1 /* 2 * Copyright (c) 1999, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "oops/objArrayKlass.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/c2compiler.hpp" 36 #include "opto/callGenerator.hpp" 37 #include "opto/castnode.hpp" 38 #include "opto/cfgnode.hpp" 39 #include "opto/convertnode.hpp" 40 #include "opto/countbitsnode.hpp" 41 #include "opto/intrinsicnode.hpp" 42 #include "opto/idealKit.hpp" 43 #include "opto/mathexactnode.hpp" 44 #include "opto/movenode.hpp" 45 #include "opto/mulnode.hpp" 46 #include "opto/narrowptrnode.hpp" 47 #include "opto/opaquenode.hpp" 48 #include "opto/parse.hpp" 49 #include "opto/runtime.hpp" 50 #include "opto/subnode.hpp" 51 #include "prims/nativeLookup.hpp" 52 #include "prims/unsafe.hpp" 53 #include "runtime/sharedRuntime.hpp" 54 #ifdef TRACE_HAVE_INTRINSICS 55 #include "trace/traceMacros.hpp" 56 #endif 57 58 class LibraryIntrinsic : public InlineCallGenerator { 59 // Extend the set of intrinsics known to the runtime: 60 public: 61 private: 62 bool _is_virtual; 63 bool _does_virtual_dispatch; 64 int8_t _predicates_count; // Intrinsic is predicated by several conditions 65 int8_t _last_predicate; // Last generated predicate 66 vmIntrinsics::ID _intrinsic_id; 67 68 public: 69 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 70 : InlineCallGenerator(m), 71 _is_virtual(is_virtual), 72 _does_virtual_dispatch(does_virtual_dispatch), 73 _predicates_count((int8_t)predicates_count), 74 _last_predicate((int8_t)-1), 75 _intrinsic_id(id) 76 { 77 } 78 virtual bool is_intrinsic() const { return true; } 79 virtual bool is_virtual() const { return _is_virtual; } 80 virtual bool is_predicated() const { return _predicates_count > 0; } 81 virtual int predicates_count() const { return _predicates_count; } 82 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 83 virtual JVMState* generate(JVMState* jvms); 84 virtual Node* generate_predicate(JVMState* jvms, int predicate); 85 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 86 }; 87 88 89 // Local helper class for LibraryIntrinsic: 90 class LibraryCallKit : public GraphKit { 91 private: 92 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 93 Node* _result; // the result node, if any 94 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 95 96 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type); 97 98 public: 99 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 100 : GraphKit(jvms), 101 _intrinsic(intrinsic), 102 _result(NULL) 103 { 104 // Check if this is a root compile. In that case we don't have a caller. 105 if (!jvms->has_method()) { 106 _reexecute_sp = sp(); 107 } else { 108 // Find out how many arguments the interpreter needs when deoptimizing 109 // and save the stack pointer value so it can used by uncommon_trap. 110 // We find the argument count by looking at the declared signature. 111 bool ignored_will_link; 112 ciSignature* declared_signature = NULL; 113 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 114 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 115 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 116 } 117 } 118 119 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 120 121 ciMethod* caller() const { return jvms()->method(); } 122 int bci() const { return jvms()->bci(); } 123 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 124 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 125 ciMethod* callee() const { return _intrinsic->method(); } 126 127 bool try_to_inline(int predicate); 128 Node* try_to_predicate(int predicate); 129 130 void push_result() { 131 // Push the result onto the stack. 132 if (!stopped() && result() != NULL) { 133 BasicType bt = result()->bottom_type()->basic_type(); 134 push_node(bt, result()); 135 } 136 } 137 138 private: 139 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 140 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 141 } 142 143 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 144 void set_result(RegionNode* region, PhiNode* value); 145 Node* result() { return _result; } 146 147 virtual int reexecute_sp() { return _reexecute_sp; } 148 149 // Helper functions to inline natives 150 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 151 Node* generate_slow_guard(Node* test, RegionNode* region); 152 Node* generate_fair_guard(Node* test, RegionNode* region); 153 Node* generate_negative_guard(Node* index, RegionNode* region, 154 // resulting CastII of index: 155 Node* *pos_index = NULL); 156 Node* generate_limit_guard(Node* offset, Node* subseq_length, 157 Node* array_length, 158 RegionNode* region); 159 void generate_string_range_check(Node* array, Node* offset, 160 Node* length, bool char_count); 161 Node* generate_current_thread(Node* &tls_output); 162 Node* load_mirror_from_klass(Node* klass); 163 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 164 RegionNode* region, int null_path, 165 int offset); 166 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 167 RegionNode* region, int null_path) { 168 int offset = java_lang_Class::klass_offset_in_bytes(); 169 return load_klass_from_mirror_common(mirror, never_see_null, 170 region, null_path, 171 offset); 172 } 173 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 174 RegionNode* region, int null_path) { 175 int offset = java_lang_Class::array_klass_offset_in_bytes(); 176 return load_klass_from_mirror_common(mirror, never_see_null, 177 region, null_path, 178 offset); 179 } 180 Node* generate_access_flags_guard(Node* kls, 181 int modifier_mask, int modifier_bits, 182 RegionNode* region); 183 Node* generate_interface_guard(Node* kls, RegionNode* region); 184 Node* generate_array_guard(Node* kls, RegionNode* region) { 185 return generate_array_guard_common(kls, region, false, false); 186 } 187 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 188 return generate_array_guard_common(kls, region, false, true); 189 } 190 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 191 return generate_array_guard_common(kls, region, true, false); 192 } 193 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 194 return generate_array_guard_common(kls, region, true, true); 195 } 196 Node* generate_array_guard_common(Node* kls, RegionNode* region, 197 bool obj_array, bool not_array); 198 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 199 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 200 bool is_virtual = false, bool is_static = false); 201 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 202 return generate_method_call(method_id, false, true); 203 } 204 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 205 return generate_method_call(method_id, true, false); 206 } 207 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 208 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 209 210 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 211 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 212 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 213 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 214 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 215 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); 216 bool inline_string_indexOfChar(); 217 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 218 bool inline_string_toBytesU(); 219 bool inline_string_getCharsU(); 220 bool inline_string_copy(bool compress); 221 bool inline_string_char_access(bool is_store); 222 Node* round_double_node(Node* n); 223 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 224 bool inline_math_native(vmIntrinsics::ID id); 225 bool inline_trig(vmIntrinsics::ID id); 226 bool inline_math(vmIntrinsics::ID id); 227 template <typename OverflowOp> 228 bool inline_math_overflow(Node* arg1, Node* arg2); 229 void inline_math_mathExact(Node* math, Node* test); 230 bool inline_math_addExactI(bool is_increment); 231 bool inline_math_addExactL(bool is_increment); 232 bool inline_math_multiplyExactI(); 233 bool inline_math_multiplyExactL(); 234 bool inline_math_negateExactI(); 235 bool inline_math_negateExactL(); 236 bool inline_math_subtractExactI(bool is_decrement); 237 bool inline_math_subtractExactL(bool is_decrement); 238 bool inline_min_max(vmIntrinsics::ID id); 239 bool inline_notify(vmIntrinsics::ID id); 240 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 241 // This returns Type::AnyPtr, RawPtr, or OopPtr. 242 int classify_unsafe_addr(Node* &base, Node* &offset); 243 Node* make_unsafe_address(Node* base, Node* offset); 244 // Helper for inline_unsafe_access. 245 // Generates the guards that check whether the result of 246 // Unsafe.getObject should be recorded in an SATB log buffer. 247 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 248 249 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind; 250 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned); 251 static bool klass_needs_init_guard(Node* kls); 252 bool inline_unsafe_allocate(); 253 bool inline_unsafe_newArray(bool uninitialized); 254 bool inline_unsafe_copyMemory(); 255 bool inline_native_currentThread(); 256 257 bool inline_native_time_funcs(address method, const char* funcName); 258 bool inline_native_isInterrupted(); 259 bool inline_native_Class_query(vmIntrinsics::ID id); 260 bool inline_native_subtype_check(); 261 bool inline_native_getLength(); 262 bool inline_array_copyOf(bool is_copyOfRange); 263 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 264 bool inline_preconditions_checkIndex(); 265 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 266 bool inline_native_clone(bool is_virtual); 267 bool inline_native_Reflection_getCallerClass(); 268 // Helper function for inlining native object hash method 269 bool inline_native_hashcode(bool is_virtual, bool is_static); 270 bool inline_native_getClass(); 271 272 // Helper functions for inlining arraycopy 273 bool inline_arraycopy(); 274 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 275 RegionNode* slow_region); 276 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 277 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 278 279 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind; 280 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store); 281 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind); 282 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind); 283 bool inline_unsafe_fence(vmIntrinsics::ID id); 284 bool inline_onspinwait(); 285 bool inline_fp_conversions(vmIntrinsics::ID id); 286 bool inline_number_methods(vmIntrinsics::ID id); 287 bool inline_reference_get(); 288 bool inline_Class_cast(); 289 bool inline_aescrypt_Block(vmIntrinsics::ID id); 290 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 291 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 292 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 293 Node* inline_counterMode_AESCrypt_predicate(); 294 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 295 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 296 bool inline_ghash_processBlocks(); 297 bool inline_sha_implCompress(vmIntrinsics::ID id); 298 bool inline_digestBase_implCompressMB(int predicate); 299 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 300 bool long_state, address stubAddr, const char *stubName, 301 Node* src_start, Node* ofs, Node* limit); 302 Node* get_state_from_sha_object(Node *sha_object); 303 Node* get_state_from_sha5_object(Node *sha_object); 304 Node* inline_digestBase_implCompressMB_predicate(int predicate); 305 bool inline_encodeISOArray(); 306 bool inline_updateCRC32(); 307 bool inline_updateBytesCRC32(); 308 bool inline_updateByteBufferCRC32(); 309 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 310 bool inline_updateBytesCRC32C(); 311 bool inline_updateDirectByteBufferCRC32C(); 312 bool inline_updateBytesAdler32(); 313 bool inline_updateByteBufferAdler32(); 314 bool inline_multiplyToLen(); 315 bool inline_hasNegatives(); 316 bool inline_squareToLen(); 317 bool inline_mulAdd(); 318 bool inline_montgomeryMultiply(); 319 bool inline_montgomerySquare(); 320 bool inline_vectorizedMismatch(); 321 322 bool inline_profileBoolean(); 323 bool inline_isCompileConstant(); 324 }; 325 326 //---------------------------make_vm_intrinsic---------------------------- 327 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 328 vmIntrinsics::ID id = m->intrinsic_id(); 329 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 330 331 if (!m->is_loaded()) { 332 // Do not attempt to inline unloaded methods. 333 return NULL; 334 } 335 336 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 337 bool is_available = false; 338 339 { 340 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 341 // the compiler must transition to '_thread_in_vm' state because both 342 // methods access VM-internal data. 343 VM_ENTRY_MARK; 344 methodHandle mh(THREAD, m->get_Method()); 345 is_available = compiler->is_intrinsic_supported(mh, is_virtual) && 346 !C->directive()->is_intrinsic_disabled(mh) && 347 !vmIntrinsics::is_disabled_by_flags(mh); 348 349 } 350 351 if (is_available) { 352 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 353 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 354 return new LibraryIntrinsic(m, is_virtual, 355 vmIntrinsics::predicates_needed(id), 356 vmIntrinsics::does_virtual_dispatch(id), 357 (vmIntrinsics::ID) id); 358 } else { 359 return NULL; 360 } 361 } 362 363 //----------------------register_library_intrinsics----------------------- 364 // Initialize this file's data structures, for each Compile instance. 365 void Compile::register_library_intrinsics() { 366 // Nothing to do here. 367 } 368 369 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 370 LibraryCallKit kit(jvms, this); 371 Compile* C = kit.C; 372 int nodes = C->unique(); 373 #ifndef PRODUCT 374 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 375 char buf[1000]; 376 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 377 tty->print_cr("Intrinsic %s", str); 378 } 379 #endif 380 ciMethod* callee = kit.callee(); 381 const int bci = kit.bci(); 382 383 // Try to inline the intrinsic. 384 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 385 kit.try_to_inline(_last_predicate)) { 386 if (C->print_intrinsics() || C->print_inlining()) { 387 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 388 } 389 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 390 if (C->log()) { 391 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 392 vmIntrinsics::name_at(intrinsic_id()), 393 (is_virtual() ? " virtual='1'" : ""), 394 C->unique() - nodes); 395 } 396 // Push the result from the inlined method onto the stack. 397 kit.push_result(); 398 C->print_inlining_update(this); 399 return kit.transfer_exceptions_into_jvms(); 400 } 401 402 // The intrinsic bailed out 403 if (C->print_intrinsics() || C->print_inlining()) { 404 if (jvms->has_method()) { 405 // Not a root compile. 406 const char* msg; 407 if (callee->intrinsic_candidate()) { 408 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 409 } else { 410 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 411 : "failed to inline (intrinsic), method not annotated"; 412 } 413 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 414 } else { 415 // Root compile 416 tty->print("Did not generate intrinsic %s%s at bci:%d in", 417 vmIntrinsics::name_at(intrinsic_id()), 418 (is_virtual() ? " (virtual)" : ""), bci); 419 } 420 } 421 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 422 C->print_inlining_update(this); 423 return NULL; 424 } 425 426 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 427 LibraryCallKit kit(jvms, this); 428 Compile* C = kit.C; 429 int nodes = C->unique(); 430 _last_predicate = predicate; 431 #ifndef PRODUCT 432 assert(is_predicated() && predicate < predicates_count(), "sanity"); 433 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 434 char buf[1000]; 435 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 436 tty->print_cr("Predicate for intrinsic %s", str); 437 } 438 #endif 439 ciMethod* callee = kit.callee(); 440 const int bci = kit.bci(); 441 442 Node* slow_ctl = kit.try_to_predicate(predicate); 443 if (!kit.failing()) { 444 if (C->print_intrinsics() || C->print_inlining()) { 445 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 446 } 447 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 448 if (C->log()) { 449 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 450 vmIntrinsics::name_at(intrinsic_id()), 451 (is_virtual() ? " virtual='1'" : ""), 452 C->unique() - nodes); 453 } 454 return slow_ctl; // Could be NULL if the check folds. 455 } 456 457 // The intrinsic bailed out 458 if (C->print_intrinsics() || C->print_inlining()) { 459 if (jvms->has_method()) { 460 // Not a root compile. 461 const char* msg = "failed to generate predicate for intrinsic"; 462 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 463 } else { 464 // Root compile 465 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 466 vmIntrinsics::name_at(intrinsic_id()), 467 (is_virtual() ? " (virtual)" : ""), bci); 468 } 469 } 470 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 471 return NULL; 472 } 473 474 bool LibraryCallKit::try_to_inline(int predicate) { 475 // Handle symbolic names for otherwise undistinguished boolean switches: 476 const bool is_store = true; 477 const bool is_compress = true; 478 const bool is_static = true; 479 const bool is_volatile = true; 480 481 if (!jvms()->has_method()) { 482 // Root JVMState has a null method. 483 assert(map()->memory()->Opcode() == Op_Parm, ""); 484 // Insert the memory aliasing node 485 set_all_memory(reset_memory()); 486 } 487 assert(merged_memory(), ""); 488 489 490 switch (intrinsic_id()) { 491 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 492 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 493 case vmIntrinsics::_getClass: return inline_native_getClass(); 494 495 case vmIntrinsics::_dsin: 496 case vmIntrinsics::_dcos: 497 case vmIntrinsics::_dtan: 498 case vmIntrinsics::_dabs: 499 case vmIntrinsics::_datan2: 500 case vmIntrinsics::_dsqrt: 501 case vmIntrinsics::_dexp: 502 case vmIntrinsics::_dlog: 503 case vmIntrinsics::_dlog10: 504 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 505 506 case vmIntrinsics::_min: 507 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 508 509 case vmIntrinsics::_notify: 510 case vmIntrinsics::_notifyAll: 511 if (InlineNotify) { 512 return inline_notify(intrinsic_id()); 513 } 514 return false; 515 516 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 517 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 518 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 519 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 520 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 521 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 522 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 523 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 524 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 525 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 526 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 527 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 528 529 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 530 531 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 532 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 533 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 534 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 535 536 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 537 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 538 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 539 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 540 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 541 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 542 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 543 544 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 545 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 546 547 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 548 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 549 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 550 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 551 552 case vmIntrinsics::_compressStringC: 553 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 554 case vmIntrinsics::_inflateStringC: 555 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 556 557 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 558 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 559 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 560 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 561 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 562 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 563 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 564 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 565 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 566 567 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 568 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 569 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 570 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 571 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 572 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 573 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 574 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 575 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 576 577 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 578 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 579 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 580 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 581 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 582 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 583 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 584 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 585 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 586 587 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 588 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 589 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 590 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 591 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 592 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 593 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 594 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 595 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 596 597 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 598 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 599 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 600 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 601 602 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 603 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 604 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 605 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 606 607 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 608 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 609 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 610 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 611 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 612 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 613 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 614 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 615 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 616 617 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 618 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 619 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 620 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 621 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 622 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 623 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 624 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 625 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 626 627 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 628 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 629 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 630 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 631 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 632 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 633 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 634 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 635 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 636 637 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 638 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 639 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 640 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 641 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 642 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 643 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 644 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 645 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 646 647 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 648 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 649 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 650 651 case vmIntrinsics::_weakCompareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 652 case vmIntrinsics::_weakCompareAndSwapObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 653 case vmIntrinsics::_weakCompareAndSwapObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 654 case vmIntrinsics::_weakCompareAndSwapObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 655 case vmIntrinsics::_weakCompareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 656 case vmIntrinsics::_weakCompareAndSwapIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 657 case vmIntrinsics::_weakCompareAndSwapIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 658 case vmIntrinsics::_weakCompareAndSwapIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 659 case vmIntrinsics::_weakCompareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 660 case vmIntrinsics::_weakCompareAndSwapLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 661 case vmIntrinsics::_weakCompareAndSwapLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 662 case vmIntrinsics::_weakCompareAndSwapLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 663 664 case vmIntrinsics::_compareAndExchangeObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 665 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 666 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 667 case vmIntrinsics::_compareAndExchangeIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 668 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 669 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 670 case vmIntrinsics::_compareAndExchangeLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 671 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 672 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 673 674 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 675 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 676 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 677 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 678 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 679 680 case vmIntrinsics::_loadFence: 681 case vmIntrinsics::_storeFence: 682 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 683 684 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 685 686 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 687 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 688 689 #ifdef TRACE_HAVE_INTRINSICS 690 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 691 #endif 692 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 693 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 694 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 695 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 696 case vmIntrinsics::_getLength: return inline_native_getLength(); 697 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 698 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 699 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 700 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 701 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 702 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 703 704 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 705 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 706 707 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 708 709 case vmIntrinsics::_isInstance: 710 case vmIntrinsics::_getModifiers: 711 case vmIntrinsics::_isInterface: 712 case vmIntrinsics::_isArray: 713 case vmIntrinsics::_isPrimitive: 714 case vmIntrinsics::_getSuperclass: 715 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 716 717 case vmIntrinsics::_floatToRawIntBits: 718 case vmIntrinsics::_floatToIntBits: 719 case vmIntrinsics::_intBitsToFloat: 720 case vmIntrinsics::_doubleToRawLongBits: 721 case vmIntrinsics::_doubleToLongBits: 722 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 723 724 case vmIntrinsics::_numberOfLeadingZeros_i: 725 case vmIntrinsics::_numberOfLeadingZeros_l: 726 case vmIntrinsics::_numberOfTrailingZeros_i: 727 case vmIntrinsics::_numberOfTrailingZeros_l: 728 case vmIntrinsics::_bitCount_i: 729 case vmIntrinsics::_bitCount_l: 730 case vmIntrinsics::_reverseBytes_i: 731 case vmIntrinsics::_reverseBytes_l: 732 case vmIntrinsics::_reverseBytes_s: 733 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 734 735 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 736 737 case vmIntrinsics::_Reference_get: return inline_reference_get(); 738 739 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 740 741 case vmIntrinsics::_aescrypt_encryptBlock: 742 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 743 744 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 745 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 746 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 747 748 case vmIntrinsics::_counterMode_AESCrypt: 749 return inline_counterMode_AESCrypt(intrinsic_id()); 750 751 case vmIntrinsics::_sha_implCompress: 752 case vmIntrinsics::_sha2_implCompress: 753 case vmIntrinsics::_sha5_implCompress: 754 return inline_sha_implCompress(intrinsic_id()); 755 756 case vmIntrinsics::_digestBase_implCompressMB: 757 return inline_digestBase_implCompressMB(predicate); 758 759 case vmIntrinsics::_multiplyToLen: 760 return inline_multiplyToLen(); 761 762 case vmIntrinsics::_squareToLen: 763 return inline_squareToLen(); 764 765 case vmIntrinsics::_mulAdd: 766 return inline_mulAdd(); 767 768 case vmIntrinsics::_montgomeryMultiply: 769 return inline_montgomeryMultiply(); 770 case vmIntrinsics::_montgomerySquare: 771 return inline_montgomerySquare(); 772 773 case vmIntrinsics::_vectorizedMismatch: 774 return inline_vectorizedMismatch(); 775 776 case vmIntrinsics::_ghash_processBlocks: 777 return inline_ghash_processBlocks(); 778 779 case vmIntrinsics::_encodeISOArray: 780 case vmIntrinsics::_encodeByteISOArray: 781 return inline_encodeISOArray(); 782 783 case vmIntrinsics::_updateCRC32: 784 return inline_updateCRC32(); 785 case vmIntrinsics::_updateBytesCRC32: 786 return inline_updateBytesCRC32(); 787 case vmIntrinsics::_updateByteBufferCRC32: 788 return inline_updateByteBufferCRC32(); 789 790 case vmIntrinsics::_updateBytesCRC32C: 791 return inline_updateBytesCRC32C(); 792 case vmIntrinsics::_updateDirectByteBufferCRC32C: 793 return inline_updateDirectByteBufferCRC32C(); 794 795 case vmIntrinsics::_updateBytesAdler32: 796 return inline_updateBytesAdler32(); 797 case vmIntrinsics::_updateByteBufferAdler32: 798 return inline_updateByteBufferAdler32(); 799 800 case vmIntrinsics::_profileBoolean: 801 return inline_profileBoolean(); 802 case vmIntrinsics::_isCompileConstant: 803 return inline_isCompileConstant(); 804 805 case vmIntrinsics::_hasNegatives: 806 return inline_hasNegatives(); 807 808 default: 809 // If you get here, it may be that someone has added a new intrinsic 810 // to the list in vmSymbols.hpp without implementing it here. 811 #ifndef PRODUCT 812 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 813 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 814 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 815 } 816 #endif 817 return false; 818 } 819 } 820 821 Node* LibraryCallKit::try_to_predicate(int predicate) { 822 if (!jvms()->has_method()) { 823 // Root JVMState has a null method. 824 assert(map()->memory()->Opcode() == Op_Parm, ""); 825 // Insert the memory aliasing node 826 set_all_memory(reset_memory()); 827 } 828 assert(merged_memory(), ""); 829 830 switch (intrinsic_id()) { 831 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 832 return inline_cipherBlockChaining_AESCrypt_predicate(false); 833 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 834 return inline_cipherBlockChaining_AESCrypt_predicate(true); 835 case vmIntrinsics::_counterMode_AESCrypt: 836 return inline_counterMode_AESCrypt_predicate(); 837 case vmIntrinsics::_digestBase_implCompressMB: 838 return inline_digestBase_implCompressMB_predicate(predicate); 839 840 default: 841 // If you get here, it may be that someone has added a new intrinsic 842 // to the list in vmSymbols.hpp without implementing it here. 843 #ifndef PRODUCT 844 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 845 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 846 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 847 } 848 #endif 849 Node* slow_ctl = control(); 850 set_control(top()); // No fast path instrinsic 851 return slow_ctl; 852 } 853 } 854 855 //------------------------------set_result------------------------------- 856 // Helper function for finishing intrinsics. 857 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 858 record_for_igvn(region); 859 set_control(_gvn.transform(region)); 860 set_result( _gvn.transform(value)); 861 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 862 } 863 864 //------------------------------generate_guard--------------------------- 865 // Helper function for generating guarded fast-slow graph structures. 866 // The given 'test', if true, guards a slow path. If the test fails 867 // then a fast path can be taken. (We generally hope it fails.) 868 // In all cases, GraphKit::control() is updated to the fast path. 869 // The returned value represents the control for the slow path. 870 // The return value is never 'top'; it is either a valid control 871 // or NULL if it is obvious that the slow path can never be taken. 872 // Also, if region and the slow control are not NULL, the slow edge 873 // is appended to the region. 874 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 875 if (stopped()) { 876 // Already short circuited. 877 return NULL; 878 } 879 880 // Build an if node and its projections. 881 // If test is true we take the slow path, which we assume is uncommon. 882 if (_gvn.type(test) == TypeInt::ZERO) { 883 // The slow branch is never taken. No need to build this guard. 884 return NULL; 885 } 886 887 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 888 889 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 890 if (if_slow == top()) { 891 // The slow branch is never taken. No need to build this guard. 892 return NULL; 893 } 894 895 if (region != NULL) 896 region->add_req(if_slow); 897 898 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 899 set_control(if_fast); 900 901 return if_slow; 902 } 903 904 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 905 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 906 } 907 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 908 return generate_guard(test, region, PROB_FAIR); 909 } 910 911 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 912 Node* *pos_index) { 913 if (stopped()) 914 return NULL; // already stopped 915 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 916 return NULL; // index is already adequately typed 917 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 918 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 919 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 920 if (is_neg != NULL && pos_index != NULL) { 921 // Emulate effect of Parse::adjust_map_after_if. 922 Node* ccast = new CastIINode(index, TypeInt::POS); 923 ccast->set_req(0, control()); 924 (*pos_index) = _gvn.transform(ccast); 925 } 926 return is_neg; 927 } 928 929 // Make sure that 'position' is a valid limit index, in [0..length]. 930 // There are two equivalent plans for checking this: 931 // A. (offset + copyLength) unsigned<= arrayLength 932 // B. offset <= (arrayLength - copyLength) 933 // We require that all of the values above, except for the sum and 934 // difference, are already known to be non-negative. 935 // Plan A is robust in the face of overflow, if offset and copyLength 936 // are both hugely positive. 937 // 938 // Plan B is less direct and intuitive, but it does not overflow at 939 // all, since the difference of two non-negatives is always 940 // representable. Whenever Java methods must perform the equivalent 941 // check they generally use Plan B instead of Plan A. 942 // For the moment we use Plan A. 943 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 944 Node* subseq_length, 945 Node* array_length, 946 RegionNode* region) { 947 if (stopped()) 948 return NULL; // already stopped 949 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 950 if (zero_offset && subseq_length->eqv_uncast(array_length)) 951 return NULL; // common case of whole-array copy 952 Node* last = subseq_length; 953 if (!zero_offset) // last += offset 954 last = _gvn.transform(new AddINode(last, offset)); 955 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 956 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 957 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 958 return is_over; 959 } 960 961 // Emit range checks for the given String.value byte array 962 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 963 if (stopped()) { 964 return; // already stopped 965 } 966 RegionNode* bailout = new RegionNode(1); 967 record_for_igvn(bailout); 968 if (char_count) { 969 // Convert char count to byte count 970 count = _gvn.transform(new LShiftINode(count, intcon(1))); 971 } 972 973 // Offset and count must not be negative 974 generate_negative_guard(offset, bailout); 975 generate_negative_guard(count, bailout); 976 // Offset + count must not exceed length of array 977 generate_limit_guard(offset, count, load_array_length(array), bailout); 978 979 if (bailout->req() > 1) { 980 PreserveJVMState pjvms(this); 981 set_control(_gvn.transform(bailout)); 982 uncommon_trap(Deoptimization::Reason_intrinsic, 983 Deoptimization::Action_maybe_recompile); 984 } 985 } 986 987 //--------------------------generate_current_thread-------------------- 988 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 989 ciKlass* thread_klass = env()->Thread_klass(); 990 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 991 Node* thread = _gvn.transform(new ThreadLocalNode()); 992 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 993 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 994 tls_output = thread; 995 return threadObj; 996 } 997 998 999 //------------------------------make_string_method_node------------------------ 1000 // Helper method for String intrinsic functions. This version is called with 1001 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 1002 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 1003 // containing the lengths of str1 and str2. 1004 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 1005 Node* result = NULL; 1006 switch (opcode) { 1007 case Op_StrIndexOf: 1008 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 1009 str1_start, cnt1, str2_start, cnt2, ae); 1010 break; 1011 case Op_StrComp: 1012 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 1013 str1_start, cnt1, str2_start, cnt2, ae); 1014 break; 1015 case Op_StrEquals: 1016 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 1017 // Use the constant length if there is one because optimized match rule may exist. 1018 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 1019 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 1020 break; 1021 default: 1022 ShouldNotReachHere(); 1023 return NULL; 1024 } 1025 1026 // All these intrinsics have checks. 1027 C->set_has_split_ifs(true); // Has chance for split-if optimization 1028 1029 return _gvn.transform(result); 1030 } 1031 1032 //------------------------------inline_string_compareTo------------------------ 1033 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 1034 Node* arg1 = argument(0); 1035 Node* arg2 = argument(1); 1036 1037 // Get start addr and length of first argument 1038 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1039 Node* arg1_cnt = load_array_length(arg1); 1040 1041 // Get start addr and length of second argument 1042 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1043 Node* arg2_cnt = load_array_length(arg2); 1044 1045 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1046 set_result(result); 1047 return true; 1048 } 1049 1050 //------------------------------inline_string_equals------------------------ 1051 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1052 Node* arg1 = argument(0); 1053 Node* arg2 = argument(1); 1054 1055 // paths (plus control) merge 1056 RegionNode* region = new RegionNode(3); 1057 Node* phi = new PhiNode(region, TypeInt::BOOL); 1058 1059 if (!stopped()) { 1060 // Get start addr and length of first argument 1061 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1062 Node* arg1_cnt = load_array_length(arg1); 1063 1064 // Get start addr and length of second argument 1065 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1066 Node* arg2_cnt = load_array_length(arg2); 1067 1068 // Check for arg1_cnt != arg2_cnt 1069 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1070 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1071 Node* if_ne = generate_slow_guard(bol, NULL); 1072 if (if_ne != NULL) { 1073 phi->init_req(2, intcon(0)); 1074 region->init_req(2, if_ne); 1075 } 1076 1077 // Check for count == 0 is done by assembler code for StrEquals. 1078 1079 if (!stopped()) { 1080 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1081 phi->init_req(1, equals); 1082 region->init_req(1, control()); 1083 } 1084 } 1085 1086 // post merge 1087 set_control(_gvn.transform(region)); 1088 record_for_igvn(region); 1089 1090 set_result(_gvn.transform(phi)); 1091 return true; 1092 } 1093 1094 //------------------------------inline_array_equals---------------------------- 1095 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1096 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1097 Node* arg1 = argument(0); 1098 Node* arg2 = argument(1); 1099 1100 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1101 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1102 return true; 1103 } 1104 1105 //------------------------------inline_hasNegatives------------------------------ 1106 bool LibraryCallKit::inline_hasNegatives() { 1107 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1108 return false; 1109 } 1110 1111 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1112 // no receiver since it is static method 1113 Node* ba = argument(0); 1114 Node* offset = argument(1); 1115 Node* len = argument(2); 1116 1117 // Range checks 1118 generate_string_range_check(ba, offset, len, false); 1119 if (stopped()) { 1120 return true; 1121 } 1122 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1123 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1124 set_result(_gvn.transform(result)); 1125 return true; 1126 } 1127 1128 bool LibraryCallKit::inline_preconditions_checkIndex() { 1129 Node* index = argument(0); 1130 Node* length = argument(1); 1131 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1132 return false; 1133 } 1134 1135 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1136 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1137 1138 { 1139 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1140 uncommon_trap(Deoptimization::Reason_intrinsic, 1141 Deoptimization::Action_make_not_entrant); 1142 } 1143 1144 if (stopped()) { 1145 return false; 1146 } 1147 1148 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1149 BoolTest::mask btest = BoolTest::lt; 1150 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1151 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1152 _gvn.set_type(rc, rc->Value(&_gvn)); 1153 if (!rc_bool->is_Con()) { 1154 record_for_igvn(rc); 1155 } 1156 set_control(_gvn.transform(new IfTrueNode(rc))); 1157 { 1158 PreserveJVMState pjvms(this); 1159 set_control(_gvn.transform(new IfFalseNode(rc))); 1160 uncommon_trap(Deoptimization::Reason_range_check, 1161 Deoptimization::Action_make_not_entrant); 1162 } 1163 1164 if (stopped()) { 1165 return false; 1166 } 1167 1168 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1169 result->set_req(0, control()); 1170 result = _gvn.transform(result); 1171 set_result(result); 1172 replace_in_map(index, result); 1173 return true; 1174 } 1175 1176 //------------------------------inline_string_indexOf------------------------ 1177 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1178 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1179 return false; 1180 } 1181 Node* src = argument(0); 1182 Node* tgt = argument(1); 1183 1184 // Make the merge point 1185 RegionNode* result_rgn = new RegionNode(4); 1186 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1187 1188 // Get start addr and length of source string 1189 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1190 Node* src_count = load_array_length(src); 1191 1192 // Get start addr and length of substring 1193 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1194 Node* tgt_count = load_array_length(tgt); 1195 1196 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1197 // Divide src size by 2 if String is UTF16 encoded 1198 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1199 } 1200 if (ae == StrIntrinsicNode::UU) { 1201 // Divide substring size by 2 if String is UTF16 encoded 1202 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1203 } 1204 1205 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1206 if (result != NULL) { 1207 result_phi->init_req(3, result); 1208 result_rgn->init_req(3, control()); 1209 } 1210 set_control(_gvn.transform(result_rgn)); 1211 record_for_igvn(result_rgn); 1212 set_result(_gvn.transform(result_phi)); 1213 1214 return true; 1215 } 1216 1217 //-----------------------------inline_string_indexOf----------------------- 1218 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1219 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1220 return false; 1221 } 1222 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1223 return false; 1224 } 1225 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1226 Node* src = argument(0); // byte[] 1227 Node* src_count = argument(1); // char count 1228 Node* tgt = argument(2); // byte[] 1229 Node* tgt_count = argument(3); // char count 1230 Node* from_index = argument(4); // char index 1231 1232 // Multiply byte array index by 2 if String is UTF16 encoded 1233 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1234 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1235 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1236 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1237 1238 // Range checks 1239 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1240 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1241 if (stopped()) { 1242 return true; 1243 } 1244 1245 RegionNode* region = new RegionNode(5); 1246 Node* phi = new PhiNode(region, TypeInt::INT); 1247 1248 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1249 if (result != NULL) { 1250 // The result is index relative to from_index if substring was found, -1 otherwise. 1251 // Generate code which will fold into cmove. 1252 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1253 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1254 1255 Node* if_lt = generate_slow_guard(bol, NULL); 1256 if (if_lt != NULL) { 1257 // result == -1 1258 phi->init_req(3, result); 1259 region->init_req(3, if_lt); 1260 } 1261 if (!stopped()) { 1262 result = _gvn.transform(new AddINode(result, from_index)); 1263 phi->init_req(4, result); 1264 region->init_req(4, control()); 1265 } 1266 } 1267 1268 set_control(_gvn.transform(region)); 1269 record_for_igvn(region); 1270 set_result(_gvn.transform(phi)); 1271 1272 return true; 1273 } 1274 1275 // Create StrIndexOfNode with fast path checks 1276 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1277 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1278 // Check for substr count > string count 1279 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1280 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1281 Node* if_gt = generate_slow_guard(bol, NULL); 1282 if (if_gt != NULL) { 1283 phi->init_req(1, intcon(-1)); 1284 region->init_req(1, if_gt); 1285 } 1286 if (!stopped()) { 1287 // Check for substr count == 0 1288 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1289 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1290 Node* if_zero = generate_slow_guard(bol, NULL); 1291 if (if_zero != NULL) { 1292 phi->init_req(2, intcon(0)); 1293 region->init_req(2, if_zero); 1294 } 1295 } 1296 if (!stopped()) { 1297 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1298 } 1299 return NULL; 1300 } 1301 1302 //-----------------------------inline_string_indexOfChar----------------------- 1303 bool LibraryCallKit::inline_string_indexOfChar() { 1304 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1305 return false; 1306 } 1307 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1308 return false; 1309 } 1310 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1311 Node* src = argument(0); // byte[] 1312 Node* tgt = argument(1); // tgt is int ch 1313 Node* from_index = argument(2); 1314 Node* max = argument(3); 1315 1316 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1317 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1318 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1319 1320 // Range checks 1321 generate_string_range_check(src, src_offset, src_count, true); 1322 if (stopped()) { 1323 return true; 1324 } 1325 1326 RegionNode* region = new RegionNode(3); 1327 Node* phi = new PhiNode(region, TypeInt::INT); 1328 1329 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1330 C->set_has_split_ifs(true); // Has chance for split-if optimization 1331 _gvn.transform(result); 1332 1333 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1334 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1335 1336 Node* if_lt = generate_slow_guard(bol, NULL); 1337 if (if_lt != NULL) { 1338 // result == -1 1339 phi->init_req(2, result); 1340 region->init_req(2, if_lt); 1341 } 1342 if (!stopped()) { 1343 result = _gvn.transform(new AddINode(result, from_index)); 1344 phi->init_req(1, result); 1345 region->init_req(1, control()); 1346 } 1347 set_control(_gvn.transform(region)); 1348 record_for_igvn(region); 1349 set_result(_gvn.transform(phi)); 1350 1351 return true; 1352 } 1353 //---------------------------inline_string_copy--------------------- 1354 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1355 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1356 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1357 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1358 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1359 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1360 bool LibraryCallKit::inline_string_copy(bool compress) { 1361 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1362 return false; 1363 } 1364 int nargs = 5; // 2 oops, 3 ints 1365 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1366 1367 Node* src = argument(0); 1368 Node* src_offset = argument(1); 1369 Node* dst = argument(2); 1370 Node* dst_offset = argument(3); 1371 Node* length = argument(4); 1372 1373 // Check for allocation before we add nodes that would confuse 1374 // tightly_coupled_allocation() 1375 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1376 1377 // Figure out the size and type of the elements we will be copying. 1378 const Type* src_type = src->Value(&_gvn); 1379 const Type* dst_type = dst->Value(&_gvn); 1380 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1381 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1382 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1383 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1384 "Unsupported array types for inline_string_copy"); 1385 1386 // Range checks 1387 generate_string_range_check(src, src_offset, length, compress && src_elem == T_BYTE); 1388 generate_string_range_check(dst, dst_offset, length, !compress && dst_elem == T_BYTE); 1389 if (stopped()) { 1390 return true; 1391 } 1392 1393 // Convert char[] offsets to byte[] offsets 1394 if (compress && src_elem == T_BYTE) { 1395 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1396 } else if (!compress && dst_elem == T_BYTE) { 1397 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1398 } 1399 1400 Node* src_start = array_element_address(src, src_offset, src_elem); 1401 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1402 // 'src_start' points to src array + scaled offset 1403 // 'dst_start' points to dst array + scaled offset 1404 Node* count = NULL; 1405 if (compress) { 1406 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1407 } else { 1408 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1409 } 1410 1411 if (alloc != NULL) { 1412 if (alloc->maybe_set_complete(&_gvn)) { 1413 // "You break it, you buy it." 1414 InitializeNode* init = alloc->initialization(); 1415 assert(init->is_complete(), "we just did this"); 1416 init->set_complete_with_arraycopy(); 1417 assert(dst->is_CheckCastPP(), "sanity"); 1418 assert(dst->in(0)->in(0) == init, "dest pinned"); 1419 } 1420 // Do not let stores that initialize this object be reordered with 1421 // a subsequent store that would make this object accessible by 1422 // other threads. 1423 // Record what AllocateNode this StoreStore protects so that 1424 // escape analysis can go from the MemBarStoreStoreNode to the 1425 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1426 // based on the escape status of the AllocateNode. 1427 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1428 } 1429 if (compress) { 1430 set_result(_gvn.transform(count)); 1431 } 1432 return true; 1433 } 1434 1435 #ifdef _LP64 1436 #define XTOP ,top() /*additional argument*/ 1437 #else //_LP64 1438 #define XTOP /*no additional argument*/ 1439 #endif //_LP64 1440 1441 //------------------------inline_string_toBytesU-------------------------- 1442 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1443 bool LibraryCallKit::inline_string_toBytesU() { 1444 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1445 return false; 1446 } 1447 // Get the arguments. 1448 Node* value = argument(0); 1449 Node* offset = argument(1); 1450 Node* length = argument(2); 1451 1452 Node* newcopy = NULL; 1453 1454 // Set the original stack and the reexecute bit for the interpreter to reexecute 1455 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1456 { PreserveReexecuteState preexecs(this); 1457 jvms()->set_should_reexecute(true); 1458 1459 // Check if a null path was taken unconditionally. 1460 value = null_check(value); 1461 1462 RegionNode* bailout = new RegionNode(1); 1463 record_for_igvn(bailout); 1464 1465 // Range checks 1466 generate_negative_guard(offset, bailout); 1467 generate_negative_guard(length, bailout); 1468 generate_limit_guard(offset, length, load_array_length(value), bailout); 1469 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1470 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1471 1472 if (bailout->req() > 1) { 1473 PreserveJVMState pjvms(this); 1474 set_control(_gvn.transform(bailout)); 1475 uncommon_trap(Deoptimization::Reason_intrinsic, 1476 Deoptimization::Action_maybe_recompile); 1477 } 1478 if (stopped()) { 1479 return true; 1480 } 1481 1482 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1483 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1484 newcopy = new_array(klass_node, size, 0); // no arguments to push 1485 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1486 1487 // Calculate starting addresses. 1488 Node* src_start = array_element_address(value, offset, T_CHAR); 1489 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1490 1491 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1492 const TypeInt* toffset = gvn().type(offset)->is_int(); 1493 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1494 1495 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1496 const char* copyfunc_name = "arraycopy"; 1497 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1498 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1499 OptoRuntime::fast_arraycopy_Type(), 1500 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1501 src_start, dst_start, ConvI2X(length) XTOP); 1502 // Do not let reads from the cloned object float above the arraycopy. 1503 if (alloc != NULL) { 1504 if (alloc->maybe_set_complete(&_gvn)) { 1505 // "You break it, you buy it." 1506 InitializeNode* init = alloc->initialization(); 1507 assert(init->is_complete(), "we just did this"); 1508 init->set_complete_with_arraycopy(); 1509 assert(newcopy->is_CheckCastPP(), "sanity"); 1510 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1511 } 1512 // Do not let stores that initialize this object be reordered with 1513 // a subsequent store that would make this object accessible by 1514 // other threads. 1515 // Record what AllocateNode this StoreStore protects so that 1516 // escape analysis can go from the MemBarStoreStoreNode to the 1517 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1518 // based on the escape status of the AllocateNode. 1519 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1520 } else { 1521 insert_mem_bar(Op_MemBarCPUOrder); 1522 } 1523 } // original reexecute is set back here 1524 1525 C->set_has_split_ifs(true); // Has chance for split-if optimization 1526 if (!stopped()) { 1527 set_result(newcopy); 1528 } 1529 return true; 1530 } 1531 1532 //------------------------inline_string_getCharsU-------------------------- 1533 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1534 bool LibraryCallKit::inline_string_getCharsU() { 1535 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1536 return false; 1537 } 1538 1539 // Get the arguments. 1540 Node* src = argument(0); 1541 Node* src_begin = argument(1); 1542 Node* src_end = argument(2); // exclusive offset (i < src_end) 1543 Node* dst = argument(3); 1544 Node* dst_begin = argument(4); 1545 1546 // Check for allocation before we add nodes that would confuse 1547 // tightly_coupled_allocation() 1548 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1549 1550 // Check if a null path was taken unconditionally. 1551 src = null_check(src); 1552 dst = null_check(dst); 1553 if (stopped()) { 1554 return true; 1555 } 1556 1557 // Get length and convert char[] offset to byte[] offset 1558 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1559 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1560 1561 // Range checks 1562 generate_string_range_check(src, src_begin, length, true); 1563 generate_string_range_check(dst, dst_begin, length, false); 1564 if (stopped()) { 1565 return true; 1566 } 1567 1568 if (!stopped()) { 1569 // Calculate starting addresses. 1570 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1571 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1572 1573 // Check if array addresses are aligned to HeapWordSize 1574 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1575 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1576 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1577 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1578 1579 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1580 const char* copyfunc_name = "arraycopy"; 1581 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1582 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1583 OptoRuntime::fast_arraycopy_Type(), 1584 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1585 src_start, dst_start, ConvI2X(length) XTOP); 1586 // Do not let reads from the cloned object float above the arraycopy. 1587 if (alloc != NULL) { 1588 if (alloc->maybe_set_complete(&_gvn)) { 1589 // "You break it, you buy it." 1590 InitializeNode* init = alloc->initialization(); 1591 assert(init->is_complete(), "we just did this"); 1592 init->set_complete_with_arraycopy(); 1593 assert(dst->is_CheckCastPP(), "sanity"); 1594 assert(dst->in(0)->in(0) == init, "dest pinned"); 1595 } 1596 // Do not let stores that initialize this object be reordered with 1597 // a subsequent store that would make this object accessible by 1598 // other threads. 1599 // Record what AllocateNode this StoreStore protects so that 1600 // escape analysis can go from the MemBarStoreStoreNode to the 1601 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1602 // based on the escape status of the AllocateNode. 1603 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1604 } else { 1605 insert_mem_bar(Op_MemBarCPUOrder); 1606 } 1607 } 1608 1609 C->set_has_split_ifs(true); // Has chance for split-if optimization 1610 return true; 1611 } 1612 1613 //----------------------inline_string_char_access---------------------------- 1614 // Store/Load char to/from byte[] array. 1615 // static void StringUTF16.putChar(byte[] val, int index, int c) 1616 // static char StringUTF16.getChar(byte[] val, int index) 1617 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1618 Node* value = argument(0); 1619 Node* index = argument(1); 1620 Node* ch = is_store ? argument(2) : NULL; 1621 1622 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1623 // correctly requires matched array shapes. 1624 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1625 "sanity: byte[] and char[] bases agree"); 1626 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1627 "sanity: byte[] and char[] scales agree"); 1628 1629 // Bail when getChar over constants is requested: constant folding would 1630 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1631 // Java method would constant fold nicely instead. 1632 if (!is_store && value->is_Con() && index->is_Con()) { 1633 return false; 1634 } 1635 1636 Node* adr = array_element_address(value, index, T_CHAR); 1637 if (is_store) { 1638 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1639 false, false, true /* mismatched */); 1640 } else { 1641 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1642 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */); 1643 set_result(ch); 1644 } 1645 return true; 1646 } 1647 1648 //--------------------------round_double_node-------------------------------- 1649 // Round a double node if necessary. 1650 Node* LibraryCallKit::round_double_node(Node* n) { 1651 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1652 n = _gvn.transform(new RoundDoubleNode(0, n)); 1653 return n; 1654 } 1655 1656 //------------------------------inline_math----------------------------------- 1657 // public static double Math.abs(double) 1658 // public static double Math.sqrt(double) 1659 // public static double Math.log(double) 1660 // public static double Math.log10(double) 1661 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1662 Node* arg = round_double_node(argument(0)); 1663 Node* n = NULL; 1664 switch (id) { 1665 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1666 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1667 default: fatal_unexpected_iid(id); break; 1668 } 1669 set_result(_gvn.transform(n)); 1670 return true; 1671 } 1672 1673 //------------------------------inline_trig---------------------------------- 1674 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1675 // argument reduction which will turn into a fast/slow diamond. 1676 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1677 Node* arg = round_double_node(argument(0)); 1678 Node* n = NULL; 1679 1680 n = _gvn.transform(n); 1681 1682 // Rounding required? Check for argument reduction! 1683 if (Matcher::strict_fp_requires_explicit_rounding) { 1684 static const double pi_4 = 0.7853981633974483; 1685 static const double neg_pi_4 = -0.7853981633974483; 1686 // pi/2 in 80-bit extended precision 1687 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1688 // -pi/2 in 80-bit extended precision 1689 // 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}; 1690 // Cutoff value for using this argument reduction technique 1691 //static const double pi_2_minus_epsilon = 1.564660403643354; 1692 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1693 1694 // Pseudocode for sin: 1695 // if (x <= Math.PI / 4.0) { 1696 // if (x >= -Math.PI / 4.0) return fsin(x); 1697 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1698 // } else { 1699 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1700 // } 1701 // return StrictMath.sin(x); 1702 1703 // Pseudocode for cos: 1704 // if (x <= Math.PI / 4.0) { 1705 // if (x >= -Math.PI / 4.0) return fcos(x); 1706 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1707 // } else { 1708 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1709 // } 1710 // return StrictMath.cos(x); 1711 1712 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1713 // requires a special machine instruction to load it. Instead we'll try 1714 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1715 // probably do the math inside the SIN encoding. 1716 1717 // Make the merge point 1718 RegionNode* r = new RegionNode(3); 1719 Node* phi = new PhiNode(r, Type::DOUBLE); 1720 1721 // Flatten arg so we need only 1 test 1722 Node *abs = _gvn.transform(new AbsDNode(arg)); 1723 // Node for PI/4 constant 1724 Node *pi4 = makecon(TypeD::make(pi_4)); 1725 // Check PI/4 : abs(arg) 1726 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1727 // Check: If PI/4 < abs(arg) then go slow 1728 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1729 // Branch either way 1730 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1731 set_control(opt_iff(r,iff)); 1732 1733 // Set fast path result 1734 phi->init_req(2, n); 1735 1736 // Slow path - non-blocking leaf call 1737 Node* call = NULL; 1738 switch (id) { 1739 case vmIntrinsics::_dtan: 1740 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1741 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1742 "Tan", NULL, arg, top()); 1743 break; 1744 } 1745 assert(control()->in(0) == call, ""); 1746 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1747 r->init_req(1, control()); 1748 phi->init_req(1, slow_result); 1749 1750 // Post-merge 1751 set_control(_gvn.transform(r)); 1752 record_for_igvn(r); 1753 n = _gvn.transform(phi); 1754 1755 C->set_has_split_ifs(true); // Has chance for split-if optimization 1756 } 1757 set_result(n); 1758 return true; 1759 } 1760 1761 //------------------------------runtime_math----------------------------- 1762 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1763 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1764 "must be (DD)D or (D)D type"); 1765 1766 // Inputs 1767 Node* a = round_double_node(argument(0)); 1768 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1769 1770 const TypePtr* no_memory_effects = NULL; 1771 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1772 no_memory_effects, 1773 a, top(), b, b ? top() : NULL); 1774 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1775 #ifdef ASSERT 1776 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1777 assert(value_top == top(), "second value must be top"); 1778 #endif 1779 1780 set_result(value); 1781 return true; 1782 } 1783 1784 //------------------------------inline_math_native----------------------------- 1785 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1786 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1787 switch (id) { 1788 // These intrinsics are not properly supported on all hardware 1789 case vmIntrinsics::_dsin: 1790 return StubRoutines::dsin() != NULL ? 1791 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1792 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1793 case vmIntrinsics::_dcos: 1794 return StubRoutines::dcos() != NULL ? 1795 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1796 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1797 case vmIntrinsics::_dtan: 1798 return StubRoutines::dtan() != NULL ? 1799 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1800 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1801 case vmIntrinsics::_dlog: 1802 return StubRoutines::dlog() != NULL ? 1803 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1804 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1805 case vmIntrinsics::_dlog10: 1806 return StubRoutines::dlog10() != NULL ? 1807 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1808 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1809 1810 // These intrinsics are supported on all hardware 1811 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1812 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1813 1814 case vmIntrinsics::_dexp: 1815 return StubRoutines::dexp() != NULL ? 1816 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1817 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1818 case vmIntrinsics::_dpow: 1819 return StubRoutines::dpow() != NULL ? 1820 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1821 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1822 #undef FN_PTR 1823 1824 // These intrinsics are not yet correctly implemented 1825 case vmIntrinsics::_datan2: 1826 return false; 1827 1828 default: 1829 fatal_unexpected_iid(id); 1830 return false; 1831 } 1832 } 1833 1834 static bool is_simple_name(Node* n) { 1835 return (n->req() == 1 // constant 1836 || (n->is_Type() && n->as_Type()->type()->singleton()) 1837 || n->is_Proj() // parameter or return value 1838 || n->is_Phi() // local of some sort 1839 ); 1840 } 1841 1842 //----------------------------inline_notify-----------------------------------* 1843 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1844 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1845 address func; 1846 if (id == vmIntrinsics::_notify) { 1847 func = OptoRuntime::monitor_notify_Java(); 1848 } else { 1849 func = OptoRuntime::monitor_notifyAll_Java(); 1850 } 1851 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1852 make_slow_call_ex(call, env()->Throwable_klass(), false); 1853 return true; 1854 } 1855 1856 1857 //----------------------------inline_min_max----------------------------------- 1858 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1859 set_result(generate_min_max(id, argument(0), argument(1))); 1860 return true; 1861 } 1862 1863 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1864 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1865 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1866 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1867 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1868 1869 { 1870 PreserveJVMState pjvms(this); 1871 PreserveReexecuteState preexecs(this); 1872 jvms()->set_should_reexecute(true); 1873 1874 set_control(slow_path); 1875 set_i_o(i_o()); 1876 1877 uncommon_trap(Deoptimization::Reason_intrinsic, 1878 Deoptimization::Action_none); 1879 } 1880 1881 set_control(fast_path); 1882 set_result(math); 1883 } 1884 1885 template <typename OverflowOp> 1886 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1887 typedef typename OverflowOp::MathOp MathOp; 1888 1889 MathOp* mathOp = new MathOp(arg1, arg2); 1890 Node* operation = _gvn.transform( mathOp ); 1891 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1892 inline_math_mathExact(operation, ofcheck); 1893 return true; 1894 } 1895 1896 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1897 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1898 } 1899 1900 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1901 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1902 } 1903 1904 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1905 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1906 } 1907 1908 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1909 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1910 } 1911 1912 bool LibraryCallKit::inline_math_negateExactI() { 1913 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1914 } 1915 1916 bool LibraryCallKit::inline_math_negateExactL() { 1917 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 1918 } 1919 1920 bool LibraryCallKit::inline_math_multiplyExactI() { 1921 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 1922 } 1923 1924 bool LibraryCallKit::inline_math_multiplyExactL() { 1925 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 1926 } 1927 1928 Node* 1929 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1930 // These are the candidate return value: 1931 Node* xvalue = x0; 1932 Node* yvalue = y0; 1933 1934 if (xvalue == yvalue) { 1935 return xvalue; 1936 } 1937 1938 bool want_max = (id == vmIntrinsics::_max); 1939 1940 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1941 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1942 if (txvalue == NULL || tyvalue == NULL) return top(); 1943 // This is not really necessary, but it is consistent with a 1944 // hypothetical MaxINode::Value method: 1945 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1946 1947 // %%% This folding logic should (ideally) be in a different place. 1948 // Some should be inside IfNode, and there to be a more reliable 1949 // transformation of ?: style patterns into cmoves. We also want 1950 // more powerful optimizations around cmove and min/max. 1951 1952 // Try to find a dominating comparison of these guys. 1953 // It can simplify the index computation for Arrays.copyOf 1954 // and similar uses of System.arraycopy. 1955 // First, compute the normalized version of CmpI(x, y). 1956 int cmp_op = Op_CmpI; 1957 Node* xkey = xvalue; 1958 Node* ykey = yvalue; 1959 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 1960 if (ideal_cmpxy->is_Cmp()) { 1961 // E.g., if we have CmpI(length - offset, count), 1962 // it might idealize to CmpI(length, count + offset) 1963 cmp_op = ideal_cmpxy->Opcode(); 1964 xkey = ideal_cmpxy->in(1); 1965 ykey = ideal_cmpxy->in(2); 1966 } 1967 1968 // Start by locating any relevant comparisons. 1969 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1970 Node* cmpxy = NULL; 1971 Node* cmpyx = NULL; 1972 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1973 Node* cmp = start_from->fast_out(k); 1974 if (cmp->outcnt() > 0 && // must have prior uses 1975 cmp->in(0) == NULL && // must be context-independent 1976 cmp->Opcode() == cmp_op) { // right kind of compare 1977 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1978 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1979 } 1980 } 1981 1982 const int NCMPS = 2; 1983 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1984 int cmpn; 1985 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1986 if (cmps[cmpn] != NULL) break; // find a result 1987 } 1988 if (cmpn < NCMPS) { 1989 // Look for a dominating test that tells us the min and max. 1990 int depth = 0; // Limit search depth for speed 1991 Node* dom = control(); 1992 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1993 if (++depth >= 100) break; 1994 Node* ifproj = dom; 1995 if (!ifproj->is_Proj()) continue; 1996 Node* iff = ifproj->in(0); 1997 if (!iff->is_If()) continue; 1998 Node* bol = iff->in(1); 1999 if (!bol->is_Bool()) continue; 2000 Node* cmp = bol->in(1); 2001 if (cmp == NULL) continue; 2002 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2003 if (cmps[cmpn] == cmp) break; 2004 if (cmpn == NCMPS) continue; 2005 BoolTest::mask btest = bol->as_Bool()->_test._test; 2006 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2007 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2008 // At this point, we know that 'x btest y' is true. 2009 switch (btest) { 2010 case BoolTest::eq: 2011 // They are proven equal, so we can collapse the min/max. 2012 // Either value is the answer. Choose the simpler. 2013 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2014 return yvalue; 2015 return xvalue; 2016 case BoolTest::lt: // x < y 2017 case BoolTest::le: // x <= y 2018 return (want_max ? yvalue : xvalue); 2019 case BoolTest::gt: // x > y 2020 case BoolTest::ge: // x >= y 2021 return (want_max ? xvalue : yvalue); 2022 } 2023 } 2024 } 2025 2026 // We failed to find a dominating test. 2027 // Let's pick a test that might GVN with prior tests. 2028 Node* best_bol = NULL; 2029 BoolTest::mask best_btest = BoolTest::illegal; 2030 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2031 Node* cmp = cmps[cmpn]; 2032 if (cmp == NULL) continue; 2033 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2034 Node* bol = cmp->fast_out(j); 2035 if (!bol->is_Bool()) continue; 2036 BoolTest::mask btest = bol->as_Bool()->_test._test; 2037 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2038 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2039 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2040 best_bol = bol->as_Bool(); 2041 best_btest = btest; 2042 } 2043 } 2044 } 2045 2046 Node* answer_if_true = NULL; 2047 Node* answer_if_false = NULL; 2048 switch (best_btest) { 2049 default: 2050 if (cmpxy == NULL) 2051 cmpxy = ideal_cmpxy; 2052 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2053 // and fall through: 2054 case BoolTest::lt: // x < y 2055 case BoolTest::le: // x <= y 2056 answer_if_true = (want_max ? yvalue : xvalue); 2057 answer_if_false = (want_max ? xvalue : yvalue); 2058 break; 2059 case BoolTest::gt: // x > y 2060 case BoolTest::ge: // x >= y 2061 answer_if_true = (want_max ? xvalue : yvalue); 2062 answer_if_false = (want_max ? yvalue : xvalue); 2063 break; 2064 } 2065 2066 jint hi, lo; 2067 if (want_max) { 2068 // We can sharpen the minimum. 2069 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2070 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2071 } else { 2072 // We can sharpen the maximum. 2073 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2074 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2075 } 2076 2077 // Use a flow-free graph structure, to avoid creating excess control edges 2078 // which could hinder other optimizations. 2079 // Since Math.min/max is often used with arraycopy, we want 2080 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2081 Node* cmov = CMoveNode::make(NULL, best_bol, 2082 answer_if_false, answer_if_true, 2083 TypeInt::make(lo, hi, widen)); 2084 2085 return _gvn.transform(cmov); 2086 2087 /* 2088 // This is not as desirable as it may seem, since Min and Max 2089 // nodes do not have a full set of optimizations. 2090 // And they would interfere, anyway, with 'if' optimizations 2091 // and with CMoveI canonical forms. 2092 switch (id) { 2093 case vmIntrinsics::_min: 2094 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2095 case vmIntrinsics::_max: 2096 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2097 default: 2098 ShouldNotReachHere(); 2099 } 2100 */ 2101 } 2102 2103 inline int 2104 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2105 const TypePtr* base_type = TypePtr::NULL_PTR; 2106 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2107 if (base_type == NULL) { 2108 // Unknown type. 2109 return Type::AnyPtr; 2110 } else if (base_type == TypePtr::NULL_PTR) { 2111 // Since this is a NULL+long form, we have to switch to a rawptr. 2112 base = _gvn.transform(new CastX2PNode(offset)); 2113 offset = MakeConX(0); 2114 return Type::RawPtr; 2115 } else if (base_type->base() == Type::RawPtr) { 2116 return Type::RawPtr; 2117 } else if (base_type->isa_oopptr()) { 2118 // Base is never null => always a heap address. 2119 if (base_type->ptr() == TypePtr::NotNull) { 2120 return Type::OopPtr; 2121 } 2122 // Offset is small => always a heap address. 2123 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2124 if (offset_type != NULL && 2125 base_type->offset() == 0 && // (should always be?) 2126 offset_type->_lo >= 0 && 2127 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2128 return Type::OopPtr; 2129 } 2130 // Otherwise, it might either be oop+off or NULL+addr. 2131 return Type::AnyPtr; 2132 } else { 2133 // No information: 2134 return Type::AnyPtr; 2135 } 2136 } 2137 2138 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2139 int kind = classify_unsafe_addr(base, offset); 2140 if (kind == Type::RawPtr) { 2141 return basic_plus_adr(top(), base, offset); 2142 } else { 2143 return basic_plus_adr(base, offset); 2144 } 2145 } 2146 2147 //--------------------------inline_number_methods----------------------------- 2148 // inline int Integer.numberOfLeadingZeros(int) 2149 // inline int Long.numberOfLeadingZeros(long) 2150 // 2151 // inline int Integer.numberOfTrailingZeros(int) 2152 // inline int Long.numberOfTrailingZeros(long) 2153 // 2154 // inline int Integer.bitCount(int) 2155 // inline int Long.bitCount(long) 2156 // 2157 // inline char Character.reverseBytes(char) 2158 // inline short Short.reverseBytes(short) 2159 // inline int Integer.reverseBytes(int) 2160 // inline long Long.reverseBytes(long) 2161 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2162 Node* arg = argument(0); 2163 Node* n = NULL; 2164 switch (id) { 2165 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2166 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2167 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2168 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2169 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2170 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2171 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2172 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2173 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2174 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2175 default: fatal_unexpected_iid(id); break; 2176 } 2177 set_result(_gvn.transform(n)); 2178 return true; 2179 } 2180 2181 //----------------------------inline_unsafe_access---------------------------- 2182 2183 // Helper that guards and inserts a pre-barrier. 2184 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2185 Node* pre_val, bool need_mem_bar) { 2186 // We could be accessing the referent field of a reference object. If so, when G1 2187 // is enabled, we need to log the value in the referent field in an SATB buffer. 2188 // This routine performs some compile time filters and generates suitable 2189 // runtime filters that guard the pre-barrier code. 2190 // Also add memory barrier for non volatile load from the referent field 2191 // to prevent commoning of loads across safepoint. 2192 if (!UseG1GC && !need_mem_bar) 2193 return; 2194 2195 // Some compile time checks. 2196 2197 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2198 const TypeX* otype = offset->find_intptr_t_type(); 2199 if (otype != NULL && otype->is_con() && 2200 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2201 // Constant offset but not the reference_offset so just return 2202 return; 2203 } 2204 2205 // We only need to generate the runtime guards for instances. 2206 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2207 if (btype != NULL) { 2208 if (btype->isa_aryptr()) { 2209 // Array type so nothing to do 2210 return; 2211 } 2212 2213 const TypeInstPtr* itype = btype->isa_instptr(); 2214 if (itype != NULL) { 2215 // Can the klass of base_oop be statically determined to be 2216 // _not_ a sub-class of Reference and _not_ Object? 2217 ciKlass* klass = itype->klass(); 2218 if ( klass->is_loaded() && 2219 !klass->is_subtype_of(env()->Reference_klass()) && 2220 !env()->Object_klass()->is_subtype_of(klass)) { 2221 return; 2222 } 2223 } 2224 } 2225 2226 // The compile time filters did not reject base_oop/offset so 2227 // we need to generate the following runtime filters 2228 // 2229 // if (offset == java_lang_ref_Reference::_reference_offset) { 2230 // if (instance_of(base, java.lang.ref.Reference)) { 2231 // pre_barrier(_, pre_val, ...); 2232 // } 2233 // } 2234 2235 float likely = PROB_LIKELY( 0.999); 2236 float unlikely = PROB_UNLIKELY(0.999); 2237 2238 IdealKit ideal(this); 2239 #define __ ideal. 2240 2241 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2242 2243 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2244 // Update graphKit memory and control from IdealKit. 2245 sync_kit(ideal); 2246 2247 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2248 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2249 2250 // Update IdealKit memory and control from graphKit. 2251 __ sync_kit(this); 2252 2253 Node* one = __ ConI(1); 2254 // is_instof == 0 if base_oop == NULL 2255 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2256 2257 // Update graphKit from IdeakKit. 2258 sync_kit(ideal); 2259 2260 // Use the pre-barrier to record the value in the referent field 2261 pre_barrier(false /* do_load */, 2262 __ ctrl(), 2263 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2264 pre_val /* pre_val */, 2265 T_OBJECT); 2266 if (need_mem_bar) { 2267 // Add memory barrier to prevent commoning reads from this field 2268 // across safepoint since GC can change its value. 2269 insert_mem_bar(Op_MemBarCPUOrder); 2270 } 2271 // Update IdealKit from graphKit. 2272 __ sync_kit(this); 2273 2274 } __ end_if(); // _ref_type != ref_none 2275 } __ end_if(); // offset == referent_offset 2276 2277 // Final sync IdealKit and GraphKit. 2278 final_sync(ideal); 2279 #undef __ 2280 } 2281 2282 2283 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2284 // Attempt to infer a sharper value type from the offset and base type. 2285 ciKlass* sharpened_klass = NULL; 2286 2287 // See if it is an instance field, with an object type. 2288 if (alias_type->field() != NULL) { 2289 if (alias_type->field()->type()->is_klass()) { 2290 sharpened_klass = alias_type->field()->type()->as_klass(); 2291 } 2292 } 2293 2294 // See if it is a narrow oop array. 2295 if (adr_type->isa_aryptr()) { 2296 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2297 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2298 if (elem_type != NULL) { 2299 sharpened_klass = elem_type->klass(); 2300 } 2301 } 2302 } 2303 2304 // The sharpened class might be unloaded if there is no class loader 2305 // contraint in place. 2306 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2307 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2308 2309 #ifndef PRODUCT 2310 if (C->print_intrinsics() || C->print_inlining()) { 2311 tty->print(" from base type: "); adr_type->dump(); 2312 tty->print(" sharpened value: "); tjp->dump(); 2313 } 2314 #endif 2315 // Sharpen the value type. 2316 return tjp; 2317 } 2318 return NULL; 2319 } 2320 2321 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2322 if (callee()->is_static()) return false; // caller must have the capability! 2323 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2324 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2325 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2326 2327 #ifndef PRODUCT 2328 { 2329 ResourceMark rm; 2330 // Check the signatures. 2331 ciSignature* sig = callee()->signature(); 2332 #ifdef ASSERT 2333 if (!is_store) { 2334 // Object getObject(Object base, int/long offset), etc. 2335 BasicType rtype = sig->return_type()->basic_type(); 2336 assert(rtype == type, "getter must return the expected value"); 2337 assert(sig->count() == 2, "oop getter has 2 arguments"); 2338 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2339 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2340 } else { 2341 // void putObject(Object base, int/long offset, Object x), etc. 2342 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2343 assert(sig->count() == 3, "oop putter has 3 arguments"); 2344 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2345 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2346 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2347 assert(vtype == type, "putter must accept the expected value"); 2348 } 2349 #endif // ASSERT 2350 } 2351 #endif //PRODUCT 2352 2353 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2354 2355 Node* receiver = argument(0); // type: oop 2356 2357 // Build address expression. 2358 Node* adr; 2359 Node* heap_base_oop = top(); 2360 Node* offset = top(); 2361 Node* val; 2362 2363 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2364 Node* base = argument(1); // type: oop 2365 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2366 offset = argument(2); // type: long 2367 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2368 // to be plain byte offsets, which are also the same as those accepted 2369 // by oopDesc::field_base. 2370 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2371 "fieldOffset must be byte-scaled"); 2372 // 32-bit machines ignore the high half! 2373 offset = ConvL2X(offset); 2374 adr = make_unsafe_address(base, offset); 2375 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) { 2376 heap_base_oop = base; 2377 } 2378 val = is_store ? argument(4) : NULL; 2379 2380 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2381 2382 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2383 // there was not enough information to nail it down. 2384 Compile::AliasType* alias_type = C->alias_type(adr_type); 2385 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2386 2387 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM || 2388 alias_type->basic_type() != T_ILLEGAL, "field, array element or unknown"); 2389 bool mismatched = false; 2390 BasicType bt = alias_type->basic_type(); 2391 if (bt != T_ILLEGAL) { 2392 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2393 // Alias type doesn't differentiate between byte[] and boolean[]). 2394 // Use address type to get the element type. 2395 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2396 } 2397 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2398 // accessing an array field with getObject is not a mismatch 2399 bt = T_OBJECT; 2400 } 2401 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2402 // Don't intrinsify mismatched object accesses 2403 return false; 2404 } 2405 mismatched = (bt != type); 2406 } 2407 2408 // First guess at the value type. 2409 const Type *value_type = Type::get_const_basic_type(type); 2410 2411 // We will need memory barriers unless we can determine a unique 2412 // alias category for this reference. (Note: If for some reason 2413 // the barriers get omitted and the unsafe reference begins to "pollute" 2414 // the alias analysis of the rest of the graph, either Compile::can_alias 2415 // or Compile::must_alias will throw a diagnostic assert.) 2416 bool need_mem_bar; 2417 switch (kind) { 2418 case Relaxed: 2419 need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2420 break; 2421 case Opaque: 2422 // Opaque uses CPUOrder membars for protection against code movement. 2423 case Acquire: 2424 case Release: 2425 case Volatile: 2426 need_mem_bar = true; 2427 break; 2428 default: 2429 ShouldNotReachHere(); 2430 } 2431 2432 // Some accesses require access atomicity for all types, notably longs and doubles. 2433 // When AlwaysAtomicAccesses is enabled, all accesses are atomic. 2434 bool requires_atomic_access = false; 2435 switch (kind) { 2436 case Relaxed: 2437 requires_atomic_access = AlwaysAtomicAccesses; 2438 break; 2439 case Opaque: 2440 // Opaque accesses are atomic. 2441 case Acquire: 2442 case Release: 2443 case Volatile: 2444 requires_atomic_access = true; 2445 break; 2446 default: 2447 ShouldNotReachHere(); 2448 } 2449 2450 // Figure out the memory ordering. 2451 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd 2452 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store); 2453 2454 // If we are reading the value of the referent field of a Reference 2455 // object (either by using Unsafe directly or through reflection) 2456 // then, if G1 is enabled, we need to record the referent in an 2457 // SATB log buffer using the pre-barrier mechanism. 2458 // Also we need to add memory barrier to prevent commoning reads 2459 // from this field across safepoint since GC can change its value. 2460 bool need_read_barrier = !is_store && 2461 offset != top() && heap_base_oop != top(); 2462 2463 if (!is_store && type == T_OBJECT) { 2464 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2465 if (tjp != NULL) { 2466 value_type = tjp; 2467 } 2468 } 2469 2470 receiver = null_check(receiver); 2471 if (stopped()) { 2472 return true; 2473 } 2474 // Heap pointers get a null-check from the interpreter, 2475 // as a courtesy. However, this is not guaranteed by Unsafe, 2476 // and it is not possible to fully distinguish unintended nulls 2477 // from intended ones in this API. 2478 2479 // We need to emit leading and trailing CPU membars (see below) in 2480 // addition to memory membars for special access modes. This is a little 2481 // too strong, but avoids the need to insert per-alias-type 2482 // volatile membars (for stores; compare Parse::do_put_xxx), which 2483 // we cannot do effectively here because we probably only have a 2484 // rough approximation of type. 2485 2486 switch(kind) { 2487 case Relaxed: 2488 case Opaque: 2489 case Acquire: 2490 break; 2491 case Release: 2492 case Volatile: 2493 if (is_store) { 2494 insert_mem_bar(Op_MemBarRelease); 2495 } else { 2496 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2497 insert_mem_bar(Op_MemBarVolatile); 2498 } 2499 } 2500 break; 2501 default: 2502 ShouldNotReachHere(); 2503 } 2504 2505 // Memory barrier to prevent normal and 'unsafe' accesses from 2506 // bypassing each other. Happens after null checks, so the 2507 // exception paths do not take memory state from the memory barrier, 2508 // so there's no problems making a strong assert about mixing users 2509 // of safe & unsafe memory. 2510 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2511 2512 if (!is_store) { 2513 Node* p = NULL; 2514 // Try to constant fold a load from a constant field 2515 ciField* field = alias_type->field(); 2516 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { 2517 // final or stable field 2518 p = make_constant_from_field(field, heap_base_oop); 2519 } 2520 if (p == NULL) { 2521 // To be valid, unsafe loads may depend on other conditions than 2522 // the one that guards them: pin the Load node 2523 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, requires_atomic_access, unaligned, mismatched); 2524 // load value 2525 switch (type) { 2526 case T_BOOLEAN: 2527 case T_CHAR: 2528 case T_BYTE: 2529 case T_SHORT: 2530 case T_INT: 2531 case T_LONG: 2532 case T_FLOAT: 2533 case T_DOUBLE: 2534 break; 2535 case T_OBJECT: 2536 if (need_read_barrier) { 2537 // We do not require a mem bar inside pre_barrier if need_mem_bar 2538 // is set: the barriers would be emitted by us. 2539 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar); 2540 } 2541 break; 2542 case T_ADDRESS: 2543 // Cast to an int type. 2544 p = _gvn.transform(new CastP2XNode(NULL, p)); 2545 p = ConvX2UL(p); 2546 break; 2547 default: 2548 fatal("unexpected type %d: %s", type, type2name(type)); 2549 break; 2550 } 2551 } 2552 // The load node has the control of the preceding MemBarCPUOrder. All 2553 // following nodes will have the control of the MemBarCPUOrder inserted at 2554 // the end of this method. So, pushing the load onto the stack at a later 2555 // point is fine. 2556 set_result(p); 2557 } else { 2558 // place effect of store into memory 2559 switch (type) { 2560 case T_DOUBLE: 2561 val = dstore_rounding(val); 2562 break; 2563 case T_ADDRESS: 2564 // Repackage the long as a pointer. 2565 val = ConvL2X(val); 2566 val = _gvn.transform(new CastX2PNode(val)); 2567 break; 2568 } 2569 2570 if (type != T_OBJECT) { 2571 (void) store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched); 2572 } else { 2573 // Possibly an oop being stored to Java heap or native memory 2574 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2575 // oop to Java heap. 2576 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2577 } else { 2578 // We can't tell at compile time if we are storing in the Java heap or outside 2579 // of it. So we need to emit code to conditionally do the proper type of 2580 // store. 2581 2582 IdealKit ideal(this); 2583 #define __ ideal. 2584 // QQQ who knows what probability is here?? 2585 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2586 // Sync IdealKit and graphKit. 2587 sync_kit(ideal); 2588 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2589 // Update IdealKit memory. 2590 __ sync_kit(this); 2591 } __ else_(); { 2592 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, requires_atomic_access, mismatched); 2593 } __ end_if(); 2594 // Final sync IdealKit and GraphKit. 2595 final_sync(ideal); 2596 #undef __ 2597 } 2598 } 2599 } 2600 2601 switch(kind) { 2602 case Relaxed: 2603 case Opaque: 2604 case Release: 2605 break; 2606 case Acquire: 2607 case Volatile: 2608 if (!is_store) { 2609 insert_mem_bar(Op_MemBarAcquire); 2610 } else { 2611 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2612 insert_mem_bar(Op_MemBarVolatile); 2613 } 2614 } 2615 break; 2616 default: 2617 ShouldNotReachHere(); 2618 } 2619 2620 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2621 2622 return true; 2623 } 2624 2625 //----------------------------inline_unsafe_load_store---------------------------- 2626 // This method serves a couple of different customers (depending on LoadStoreKind): 2627 // 2628 // LS_cmp_swap: 2629 // 2630 // boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2631 // boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2632 // boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2633 // 2634 // LS_cmp_swap_weak: 2635 // 2636 // boolean weakCompareAndSwapObject( Object o, long offset, Object expected, Object x); 2637 // boolean weakCompareAndSwapObjectAcquire(Object o, long offset, Object expected, Object x); 2638 // boolean weakCompareAndSwapObjectRelease(Object o, long offset, Object expected, Object x); 2639 // 2640 // boolean weakCompareAndSwapInt( Object o, long offset, int expected, int x); 2641 // boolean weakCompareAndSwapIntAcquire( Object o, long offset, int expected, int x); 2642 // boolean weakCompareAndSwapIntRelease( Object o, long offset, int expected, int x); 2643 // 2644 // boolean weakCompareAndSwapLong( Object o, long offset, long expected, long x); 2645 // boolean weakCompareAndSwapLongAcquire( Object o, long offset, long expected, long x); 2646 // boolean weakCompareAndSwapLongRelease( Object o, long offset, long expected, long x); 2647 // 2648 // LS_cmp_exchange: 2649 // 2650 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x); 2651 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x); 2652 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x); 2653 // 2654 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2655 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2656 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2657 // 2658 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2659 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2660 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2661 // 2662 // LS_get_add: 2663 // 2664 // int getAndAddInt( Object o, long offset, int delta) 2665 // long getAndAddLong(Object o, long offset, long delta) 2666 // 2667 // LS_get_set: 2668 // 2669 // int getAndSet(Object o, long offset, int newValue) 2670 // long getAndSet(Object o, long offset, long newValue) 2671 // Object getAndSet(Object o, long offset, Object newValue) 2672 // 2673 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2674 // This basic scheme here is the same as inline_unsafe_access, but 2675 // differs in enough details that combining them would make the code 2676 // overly confusing. (This is a true fact! I originally combined 2677 // them, but even I was confused by it!) As much code/comments as 2678 // possible are retained from inline_unsafe_access though to make 2679 // the correspondences clearer. - dl 2680 2681 if (callee()->is_static()) return false; // caller must have the capability! 2682 2683 #ifndef PRODUCT 2684 BasicType rtype; 2685 { 2686 ResourceMark rm; 2687 // Check the signatures. 2688 ciSignature* sig = callee()->signature(); 2689 rtype = sig->return_type()->basic_type(); 2690 switch(kind) { 2691 case LS_get_add: 2692 case LS_get_set: { 2693 // Check the signatures. 2694 #ifdef ASSERT 2695 assert(rtype == type, "get and set must return the expected type"); 2696 assert(sig->count() == 3, "get and set has 3 arguments"); 2697 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2698 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2699 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2700 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2701 #endif // ASSERT 2702 break; 2703 } 2704 case LS_cmp_swap: 2705 case LS_cmp_swap_weak: { 2706 // Check the signatures. 2707 #ifdef ASSERT 2708 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2709 assert(sig->count() == 4, "CAS has 4 arguments"); 2710 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2711 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2712 #endif // ASSERT 2713 break; 2714 } 2715 case LS_cmp_exchange: { 2716 // Check the signatures. 2717 #ifdef ASSERT 2718 assert(rtype == type, "CAS must return the expected type"); 2719 assert(sig->count() == 4, "CAS has 4 arguments"); 2720 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2721 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2722 #endif // ASSERT 2723 break; 2724 } 2725 default: 2726 ShouldNotReachHere(); 2727 } 2728 } 2729 #endif //PRODUCT 2730 2731 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2732 2733 // Get arguments: 2734 Node* receiver = NULL; 2735 Node* base = NULL; 2736 Node* offset = NULL; 2737 Node* oldval = NULL; 2738 Node* newval = NULL; 2739 switch(kind) { 2740 case LS_cmp_swap: 2741 case LS_cmp_swap_weak: 2742 case LS_cmp_exchange: { 2743 const bool two_slot_type = type2size[type] == 2; 2744 receiver = argument(0); // type: oop 2745 base = argument(1); // type: oop 2746 offset = argument(2); // type: long 2747 oldval = argument(4); // type: oop, int, or long 2748 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2749 break; 2750 } 2751 case LS_get_add: 2752 case LS_get_set: { 2753 receiver = argument(0); // type: oop 2754 base = argument(1); // type: oop 2755 offset = argument(2); // type: long 2756 oldval = NULL; 2757 newval = argument(4); // type: oop, int, or long 2758 break; 2759 } 2760 default: 2761 ShouldNotReachHere(); 2762 } 2763 2764 // Null check receiver. 2765 receiver = null_check(receiver); 2766 if (stopped()) { 2767 return true; 2768 } 2769 2770 // Build field offset expression. 2771 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2772 // to be plain byte offsets, which are also the same as those accepted 2773 // by oopDesc::field_base. 2774 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2775 // 32-bit machines ignore the high half of long offsets 2776 offset = ConvL2X(offset); 2777 Node* adr = make_unsafe_address(base, offset); 2778 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2779 2780 Compile::AliasType* alias_type = C->alias_type(adr_type); 2781 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM || 2782 alias_type->basic_type() != T_ILLEGAL, "field, array element or unknown"); 2783 BasicType bt = alias_type->basic_type(); 2784 if (bt != T_ILLEGAL && 2785 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2786 // Don't intrinsify mismatched object accesses. 2787 return false; 2788 } 2789 2790 // For CAS, unlike inline_unsafe_access, there seems no point in 2791 // trying to refine types. Just use the coarse types here. 2792 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2793 const Type *value_type = Type::get_const_basic_type(type); 2794 2795 switch (kind) { 2796 case LS_get_set: 2797 case LS_cmp_exchange: { 2798 if (type == T_OBJECT) { 2799 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2800 if (tjp != NULL) { 2801 value_type = tjp; 2802 } 2803 } 2804 break; 2805 } 2806 case LS_cmp_swap: 2807 case LS_cmp_swap_weak: 2808 case LS_get_add: 2809 break; 2810 default: 2811 ShouldNotReachHere(); 2812 } 2813 2814 int alias_idx = C->get_alias_index(adr_type); 2815 2816 // Memory-model-wise, a LoadStore acts like a little synchronized 2817 // block, so needs barriers on each side. These don't translate 2818 // into actual barriers on most machines, but we still need rest of 2819 // compiler to respect ordering. 2820 2821 switch (access_kind) { 2822 case Relaxed: 2823 case Acquire: 2824 break; 2825 case Release: 2826 insert_mem_bar(Op_MemBarRelease); 2827 break; 2828 case Volatile: 2829 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2830 insert_mem_bar(Op_MemBarVolatile); 2831 } else { 2832 insert_mem_bar(Op_MemBarRelease); 2833 } 2834 break; 2835 default: 2836 ShouldNotReachHere(); 2837 } 2838 insert_mem_bar(Op_MemBarCPUOrder); 2839 2840 // Figure out the memory ordering. 2841 MemNode::MemOrd mo = access_kind_to_memord(access_kind); 2842 2843 // 4984716: MemBars must be inserted before this 2844 // memory node in order to avoid a false 2845 // dependency which will confuse the scheduler. 2846 Node *mem = memory(alias_idx); 2847 2848 // For now, we handle only those cases that actually exist: ints, 2849 // longs, and Object. Adding others should be straightforward. 2850 Node* load_store = NULL; 2851 switch(type) { 2852 case T_INT: 2853 switch(kind) { 2854 case LS_get_add: 2855 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2856 break; 2857 case LS_get_set: 2858 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2859 break; 2860 case LS_cmp_swap_weak: 2861 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo)); 2862 break; 2863 case LS_cmp_swap: 2864 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo)); 2865 break; 2866 case LS_cmp_exchange: 2867 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo)); 2868 break; 2869 default: 2870 ShouldNotReachHere(); 2871 } 2872 break; 2873 case T_LONG: 2874 switch(kind) { 2875 case LS_get_add: 2876 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2877 break; 2878 case LS_get_set: 2879 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2880 break; 2881 case LS_cmp_swap_weak: 2882 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo)); 2883 break; 2884 case LS_cmp_swap: 2885 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo)); 2886 break; 2887 case LS_cmp_exchange: 2888 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo)); 2889 break; 2890 default: 2891 ShouldNotReachHere(); 2892 } 2893 break; 2894 case T_OBJECT: 2895 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2896 // could be delayed during Parse (for example, in adjust_map_after_if()). 2897 // Execute transformation here to avoid barrier generation in such case. 2898 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2899 newval = _gvn.makecon(TypePtr::NULL_PTR); 2900 2901 // Reference stores need a store barrier. 2902 switch(kind) { 2903 case LS_get_set: { 2904 // If pre-barrier must execute before the oop store, old value will require do_load here. 2905 if (!can_move_pre_barrier()) { 2906 pre_barrier(true /* do_load*/, 2907 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2908 NULL /* pre_val*/, 2909 T_OBJECT); 2910 } // Else move pre_barrier to use load_store value, see below. 2911 break; 2912 } 2913 case LS_cmp_swap_weak: 2914 case LS_cmp_swap: 2915 case LS_cmp_exchange: { 2916 // Same as for newval above: 2917 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2918 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2919 } 2920 // The only known value which might get overwritten is oldval. 2921 pre_barrier(false /* do_load */, 2922 control(), NULL, NULL, max_juint, NULL, NULL, 2923 oldval /* pre_val */, 2924 T_OBJECT); 2925 break; 2926 } 2927 default: 2928 ShouldNotReachHere(); 2929 } 2930 2931 #ifdef _LP64 2932 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2933 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2934 2935 switch(kind) { 2936 case LS_get_set: 2937 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop())); 2938 break; 2939 case LS_cmp_swap_weak: { 2940 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2941 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo)); 2942 break; 2943 } 2944 case LS_cmp_swap: { 2945 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2946 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo)); 2947 break; 2948 } 2949 case LS_cmp_exchange: { 2950 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2951 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 2952 break; 2953 } 2954 default: 2955 ShouldNotReachHere(); 2956 } 2957 } else 2958 #endif 2959 switch (kind) { 2960 case LS_get_set: 2961 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2962 break; 2963 case LS_cmp_swap_weak: 2964 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo)); 2965 break; 2966 case LS_cmp_swap: 2967 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo)); 2968 break; 2969 case LS_cmp_exchange: 2970 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo)); 2971 break; 2972 default: 2973 ShouldNotReachHere(); 2974 } 2975 2976 // Emit the post barrier only when the actual store happened. This makes sense 2977 // to check only for LS_cmp_* that can fail to set the value. 2978 // LS_cmp_exchange does not produce any branches by default, so there is no 2979 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with 2980 // CompareAndExchange and move branches here, it would make sense to conditionalize 2981 // post_barriers for LS_cmp_exchange as well. 2982 // 2983 // CAS success path is marked more likely since we anticipate this is a performance 2984 // critical path, while CAS failure path can use the penalty for going through unlikely 2985 // path as backoff. Which is still better than doing a store barrier there. 2986 switch (kind) { 2987 case LS_get_set: 2988 case LS_cmp_exchange: { 2989 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2990 break; 2991 } 2992 case LS_cmp_swap_weak: 2993 case LS_cmp_swap: { 2994 IdealKit ideal(this); 2995 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); { 2996 sync_kit(ideal); 2997 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2998 ideal.sync_kit(this); 2999 } ideal.end_if(); 3000 final_sync(ideal); 3001 break; 3002 } 3003 default: 3004 ShouldNotReachHere(); 3005 } 3006 break; 3007 default: 3008 fatal("unexpected type %d: %s", type, type2name(type)); 3009 break; 3010 } 3011 3012 // SCMemProjNodes represent the memory state of a LoadStore. Their 3013 // main role is to prevent LoadStore nodes from being optimized away 3014 // when their results aren't used. 3015 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 3016 set_memory(proj, alias_idx); 3017 3018 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) { 3019 #ifdef _LP64 3020 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3021 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 3022 } 3023 #endif 3024 if (can_move_pre_barrier()) { 3025 // Don't need to load pre_val. The old value is returned by load_store. 3026 // The pre_barrier can execute after the xchg as long as no safepoint 3027 // gets inserted between them. 3028 pre_barrier(false /* do_load */, 3029 control(), NULL, NULL, max_juint, NULL, NULL, 3030 load_store /* pre_val */, 3031 T_OBJECT); 3032 } 3033 } 3034 3035 // Add the trailing membar surrounding the access 3036 insert_mem_bar(Op_MemBarCPUOrder); 3037 3038 switch (access_kind) { 3039 case Relaxed: 3040 case Release: 3041 break; // do nothing 3042 case Acquire: 3043 case Volatile: 3044 insert_mem_bar(Op_MemBarAcquire); 3045 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code 3046 break; 3047 default: 3048 ShouldNotReachHere(); 3049 } 3050 3051 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3052 set_result(load_store); 3053 return true; 3054 } 3055 3056 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) { 3057 MemNode::MemOrd mo = MemNode::unset; 3058 switch(kind) { 3059 case Opaque: 3060 case Relaxed: mo = MemNode::unordered; break; 3061 case Acquire: mo = MemNode::acquire; break; 3062 case Release: mo = MemNode::release; break; 3063 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break; 3064 default: 3065 ShouldNotReachHere(); 3066 } 3067 guarantee(mo != MemNode::unset, "Should select memory ordering"); 3068 return mo; 3069 } 3070 3071 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) { 3072 MemNode::MemOrd mo = MemNode::unset; 3073 switch(kind) { 3074 case Opaque: 3075 case Relaxed: mo = MemNode::unordered; break; 3076 case Acquire: mo = MemNode::acquire; break; 3077 case Release: mo = MemNode::release; break; 3078 case Volatile: mo = MemNode::seqcst; break; 3079 default: 3080 ShouldNotReachHere(); 3081 } 3082 guarantee(mo != MemNode::unset, "Should select memory ordering"); 3083 return mo; 3084 } 3085 3086 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3087 // Regardless of form, don't allow previous ld/st to move down, 3088 // then issue acquire, release, or volatile mem_bar. 3089 insert_mem_bar(Op_MemBarCPUOrder); 3090 switch(id) { 3091 case vmIntrinsics::_loadFence: 3092 insert_mem_bar(Op_LoadFence); 3093 return true; 3094 case vmIntrinsics::_storeFence: 3095 insert_mem_bar(Op_StoreFence); 3096 return true; 3097 case vmIntrinsics::_fullFence: 3098 insert_mem_bar(Op_MemBarVolatile); 3099 return true; 3100 default: 3101 fatal_unexpected_iid(id); 3102 return false; 3103 } 3104 } 3105 3106 bool LibraryCallKit::inline_onspinwait() { 3107 insert_mem_bar(Op_OnSpinWait); 3108 return true; 3109 } 3110 3111 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3112 if (!kls->is_Con()) { 3113 return true; 3114 } 3115 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3116 if (klsptr == NULL) { 3117 return true; 3118 } 3119 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3120 // don't need a guard for a klass that is already initialized 3121 return !ik->is_initialized(); 3122 } 3123 3124 //----------------------------inline_unsafe_allocate--------------------------- 3125 // public native Object Unsafe.allocateInstance(Class<?> cls); 3126 bool LibraryCallKit::inline_unsafe_allocate() { 3127 if (callee()->is_static()) return false; // caller must have the capability! 3128 3129 null_check_receiver(); // null-check, then ignore 3130 Node* cls = null_check(argument(1)); 3131 if (stopped()) return true; 3132 3133 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3134 kls = null_check(kls); 3135 if (stopped()) return true; // argument was like int.class 3136 3137 Node* test = NULL; 3138 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3139 // Note: The argument might still be an illegal value like 3140 // Serializable.class or Object[].class. The runtime will handle it. 3141 // But we must make an explicit check for initialization. 3142 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3143 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3144 // can generate code to load it as unsigned byte. 3145 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3146 Node* bits = intcon(InstanceKlass::fully_initialized); 3147 test = _gvn.transform(new SubINode(inst, bits)); 3148 // The 'test' is non-zero if we need to take a slow path. 3149 } 3150 3151 Node* obj = new_instance(kls, test); 3152 set_result(obj); 3153 return true; 3154 } 3155 3156 //------------------------inline_native_time_funcs-------------- 3157 // inline code for System.currentTimeMillis() and System.nanoTime() 3158 // these have the same type and signature 3159 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3160 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3161 const TypePtr* no_memory_effects = NULL; 3162 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3163 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3164 #ifdef ASSERT 3165 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3166 assert(value_top == top(), "second value must be top"); 3167 #endif 3168 set_result(value); 3169 return true; 3170 } 3171 3172 //------------------------inline_native_currentThread------------------ 3173 bool LibraryCallKit::inline_native_currentThread() { 3174 Node* junk = NULL; 3175 set_result(generate_current_thread(junk)); 3176 return true; 3177 } 3178 3179 //------------------------inline_native_isInterrupted------------------ 3180 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3181 bool LibraryCallKit::inline_native_isInterrupted() { 3182 // Add a fast path to t.isInterrupted(clear_int): 3183 // (t == Thread.current() && 3184 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3185 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3186 // So, in the common case that the interrupt bit is false, 3187 // we avoid making a call into the VM. Even if the interrupt bit 3188 // is true, if the clear_int argument is false, we avoid the VM call. 3189 // However, if the receiver is not currentThread, we must call the VM, 3190 // because there must be some locking done around the operation. 3191 3192 // We only go to the fast case code if we pass two guards. 3193 // Paths which do not pass are accumulated in the slow_region. 3194 3195 enum { 3196 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3197 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3198 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3199 PATH_LIMIT 3200 }; 3201 3202 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3203 // out of the function. 3204 insert_mem_bar(Op_MemBarCPUOrder); 3205 3206 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3207 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3208 3209 RegionNode* slow_region = new RegionNode(1); 3210 record_for_igvn(slow_region); 3211 3212 // (a) Receiving thread must be the current thread. 3213 Node* rec_thr = argument(0); 3214 Node* tls_ptr = NULL; 3215 Node* cur_thr = generate_current_thread(tls_ptr); 3216 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3217 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3218 3219 generate_slow_guard(bol_thr, slow_region); 3220 3221 // (b) Interrupt bit on TLS must be false. 3222 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3223 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3224 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3225 3226 // Set the control input on the field _interrupted read to prevent it floating up. 3227 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3228 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3229 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3230 3231 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3232 3233 // First fast path: if (!TLS._interrupted) return false; 3234 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3235 result_rgn->init_req(no_int_result_path, false_bit); 3236 result_val->init_req(no_int_result_path, intcon(0)); 3237 3238 // drop through to next case 3239 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3240 3241 #ifndef TARGET_OS_FAMILY_windows 3242 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3243 Node* clr_arg = argument(1); 3244 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3245 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3246 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3247 3248 // Second fast path: ... else if (!clear_int) return true; 3249 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3250 result_rgn->init_req(no_clear_result_path, false_arg); 3251 result_val->init_req(no_clear_result_path, intcon(1)); 3252 3253 // drop through to next case 3254 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3255 #else 3256 // To return true on Windows you must read the _interrupted field 3257 // and check the event state i.e. take the slow path. 3258 #endif // TARGET_OS_FAMILY_windows 3259 3260 // (d) Otherwise, go to the slow path. 3261 slow_region->add_req(control()); 3262 set_control( _gvn.transform(slow_region)); 3263 3264 if (stopped()) { 3265 // There is no slow path. 3266 result_rgn->init_req(slow_result_path, top()); 3267 result_val->init_req(slow_result_path, top()); 3268 } else { 3269 // non-virtual because it is a private non-static 3270 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3271 3272 Node* slow_val = set_results_for_java_call(slow_call); 3273 // this->control() comes from set_results_for_java_call 3274 3275 Node* fast_io = slow_call->in(TypeFunc::I_O); 3276 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3277 3278 // These two phis are pre-filled with copies of of the fast IO and Memory 3279 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3280 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3281 3282 result_rgn->init_req(slow_result_path, control()); 3283 result_io ->init_req(slow_result_path, i_o()); 3284 result_mem->init_req(slow_result_path, reset_memory()); 3285 result_val->init_req(slow_result_path, slow_val); 3286 3287 set_all_memory(_gvn.transform(result_mem)); 3288 set_i_o( _gvn.transform(result_io)); 3289 } 3290 3291 C->set_has_split_ifs(true); // Has chance for split-if optimization 3292 set_result(result_rgn, result_val); 3293 return true; 3294 } 3295 3296 //---------------------------load_mirror_from_klass---------------------------- 3297 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3298 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3299 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3300 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3301 } 3302 3303 //-----------------------load_klass_from_mirror_common------------------------- 3304 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3305 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3306 // and branch to the given path on the region. 3307 // If never_see_null, take an uncommon trap on null, so we can optimistically 3308 // compile for the non-null case. 3309 // If the region is NULL, force never_see_null = true. 3310 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3311 bool never_see_null, 3312 RegionNode* region, 3313 int null_path, 3314 int offset) { 3315 if (region == NULL) never_see_null = true; 3316 Node* p = basic_plus_adr(mirror, offset); 3317 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3318 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3319 Node* null_ctl = top(); 3320 kls = null_check_oop(kls, &null_ctl, never_see_null); 3321 if (region != NULL) { 3322 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3323 region->init_req(null_path, null_ctl); 3324 } else { 3325 assert(null_ctl == top(), "no loose ends"); 3326 } 3327 return kls; 3328 } 3329 3330 //--------------------(inline_native_Class_query helpers)--------------------- 3331 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3332 // Fall through if (mods & mask) == bits, take the guard otherwise. 3333 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3334 // Branch around if the given klass has the given modifier bit set. 3335 // Like generate_guard, adds a new path onto the region. 3336 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3337 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3338 Node* mask = intcon(modifier_mask); 3339 Node* bits = intcon(modifier_bits); 3340 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3341 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3342 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3343 return generate_fair_guard(bol, region); 3344 } 3345 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3346 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3347 } 3348 3349 //-------------------------inline_native_Class_query------------------- 3350 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3351 const Type* return_type = TypeInt::BOOL; 3352 Node* prim_return_value = top(); // what happens if it's a primitive class? 3353 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3354 bool expect_prim = false; // most of these guys expect to work on refs 3355 3356 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3357 3358 Node* mirror = argument(0); 3359 Node* obj = top(); 3360 3361 switch (id) { 3362 case vmIntrinsics::_isInstance: 3363 // nothing is an instance of a primitive type 3364 prim_return_value = intcon(0); 3365 obj = argument(1); 3366 break; 3367 case vmIntrinsics::_getModifiers: 3368 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3369 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3370 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3371 break; 3372 case vmIntrinsics::_isInterface: 3373 prim_return_value = intcon(0); 3374 break; 3375 case vmIntrinsics::_isArray: 3376 prim_return_value = intcon(0); 3377 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3378 break; 3379 case vmIntrinsics::_isPrimitive: 3380 prim_return_value = intcon(1); 3381 expect_prim = true; // obviously 3382 break; 3383 case vmIntrinsics::_getSuperclass: 3384 prim_return_value = null(); 3385 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3386 break; 3387 case vmIntrinsics::_getClassAccessFlags: 3388 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3389 return_type = TypeInt::INT; // not bool! 6297094 3390 break; 3391 default: 3392 fatal_unexpected_iid(id); 3393 break; 3394 } 3395 3396 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3397 if (mirror_con == NULL) return false; // cannot happen? 3398 3399 #ifndef PRODUCT 3400 if (C->print_intrinsics() || C->print_inlining()) { 3401 ciType* k = mirror_con->java_mirror_type(); 3402 if (k) { 3403 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3404 k->print_name(); 3405 tty->cr(); 3406 } 3407 } 3408 #endif 3409 3410 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3411 RegionNode* region = new RegionNode(PATH_LIMIT); 3412 record_for_igvn(region); 3413 PhiNode* phi = new PhiNode(region, return_type); 3414 3415 // The mirror will never be null of Reflection.getClassAccessFlags, however 3416 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3417 // if it is. See bug 4774291. 3418 3419 // For Reflection.getClassAccessFlags(), the null check occurs in 3420 // the wrong place; see inline_unsafe_access(), above, for a similar 3421 // situation. 3422 mirror = null_check(mirror); 3423 // If mirror or obj is dead, only null-path is taken. 3424 if (stopped()) return true; 3425 3426 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3427 3428 // Now load the mirror's klass metaobject, and null-check it. 3429 // Side-effects region with the control path if the klass is null. 3430 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3431 // If kls is null, we have a primitive mirror. 3432 phi->init_req(_prim_path, prim_return_value); 3433 if (stopped()) { set_result(region, phi); return true; } 3434 bool safe_for_replace = (region->in(_prim_path) == top()); 3435 3436 Node* p; // handy temp 3437 Node* null_ctl; 3438 3439 // Now that we have the non-null klass, we can perform the real query. 3440 // For constant classes, the query will constant-fold in LoadNode::Value. 3441 Node* query_value = top(); 3442 switch (id) { 3443 case vmIntrinsics::_isInstance: 3444 // nothing is an instance of a primitive type 3445 query_value = gen_instanceof(obj, kls, safe_for_replace); 3446 break; 3447 3448 case vmIntrinsics::_getModifiers: 3449 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3450 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3451 break; 3452 3453 case vmIntrinsics::_isInterface: 3454 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3455 if (generate_interface_guard(kls, region) != NULL) 3456 // A guard was added. If the guard is taken, it was an interface. 3457 phi->add_req(intcon(1)); 3458 // If we fall through, it's a plain class. 3459 query_value = intcon(0); 3460 break; 3461 3462 case vmIntrinsics::_isArray: 3463 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3464 if (generate_array_guard(kls, region) != NULL) 3465 // A guard was added. If the guard is taken, it was an array. 3466 phi->add_req(intcon(1)); 3467 // If we fall through, it's a plain class. 3468 query_value = intcon(0); 3469 break; 3470 3471 case vmIntrinsics::_isPrimitive: 3472 query_value = intcon(0); // "normal" path produces false 3473 break; 3474 3475 case vmIntrinsics::_getSuperclass: 3476 // The rules here are somewhat unfortunate, but we can still do better 3477 // with random logic than with a JNI call. 3478 // Interfaces store null or Object as _super, but must report null. 3479 // Arrays store an intermediate super as _super, but must report Object. 3480 // Other types can report the actual _super. 3481 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3482 if (generate_interface_guard(kls, region) != NULL) 3483 // A guard was added. If the guard is taken, it was an interface. 3484 phi->add_req(null()); 3485 if (generate_array_guard(kls, region) != NULL) 3486 // A guard was added. If the guard is taken, it was an array. 3487 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3488 // If we fall through, it's a plain class. Get its _super. 3489 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3490 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3491 null_ctl = top(); 3492 kls = null_check_oop(kls, &null_ctl); 3493 if (null_ctl != top()) { 3494 // If the guard is taken, Object.superClass is null (both klass and mirror). 3495 region->add_req(null_ctl); 3496 phi ->add_req(null()); 3497 } 3498 if (!stopped()) { 3499 query_value = load_mirror_from_klass(kls); 3500 } 3501 break; 3502 3503 case vmIntrinsics::_getClassAccessFlags: 3504 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3505 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3506 break; 3507 3508 default: 3509 fatal_unexpected_iid(id); 3510 break; 3511 } 3512 3513 // Fall-through is the normal case of a query to a real class. 3514 phi->init_req(1, query_value); 3515 region->init_req(1, control()); 3516 3517 C->set_has_split_ifs(true); // Has chance for split-if optimization 3518 set_result(region, phi); 3519 return true; 3520 } 3521 3522 //-------------------------inline_Class_cast------------------- 3523 bool LibraryCallKit::inline_Class_cast() { 3524 Node* mirror = argument(0); // Class 3525 Node* obj = argument(1); 3526 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3527 if (mirror_con == NULL) { 3528 return false; // dead path (mirror->is_top()). 3529 } 3530 if (obj == NULL || obj->is_top()) { 3531 return false; // dead path 3532 } 3533 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3534 3535 // First, see if Class.cast() can be folded statically. 3536 // java_mirror_type() returns non-null for compile-time Class constants. 3537 ciType* tm = mirror_con->java_mirror_type(); 3538 if (tm != NULL && tm->is_klass() && 3539 tp != NULL && tp->klass() != NULL) { 3540 if (!tp->klass()->is_loaded()) { 3541 // Don't use intrinsic when class is not loaded. 3542 return false; 3543 } else { 3544 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3545 if (static_res == Compile::SSC_always_true) { 3546 // isInstance() is true - fold the code. 3547 set_result(obj); 3548 return true; 3549 } else if (static_res == Compile::SSC_always_false) { 3550 // Don't use intrinsic, have to throw ClassCastException. 3551 // If the reference is null, the non-intrinsic bytecode will 3552 // be optimized appropriately. 3553 return false; 3554 } 3555 } 3556 } 3557 3558 // Bailout intrinsic and do normal inlining if exception path is frequent. 3559 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3560 return false; 3561 } 3562 3563 // Generate dynamic checks. 3564 // Class.cast() is java implementation of _checkcast bytecode. 3565 // Do checkcast (Parse::do_checkcast()) optimizations here. 3566 3567 mirror = null_check(mirror); 3568 // If mirror is dead, only null-path is taken. 3569 if (stopped()) { 3570 return true; 3571 } 3572 3573 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3574 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3575 RegionNode* region = new RegionNode(PATH_LIMIT); 3576 record_for_igvn(region); 3577 3578 // Now load the mirror's klass metaobject, and null-check it. 3579 // If kls is null, we have a primitive mirror and 3580 // nothing is an instance of a primitive type. 3581 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3582 3583 Node* res = top(); 3584 if (!stopped()) { 3585 Node* bad_type_ctrl = top(); 3586 // Do checkcast optimizations. 3587 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3588 region->init_req(_bad_type_path, bad_type_ctrl); 3589 } 3590 if (region->in(_prim_path) != top() || 3591 region->in(_bad_type_path) != top()) { 3592 // Let Interpreter throw ClassCastException. 3593 PreserveJVMState pjvms(this); 3594 set_control(_gvn.transform(region)); 3595 uncommon_trap(Deoptimization::Reason_intrinsic, 3596 Deoptimization::Action_maybe_recompile); 3597 } 3598 if (!stopped()) { 3599 set_result(res); 3600 } 3601 return true; 3602 } 3603 3604 3605 //--------------------------inline_native_subtype_check------------------------ 3606 // This intrinsic takes the JNI calls out of the heart of 3607 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3608 bool LibraryCallKit::inline_native_subtype_check() { 3609 // Pull both arguments off the stack. 3610 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3611 args[0] = argument(0); 3612 args[1] = argument(1); 3613 Node* klasses[2]; // corresponding Klasses: superk, subk 3614 klasses[0] = klasses[1] = top(); 3615 3616 enum { 3617 // A full decision tree on {superc is prim, subc is prim}: 3618 _prim_0_path = 1, // {P,N} => false 3619 // {P,P} & superc!=subc => false 3620 _prim_same_path, // {P,P} & superc==subc => true 3621 _prim_1_path, // {N,P} => false 3622 _ref_subtype_path, // {N,N} & subtype check wins => true 3623 _both_ref_path, // {N,N} & subtype check loses => false 3624 PATH_LIMIT 3625 }; 3626 3627 RegionNode* region = new RegionNode(PATH_LIMIT); 3628 Node* phi = new PhiNode(region, TypeInt::BOOL); 3629 record_for_igvn(region); 3630 3631 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3632 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3633 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3634 3635 // First null-check both mirrors and load each mirror's klass metaobject. 3636 int which_arg; 3637 for (which_arg = 0; which_arg <= 1; which_arg++) { 3638 Node* arg = args[which_arg]; 3639 arg = null_check(arg); 3640 if (stopped()) break; 3641 args[which_arg] = arg; 3642 3643 Node* p = basic_plus_adr(arg, class_klass_offset); 3644 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3645 klasses[which_arg] = _gvn.transform(kls); 3646 } 3647 3648 // Having loaded both klasses, test each for null. 3649 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3650 for (which_arg = 0; which_arg <= 1; which_arg++) { 3651 Node* kls = klasses[which_arg]; 3652 Node* null_ctl = top(); 3653 kls = null_check_oop(kls, &null_ctl, never_see_null); 3654 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3655 region->init_req(prim_path, null_ctl); 3656 if (stopped()) break; 3657 klasses[which_arg] = kls; 3658 } 3659 3660 if (!stopped()) { 3661 // now we have two reference types, in klasses[0..1] 3662 Node* subk = klasses[1]; // the argument to isAssignableFrom 3663 Node* superk = klasses[0]; // the receiver 3664 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3665 // now we have a successful reference subtype check 3666 region->set_req(_ref_subtype_path, control()); 3667 } 3668 3669 // If both operands are primitive (both klasses null), then 3670 // we must return true when they are identical primitives. 3671 // It is convenient to test this after the first null klass check. 3672 set_control(region->in(_prim_0_path)); // go back to first null check 3673 if (!stopped()) { 3674 // Since superc is primitive, make a guard for the superc==subc case. 3675 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3676 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3677 generate_guard(bol_eq, region, PROB_FAIR); 3678 if (region->req() == PATH_LIMIT+1) { 3679 // A guard was added. If the added guard is taken, superc==subc. 3680 region->swap_edges(PATH_LIMIT, _prim_same_path); 3681 region->del_req(PATH_LIMIT); 3682 } 3683 region->set_req(_prim_0_path, control()); // Not equal after all. 3684 } 3685 3686 // these are the only paths that produce 'true': 3687 phi->set_req(_prim_same_path, intcon(1)); 3688 phi->set_req(_ref_subtype_path, intcon(1)); 3689 3690 // pull together the cases: 3691 assert(region->req() == PATH_LIMIT, "sane region"); 3692 for (uint i = 1; i < region->req(); i++) { 3693 Node* ctl = region->in(i); 3694 if (ctl == NULL || ctl == top()) { 3695 region->set_req(i, top()); 3696 phi ->set_req(i, top()); 3697 } else if (phi->in(i) == NULL) { 3698 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3699 } 3700 } 3701 3702 set_control(_gvn.transform(region)); 3703 set_result(_gvn.transform(phi)); 3704 return true; 3705 } 3706 3707 //---------------------generate_array_guard_common------------------------ 3708 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3709 bool obj_array, bool not_array) { 3710 3711 if (stopped()) { 3712 return NULL; 3713 } 3714 3715 // If obj_array/non_array==false/false: 3716 // Branch around if the given klass is in fact an array (either obj or prim). 3717 // If obj_array/non_array==false/true: 3718 // Branch around if the given klass is not an array klass of any kind. 3719 // If obj_array/non_array==true/true: 3720 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3721 // If obj_array/non_array==true/false: 3722 // Branch around if the kls is an oop array (Object[] or subtype) 3723 // 3724 // Like generate_guard, adds a new path onto the region. 3725 jint layout_con = 0; 3726 Node* layout_val = get_layout_helper(kls, layout_con); 3727 if (layout_val == NULL) { 3728 bool query = (obj_array 3729 ? Klass::layout_helper_is_objArray(layout_con) 3730 : Klass::layout_helper_is_array(layout_con)); 3731 if (query == not_array) { 3732 return NULL; // never a branch 3733 } else { // always a branch 3734 Node* always_branch = control(); 3735 if (region != NULL) 3736 region->add_req(always_branch); 3737 set_control(top()); 3738 return always_branch; 3739 } 3740 } 3741 // Now test the correct condition. 3742 jint nval = (obj_array 3743 ? ((jint)Klass::_lh_array_tag_type_value 3744 << Klass::_lh_array_tag_shift) 3745 : Klass::_lh_neutral_value); 3746 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3747 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3748 // invert the test if we are looking for a non-array 3749 if (not_array) btest = BoolTest(btest).negate(); 3750 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3751 return generate_fair_guard(bol, region); 3752 } 3753 3754 3755 //-----------------------inline_native_newArray-------------------------- 3756 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3757 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 3758 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 3759 Node* mirror; 3760 Node* count_val; 3761 if (uninitialized) { 3762 mirror = argument(1); 3763 count_val = argument(2); 3764 } else { 3765 mirror = argument(0); 3766 count_val = argument(1); 3767 } 3768 3769 mirror = null_check(mirror); 3770 // If mirror or obj is dead, only null-path is taken. 3771 if (stopped()) return true; 3772 3773 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3774 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3775 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3776 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3777 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3778 3779 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3780 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3781 result_reg, _slow_path); 3782 Node* normal_ctl = control(); 3783 Node* no_array_ctl = result_reg->in(_slow_path); 3784 3785 // Generate code for the slow case. We make a call to newArray(). 3786 set_control(no_array_ctl); 3787 if (!stopped()) { 3788 // Either the input type is void.class, or else the 3789 // array klass has not yet been cached. Either the 3790 // ensuing call will throw an exception, or else it 3791 // will cache the array klass for next time. 3792 PreserveJVMState pjvms(this); 3793 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3794 Node* slow_result = set_results_for_java_call(slow_call); 3795 // this->control() comes from set_results_for_java_call 3796 result_reg->set_req(_slow_path, control()); 3797 result_val->set_req(_slow_path, slow_result); 3798 result_io ->set_req(_slow_path, i_o()); 3799 result_mem->set_req(_slow_path, reset_memory()); 3800 } 3801 3802 set_control(normal_ctl); 3803 if (!stopped()) { 3804 // Normal case: The array type has been cached in the java.lang.Class. 3805 // The following call works fine even if the array type is polymorphic. 3806 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3807 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3808 result_reg->init_req(_normal_path, control()); 3809 result_val->init_req(_normal_path, obj); 3810 result_io ->init_req(_normal_path, i_o()); 3811 result_mem->init_req(_normal_path, reset_memory()); 3812 3813 if (uninitialized) { 3814 // Mark the allocation so that zeroing is skipped 3815 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 3816 alloc->maybe_set_complete(&_gvn); 3817 } 3818 } 3819 3820 // Return the combined state. 3821 set_i_o( _gvn.transform(result_io) ); 3822 set_all_memory( _gvn.transform(result_mem)); 3823 3824 C->set_has_split_ifs(true); // Has chance for split-if optimization 3825 set_result(result_reg, result_val); 3826 return true; 3827 } 3828 3829 //----------------------inline_native_getLength-------------------------- 3830 // public static native int java.lang.reflect.Array.getLength(Object array); 3831 bool LibraryCallKit::inline_native_getLength() { 3832 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3833 3834 Node* array = null_check(argument(0)); 3835 // If array is dead, only null-path is taken. 3836 if (stopped()) return true; 3837 3838 // Deoptimize if it is a non-array. 3839 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3840 3841 if (non_array != NULL) { 3842 PreserveJVMState pjvms(this); 3843 set_control(non_array); 3844 uncommon_trap(Deoptimization::Reason_intrinsic, 3845 Deoptimization::Action_maybe_recompile); 3846 } 3847 3848 // If control is dead, only non-array-path is taken. 3849 if (stopped()) return true; 3850 3851 // The works fine even if the array type is polymorphic. 3852 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3853 Node* result = load_array_length(array); 3854 3855 C->set_has_split_ifs(true); // Has chance for split-if optimization 3856 set_result(result); 3857 return true; 3858 } 3859 3860 //------------------------inline_array_copyOf---------------------------- 3861 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3862 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3863 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3864 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3865 3866 // Get the arguments. 3867 Node* original = argument(0); 3868 Node* start = is_copyOfRange? argument(1): intcon(0); 3869 Node* end = is_copyOfRange? argument(2): argument(1); 3870 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3871 3872 Node* newcopy = NULL; 3873 3874 // Set the original stack and the reexecute bit for the interpreter to reexecute 3875 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3876 { PreserveReexecuteState preexecs(this); 3877 jvms()->set_should_reexecute(true); 3878 3879 array_type_mirror = null_check(array_type_mirror); 3880 original = null_check(original); 3881 3882 // Check if a null path was taken unconditionally. 3883 if (stopped()) return true; 3884 3885 Node* orig_length = load_array_length(original); 3886 3887 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3888 klass_node = null_check(klass_node); 3889 3890 RegionNode* bailout = new RegionNode(1); 3891 record_for_igvn(bailout); 3892 3893 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3894 // Bail out if that is so. 3895 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3896 if (not_objArray != NULL) { 3897 // Improve the klass node's type from the new optimistic assumption: 3898 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3899 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3900 Node* cast = new CastPPNode(klass_node, akls); 3901 cast->init_req(0, control()); 3902 klass_node = _gvn.transform(cast); 3903 } 3904 3905 // Bail out if either start or end is negative. 3906 generate_negative_guard(start, bailout, &start); 3907 generate_negative_guard(end, bailout, &end); 3908 3909 Node* length = end; 3910 if (_gvn.type(start) != TypeInt::ZERO) { 3911 length = _gvn.transform(new SubINode(end, start)); 3912 } 3913 3914 // Bail out if length is negative. 3915 // Without this the new_array would throw 3916 // NegativeArraySizeException but IllegalArgumentException is what 3917 // should be thrown 3918 generate_negative_guard(length, bailout, &length); 3919 3920 if (bailout->req() > 1) { 3921 PreserveJVMState pjvms(this); 3922 set_control(_gvn.transform(bailout)); 3923 uncommon_trap(Deoptimization::Reason_intrinsic, 3924 Deoptimization::Action_maybe_recompile); 3925 } 3926 3927 if (!stopped()) { 3928 // How many elements will we copy from the original? 3929 // The answer is MinI(orig_length - start, length). 3930 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3931 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3932 3933 // Generate a direct call to the right arraycopy function(s). 3934 // We know the copy is disjoint but we might not know if the 3935 // oop stores need checking. 3936 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3937 // This will fail a store-check if x contains any non-nulls. 3938 3939 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3940 // loads/stores but it is legal only if we're sure the 3941 // Arrays.copyOf would succeed. So we need all input arguments 3942 // to the copyOf to be validated, including that the copy to the 3943 // new array won't trigger an ArrayStoreException. That subtype 3944 // check can be optimized if we know something on the type of 3945 // the input array from type speculation. 3946 if (_gvn.type(klass_node)->singleton()) { 3947 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3948 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3949 3950 int test = C->static_subtype_check(superk, subk); 3951 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3952 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3953 if (t_original->speculative_type() != NULL) { 3954 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3955 } 3956 } 3957 } 3958 3959 bool validated = false; 3960 // Reason_class_check rather than Reason_intrinsic because we 3961 // want to intrinsify even if this traps. 3962 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3963 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3964 klass_node); 3965 3966 if (not_subtype_ctrl != top()) { 3967 PreserveJVMState pjvms(this); 3968 set_control(not_subtype_ctrl); 3969 uncommon_trap(Deoptimization::Reason_class_check, 3970 Deoptimization::Action_make_not_entrant); 3971 assert(stopped(), "Should be stopped"); 3972 } 3973 validated = true; 3974 } 3975 3976 if (!stopped()) { 3977 newcopy = new_array(klass_node, length, 0); // no arguments to push 3978 3979 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3980 load_object_klass(original), klass_node); 3981 if (!is_copyOfRange) { 3982 ac->set_copyof(validated); 3983 } else { 3984 ac->set_copyofrange(validated); 3985 } 3986 Node* n = _gvn.transform(ac); 3987 if (n == ac) { 3988 ac->connect_outputs(this); 3989 } else { 3990 assert(validated, "shouldn't transform if all arguments not validated"); 3991 set_all_memory(n); 3992 } 3993 } 3994 } 3995 } // original reexecute is set back here 3996 3997 C->set_has_split_ifs(true); // Has chance for split-if optimization 3998 if (!stopped()) { 3999 set_result(newcopy); 4000 } 4001 return true; 4002 } 4003 4004 4005 //----------------------generate_virtual_guard--------------------------- 4006 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4007 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4008 RegionNode* slow_region) { 4009 ciMethod* method = callee(); 4010 int vtable_index = method->vtable_index(); 4011 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4012 "bad index %d", vtable_index); 4013 // Get the Method* out of the appropriate vtable entry. 4014 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 4015 vtable_index*vtableEntry::size_in_bytes() + 4016 vtableEntry::method_offset_in_bytes(); 4017 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4018 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4019 4020 // Compare the target method with the expected method (e.g., Object.hashCode). 4021 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4022 4023 Node* native_call = makecon(native_call_addr); 4024 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4025 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4026 4027 return generate_slow_guard(test_native, slow_region); 4028 } 4029 4030 //-----------------------generate_method_call---------------------------- 4031 // Use generate_method_call to make a slow-call to the real 4032 // method if the fast path fails. An alternative would be to 4033 // use a stub like OptoRuntime::slow_arraycopy_Java. 4034 // This only works for expanding the current library call, 4035 // not another intrinsic. (E.g., don't use this for making an 4036 // arraycopy call inside of the copyOf intrinsic.) 4037 CallJavaNode* 4038 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4039 // When compiling the intrinsic method itself, do not use this technique. 4040 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4041 4042 ciMethod* method = callee(); 4043 // ensure the JVMS we have will be correct for this call 4044 guarantee(method_id == method->intrinsic_id(), "must match"); 4045 4046 const TypeFunc* tf = TypeFunc::make(method); 4047 CallJavaNode* slow_call; 4048 if (is_static) { 4049 assert(!is_virtual, ""); 4050 slow_call = new CallStaticJavaNode(C, tf, 4051 SharedRuntime::get_resolve_static_call_stub(), 4052 method, bci()); 4053 } else if (is_virtual) { 4054 null_check_receiver(); 4055 int vtable_index = Method::invalid_vtable_index; 4056 if (UseInlineCaches) { 4057 // Suppress the vtable call 4058 } else { 4059 // hashCode and clone are not a miranda methods, 4060 // so the vtable index is fixed. 4061 // No need to use the linkResolver to get it. 4062 vtable_index = method->vtable_index(); 4063 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4064 "bad index %d", vtable_index); 4065 } 4066 slow_call = new CallDynamicJavaNode(tf, 4067 SharedRuntime::get_resolve_virtual_call_stub(), 4068 method, vtable_index, bci()); 4069 } else { // neither virtual nor static: opt_virtual 4070 null_check_receiver(); 4071 slow_call = new CallStaticJavaNode(C, tf, 4072 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4073 method, bci()); 4074 slow_call->set_optimized_virtual(true); 4075 } 4076 set_arguments_for_java_call(slow_call); 4077 set_edges_for_java_call(slow_call); 4078 return slow_call; 4079 } 4080 4081 4082 /** 4083 * Build special case code for calls to hashCode on an object. This call may 4084 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4085 * slightly different code. 4086 */ 4087 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4088 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4089 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4090 4091 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4092 4093 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4094 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4095 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4096 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4097 Node* obj = NULL; 4098 if (!is_static) { 4099 // Check for hashing null object 4100 obj = null_check_receiver(); 4101 if (stopped()) return true; // unconditionally null 4102 result_reg->init_req(_null_path, top()); 4103 result_val->init_req(_null_path, top()); 4104 } else { 4105 // Do a null check, and return zero if null. 4106 // System.identityHashCode(null) == 0 4107 obj = argument(0); 4108 Node* null_ctl = top(); 4109 obj = null_check_oop(obj, &null_ctl); 4110 result_reg->init_req(_null_path, null_ctl); 4111 result_val->init_req(_null_path, _gvn.intcon(0)); 4112 } 4113 4114 // Unconditionally null? Then return right away. 4115 if (stopped()) { 4116 set_control( result_reg->in(_null_path)); 4117 if (!stopped()) 4118 set_result(result_val->in(_null_path)); 4119 return true; 4120 } 4121 4122 // We only go to the fast case code if we pass a number of guards. The 4123 // paths which do not pass are accumulated in the slow_region. 4124 RegionNode* slow_region = new RegionNode(1); 4125 record_for_igvn(slow_region); 4126 4127 // If this is a virtual call, we generate a funny guard. We pull out 4128 // the vtable entry corresponding to hashCode() from the target object. 4129 // If the target method which we are calling happens to be the native 4130 // Object hashCode() method, we pass the guard. We do not need this 4131 // guard for non-virtual calls -- the caller is known to be the native 4132 // Object hashCode(). 4133 if (is_virtual) { 4134 // After null check, get the object's klass. 4135 Node* obj_klass = load_object_klass(obj); 4136 generate_virtual_guard(obj_klass, slow_region); 4137 } 4138 4139 // Get the header out of the object, use LoadMarkNode when available 4140 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4141 // The control of the load must be NULL. Otherwise, the load can move before 4142 // the null check after castPP removal. 4143 Node* no_ctrl = NULL; 4144 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4145 4146 // Test the header to see if it is unlocked. 4147 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4148 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4149 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4150 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4151 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4152 4153 generate_slow_guard(test_unlocked, slow_region); 4154 4155 // Get the hash value and check to see that it has been properly assigned. 4156 // We depend on hash_mask being at most 32 bits and avoid the use of 4157 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4158 // vm: see markOop.hpp. 4159 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4160 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4161 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4162 // This hack lets the hash bits live anywhere in the mark object now, as long 4163 // as the shift drops the relevant bits into the low 32 bits. Note that 4164 // Java spec says that HashCode is an int so there's no point in capturing 4165 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4166 hshifted_header = ConvX2I(hshifted_header); 4167 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4168 4169 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4170 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4171 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4172 4173 generate_slow_guard(test_assigned, slow_region); 4174 4175 Node* init_mem = reset_memory(); 4176 // fill in the rest of the null path: 4177 result_io ->init_req(_null_path, i_o()); 4178 result_mem->init_req(_null_path, init_mem); 4179 4180 result_val->init_req(_fast_path, hash_val); 4181 result_reg->init_req(_fast_path, control()); 4182 result_io ->init_req(_fast_path, i_o()); 4183 result_mem->init_req(_fast_path, init_mem); 4184 4185 // Generate code for the slow case. We make a call to hashCode(). 4186 set_control(_gvn.transform(slow_region)); 4187 if (!stopped()) { 4188 // No need for PreserveJVMState, because we're using up the present state. 4189 set_all_memory(init_mem); 4190 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4191 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4192 Node* slow_result = set_results_for_java_call(slow_call); 4193 // this->control() comes from set_results_for_java_call 4194 result_reg->init_req(_slow_path, control()); 4195 result_val->init_req(_slow_path, slow_result); 4196 result_io ->set_req(_slow_path, i_o()); 4197 result_mem ->set_req(_slow_path, reset_memory()); 4198 } 4199 4200 // Return the combined state. 4201 set_i_o( _gvn.transform(result_io) ); 4202 set_all_memory( _gvn.transform(result_mem)); 4203 4204 set_result(result_reg, result_val); 4205 return true; 4206 } 4207 4208 //---------------------------inline_native_getClass---------------------------- 4209 // public final native Class<?> java.lang.Object.getClass(); 4210 // 4211 // Build special case code for calls to getClass on an object. 4212 bool LibraryCallKit::inline_native_getClass() { 4213 Node* obj = null_check_receiver(); 4214 if (stopped()) return true; 4215 set_result(load_mirror_from_klass(load_object_klass(obj))); 4216 return true; 4217 } 4218 4219 //-----------------inline_native_Reflection_getCallerClass--------------------- 4220 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4221 // 4222 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4223 // 4224 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4225 // in that it must skip particular security frames and checks for 4226 // caller sensitive methods. 4227 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4228 #ifndef PRODUCT 4229 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4230 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4231 } 4232 #endif 4233 4234 if (!jvms()->has_method()) { 4235 #ifndef PRODUCT 4236 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4237 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4238 } 4239 #endif 4240 return false; 4241 } 4242 4243 // Walk back up the JVM state to find the caller at the required 4244 // depth. 4245 JVMState* caller_jvms = jvms(); 4246 4247 // Cf. JVM_GetCallerClass 4248 // NOTE: Start the loop at depth 1 because the current JVM state does 4249 // not include the Reflection.getCallerClass() frame. 4250 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4251 ciMethod* m = caller_jvms->method(); 4252 switch (n) { 4253 case 0: 4254 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4255 break; 4256 case 1: 4257 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4258 if (!m->caller_sensitive()) { 4259 #ifndef PRODUCT 4260 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4261 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4262 } 4263 #endif 4264 return false; // bail-out; let JVM_GetCallerClass do the work 4265 } 4266 break; 4267 default: 4268 if (!m->is_ignored_by_security_stack_walk()) { 4269 // We have reached the desired frame; return the holder class. 4270 // Acquire method holder as java.lang.Class and push as constant. 4271 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4272 ciInstance* caller_mirror = caller_klass->java_mirror(); 4273 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4274 4275 #ifndef PRODUCT 4276 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4277 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()); 4278 tty->print_cr(" JVM state at this point:"); 4279 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4280 ciMethod* m = jvms()->of_depth(i)->method(); 4281 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4282 } 4283 } 4284 #endif 4285 return true; 4286 } 4287 break; 4288 } 4289 } 4290 4291 #ifndef PRODUCT 4292 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4293 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4294 tty->print_cr(" JVM state at this point:"); 4295 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4296 ciMethod* m = jvms()->of_depth(i)->method(); 4297 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4298 } 4299 } 4300 #endif 4301 4302 return false; // bail-out; let JVM_GetCallerClass do the work 4303 } 4304 4305 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4306 Node* arg = argument(0); 4307 Node* result = NULL; 4308 4309 switch (id) { 4310 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4311 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4312 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4313 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4314 4315 case vmIntrinsics::_doubleToLongBits: { 4316 // two paths (plus control) merge in a wood 4317 RegionNode *r = new RegionNode(3); 4318 Node *phi = new PhiNode(r, TypeLong::LONG); 4319 4320 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4321 // Build the boolean node 4322 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4323 4324 // Branch either way. 4325 // NaN case is less traveled, which makes all the difference. 4326 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4327 Node *opt_isnan = _gvn.transform(ifisnan); 4328 assert( opt_isnan->is_If(), "Expect an IfNode"); 4329 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4330 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4331 4332 set_control(iftrue); 4333 4334 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4335 Node *slow_result = longcon(nan_bits); // return NaN 4336 phi->init_req(1, _gvn.transform( slow_result )); 4337 r->init_req(1, iftrue); 4338 4339 // Else fall through 4340 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4341 set_control(iffalse); 4342 4343 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4344 r->init_req(2, iffalse); 4345 4346 // Post merge 4347 set_control(_gvn.transform(r)); 4348 record_for_igvn(r); 4349 4350 C->set_has_split_ifs(true); // Has chance for split-if optimization 4351 result = phi; 4352 assert(result->bottom_type()->isa_long(), "must be"); 4353 break; 4354 } 4355 4356 case vmIntrinsics::_floatToIntBits: { 4357 // two paths (plus control) merge in a wood 4358 RegionNode *r = new RegionNode(3); 4359 Node *phi = new PhiNode(r, TypeInt::INT); 4360 4361 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4362 // Build the boolean node 4363 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4364 4365 // Branch either way. 4366 // NaN case is less traveled, which makes all the difference. 4367 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4368 Node *opt_isnan = _gvn.transform(ifisnan); 4369 assert( opt_isnan->is_If(), "Expect an IfNode"); 4370 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4371 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4372 4373 set_control(iftrue); 4374 4375 static const jint nan_bits = 0x7fc00000; 4376 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4377 phi->init_req(1, _gvn.transform( slow_result )); 4378 r->init_req(1, iftrue); 4379 4380 // Else fall through 4381 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4382 set_control(iffalse); 4383 4384 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4385 r->init_req(2, iffalse); 4386 4387 // Post merge 4388 set_control(_gvn.transform(r)); 4389 record_for_igvn(r); 4390 4391 C->set_has_split_ifs(true); // Has chance for split-if optimization 4392 result = phi; 4393 assert(result->bottom_type()->isa_int(), "must be"); 4394 break; 4395 } 4396 4397 default: 4398 fatal_unexpected_iid(id); 4399 break; 4400 } 4401 set_result(_gvn.transform(result)); 4402 return true; 4403 } 4404 4405 //----------------------inline_unsafe_copyMemory------------------------- 4406 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4407 bool LibraryCallKit::inline_unsafe_copyMemory() { 4408 if (callee()->is_static()) return false; // caller must have the capability! 4409 null_check_receiver(); // null-check receiver 4410 if (stopped()) return true; 4411 4412 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4413 4414 Node* src_ptr = argument(1); // type: oop 4415 Node* src_off = ConvL2X(argument(2)); // type: long 4416 Node* dst_ptr = argument(4); // type: oop 4417 Node* dst_off = ConvL2X(argument(5)); // type: long 4418 Node* size = ConvL2X(argument(7)); // type: long 4419 4420 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4421 "fieldOffset must be byte-scaled"); 4422 4423 Node* src = make_unsafe_address(src_ptr, src_off); 4424 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4425 4426 // Conservatively insert a memory barrier on all memory slices. 4427 // Do not let writes of the copy source or destination float below the copy. 4428 insert_mem_bar(Op_MemBarCPUOrder); 4429 4430 // Call it. Note that the length argument is not scaled. 4431 make_runtime_call(RC_LEAF|RC_NO_FP, 4432 OptoRuntime::fast_arraycopy_Type(), 4433 StubRoutines::unsafe_arraycopy(), 4434 "unsafe_arraycopy", 4435 TypeRawPtr::BOTTOM, 4436 src, dst, size XTOP); 4437 4438 // Do not let reads of the copy destination float above the copy. 4439 insert_mem_bar(Op_MemBarCPUOrder); 4440 4441 return true; 4442 } 4443 4444 //------------------------clone_coping----------------------------------- 4445 // Helper function for inline_native_clone. 4446 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4447 assert(obj_size != NULL, ""); 4448 Node* raw_obj = alloc_obj->in(1); 4449 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4450 4451 AllocateNode* alloc = NULL; 4452 if (ReduceBulkZeroing) { 4453 // We will be completely responsible for initializing this object - 4454 // mark Initialize node as complete. 4455 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4456 // The object was just allocated - there should be no any stores! 4457 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4458 // Mark as complete_with_arraycopy so that on AllocateNode 4459 // expansion, we know this AllocateNode is initialized by an array 4460 // copy and a StoreStore barrier exists after the array copy. 4461 alloc->initialization()->set_complete_with_arraycopy(); 4462 } 4463 4464 // Copy the fastest available way. 4465 // TODO: generate fields copies for small objects instead. 4466 Node* src = obj; 4467 Node* dest = alloc_obj; 4468 Node* size = _gvn.transform(obj_size); 4469 4470 // Exclude the header but include array length to copy by 8 bytes words. 4471 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4472 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4473 instanceOopDesc::base_offset_in_bytes(); 4474 // base_off: 4475 // 8 - 32-bit VM 4476 // 12 - 64-bit VM, compressed klass 4477 // 16 - 64-bit VM, normal klass 4478 if (base_off % BytesPerLong != 0) { 4479 assert(UseCompressedClassPointers, ""); 4480 if (is_array) { 4481 // Exclude length to copy by 8 bytes words. 4482 base_off += sizeof(int); 4483 } else { 4484 // Include klass to copy by 8 bytes words. 4485 base_off = instanceOopDesc::klass_offset_in_bytes(); 4486 } 4487 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4488 } 4489 src = basic_plus_adr(src, base_off); 4490 dest = basic_plus_adr(dest, base_off); 4491 4492 // Compute the length also, if needed: 4493 Node* countx = size; 4494 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4495 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4496 4497 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4498 4499 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4500 ac->set_clonebasic(); 4501 Node* n = _gvn.transform(ac); 4502 if (n == ac) { 4503 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4504 } else { 4505 set_all_memory(n); 4506 } 4507 4508 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4509 if (card_mark) { 4510 assert(!is_array, ""); 4511 // Put in store barrier for any and all oops we are sticking 4512 // into this object. (We could avoid this if we could prove 4513 // that the object type contains no oop fields at all.) 4514 Node* no_particular_value = NULL; 4515 Node* no_particular_field = NULL; 4516 int raw_adr_idx = Compile::AliasIdxRaw; 4517 post_barrier(control(), 4518 memory(raw_adr_type), 4519 alloc_obj, 4520 no_particular_field, 4521 raw_adr_idx, 4522 no_particular_value, 4523 T_OBJECT, 4524 false); 4525 } 4526 4527 // Do not let reads from the cloned object float above the arraycopy. 4528 if (alloc != NULL) { 4529 // Do not let stores that initialize this object be reordered with 4530 // a subsequent store that would make this object accessible by 4531 // other threads. 4532 // Record what AllocateNode this StoreStore protects so that 4533 // escape analysis can go from the MemBarStoreStoreNode to the 4534 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4535 // based on the escape status of the AllocateNode. 4536 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4537 } else { 4538 insert_mem_bar(Op_MemBarCPUOrder); 4539 } 4540 } 4541 4542 //------------------------inline_native_clone---------------------------- 4543 // protected native Object java.lang.Object.clone(); 4544 // 4545 // Here are the simple edge cases: 4546 // null receiver => normal trap 4547 // virtual and clone was overridden => slow path to out-of-line clone 4548 // not cloneable or finalizer => slow path to out-of-line Object.clone 4549 // 4550 // The general case has two steps, allocation and copying. 4551 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4552 // 4553 // Copying also has two cases, oop arrays and everything else. 4554 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4555 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4556 // 4557 // These steps fold up nicely if and when the cloned object's klass 4558 // can be sharply typed as an object array, a type array, or an instance. 4559 // 4560 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4561 PhiNode* result_val; 4562 4563 // Set the reexecute bit for the interpreter to reexecute 4564 // the bytecode that invokes Object.clone if deoptimization happens. 4565 { PreserveReexecuteState preexecs(this); 4566 jvms()->set_should_reexecute(true); 4567 4568 Node* obj = null_check_receiver(); 4569 if (stopped()) return true; 4570 4571 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4572 4573 // If we are going to clone an instance, we need its exact type to 4574 // know the number and types of fields to convert the clone to 4575 // loads/stores. Maybe a speculative type can help us. 4576 if (!obj_type->klass_is_exact() && 4577 obj_type->speculative_type() != NULL && 4578 obj_type->speculative_type()->is_instance_klass()) { 4579 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4580 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4581 !spec_ik->has_injected_fields()) { 4582 ciKlass* k = obj_type->klass(); 4583 if (!k->is_instance_klass() || 4584 k->as_instance_klass()->is_interface() || 4585 k->as_instance_klass()->has_subklass()) { 4586 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4587 } 4588 } 4589 } 4590 4591 Node* obj_klass = load_object_klass(obj); 4592 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4593 const TypeOopPtr* toop = ((tklass != NULL) 4594 ? tklass->as_instance_type() 4595 : TypeInstPtr::NOTNULL); 4596 4597 // Conservatively insert a memory barrier on all memory slices. 4598 // Do not let writes into the original float below the clone. 4599 insert_mem_bar(Op_MemBarCPUOrder); 4600 4601 // paths into result_reg: 4602 enum { 4603 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4604 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4605 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4606 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4607 PATH_LIMIT 4608 }; 4609 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4610 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4611 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4612 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4613 record_for_igvn(result_reg); 4614 4615 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4616 int raw_adr_idx = Compile::AliasIdxRaw; 4617 4618 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4619 if (array_ctl != NULL) { 4620 // It's an array. 4621 PreserveJVMState pjvms(this); 4622 set_control(array_ctl); 4623 Node* obj_length = load_array_length(obj); 4624 Node* obj_size = NULL; 4625 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4626 4627 if (!use_ReduceInitialCardMarks()) { 4628 // If it is an oop array, it requires very special treatment, 4629 // because card marking is required on each card of the array. 4630 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4631 if (is_obja != NULL) { 4632 PreserveJVMState pjvms2(this); 4633 set_control(is_obja); 4634 // Generate a direct call to the right arraycopy function(s). 4635 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4636 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4637 ac->set_cloneoop(); 4638 Node* n = _gvn.transform(ac); 4639 assert(n == ac, "cannot disappear"); 4640 ac->connect_outputs(this); 4641 4642 result_reg->init_req(_objArray_path, control()); 4643 result_val->init_req(_objArray_path, alloc_obj); 4644 result_i_o ->set_req(_objArray_path, i_o()); 4645 result_mem ->set_req(_objArray_path, reset_memory()); 4646 } 4647 } 4648 // Otherwise, there are no card marks to worry about. 4649 // (We can dispense with card marks if we know the allocation 4650 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4651 // causes the non-eden paths to take compensating steps to 4652 // simulate a fresh allocation, so that no further 4653 // card marks are required in compiled code to initialize 4654 // the object.) 4655 4656 if (!stopped()) { 4657 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4658 4659 // Present the results of the copy. 4660 result_reg->init_req(_array_path, control()); 4661 result_val->init_req(_array_path, alloc_obj); 4662 result_i_o ->set_req(_array_path, i_o()); 4663 result_mem ->set_req(_array_path, reset_memory()); 4664 } 4665 } 4666 4667 // We only go to the instance fast case code if we pass a number of guards. 4668 // The paths which do not pass are accumulated in the slow_region. 4669 RegionNode* slow_region = new RegionNode(1); 4670 record_for_igvn(slow_region); 4671 if (!stopped()) { 4672 // It's an instance (we did array above). Make the slow-path tests. 4673 // If this is a virtual call, we generate a funny guard. We grab 4674 // the vtable entry corresponding to clone() from the target object. 4675 // If the target method which we are calling happens to be the 4676 // Object clone() method, we pass the guard. We do not need this 4677 // guard for non-virtual calls; the caller is known to be the native 4678 // Object clone(). 4679 if (is_virtual) { 4680 generate_virtual_guard(obj_klass, slow_region); 4681 } 4682 4683 // The object must be easily cloneable and must not have a finalizer. 4684 // Both of these conditions may be checked in a single test. 4685 // We could optimize the test further, but we don't care. 4686 generate_access_flags_guard(obj_klass, 4687 // Test both conditions: 4688 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4689 // Must be cloneable but not finalizer: 4690 JVM_ACC_IS_CLONEABLE_FAST, 4691 slow_region); 4692 } 4693 4694 if (!stopped()) { 4695 // It's an instance, and it passed the slow-path tests. 4696 PreserveJVMState pjvms(this); 4697 Node* obj_size = NULL; 4698 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4699 // is reexecuted if deoptimization occurs and there could be problems when merging 4700 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4701 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4702 4703 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4704 4705 // Present the results of the slow call. 4706 result_reg->init_req(_instance_path, control()); 4707 result_val->init_req(_instance_path, alloc_obj); 4708 result_i_o ->set_req(_instance_path, i_o()); 4709 result_mem ->set_req(_instance_path, reset_memory()); 4710 } 4711 4712 // Generate code for the slow case. We make a call to clone(). 4713 set_control(_gvn.transform(slow_region)); 4714 if (!stopped()) { 4715 PreserveJVMState pjvms(this); 4716 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4717 Node* slow_result = set_results_for_java_call(slow_call); 4718 // this->control() comes from set_results_for_java_call 4719 result_reg->init_req(_slow_path, control()); 4720 result_val->init_req(_slow_path, slow_result); 4721 result_i_o ->set_req(_slow_path, i_o()); 4722 result_mem ->set_req(_slow_path, reset_memory()); 4723 } 4724 4725 // Return the combined state. 4726 set_control( _gvn.transform(result_reg)); 4727 set_i_o( _gvn.transform(result_i_o)); 4728 set_all_memory( _gvn.transform(result_mem)); 4729 } // original reexecute is set back here 4730 4731 set_result(_gvn.transform(result_val)); 4732 return true; 4733 } 4734 4735 // If we have a tighly coupled allocation, the arraycopy may take care 4736 // of the array initialization. If one of the guards we insert between 4737 // the allocation and the arraycopy causes a deoptimization, an 4738 // unitialized array will escape the compiled method. To prevent that 4739 // we set the JVM state for uncommon traps between the allocation and 4740 // the arraycopy to the state before the allocation so, in case of 4741 // deoptimization, we'll reexecute the allocation and the 4742 // initialization. 4743 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4744 if (alloc != NULL) { 4745 ciMethod* trap_method = alloc->jvms()->method(); 4746 int trap_bci = alloc->jvms()->bci(); 4747 4748 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4749 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4750 // Make sure there's no store between the allocation and the 4751 // arraycopy otherwise visible side effects could be rexecuted 4752 // in case of deoptimization and cause incorrect execution. 4753 bool no_interfering_store = true; 4754 Node* mem = alloc->in(TypeFunc::Memory); 4755 if (mem->is_MergeMem()) { 4756 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4757 Node* n = mms.memory(); 4758 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4759 assert(n->is_Store(), "what else?"); 4760 no_interfering_store = false; 4761 break; 4762 } 4763 } 4764 } else { 4765 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4766 Node* n = mms.memory(); 4767 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4768 assert(n->is_Store(), "what else?"); 4769 no_interfering_store = false; 4770 break; 4771 } 4772 } 4773 } 4774 4775 if (no_interfering_store) { 4776 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4777 uint size = alloc->req(); 4778 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4779 old_jvms->set_map(sfpt); 4780 for (uint i = 0; i < size; i++) { 4781 sfpt->init_req(i, alloc->in(i)); 4782 } 4783 // re-push array length for deoptimization 4784 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4785 old_jvms->set_sp(old_jvms->sp()+1); 4786 old_jvms->set_monoff(old_jvms->monoff()+1); 4787 old_jvms->set_scloff(old_jvms->scloff()+1); 4788 old_jvms->set_endoff(old_jvms->endoff()+1); 4789 old_jvms->set_should_reexecute(true); 4790 4791 sfpt->set_i_o(map()->i_o()); 4792 sfpt->set_memory(map()->memory()); 4793 sfpt->set_control(map()->control()); 4794 4795 JVMState* saved_jvms = jvms(); 4796 saved_reexecute_sp = _reexecute_sp; 4797 4798 set_jvms(sfpt->jvms()); 4799 _reexecute_sp = jvms()->sp(); 4800 4801 return saved_jvms; 4802 } 4803 } 4804 } 4805 return NULL; 4806 } 4807 4808 // In case of a deoptimization, we restart execution at the 4809 // allocation, allocating a new array. We would leave an uninitialized 4810 // array in the heap that GCs wouldn't expect. Move the allocation 4811 // after the traps so we don't allocate the array if we 4812 // deoptimize. This is possible because tightly_coupled_allocation() 4813 // guarantees there's no observer of the allocated array at this point 4814 // and the control flow is simple enough. 4815 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4816 if (saved_jvms != NULL && !stopped()) { 4817 assert(alloc != NULL, "only with a tightly coupled allocation"); 4818 // restore JVM state to the state at the arraycopy 4819 saved_jvms->map()->set_control(map()->control()); 4820 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4821 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4822 // If we've improved the types of some nodes (null check) while 4823 // emitting the guards, propagate them to the current state 4824 map()->replaced_nodes().apply(saved_jvms->map()); 4825 set_jvms(saved_jvms); 4826 _reexecute_sp = saved_reexecute_sp; 4827 4828 // Remove the allocation from above the guards 4829 CallProjections callprojs; 4830 alloc->extract_projections(&callprojs, true); 4831 InitializeNode* init = alloc->initialization(); 4832 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4833 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4834 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4835 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4836 4837 // move the allocation here (after the guards) 4838 _gvn.hash_delete(alloc); 4839 alloc->set_req(TypeFunc::Control, control()); 4840 alloc->set_req(TypeFunc::I_O, i_o()); 4841 Node *mem = reset_memory(); 4842 set_all_memory(mem); 4843 alloc->set_req(TypeFunc::Memory, mem); 4844 set_control(init->proj_out(TypeFunc::Control)); 4845 set_i_o(callprojs.fallthrough_ioproj); 4846 4847 // Update memory as done in GraphKit::set_output_for_allocation() 4848 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4849 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4850 if (ary_type->isa_aryptr() && length_type != NULL) { 4851 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4852 } 4853 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4854 int elemidx = C->get_alias_index(telemref); 4855 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4856 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4857 4858 Node* allocx = _gvn.transform(alloc); 4859 assert(allocx == alloc, "where has the allocation gone?"); 4860 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4861 4862 _gvn.hash_delete(dest); 4863 dest->set_req(0, control()); 4864 Node* destx = _gvn.transform(dest); 4865 assert(destx == dest, "where has the allocation result gone?"); 4866 } 4867 } 4868 4869 4870 //------------------------------inline_arraycopy----------------------- 4871 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4872 // Object dest, int destPos, 4873 // int length); 4874 bool LibraryCallKit::inline_arraycopy() { 4875 // Get the arguments. 4876 Node* src = argument(0); // type: oop 4877 Node* src_offset = argument(1); // type: int 4878 Node* dest = argument(2); // type: oop 4879 Node* dest_offset = argument(3); // type: int 4880 Node* length = argument(4); // type: int 4881 4882 4883 // Check for allocation before we add nodes that would confuse 4884 // tightly_coupled_allocation() 4885 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4886 4887 int saved_reexecute_sp = -1; 4888 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4889 // See arraycopy_restore_alloc_state() comment 4890 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4891 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4892 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4893 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4894 4895 // The following tests must be performed 4896 // (1) src and dest are arrays. 4897 // (2) src and dest arrays must have elements of the same BasicType 4898 // (3) src and dest must not be null. 4899 // (4) src_offset must not be negative. 4900 // (5) dest_offset must not be negative. 4901 // (6) length must not be negative. 4902 // (7) src_offset + length must not exceed length of src. 4903 // (8) dest_offset + length must not exceed length of dest. 4904 // (9) each element of an oop array must be assignable 4905 4906 // (3) src and dest must not be null. 4907 // always do this here because we need the JVM state for uncommon traps 4908 Node* null_ctl = top(); 4909 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4910 assert(null_ctl->is_top(), "no null control here"); 4911 dest = null_check(dest, T_ARRAY); 4912 4913 if (!can_emit_guards) { 4914 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4915 // guards but the arraycopy node could still take advantage of a 4916 // tightly allocated allocation. tightly_coupled_allocation() is 4917 // called again to make sure it takes the null check above into 4918 // account: the null check is mandatory and if it caused an 4919 // uncommon trap to be emitted then the allocation can't be 4920 // considered tightly coupled in this context. 4921 alloc = tightly_coupled_allocation(dest, NULL); 4922 } 4923 4924 bool validated = false; 4925 4926 const Type* src_type = _gvn.type(src); 4927 const Type* dest_type = _gvn.type(dest); 4928 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4929 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4930 4931 // Do we have the type of src? 4932 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4933 // Do we have the type of dest? 4934 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4935 // Is the type for src from speculation? 4936 bool src_spec = false; 4937 // Is the type for dest from speculation? 4938 bool dest_spec = false; 4939 4940 if ((!has_src || !has_dest) && can_emit_guards) { 4941 // We don't have sufficient type information, let's see if 4942 // speculative types can help. We need to have types for both src 4943 // and dest so that it pays off. 4944 4945 // Do we already have or could we have type information for src 4946 bool could_have_src = has_src; 4947 // Do we already have or could we have type information for dest 4948 bool could_have_dest = has_dest; 4949 4950 ciKlass* src_k = NULL; 4951 if (!has_src) { 4952 src_k = src_type->speculative_type_not_null(); 4953 if (src_k != NULL && src_k->is_array_klass()) { 4954 could_have_src = true; 4955 } 4956 } 4957 4958 ciKlass* dest_k = NULL; 4959 if (!has_dest) { 4960 dest_k = dest_type->speculative_type_not_null(); 4961 if (dest_k != NULL && dest_k->is_array_klass()) { 4962 could_have_dest = true; 4963 } 4964 } 4965 4966 if (could_have_src && could_have_dest) { 4967 // This is going to pay off so emit the required guards 4968 if (!has_src) { 4969 src = maybe_cast_profiled_obj(src, src_k, true); 4970 src_type = _gvn.type(src); 4971 top_src = src_type->isa_aryptr(); 4972 has_src = (top_src != NULL && top_src->klass() != NULL); 4973 src_spec = true; 4974 } 4975 if (!has_dest) { 4976 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4977 dest_type = _gvn.type(dest); 4978 top_dest = dest_type->isa_aryptr(); 4979 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4980 dest_spec = true; 4981 } 4982 } 4983 } 4984 4985 if (has_src && has_dest && can_emit_guards) { 4986 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4987 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4988 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4989 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4990 4991 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4992 // If both arrays are object arrays then having the exact types 4993 // for both will remove the need for a subtype check at runtime 4994 // before the call and may make it possible to pick a faster copy 4995 // routine (without a subtype check on every element) 4996 // Do we have the exact type of src? 4997 bool could_have_src = src_spec; 4998 // Do we have the exact type of dest? 4999 bool could_have_dest = dest_spec; 5000 ciKlass* src_k = top_src->klass(); 5001 ciKlass* dest_k = top_dest->klass(); 5002 if (!src_spec) { 5003 src_k = src_type->speculative_type_not_null(); 5004 if (src_k != NULL && src_k->is_array_klass()) { 5005 could_have_src = true; 5006 } 5007 } 5008 if (!dest_spec) { 5009 dest_k = dest_type->speculative_type_not_null(); 5010 if (dest_k != NULL && dest_k->is_array_klass()) { 5011 could_have_dest = true; 5012 } 5013 } 5014 if (could_have_src && could_have_dest) { 5015 // If we can have both exact types, emit the missing guards 5016 if (could_have_src && !src_spec) { 5017 src = maybe_cast_profiled_obj(src, src_k, true); 5018 } 5019 if (could_have_dest && !dest_spec) { 5020 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5021 } 5022 } 5023 } 5024 } 5025 5026 ciMethod* trap_method = method(); 5027 int trap_bci = bci(); 5028 if (saved_jvms != NULL) { 5029 trap_method = alloc->jvms()->method(); 5030 trap_bci = alloc->jvms()->bci(); 5031 } 5032 5033 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5034 can_emit_guards && 5035 !src->is_top() && !dest->is_top()) { 5036 // validate arguments: enables transformation the ArrayCopyNode 5037 validated = true; 5038 5039 RegionNode* slow_region = new RegionNode(1); 5040 record_for_igvn(slow_region); 5041 5042 // (1) src and dest are arrays. 5043 generate_non_array_guard(load_object_klass(src), slow_region); 5044 generate_non_array_guard(load_object_klass(dest), slow_region); 5045 5046 // (2) src and dest arrays must have elements of the same BasicType 5047 // done at macro expansion or at Ideal transformation time 5048 5049 // (4) src_offset must not be negative. 5050 generate_negative_guard(src_offset, slow_region); 5051 5052 // (5) dest_offset must not be negative. 5053 generate_negative_guard(dest_offset, slow_region); 5054 5055 // (7) src_offset + length must not exceed length of src. 5056 generate_limit_guard(src_offset, length, 5057 load_array_length(src), 5058 slow_region); 5059 5060 // (8) dest_offset + length must not exceed length of dest. 5061 generate_limit_guard(dest_offset, length, 5062 load_array_length(dest), 5063 slow_region); 5064 5065 // (9) each element of an oop array must be assignable 5066 Node* src_klass = load_object_klass(src); 5067 Node* dest_klass = load_object_klass(dest); 5068 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5069 5070 if (not_subtype_ctrl != top()) { 5071 PreserveJVMState pjvms(this); 5072 set_control(not_subtype_ctrl); 5073 uncommon_trap(Deoptimization::Reason_intrinsic, 5074 Deoptimization::Action_make_not_entrant); 5075 assert(stopped(), "Should be stopped"); 5076 } 5077 { 5078 PreserveJVMState pjvms(this); 5079 set_control(_gvn.transform(slow_region)); 5080 uncommon_trap(Deoptimization::Reason_intrinsic, 5081 Deoptimization::Action_make_not_entrant); 5082 assert(stopped(), "Should be stopped"); 5083 } 5084 } 5085 5086 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5087 5088 if (stopped()) { 5089 return true; 5090 } 5091 5092 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5093 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5094 // so the compiler has a chance to eliminate them: during macro expansion, 5095 // we have to set their control (CastPP nodes are eliminated). 5096 load_object_klass(src), load_object_klass(dest), 5097 load_array_length(src), load_array_length(dest)); 5098 5099 ac->set_arraycopy(validated); 5100 5101 Node* n = _gvn.transform(ac); 5102 if (n == ac) { 5103 ac->connect_outputs(this); 5104 } else { 5105 assert(validated, "shouldn't transform if all arguments not validated"); 5106 set_all_memory(n); 5107 } 5108 5109 return true; 5110 } 5111 5112 5113 // Helper function which determines if an arraycopy immediately follows 5114 // an allocation, with no intervening tests or other escapes for the object. 5115 AllocateArrayNode* 5116 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5117 RegionNode* slow_region) { 5118 if (stopped()) return NULL; // no fast path 5119 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5120 5121 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5122 if (alloc == NULL) return NULL; 5123 5124 Node* rawmem = memory(Compile::AliasIdxRaw); 5125 // Is the allocation's memory state untouched? 5126 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5127 // Bail out if there have been raw-memory effects since the allocation. 5128 // (Example: There might have been a call or safepoint.) 5129 return NULL; 5130 } 5131 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5132 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5133 return NULL; 5134 } 5135 5136 // There must be no unexpected observers of this allocation. 5137 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5138 Node* obs = ptr->fast_out(i); 5139 if (obs != this->map()) { 5140 return NULL; 5141 } 5142 } 5143 5144 // This arraycopy must unconditionally follow the allocation of the ptr. 5145 Node* alloc_ctl = ptr->in(0); 5146 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5147 5148 Node* ctl = control(); 5149 while (ctl != alloc_ctl) { 5150 // There may be guards which feed into the slow_region. 5151 // Any other control flow means that we might not get a chance 5152 // to finish initializing the allocated object. 5153 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5154 IfNode* iff = ctl->in(0)->as_If(); 5155 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5156 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5157 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5158 ctl = iff->in(0); // This test feeds the known slow_region. 5159 continue; 5160 } 5161 // One more try: Various low-level checks bottom out in 5162 // uncommon traps. If the debug-info of the trap omits 5163 // any reference to the allocation, as we've already 5164 // observed, then there can be no objection to the trap. 5165 bool found_trap = false; 5166 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5167 Node* obs = not_ctl->fast_out(j); 5168 if (obs->in(0) == not_ctl && obs->is_Call() && 5169 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5170 found_trap = true; break; 5171 } 5172 } 5173 if (found_trap) { 5174 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5175 continue; 5176 } 5177 } 5178 return NULL; 5179 } 5180 5181 // If we get this far, we have an allocation which immediately 5182 // precedes the arraycopy, and we can take over zeroing the new object. 5183 // The arraycopy will finish the initialization, and provide 5184 // a new control state to which we will anchor the destination pointer. 5185 5186 return alloc; 5187 } 5188 5189 //-------------inline_encodeISOArray----------------------------------- 5190 // encode char[] to byte[] in ISO_8859_1 5191 bool LibraryCallKit::inline_encodeISOArray() { 5192 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5193 // no receiver since it is static method 5194 Node *src = argument(0); 5195 Node *src_offset = argument(1); 5196 Node *dst = argument(2); 5197 Node *dst_offset = argument(3); 5198 Node *length = argument(4); 5199 5200 const Type* src_type = src->Value(&_gvn); 5201 const Type* dst_type = dst->Value(&_gvn); 5202 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5203 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5204 if (top_src == NULL || top_src->klass() == NULL || 5205 top_dest == NULL || top_dest->klass() == NULL) { 5206 // failed array check 5207 return false; 5208 } 5209 5210 // Figure out the size and type of the elements we will be copying. 5211 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5212 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5213 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5214 return false; 5215 } 5216 5217 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5218 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5219 // 'src_start' points to src array + scaled offset 5220 // 'dst_start' points to dst array + scaled offset 5221 5222 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5223 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5224 enc = _gvn.transform(enc); 5225 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5226 set_memory(res_mem, mtype); 5227 set_result(enc); 5228 return true; 5229 } 5230 5231 //-------------inline_multiplyToLen----------------------------------- 5232 bool LibraryCallKit::inline_multiplyToLen() { 5233 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5234 5235 address stubAddr = StubRoutines::multiplyToLen(); 5236 if (stubAddr == NULL) { 5237 return false; // Intrinsic's stub is not implemented on this platform 5238 } 5239 const char* stubName = "multiplyToLen"; 5240 5241 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5242 5243 // no receiver because it is a static method 5244 Node* x = argument(0); 5245 Node* xlen = argument(1); 5246 Node* y = argument(2); 5247 Node* ylen = argument(3); 5248 Node* z = argument(4); 5249 5250 const Type* x_type = x->Value(&_gvn); 5251 const Type* y_type = y->Value(&_gvn); 5252 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5253 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5254 if (top_x == NULL || top_x->klass() == NULL || 5255 top_y == NULL || top_y->klass() == NULL) { 5256 // failed array check 5257 return false; 5258 } 5259 5260 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5261 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5262 if (x_elem != T_INT || y_elem != T_INT) { 5263 return false; 5264 } 5265 5266 // Set the original stack and the reexecute bit for the interpreter to reexecute 5267 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5268 // on the return from z array allocation in runtime. 5269 { PreserveReexecuteState preexecs(this); 5270 jvms()->set_should_reexecute(true); 5271 5272 Node* x_start = array_element_address(x, intcon(0), x_elem); 5273 Node* y_start = array_element_address(y, intcon(0), y_elem); 5274 // 'x_start' points to x array + scaled xlen 5275 // 'y_start' points to y array + scaled ylen 5276 5277 // Allocate the result array 5278 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5279 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5280 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5281 5282 IdealKit ideal(this); 5283 5284 #define __ ideal. 5285 Node* one = __ ConI(1); 5286 Node* zero = __ ConI(0); 5287 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5288 __ set(need_alloc, zero); 5289 __ set(z_alloc, z); 5290 __ if_then(z, BoolTest::eq, null()); { 5291 __ increment (need_alloc, one); 5292 } __ else_(); { 5293 // Update graphKit memory and control from IdealKit. 5294 sync_kit(ideal); 5295 Node* zlen_arg = load_array_length(z); 5296 // Update IdealKit memory and control from graphKit. 5297 __ sync_kit(this); 5298 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5299 __ increment (need_alloc, one); 5300 } __ end_if(); 5301 } __ end_if(); 5302 5303 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5304 // Update graphKit memory and control from IdealKit. 5305 sync_kit(ideal); 5306 Node * narr = new_array(klass_node, zlen, 1); 5307 // Update IdealKit memory and control from graphKit. 5308 __ sync_kit(this); 5309 __ set(z_alloc, narr); 5310 } __ end_if(); 5311 5312 sync_kit(ideal); 5313 z = __ value(z_alloc); 5314 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5315 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5316 // Final sync IdealKit and GraphKit. 5317 final_sync(ideal); 5318 #undef __ 5319 5320 Node* z_start = array_element_address(z, intcon(0), T_INT); 5321 5322 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5323 OptoRuntime::multiplyToLen_Type(), 5324 stubAddr, stubName, TypePtr::BOTTOM, 5325 x_start, xlen, y_start, ylen, z_start, zlen); 5326 } // original reexecute is set back here 5327 5328 C->set_has_split_ifs(true); // Has chance for split-if optimization 5329 set_result(z); 5330 return true; 5331 } 5332 5333 //-------------inline_squareToLen------------------------------------ 5334 bool LibraryCallKit::inline_squareToLen() { 5335 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5336 5337 address stubAddr = StubRoutines::squareToLen(); 5338 if (stubAddr == NULL) { 5339 return false; // Intrinsic's stub is not implemented on this platform 5340 } 5341 const char* stubName = "squareToLen"; 5342 5343 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5344 5345 Node* x = argument(0); 5346 Node* len = argument(1); 5347 Node* z = argument(2); 5348 Node* zlen = argument(3); 5349 5350 const Type* x_type = x->Value(&_gvn); 5351 const Type* z_type = z->Value(&_gvn); 5352 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5353 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5354 if (top_x == NULL || top_x->klass() == NULL || 5355 top_z == NULL || top_z->klass() == NULL) { 5356 // failed array check 5357 return false; 5358 } 5359 5360 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5361 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5362 if (x_elem != T_INT || z_elem != T_INT) { 5363 return false; 5364 } 5365 5366 5367 Node* x_start = array_element_address(x, intcon(0), x_elem); 5368 Node* z_start = array_element_address(z, intcon(0), z_elem); 5369 5370 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5371 OptoRuntime::squareToLen_Type(), 5372 stubAddr, stubName, TypePtr::BOTTOM, 5373 x_start, len, z_start, zlen); 5374 5375 set_result(z); 5376 return true; 5377 } 5378 5379 //-------------inline_mulAdd------------------------------------------ 5380 bool LibraryCallKit::inline_mulAdd() { 5381 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5382 5383 address stubAddr = StubRoutines::mulAdd(); 5384 if (stubAddr == NULL) { 5385 return false; // Intrinsic's stub is not implemented on this platform 5386 } 5387 const char* stubName = "mulAdd"; 5388 5389 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5390 5391 Node* out = argument(0); 5392 Node* in = argument(1); 5393 Node* offset = argument(2); 5394 Node* len = argument(3); 5395 Node* k = argument(4); 5396 5397 const Type* out_type = out->Value(&_gvn); 5398 const Type* in_type = in->Value(&_gvn); 5399 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5400 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5401 if (top_out == NULL || top_out->klass() == NULL || 5402 top_in == NULL || top_in->klass() == NULL) { 5403 // failed array check 5404 return false; 5405 } 5406 5407 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5408 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5409 if (out_elem != T_INT || in_elem != T_INT) { 5410 return false; 5411 } 5412 5413 Node* outlen = load_array_length(out); 5414 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5415 Node* out_start = array_element_address(out, intcon(0), out_elem); 5416 Node* in_start = array_element_address(in, intcon(0), in_elem); 5417 5418 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5419 OptoRuntime::mulAdd_Type(), 5420 stubAddr, stubName, TypePtr::BOTTOM, 5421 out_start,in_start, new_offset, len, k); 5422 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5423 set_result(result); 5424 return true; 5425 } 5426 5427 //-------------inline_montgomeryMultiply----------------------------------- 5428 bool LibraryCallKit::inline_montgomeryMultiply() { 5429 address stubAddr = StubRoutines::montgomeryMultiply(); 5430 if (stubAddr == NULL) { 5431 return false; // Intrinsic's stub is not implemented on this platform 5432 } 5433 5434 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5435 const char* stubName = "montgomery_square"; 5436 5437 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5438 5439 Node* a = argument(0); 5440 Node* b = argument(1); 5441 Node* n = argument(2); 5442 Node* len = argument(3); 5443 Node* inv = argument(4); 5444 Node* m = argument(6); 5445 5446 const Type* a_type = a->Value(&_gvn); 5447 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5448 const Type* b_type = b->Value(&_gvn); 5449 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5450 const Type* n_type = a->Value(&_gvn); 5451 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5452 const Type* m_type = a->Value(&_gvn); 5453 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5454 if (top_a == NULL || top_a->klass() == NULL || 5455 top_b == NULL || top_b->klass() == NULL || 5456 top_n == NULL || top_n->klass() == NULL || 5457 top_m == NULL || top_m->klass() == NULL) { 5458 // failed array check 5459 return false; 5460 } 5461 5462 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5463 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5464 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5465 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5466 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5467 return false; 5468 } 5469 5470 // Make the call 5471 { 5472 Node* a_start = array_element_address(a, intcon(0), a_elem); 5473 Node* b_start = array_element_address(b, intcon(0), b_elem); 5474 Node* n_start = array_element_address(n, intcon(0), n_elem); 5475 Node* m_start = array_element_address(m, intcon(0), m_elem); 5476 5477 Node* call = make_runtime_call(RC_LEAF, 5478 OptoRuntime::montgomeryMultiply_Type(), 5479 stubAddr, stubName, TypePtr::BOTTOM, 5480 a_start, b_start, n_start, len, inv, top(), 5481 m_start); 5482 set_result(m); 5483 } 5484 5485 return true; 5486 } 5487 5488 bool LibraryCallKit::inline_montgomerySquare() { 5489 address stubAddr = StubRoutines::montgomerySquare(); 5490 if (stubAddr == NULL) { 5491 return false; // Intrinsic's stub is not implemented on this platform 5492 } 5493 5494 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5495 const char* stubName = "montgomery_square"; 5496 5497 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5498 5499 Node* a = argument(0); 5500 Node* n = argument(1); 5501 Node* len = argument(2); 5502 Node* inv = argument(3); 5503 Node* m = argument(5); 5504 5505 const Type* a_type = a->Value(&_gvn); 5506 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5507 const Type* n_type = a->Value(&_gvn); 5508 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5509 const Type* m_type = a->Value(&_gvn); 5510 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5511 if (top_a == NULL || top_a->klass() == NULL || 5512 top_n == NULL || top_n->klass() == NULL || 5513 top_m == NULL || top_m->klass() == NULL) { 5514 // failed array check 5515 return false; 5516 } 5517 5518 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5519 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5520 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5521 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5522 return false; 5523 } 5524 5525 // Make the call 5526 { 5527 Node* a_start = array_element_address(a, intcon(0), a_elem); 5528 Node* n_start = array_element_address(n, intcon(0), n_elem); 5529 Node* m_start = array_element_address(m, intcon(0), m_elem); 5530 5531 Node* call = make_runtime_call(RC_LEAF, 5532 OptoRuntime::montgomerySquare_Type(), 5533 stubAddr, stubName, TypePtr::BOTTOM, 5534 a_start, n_start, len, inv, top(), 5535 m_start); 5536 set_result(m); 5537 } 5538 5539 return true; 5540 } 5541 5542 //-------------inline_vectorizedMismatch------------------------------ 5543 bool LibraryCallKit::inline_vectorizedMismatch() { 5544 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5545 5546 address stubAddr = StubRoutines::vectorizedMismatch(); 5547 if (stubAddr == NULL) { 5548 return false; // Intrinsic's stub is not implemented on this platform 5549 } 5550 const char* stubName = "vectorizedMismatch"; 5551 int size_l = callee()->signature()->size(); 5552 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5553 5554 Node* obja = argument(0); 5555 Node* aoffset = argument(1); 5556 Node* objb = argument(3); 5557 Node* boffset = argument(4); 5558 Node* length = argument(6); 5559 Node* scale = argument(7); 5560 5561 const Type* a_type = obja->Value(&_gvn); 5562 const Type* b_type = objb->Value(&_gvn); 5563 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5564 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5565 if (top_a == NULL || top_a->klass() == NULL || 5566 top_b == NULL || top_b->klass() == NULL) { 5567 // failed array check 5568 return false; 5569 } 5570 5571 Node* call; 5572 jvms()->set_should_reexecute(true); 5573 5574 Node* obja_adr = make_unsafe_address(obja, aoffset); 5575 Node* objb_adr = make_unsafe_address(objb, boffset); 5576 5577 call = make_runtime_call(RC_LEAF, 5578 OptoRuntime::vectorizedMismatch_Type(), 5579 stubAddr, stubName, TypePtr::BOTTOM, 5580 obja_adr, objb_adr, length, scale); 5581 5582 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5583 set_result(result); 5584 return true; 5585 } 5586 5587 /** 5588 * Calculate CRC32 for byte. 5589 * int java.util.zip.CRC32.update(int crc, int b) 5590 */ 5591 bool LibraryCallKit::inline_updateCRC32() { 5592 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5593 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5594 // no receiver since it is static method 5595 Node* crc = argument(0); // type: int 5596 Node* b = argument(1); // type: int 5597 5598 /* 5599 * int c = ~ crc; 5600 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5601 * b = b ^ (c >>> 8); 5602 * crc = ~b; 5603 */ 5604 5605 Node* M1 = intcon(-1); 5606 crc = _gvn.transform(new XorINode(crc, M1)); 5607 Node* result = _gvn.transform(new XorINode(crc, b)); 5608 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5609 5610 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5611 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5612 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5613 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5614 5615 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5616 result = _gvn.transform(new XorINode(crc, result)); 5617 result = _gvn.transform(new XorINode(result, M1)); 5618 set_result(result); 5619 return true; 5620 } 5621 5622 /** 5623 * Calculate CRC32 for byte[] array. 5624 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5625 */ 5626 bool LibraryCallKit::inline_updateBytesCRC32() { 5627 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5628 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5629 // no receiver since it is static method 5630 Node* crc = argument(0); // type: int 5631 Node* src = argument(1); // type: oop 5632 Node* offset = argument(2); // type: int 5633 Node* length = argument(3); // type: int 5634 5635 const Type* src_type = src->Value(&_gvn); 5636 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5637 if (top_src == NULL || top_src->klass() == NULL) { 5638 // failed array check 5639 return false; 5640 } 5641 5642 // Figure out the size and type of the elements we will be copying. 5643 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5644 if (src_elem != T_BYTE) { 5645 return false; 5646 } 5647 5648 // 'src_start' points to src array + scaled offset 5649 Node* src_start = array_element_address(src, offset, src_elem); 5650 5651 // We assume that range check is done by caller. 5652 // TODO: generate range check (offset+length < src.length) in debug VM. 5653 5654 // Call the stub. 5655 address stubAddr = StubRoutines::updateBytesCRC32(); 5656 const char *stubName = "updateBytesCRC32"; 5657 5658 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5659 stubAddr, stubName, TypePtr::BOTTOM, 5660 crc, src_start, length); 5661 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5662 set_result(result); 5663 return true; 5664 } 5665 5666 /** 5667 * Calculate CRC32 for ByteBuffer. 5668 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5669 */ 5670 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5671 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5672 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5673 // no receiver since it is static method 5674 Node* crc = argument(0); // type: int 5675 Node* src = argument(1); // type: long 5676 Node* offset = argument(3); // type: int 5677 Node* length = argument(4); // type: int 5678 5679 src = ConvL2X(src); // adjust Java long to machine word 5680 Node* base = _gvn.transform(new CastX2PNode(src)); 5681 offset = ConvI2X(offset); 5682 5683 // 'src_start' points to src array + scaled offset 5684 Node* src_start = basic_plus_adr(top(), base, offset); 5685 5686 // Call the stub. 5687 address stubAddr = StubRoutines::updateBytesCRC32(); 5688 const char *stubName = "updateBytesCRC32"; 5689 5690 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5691 stubAddr, stubName, TypePtr::BOTTOM, 5692 crc, src_start, length); 5693 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5694 set_result(result); 5695 return true; 5696 } 5697 5698 //------------------------------get_table_from_crc32c_class----------------------- 5699 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5700 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5701 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5702 5703 return table; 5704 } 5705 5706 //------------------------------inline_updateBytesCRC32C----------------------- 5707 // 5708 // Calculate CRC32C for byte[] array. 5709 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5710 // 5711 bool LibraryCallKit::inline_updateBytesCRC32C() { 5712 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5713 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5714 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5715 // no receiver since it is a static method 5716 Node* crc = argument(0); // type: int 5717 Node* src = argument(1); // type: oop 5718 Node* offset = argument(2); // type: int 5719 Node* end = argument(3); // type: int 5720 5721 Node* length = _gvn.transform(new SubINode(end, offset)); 5722 5723 const Type* src_type = src->Value(&_gvn); 5724 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5725 if (top_src == NULL || top_src->klass() == NULL) { 5726 // failed array check 5727 return false; 5728 } 5729 5730 // Figure out the size and type of the elements we will be copying. 5731 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5732 if (src_elem != T_BYTE) { 5733 return false; 5734 } 5735 5736 // 'src_start' points to src array + scaled offset 5737 Node* src_start = array_element_address(src, offset, src_elem); 5738 5739 // static final int[] byteTable in class CRC32C 5740 Node* table = get_table_from_crc32c_class(callee()->holder()); 5741 Node* table_start = array_element_address(table, intcon(0), T_INT); 5742 5743 // We assume that range check is done by caller. 5744 // TODO: generate range check (offset+length < src.length) in debug VM. 5745 5746 // Call the stub. 5747 address stubAddr = StubRoutines::updateBytesCRC32C(); 5748 const char *stubName = "updateBytesCRC32C"; 5749 5750 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5751 stubAddr, stubName, TypePtr::BOTTOM, 5752 crc, src_start, length, table_start); 5753 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5754 set_result(result); 5755 return true; 5756 } 5757 5758 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5759 // 5760 // Calculate CRC32C for DirectByteBuffer. 5761 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5762 // 5763 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5764 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5765 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5766 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5767 // no receiver since it is a static method 5768 Node* crc = argument(0); // type: int 5769 Node* src = argument(1); // type: long 5770 Node* offset = argument(3); // type: int 5771 Node* end = argument(4); // type: int 5772 5773 Node* length = _gvn.transform(new SubINode(end, offset)); 5774 5775 src = ConvL2X(src); // adjust Java long to machine word 5776 Node* base = _gvn.transform(new CastX2PNode(src)); 5777 offset = ConvI2X(offset); 5778 5779 // 'src_start' points to src array + scaled offset 5780 Node* src_start = basic_plus_adr(top(), base, offset); 5781 5782 // static final int[] byteTable in class CRC32C 5783 Node* table = get_table_from_crc32c_class(callee()->holder()); 5784 Node* table_start = array_element_address(table, intcon(0), T_INT); 5785 5786 // Call the stub. 5787 address stubAddr = StubRoutines::updateBytesCRC32C(); 5788 const char *stubName = "updateBytesCRC32C"; 5789 5790 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5791 stubAddr, stubName, TypePtr::BOTTOM, 5792 crc, src_start, length, table_start); 5793 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5794 set_result(result); 5795 return true; 5796 } 5797 5798 //------------------------------inline_updateBytesAdler32---------------------- 5799 // 5800 // Calculate Adler32 checksum for byte[] array. 5801 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5802 // 5803 bool LibraryCallKit::inline_updateBytesAdler32() { 5804 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5805 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5806 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5807 // no receiver since it is static method 5808 Node* crc = argument(0); // type: int 5809 Node* src = argument(1); // type: oop 5810 Node* offset = argument(2); // type: int 5811 Node* length = argument(3); // type: int 5812 5813 const Type* src_type = src->Value(&_gvn); 5814 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5815 if (top_src == NULL || top_src->klass() == NULL) { 5816 // failed array check 5817 return false; 5818 } 5819 5820 // Figure out the size and type of the elements we will be copying. 5821 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5822 if (src_elem != T_BYTE) { 5823 return false; 5824 } 5825 5826 // 'src_start' points to src array + scaled offset 5827 Node* src_start = array_element_address(src, offset, src_elem); 5828 5829 // We assume that range check is done by caller. 5830 // TODO: generate range check (offset+length < src.length) in debug VM. 5831 5832 // Call the stub. 5833 address stubAddr = StubRoutines::updateBytesAdler32(); 5834 const char *stubName = "updateBytesAdler32"; 5835 5836 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5837 stubAddr, stubName, TypePtr::BOTTOM, 5838 crc, src_start, length); 5839 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5840 set_result(result); 5841 return true; 5842 } 5843 5844 //------------------------------inline_updateByteBufferAdler32--------------- 5845 // 5846 // Calculate Adler32 checksum for DirectByteBuffer. 5847 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5848 // 5849 bool LibraryCallKit::inline_updateByteBufferAdler32() { 5850 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5851 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5852 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5853 // no receiver since it is static method 5854 Node* crc = argument(0); // type: int 5855 Node* src = argument(1); // type: long 5856 Node* offset = argument(3); // type: int 5857 Node* length = argument(4); // type: int 5858 5859 src = ConvL2X(src); // adjust Java long to machine word 5860 Node* base = _gvn.transform(new CastX2PNode(src)); 5861 offset = ConvI2X(offset); 5862 5863 // 'src_start' points to src array + scaled offset 5864 Node* src_start = basic_plus_adr(top(), base, offset); 5865 5866 // Call the stub. 5867 address stubAddr = StubRoutines::updateBytesAdler32(); 5868 const char *stubName = "updateBytesAdler32"; 5869 5870 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5871 stubAddr, stubName, TypePtr::BOTTOM, 5872 crc, src_start, length); 5873 5874 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5875 set_result(result); 5876 return true; 5877 } 5878 5879 //----------------------------inline_reference_get---------------------------- 5880 // public T java.lang.ref.Reference.get(); 5881 bool LibraryCallKit::inline_reference_get() { 5882 const int referent_offset = java_lang_ref_Reference::referent_offset; 5883 guarantee(referent_offset > 0, "should have already been set"); 5884 5885 // Get the argument: 5886 Node* reference_obj = null_check_receiver(); 5887 if (stopped()) return true; 5888 5889 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5890 5891 ciInstanceKlass* klass = env()->Object_klass(); 5892 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5893 5894 Node* no_ctrl = NULL; 5895 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5896 5897 // Use the pre-barrier to record the value in the referent field 5898 pre_barrier(false /* do_load */, 5899 control(), 5900 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5901 result /* pre_val */, 5902 T_OBJECT); 5903 5904 // Add memory barrier to prevent commoning reads from this field 5905 // across safepoint since GC can change its value. 5906 insert_mem_bar(Op_MemBarCPUOrder); 5907 5908 set_result(result); 5909 return true; 5910 } 5911 5912 5913 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5914 bool is_exact=true, bool is_static=false, 5915 ciInstanceKlass * fromKls=NULL) { 5916 if (fromKls == NULL) { 5917 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5918 assert(tinst != NULL, "obj is null"); 5919 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5920 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5921 fromKls = tinst->klass()->as_instance_klass(); 5922 } else { 5923 assert(is_static, "only for static field access"); 5924 } 5925 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5926 ciSymbol::make(fieldTypeString), 5927 is_static); 5928 5929 assert (field != NULL, "undefined field"); 5930 if (field == NULL) return (Node *) NULL; 5931 5932 if (is_static) { 5933 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5934 fromObj = makecon(tip); 5935 } 5936 5937 // Next code copied from Parse::do_get_xxx(): 5938 5939 // Compute address and memory type. 5940 int offset = field->offset_in_bytes(); 5941 bool is_vol = field->is_volatile(); 5942 ciType* field_klass = field->type(); 5943 assert(field_klass->is_loaded(), "should be loaded"); 5944 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5945 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5946 BasicType bt = field->layout_type(); 5947 5948 // Build the resultant type of the load 5949 const Type *type; 5950 if (bt == T_OBJECT) { 5951 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5952 } else { 5953 type = Type::get_const_basic_type(bt); 5954 } 5955 5956 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5957 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5958 } 5959 // Build the load. 5960 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5961 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5962 // If reference is volatile, prevent following memory ops from 5963 // floating up past the volatile read. Also prevents commoning 5964 // another volatile read. 5965 if (is_vol) { 5966 // Memory barrier includes bogus read of value to force load BEFORE membar 5967 insert_mem_bar(Op_MemBarAcquire, loadedField); 5968 } 5969 return loadedField; 5970 } 5971 5972 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5973 bool is_exact = true, bool is_static = false, 5974 ciInstanceKlass * fromKls = NULL) { 5975 if (fromKls == NULL) { 5976 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5977 assert(tinst != NULL, "obj is null"); 5978 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5979 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5980 fromKls = tinst->klass()->as_instance_klass(); 5981 } 5982 else { 5983 assert(is_static, "only for static field access"); 5984 } 5985 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5986 ciSymbol::make(fieldTypeString), 5987 is_static); 5988 5989 assert(field != NULL, "undefined field"); 5990 assert(!field->is_volatile(), "not defined for volatile fields"); 5991 5992 if (is_static) { 5993 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5994 fromObj = makecon(tip); 5995 } 5996 5997 // Next code copied from Parse::do_get_xxx(): 5998 5999 // Compute address and memory type. 6000 int offset = field->offset_in_bytes(); 6001 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 6002 6003 return adr; 6004 } 6005 6006 //------------------------------inline_aescrypt_Block----------------------- 6007 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 6008 address stubAddr = NULL; 6009 const char *stubName; 6010 assert(UseAES, "need AES instruction support"); 6011 6012 switch(id) { 6013 case vmIntrinsics::_aescrypt_encryptBlock: 6014 stubAddr = StubRoutines::aescrypt_encryptBlock(); 6015 stubName = "aescrypt_encryptBlock"; 6016 break; 6017 case vmIntrinsics::_aescrypt_decryptBlock: 6018 stubAddr = StubRoutines::aescrypt_decryptBlock(); 6019 stubName = "aescrypt_decryptBlock"; 6020 break; 6021 } 6022 if (stubAddr == NULL) return false; 6023 6024 Node* aescrypt_object = argument(0); 6025 Node* src = argument(1); 6026 Node* src_offset = argument(2); 6027 Node* dest = argument(3); 6028 Node* dest_offset = argument(4); 6029 6030 // (1) src and dest are arrays. 6031 const Type* src_type = src->Value(&_gvn); 6032 const Type* dest_type = dest->Value(&_gvn); 6033 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6034 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6035 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6036 6037 // for the quick and dirty code we will skip all the checks. 6038 // we are just trying to get the call to be generated. 6039 Node* src_start = src; 6040 Node* dest_start = dest; 6041 if (src_offset != NULL || dest_offset != NULL) { 6042 assert(src_offset != NULL && dest_offset != NULL, ""); 6043 src_start = array_element_address(src, src_offset, T_BYTE); 6044 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6045 } 6046 6047 // now need to get the start of its expanded key array 6048 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6049 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6050 if (k_start == NULL) return false; 6051 6052 if (Matcher::pass_original_key_for_aes()) { 6053 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6054 // compatibility issues between Java key expansion and SPARC crypto instructions 6055 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6056 if (original_k_start == NULL) return false; 6057 6058 // Call the stub. 6059 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6060 stubAddr, stubName, TypePtr::BOTTOM, 6061 src_start, dest_start, k_start, original_k_start); 6062 } else { 6063 // Call the stub. 6064 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6065 stubAddr, stubName, TypePtr::BOTTOM, 6066 src_start, dest_start, k_start); 6067 } 6068 6069 return true; 6070 } 6071 6072 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 6073 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 6074 address stubAddr = NULL; 6075 const char *stubName = NULL; 6076 6077 assert(UseAES, "need AES instruction support"); 6078 6079 switch(id) { 6080 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6081 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6082 stubName = "cipherBlockChaining_encryptAESCrypt"; 6083 break; 6084 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6085 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6086 stubName = "cipherBlockChaining_decryptAESCrypt"; 6087 break; 6088 } 6089 if (stubAddr == NULL) return false; 6090 6091 Node* cipherBlockChaining_object = argument(0); 6092 Node* src = argument(1); 6093 Node* src_offset = argument(2); 6094 Node* len = argument(3); 6095 Node* dest = argument(4); 6096 Node* dest_offset = argument(5); 6097 6098 // (1) src and dest are arrays. 6099 const Type* src_type = src->Value(&_gvn); 6100 const Type* dest_type = dest->Value(&_gvn); 6101 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6102 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6103 assert (top_src != NULL && top_src->klass() != NULL 6104 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6105 6106 // checks are the responsibility of the caller 6107 Node* src_start = src; 6108 Node* dest_start = dest; 6109 if (src_offset != NULL || dest_offset != NULL) { 6110 assert(src_offset != NULL && dest_offset != NULL, ""); 6111 src_start = array_element_address(src, src_offset, T_BYTE); 6112 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6113 } 6114 6115 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6116 // (because of the predicated logic executed earlier). 6117 // so we cast it here safely. 6118 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6119 6120 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6121 if (embeddedCipherObj == NULL) return false; 6122 6123 // cast it to what we know it will be at runtime 6124 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6125 assert(tinst != NULL, "CBC obj is null"); 6126 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6127 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6128 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6129 6130 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6131 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6132 const TypeOopPtr* xtype = aklass->as_instance_type(); 6133 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6134 aescrypt_object = _gvn.transform(aescrypt_object); 6135 6136 // we need to get the start of the aescrypt_object's expanded key array 6137 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6138 if (k_start == NULL) return false; 6139 6140 // similarly, get the start address of the r vector 6141 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6142 if (objRvec == NULL) return false; 6143 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6144 6145 Node* cbcCrypt; 6146 if (Matcher::pass_original_key_for_aes()) { 6147 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6148 // compatibility issues between Java key expansion and SPARC crypto instructions 6149 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6150 if (original_k_start == NULL) return false; 6151 6152 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6153 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6154 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6155 stubAddr, stubName, TypePtr::BOTTOM, 6156 src_start, dest_start, k_start, r_start, len, original_k_start); 6157 } else { 6158 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6159 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6160 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6161 stubAddr, stubName, TypePtr::BOTTOM, 6162 src_start, dest_start, k_start, r_start, len); 6163 } 6164 6165 // return cipher length (int) 6166 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6167 set_result(retvalue); 6168 return true; 6169 } 6170 6171 //------------------------------inline_counterMode_AESCrypt----------------------- 6172 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6173 assert(UseAES, "need AES instruction support"); 6174 if (!UseAESCTRIntrinsics) return false; 6175 6176 address stubAddr = NULL; 6177 const char *stubName = NULL; 6178 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6179 stubAddr = StubRoutines::counterMode_AESCrypt(); 6180 stubName = "counterMode_AESCrypt"; 6181 } 6182 if (stubAddr == NULL) return false; 6183 6184 Node* counterMode_object = argument(0); 6185 Node* src = argument(1); 6186 Node* src_offset = argument(2); 6187 Node* len = argument(3); 6188 Node* dest = argument(4); 6189 Node* dest_offset = argument(5); 6190 6191 // (1) src and dest are arrays. 6192 const Type* src_type = src->Value(&_gvn); 6193 const Type* dest_type = dest->Value(&_gvn); 6194 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6195 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6196 assert(top_src != NULL && top_src->klass() != NULL && 6197 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6198 6199 // checks are the responsibility of the caller 6200 Node* src_start = src; 6201 Node* dest_start = dest; 6202 if (src_offset != NULL || dest_offset != NULL) { 6203 assert(src_offset != NULL && dest_offset != NULL, ""); 6204 src_start = array_element_address(src, src_offset, T_BYTE); 6205 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6206 } 6207 6208 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6209 // (because of the predicated logic executed earlier). 6210 // so we cast it here safely. 6211 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6212 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6213 if (embeddedCipherObj == NULL) return false; 6214 // cast it to what we know it will be at runtime 6215 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6216 assert(tinst != NULL, "CTR obj is null"); 6217 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6218 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6219 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6220 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6221 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6222 const TypeOopPtr* xtype = aklass->as_instance_type(); 6223 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6224 aescrypt_object = _gvn.transform(aescrypt_object); 6225 // we need to get the start of the aescrypt_object's expanded key array 6226 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6227 if (k_start == NULL) return false; 6228 // similarly, get the start address of the r vector 6229 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6230 if (obj_counter == NULL) return false; 6231 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6232 6233 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6234 if (saved_encCounter == NULL) return false; 6235 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6236 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6237 6238 Node* ctrCrypt; 6239 if (Matcher::pass_original_key_for_aes()) { 6240 // no SPARC version for AES/CTR intrinsics now. 6241 return false; 6242 } 6243 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6244 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6245 OptoRuntime::counterMode_aescrypt_Type(), 6246 stubAddr, stubName, TypePtr::BOTTOM, 6247 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6248 6249 // return cipher length (int) 6250 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6251 set_result(retvalue); 6252 return true; 6253 } 6254 6255 //------------------------------get_key_start_from_aescrypt_object----------------------- 6256 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6257 #ifdef PPC64 6258 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6259 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6260 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6261 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6262 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6263 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6264 if (objSessionK == NULL) { 6265 return (Node *) NULL; 6266 } 6267 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6268 #else 6269 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6270 #endif // PPC64 6271 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6272 if (objAESCryptKey == NULL) return (Node *) NULL; 6273 6274 // now have the array, need to get the start address of the K array 6275 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6276 return k_start; 6277 } 6278 6279 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6280 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6281 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6282 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6283 if (objAESCryptKey == NULL) return (Node *) NULL; 6284 6285 // now have the array, need to get the start address of the lastKey array 6286 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6287 return original_k_start; 6288 } 6289 6290 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6291 // Return node representing slow path of predicate check. 6292 // the pseudo code we want to emulate with this predicate is: 6293 // for encryption: 6294 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6295 // for decryption: 6296 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6297 // note cipher==plain is more conservative than the original java code but that's OK 6298 // 6299 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6300 // The receiver was checked for NULL already. 6301 Node* objCBC = argument(0); 6302 6303 // Load embeddedCipher field of CipherBlockChaining object. 6304 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6305 6306 // get AESCrypt klass for instanceOf check 6307 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6308 // will have same classloader as CipherBlockChaining object 6309 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6310 assert(tinst != NULL, "CBCobj is null"); 6311 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6312 6313 // we want to do an instanceof comparison against the AESCrypt class 6314 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6315 if (!klass_AESCrypt->is_loaded()) { 6316 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6317 Node* ctrl = control(); 6318 set_control(top()); // no regular fast path 6319 return ctrl; 6320 } 6321 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6322 6323 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6324 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6325 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6326 6327 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6328 6329 // for encryption, we are done 6330 if (!decrypting) 6331 return instof_false; // even if it is NULL 6332 6333 // for decryption, we need to add a further check to avoid 6334 // taking the intrinsic path when cipher and plain are the same 6335 // see the original java code for why. 6336 RegionNode* region = new RegionNode(3); 6337 region->init_req(1, instof_false); 6338 Node* src = argument(1); 6339 Node* dest = argument(4); 6340 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6341 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6342 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6343 region->init_req(2, src_dest_conjoint); 6344 6345 record_for_igvn(region); 6346 return _gvn.transform(region); 6347 } 6348 6349 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6350 // Return node representing slow path of predicate check. 6351 // the pseudo code we want to emulate with this predicate is: 6352 // for encryption: 6353 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6354 // for decryption: 6355 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6356 // note cipher==plain is more conservative than the original java code but that's OK 6357 // 6358 6359 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6360 // The receiver was checked for NULL already. 6361 Node* objCTR = argument(0); 6362 6363 // Load embeddedCipher field of CipherBlockChaining object. 6364 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6365 6366 // get AESCrypt klass for instanceOf check 6367 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6368 // will have same classloader as CipherBlockChaining object 6369 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6370 assert(tinst != NULL, "CTRobj is null"); 6371 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6372 6373 // we want to do an instanceof comparison against the AESCrypt class 6374 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6375 if (!klass_AESCrypt->is_loaded()) { 6376 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6377 Node* ctrl = control(); 6378 set_control(top()); // no regular fast path 6379 return ctrl; 6380 } 6381 6382 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6383 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6384 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6385 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6386 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6387 6388 return instof_false; // even if it is NULL 6389 } 6390 6391 //------------------------------inline_ghash_processBlocks 6392 bool LibraryCallKit::inline_ghash_processBlocks() { 6393 address stubAddr; 6394 const char *stubName; 6395 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6396 6397 stubAddr = StubRoutines::ghash_processBlocks(); 6398 stubName = "ghash_processBlocks"; 6399 6400 Node* data = argument(0); 6401 Node* offset = argument(1); 6402 Node* len = argument(2); 6403 Node* state = argument(3); 6404 Node* subkeyH = argument(4); 6405 6406 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6407 assert(state_start, "state is NULL"); 6408 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6409 assert(subkeyH_start, "subkeyH is NULL"); 6410 Node* data_start = array_element_address(data, offset, T_BYTE); 6411 assert(data_start, "data is NULL"); 6412 6413 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6414 OptoRuntime::ghash_processBlocks_Type(), 6415 stubAddr, stubName, TypePtr::BOTTOM, 6416 state_start, subkeyH_start, data_start, len); 6417 return true; 6418 } 6419 6420 //------------------------------inline_sha_implCompress----------------------- 6421 // 6422 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6423 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6424 // 6425 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6426 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6427 // 6428 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6429 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6430 // 6431 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6432 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6433 6434 Node* sha_obj = argument(0); 6435 Node* src = argument(1); // type oop 6436 Node* ofs = argument(2); // type int 6437 6438 const Type* src_type = src->Value(&_gvn); 6439 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6440 if (top_src == NULL || top_src->klass() == NULL) { 6441 // failed array check 6442 return false; 6443 } 6444 // Figure out the size and type of the elements we will be copying. 6445 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6446 if (src_elem != T_BYTE) { 6447 return false; 6448 } 6449 // 'src_start' points to src array + offset 6450 Node* src_start = array_element_address(src, ofs, src_elem); 6451 Node* state = NULL; 6452 address stubAddr; 6453 const char *stubName; 6454 6455 switch(id) { 6456 case vmIntrinsics::_sha_implCompress: 6457 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6458 state = get_state_from_sha_object(sha_obj); 6459 stubAddr = StubRoutines::sha1_implCompress(); 6460 stubName = "sha1_implCompress"; 6461 break; 6462 case vmIntrinsics::_sha2_implCompress: 6463 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6464 state = get_state_from_sha_object(sha_obj); 6465 stubAddr = StubRoutines::sha256_implCompress(); 6466 stubName = "sha256_implCompress"; 6467 break; 6468 case vmIntrinsics::_sha5_implCompress: 6469 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6470 state = get_state_from_sha5_object(sha_obj); 6471 stubAddr = StubRoutines::sha512_implCompress(); 6472 stubName = "sha512_implCompress"; 6473 break; 6474 default: 6475 fatal_unexpected_iid(id); 6476 return false; 6477 } 6478 if (state == NULL) return false; 6479 6480 // Call the stub. 6481 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6482 stubAddr, stubName, TypePtr::BOTTOM, 6483 src_start, state); 6484 6485 return true; 6486 } 6487 6488 //------------------------------inline_digestBase_implCompressMB----------------------- 6489 // 6490 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6491 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6492 // 6493 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6494 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6495 "need SHA1/SHA256/SHA512 instruction support"); 6496 assert((uint)predicate < 3, "sanity"); 6497 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6498 6499 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6500 Node* src = argument(1); // byte[] array 6501 Node* ofs = argument(2); // type int 6502 Node* limit = argument(3); // type int 6503 6504 const Type* src_type = src->Value(&_gvn); 6505 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6506 if (top_src == NULL || top_src->klass() == NULL) { 6507 // failed array check 6508 return false; 6509 } 6510 // Figure out the size and type of the elements we will be copying. 6511 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6512 if (src_elem != T_BYTE) { 6513 return false; 6514 } 6515 // 'src_start' points to src array + offset 6516 Node* src_start = array_element_address(src, ofs, src_elem); 6517 6518 const char* klass_SHA_name = NULL; 6519 const char* stub_name = NULL; 6520 address stub_addr = NULL; 6521 bool long_state = false; 6522 6523 switch (predicate) { 6524 case 0: 6525 if (UseSHA1Intrinsics) { 6526 klass_SHA_name = "sun/security/provider/SHA"; 6527 stub_name = "sha1_implCompressMB"; 6528 stub_addr = StubRoutines::sha1_implCompressMB(); 6529 } 6530 break; 6531 case 1: 6532 if (UseSHA256Intrinsics) { 6533 klass_SHA_name = "sun/security/provider/SHA2"; 6534 stub_name = "sha256_implCompressMB"; 6535 stub_addr = StubRoutines::sha256_implCompressMB(); 6536 } 6537 break; 6538 case 2: 6539 if (UseSHA512Intrinsics) { 6540 klass_SHA_name = "sun/security/provider/SHA5"; 6541 stub_name = "sha512_implCompressMB"; 6542 stub_addr = StubRoutines::sha512_implCompressMB(); 6543 long_state = true; 6544 } 6545 break; 6546 default: 6547 fatal("unknown SHA intrinsic predicate: %d", predicate); 6548 } 6549 if (klass_SHA_name != NULL) { 6550 // get DigestBase klass to lookup for SHA klass 6551 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6552 assert(tinst != NULL, "digestBase_obj is not instance???"); 6553 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6554 6555 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6556 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6557 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6558 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6559 } 6560 return false; 6561 } 6562 //------------------------------inline_sha_implCompressMB----------------------- 6563 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6564 bool long_state, address stubAddr, const char *stubName, 6565 Node* src_start, Node* ofs, Node* limit) { 6566 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6567 const TypeOopPtr* xtype = aklass->as_instance_type(); 6568 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6569 sha_obj = _gvn.transform(sha_obj); 6570 6571 Node* state; 6572 if (long_state) { 6573 state = get_state_from_sha5_object(sha_obj); 6574 } else { 6575 state = get_state_from_sha_object(sha_obj); 6576 } 6577 if (state == NULL) return false; 6578 6579 // Call the stub. 6580 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6581 OptoRuntime::digestBase_implCompressMB_Type(), 6582 stubAddr, stubName, TypePtr::BOTTOM, 6583 src_start, state, ofs, limit); 6584 // return ofs (int) 6585 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6586 set_result(result); 6587 6588 return true; 6589 } 6590 6591 //------------------------------get_state_from_sha_object----------------------- 6592 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6593 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6594 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6595 if (sha_state == NULL) return (Node *) NULL; 6596 6597 // now have the array, need to get the start address of the state array 6598 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6599 return state; 6600 } 6601 6602 //------------------------------get_state_from_sha5_object----------------------- 6603 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6604 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6605 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6606 if (sha_state == NULL) return (Node *) NULL; 6607 6608 // now have the array, need to get the start address of the state array 6609 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6610 return state; 6611 } 6612 6613 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6614 // Return node representing slow path of predicate check. 6615 // the pseudo code we want to emulate with this predicate is: 6616 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6617 // 6618 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6619 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6620 "need SHA1/SHA256/SHA512 instruction support"); 6621 assert((uint)predicate < 3, "sanity"); 6622 6623 // The receiver was checked for NULL already. 6624 Node* digestBaseObj = argument(0); 6625 6626 // get DigestBase klass for instanceOf check 6627 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6628 assert(tinst != NULL, "digestBaseObj is null"); 6629 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6630 6631 const char* klass_SHA_name = NULL; 6632 switch (predicate) { 6633 case 0: 6634 if (UseSHA1Intrinsics) { 6635 // we want to do an instanceof comparison against the SHA class 6636 klass_SHA_name = "sun/security/provider/SHA"; 6637 } 6638 break; 6639 case 1: 6640 if (UseSHA256Intrinsics) { 6641 // we want to do an instanceof comparison against the SHA2 class 6642 klass_SHA_name = "sun/security/provider/SHA2"; 6643 } 6644 break; 6645 case 2: 6646 if (UseSHA512Intrinsics) { 6647 // we want to do an instanceof comparison against the SHA5 class 6648 klass_SHA_name = "sun/security/provider/SHA5"; 6649 } 6650 break; 6651 default: 6652 fatal("unknown SHA intrinsic predicate: %d", predicate); 6653 } 6654 6655 ciKlass* klass_SHA = NULL; 6656 if (klass_SHA_name != NULL) { 6657 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6658 } 6659 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6660 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6661 Node* ctrl = control(); 6662 set_control(top()); // no intrinsic path 6663 return ctrl; 6664 } 6665 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6666 6667 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6668 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6669 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6670 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6671 6672 return instof_false; // even if it is NULL 6673 } 6674 6675 bool LibraryCallKit::inline_profileBoolean() { 6676 Node* counts = argument(1); 6677 const TypeAryPtr* ary = NULL; 6678 ciArray* aobj = NULL; 6679 if (counts->is_Con() 6680 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6681 && (aobj = ary->const_oop()->as_array()) != NULL 6682 && (aobj->length() == 2)) { 6683 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6684 jint false_cnt = aobj->element_value(0).as_int(); 6685 jint true_cnt = aobj->element_value(1).as_int(); 6686 6687 if (C->log() != NULL) { 6688 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6689 false_cnt, true_cnt); 6690 } 6691 6692 if (false_cnt + true_cnt == 0) { 6693 // According to profile, never executed. 6694 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6695 Deoptimization::Action_reinterpret); 6696 return true; 6697 } 6698 6699 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6700 // is a number of each value occurrences. 6701 Node* result = argument(0); 6702 if (false_cnt == 0 || true_cnt == 0) { 6703 // According to profile, one value has been never seen. 6704 int expected_val = (false_cnt == 0) ? 1 : 0; 6705 6706 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6707 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6708 6709 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6710 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6711 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6712 6713 { // Slow path: uncommon trap for never seen value and then reexecute 6714 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6715 // the value has been seen at least once. 6716 PreserveJVMState pjvms(this); 6717 PreserveReexecuteState preexecs(this); 6718 jvms()->set_should_reexecute(true); 6719 6720 set_control(slow_path); 6721 set_i_o(i_o()); 6722 6723 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6724 Deoptimization::Action_reinterpret); 6725 } 6726 // The guard for never seen value enables sharpening of the result and 6727 // returning a constant. It allows to eliminate branches on the same value 6728 // later on. 6729 set_control(fast_path); 6730 result = intcon(expected_val); 6731 } 6732 // Stop profiling. 6733 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6734 // By replacing method body with profile data (represented as ProfileBooleanNode 6735 // on IR level) we effectively disable profiling. 6736 // It enables full speed execution once optimized code is generated. 6737 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6738 C->record_for_igvn(profile); 6739 set_result(profile); 6740 return true; 6741 } else { 6742 // Continue profiling. 6743 // Profile data isn't available at the moment. So, execute method's bytecode version. 6744 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6745 // is compiled and counters aren't available since corresponding MethodHandle 6746 // isn't a compile-time constant. 6747 return false; 6748 } 6749 } 6750 6751 bool LibraryCallKit::inline_isCompileConstant() { 6752 Node* n = argument(0); 6753 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6754 return true; 6755 }