(); 999 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1000 Node* thread = _gvn.transform(new (C) ThreadLocalNode()); 1001 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1002 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 1003 tls_output = thread; 1004 return threadObj; 1005 } 1006 1007 1008 //------------------------------make_string_method_node------------------------ 1009 // Helper method for String intrinsic functions. This version is called 1010 // with str1 and str2 pointing to String object nodes. 1011 // 1012 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1013 Node* no_ctrl = NULL; 1014 1015 // Get start addr of string 1016 Node* str1_value = load_String_value(no_ctrl, str1); 1017 Node* str1_offset = load_String_offset(no_ctrl, str1); 1018 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1019 1020 // Get length of string 1 1021 Node* str1_len = load_String_length(no_ctrl, str1); 1022 1023 Node* str2_value = load_String_value(no_ctrl, str2); 1024 Node* str2_offset = load_String_offset(no_ctrl, str2); 1025 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1026 1027 Node* str2_len = NULL; 1028 Node* result = NULL; 1029 1030 switch (opcode) { 1031 case Op_StrIndexOf: 1032 // Get length of string 2 1033 str2_len = load_String_length(no_ctrl, str2); 1034 1035 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1036 str1_start, str1_len, str2_start, str2_len); 1037 break; 1038 case Op_StrComp: 1039 // Get length of string 2 1040 str2_len = load_String_length(no_ctrl, str2); 1041 1042 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1043 str1_start, str1_len, str2_start, str2_len); 1044 break; 1045 case Op_StrEquals: 1046 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1047 str1_start, str2_start, str1_len); 1048 break; 1049 default: 1050 ShouldNotReachHere(); 1051 return NULL; 1052 } 1053 1054 // All these intrinsics have checks. 1055 C->set_has_split_ifs(true); // Has chance for split-if optimization 1056 1057 return _gvn.transform(result); 1058 } 1059 1060 // Helper method for String intrinsic functions. This version is called 1061 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1062 // to Int nodes containing the lenghts of str1 and str2. 1063 // 1064 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1065 Node* result = NULL; 1066 switch (opcode) { 1067 case Op_StrIndexOf: 1068 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1069 str1_start, cnt1, str2_start, cnt2); 1070 break; 1071 case Op_StrComp: 1072 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1073 str1_start, cnt1, str2_start, cnt2); 1074 break; 1075 case Op_StrEquals: 1076 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1077 str1_start, str2_start, cnt1); 1078 break; 1079 default: 1080 ShouldNotReachHere(); 1081 return NULL; 1082 } 1083 1084 // All these intrinsics have checks. 1085 C->set_has_split_ifs(true); // Has chance for split-if optimization 1086 1087 return _gvn.transform(result); 1088 } 1089 1090 //------------------------------inline_string_compareTo------------------------ 1091 // public int java.lang.String.compareTo(String anotherString); 1092 bool LibraryCallKit::inline_string_compareTo() { 1093 Node* receiver = null_check(argument(0)); 1094 Node* arg = null_check(argument(1)); 1095 if (stopped()) { 1096 return true; 1097 } 1098 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1099 return true; 1100 } 1101 1102 //------------------------------inline_string_equals------------------------ 1103 bool LibraryCallKit::inline_string_equals() { 1104 Node* receiver = null_check_receiver(); 1105 // NOTE: Do not null check argument for String.equals() because spec 1106 // allows to specify NULL as argument. 1107 Node* argument = this->argument(1); 1108 if (stopped()) { 1109 return true; 1110 } 1111 1112 // paths (plus control) merge 1113 RegionNode* region = new (C) RegionNode(5); 1114 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 1115 1116 // does source == target string? 1117 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument)); 1118 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 1119 1120 Node* if_eq = generate_slow_guard(bol, NULL); 1121 if (if_eq != NULL) { 1122 // receiver == argument 1123 phi->init_req(2, intcon(1)); 1124 region->init_req(2, if_eq); 1125 } 1126 1127 // get String klass for instanceOf 1128 ciInstanceKlass* klass = env()->String_klass(); 1129 1130 if (!stopped()) { 1131 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1132 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1))); 1133 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 1134 1135 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1136 //instanceOf == true, fallthrough 1137 1138 if (inst_false != NULL) { 1139 phi->init_req(3, intcon(0)); 1140 region->init_req(3, inst_false); 1141 } 1142 } 1143 1144 if (!stopped()) { 1145 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1146 1147 // Properly cast the argument to String 1148 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type)); 1149 // This path is taken only when argument's type is String:NotNull. 1150 argument = cast_not_null(argument, false); 1151 1152 Node* no_ctrl = NULL; 1153 1154 // Get start addr of receiver 1155 Node* receiver_val = load_String_value(no_ctrl, receiver); 1156 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1157 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1158 1159 // Get length of receiver 1160 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1161 1162 // Get start addr of argument 1163 Node* argument_val = load_String_value(no_ctrl, argument); 1164 Node* argument_offset = load_String_offset(no_ctrl, argument); 1165 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1166 1167 // Get length of argument 1168 Node* argument_cnt = load_String_length(no_ctrl, argument); 1169 1170 // Check for receiver count != argument count 1171 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt)); 1172 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne)); 1173 Node* if_ne = generate_slow_guard(bol, NULL); 1174 if (if_ne != NULL) { 1175 phi->init_req(4, intcon(0)); 1176 region->init_req(4, if_ne); 1177 } 1178 1179 // Check for count == 0 is done by assembler code for StrEquals. 1180 1181 if (!stopped()) { 1182 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1183 phi->init_req(1, equals); 1184 region->init_req(1, control()); 1185 } 1186 } 1187 1188 // post merge 1189 set_control(_gvn.transform(region)); 1190 record_for_igvn(region); 1191 1192 set_result(_gvn.transform(phi)); 1193 return true; 1194 } 1195 1196 //------------------------------inline_array_equals---------------------------- 1197 bool LibraryCallKit::inline_array_equals() { 1198 Node* arg1 = argument(0); 1199 Node* arg2 = argument(1); 1200 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1201 return true; 1202 } 1203 1204 // Java version of String.indexOf(constant string) 1205 // class StringDecl { 1206 // StringDecl(char[] ca) { 1207 // offset = 0; 1208 // count = ca.length; 1209 // value = ca; 1210 // } 1211 // int offset; 1212 // int count; 1213 // char[] value; 1214 // } 1215 // 1216 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1217 // int targetOffset, int cache_i, int md2) { 1218 // int cache = cache_i; 1219 // int sourceOffset = string_object.offset; 1220 // int sourceCount = string_object.count; 1221 // int targetCount = target_object.length; 1222 // 1223 // int targetCountLess1 = targetCount - 1; 1224 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1225 // 1226 // char[] source = string_object.value; 1227 // char[] target = target_object; 1228 // int lastChar = target[targetCountLess1]; 1229 // 1230 // outer_loop: 1231 // for (int i = sourceOffset; i < sourceEnd; ) { 1232 // int src = source[i + targetCountLess1]; 1233 // if (src == lastChar) { 1234 // // With random strings and a 4-character alphabet, 1235 // // reverse matching at this point sets up 0.8% fewer 1236 // // frames, but (paradoxically) makes 0.3% more probes. 1237 // // Since those probes are nearer the lastChar probe, 1238 // // there is may be a net D$ win with reverse matching. 1239 // // But, reversing loop inhibits unroll of inner loop 1240 // // for unknown reason. So, does running outer loop from 1241 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1242 // for (int j = 0; j < targetCountLess1; j++) { 1243 // if (target[targetOffset + j] != source[i+j]) { 1244 // if ((cache & (1 << source[i+j])) == 0) { 1245 // if (md2 < j+1) { 1246 // i += j+1; 1247 // continue outer_loop; 1248 // } 1249 // } 1250 // i += md2; 1251 // continue outer_loop; 1252 // } 1253 // } 1254 // return i - sourceOffset; 1255 // } 1256 // if ((cache & (1 << src)) == 0) { 1257 // i += targetCountLess1; 1258 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1259 // i++; 1260 // } 1261 // return -1; 1262 // } 1263 1264 //------------------------------string_indexOf------------------------ 1265 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1266 jint cache_i, jint md2_i) { 1267 1268 Node* no_ctrl = NULL; 1269 float likely = PROB_LIKELY(0.9); 1270 float unlikely = PROB_UNLIKELY(0.9); 1271 1272 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1273 1274 Node* source = load_String_value(no_ctrl, string_object); 1275 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1276 Node* sourceCount = load_String_length(no_ctrl, string_object); 1277 1278 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1279 jint target_length = target_array->length(); 1280 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1281 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1282 1283 IdealKit kit(this, false, true); 1284 #define __ kit. 1285 Node* zero = __ ConI(0); 1286 Node* one = __ ConI(1); 1287 Node* cache = __ ConI(cache_i); 1288 Node* md2 = __ ConI(md2_i); 1289 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1290 Node* targetCount = __ ConI(target_length); 1291 Node* targetCountLess1 = __ ConI(target_length - 1); 1292 Node* targetOffset = __ ConI(targetOffset_i); 1293 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1294 1295 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1296 Node* outer_loop = __ make_label(2 /* goto */); 1297 Node* return_ = __ make_label(1); 1298 1299 __ set(rtn,__ ConI(-1)); 1300 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1301 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1302 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1303 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1304 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1305 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1306 Node* tpj = __ AddI(targetOffset, __ value(j)); 1307 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1308 Node* ipj = __ AddI(__ value(i), __ value(j)); 1309 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1310 __ if_then(targ, BoolTest::ne, src2); { 1311 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1312 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1313 __ increment(i, __ AddI(__ value(j), one)); 1314 __ goto_(outer_loop); 1315 } __ end_if(); __ dead(j); 1316 }__ end_if(); __ dead(j); 1317 __ increment(i, md2); 1318 __ goto_(outer_loop); 1319 }__ end_if(); 1320 __ increment(j, one); 1321 }__ end_loop(); __ dead(j); 1322 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1323 __ goto_(return_); 1324 }__ end_if(); 1325 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1326 __ increment(i, targetCountLess1); 1327 }__ end_if(); 1328 __ increment(i, one); 1329 __ bind(outer_loop); 1330 }__ end_loop(); __ dead(i); 1331 __ bind(return_); 1332 1333 // Final sync IdealKit and GraphKit. 1334 final_sync(kit); 1335 Node* result = __ value(rtn); 1336 #undef __ 1337 C->set_has_loops(true); 1338 return result; 1339 } 1340 1341 //------------------------------inline_string_indexOf------------------------ 1342 bool LibraryCallKit::inline_string_indexOf() { 1343 Node* receiver = argument(0); 1344 Node* arg = argument(1); 1345 1346 Node* result; 1347 // Disable the use of pcmpestri until it can be guaranteed that 1348 // the load doesn't cross into the uncommited space. 1349 if (Matcher::has_match_rule(Op_StrIndexOf) && 1350 UseSSE42Intrinsics) { 1351 // Generate SSE4.2 version of indexOf 1352 // We currently only have match rules that use SSE4.2 1353 1354 receiver = null_check(receiver); 1355 arg = null_check(arg); 1356 if (stopped()) { 1357 return true; 1358 } 1359 1360 ciInstanceKlass* str_klass = env()->String_klass(); 1361 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1362 1363 // Make the merge point 1364 RegionNode* result_rgn = new (C) RegionNode(4); 1365 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT); 1366 Node* no_ctrl = NULL; 1367 1368 // Get start addr of source string 1369 Node* source = load_String_value(no_ctrl, receiver); 1370 Node* source_offset = load_String_offset(no_ctrl, receiver); 1371 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1372 1373 // Get length of source string 1374 Node* source_cnt = load_String_length(no_ctrl, receiver); 1375 1376 // Get start addr of substring 1377 Node* substr = load_String_value(no_ctrl, arg); 1378 Node* substr_offset = load_String_offset(no_ctrl, arg); 1379 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1380 1381 // Get length of source string 1382 Node* substr_cnt = load_String_length(no_ctrl, arg); 1383 1384 // Check for substr count > string count 1385 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt)); 1386 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt)); 1387 Node* if_gt = generate_slow_guard(bol, NULL); 1388 if (if_gt != NULL) { 1389 result_phi->init_req(2, intcon(-1)); 1390 result_rgn->init_req(2, if_gt); 1391 } 1392 1393 if (!stopped()) { 1394 // Check for substr count == 0 1395 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0))); 1396 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 1397 Node* if_zero = generate_slow_guard(bol, NULL); 1398 if (if_zero != NULL) { 1399 result_phi->init_req(3, intcon(0)); 1400 result_rgn->init_req(3, if_zero); 1401 } 1402 } 1403 1404 if (!stopped()) { 1405 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1406 result_phi->init_req(1, result); 1407 result_rgn->init_req(1, control()); 1408 } 1409 set_control(_gvn.transform(result_rgn)); 1410 record_for_igvn(result_rgn); 1411 result = _gvn.transform(result_phi); 1412 1413 } else { // Use LibraryCallKit::string_indexOf 1414 // don't intrinsify if argument isn't a constant string. 1415 if (!arg->is_Con()) { 1416 return false; 1417 } 1418 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1419 if (str_type == NULL) { 1420 return false; 1421 } 1422 ciInstanceKlass* klass = env()->String_klass(); 1423 ciObject* str_const = str_type->const_oop(); 1424 if (str_const == NULL || str_const->klass() != klass) { 1425 return false; 1426 } 1427 ciInstance* str = str_const->as_instance(); 1428 assert(str != NULL, "must be instance"); 1429 1430 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1431 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1432 1433 int o; 1434 int c; 1435 if (java_lang_String::has_offset_field()) { 1436 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1437 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1438 } else { 1439 o = 0; 1440 c = pat->length(); 1441 } 1442 1443 // constant strings have no offset and count == length which 1444 // simplifies the resulting code somewhat so lets optimize for that. 1445 if (o != 0 || c != pat->length()) { 1446 return false; 1447 } 1448 1449 receiver = null_check(receiver, T_OBJECT); 1450 // NOTE: No null check on the argument is needed since it's a constant String oop. 1451 if (stopped()) { 1452 return true; 1453 } 1454 1455 // The null string as a pattern always returns 0 (match at beginning of string) 1456 if (c == 0) { 1457 set_result(intcon(0)); 1458 return true; 1459 } 1460 1461 // Generate default indexOf 1462 jchar lastChar = pat->char_at(o + (c - 1)); 1463 int cache = 0; 1464 int i; 1465 for (i = 0; i < c - 1; i++) { 1466 assert(i < pat->length(), "out of range"); 1467 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1468 } 1469 1470 int md2 = c; 1471 for (i = 0; i < c - 1; i++) { 1472 assert(i < pat->length(), "out of range"); 1473 if (pat->char_at(o + i) == lastChar) { 1474 md2 = (c - 1) - i; 1475 } 1476 } 1477 1478 result = string_indexOf(receiver, pat, o, cache, md2); 1479 } 1480 set_result(result); 1481 return true; 1482 } 1483 1484 //--------------------------round_double_node-------------------------------- 1485 // Round a double node if necessary. 1486 Node* LibraryCallKit::round_double_node(Node* n) { 1487 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1488 n = _gvn.transform(new (C) RoundDoubleNode(0, n)); 1489 return n; 1490 } 1491 1492 //------------------------------inline_math----------------------------------- 1493 // public static double Math.abs(double) 1494 // public static double Math.sqrt(double) 1495 // public static double Math.log(double) 1496 // public static double Math.log10(double) 1497 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1498 Node* arg = round_double_node(argument(0)); 1499 Node* n; 1500 switch (id) { 1501 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break; 1502 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break; 1503 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break; 1504 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break; 1505 default: fatal_unexpected_iid(id); break; 1506 } 1507 set_result(_gvn.transform(n)); 1508 return true; 1509 } 1510 1511 //------------------------------inline_trig---------------------------------- 1512 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1513 // argument reduction which will turn into a fast/slow diamond. 1514 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1515 Node* arg = round_double_node(argument(0)); 1516 Node* n = NULL; 1517 1518 switch (id) { 1519 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break; 1520 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break; 1521 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break; 1522 default: fatal_unexpected_iid(id); break; 1523 } 1524 n = _gvn.transform(n); 1525 1526 // Rounding required? Check for argument reduction! 1527 if (Matcher::strict_fp_requires_explicit_rounding) { 1528 static const double pi_4 = 0.7853981633974483; 1529 static const double neg_pi_4 = -0.7853981633974483; 1530 // pi/2 in 80-bit extended precision 1531 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1532 // -pi/2 in 80-bit extended precision 1533 // 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}; 1534 // Cutoff value for using this argument reduction technique 1535 //static const double pi_2_minus_epsilon = 1.564660403643354; 1536 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1537 1538 // Pseudocode for sin: 1539 // if (x <= Math.PI / 4.0) { 1540 // if (x >= -Math.PI / 4.0) return fsin(x); 1541 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1542 // } else { 1543 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1544 // } 1545 // return StrictMath.sin(x); 1546 1547 // Pseudocode for cos: 1548 // if (x <= Math.PI / 4.0) { 1549 // if (x >= -Math.PI / 4.0) return fcos(x); 1550 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1551 // } else { 1552 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1553 // } 1554 // return StrictMath.cos(x); 1555 1556 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1557 // requires a special machine instruction to load it. Instead we'll try 1558 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1559 // probably do the math inside the SIN encoding. 1560 1561 // Make the merge point 1562 RegionNode* r = new (C) RegionNode(3); 1563 Node* phi = new (C) PhiNode(r, Type::DOUBLE); 1564 1565 // Flatten arg so we need only 1 test 1566 Node *abs = _gvn.transform(new (C) AbsDNode(arg)); 1567 // Node for PI/4 constant 1568 Node *pi4 = makecon(TypeD::make(pi_4)); 1569 // Check PI/4 : abs(arg) 1570 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs)); 1571 // Check: If PI/4 < abs(arg) then go slow 1572 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt )); 1573 // Branch either way 1574 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1575 set_control(opt_iff(r,iff)); 1576 1577 // Set fast path result 1578 phi->init_req(2, n); 1579 1580 // Slow path - non-blocking leaf call 1581 Node* call = NULL; 1582 switch (id) { 1583 case vmIntrinsics::_dsin: 1584 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1585 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1586 "Sin", NULL, arg, top()); 1587 break; 1588 case vmIntrinsics::_dcos: 1589 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1590 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1591 "Cos", NULL, arg, top()); 1592 break; 1593 case vmIntrinsics::_dtan: 1594 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1595 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1596 "Tan", NULL, arg, top()); 1597 break; 1598 } 1599 assert(control()->in(0) == call, ""); 1600 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1601 r->init_req(1, control()); 1602 phi->init_req(1, slow_result); 1603 1604 // Post-merge 1605 set_control(_gvn.transform(r)); 1606 record_for_igvn(r); 1607 n = _gvn.transform(phi); 1608 1609 C->set_has_split_ifs(true); // Has chance for split-if optimization 1610 } 1611 set_result(n); 1612 return true; 1613 } 1614 1615 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1616 //------------------- 1617 //result=(result.isNaN())? funcAddr():result; 1618 // Check: If isNaN() by checking result!=result? then either trap 1619 // or go to runtime 1620 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result)); 1621 // Build the boolean node 1622 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq)); 1623 1624 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1625 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1626 // The pow or exp intrinsic returned a NaN, which requires a call 1627 // to the runtime. Recompile with the runtime call. 1628 uncommon_trap(Deoptimization::Reason_intrinsic, 1629 Deoptimization::Action_make_not_entrant); 1630 } 1631 set_result(result); 1632 } else { 1633 // If this inlining ever returned NaN in the past, we compile a call 1634 // to the runtime to properly handle corner cases 1635 1636 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1637 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff)); 1638 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff)); 1639 1640 if (!if_slow->is_top()) { 1641 RegionNode* result_region = new (C) RegionNode(3); 1642 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE); 1643 1644 result_region->init_req(1, if_fast); 1645 result_val->init_req(1, result); 1646 1647 set_control(if_slow); 1648 1649 const TypePtr* no_memory_effects = NULL; 1650 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1651 no_memory_effects, 1652 x, top(), y, y ? top() : NULL); 1653 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0)); 1654 #ifdef ASSERT 1655 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1)); 1656 assert(value_top == top(), "second value must be top"); 1657 #endif 1658 1659 result_region->init_req(2, control()); 1660 result_val->init_req(2, value); 1661 set_result(result_region, result_val); 1662 } else { 1663 set_result(result); 1664 } 1665 } 1666 } 1667 1668 //------------------------------inline_exp------------------------------------- 1669 // Inline exp instructions, if possible. The Intel hardware only misses 1670 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1671 bool LibraryCallKit::inline_exp() { 1672 Node* arg = round_double_node(argument(0)); 1673 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg)); 1674 1675 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1676 1677 C->set_has_split_ifs(true); // Has chance for split-if optimization 1678 return true; 1679 } 1680 1681 //------------------------------inline_pow------------------------------------- 1682 // Inline power instructions, if possible. 1683 bool LibraryCallKit::inline_pow() { 1684 // Pseudocode for pow 1685 // if (x <= 0.0) { 1686 // long longy = (long)y; 1687 // if ((double)longy == y) { // if y is long 1688 // if (y + 1 == y) longy = 0; // huge number: even 1689 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1690 // } else { 1691 // result = NaN; 1692 // } 1693 // } else { 1694 // result = DPow(x,y); 1695 // } 1696 // if (result != result)? { 1697 // result = uncommon_trap() or runtime_call(); 1698 // } 1699 // return result; 1700 1701 Node* x = round_double_node(argument(0)); 1702 Node* y = round_double_node(argument(2)); 1703 1704 Node* result = NULL; 1705 1706 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1707 // Short form: skip the fancy tests and just check for NaN result. 1708 result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1709 } else { 1710 // If this inlining ever returned NaN in the past, include all 1711 // checks + call to the runtime. 1712 1713 // Set the merge point for If node with condition of (x <= 0.0) 1714 // There are four possible paths to region node and phi node 1715 RegionNode *r = new (C) RegionNode(4); 1716 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1717 1718 // Build the first if node: if (x <= 0.0) 1719 // Node for 0 constant 1720 Node *zeronode = makecon(TypeD::ZERO); 1721 // Check x:0 1722 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode)); 1723 // Check: If (x<=0) then go complex path 1724 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le )); 1725 // Branch either way 1726 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1727 // Fast path taken; set region slot 3 1728 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1)); 1729 r->init_req(3,fast_taken); // Capture fast-control 1730 1731 // Fast path not-taken, i.e. slow path 1732 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1)); 1733 1734 // Set fast path result 1735 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1736 phi->init_req(3, fast_result); 1737 1738 // Complex path 1739 // Build the second if node (if y is long) 1740 // Node for (long)y 1741 Node *longy = _gvn.transform(new (C) ConvD2LNode(y)); 1742 // Node for (double)((long) y) 1743 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy)); 1744 // Check (double)((long) y) : y 1745 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y)); 1746 // Check if (y isn't long) then go to slow path 1747 1748 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne )); 1749 // Branch either way 1750 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1751 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2)); 1752 1753 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2)); 1754 1755 // Calculate DPow(abs(x), y)*(1 & (long)y) 1756 // Node for constant 1 1757 Node *conone = longcon(1); 1758 // 1& (long)y 1759 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy)); 1760 1761 // A huge number is always even. Detect a huge number by checking 1762 // if y + 1 == y and set integer to be tested for parity to 0. 1763 // Required for corner case: 1764 // (long)9.223372036854776E18 = max_jlong 1765 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1766 // max_jlong is odd but 9.223372036854776E18 is even 1767 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1)))); 1768 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y)); 1769 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq )); 1770 Node* correctedsign = NULL; 1771 if (ConditionalMoveLimit != 0) { 1772 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1773 } else { 1774 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1775 RegionNode *r = new (C) RegionNode(3); 1776 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 1777 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1))); 1778 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1))); 1779 phi->init_req(1, signnode); 1780 phi->init_req(2, longcon(0)); 1781 correctedsign = _gvn.transform(phi); 1782 ylong_path = _gvn.transform(r); 1783 record_for_igvn(r); 1784 } 1785 1786 // zero node 1787 Node *conzero = longcon(0); 1788 // Check (1&(long)y)==0? 1789 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero)); 1790 // Check if (1&(long)y)!=0?, if so the result is negative 1791 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne )); 1792 // abs(x) 1793 Node *absx=_gvn.transform(new (C) AbsDNode(x)); 1794 // abs(x)^y 1795 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y)); 1796 // -abs(x)^y 1797 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy)); 1798 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1799 Node *signresult = NULL; 1800 if (ConditionalMoveLimit != 0) { 1801 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1802 } else { 1803 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1804 RegionNode *r = new (C) RegionNode(3); 1805 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1806 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven))); 1807 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven))); 1808 phi->init_req(1, absxpowy); 1809 phi->init_req(2, negabsxpowy); 1810 signresult = _gvn.transform(phi); 1811 ylong_path = _gvn.transform(r); 1812 record_for_igvn(r); 1813 } 1814 // Set complex path fast result 1815 r->init_req(2, ylong_path); 1816 phi->init_req(2, signresult); 1817 1818 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1819 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1820 r->init_req(1,slow_path); 1821 phi->init_req(1,slow_result); 1822 1823 // Post merge 1824 set_control(_gvn.transform(r)); 1825 record_for_igvn(r); 1826 result = _gvn.transform(phi); 1827 } 1828 1829 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1830 1831 C->set_has_split_ifs(true); // Has chance for split-if optimization 1832 return true; 1833 } 1834 1835 //------------------------------runtime_math----------------------------- 1836 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1837 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1838 "must be (DD)D or (D)D type"); 1839 1840 // Inputs 1841 Node* a = round_double_node(argument(0)); 1842 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1843 1844 const TypePtr* no_memory_effects = NULL; 1845 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1846 no_memory_effects, 1847 a, top(), b, b ? top() : NULL); 1848 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0)); 1849 #ifdef ASSERT 1850 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1)); 1851 assert(value_top == top(), "second value must be top"); 1852 #endif 1853 1854 set_result(value); 1855 return true; 1856 } 1857 1858 //------------------------------inline_math_native----------------------------- 1859 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1860 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1861 switch (id) { 1862 // These intrinsics are not properly supported on all hardware 1863 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1864 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1865 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1866 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1867 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1868 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1869 1870 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1871 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1872 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1873 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1874 1875 // These intrinsics are supported on all hardware 1876 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false; 1877 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1878 1879 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1880 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1881 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1882 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1883 #undef FN_PTR 1884 1885 // These intrinsics are not yet correctly implemented 1886 case vmIntrinsics::_datan2: 1887 return false; 1888 1889 default: 1890 fatal_unexpected_iid(id); 1891 return false; 1892 } 1893 } 1894 1895 static bool is_simple_name(Node* n) { 1896 return (n->req() == 1 // constant 1897 || (n->is_Type() && n->as_Type()->type()->singleton()) 1898 || n->is_Proj() // parameter or return value 1899 || n->is_Phi() // local of some sort 1900 ); 1901 } 1902 1903 //----------------------------inline_min_max----------------------------------- 1904 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1905 set_result(generate_min_max(id, argument(0), argument(1))); 1906 return true; 1907 } 1908 1909 Node* 1910 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1911 // These are the candidate return value: 1912 Node* xvalue = x0; 1913 Node* yvalue = y0; 1914 1915 if (xvalue == yvalue) { 1916 return xvalue; 1917 } 1918 1919 bool want_max = (id == vmIntrinsics::_max); 1920 1921 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1922 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1923 if (txvalue == NULL || tyvalue == NULL) return top(); 1924 // This is not really necessary, but it is consistent with a 1925 // hypothetical MaxINode::Value method: 1926 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1927 1928 // %%% This folding logic should (ideally) be in a different place. 1929 // Some should be inside IfNode, and there to be a more reliable 1930 // transformation of ?: style patterns into cmoves. We also want 1931 // more powerful optimizations around cmove and min/max. 1932 1933 // Try to find a dominating comparison of these guys. 1934 // It can simplify the index computation for Arrays.copyOf 1935 // and similar uses of System.arraycopy. 1936 // First, compute the normalized version of CmpI(x, y). 1937 int cmp_op = Op_CmpI; 1938 Node* xkey = xvalue; 1939 Node* ykey = yvalue; 1940 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey)); 1941 if (ideal_cmpxy->is_Cmp()) { 1942 // E.g., if we have CmpI(length - offset, count), 1943 // it might idealize to CmpI(length, count + offset) 1944 cmp_op = ideal_cmpxy->Opcode(); 1945 xkey = ideal_cmpxy->in(1); 1946 ykey = ideal_cmpxy->in(2); 1947 } 1948 1949 // Start by locating any relevant comparisons. 1950 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1951 Node* cmpxy = NULL; 1952 Node* cmpyx = NULL; 1953 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1954 Node* cmp = start_from->fast_out(k); 1955 if (cmp->outcnt() > 0 && // must have prior uses 1956 cmp->in(0) == NULL && // must be context-independent 1957 cmp->Opcode() == cmp_op) { // right kind of compare 1958 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1959 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1960 } 1961 } 1962 1963 const int NCMPS = 2; 1964 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1965 int cmpn; 1966 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1967 if (cmps[cmpn] != NULL) break; // find a result 1968 } 1969 if (cmpn < NCMPS) { 1970 // Look for a dominating test that tells us the min and max. 1971 int depth = 0; // Limit search depth for speed 1972 Node* dom = control(); 1973 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1974 if (++depth >= 100) break; 1975 Node* ifproj = dom; 1976 if (!ifproj->is_Proj()) continue; 1977 Node* iff = ifproj->in(0); 1978 if (!iff->is_If()) continue; 1979 Node* bol = iff->in(1); 1980 if (!bol->is_Bool()) continue; 1981 Node* cmp = bol->in(1); 1982 if (cmp == NULL) continue; 1983 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1984 if (cmps[cmpn] == cmp) break; 1985 if (cmpn == NCMPS) continue; 1986 BoolTest::mask btest = bol->as_Bool()->_test._test; 1987 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1988 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1989 // At this point, we know that 'x btest y' is true. 1990 switch (btest) { 1991 case BoolTest::eq: 1992 // They are proven equal, so we can collapse the min/max. 1993 // Either value is the answer. Choose the simpler. 1994 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1995 return yvalue; 1996 return xvalue; 1997 case BoolTest::lt: // x < y 1998 case BoolTest::le: // x <= y 1999 return (want_max ? yvalue : xvalue); 2000 case BoolTest::gt: // x > y 2001 case BoolTest::ge: // x >= y 2002 return (want_max ? xvalue : yvalue); 2003 } 2004 } 2005 } 2006 2007 // We failed to find a dominating test. 2008 // Let's pick a test that might GVN with prior tests. 2009 Node* best_bol = NULL; 2010 BoolTest::mask best_btest = BoolTest::illegal; 2011 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2012 Node* cmp = cmps[cmpn]; 2013 if (cmp == NULL) continue; 2014 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2015 Node* bol = cmp->fast_out(j); 2016 if (!bol->is_Bool()) continue; 2017 BoolTest::mask btest = bol->as_Bool()->_test._test; 2018 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2019 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2020 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2021 best_bol = bol->as_Bool(); 2022 best_btest = btest; 2023 } 2024 } 2025 } 2026 2027 Node* answer_if_true = NULL; 2028 Node* answer_if_false = NULL; 2029 switch (best_btest) { 2030 default: 2031 if (cmpxy == NULL) 2032 cmpxy = ideal_cmpxy; 2033 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt)); 2034 // and fall through: 2035 case BoolTest::lt: // x < y 2036 case BoolTest::le: // x <= y 2037 answer_if_true = (want_max ? yvalue : xvalue); 2038 answer_if_false = (want_max ? xvalue : yvalue); 2039 break; 2040 case BoolTest::gt: // x > y 2041 case BoolTest::ge: // x >= y 2042 answer_if_true = (want_max ? xvalue : yvalue); 2043 answer_if_false = (want_max ? yvalue : xvalue); 2044 break; 2045 } 2046 2047 jint hi, lo; 2048 if (want_max) { 2049 // We can sharpen the minimum. 2050 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2051 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2052 } else { 2053 // We can sharpen the maximum. 2054 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2055 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2056 } 2057 2058 // Use a flow-free graph structure, to avoid creating excess control edges 2059 // which could hinder other optimizations. 2060 // Since Math.min/max is often used with arraycopy, we want 2061 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2062 Node* cmov = CMoveNode::make(C, NULL, best_bol, 2063 answer_if_false, answer_if_true, 2064 TypeInt::make(lo, hi, widen)); 2065 2066 return _gvn.transform(cmov); 2067 2068 /* 2069 // This is not as desirable as it may seem, since Min and Max 2070 // nodes do not have a full set of optimizations. 2071 // And they would interfere, anyway, with 'if' optimizations 2072 // and with CMoveI canonical forms. 2073 switch (id) { 2074 case vmIntrinsics::_min: 2075 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2076 case vmIntrinsics::_max: 2077 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2078 default: 2079 ShouldNotReachHere(); 2080 } 2081 */ 2082 } 2083 2084 inline int 2085 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2086 const TypePtr* base_type = TypePtr::NULL_PTR; 2087 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2088 if (base_type == NULL) { 2089 // Unknown type. 2090 return Type::AnyPtr; 2091 } else if (base_type == TypePtr::NULL_PTR) { 2092 // Since this is a NULL+long form, we have to switch to a rawptr. 2093 base = _gvn.transform(new (C) CastX2PNode(offset)); 2094 offset = MakeConX(0); 2095 return Type::RawPtr; 2096 } else if (base_type->base() == Type::RawPtr) { 2097 return Type::RawPtr; 2098 } else if (base_type->isa_oopptr()) { 2099 // Base is never null => always a heap address. 2100 if (base_type->ptr() == TypePtr::NotNull) { 2101 return Type::OopPtr; 2102 } 2103 // Offset is small => always a heap address. 2104 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2105 if (offset_type != NULL && 2106 base_type->offset() == 0 && // (should always be?) 2107 offset_type->_lo >= 0 && 2108 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2109 return Type::OopPtr; 2110 } 2111 // Otherwise, it might either be oop+off or NULL+addr. 2112 return Type::AnyPtr; 2113 } else { 2114 // No information: 2115 return Type::AnyPtr; 2116 } 2117 } 2118 2119 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2120 int kind = classify_unsafe_addr(base, offset); 2121 if (kind == Type::RawPtr) { 2122 return basic_plus_adr(top(), base, offset); 2123 } else { 2124 return basic_plus_adr(base, offset); 2125 } 2126 } 2127 2128 //--------------------------inline_number_methods----------------------------- 2129 // inline int Integer.numberOfLeadingZeros(int) 2130 // inline int Long.numberOfLeadingZeros(long) 2131 // 2132 // inline int Integer.numberOfTrailingZeros(int) 2133 // inline int Long.numberOfTrailingZeros(long) 2134 // 2135 // inline int Integer.bitCount(int) 2136 // inline int Long.bitCount(long) 2137 // 2138 // inline char Character.reverseBytes(char) 2139 // inline short Short.reverseBytes(short) 2140 // inline int Integer.reverseBytes(int) 2141 // inline long Long.reverseBytes(long) 2142 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2143 Node* arg = argument(0); 2144 Node* n; 2145 switch (id) { 2146 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break; 2147 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break; 2148 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break; 2149 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break; 2150 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break; 2151 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break; 2152 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break; 2153 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break; 2154 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break; 2155 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break; 2156 default: fatal_unexpected_iid(id); break; 2157 } 2158 set_result(_gvn.transform(n)); 2159 return true; 2160 } 2161 2162 //----------------------------inline_unsafe_access---------------------------- 2163 2164 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2165 2166 // Helper that guards and inserts a pre-barrier. 2167 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2168 Node* pre_val, bool need_mem_bar) { 2169 // We could be accessing the referent field of a reference object. If so, when G1 2170 // is enabled, we need to log the value in the referent field in an SATB buffer. 2171 // This routine performs some compile time filters and generates suitable 2172 // runtime filters that guard the pre-barrier code. 2173 // Also add memory barrier for non volatile load from the referent field 2174 // to prevent commoning of loads across safepoint. 2175 if (!UseG1GC && !need_mem_bar) 2176 return; 2177 2178 // Some compile time checks. 2179 2180 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2181 const TypeX* otype = offset->find_intptr_t_type(); 2182 if (otype != NULL && otype->is_con() && 2183 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2184 // Constant offset but not the reference_offset so just return 2185 return; 2186 } 2187 2188 // We only need to generate the runtime guards for instances. 2189 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2190 if (btype != NULL) { 2191 if (btype->isa_aryptr()) { 2192 // Array type so nothing to do 2193 return; 2194 } 2195 2196 const TypeInstPtr* itype = btype->isa_instptr(); 2197 if (itype != NULL) { 2198 // Can the klass of base_oop be statically determined to be 2199 // _not_ a sub-class of Reference and _not_ Object? 2200 ciKlass* klass = itype->klass(); 2201 if ( klass->is_loaded() && 2202 !klass->is_subtype_of(env()->Reference_klass()) && 2203 !env()->Object_klass()->is_subtype_of(klass)) { 2204 return; 2205 } 2206 } 2207 } 2208 2209 // The compile time filters did not reject base_oop/offset so 2210 // we need to generate the following runtime filters 2211 // 2212 // if (offset == java_lang_ref_Reference::_reference_offset) { 2213 // if (instance_of(base, java.lang.ref.Reference)) { 2214 // pre_barrier(_, pre_val, ...); 2215 // } 2216 // } 2217 2218 float likely = PROB_LIKELY( 0.999); 2219 float unlikely = PROB_UNLIKELY(0.999); 2220 2221 IdealKit ideal(this); 2222 #define __ ideal. 2223 2224 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2225 2226 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2227 // Update graphKit memory and control from IdealKit. 2228 sync_kit(ideal); 2229 2230 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2231 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2232 2233 // Update IdealKit memory and control from graphKit. 2234 __ sync_kit(this); 2235 2236 Node* one = __ ConI(1); 2237 // is_instof == 0 if base_oop == NULL 2238 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2239 2240 // Update graphKit from IdeakKit. 2241 sync_kit(ideal); 2242 2243 // Use the pre-barrier to record the value in the referent field 2244 pre_barrier(false /* do_load */, 2245 __ ctrl(), 2246 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2247 pre_val /* pre_val */, 2248 T_OBJECT); 2249 if (need_mem_bar) { 2250 // Add memory barrier to prevent commoning reads from this field 2251 // across safepoint since GC can change its value. 2252 insert_mem_bar(Op_MemBarCPUOrder); 2253 } 2254 // Update IdealKit from graphKit. 2255 __ sync_kit(this); 2256 2257 } __ end_if(); // _ref_type != ref_none 2258 } __ end_if(); // offset == referent_offset 2259 2260 // Final sync IdealKit and GraphKit. 2261 final_sync(ideal); 2262 #undef __ 2263 } 2264 2265 2266 // Interpret Unsafe.fieldOffset cookies correctly: 2267 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2268 2269 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2270 // Attempt to infer a sharper value type from the offset and base type. 2271 ciKlass* sharpened_klass = NULL; 2272 2273 // See if it is an instance field, with an object type. 2274 if (alias_type->field() != NULL) { 2275 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2276 if (alias_type->field()->type()->is_klass()) { 2277 sharpened_klass = alias_type->field()->type()->as_klass(); 2278 } 2279 } 2280 2281 // See if it is a narrow oop array. 2282 if (adr_type->isa_aryptr()) { 2283 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2284 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2285 if (elem_type != NULL) { 2286 sharpened_klass = elem_type->klass(); 2287 } 2288 } 2289 } 2290 2291 // The sharpened class might be unloaded if there is no class loader 2292 // contraint in place. 2293 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2294 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2295 2296 #ifndef PRODUCT 2297 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2298 tty->print(" from base type: "); adr_type->dump(); 2299 tty->print(" sharpened value: "); tjp->dump(); 2300 } 2301 #endif 2302 // Sharpen the value type. 2303 return tjp; 2304 } 2305 return NULL; 2306 } 2307 2308 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2309 if (callee()->is_static()) return false; // caller must have the capability! 2310 2311 #ifndef PRODUCT 2312 { 2313 ResourceMark rm; 2314 // Check the signatures. 2315 ciSignature* sig = callee()->signature(); 2316 #ifdef ASSERT 2317 if (!is_store) { 2318 // Object getObject(Object base, int/long offset), etc. 2319 BasicType rtype = sig->return_type()->basic_type(); 2320 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2321 rtype = T_ADDRESS; // it is really a C void* 2322 assert(rtype == type, "getter must return the expected value"); 2323 if (!is_native_ptr) { 2324 assert(sig->count() == 2, "oop getter has 2 arguments"); 2325 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2326 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2327 } else { 2328 assert(sig->count() == 1, "native getter has 1 argument"); 2329 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2330 } 2331 } else { 2332 // void putObject(Object base, int/long offset, Object x), etc. 2333 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2334 if (!is_native_ptr) { 2335 assert(sig->count() == 3, "oop putter has 3 arguments"); 2336 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2337 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2338 } else { 2339 assert(sig->count() == 2, "native putter has 2 arguments"); 2340 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2341 } 2342 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2343 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2344 vtype = T_ADDRESS; // it is really a C void* 2345 assert(vtype == type, "putter must accept the expected value"); 2346 } 2347 #endif // ASSERT 2348 } 2349 #endif //PRODUCT 2350 2351 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2352 2353 Node* receiver = argument(0); // type: oop 2354 2355 // Build address expression. See the code in inline_unsafe_prefetch. 2356 Node* adr; 2357 Node* heap_base_oop = top(); 2358 Node* offset = top(); 2359 Node* val; 2360 2361 if (!is_native_ptr) { 2362 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2363 Node* base = argument(1); // type: oop 2364 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2365 offset = argument(2); // type: long 2366 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2367 // to be plain byte offsets, which are also the same as those accepted 2368 // by oopDesc::field_base. 2369 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2370 "fieldOffset must be byte-scaled"); 2371 // 32-bit machines ignore the high half! 2372 offset = ConvL2X(offset); 2373 adr = make_unsafe_address(base, offset); 2374 heap_base_oop = base; 2375 val = is_store ? argument(4) : NULL; 2376 } else { 2377 Node* ptr = argument(1); // type: long 2378 ptr = ConvL2X(ptr); // adjust Java long to machine word 2379 adr = make_unsafe_address(NULL, ptr); 2380 val = is_store ? argument(3) : NULL; 2381 } 2382 2383 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2384 2385 // First guess at the value type. 2386 const Type *value_type = Type::get_const_basic_type(type); 2387 2388 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2389 // there was not enough information to nail it down. 2390 Compile::AliasType* alias_type = C->alias_type(adr_type); 2391 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2392 2393 // We will need memory barriers unless we can determine a unique 2394 // alias category for this reference. (Note: If for some reason 2395 // the barriers get omitted and the unsafe reference begins to "pollute" 2396 // the alias analysis of the rest of the graph, either Compile::can_alias 2397 // or Compile::must_alias will throw a diagnostic assert.) 2398 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2399 2400 // If we are reading the value of the referent field of a Reference 2401 // object (either by using Unsafe directly or through reflection) 2402 // then, if G1 is enabled, we need to record the referent in an 2403 // SATB log buffer using the pre-barrier mechanism. 2404 // Also we need to add memory barrier to prevent commoning reads 2405 // from this field across safepoint since GC can change its value. 2406 bool need_read_barrier = !is_native_ptr && !is_store && 2407 offset != top() && heap_base_oop != top(); 2408 2409 if (!is_store && type == T_OBJECT) { 2410 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2411 if (tjp != NULL) { 2412 value_type = tjp; 2413 } 2414 } 2415 2416 receiver = null_check(receiver); 2417 if (stopped()) { 2418 return true; 2419 } 2420 // Heap pointers get a null-check from the interpreter, 2421 // as a courtesy. However, this is not guaranteed by Unsafe, 2422 // and it is not possible to fully distinguish unintended nulls 2423 // from intended ones in this API. 2424 2425 if (is_volatile) { 2426 // We need to emit leading and trailing CPU membars (see below) in 2427 // addition to memory membars when is_volatile. This is a little 2428 // too strong, but avoids the need to insert per-alias-type 2429 // volatile membars (for stores; compare Parse::do_put_xxx), which 2430 // we cannot do effectively here because we probably only have a 2431 // rough approximation of type. 2432 need_mem_bar = true; 2433 // For Stores, place a memory ordering barrier now. 2434 if (is_store) 2435 insert_mem_bar(Op_MemBarRelease); 2436 } 2437 2438 // Memory barrier to prevent normal and 'unsafe' accesses from 2439 // bypassing each other. Happens after null checks, so the 2440 // exception paths do not take memory state from the memory barrier, 2441 // so there's no problems making a strong assert about mixing users 2442 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2443 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2444 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2445 2446 if (!is_store) { 2447 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2448 // load value 2449 switch (type) { 2450 case T_BOOLEAN: 2451 case T_CHAR: 2452 case T_BYTE: 2453 case T_SHORT: 2454 case T_INT: 2455 case T_LONG: 2456 case T_FLOAT: 2457 case T_DOUBLE: 2458 break; 2459 case T_OBJECT: 2460 if (need_read_barrier) { 2461 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2462 } 2463 break; 2464 case T_ADDRESS: 2465 // Cast to an int type. 2466 p = _gvn.transform(new (C) CastP2XNode(NULL, p)); 2467 p = ConvX2L(p); 2468 break; 2469 default: 2470 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2471 break; 2472 } 2473 // The load node has the control of the preceding MemBarCPUOrder. All 2474 // following nodes will have the control of the MemBarCPUOrder inserted at 2475 // the end of this method. So, pushing the load onto the stack at a later 2476 // point is fine. 2477 set_result(p); 2478 } else { 2479 // place effect of store into memory 2480 switch (type) { 2481 case T_DOUBLE: 2482 val = dstore_rounding(val); 2483 break; 2484 case T_ADDRESS: 2485 // Repackage the long as a pointer. 2486 val = ConvL2X(val); 2487 val = _gvn.transform(new (C) CastX2PNode(val)); 2488 break; 2489 } 2490 2491 if (type != T_OBJECT ) { 2492 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2493 } else { 2494 // Possibly an oop being stored to Java heap or native memory 2495 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2496 // oop to Java heap. 2497 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2498 } else { 2499 // We can't tell at compile time if we are storing in the Java heap or outside 2500 // of it. So we need to emit code to conditionally do the proper type of 2501 // store. 2502 2503 IdealKit ideal(this); 2504 #define __ ideal. 2505 // QQQ who knows what probability is here?? 2506 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2507 // Sync IdealKit and graphKit. 2508 sync_kit(ideal); 2509 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2510 // Update IdealKit memory. 2511 __ sync_kit(this); 2512 } __ else_(); { 2513 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile); 2514 } __ end_if(); 2515 // Final sync IdealKit and GraphKit. 2516 final_sync(ideal); 2517 #undef __ 2518 } 2519 } 2520 } 2521 2522 if (is_volatile) { 2523 if (!is_store) 2524 insert_mem_bar(Op_MemBarAcquire); 2525 else 2526 insert_mem_bar(Op_MemBarVolatile); 2527 } 2528 2529 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2530 2531 return true; 2532 } 2533 2534 //----------------------------inline_unsafe_prefetch---------------------------- 2535 2536 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2537 #ifndef PRODUCT 2538 { 2539 ResourceMark rm; 2540 // Check the signatures. 2541 ciSignature* sig = callee()->signature(); 2542 #ifdef ASSERT 2543 // Object getObject(Object base, int/long offset), etc. 2544 BasicType rtype = sig->return_type()->basic_type(); 2545 if (!is_native_ptr) { 2546 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2547 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2548 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2549 } else { 2550 assert(sig->count() == 1, "native prefetch has 1 argument"); 2551 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2552 } 2553 #endif // ASSERT 2554 } 2555 #endif // !PRODUCT 2556 2557 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2558 2559 const int idx = is_static ? 0 : 1; 2560 if (!is_static) { 2561 null_check_receiver(); 2562 if (stopped()) { 2563 return true; 2564 } 2565 } 2566 2567 // Build address expression. See the code in inline_unsafe_access. 2568 Node *adr; 2569 if (!is_native_ptr) { 2570 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2571 Node* base = argument(idx + 0); // type: oop 2572 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2573 Node* offset = argument(idx + 1); // type: long 2574 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2575 // to be plain byte offsets, which are also the same as those accepted 2576 // by oopDesc::field_base. 2577 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2578 "fieldOffset must be byte-scaled"); 2579 // 32-bit machines ignore the high half! 2580 offset = ConvL2X(offset); 2581 adr = make_unsafe_address(base, offset); 2582 } else { 2583 Node* ptr = argument(idx + 0); // type: long 2584 ptr = ConvL2X(ptr); // adjust Java long to machine word 2585 adr = make_unsafe_address(NULL, ptr); 2586 } 2587 2588 // Generate the read or write prefetch 2589 Node *prefetch; 2590 if (is_store) { 2591 prefetch = new (C) PrefetchWriteNode(i_o(), adr); 2592 } else { 2593 prefetch = new (C) PrefetchReadNode(i_o(), adr); 2594 } 2595 prefetch->init_req(0, control()); 2596 set_i_o(_gvn.transform(prefetch)); 2597 2598 return true; 2599 } 2600 2601 //----------------------------inline_unsafe_load_store---------------------------- 2602 // This method serves a couple of different customers (depending on LoadStoreKind): 2603 // 2604 // LS_cmpxchg: 2605 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2606 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2607 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2608 // 2609 // LS_xadd: 2610 // public int getAndAddInt( Object o, long offset, int delta) 2611 // public long getAndAddLong(Object o, long offset, long delta) 2612 // 2613 // LS_xchg: 2614 // int getAndSet(Object o, long offset, int newValue) 2615 // long getAndSet(Object o, long offset, long newValue) 2616 // Object getAndSet(Object o, long offset, Object newValue) 2617 // 2618 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2619 // This basic scheme here is the same as inline_unsafe_access, but 2620 // differs in enough details that combining them would make the code 2621 // overly confusing. (This is a true fact! I originally combined 2622 // them, but even I was confused by it!) As much code/comments as 2623 // possible are retained from inline_unsafe_access though to make 2624 // the correspondences clearer. - dl 2625 2626 if (callee()->is_static()) return false; // caller must have the capability! 2627 2628 #ifndef PRODUCT 2629 BasicType rtype; 2630 { 2631 ResourceMark rm; 2632 // Check the signatures. 2633 ciSignature* sig = callee()->signature(); 2634 rtype = sig->return_type()->basic_type(); 2635 if (kind == LS_xadd || kind == LS_xchg) { 2636 // Check the signatures. 2637 #ifdef ASSERT 2638 assert(rtype == type, "get and set must return the expected type"); 2639 assert(sig->count() == 3, "get and set has 3 arguments"); 2640 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2641 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2642 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2643 #endif // ASSERT 2644 } else if (kind == LS_cmpxchg) { 2645 // Check the signatures. 2646 #ifdef ASSERT 2647 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2648 assert(sig->count() == 4, "CAS has 4 arguments"); 2649 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2650 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2651 #endif // ASSERT 2652 } else { 2653 ShouldNotReachHere(); 2654 } 2655 } 2656 #endif //PRODUCT 2657 2658 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2659 2660 // Get arguments: 2661 Node* receiver = NULL; 2662 Node* base = NULL; 2663 Node* offset = NULL; 2664 Node* oldval = NULL; 2665 Node* newval = NULL; 2666 if (kind == LS_cmpxchg) { 2667 const bool two_slot_type = type2size[type] == 2; 2668 receiver = argument(0); // type: oop 2669 base = argument(1); // type: oop 2670 offset = argument(2); // type: long 2671 oldval = argument(4); // type: oop, int, or long 2672 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2673 } else if (kind == LS_xadd || kind == LS_xchg){ 2674 receiver = argument(0); // type: oop 2675 base = argument(1); // type: oop 2676 offset = argument(2); // type: long 2677 oldval = NULL; 2678 newval = argument(4); // type: oop, int, or long 2679 } 2680 2681 // Null check receiver. 2682 receiver = null_check(receiver); 2683 if (stopped()) { 2684 return true; 2685 } 2686 2687 // Build field offset expression. 2688 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2689 // to be plain byte offsets, which are also the same as those accepted 2690 // by oopDesc::field_base. 2691 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2692 // 32-bit machines ignore the high half of long offsets 2693 offset = ConvL2X(offset); 2694 Node* adr = make_unsafe_address(base, offset); 2695 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2696 2697 // For CAS, unlike inline_unsafe_access, there seems no point in 2698 // trying to refine types. Just use the coarse types here. 2699 const Type *value_type = Type::get_const_basic_type(type); 2700 Compile::AliasType* alias_type = C->alias_type(adr_type); 2701 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2702 2703 if (kind == LS_xchg && type == T_OBJECT) { 2704 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2705 if (tjp != NULL) { 2706 value_type = tjp; 2707 } 2708 } 2709 2710 int alias_idx = C->get_alias_index(adr_type); 2711 2712 // Memory-model-wise, a LoadStore acts like a little synchronized 2713 // block, so needs barriers on each side. These don't translate 2714 // into actual barriers on most machines, but we still need rest of 2715 // compiler to respect ordering. 2716 2717 insert_mem_bar(Op_MemBarRelease); 2718 insert_mem_bar(Op_MemBarCPUOrder); 2719 2720 // 4984716: MemBars must be inserted before this 2721 // memory node in order to avoid a false 2722 // dependency which will confuse the scheduler. 2723 Node *mem = memory(alias_idx); 2724 2725 // For now, we handle only those cases that actually exist: ints, 2726 // longs, and Object. Adding others should be straightforward. 2727 Node* load_store; 2728 switch(type) { 2729 case T_INT: 2730 if (kind == LS_xadd) { 2731 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type)); 2732 } else if (kind == LS_xchg) { 2733 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type)); 2734 } else if (kind == LS_cmpxchg) { 2735 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2736 } else { 2737 ShouldNotReachHere(); 2738 } 2739 break; 2740 case T_LONG: 2741 if (kind == LS_xadd) { 2742 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2743 } else if (kind == LS_xchg) { 2744 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2745 } else if (kind == LS_cmpxchg) { 2746 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2747 } else { 2748 ShouldNotReachHere(); 2749 } 2750 break; 2751 case T_OBJECT: 2752 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2753 // could be delayed during Parse (for example, in adjust_map_after_if()). 2754 // Execute transformation here to avoid barrier generation in such case. 2755 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2756 newval = _gvn.makecon(TypePtr::NULL_PTR); 2757 2758 // Reference stores need a store barrier. 2759 if (kind == LS_xchg) { 2760 // If pre-barrier must execute before the oop store, old value will require do_load here. 2761 if (!can_move_pre_barrier()) { 2762 pre_barrier(true /* do_load*/, 2763 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2764 NULL /* pre_val*/, 2765 T_OBJECT); 2766 } // Else move pre_barrier to use load_store value, see below. 2767 } else if (kind == LS_cmpxchg) { 2768 // Same as for newval above: 2769 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2770 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2771 } 2772 // The only known value which might get overwritten is oldval. 2773 pre_barrier(false /* do_load */, 2774 control(), NULL, NULL, max_juint, NULL, NULL, 2775 oldval /* pre_val */, 2776 T_OBJECT); 2777 } else { 2778 ShouldNotReachHere(); 2779 } 2780 2781 #ifdef _LP64 2782 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2783 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2784 if (kind == LS_xchg) { 2785 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr, 2786 newval_enc, adr_type, value_type->make_narrowoop())); 2787 } else { 2788 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2789 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2790 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr, 2791 newval_enc, oldval_enc)); 2792 } 2793 } else 2794 #endif 2795 { 2796 if (kind == LS_xchg) { 2797 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2798 } else { 2799 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2800 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2801 } 2802 } 2803 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2804 break; 2805 default: 2806 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2807 break; 2808 } 2809 2810 // SCMemProjNodes represent the memory state of a LoadStore. Their 2811 // main role is to prevent LoadStore nodes from being optimized away 2812 // when their results aren't used. 2813 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store)); 2814 set_memory(proj, alias_idx); 2815 2816 if (type == T_OBJECT && kind == LS_xchg) { 2817 #ifdef _LP64 2818 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2819 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type())); 2820 } 2821 #endif 2822 if (can_move_pre_barrier()) { 2823 // Don't need to load pre_val. The old value is returned by load_store. 2824 // The pre_barrier can execute after the xchg as long as no safepoint 2825 // gets inserted between them. 2826 pre_barrier(false /* do_load */, 2827 control(), NULL, NULL, max_juint, NULL, NULL, 2828 load_store /* pre_val */, 2829 T_OBJECT); 2830 } 2831 } 2832 2833 // Add the trailing membar surrounding the access 2834 insert_mem_bar(Op_MemBarCPUOrder); 2835 insert_mem_bar(Op_MemBarAcquire); 2836 2837 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2838 set_result(load_store); 2839 return true; 2840 } 2841 2842 //----------------------------inline_unsafe_ordered_store---------------------- 2843 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2844 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 2845 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 2846 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2847 // This is another variant of inline_unsafe_access, differing in 2848 // that it always issues store-store ("release") barrier and ensures 2849 // store-atomicity (which only matters for "long"). 2850 2851 if (callee()->is_static()) return false; // caller must have the capability! 2852 2853 #ifndef PRODUCT 2854 { 2855 ResourceMark rm; 2856 // Check the signatures. 2857 ciSignature* sig = callee()->signature(); 2858 #ifdef ASSERT 2859 BasicType rtype = sig->return_type()->basic_type(); 2860 assert(rtype == T_VOID, "must return void"); 2861 assert(sig->count() == 3, "has 3 arguments"); 2862 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2863 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2864 #endif // ASSERT 2865 } 2866 #endif //PRODUCT 2867 2868 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2869 2870 // Get arguments: 2871 Node* receiver = argument(0); // type: oop 2872 Node* base = argument(1); // type: oop 2873 Node* offset = argument(2); // type: long 2874 Node* val = argument(4); // type: oop, int, or long 2875 2876 // Null check receiver. 2877 receiver = null_check(receiver); 2878 if (stopped()) { 2879 return true; 2880 } 2881 2882 // Build field offset expression. 2883 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2884 // 32-bit machines ignore the high half of long offsets 2885 offset = ConvL2X(offset); 2886 Node* adr = make_unsafe_address(base, offset); 2887 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2888 const Type *value_type = Type::get_const_basic_type(type); 2889 Compile::AliasType* alias_type = C->alias_type(adr_type); 2890 2891 insert_mem_bar(Op_MemBarRelease); 2892 insert_mem_bar(Op_MemBarCPUOrder); 2893 // Ensure that the store is atomic for longs: 2894 const bool require_atomic_access = true; 2895 Node* store; 2896 if (type == T_OBJECT) // reference stores need a store barrier. 2897 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2898 else { 2899 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2900 } 2901 insert_mem_bar(Op_MemBarCPUOrder); 2902 return true; 2903 } 2904 2905 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2906 // Regardless of form, don't allow previous ld/st to move down, 2907 // then issue acquire, release, or volatile mem_bar. 2908 insert_mem_bar(Op_MemBarCPUOrder); 2909 switch(id) { 2910 case vmIntrinsics::_loadFence: 2911 insert_mem_bar(Op_MemBarAcquire); 2912 return true; 2913 case vmIntrinsics::_storeFence: 2914 insert_mem_bar(Op_MemBarRelease); 2915 return true; 2916 case vmIntrinsics::_fullFence: 2917 insert_mem_bar(Op_MemBarVolatile); 2918 return true; 2919 default: 2920 fatal_unexpected_iid(id); 2921 return false; 2922 } 2923 } 2924 2925 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 2926 if (!kls->is_Con()) { 2927 return true; 2928 } 2929 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 2930 if (klsptr == NULL) { 2931 return true; 2932 } 2933 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 2934 // don't need a guard for a klass that is already initialized 2935 return !ik->is_initialized(); 2936 } 2937 2938 //----------------------------inline_unsafe_allocate--------------------------- 2939 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 2940 bool LibraryCallKit::inline_unsafe_allocate() { 2941 if (callee()->is_static()) return false; // caller must have the capability! 2942 2943 null_check_receiver(); // null-check, then ignore 2944 Node* cls = null_check(argument(1)); 2945 if (stopped()) return true; 2946 2947 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2948 kls = null_check(kls); 2949 if (stopped()) return true; // argument was like int.class 2950 2951 Node* test = NULL; 2952 if (LibraryCallKit::klass_needs_init_guard(kls)) { 2953 // Note: The argument might still be an illegal value like 2954 // Serializable.class or Object[].class. The runtime will handle it. 2955 // But we must make an explicit check for initialization. 2956 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 2957 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 2958 // can generate code to load it as unsigned byte. 2959 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN); 2960 Node* bits = intcon(InstanceKlass::fully_initialized); 2961 test = _gvn.transform(new (C) SubINode(inst, bits)); 2962 // The 'test' is non-zero if we need to take a slow path. 2963 } 2964 2965 Node* obj = new_instance(kls, test); 2966 set_result(obj); 2967 return true; 2968 } 2969 2970 #ifdef TRACE_HAVE_INTRINSICS 2971 /* 2972 * oop -> myklass 2973 * myklass->trace_id |= USED 2974 * return myklass->trace_id & ~0x3 2975 */ 2976 bool LibraryCallKit::inline_native_classID() { 2977 null_check_receiver(); // null-check, then ignore 2978 Node* cls = null_check(argument(1), T_OBJECT); 2979 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2980 kls = null_check(kls, T_OBJECT); 2981 ByteSize offset = TRACE_ID_OFFSET; 2982 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 2983 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG); 2984 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 2985 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits)); 2986 Node* clsused = longcon(0x01l); // set the class bit 2987 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused)); 2988 2989 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 2990 store_to_memory(control(), insp, orl, T_LONG, adr_type); 2991 set_result(andl); 2992 return true; 2993 } 2994 2995 bool LibraryCallKit::inline_native_threadID() { 2996 Node* tls_ptr = NULL; 2997 Node* cur_thr = generate_current_thread(tls_ptr); 2998 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2999 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 3000 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3001 3002 Node* threadid = NULL; 3003 size_t thread_id_size = OSThread::thread_id_size(); 3004 if (thread_id_size == (size_t) BytesPerLong) { 3005 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG)); 3006 } else if (thread_id_size == (size_t) BytesPerInt) { 3007 threadid = make_load(control(), p, TypeInt::INT, T_INT); 3008 } else { 3009 ShouldNotReachHere(); 3010 } 3011 set_result(threadid); 3012 return true; 3013 } 3014 #endif 3015 3016 //------------------------inline_native_time_funcs-------------- 3017 // inline code for System.currentTimeMillis() and System.nanoTime() 3018 // these have the same type and signature 3019 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3020 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3021 const TypePtr* no_memory_effects = NULL; 3022 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3023 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0)); 3024 #ifdef ASSERT 3025 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1)); 3026 assert(value_top == top(), "second value must be top"); 3027 #endif 3028 set_result(value); 3029 return true; 3030 } 3031 3032 //------------------------inline_native_currentThread------------------ 3033 bool LibraryCallKit::inline_native_currentThread() { 3034 Node* junk = NULL; 3035 set_result(generate_current_thread(junk)); 3036 return true; 3037 } 3038 3039 //------------------------inline_native_isInterrupted------------------ 3040 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3041 bool LibraryCallKit::inline_native_isInterrupted() { 3042 // Add a fast path to t.isInterrupted(clear_int): 3043 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 3044 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3045 // So, in the common case that the interrupt bit is false, 3046 // we avoid making a call into the VM. Even if the interrupt bit 3047 // is true, if the clear_int argument is false, we avoid the VM call. 3048 // However, if the receiver is not currentThread, we must call the VM, 3049 // because there must be some locking done around the operation. 3050 3051 // We only go to the fast case code if we pass two guards. 3052 // Paths which do not pass are accumulated in the slow_region. 3053 3054 enum { 3055 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3056 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3057 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3058 PATH_LIMIT 3059 }; 3060 3061 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3062 // out of the function. 3063 insert_mem_bar(Op_MemBarCPUOrder); 3064 3065 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT); 3066 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL); 3067 3068 RegionNode* slow_region = new (C) RegionNode(1); 3069 record_for_igvn(slow_region); 3070 3071 // (a) Receiving thread must be the current thread. 3072 Node* rec_thr = argument(0); 3073 Node* tls_ptr = NULL; 3074 Node* cur_thr = generate_current_thread(tls_ptr); 3075 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr)); 3076 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne)); 3077 3078 generate_slow_guard(bol_thr, slow_region); 3079 3080 // (b) Interrupt bit on TLS must be false. 3081 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3082 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 3083 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3084 3085 // Set the control input on the field _interrupted read to prevent it floating up. 3086 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 3087 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0))); 3088 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne)); 3089 3090 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3091 3092 // First fast path: if (!TLS._interrupted) return false; 3093 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit)); 3094 result_rgn->init_req(no_int_result_path, false_bit); 3095 result_val->init_req(no_int_result_path, intcon(0)); 3096 3097 // drop through to next case 3098 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit))); 3099 3100 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3101 Node* clr_arg = argument(1); 3102 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0))); 3103 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne)); 3104 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3105 3106 // Second fast path: ... else if (!clear_int) return true; 3107 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg)); 3108 result_rgn->init_req(no_clear_result_path, false_arg); 3109 result_val->init_req(no_clear_result_path, intcon(1)); 3110 3111 // drop through to next case 3112 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg))); 3113 3114 // (d) Otherwise, go to the slow path. 3115 slow_region->add_req(control()); 3116 set_control( _gvn.transform(slow_region)); 3117 3118 if (stopped()) { 3119 // There is no slow path. 3120 result_rgn->init_req(slow_result_path, top()); 3121 result_val->init_req(slow_result_path, top()); 3122 } else { 3123 // non-virtual because it is a private non-static 3124 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3125 3126 Node* slow_val = set_results_for_java_call(slow_call); 3127 // this->control() comes from set_results_for_java_call 3128 3129 Node* fast_io = slow_call->in(TypeFunc::I_O); 3130 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3131 3132 // These two phis are pre-filled with copies of of the fast IO and Memory 3133 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3134 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3135 3136 result_rgn->init_req(slow_result_path, control()); 3137 result_io ->init_req(slow_result_path, i_o()); 3138 result_mem->init_req(slow_result_path, reset_memory()); 3139 result_val->init_req(slow_result_path, slow_val); 3140 3141 set_all_memory(_gvn.transform(result_mem)); 3142 set_i_o( _gvn.transform(result_io)); 3143 } 3144 3145 C->set_has_split_ifs(true); // Has chance for split-if optimization 3146 set_result(result_rgn, result_val); 3147 return true; 3148 } 3149 3150 //---------------------------load_mirror_from_klass---------------------------- 3151 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3152 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3153 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3154 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 3155 } 3156 3157 //-----------------------load_klass_from_mirror_common------------------------- 3158 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3159 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3160 // and branch to the given path on the region. 3161 // If never_see_null, take an uncommon trap on null, so we can optimistically 3162 // compile for the non-null case. 3163 // If the region is NULL, force never_see_null = true. 3164 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3165 bool never_see_null, 3166 RegionNode* region, 3167 int null_path, 3168 int offset) { 3169 if (region == NULL) never_see_null = true; 3170 Node* p = basic_plus_adr(mirror, offset); 3171 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3172 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3173 Node* null_ctl = top(); 3174 kls = null_check_oop(kls, &null_ctl, never_see_null); 3175 if (region != NULL) { 3176 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3177 region->init_req(null_path, null_ctl); 3178 } else { 3179 assert(null_ctl == top(), "no loose ends"); 3180 } 3181 return kls; 3182 } 3183 3184 //--------------------(inline_native_Class_query helpers)--------------------- 3185 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3186 // Fall through if (mods & mask) == bits, take the guard otherwise. 3187 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3188 // Branch around if the given klass has the given modifier bit set. 3189 // Like generate_guard, adds a new path onto the region. 3190 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3191 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 3192 Node* mask = intcon(modifier_mask); 3193 Node* bits = intcon(modifier_bits); 3194 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask)); 3195 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits)); 3196 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 3197 return generate_fair_guard(bol, region); 3198 } 3199 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3200 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3201 } 3202 3203 //-------------------------inline_native_Class_query------------------- 3204 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3205 const Type* return_type = TypeInt::BOOL; 3206 Node* prim_return_value = top(); // what happens if it's a primitive class? 3207 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3208 bool expect_prim = false; // most of these guys expect to work on refs 3209 3210 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3211 3212 Node* mirror = argument(0); 3213 Node* obj = top(); 3214 3215 switch (id) { 3216 case vmIntrinsics::_isInstance: 3217 // nothing is an instance of a primitive type 3218 prim_return_value = intcon(0); 3219 obj = argument(1); 3220 break; 3221 case vmIntrinsics::_getModifiers: 3222 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3223 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3224 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3225 break; 3226 case vmIntrinsics::_isInterface: 3227 prim_return_value = intcon(0); 3228 break; 3229 case vmIntrinsics::_isArray: 3230 prim_return_value = intcon(0); 3231 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3232 break; 3233 case vmIntrinsics::_isPrimitive: 3234 prim_return_value = intcon(1); 3235 expect_prim = true; // obviously 3236 break; 3237 case vmIntrinsics::_getSuperclass: 3238 prim_return_value = null(); 3239 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3240 break; 3241 case vmIntrinsics::_getComponentType: 3242 prim_return_value = null(); 3243 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3244 break; 3245 case vmIntrinsics::_getClassAccessFlags: 3246 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3247 return_type = TypeInt::INT; // not bool! 6297094 3248 break; 3249 default: 3250 fatal_unexpected_iid(id); 3251 break; 3252 } 3253 3254 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3255 if (mirror_con == NULL) return false; // cannot happen? 3256 3257 #ifndef PRODUCT 3258 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 3259 ciType* k = mirror_con->java_mirror_type(); 3260 if (k) { 3261 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3262 k->print_name(); 3263 tty->cr(); 3264 } 3265 } 3266 #endif 3267 3268 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3269 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3270 record_for_igvn(region); 3271 PhiNode* phi = new (C) PhiNode(region, return_type); 3272 3273 // The mirror will never be null of Reflection.getClassAccessFlags, however 3274 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3275 // if it is. See bug 4774291. 3276 3277 // For Reflection.getClassAccessFlags(), the null check occurs in 3278 // the wrong place; see inline_unsafe_access(), above, for a similar 3279 // situation. 3280 mirror = null_check(mirror); 3281 // If mirror or obj is dead, only null-path is taken. 3282 if (stopped()) return true; 3283 3284 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3285 3286 // Now load the mirror's klass metaobject, and null-check it. 3287 // Side-effects region with the control path if the klass is null. 3288 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3289 // If kls is null, we have a primitive mirror. 3290 phi->init_req(_prim_path, prim_return_value); 3291 if (stopped()) { set_result(region, phi); return true; } 3292 3293 Node* p; // handy temp 3294 Node* null_ctl; 3295 3296 // Now that we have the non-null klass, we can perform the real query. 3297 // For constant classes, the query will constant-fold in LoadNode::Value. 3298 Node* query_value = top(); 3299 switch (id) { 3300 case vmIntrinsics::_isInstance: 3301 // nothing is an instance of a primitive type 3302 query_value = gen_instanceof(obj, kls); 3303 break; 3304 3305 case vmIntrinsics::_getModifiers: 3306 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3307 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3308 break; 3309 3310 case vmIntrinsics::_isInterface: 3311 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3312 if (generate_interface_guard(kls, region) != NULL) 3313 // A guard was added. If the guard is taken, it was an interface. 3314 phi->add_req(intcon(1)); 3315 // If we fall through, it's a plain class. 3316 query_value = intcon(0); 3317 break; 3318 3319 case vmIntrinsics::_isArray: 3320 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3321 if (generate_array_guard(kls, region) != NULL) 3322 // A guard was added. If the guard is taken, it was an array. 3323 phi->add_req(intcon(1)); 3324 // If we fall through, it's a plain class. 3325 query_value = intcon(0); 3326 break; 3327 3328 case vmIntrinsics::_isPrimitive: 3329 query_value = intcon(0); // "normal" path produces false 3330 break; 3331 3332 case vmIntrinsics::_getSuperclass: 3333 // The rules here are somewhat unfortunate, but we can still do better 3334 // with random logic than with a JNI call. 3335 // Interfaces store null or Object as _super, but must report null. 3336 // Arrays store an intermediate super as _super, but must report Object. 3337 // Other types can report the actual _super. 3338 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3339 if (generate_interface_guard(kls, region) != NULL) 3340 // A guard was added. If the guard is taken, it was an interface. 3341 phi->add_req(null()); 3342 if (generate_array_guard(kls, region) != NULL) 3343 // A guard was added. If the guard is taken, it was an array. 3344 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3345 // If we fall through, it's a plain class. Get its _super. 3346 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3347 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3348 null_ctl = top(); 3349 kls = null_check_oop(kls, &null_ctl); 3350 if (null_ctl != top()) { 3351 // If the guard is taken, Object.superClass is null (both klass and mirror). 3352 region->add_req(null_ctl); 3353 phi ->add_req(null()); 3354 } 3355 if (!stopped()) { 3356 query_value = load_mirror_from_klass(kls); 3357 } 3358 break; 3359 3360 case vmIntrinsics::_getComponentType: 3361 if (generate_array_guard(kls, region) != NULL) { 3362 // Be sure to pin the oop load to the guard edge just created: 3363 Node* is_array_ctrl = region->in(region->req()-1); 3364 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); 3365 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 3366 phi->add_req(cmo); 3367 } 3368 query_value = null(); // non-array case is null 3369 break; 3370 3371 case vmIntrinsics::_getClassAccessFlags: 3372 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3373 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3374 break; 3375 3376 default: 3377 fatal_unexpected_iid(id); 3378 break; 3379 } 3380 3381 // Fall-through is the normal case of a query to a real class. 3382 phi->init_req(1, query_value); 3383 region->init_req(1, control()); 3384 3385 C->set_has_split_ifs(true); // Has chance for split-if optimization 3386 set_result(region, phi); 3387 return true; 3388 } 3389 3390 //--------------------------inline_native_subtype_check------------------------ 3391 // This intrinsic takes the JNI calls out of the heart of 3392 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3393 bool LibraryCallKit::inline_native_subtype_check() { 3394 // Pull both arguments off the stack. 3395 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3396 args[0] = argument(0); 3397 args[1] = argument(1); 3398 Node* klasses[2]; // corresponding Klasses: superk, subk 3399 klasses[0] = klasses[1] = top(); 3400 3401 enum { 3402 // A full decision tree on {superc is prim, subc is prim}: 3403 _prim_0_path = 1, // {P,N} => false 3404 // {P,P} & superc!=subc => false 3405 _prim_same_path, // {P,P} & superc==subc => true 3406 _prim_1_path, // {N,P} => false 3407 _ref_subtype_path, // {N,N} & subtype check wins => true 3408 _both_ref_path, // {N,N} & subtype check loses => false 3409 PATH_LIMIT 3410 }; 3411 3412 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3413 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 3414 record_for_igvn(region); 3415 3416 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3417 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3418 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3419 3420 // First null-check both mirrors and load each mirror's klass metaobject. 3421 int which_arg; 3422 for (which_arg = 0; which_arg <= 1; which_arg++) { 3423 Node* arg = args[which_arg]; 3424 arg = null_check(arg); 3425 if (stopped()) break; 3426 args[which_arg] = arg; 3427 3428 Node* p = basic_plus_adr(arg, class_klass_offset); 3429 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3430 klasses[which_arg] = _gvn.transform(kls); 3431 } 3432 3433 // Having loaded both klasses, test each for null. 3434 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3435 for (which_arg = 0; which_arg <= 1; which_arg++) { 3436 Node* kls = klasses[which_arg]; 3437 Node* null_ctl = top(); 3438 kls = null_check_oop(kls, &null_ctl, never_see_null); 3439 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3440 region->init_req(prim_path, null_ctl); 3441 if (stopped()) break; 3442 klasses[which_arg] = kls; 3443 } 3444 3445 if (!stopped()) { 3446 // now we have two reference types, in klasses[0..1] 3447 Node* subk = klasses[1]; // the argument to isAssignableFrom 3448 Node* superk = klasses[0]; // the receiver 3449 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3450 // now we have a successful reference subtype check 3451 region->set_req(_ref_subtype_path, control()); 3452 } 3453 3454 // If both operands are primitive (both klasses null), then 3455 // we must return true when they are identical primitives. 3456 // It is convenient to test this after the first null klass check. 3457 set_control(region->in(_prim_0_path)); // go back to first null check 3458 if (!stopped()) { 3459 // Since superc is primitive, make a guard for the superc==subc case. 3460 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1])); 3461 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq)); 3462 generate_guard(bol_eq, region, PROB_FAIR); 3463 if (region->req() == PATH_LIMIT+1) { 3464 // A guard was added. If the added guard is taken, superc==subc. 3465 region->swap_edges(PATH_LIMIT, _prim_same_path); 3466 region->del_req(PATH_LIMIT); 3467 } 3468 region->set_req(_prim_0_path, control()); // Not equal after all. 3469 } 3470 3471 // these are the only paths that produce 'true': 3472 phi->set_req(_prim_same_path, intcon(1)); 3473 phi->set_req(_ref_subtype_path, intcon(1)); 3474 3475 // pull together the cases: 3476 assert(region->req() == PATH_LIMIT, "sane region"); 3477 for (uint i = 1; i < region->req(); i++) { 3478 Node* ctl = region->in(i); 3479 if (ctl == NULL || ctl == top()) { 3480 region->set_req(i, top()); 3481 phi ->set_req(i, top()); 3482 } else if (phi->in(i) == NULL) { 3483 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3484 } 3485 } 3486 3487 set_control(_gvn.transform(region)); 3488 set_result(_gvn.transform(phi)); 3489 return true; 3490 } 3491 3492 //---------------------generate_array_guard_common------------------------ 3493 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3494 bool obj_array, bool not_array) { 3495 // If obj_array/non_array==false/false: 3496 // Branch around if the given klass is in fact an array (either obj or prim). 3497 // If obj_array/non_array==false/true: 3498 // Branch around if the given klass is not an array klass of any kind. 3499 // If obj_array/non_array==true/true: 3500 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3501 // If obj_array/non_array==true/false: 3502 // Branch around if the kls is an oop array (Object[] or subtype) 3503 // 3504 // Like generate_guard, adds a new path onto the region. 3505 jint layout_con = 0; 3506 Node* layout_val = get_layout_helper(kls, layout_con); 3507 if (layout_val == NULL) { 3508 bool query = (obj_array 3509 ? Klass::layout_helper_is_objArray(layout_con) 3510 : Klass::layout_helper_is_array(layout_con)); 3511 if (query == not_array) { 3512 return NULL; // never a branch 3513 } else { // always a branch 3514 Node* always_branch = control(); 3515 if (region != NULL) 3516 region->add_req(always_branch); 3517 set_control(top()); 3518 return always_branch; 3519 } 3520 } 3521 // Now test the correct condition. 3522 jint nval = (obj_array 3523 ? ((jint)Klass::_lh_array_tag_type_value 3524 << Klass::_lh_array_tag_shift) 3525 : Klass::_lh_neutral_value); 3526 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval))); 3527 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3528 // invert the test if we are looking for a non-array 3529 if (not_array) btest = BoolTest(btest).negate(); 3530 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest)); 3531 return generate_fair_guard(bol, region); 3532 } 3533 3534 3535 //-----------------------inline_native_newArray-------------------------- 3536 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3537 bool LibraryCallKit::inline_native_newArray() { 3538 Node* mirror = argument(0); 3539 Node* count_val = argument(1); 3540 3541 mirror = null_check(mirror); 3542 // If mirror or obj is dead, only null-path is taken. 3543 if (stopped()) return true; 3544 3545 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3546 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3547 PhiNode* result_val = new(C) PhiNode(result_reg, 3548 TypeInstPtr::NOTNULL); 3549 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3550 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3551 TypePtr::BOTTOM); 3552 3553 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3554 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3555 result_reg, _slow_path); 3556 Node* normal_ctl = control(); 3557 Node* no_array_ctl = result_reg->in(_slow_path); 3558 3559 // Generate code for the slow case. We make a call to newArray(). 3560 set_control(no_array_ctl); 3561 if (!stopped()) { 3562 // Either the input type is void.class, or else the 3563 // array klass has not yet been cached. Either the 3564 // ensuing call will throw an exception, or else it 3565 // will cache the array klass for next time. 3566 PreserveJVMState pjvms(this); 3567 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3568 Node* slow_result = set_results_for_java_call(slow_call); 3569 // this->control() comes from set_results_for_java_call 3570 result_reg->set_req(_slow_path, control()); 3571 result_val->set_req(_slow_path, slow_result); 3572 result_io ->set_req(_slow_path, i_o()); 3573 result_mem->set_req(_slow_path, reset_memory()); 3574 } 3575 3576 set_control(normal_ctl); 3577 if (!stopped()) { 3578 // Normal case: The array type has been cached in the java.lang.Class. 3579 // The following call works fine even if the array type is polymorphic. 3580 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3581 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3582 result_reg->init_req(_normal_path, control()); 3583 result_val->init_req(_normal_path, obj); 3584 result_io ->init_req(_normal_path, i_o()); 3585 result_mem->init_req(_normal_path, reset_memory()); 3586 } 3587 3588 // Return the combined state. 3589 set_i_o( _gvn.transform(result_io) ); 3590 set_all_memory( _gvn.transform(result_mem)); 3591 3592 C->set_has_split_ifs(true); // Has chance for split-if optimization 3593 set_result(result_reg, result_val); 3594 return true; 3595 } 3596 3597 //----------------------inline_native_getLength-------------------------- 3598 // public static native int java.lang.reflect.Array.getLength(Object array); 3599 bool LibraryCallKit::inline_native_getLength() { 3600 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3601 3602 Node* array = null_check(argument(0)); 3603 // If array is dead, only null-path is taken. 3604 if (stopped()) return true; 3605 3606 // Deoptimize if it is a non-array. 3607 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3608 3609 if (non_array != NULL) { 3610 PreserveJVMState pjvms(this); 3611 set_control(non_array); 3612 uncommon_trap(Deoptimization::Reason_intrinsic, 3613 Deoptimization::Action_maybe_recompile); 3614 } 3615 3616 // If control is dead, only non-array-path is taken. 3617 if (stopped()) return true; 3618 3619 // The works fine even if the array type is polymorphic. 3620 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3621 Node* result = load_array_length(array); 3622 3623 C->set_has_split_ifs(true); // Has chance for split-if optimization 3624 set_result(result); 3625 return true; 3626 } 3627 3628 //------------------------inline_array_copyOf---------------------------- 3629 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3630 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3631 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3632 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3633 3634 // Get the arguments. 3635 Node* original = argument(0); 3636 Node* start = is_copyOfRange? argument(1): intcon(0); 3637 Node* end = is_copyOfRange? argument(2): argument(1); 3638 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3639 3640 Node* newcopy; 3641 3642 // Set the original stack and the reexecute bit for the interpreter to reexecute 3643 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3644 { PreserveReexecuteState preexecs(this); 3645 jvms()->set_should_reexecute(true); 3646 3647 array_type_mirror = null_check(array_type_mirror); 3648 original = null_check(original); 3649 3650 // Check if a null path was taken unconditionally. 3651 if (stopped()) return true; 3652 3653 Node* orig_length = load_array_length(original); 3654 3655 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3656 klass_node = null_check(klass_node); 3657 3658 RegionNode* bailout = new (C) RegionNode(1); 3659 record_for_igvn(bailout); 3660 3661 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3662 // Bail out if that is so. 3663 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3664 if (not_objArray != NULL) { 3665 // Improve the klass node's type from the new optimistic assumption: 3666 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3667 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3668 Node* cast = new (C) CastPPNode(klass_node, akls); 3669 cast->init_req(0, control()); 3670 klass_node = _gvn.transform(cast); 3671 } 3672 3673 // Bail out if either start or end is negative. 3674 generate_negative_guard(start, bailout, &start); 3675 generate_negative_guard(end, bailout, &end); 3676 3677 Node* length = end; 3678 if (_gvn.type(start) != TypeInt::ZERO) { 3679 length = _gvn.transform(new (C) SubINode(end, start)); 3680 } 3681 3682 // Bail out if length is negative. 3683 // Without this the new_array would throw 3684 // NegativeArraySizeException but IllegalArgumentException is what 3685 // should be thrown 3686 generate_negative_guard(length, bailout, &length); 3687 3688 if (bailout->req() > 1) { 3689 PreserveJVMState pjvms(this); 3690 set_control(_gvn.transform(bailout)); 3691 uncommon_trap(Deoptimization::Reason_intrinsic, 3692 Deoptimization::Action_maybe_recompile); 3693 } 3694 3695 if (!stopped()) { 3696 // How many elements will we copy from the original? 3697 // The answer is MinI(orig_length - start, length). 3698 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start)); 3699 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3700 3701 newcopy = new_array(klass_node, length, 0); // no argments to push 3702 3703 // Generate a direct call to the right arraycopy function(s). 3704 // We know the copy is disjoint but we might not know if the 3705 // oop stores need checking. 3706 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3707 // This will fail a store-check if x contains any non-nulls. 3708 bool disjoint_bases = true; 3709 // if start > orig_length then the length of the copy may be 3710 // negative. 3711 bool length_never_negative = !is_copyOfRange; 3712 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3713 original, start, newcopy, intcon(0), moved, 3714 disjoint_bases, length_never_negative); 3715 } 3716 } // original reexecute is set back here 3717 3718 C->set_has_split_ifs(true); // Has chance for split-if optimization 3719 if (!stopped()) { 3720 set_result(newcopy); 3721 } 3722 return true; 3723 } 3724 3725 3726 //----------------------generate_virtual_guard--------------------------- 3727 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3728 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3729 RegionNode* slow_region) { 3730 ciMethod* method = callee(); 3731 int vtable_index = method->vtable_index(); 3732 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3733 err_msg_res("bad index %d", vtable_index)); 3734 // Get the Method* out of the appropriate vtable entry. 3735 int entry_offset = (InstanceKlass::vtable_start_offset() + 3736 vtable_index*vtableEntry::size()) * wordSize + 3737 vtableEntry::method_offset_in_bytes(); 3738 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3739 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS); 3740 3741 // Compare the target method with the expected method (e.g., Object.hashCode). 3742 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3743 3744 Node* native_call = makecon(native_call_addr); 3745 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call)); 3746 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne)); 3747 3748 return generate_slow_guard(test_native, slow_region); 3749 } 3750 3751 //-----------------------generate_method_call---------------------------- 3752 // Use generate_method_call to make a slow-call to the real 3753 // method if the fast path fails. An alternative would be to 3754 // use a stub like OptoRuntime::slow_arraycopy_Java. 3755 // This only works for expanding the current library call, 3756 // not another intrinsic. (E.g., don't use this for making an 3757 // arraycopy call inside of the copyOf intrinsic.) 3758 CallJavaNode* 3759 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3760 // When compiling the intrinsic method itself, do not use this technique. 3761 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3762 3763 ciMethod* method = callee(); 3764 // ensure the JVMS we have will be correct for this call 3765 guarantee(method_id == method->intrinsic_id(), "must match"); 3766 3767 const TypeFunc* tf = TypeFunc::make(method); 3768 CallJavaNode* slow_call; 3769 if (is_static) { 3770 assert(!is_virtual, ""); 3771 slow_call = new(C) CallStaticJavaNode(C, tf, 3772 SharedRuntime::get_resolve_static_call_stub(), 3773 method, bci()); 3774 } else if (is_virtual) { 3775 null_check_receiver(); 3776 int vtable_index = Method::invalid_vtable_index; 3777 if (UseInlineCaches) { 3778 // Suppress the vtable call 3779 } else { 3780 // hashCode and clone are not a miranda methods, 3781 // so the vtable index is fixed. 3782 // No need to use the linkResolver to get it. 3783 vtable_index = method->vtable_index(); 3784 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3785 err_msg_res("bad index %d", vtable_index)); 3786 } 3787 slow_call = new(C) CallDynamicJavaNode(tf, 3788 SharedRuntime::get_resolve_virtual_call_stub(), 3789 method, vtable_index, bci()); 3790 } else { // neither virtual nor static: opt_virtual 3791 null_check_receiver(); 3792 slow_call = new(C) CallStaticJavaNode(C, tf, 3793 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3794 method, bci()); 3795 slow_call->set_optimized_virtual(true); 3796 } 3797 set_arguments_for_java_call(slow_call); 3798 set_edges_for_java_call(slow_call); 3799 return slow_call; 3800 } 3801 3802 3803 //------------------------------inline_native_hashcode-------------------- 3804 // Build special case code for calls to hashCode on an object. 3805 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3806 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3807 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3808 3809 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3810 3811 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3812 PhiNode* result_val = new(C) PhiNode(result_reg, 3813 TypeInt::INT); 3814 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3815 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3816 TypePtr::BOTTOM); 3817 Node* obj = NULL; 3818 if (!is_static) { 3819 // Check for hashing null object 3820 obj = null_check_receiver(); 3821 if (stopped()) return true; // unconditionally null 3822 result_reg->init_req(_null_path, top()); 3823 result_val->init_req(_null_path, top()); 3824 } else { 3825 // Do a null check, and return zero if null. 3826 // System.identityHashCode(null) == 0 3827 obj = argument(0); 3828 Node* null_ctl = top(); 3829 obj = null_check_oop(obj, &null_ctl); 3830 result_reg->init_req(_null_path, null_ctl); 3831 result_val->init_req(_null_path, _gvn.intcon(0)); 3832 } 3833 3834 // Unconditionally null? Then return right away. 3835 if (stopped()) { 3836 set_control( result_reg->in(_null_path)); 3837 if (!stopped()) 3838 set_result(result_val->in(_null_path)); 3839 return true; 3840 } 3841 3842 // After null check, get the object's klass. 3843 Node* obj_klass = load_object_klass(obj); 3844 3845 // This call may be virtual (invokevirtual) or bound (invokespecial). 3846 // For each case we generate slightly different code. 3847 3848 // We only go to the fast case code if we pass a number of guards. The 3849 // paths which do not pass are accumulated in the slow_region. 3850 RegionNode* slow_region = new (C) RegionNode(1); 3851 record_for_igvn(slow_region); 3852 3853 // If this is a virtual call, we generate a funny guard. We pull out 3854 // the vtable entry corresponding to hashCode() from the target object. 3855 // If the target method which we are calling happens to be the native 3856 // Object hashCode() method, we pass the guard. We do not need this 3857 // guard for non-virtual calls -- the caller is known to be the native 3858 // Object hashCode(). 3859 if (is_virtual) { 3860 generate_virtual_guard(obj_klass, slow_region); 3861 } 3862 3863 // Get the header out of the object, use LoadMarkNode when available 3864 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3865 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type()); 3866 3867 // Test the header to see if it is unlocked. 3868 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3869 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask)); 3870 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3871 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val)); 3872 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne)); 3873 3874 generate_slow_guard(test_unlocked, slow_region); 3875 3876 // Get the hash value and check to see that it has been properly assigned. 3877 // We depend on hash_mask being at most 32 bits and avoid the use of 3878 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3879 // vm: see markOop.hpp. 3880 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3881 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3882 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift)); 3883 // This hack lets the hash bits live anywhere in the mark object now, as long 3884 // as the shift drops the relevant bits into the low 32 bits. Note that 3885 // Java spec says that HashCode is an int so there's no point in capturing 3886 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3887 hshifted_header = ConvX2I(hshifted_header); 3888 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask)); 3889 3890 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3891 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val)); 3892 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq)); 3893 3894 generate_slow_guard(test_assigned, slow_region); 3895 3896 Node* init_mem = reset_memory(); 3897 // fill in the rest of the null path: 3898 result_io ->init_req(_null_path, i_o()); 3899 result_mem->init_req(_null_path, init_mem); 3900 3901 result_val->init_req(_fast_path, hash_val); 3902 result_reg->init_req(_fast_path, control()); 3903 result_io ->init_req(_fast_path, i_o()); 3904 result_mem->init_req(_fast_path, init_mem); 3905 3906 // Generate code for the slow case. We make a call to hashCode(). 3907 set_control(_gvn.transform(slow_region)); 3908 if (!stopped()) { 3909 // No need for PreserveJVMState, because we're using up the present state. 3910 set_all_memory(init_mem); 3911 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 3912 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3913 Node* slow_result = set_results_for_java_call(slow_call); 3914 // this->control() comes from set_results_for_java_call 3915 result_reg->init_req(_slow_path, control()); 3916 result_val->init_req(_slow_path, slow_result); 3917 result_io ->set_req(_slow_path, i_o()); 3918 result_mem ->set_req(_slow_path, reset_memory()); 3919 } 3920 3921 // Return the combined state. 3922 set_i_o( _gvn.transform(result_io) ); 3923 set_all_memory( _gvn.transform(result_mem)); 3924 3925 set_result(result_reg, result_val); 3926 return true; 3927 } 3928 3929 //---------------------------inline_native_getClass---------------------------- 3930 // public final native Class<?> java.lang.Object.getClass(); 3931 // 3932 // Build special case code for calls to getClass on an object. 3933 bool LibraryCallKit::inline_native_getClass() { 3934 Node* obj = null_check_receiver(); 3935 if (stopped()) return true; 3936 set_result(load_mirror_from_klass(load_object_klass(obj))); 3937 return true; 3938 } 3939 3940 //-----------------inline_native_Reflection_getCallerClass--------------------- 3941 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 3942 // 3943 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3944 // 3945 // NOTE: This code must perform the same logic as JVM_GetCallerClass 3946 // in that it must skip particular security frames and checks for 3947 // caller sensitive methods. 3948 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3949 #ifndef PRODUCT 3950 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3951 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3952 } 3953 #endif 3954 3955 if (!jvms()->has_method()) { 3956 #ifndef PRODUCT 3957 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3958 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3959 } 3960 #endif 3961 return false; 3962 } 3963 3964 // Walk back up the JVM state to find the caller at the required 3965 // depth. 3966 JVMState* caller_jvms = jvms(); 3967 3968 // Cf. JVM_GetCallerClass 3969 // NOTE: Start the loop at depth 1 because the current JVM state does 3970 // not include the Reflection.getCallerClass() frame. 3971 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 3972 ciMethod* m = caller_jvms->method(); 3973 switch (n) { 3974 case 0: 3975 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 3976 break; 3977 case 1: 3978 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 3979 if (!m->caller_sensitive()) { 3980 #ifndef PRODUCT 3981 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3982 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 3983 } 3984 #endif 3985 return false; // bail-out; let JVM_GetCallerClass do the work 3986 } 3987 break; 3988 default: 3989 if (!m->is_ignored_by_security_stack_walk()) { 3990 // We have reached the desired frame; return the holder class. 3991 // Acquire method holder as java.lang.Class and push as constant. 3992 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 3993 ciInstance* caller_mirror = caller_klass->java_mirror(); 3994 set_result(makecon(TypeInstPtr::make(caller_mirror))); 3995 3996 #ifndef PRODUCT 3997 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3998 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()); 3999 tty->print_cr(" JVM state at this point:"); 4000 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4001 ciMethod* m = jvms()->of_depth(i)->method(); 4002 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4003 } 4004 } 4005 #endif 4006 return true; 4007 } 4008 break; 4009 } 4010 } 4011 4012 #ifndef PRODUCT 4013 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 4014 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4015 tty->print_cr(" JVM state at this point:"); 4016 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4017 ciMethod* m = jvms()->of_depth(i)->method(); 4018 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4019 } 4020 } 4021 #endif 4022 4023 return false; // bail-out; let JVM_GetCallerClass do the work 4024 } 4025 4026 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4027 Node* arg = argument(0); 4028 Node* result; 4029 4030 switch (id) { 4031 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break; 4032 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break; 4033 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break; 4034 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break; 4035 4036 case vmIntrinsics::_doubleToLongBits: { 4037 // two paths (plus control) merge in a wood 4038 RegionNode *r = new (C) RegionNode(3); 4039 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 4040 4041 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg)); 4042 // Build the boolean node 4043 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4044 4045 // Branch either way. 4046 // NaN case is less traveled, which makes all the difference. 4047 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4048 Node *opt_isnan = _gvn.transform(ifisnan); 4049 assert( opt_isnan->is_If(), "Expect an IfNode"); 4050 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4051 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4052 4053 set_control(iftrue); 4054 4055 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4056 Node *slow_result = longcon(nan_bits); // return NaN 4057 phi->init_req(1, _gvn.transform( slow_result )); 4058 r->init_req(1, iftrue); 4059 4060 // Else fall through 4061 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4062 set_control(iffalse); 4063 4064 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg))); 4065 r->init_req(2, iffalse); 4066 4067 // Post merge 4068 set_control(_gvn.transform(r)); 4069 record_for_igvn(r); 4070 4071 C->set_has_split_ifs(true); // Has chance for split-if optimization 4072 result = phi; 4073 assert(result->bottom_type()->isa_long(), "must be"); 4074 break; 4075 } 4076 4077 case vmIntrinsics::_floatToIntBits: { 4078 // two paths (plus control) merge in a wood 4079 RegionNode *r = new (C) RegionNode(3); 4080 Node *phi = new (C) PhiNode(r, TypeInt::INT); 4081 4082 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg)); 4083 // Build the boolean node 4084 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4085 4086 // Branch either way. 4087 // NaN case is less traveled, which makes all the difference. 4088 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4089 Node *opt_isnan = _gvn.transform(ifisnan); 4090 assert( opt_isnan->is_If(), "Expect an IfNode"); 4091 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4092 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4093 4094 set_control(iftrue); 4095 4096 static const jint nan_bits = 0x7fc00000; 4097 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4098 phi->init_req(1, _gvn.transform( slow_result )); 4099 r->init_req(1, iftrue); 4100 4101 // Else fall through 4102 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4103 set_control(iffalse); 4104 4105 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg))); 4106 r->init_req(2, iffalse); 4107 4108 // Post merge 4109 set_control(_gvn.transform(r)); 4110 record_for_igvn(r); 4111 4112 C->set_has_split_ifs(true); // Has chance for split-if optimization 4113 result = phi; 4114 assert(result->bottom_type()->isa_int(), "must be"); 4115 break; 4116 } 4117 4118 default: 4119 fatal_unexpected_iid(id); 4120 break; 4121 } 4122 set_result(_gvn.transform(result)); 4123 return true; 4124 } 4125 4126 #ifdef _LP64 4127 #define XTOP ,top() /*additional argument*/ 4128 #else //_LP64 4129 #define XTOP /*no additional argument*/ 4130 #endif //_LP64 4131 4132 //----------------------inline_unsafe_copyMemory------------------------- 4133 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4134 bool LibraryCallKit::inline_unsafe_copyMemory() { 4135 if (callee()->is_static()) return false; // caller must have the capability! 4136 null_check_receiver(); // null-check receiver 4137 if (stopped()) return true; 4138 4139 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4140 4141 Node* src_ptr = argument(1); // type: oop 4142 Node* src_off = ConvL2X(argument(2)); // type: long 4143 Node* dst_ptr = argument(4); // type: oop 4144 Node* dst_off = ConvL2X(argument(5)); // type: long 4145 Node* size = ConvL2X(argument(7)); // type: long 4146 4147 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4148 "fieldOffset must be byte-scaled"); 4149 4150 Node* src = make_unsafe_address(src_ptr, src_off); 4151 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4152 4153 // Conservatively insert a memory barrier on all memory slices. 4154 // Do not let writes of the copy source or destination float below the copy. 4155 insert_mem_bar(Op_MemBarCPUOrder); 4156 4157 // Call it. Note that the length argument is not scaled. 4158 make_runtime_call(RC_LEAF|RC_NO_FP, 4159 OptoRuntime::fast_arraycopy_Type(), 4160 StubRoutines::unsafe_arraycopy(), 4161 "unsafe_arraycopy", 4162 TypeRawPtr::BOTTOM, 4163 src, dst, size XTOP); 4164 4165 // Do not let reads of the copy destination float above the copy. 4166 insert_mem_bar(Op_MemBarCPUOrder); 4167 4168 return true; 4169 } 4170 4171 //------------------------clone_coping----------------------------------- 4172 // Helper function for inline_native_clone. 4173 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4174 assert(obj_size != NULL, ""); 4175 Node* raw_obj = alloc_obj->in(1); 4176 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4177 4178 AllocateNode* alloc = NULL; 4179 if (ReduceBulkZeroing) { 4180 // We will be completely responsible for initializing this object - 4181 // mark Initialize node as complete. 4182 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4183 // The object was just allocated - there should be no any stores! 4184 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4185 // Mark as complete_with_arraycopy so that on AllocateNode 4186 // expansion, we know this AllocateNode is initialized by an array 4187 // copy and a StoreStore barrier exists after the array copy. 4188 alloc->initialization()->set_complete_with_arraycopy(); 4189 } 4190 4191 // Copy the fastest available way. 4192 // TODO: generate fields copies for small objects instead. 4193 Node* src = obj; 4194 Node* dest = alloc_obj; 4195 Node* size = _gvn.transform(obj_size); 4196 4197 // Exclude the header but include array length to copy by 8 bytes words. 4198 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4199 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4200 instanceOopDesc::base_offset_in_bytes(); 4201 // base_off: 4202 // 8 - 32-bit VM 4203 // 12 - 64-bit VM, compressed klass 4204 // 16 - 64-bit VM, normal klass 4205 if (base_off % BytesPerLong != 0) { 4206 assert(UseCompressedKlassPointers, ""); 4207 if (is_array) { 4208 // Exclude length to copy by 8 bytes words. 4209 base_off += sizeof(int); 4210 } else { 4211 // Include klass to copy by 8 bytes words. 4212 base_off = instanceOopDesc::klass_offset_in_bytes(); 4213 } 4214 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4215 } 4216 src = basic_plus_adr(src, base_off); 4217 dest = basic_plus_adr(dest, base_off); 4218 4219 // Compute the length also, if needed: 4220 Node* countx = size; 4221 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off))); 4222 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4223 4224 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4225 bool disjoint_bases = true; 4226 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4227 src, NULL, dest, NULL, countx, 4228 /*dest_uninitialized*/true); 4229 4230 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4231 if (card_mark) { 4232 assert(!is_array, ""); 4233 // Put in store barrier for any and all oops we are sticking 4234 // into this object. (We could avoid this if we could prove 4235 // that the object type contains no oop fields at all.) 4236 Node* no_particular_value = NULL; 4237 Node* no_particular_field = NULL; 4238 int raw_adr_idx = Compile::AliasIdxRaw; 4239 post_barrier(control(), 4240 memory(raw_adr_type), 4241 alloc_obj, 4242 no_particular_field, 4243 raw_adr_idx, 4244 no_particular_value, 4245 T_OBJECT, 4246 false); 4247 } 4248 4249 // Do not let reads from the cloned object float above the arraycopy. 4250 if (alloc != NULL) { 4251 // Do not let stores that initialize this object be reordered with 4252 // a subsequent store that would make this object accessible by 4253 // other threads. 4254 // Record what AllocateNode this StoreStore protects so that 4255 // escape analysis can go from the MemBarStoreStoreNode to the 4256 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4257 // based on the escape status of the AllocateNode. 4258 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4259 } else { 4260 insert_mem_bar(Op_MemBarCPUOrder); 4261 } 4262 } 4263 4264 //------------------------inline_native_clone---------------------------- 4265 // protected native Object java.lang.Object.clone(); 4266 // 4267 // Here are the simple edge cases: 4268 // null receiver => normal trap 4269 // virtual and clone was overridden => slow path to out-of-line clone 4270 // not cloneable or finalizer => slow path to out-of-line Object.clone 4271 // 4272 // The general case has two steps, allocation and copying. 4273 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4274 // 4275 // Copying also has two cases, oop arrays and everything else. 4276 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4277 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4278 // 4279 // These steps fold up nicely if and when the cloned object's klass 4280 // can be sharply typed as an object array, a type array, or an instance. 4281 // 4282 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4283 PhiNode* result_val; 4284 4285 // Set the reexecute bit for the interpreter to reexecute 4286 // the bytecode that invokes Object.clone if deoptimization happens. 4287 { PreserveReexecuteState preexecs(this); 4288 jvms()->set_should_reexecute(true); 4289 4290 Node* obj = null_check_receiver(); 4291 if (stopped()) return true; 4292 4293 Node* obj_klass = load_object_klass(obj); 4294 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4295 const TypeOopPtr* toop = ((tklass != NULL) 4296 ? tklass->as_instance_type() 4297 : TypeInstPtr::NOTNULL); 4298 4299 // Conservatively insert a memory barrier on all memory slices. 4300 // Do not let writes into the original float below the clone. 4301 insert_mem_bar(Op_MemBarCPUOrder); 4302 4303 // paths into result_reg: 4304 enum { 4305 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4306 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4307 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4308 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4309 PATH_LIMIT 4310 }; 4311 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4312 result_val = new(C) PhiNode(result_reg, 4313 TypeInstPtr::NOTNULL); 4314 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO); 4315 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 4316 TypePtr::BOTTOM); 4317 record_for_igvn(result_reg); 4318 4319 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4320 int raw_adr_idx = Compile::AliasIdxRaw; 4321 4322 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4323 if (array_ctl != NULL) { 4324 // It's an array. 4325 PreserveJVMState pjvms(this); 4326 set_control(array_ctl); 4327 Node* obj_length = load_array_length(obj); 4328 Node* obj_size = NULL; 4329 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4330 4331 if (!use_ReduceInitialCardMarks()) { 4332 // If it is an oop array, it requires very special treatment, 4333 // because card marking is required on each card of the array. 4334 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4335 if (is_obja != NULL) { 4336 PreserveJVMState pjvms2(this); 4337 set_control(is_obja); 4338 // Generate a direct call to the right arraycopy function(s). 4339 bool disjoint_bases = true; 4340 bool length_never_negative = true; 4341 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4342 obj, intcon(0), alloc_obj, intcon(0), 4343 obj_length, 4344 disjoint_bases, length_never_negative); 4345 result_reg->init_req(_objArray_path, control()); 4346 result_val->init_req(_objArray_path, alloc_obj); 4347 result_i_o ->set_req(_objArray_path, i_o()); 4348 result_mem ->set_req(_objArray_path, reset_memory()); 4349 } 4350 } 4351 // Otherwise, there are no card marks to worry about. 4352 // (We can dispense with card marks if we know the allocation 4353 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4354 // causes the non-eden paths to take compensating steps to 4355 // simulate a fresh allocation, so that no further 4356 // card marks are required in compiled code to initialize 4357 // the object.) 4358 4359 if (!stopped()) { 4360 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4361 4362 // Present the results of the copy. 4363 result_reg->init_req(_array_path, control()); 4364 result_val->init_req(_array_path, alloc_obj); 4365 result_i_o ->set_req(_array_path, i_o()); 4366 result_mem ->set_req(_array_path, reset_memory()); 4367 } 4368 } 4369 4370 // We only go to the instance fast case code if we pass a number of guards. 4371 // The paths which do not pass are accumulated in the slow_region. 4372 RegionNode* slow_region = new (C) RegionNode(1); 4373 record_for_igvn(slow_region); 4374 if (!stopped()) { 4375 // It's an instance (we did array above). Make the slow-path tests. 4376 // If this is a virtual call, we generate a funny guard. We grab 4377 // the vtable entry corresponding to clone() from the target object. 4378 // If the target method which we are calling happens to be the 4379 // Object clone() method, we pass the guard. We do not need this 4380 // guard for non-virtual calls; the caller is known to be the native 4381 // Object clone(). 4382 if (is_virtual) { 4383 generate_virtual_guard(obj_klass, slow_region); 4384 } 4385 4386 // The object must be cloneable and must not have a finalizer. 4387 // Both of these conditions may be checked in a single test. 4388 // We could optimize the cloneable test further, but we don't care. 4389 generate_access_flags_guard(obj_klass, 4390 // Test both conditions: 4391 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4392 // Must be cloneable but not finalizer: 4393 JVM_ACC_IS_CLONEABLE, 4394 slow_region); 4395 } 4396 4397 if (!stopped()) { 4398 // It's an instance, and it passed the slow-path tests. 4399 PreserveJVMState pjvms(this); 4400 Node* obj_size = NULL; 4401 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size); 4402 4403 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4404 4405 // Present the results of the slow call. 4406 result_reg->init_req(_instance_path, control()); 4407 result_val->init_req(_instance_path, alloc_obj); 4408 result_i_o ->set_req(_instance_path, i_o()); 4409 result_mem ->set_req(_instance_path, reset_memory()); 4410 } 4411 4412 // Generate code for the slow case. We make a call to clone(). 4413 set_control(_gvn.transform(slow_region)); 4414 if (!stopped()) { 4415 PreserveJVMState pjvms(this); 4416 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4417 Node* slow_result = set_results_for_java_call(slow_call); 4418 // this->control() comes from set_results_for_java_call 4419 result_reg->init_req(_slow_path, control()); 4420 result_val->init_req(_slow_path, slow_result); 4421 result_i_o ->set_req(_slow_path, i_o()); 4422 result_mem ->set_req(_slow_path, reset_memory()); 4423 } 4424 4425 // Return the combined state. 4426 set_control( _gvn.transform(result_reg)); 4427 set_i_o( _gvn.transform(result_i_o)); 4428 set_all_memory( _gvn.transform(result_mem)); 4429 } // original reexecute is set back here 4430 4431 set_result(_gvn.transform(result_val)); 4432 return true; 4433 } 4434 4435 //------------------------------basictype2arraycopy---------------------------- 4436 address LibraryCallKit::basictype2arraycopy(BasicType t, 4437 Node* src_offset, 4438 Node* dest_offset, 4439 bool disjoint_bases, 4440 const char* &name, 4441 bool dest_uninitialized) { 4442 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4443 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4444 4445 bool aligned = false; 4446 bool disjoint = disjoint_bases; 4447 4448 // if the offsets are the same, we can treat the memory regions as 4449 // disjoint, because either the memory regions are in different arrays, 4450 // or they are identical (which we can treat as disjoint.) We can also 4451 // treat a copy with a destination index less that the source index 4452 // as disjoint since a low->high copy will work correctly in this case. 4453 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4454 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4455 // both indices are constants 4456 int s_offs = src_offset_inttype->get_con(); 4457 int d_offs = dest_offset_inttype->get_con(); 4458 int element_size = type2aelembytes(t); 4459 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4460 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4461 if (s_offs >= d_offs) disjoint = true; 4462 } else if (src_offset == dest_offset && src_offset != NULL) { 4463 // This can occur if the offsets are identical non-constants. 4464 disjoint = true; 4465 } 4466 4467 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4468 } 4469 4470 4471 //------------------------------inline_arraycopy----------------------- 4472 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4473 // Object dest, int destPos, 4474 // int length); 4475 bool LibraryCallKit::inline_arraycopy() { 4476 // Get the arguments. 4477 Node* src = argument(0); // type: oop 4478 Node* src_offset = argument(1); // type: int 4479 Node* dest = argument(2); // type: oop 4480 Node* dest_offset = argument(3); // type: int 4481 Node* length = argument(4); // type: int 4482 4483 // Compile time checks. If any of these checks cannot be verified at compile time, 4484 // we do not make a fast path for this call. Instead, we let the call remain as it 4485 // is. The checks we choose to mandate at compile time are: 4486 // 4487 // (1) src and dest are arrays. 4488 const Type* src_type = src->Value(&_gvn); 4489 const Type* dest_type = dest->Value(&_gvn); 4490 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4491 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4492 if (top_src == NULL || top_src->klass() == NULL || 4493 top_dest == NULL || top_dest->klass() == NULL) { 4494 // Conservatively insert a memory barrier on all memory slices. 4495 // Do not let writes into the source float below the arraycopy. 4496 insert_mem_bar(Op_MemBarCPUOrder); 4497 4498 // Call StubRoutines::generic_arraycopy stub. 4499 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4500 src, src_offset, dest, dest_offset, length); 4501 4502 // Do not let reads from the destination float above the arraycopy. 4503 // Since we cannot type the arrays, we don't know which slices 4504 // might be affected. We could restrict this barrier only to those 4505 // memory slices which pertain to array elements--but don't bother. 4506 if (!InsertMemBarAfterArraycopy) 4507 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4508 insert_mem_bar(Op_MemBarCPUOrder); 4509 return true; 4510 } 4511 4512 // (2) src and dest arrays must have elements of the same BasicType 4513 // Figure out the size and type of the elements we will be copying. 4514 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4515 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4516 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4517 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4518 4519 if (src_elem != dest_elem || dest_elem == T_VOID) { 4520 // The component types are not the same or are not recognized. Punt. 4521 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4522 generate_slow_arraycopy(TypePtr::BOTTOM, 4523 src, src_offset, dest, dest_offset, length, 4524 /*dest_uninitialized*/false); 4525 return true; 4526 } 4527 4528 //--------------------------------------------------------------------------- 4529 // We will make a fast path for this call to arraycopy. 4530 4531 // We have the following tests left to perform: 4532 // 4533 // (3) src and dest must not be null. 4534 // (4) src_offset must not be negative. 4535 // (5) dest_offset must not be negative. 4536 // (6) length must not be negative. 4537 // (7) src_offset + length must not exceed length of src. 4538 // (8) dest_offset + length must not exceed length of dest. 4539 // (9) each element of an oop array must be assignable 4540 4541 RegionNode* slow_region = new (C) RegionNode(1); 4542 record_for_igvn(slow_region); 4543 4544 // (3) operands must not be null 4545 // We currently perform our null checks with the null_check routine. 4546 // This means that the null exceptions will be reported in the caller 4547 // rather than (correctly) reported inside of the native arraycopy call. 4548 // This should be corrected, given time. We do our null check with the 4549 // stack pointer restored. 4550 src = null_check(src, T_ARRAY); 4551 dest = null_check(dest, T_ARRAY); 4552 4553 // (4) src_offset must not be negative. 4554 generate_negative_guard(src_offset, slow_region); 4555 4556 // (5) dest_offset must not be negative. 4557 generate_negative_guard(dest_offset, slow_region); 4558 4559 // (6) length must not be negative (moved to generate_arraycopy()). 4560 // generate_negative_guard(length, slow_region); 4561 4562 // (7) src_offset + length must not exceed length of src. 4563 generate_limit_guard(src_offset, length, 4564 load_array_length(src), 4565 slow_region); 4566 4567 // (8) dest_offset + length must not exceed length of dest. 4568 generate_limit_guard(dest_offset, length, 4569 load_array_length(dest), 4570 slow_region); 4571 4572 // (9) each element of an oop array must be assignable 4573 // The generate_arraycopy subroutine checks this. 4574 4575 // This is where the memory effects are placed: 4576 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4577 generate_arraycopy(adr_type, dest_elem, 4578 src, src_offset, dest, dest_offset, length, 4579 false, false, slow_region); 4580 4581 return true; 4582 } 4583 4584 //-----------------------------generate_arraycopy---------------------- 4585 // Generate an optimized call to arraycopy. 4586 // Caller must guard against non-arrays. 4587 // Caller must determine a common array basic-type for both arrays. 4588 // Caller must validate offsets against array bounds. 4589 // The slow_region has already collected guard failure paths 4590 // (such as out of bounds length or non-conformable array types). 4591 // The generated code has this shape, in general: 4592 // 4593 // if (length == 0) return // via zero_path 4594 // slowval = -1 4595 // if (types unknown) { 4596 // slowval = call generic copy loop 4597 // if (slowval == 0) return // via checked_path 4598 // } else if (indexes in bounds) { 4599 // if ((is object array) && !(array type check)) { 4600 // slowval = call checked copy loop 4601 // if (slowval == 0) return // via checked_path 4602 // } else { 4603 // call bulk copy loop 4604 // return // via fast_path 4605 // } 4606 // } 4607 // // adjust params for remaining work: 4608 // if (slowval != -1) { 4609 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4610 // } 4611 // slow_region: 4612 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4613 // return // via slow_call_path 4614 // 4615 // This routine is used from several intrinsics: System.arraycopy, 4616 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4617 // 4618 void 4619 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4620 BasicType basic_elem_type, 4621 Node* src, Node* src_offset, 4622 Node* dest, Node* dest_offset, 4623 Node* copy_length, 4624 bool disjoint_bases, 4625 bool length_never_negative, 4626 RegionNode* slow_region) { 4627 4628 if (slow_region == NULL) { 4629 slow_region = new(C) RegionNode(1); 4630 record_for_igvn(slow_region); 4631 } 4632 4633 Node* original_dest = dest; 4634 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4635 bool dest_uninitialized = false; 4636 4637 // See if this is the initialization of a newly-allocated array. 4638 // If so, we will take responsibility here for initializing it to zero. 4639 // (Note: Because tightly_coupled_allocation performs checks on the 4640 // out-edges of the dest, we need to avoid making derived pointers 4641 // from it until we have checked its uses.) 4642 if (ReduceBulkZeroing 4643 && !ZeroTLAB // pointless if already zeroed 4644 && basic_elem_type != T_CONFLICT // avoid corner case 4645 && !src->eqv_uncast(dest) 4646 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4647 != NULL) 4648 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4649 && alloc->maybe_set_complete(&_gvn)) { 4650 // "You break it, you buy it." 4651 InitializeNode* init = alloc->initialization(); 4652 assert(init->is_complete(), "we just did this"); 4653 init->set_complete_with_arraycopy(); 4654 assert(dest->is_CheckCastPP(), "sanity"); 4655 assert(dest->in(0)->in(0) == init, "dest pinned"); 4656 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4657 // From this point on, every exit path is responsible for 4658 // initializing any non-copied parts of the object to zero. 4659 // Also, if this flag is set we make sure that arraycopy interacts properly 4660 // with G1, eliding pre-barriers. See CR 6627983. 4661 dest_uninitialized = true; 4662 } else { 4663 // No zeroing elimination here. 4664 alloc = NULL; 4665 //original_dest = dest; 4666 //dest_uninitialized = false; 4667 } 4668 4669 // Results are placed here: 4670 enum { fast_path = 1, // normal void-returning assembly stub 4671 checked_path = 2, // special assembly stub with cleanup 4672 slow_call_path = 3, // something went wrong; call the VM 4673 zero_path = 4, // bypass when length of copy is zero 4674 bcopy_path = 5, // copy primitive array by 64-bit blocks 4675 PATH_LIMIT = 6 4676 }; 4677 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT); 4678 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO); 4679 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type); 4680 record_for_igvn(result_region); 4681 _gvn.set_type_bottom(result_i_o); 4682 _gvn.set_type_bottom(result_memory); 4683 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4684 4685 // The slow_control path: 4686 Node* slow_control; 4687 Node* slow_i_o = i_o(); 4688 Node* slow_mem = memory(adr_type); 4689 debug_only(slow_control = (Node*) badAddress); 4690 4691 // Checked control path: 4692 Node* checked_control = top(); 4693 Node* checked_mem = NULL; 4694 Node* checked_i_o = NULL; 4695 Node* checked_value = NULL; 4696 4697 if (basic_elem_type == T_CONFLICT) { 4698 assert(!dest_uninitialized, ""); 4699 Node* cv = generate_generic_arraycopy(adr_type, 4700 src, src_offset, dest, dest_offset, 4701 copy_length, dest_uninitialized); 4702 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4703 checked_control = control(); 4704 checked_i_o = i_o(); 4705 checked_mem = memory(adr_type); 4706 checked_value = cv; 4707 set_control(top()); // no fast path 4708 } 4709 4710 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4711 if (not_pos != NULL) { 4712 PreserveJVMState pjvms(this); 4713 set_control(not_pos); 4714 4715 // (6) length must not be negative. 4716 if (!length_never_negative) { 4717 generate_negative_guard(copy_length, slow_region); 4718 } 4719 4720 // copy_length is 0. 4721 if (!stopped() && dest_uninitialized) { 4722 Node* dest_length = alloc->in(AllocateNode::ALength); 4723 if (copy_length->eqv_uncast(dest_length) 4724 || _gvn.find_int_con(dest_length, 1) <= 0) { 4725 // There is no zeroing to do. No need for a secondary raw memory barrier. 4726 } else { 4727 // Clear the whole thing since there are no source elements to copy. 4728 generate_clear_array(adr_type, dest, basic_elem_type, 4729 intcon(0), NULL, 4730 alloc->in(AllocateNode::AllocSize)); 4731 // Use a secondary InitializeNode as raw memory barrier. 4732 // Currently it is needed only on this path since other 4733 // paths have stub or runtime calls as raw memory barriers. 4734 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4735 Compile::AliasIdxRaw, 4736 top())->as_Initialize(); 4737 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4738 } 4739 } 4740 4741 // Present the results of the fast call. 4742 result_region->init_req(zero_path, control()); 4743 result_i_o ->init_req(zero_path, i_o()); 4744 result_memory->init_req(zero_path, memory(adr_type)); 4745 } 4746 4747 if (!stopped() && dest_uninitialized) { 4748 // We have to initialize the *uncopied* part of the array to zero. 4749 // The copy destination is the slice dest[off..off+len]. The other slices 4750 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4751 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4752 Node* dest_length = alloc->in(AllocateNode::ALength); 4753 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset, 4754 copy_length)); 4755 4756 // If there is a head section that needs zeroing, do it now. 4757 if (find_int_con(dest_offset, -1) != 0) { 4758 generate_clear_array(adr_type, dest, basic_elem_type, 4759 intcon(0), dest_offset, 4760 NULL); 4761 } 4762 4763 // Next, perform a dynamic check on the tail length. 4764 // It is often zero, and we can win big if we prove this. 4765 // There are two wins: Avoid generating the ClearArray 4766 // with its attendant messy index arithmetic, and upgrade 4767 // the copy to a more hardware-friendly word size of 64 bits. 4768 Node* tail_ctl = NULL; 4769 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 4770 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length)); 4771 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt)); 4772 tail_ctl = generate_slow_guard(bol_lt, NULL); 4773 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4774 } 4775 4776 // At this point, let's assume there is no tail. 4777 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4778 // There is no tail. Try an upgrade to a 64-bit copy. 4779 bool didit = false; 4780 { PreserveJVMState pjvms(this); 4781 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4782 src, src_offset, dest, dest_offset, 4783 dest_size, dest_uninitialized); 4784 if (didit) { 4785 // Present the results of the block-copying fast call. 4786 result_region->init_req(bcopy_path, control()); 4787 result_i_o ->init_req(bcopy_path, i_o()); 4788 result_memory->init_req(bcopy_path, memory(adr_type)); 4789 } 4790 } 4791 if (didit) 4792 set_control(top()); // no regular fast path 4793 } 4794 4795 // Clear the tail, if any. 4796 if (tail_ctl != NULL) { 4797 Node* notail_ctl = stopped() ? NULL : control(); 4798 set_control(tail_ctl); 4799 if (notail_ctl == NULL) { 4800 generate_clear_array(adr_type, dest, basic_elem_type, 4801 dest_tail, NULL, 4802 dest_size); 4803 } else { 4804 // Make a local merge. 4805 Node* done_ctl = new(C) RegionNode(3); 4806 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type); 4807 done_ctl->init_req(1, notail_ctl); 4808 done_mem->init_req(1, memory(adr_type)); 4809 generate_clear_array(adr_type, dest, basic_elem_type, 4810 dest_tail, NULL, 4811 dest_size); 4812 done_ctl->init_req(2, control()); 4813 done_mem->init_req(2, memory(adr_type)); 4814 set_control( _gvn.transform(done_ctl)); 4815 set_memory( _gvn.transform(done_mem), adr_type ); 4816 } 4817 } 4818 } 4819 4820 BasicType copy_type = basic_elem_type; 4821 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 4822 if (!stopped() && copy_type == T_OBJECT) { 4823 // If src and dest have compatible element types, we can copy bits. 4824 // Types S[] and D[] are compatible if D is a supertype of S. 4825 // 4826 // If they are not, we will use checked_oop_disjoint_arraycopy, 4827 // which performs a fast optimistic per-oop check, and backs off 4828 // further to JVM_ArrayCopy on the first per-oop check that fails. 4829 // (Actually, we don't move raw bits only; the GC requires card marks.) 4830 4831 // Get the Klass* for both src and dest 4832 Node* src_klass = load_object_klass(src); 4833 Node* dest_klass = load_object_klass(dest); 4834 4835 // Generate the subtype check. 4836 // This might fold up statically, or then again it might not. 4837 // 4838 // Non-static example: Copying List<String>.elements to a new String[]. 4839 // The backing store for a List<String> is always an Object[], 4840 // but its elements are always type String, if the generic types 4841 // are correct at the source level. 4842 // 4843 // Test S[] against D[], not S against D, because (probably) 4844 // the secondary supertype cache is less busy for S[] than S. 4845 // This usually only matters when D is an interface. 4846 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4847 // Plug failing path into checked_oop_disjoint_arraycopy 4848 if (not_subtype_ctrl != top()) { 4849 PreserveJVMState pjvms(this); 4850 set_control(not_subtype_ctrl); 4851 // (At this point we can assume disjoint_bases, since types differ.) 4852 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 4853 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 4854 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 4855 Node* dest_elem_klass = _gvn.transform(n1); 4856 Node* cv = generate_checkcast_arraycopy(adr_type, 4857 dest_elem_klass, 4858 src, src_offset, dest, dest_offset, 4859 ConvI2X(copy_length), dest_uninitialized); 4860 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4861 checked_control = control(); 4862 checked_i_o = i_o(); 4863 checked_mem = memory(adr_type); 4864 checked_value = cv; 4865 } 4866 // At this point we know we do not need type checks on oop stores. 4867 4868 // Let's see if we need card marks: 4869 if (alloc != NULL && use_ReduceInitialCardMarks()) { 4870 // If we do not need card marks, copy using the jint or jlong stub. 4871 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 4872 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 4873 "sizes agree"); 4874 } 4875 } 4876 4877 if (!stopped()) { 4878 // Generate the fast path, if possible. 4879 PreserveJVMState pjvms(this); 4880 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 4881 src, src_offset, dest, dest_offset, 4882 ConvI2X(copy_length), dest_uninitialized); 4883 4884 // Present the results of the fast call. 4885 result_region->init_req(fast_path, control()); 4886 result_i_o ->init_req(fast_path, i_o()); 4887 result_memory->init_req(fast_path, memory(adr_type)); 4888 } 4889 4890 // Here are all the slow paths up to this point, in one bundle: 4891 slow_control = top(); 4892 if (slow_region != NULL) 4893 slow_control = _gvn.transform(slow_region); 4894 DEBUG_ONLY(slow_region = (RegionNode*)badAddress); 4895 4896 set_control(checked_control); 4897 if (!stopped()) { 4898 // Clean up after the checked call. 4899 // The returned value is either 0 or -1^K, 4900 // where K = number of partially transferred array elements. 4901 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0))); 4902 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 4903 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4904 4905 // If it is 0, we are done, so transfer to the end. 4906 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff)); 4907 result_region->init_req(checked_path, checks_done); 4908 result_i_o ->init_req(checked_path, checked_i_o); 4909 result_memory->init_req(checked_path, checked_mem); 4910 4911 // If it is not zero, merge into the slow call. 4912 set_control( _gvn.transform(new(C) IfFalseNode(iff) )); 4913 RegionNode* slow_reg2 = new(C) RegionNode(3); 4914 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO); 4915 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type); 4916 record_for_igvn(slow_reg2); 4917 slow_reg2 ->init_req(1, slow_control); 4918 slow_i_o2 ->init_req(1, slow_i_o); 4919 slow_mem2 ->init_req(1, slow_mem); 4920 slow_reg2 ->init_req(2, control()); 4921 slow_i_o2 ->init_req(2, checked_i_o); 4922 slow_mem2 ->init_req(2, checked_mem); 4923 4924 slow_control = _gvn.transform(slow_reg2); 4925 slow_i_o = _gvn.transform(slow_i_o2); 4926 slow_mem = _gvn.transform(slow_mem2); 4927 4928 if (alloc != NULL) { 4929 // We'll restart from the very beginning, after zeroing the whole thing. 4930 // This can cause double writes, but that's OK since dest is brand new. 4931 // So we ignore the low 31 bits of the value returned from the stub. 4932 } else { 4933 // We must continue the copy exactly where it failed, or else 4934 // another thread might see the wrong number of writes to dest. 4935 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1))); 4936 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT); 4937 slow_offset->init_req(1, intcon(0)); 4938 slow_offset->init_req(2, checked_offset); 4939 slow_offset = _gvn.transform(slow_offset); 4940 4941 // Adjust the arguments by the conditionally incoming offset. 4942 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset)); 4943 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset)); 4944 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset)); 4945 4946 // Tweak the node variables to adjust the code produced below: 4947 src_offset = src_off_plus; 4948 dest_offset = dest_off_plus; 4949 copy_length = length_minus; 4950 } 4951 } 4952 4953 set_control(slow_control); 4954 if (!stopped()) { 4955 // Generate the slow path, if needed. 4956 PreserveJVMState pjvms(this); // replace_in_map may trash the map 4957 4958 set_memory(slow_mem, adr_type); 4959 set_i_o(slow_i_o); 4960 4961 if (dest_uninitialized) { 4962 generate_clear_array(adr_type, dest, basic_elem_type, 4963 intcon(0), NULL, 4964 alloc->in(AllocateNode::AllocSize)); 4965 } 4966 4967 generate_slow_arraycopy(adr_type, 4968 src, src_offset, dest, dest_offset, 4969 copy_length, /*dest_uninitialized*/false); 4970 4971 result_region->init_req(slow_call_path, control()); 4972 result_i_o ->init_req(slow_call_path, i_o()); 4973 result_memory->init_req(slow_call_path, memory(adr_type)); 4974 } 4975 4976 // Remove unused edges. 4977 for (uint i = 1; i < result_region->req(); i++) { 4978 if (result_region->in(i) == NULL) 4979 result_region->init_req(i, top()); 4980 } 4981 4982 // Finished; return the combined state. 4983 set_control( _gvn.transform(result_region)); 4984 set_i_o( _gvn.transform(result_i_o) ); 4985 set_memory( _gvn.transform(result_memory), adr_type ); 4986 4987 // The memory edges above are precise in order to model effects around 4988 // array copies accurately to allow value numbering of field loads around 4989 // arraycopy. Such field loads, both before and after, are common in Java 4990 // collections and similar classes involving header/array data structures. 4991 // 4992 // But with low number of register or when some registers are used or killed 4993 // by arraycopy calls it causes registers spilling on stack. See 6544710. 4994 // The next memory barrier is added to avoid it. If the arraycopy can be 4995 // optimized away (which it can, sometimes) then we can manually remove 4996 // the membar also. 4997 // 4998 // Do not let reads from the cloned object float above the arraycopy. 4999 if (alloc != NULL) { 5000 // Do not let stores that initialize this object be reordered with 5001 // a subsequent store that would make this object accessible by 5002 // other threads. 5003 // Record what AllocateNode this StoreStore protects so that 5004 // escape analysis can go from the MemBarStoreStoreNode to the 5005 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5006 // based on the escape status of the AllocateNode. 5007 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5008 } else if (InsertMemBarAfterArraycopy) 5009 insert_mem_bar(Op_MemBarCPUOrder); 5010 } 5011 5012 5013 // Helper function which determines if an arraycopy immediately follows 5014 // an allocation, with no intervening tests or other escapes for the object. 5015 AllocateArrayNode* 5016 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5017 RegionNode* slow_region) { 5018 if (stopped()) return NULL; // no fast path 5019 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5020 5021 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5022 if (alloc == NULL) return NULL; 5023 5024 Node* rawmem = memory(Compile::AliasIdxRaw); 5025 // Is the allocation's memory state untouched? 5026 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5027 // Bail out if there have been raw-memory effects since the allocation. 5028 // (Example: There might have been a call or safepoint.) 5029 return NULL; 5030 } 5031 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5032 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5033 return NULL; 5034 } 5035 5036 // There must be no unexpected observers of this allocation. 5037 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5038 Node* obs = ptr->fast_out(i); 5039 if (obs != this->map()) { 5040 return NULL; 5041 } 5042 } 5043 5044 // This arraycopy must unconditionally follow the allocation of the ptr. 5045 Node* alloc_ctl = ptr->in(0); 5046 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5047 5048 Node* ctl = control(); 5049 while (ctl != alloc_ctl) { 5050 // There may be guards which feed into the slow_region. 5051 // Any other control flow means that we might not get a chance 5052 // to finish initializing the allocated object. 5053 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5054 IfNode* iff = ctl->in(0)->as_If(); 5055 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5056 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5057 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5058 ctl = iff->in(0); // This test feeds the known slow_region. 5059 continue; 5060 } 5061 // One more try: Various low-level checks bottom out in 5062 // uncommon traps. If the debug-info of the trap omits 5063 // any reference to the allocation, as we've already 5064 // observed, then there can be no objection to the trap. 5065 bool found_trap = false; 5066 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5067 Node* obs = not_ctl->fast_out(j); 5068 if (obs->in(0) == not_ctl && obs->is_Call() && 5069 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5070 found_trap = true; break; 5071 } 5072 } 5073 if (found_trap) { 5074 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5075 continue; 5076 } 5077 } 5078 return NULL; 5079 } 5080 5081 // If we get this far, we have an allocation which immediately 5082 // precedes the arraycopy, and we can take over zeroing the new object. 5083 // The arraycopy will finish the initialization, and provide 5084 // a new control state to which we will anchor the destination pointer. 5085 5086 return alloc; 5087 } 5088 5089 // Helper for initialization of arrays, creating a ClearArray. 5090 // It writes zero bits in [start..end), within the body of an array object. 5091 // The memory effects are all chained onto the 'adr_type' alias category. 5092 // 5093 // Since the object is otherwise uninitialized, we are free 5094 // to put a little "slop" around the edges of the cleared area, 5095 // as long as it does not go back into the array's header, 5096 // or beyond the array end within the heap. 5097 // 5098 // The lower edge can be rounded down to the nearest jint and the 5099 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5100 // 5101 // Arguments: 5102 // adr_type memory slice where writes are generated 5103 // dest oop of the destination array 5104 // basic_elem_type element type of the destination 5105 // slice_idx array index of first element to store 5106 // slice_len number of elements to store (or NULL) 5107 // dest_size total size in bytes of the array object 5108 // 5109 // Exactly one of slice_len or dest_size must be non-NULL. 5110 // If dest_size is non-NULL, zeroing extends to the end of the object. 5111 // If slice_len is non-NULL, the slice_idx value must be a constant. 5112 void 5113 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5114 Node* dest, 5115 BasicType basic_elem_type, 5116 Node* slice_idx, 5117 Node* slice_len, 5118 Node* dest_size) { 5119 // one or the other but not both of slice_len and dest_size: 5120 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5121 if (slice_len == NULL) slice_len = top(); 5122 if (dest_size == NULL) dest_size = top(); 5123 5124 // operate on this memory slice: 5125 Node* mem = memory(adr_type); // memory slice to operate on 5126 5127 // scaling and rounding of indexes: 5128 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5129 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5130 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5131 int bump_bit = (-1 << scale) & BytesPerInt; 5132 5133 // determine constant starts and ends 5134 const intptr_t BIG_NEG = -128; 5135 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5136 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5137 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5138 if (slice_len_con == 0) { 5139 return; // nothing to do here 5140 } 5141 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5142 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5143 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5144 assert(end_con < 0, "not two cons"); 5145 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5146 BytesPerLong); 5147 } 5148 5149 if (start_con >= 0 && end_con >= 0) { 5150 // Constant start and end. Simple. 5151 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5152 start_con, end_con, &_gvn); 5153 } else if (start_con >= 0 && dest_size != top()) { 5154 // Constant start, pre-rounded end after the tail of the array. 5155 Node* end = dest_size; 5156 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5157 start_con, end, &_gvn); 5158 } else if (start_con >= 0 && slice_len != top()) { 5159 // Constant start, non-constant end. End needs rounding up. 5160 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5161 intptr_t end_base = abase + (slice_idx_con << scale); 5162 int end_round = (-1 << scale) & (BytesPerLong - 1); 5163 Node* end = ConvI2X(slice_len); 5164 if (scale != 0) 5165 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) )); 5166 end_base += end_round; 5167 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base))); 5168 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round))); 5169 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5170 start_con, end, &_gvn); 5171 } else if (start_con < 0 && dest_size != top()) { 5172 // Non-constant start, pre-rounded end after the tail of the array. 5173 // This is almost certainly a "round-to-end" operation. 5174 Node* start = slice_idx; 5175 start = ConvI2X(start); 5176 if (scale != 0) 5177 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) )); 5178 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase))); 5179 if ((bump_bit | clear_low) != 0) { 5180 int to_clear = (bump_bit | clear_low); 5181 // Align up mod 8, then store a jint zero unconditionally 5182 // just before the mod-8 boundary. 5183 if (((abase + bump_bit) & ~to_clear) - bump_bit 5184 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5185 bump_bit = 0; 5186 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5187 } else { 5188 // Bump 'start' up to (or past) the next jint boundary: 5189 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit))); 5190 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5191 } 5192 // Round bumped 'start' down to jlong boundary in body of array. 5193 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear))); 5194 if (bump_bit != 0) { 5195 // Store a zero to the immediately preceding jint: 5196 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit))); 5197 Node* p1 = basic_plus_adr(dest, x1); 5198 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 5199 mem = _gvn.transform(mem); 5200 } 5201 } 5202 Node* end = dest_size; // pre-rounded 5203 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5204 start, end, &_gvn); 5205 } else { 5206 // Non-constant start, unrounded non-constant end. 5207 // (Nobody zeroes a random midsection of an array using this routine.) 5208 ShouldNotReachHere(); // fix caller 5209 } 5210 5211 // Done. 5212 set_memory(mem, adr_type); 5213 } 5214 5215 5216 bool 5217 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5218 BasicType basic_elem_type, 5219 AllocateNode* alloc, 5220 Node* src, Node* src_offset, 5221 Node* dest, Node* dest_offset, 5222 Node* dest_size, bool dest_uninitialized) { 5223 // See if there is an advantage from block transfer. 5224 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5225 if (scale >= LogBytesPerLong) 5226 return false; // it is already a block transfer 5227 5228 // Look at the alignment of the starting offsets. 5229 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5230 5231 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5232 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5233 if (src_off_con < 0 || dest_off_con < 0) 5234 // At present, we can only understand constants. 5235 return false; 5236 5237 intptr_t src_off = abase + (src_off_con << scale); 5238 intptr_t dest_off = abase + (dest_off_con << scale); 5239 5240 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5241 // Non-aligned; too bad. 5242 // One more chance: Pick off an initial 32-bit word. 5243 // This is a common case, since abase can be odd mod 8. 5244 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5245 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5246 Node* sptr = basic_plus_adr(src, src_off); 5247 Node* dptr = basic_plus_adr(dest, dest_off); 5248 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 5249 store_to_memory(control(), dptr, sval, T_INT, adr_type); 5250 src_off += BytesPerInt; 5251 dest_off += BytesPerInt; 5252 } else { 5253 return false; 5254 } 5255 } 5256 assert(src_off % BytesPerLong == 0, ""); 5257 assert(dest_off % BytesPerLong == 0, ""); 5258 5259 // Do this copy by giant steps. 5260 Node* sptr = basic_plus_adr(src, src_off); 5261 Node* dptr = basic_plus_adr(dest, dest_off); 5262 Node* countx = dest_size; 5263 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off))); 5264 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong))); 5265 5266 bool disjoint_bases = true; // since alloc != NULL 5267 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5268 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5269 5270 return true; 5271 } 5272 5273 5274 // Helper function; generates code for the slow case. 5275 // We make a call to a runtime method which emulates the native method, 5276 // but without the native wrapper overhead. 5277 void 5278 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5279 Node* src, Node* src_offset, 5280 Node* dest, Node* dest_offset, 5281 Node* copy_length, bool dest_uninitialized) { 5282 assert(!dest_uninitialized, "Invariant"); 5283 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5284 OptoRuntime::slow_arraycopy_Type(), 5285 OptoRuntime::slow_arraycopy_Java(), 5286 "slow_arraycopy", adr_type, 5287 src, src_offset, dest, dest_offset, 5288 copy_length); 5289 5290 // Handle exceptions thrown by this fellow: 5291 make_slow_call_ex(call, env()->Throwable_klass(), false); 5292 } 5293 5294 // Helper function; generates code for cases requiring runtime checks. 5295 Node* 5296 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5297 Node* dest_elem_klass, 5298 Node* src, Node* src_offset, 5299 Node* dest, Node* dest_offset, 5300 Node* copy_length, bool dest_uninitialized) { 5301 if (stopped()) return NULL; 5302 5303 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5304 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5305 return NULL; 5306 } 5307 5308 // Pick out the parameters required to perform a store-check 5309 // for the target array. This is an optimistic check. It will 5310 // look in each non-null element's class, at the desired klass's 5311 // super_check_offset, for the desired klass. 5312 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5313 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5314 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr()); 5315 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5316 Node* check_value = dest_elem_klass; 5317 5318 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5319 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5320 5321 // (We know the arrays are never conjoint, because their types differ.) 5322 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5323 OptoRuntime::checkcast_arraycopy_Type(), 5324 copyfunc_addr, "checkcast_arraycopy", adr_type, 5325 // five arguments, of which two are 5326 // intptr_t (jlong in LP64) 5327 src_start, dest_start, 5328 copy_length XTOP, 5329 check_offset XTOP, 5330 check_value); 5331 5332 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5333 } 5334 5335 5336 // Helper function; generates code for cases requiring runtime checks. 5337 Node* 5338 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5339 Node* src, Node* src_offset, 5340 Node* dest, Node* dest_offset, 5341 Node* copy_length, bool dest_uninitialized) { 5342 assert(!dest_uninitialized, "Invariant"); 5343 if (stopped()) return NULL; 5344 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5345 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5346 return NULL; 5347 } 5348 5349 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5350 OptoRuntime::generic_arraycopy_Type(), 5351 copyfunc_addr, "generic_arraycopy", adr_type, 5352 src, src_offset, dest, dest_offset, copy_length); 5353 5354 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5355 } 5356 5357 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5358 void 5359 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5360 BasicType basic_elem_type, 5361 bool disjoint_bases, 5362 Node* src, Node* src_offset, 5363 Node* dest, Node* dest_offset, 5364 Node* copy_length, bool dest_uninitialized) { 5365 if (stopped()) return; // nothing to do 5366 5367 Node* src_start = src; 5368 Node* dest_start = dest; 5369 if (src_offset != NULL || dest_offset != NULL) { 5370 assert(src_offset != NULL && dest_offset != NULL, ""); 5371 src_start = array_element_address(src, src_offset, basic_elem_type); 5372 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5373 } 5374 5375 // Figure out which arraycopy runtime method to call. 5376 const char* copyfunc_name = "arraycopy"; 5377 address copyfunc_addr = 5378 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5379 disjoint_bases, copyfunc_name, dest_uninitialized); 5380 5381 // Call it. Note that the count_ix value is not scaled to a byte-size. 5382 make_runtime_call(RC_LEAF|RC_NO_FP, 5383 OptoRuntime::fast_arraycopy_Type(), 5384 copyfunc_addr, copyfunc_name, adr_type, 5385 src_start, dest_start, copy_length XTOP); 5386 } 5387 5388 //-------------inline_encodeISOArray----------------------------------- 5389 // encode char[] to byte[] in ISO_8859_1 5390 bool LibraryCallKit::inline_encodeISOArray() { 5391 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5392 // no receiver since it is static method 5393 Node *src = argument(0); 5394 Node *src_offset = argument(1); 5395 Node *dst = argument(2); 5396 Node *dst_offset = argument(3); 5397 Node *length = argument(4); 5398 5399 const Type* src_type = src->Value(&_gvn); 5400 const Type* dst_type = dst->Value(&_gvn); 5401 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5402 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5403 if (top_src == NULL || top_src->klass() == NULL || 5404 top_dest == NULL || top_dest->klass() == NULL) { 5405 // failed array check 5406 return false; 5407 } 5408 5409 // Figure out the size and type of the elements we will be copying. 5410 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5411 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5412 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5413 return false; 5414 } 5415 Node* src_start = array_element_address(src, src_offset, src_elem); 5416 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5417 // 'src_start' points to src array + scaled offset 5418 // 'dst_start' points to dst array + scaled offset 5419 5420 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5421 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5422 enc = _gvn.transform(enc); 5423 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc)); 5424 set_memory(res_mem, mtype); 5425 set_result(enc); 5426 return true; 5427 } 5428 5429 /** 5430 * Calculate CRC32 for byte. 5431 * int java.util.zip.CRC32.update(int crc, int b) 5432 */ 5433 bool LibraryCallKit::inline_updateCRC32() { 5434 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5435 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5436 // no receiver since it is static method 5437 Node* crc = argument(0); // type: int 5438 Node* b = argument(1); // type: int 5439 5440 /* 5441 * int c = ~ crc; 5442 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5443 * b = b ^ (c >>> 8); 5444 * crc = ~b; 5445 */ 5446 5447 Node* M1 = intcon(-1); 5448 crc = _gvn.transform(new (C) XorINode(crc, M1)); 5449 Node* result = _gvn.transform(new (C) XorINode(crc, b)); 5450 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF))); 5451 5452 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5453 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2))); 5454 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5455 result = make_load(control(), adr, TypeInt::INT, T_INT); 5456 5457 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8))); 5458 result = _gvn.transform(new (C) XorINode(crc, result)); 5459 result = _gvn.transform(new (C) XorINode(result, M1)); 5460 set_result(result); 5461 return true; 5462 } 5463 5464 /** 5465 * Calculate CRC32 for byte[] array. 5466 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5467 */ 5468 bool LibraryCallKit::inline_updateBytesCRC32() { 5469 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5470 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5471 // no receiver since it is static method 5472 Node* crc = argument(0); // type: int 5473 Node* src = argument(1); // type: oop 5474 Node* offset = argument(2); // type: int 5475 Node* length = argument(3); // type: int 5476 5477 const Type* src_type = src->Value(&_gvn); 5478 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5479 if (top_src == NULL || top_src->klass() == NULL) { 5480 // failed array check 5481 return false; 5482 } 5483 5484 // Figure out the size and type of the elements we will be copying. 5485 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5486 if (src_elem != T_BYTE) { 5487 return false; 5488 } 5489 5490 // 'src_start' points to src array + scaled offset 5491 Node* src_start = array_element_address(src, offset, src_elem); 5492 5493 // We assume that range check is done by caller. 5494 // TODO: generate range check (offset+length < src.length) in debug VM. 5495 5496 // Call the stub. 5497 address stubAddr = StubRoutines::updateBytesCRC32(); 5498 const char *stubName = "updateBytesCRC32"; 5499 5500 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5501 stubAddr, stubName, TypePtr::BOTTOM, 5502 crc, src_start, length); 5503 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5504 set_result(result); 5505 return true; 5506 } 5507 5508 /** 5509 * Calculate CRC32 for ByteBuffer. 5510 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5511 */ 5512 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5513 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5514 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5515 // no receiver since it is static method 5516 Node* crc = argument(0); // type: int 5517 Node* src = argument(1); // type: long 5518 Node* offset = argument(3); // type: int 5519 Node* length = argument(4); // type: int 5520 5521 src = ConvL2X(src); // adjust Java long to machine word 5522 Node* base = _gvn.transform(new (C) CastX2PNode(src)); 5523 offset = ConvI2X(offset); 5524 5525 // 'src_start' points to src array + scaled offset 5526 Node* src_start = basic_plus_adr(top(), base, offset); 5527 5528 // Call the stub. 5529 address stubAddr = StubRoutines::updateBytesCRC32(); 5530 const char *stubName = "updateBytesCRC32"; 5531 5532 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5533 stubAddr, stubName, TypePtr::BOTTOM, 5534 crc, src_start, length); 5535 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5536 set_result(result); 5537 return true; 5538 } 5539 5540 //----------------------------inline_reference_get---------------------------- 5541 // public T java.lang.ref.Reference.get(); 5542 bool LibraryCallKit::inline_reference_get() { 5543 const int referent_offset = java_lang_ref_Reference::referent_offset; 5544 guarantee(referent_offset > 0, "should have already been set"); 5545 5546 // Get the argument: 5547 Node* reference_obj = null_check_receiver(); 5548 if (stopped()) return true; 5549 5550 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5551 5552 ciInstanceKlass* klass = env()->Object_klass(); 5553 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5554 5555 Node* no_ctrl = NULL; 5556 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT); 5557 5558 // Use the pre-barrier to record the value in the referent field 5559 pre_barrier(false /* do_load */, 5560 control(), 5561 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5562 result /* pre_val */, 5563 T_OBJECT); 5564 5565 // Add memory barrier to prevent commoning reads from this field 5566 // across safepoint since GC can change its value. 5567 insert_mem_bar(Op_MemBarCPUOrder); 5568 5569 set_result(result); 5570 return true; 5571 } 5572 5573 5574 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5575 bool is_exact=true, bool is_static=false) { 5576 5577 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5578 assert(tinst != NULL, "obj is null"); 5579 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5580 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5581 5582 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5583 ciSymbol::make(fieldTypeString), 5584 is_static); 5585 if (field == NULL) return (Node *) NULL; 5586 assert (field != NULL, "undefined field"); 5587 5588 // Next code copied from Parse::do_get_xxx(): 5589 5590 // Compute address and memory type. 5591 int offset = field->offset_in_bytes(); 5592 bool is_vol = field->is_volatile(); 5593 ciType* field_klass = field->type(); 5594 assert(field_klass->is_loaded(), "should be loaded"); 5595 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5596 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5597 BasicType bt = field->layout_type(); 5598 5599 // Build the resultant type of the load 5600 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5601 5602 // Build the load. 5603 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol); 5604 return loadedField; 5605 } 5606 5607 5608 //------------------------------inline_aescrypt_Block----------------------- 5609 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5610 address stubAddr; 5611 const char *stubName; 5612 assert(UseAES, "need AES instruction support"); 5613 5614 switch(id) { 5615 case vmIntrinsics::_aescrypt_encryptBlock: 5616 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5617 stubName = "aescrypt_encryptBlock"; 5618 break; 5619 case vmIntrinsics::_aescrypt_decryptBlock: 5620 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5621 stubName = "aescrypt_decryptBlock"; 5622 break; 5623 } 5624 if (stubAddr == NULL) return false; 5625 5626 Node* aescrypt_object = argument(0); 5627 Node* src = argument(1); 5628 Node* src_offset = argument(2); 5629 Node* dest = argument(3); 5630 Node* dest_offset = argument(4); 5631 5632 // (1) src and dest are arrays. 5633 const Type* src_type = src->Value(&_gvn); 5634 const Type* dest_type = dest->Value(&_gvn); 5635 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5636 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5637 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5638 5639 // for the quick and dirty code we will skip all the checks. 5640 // we are just trying to get the call to be generated. 5641 Node* src_start = src; 5642 Node* dest_start = dest; 5643 if (src_offset != NULL || dest_offset != NULL) { 5644 assert(src_offset != NULL && dest_offset != NULL, ""); 5645 src_start = array_element_address(src, src_offset, T_BYTE); 5646 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5647 } 5648 5649 // now need to get the start of its expanded key array 5650 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5651 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5652 if (k_start == NULL) return false; 5653 5654 // Call the stub. 5655 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5656 stubAddr, stubName, TypePtr::BOTTOM, 5657 src_start, dest_start, k_start); 5658 5659 return true; 5660 } 5661 5662 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5663 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5664 address stubAddr; 5665 const char *stubName; 5666 5667 assert(UseAES, "need AES instruction support"); 5668 5669 switch(id) { 5670 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5671 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5672 stubName = "cipherBlockChaining_encryptAESCrypt"; 5673 break; 5674 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5675 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5676 stubName = "cipherBlockChaining_decryptAESCrypt"; 5677 break; 5678 } 5679 if (stubAddr == NULL) return false; 5680 5681 Node* cipherBlockChaining_object = argument(0); 5682 Node* src = argument(1); 5683 Node* src_offset = argument(2); 5684 Node* len = argument(3); 5685 Node* dest = argument(4); 5686 Node* dest_offset = argument(5); 5687 5688 // (1) src and dest are arrays. 5689 const Type* src_type = src->Value(&_gvn); 5690 const Type* dest_type = dest->Value(&_gvn); 5691 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5692 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5693 assert (top_src != NULL && top_src->klass() != NULL 5694 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5695 5696 // checks are the responsibility of the caller 5697 Node* src_start = src; 5698 Node* dest_start = dest; 5699 if (src_offset != NULL || dest_offset != NULL) { 5700 assert(src_offset != NULL && dest_offset != NULL, ""); 5701 src_start = array_element_address(src, src_offset, T_BYTE); 5702 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5703 } 5704 5705 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5706 // (because of the predicated logic executed earlier). 5707 // so we cast it here safely. 5708 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5709 5710 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5711 if (embeddedCipherObj == NULL) return false; 5712 5713 // cast it to what we know it will be at runtime 5714 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5715 assert(tinst != NULL, "CBC obj is null"); 5716 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5717 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5718 if (!klass_AESCrypt->is_loaded()) return false; 5719 5720 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5721 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5722 const TypeOopPtr* xtype = aklass->as_instance_type(); 5723 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype); 5724 aescrypt_object = _gvn.transform(aescrypt_object); 5725 5726 // we need to get the start of the aescrypt_object's expanded key array 5727 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5728 if (k_start == NULL) return false; 5729 5730 // similarly, get the start address of the r vector 5731 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5732 if (objRvec == NULL) return false; 5733 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5734 5735 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5736 make_runtime_call(RC_LEAF|RC_NO_FP, 5737 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5738 stubAddr, stubName, TypePtr::BOTTOM, 5739 src_start, dest_start, k_start, r_start, len); 5740 5741 // return is void so no result needs to be pushed 5742 5743 return true; 5744 } 5745 5746 //------------------------------get_key_start_from_aescrypt_object----------------------- 5747 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5748 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5749 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5750 if (objAESCryptKey == NULL) return (Node *) NULL; 5751 5752 // now have the array, need to get the start address of the K array 5753 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5754 return k_start; 5755 } 5756 5757 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5758 // Return node representing slow path of predicate check. 5759 // the pseudo code we want to emulate with this predicate is: 5760 // for encryption: 5761 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5762 // for decryption: 5763 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5764 // note cipher==plain is more conservative than the original java code but that's OK 5765 // 5766 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5767 // First, check receiver for NULL since it is virtual method. 5768 Node* objCBC = argument(0); 5769 objCBC = null_check(objCBC); 5770 5771 if (stopped()) return NULL; // Always NULL 5772 5773 // Load embeddedCipher field of CipherBlockChaining object. 5774 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5775 5776 // get AESCrypt klass for instanceOf check 5777 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5778 // will have same classloader as CipherBlockChaining object 5779 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5780 assert(tinst != NULL, "CBCobj is null"); 5781 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5782 5783 // we want to do an instanceof comparison against the AESCrypt class 5784 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5785 if (!klass_AESCrypt->is_loaded()) { 5786 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5787 Node* ctrl = control(); 5788 set_control(top()); // no regular fast path 5789 return ctrl; 5790 } 5791 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5792 5793 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5794 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1))); 5795 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); 5796 5797 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5798 5799 // for encryption, we are done 5800 if (!decrypting) 5801 return instof_false; // even if it is NULL 5802 5803 // for decryption, we need to add a further check to avoid 5804 // taking the intrinsic path when cipher and plain are the same 5805 // see the original java code for why. 5806 RegionNode* region = new(C) RegionNode(3); 5807 region->init_req(1, instof_false); 5808 Node* src = argument(1); 5809 Node* dest = argument(4); 5810 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest)); 5811 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq)); 5812 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5813 region->init_req(2, src_dest_conjoint); 5814 5815 record_for_igvn(region); 5816 return _gvn.transform(region); 5817 }