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 "oops/objArrayKlass.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/c2compiler.hpp" 35 #include "opto/callGenerator.hpp" 36 #include "opto/castnode.hpp" 37 #include "opto/cfgnode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/countbitsnode.hpp" 40 #include "opto/intrinsicnode.hpp" 41 #include "opto/idealKit.hpp" 42 #include "opto/mathexactnode.hpp" 43 #include "opto/movenode.hpp" 44 #include "opto/mulnode.hpp" 45 #include "opto/narrowptrnode.hpp" 46 #include "opto/opaquenode.hpp" 47 #include "opto/parse.hpp" 48 #include "opto/runtime.hpp" 49 #include "opto/subnode.hpp" 50 #include "prims/nativeLookup.hpp" 51 #include "runtime/sharedRuntime.hpp" 52 #include "trace/traceMacros.hpp" 53 54 class LibraryIntrinsic : public InlineCallGenerator { 55 // Extend the set of intrinsics known to the runtime: 56 public: 57 private: 58 bool _is_virtual; 59 bool _does_virtual_dispatch; 60 int8_t _predicates_count; // Intrinsic is predicated by several conditions 61 int8_t _last_predicate; // Last generated predicate 62 vmIntrinsics::ID _intrinsic_id; 63 64 public: 65 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 66 : InlineCallGenerator(m), 67 _is_virtual(is_virtual), 68 _does_virtual_dispatch(does_virtual_dispatch), 69 _predicates_count((int8_t)predicates_count), 70 _last_predicate((int8_t)-1), 71 _intrinsic_id(id) 72 { 73 } 74 virtual bool is_intrinsic() const { return true; } 75 virtual bool is_virtual() const { return _is_virtual; } 76 virtual bool is_predicated() const { return _predicates_count > 0; } 77 virtual int predicates_count() const { return _predicates_count; } 78 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 79 virtual JVMState* generate(JVMState* jvms); 80 virtual Node* generate_predicate(JVMState* jvms, int predicate); 81 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 82 }; 83 84 85 // Local helper class for LibraryIntrinsic: 86 class LibraryCallKit : public GraphKit { 87 private: 88 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 89 Node* _result; // the result node, if any 90 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 91 92 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 93 94 public: 95 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 96 : GraphKit(jvms), 97 _intrinsic(intrinsic), 98 _result(NULL) 99 { 100 // Check if this is a root compile. In that case we don't have a caller. 101 if (!jvms->has_method()) { 102 _reexecute_sp = sp(); 103 } else { 104 // Find out how many arguments the interpreter needs when deoptimizing 105 // and save the stack pointer value so it can used by uncommon_trap. 106 // We find the argument count by looking at the declared signature. 107 bool ignored_will_link; 108 ciSignature* declared_signature = NULL; 109 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 110 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 111 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 112 } 113 } 114 115 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 116 117 ciMethod* caller() const { return jvms()->method(); } 118 int bci() const { return jvms()->bci(); } 119 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 120 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 121 ciMethod* callee() const { return _intrinsic->method(); } 122 123 bool try_to_inline(int predicate); 124 Node* try_to_predicate(int predicate); 125 126 void push_result() { 127 // Push the result onto the stack. 128 if (!stopped() && result() != NULL) { 129 BasicType bt = result()->bottom_type()->basic_type(); 130 push_node(bt, result()); 131 } 132 } 133 134 private: 135 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 136 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 137 } 138 139 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 140 void set_result(RegionNode* region, PhiNode* value); 141 Node* result() { return _result; } 142 143 virtual int reexecute_sp() { return _reexecute_sp; } 144 145 // Helper functions to inline natives 146 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 147 Node* generate_slow_guard(Node* test, RegionNode* region); 148 Node* generate_fair_guard(Node* test, RegionNode* region); 149 Node* generate_negative_guard(Node* index, RegionNode* region, 150 // resulting CastII of index: 151 Node* *pos_index = NULL); 152 Node* generate_limit_guard(Node* offset, Node* subseq_length, 153 Node* array_length, 154 RegionNode* region); 155 void generate_string_range_check(Node* array, Node* offset, 156 Node* length, bool char_count); 157 Node* generate_current_thread(Node* &tls_output); 158 Node* load_mirror_from_klass(Node* klass); 159 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path, 161 int offset); 162 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 163 RegionNode* region, int null_path) { 164 int offset = java_lang_Class::klass_offset_in_bytes(); 165 return load_klass_from_mirror_common(mirror, never_see_null, 166 region, null_path, 167 offset); 168 } 169 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 170 RegionNode* region, int null_path) { 171 int offset = java_lang_Class::array_klass_offset_in_bytes(); 172 return load_klass_from_mirror_common(mirror, never_see_null, 173 region, null_path, 174 offset); 175 } 176 Node* generate_access_flags_guard(Node* kls, 177 int modifier_mask, int modifier_bits, 178 RegionNode* region); 179 Node* generate_interface_guard(Node* kls, RegionNode* region); 180 Node* generate_array_guard(Node* kls, RegionNode* region) { 181 return generate_array_guard_common(kls, region, false, false); 182 } 183 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 184 return generate_array_guard_common(kls, region, false, true); 185 } 186 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 187 return generate_array_guard_common(kls, region, true, false); 188 } 189 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 190 return generate_array_guard_common(kls, region, true, true); 191 } 192 Node* generate_array_guard_common(Node* kls, RegionNode* region, 193 bool obj_array, bool not_array); 194 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 195 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 196 bool is_virtual = false, bool is_static = false); 197 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 198 return generate_method_call(method_id, false, true); 199 } 200 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 201 return generate_method_call(method_id, true, false); 202 } 203 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 204 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 205 206 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 207 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 208 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 209 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 210 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 211 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); 212 bool inline_string_indexOfChar(); 213 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 214 bool inline_string_toBytesU(); 215 bool inline_string_getCharsU(); 216 bool inline_string_copy(bool compress); 217 bool inline_string_char_access(bool is_store); 218 Node* round_double_node(Node* n); 219 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 220 bool inline_math_native(vmIntrinsics::ID id); 221 bool inline_trig(vmIntrinsics::ID id); 222 bool inline_math(vmIntrinsics::ID id); 223 template <typename OverflowOp> 224 bool inline_math_overflow(Node* arg1, Node* arg2); 225 void inline_math_mathExact(Node* math, Node* test); 226 bool inline_math_addExactI(bool is_increment); 227 bool inline_math_addExactL(bool is_increment); 228 bool inline_math_multiplyExactI(); 229 bool inline_math_multiplyExactL(); 230 bool inline_math_negateExactI(); 231 bool inline_math_negateExactL(); 232 bool inline_math_subtractExactI(bool is_decrement); 233 bool inline_math_subtractExactL(bool is_decrement); 234 bool inline_min_max(vmIntrinsics::ID id); 235 bool inline_notify(vmIntrinsics::ID id); 236 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 237 // This returns Type::AnyPtr, RawPtr, or OopPtr. 238 int classify_unsafe_addr(Node* &base, Node* &offset); 239 Node* make_unsafe_address(Node* base, Node* offset); 240 // Helper for inline_unsafe_access. 241 // Generates the guards that check whether the result of 242 // Unsafe.getObject should be recorded in an SATB log buffer. 243 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 244 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned); 245 static bool klass_needs_init_guard(Node* kls); 246 bool inline_unsafe_allocate(); 247 bool inline_unsafe_copyMemory(); 248 bool inline_native_currentThread(); 249 #ifdef TRACE_HAVE_INTRINSICS 250 bool inline_native_classID(); 251 bool inline_native_threadID(); 252 #endif 253 bool inline_native_time_funcs(address method, const char* funcName); 254 bool inline_native_isInterrupted(); 255 bool inline_native_Class_query(vmIntrinsics::ID id); 256 bool inline_native_subtype_check(); 257 258 bool inline_native_newArray(); 259 bool inline_native_getLength(); 260 bool inline_array_copyOf(bool is_copyOfRange); 261 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 262 bool inline_objects_checkIndex(); 263 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 264 bool inline_native_clone(bool is_virtual); 265 bool inline_native_Reflection_getCallerClass(); 266 // Helper function for inlining native object hash method 267 bool inline_native_hashcode(bool is_virtual, bool is_static); 268 bool inline_native_getClass(); 269 270 // Helper functions for inlining arraycopy 271 bool inline_arraycopy(); 272 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 273 RegionNode* slow_region); 274 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, 275 int& saved_reexecute_sp); 276 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, 277 Node* dest, 278 JVMState* saved_jvms, 279 int saved_reexecute_sp); 280 281 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 282 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 283 bool inline_unsafe_ordered_store(BasicType type); 284 bool inline_unsafe_fence(vmIntrinsics::ID id); 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 bool inline_deoptimize(); 326 }; 327 328 //---------------------------make_vm_intrinsic---------------------------- 329 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 330 vmIntrinsics::ID id = m->intrinsic_id(); 331 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 332 333 if (!m->is_loaded()) { 334 // Do not attempt to inline unloaded methods. 335 return NULL; 336 } 337 338 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 339 bool is_available = false; 340 341 { 342 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 343 // the compiler must transition to '_thread_in_vm' state because both 344 // methods access VM-internal data. 345 VM_ENTRY_MARK; 346 methodHandle mh(THREAD, m->get_Method()); 347 is_available = compiler->is_intrinsic_supported(mh, is_virtual) && 348 !C->directive()->is_intrinsic_disabled(mh) && 349 !vmIntrinsics::is_disabled_by_flags(mh); 350 351 } 352 353 if (is_available) { 354 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 355 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 356 return new LibraryIntrinsic(m, is_virtual, 357 vmIntrinsics::predicates_needed(id), 358 vmIntrinsics::does_virtual_dispatch(id), 359 (vmIntrinsics::ID) id); 360 } else { 361 return NULL; 362 } 363 } 364 365 //----------------------register_library_intrinsics----------------------- 366 // Initialize this file's data structures, for each Compile instance. 367 void Compile::register_library_intrinsics() { 368 // Nothing to do here. 369 } 370 371 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 372 LibraryCallKit kit(jvms, this); 373 Compile* C = kit.C; 374 int nodes = C->unique(); 375 #ifndef PRODUCT 376 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 377 char buf[1000]; 378 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 379 tty->print_cr("Intrinsic %s", str); 380 } 381 #endif 382 ciMethod* callee = kit.callee(); 383 const int bci = kit.bci(); 384 385 // Try to inline the intrinsic. 386 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 387 kit.try_to_inline(_last_predicate)) { 388 if (C->print_intrinsics() || C->print_inlining()) { 389 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 390 } 391 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 392 if (C->log()) { 393 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 394 vmIntrinsics::name_at(intrinsic_id()), 395 (is_virtual() ? " virtual='1'" : ""), 396 C->unique() - nodes); 397 } 398 // Push the result from the inlined method onto the stack. 399 kit.push_result(); 400 C->print_inlining_update(this); 401 return kit.transfer_exceptions_into_jvms(); 402 } 403 404 // The intrinsic bailed out 405 if (C->print_intrinsics() || C->print_inlining()) { 406 if (jvms->has_method()) { 407 // Not a root compile. 408 const char* msg; 409 if (callee->intrinsic_candidate()) { 410 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 411 } else { 412 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 413 : "failed to inline (intrinsic), method not annotated"; 414 } 415 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 416 } else { 417 // Root compile 418 tty->print("Did not generate intrinsic %s%s at bci:%d in", 419 vmIntrinsics::name_at(intrinsic_id()), 420 (is_virtual() ? " (virtual)" : ""), bci); 421 } 422 } 423 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 424 C->print_inlining_update(this); 425 return NULL; 426 } 427 428 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 429 LibraryCallKit kit(jvms, this); 430 Compile* C = kit.C; 431 int nodes = C->unique(); 432 _last_predicate = predicate; 433 #ifndef PRODUCT 434 assert(is_predicated() && predicate < predicates_count(), "sanity"); 435 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 436 char buf[1000]; 437 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 438 tty->print_cr("Predicate for intrinsic %s", str); 439 } 440 #endif 441 ciMethod* callee = kit.callee(); 442 const int bci = kit.bci(); 443 444 Node* slow_ctl = kit.try_to_predicate(predicate); 445 if (!kit.failing()) { 446 if (C->print_intrinsics() || C->print_inlining()) { 447 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 448 } 449 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 450 if (C->log()) { 451 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 452 vmIntrinsics::name_at(intrinsic_id()), 453 (is_virtual() ? " virtual='1'" : ""), 454 C->unique() - nodes); 455 } 456 return slow_ctl; // Could be NULL if the check folds. 457 } 458 459 // The intrinsic bailed out 460 if (C->print_intrinsics() || C->print_inlining()) { 461 if (jvms->has_method()) { 462 // Not a root compile. 463 const char* msg = "failed to generate predicate for intrinsic"; 464 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 465 } else { 466 // Root compile 467 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 468 vmIntrinsics::name_at(intrinsic_id()), 469 (is_virtual() ? " (virtual)" : ""), bci); 470 } 471 } 472 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 473 return NULL; 474 } 475 476 bool LibraryCallKit::try_to_inline(int predicate) { 477 // Handle symbolic names for otherwise undistinguished boolean switches: 478 const bool is_store = true; 479 const bool is_compress = true; 480 const bool is_native_ptr = true; 481 const bool is_static = true; 482 const bool is_volatile = true; 483 484 if (!jvms()->has_method()) { 485 // Root JVMState has a null method. 486 assert(map()->memory()->Opcode() == Op_Parm, ""); 487 // Insert the memory aliasing node 488 set_all_memory(reset_memory()); 489 } 490 assert(merged_memory(), ""); 491 492 493 switch (intrinsic_id()) { 494 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 495 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 496 case vmIntrinsics::_getClass: return inline_native_getClass(); 497 498 case vmIntrinsics::_dsin: 499 case vmIntrinsics::_dcos: 500 case vmIntrinsics::_dtan: 501 case vmIntrinsics::_dabs: 502 case vmIntrinsics::_datan2: 503 case vmIntrinsics::_dsqrt: 504 case vmIntrinsics::_dexp: 505 case vmIntrinsics::_dlog: 506 case vmIntrinsics::_dlog10: 507 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 508 509 case vmIntrinsics::_min: 510 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 511 512 case vmIntrinsics::_notify: 513 case vmIntrinsics::_notifyAll: 514 if (InlineNotify) { 515 return inline_notify(intrinsic_id()); 516 } 517 return false; 518 519 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 520 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 521 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 522 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 523 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 524 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 525 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 526 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 527 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 528 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 529 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 530 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 531 532 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 533 534 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 535 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 536 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 537 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 538 539 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 540 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 541 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 542 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 543 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 544 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 545 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 546 547 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 548 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 549 550 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 551 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 552 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 553 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 554 555 case vmIntrinsics::_compressStringC: 556 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 557 case vmIntrinsics::_inflateStringC: 558 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 559 560 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false); 561 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false); 562 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 563 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 564 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 565 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false); 566 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false); 567 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 568 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 569 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false); 570 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false); 571 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false); 572 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false); 573 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false); 574 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false); 575 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false); 576 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 577 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 578 579 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 580 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 581 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 582 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false); 583 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false); 584 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 585 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 586 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false); 587 588 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false); 589 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false); 590 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false); 591 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false); 592 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false); 593 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 594 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 595 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false); 596 597 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false); 598 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false); 599 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false); 600 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false); 601 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false); 602 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false); 603 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false); 604 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false); 605 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false); 606 607 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false); 608 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false); 609 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false); 610 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false); 611 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false); 612 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false); 613 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false); 614 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false); 615 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false); 616 617 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, true); 618 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, true); 619 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, true); 620 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, true); 621 622 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, true); 623 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, true); 624 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, true); 625 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, true); 626 627 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 628 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 629 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 630 631 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 632 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 633 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 634 635 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 636 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 637 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 638 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 639 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 640 641 case vmIntrinsics::_loadFence: 642 case vmIntrinsics::_storeFence: 643 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 644 645 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 646 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 647 648 #ifdef TRACE_HAVE_INTRINSICS 649 case vmIntrinsics::_classID: return inline_native_classID(); 650 case vmIntrinsics::_threadID: return inline_native_threadID(); 651 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 652 #endif 653 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 654 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 655 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 656 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 657 case vmIntrinsics::_newArray: return inline_native_newArray(); 658 case vmIntrinsics::_getLength: return inline_native_getLength(); 659 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 660 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 661 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 662 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 663 case vmIntrinsics::_Objects_checkIndex: return inline_objects_checkIndex(); 664 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 665 666 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 667 668 case vmIntrinsics::_isInstance: 669 case vmIntrinsics::_getModifiers: 670 case vmIntrinsics::_isInterface: 671 case vmIntrinsics::_isArray: 672 case vmIntrinsics::_isPrimitive: 673 case vmIntrinsics::_getSuperclass: 674 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 675 676 case vmIntrinsics::_floatToRawIntBits: 677 case vmIntrinsics::_floatToIntBits: 678 case vmIntrinsics::_intBitsToFloat: 679 case vmIntrinsics::_doubleToRawLongBits: 680 case vmIntrinsics::_doubleToLongBits: 681 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 682 683 case vmIntrinsics::_numberOfLeadingZeros_i: 684 case vmIntrinsics::_numberOfLeadingZeros_l: 685 case vmIntrinsics::_numberOfTrailingZeros_i: 686 case vmIntrinsics::_numberOfTrailingZeros_l: 687 case vmIntrinsics::_bitCount_i: 688 case vmIntrinsics::_bitCount_l: 689 case vmIntrinsics::_reverseBytes_i: 690 case vmIntrinsics::_reverseBytes_l: 691 case vmIntrinsics::_reverseBytes_s: 692 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 693 694 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 695 696 case vmIntrinsics::_Reference_get: return inline_reference_get(); 697 698 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 699 700 case vmIntrinsics::_aescrypt_encryptBlock: 701 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 702 703 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 704 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 705 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 706 707 case vmIntrinsics::_counterMode_AESCrypt: 708 return inline_counterMode_AESCrypt(intrinsic_id()); 709 710 case vmIntrinsics::_sha_implCompress: 711 case vmIntrinsics::_sha2_implCompress: 712 case vmIntrinsics::_sha5_implCompress: 713 return inline_sha_implCompress(intrinsic_id()); 714 715 case vmIntrinsics::_digestBase_implCompressMB: 716 return inline_digestBase_implCompressMB(predicate); 717 718 case vmIntrinsics::_multiplyToLen: 719 return inline_multiplyToLen(); 720 721 case vmIntrinsics::_squareToLen: 722 return inline_squareToLen(); 723 724 case vmIntrinsics::_mulAdd: 725 return inline_mulAdd(); 726 727 case vmIntrinsics::_montgomeryMultiply: 728 return inline_montgomeryMultiply(); 729 case vmIntrinsics::_montgomerySquare: 730 return inline_montgomerySquare(); 731 732 case vmIntrinsics::_vectorizedMismatch: 733 return inline_vectorizedMismatch(); 734 735 case vmIntrinsics::_ghash_processBlocks: 736 return inline_ghash_processBlocks(); 737 738 case vmIntrinsics::_encodeISOArray: 739 case vmIntrinsics::_encodeByteISOArray: 740 return inline_encodeISOArray(); 741 742 case vmIntrinsics::_updateCRC32: 743 return inline_updateCRC32(); 744 case vmIntrinsics::_updateBytesCRC32: 745 return inline_updateBytesCRC32(); 746 case vmIntrinsics::_updateByteBufferCRC32: 747 return inline_updateByteBufferCRC32(); 748 749 case vmIntrinsics::_updateBytesCRC32C: 750 return inline_updateBytesCRC32C(); 751 case vmIntrinsics::_updateDirectByteBufferCRC32C: 752 return inline_updateDirectByteBufferCRC32C(); 753 754 case vmIntrinsics::_updateBytesAdler32: 755 return inline_updateBytesAdler32(); 756 case vmIntrinsics::_updateByteBufferAdler32: 757 return inline_updateByteBufferAdler32(); 758 759 case vmIntrinsics::_profileBoolean: 760 return inline_profileBoolean(); 761 case vmIntrinsics::_isCompileConstant: 762 return inline_isCompileConstant(); 763 764 case vmIntrinsics::_hasNegatives: 765 return inline_hasNegatives(); 766 767 case vmIntrinsics::_deoptimize: 768 return inline_deoptimize(); 769 770 default: 771 // If you get here, it may be that someone has added a new intrinsic 772 // to the list in vmSymbols.hpp without implementing it here. 773 #ifndef PRODUCT 774 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 775 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 776 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 777 } 778 #endif 779 return false; 780 } 781 } 782 783 Node* LibraryCallKit::try_to_predicate(int predicate) { 784 if (!jvms()->has_method()) { 785 // Root JVMState has a null method. 786 assert(map()->memory()->Opcode() == Op_Parm, ""); 787 // Insert the memory aliasing node 788 set_all_memory(reset_memory()); 789 } 790 assert(merged_memory(), ""); 791 792 switch (intrinsic_id()) { 793 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 794 return inline_cipherBlockChaining_AESCrypt_predicate(false); 795 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 796 return inline_cipherBlockChaining_AESCrypt_predicate(true); 797 case vmIntrinsics::_counterMode_AESCrypt: 798 return inline_counterMode_AESCrypt_predicate(); 799 case vmIntrinsics::_digestBase_implCompressMB: 800 return inline_digestBase_implCompressMB_predicate(predicate); 801 802 default: 803 // If you get here, it may be that someone has added a new intrinsic 804 // to the list in vmSymbols.hpp without implementing it here. 805 #ifndef PRODUCT 806 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 807 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 808 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 809 } 810 #endif 811 Node* slow_ctl = control(); 812 set_control(top()); // No fast path instrinsic 813 return slow_ctl; 814 } 815 } 816 817 //------------------------------set_result------------------------------- 818 // Helper function for finishing intrinsics. 819 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 820 record_for_igvn(region); 821 set_control(_gvn.transform(region)); 822 set_result( _gvn.transform(value)); 823 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 824 } 825 826 //------------------------------generate_guard--------------------------- 827 // Helper function for generating guarded fast-slow graph structures. 828 // The given 'test', if true, guards a slow path. If the test fails 829 // then a fast path can be taken. (We generally hope it fails.) 830 // In all cases, GraphKit::control() is updated to the fast path. 831 // The returned value represents the control for the slow path. 832 // The return value is never 'top'; it is either a valid control 833 // or NULL if it is obvious that the slow path can never be taken. 834 // Also, if region and the slow control are not NULL, the slow edge 835 // is appended to the region. 836 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 837 if (stopped()) { 838 // Already short circuited. 839 return NULL; 840 } 841 842 // Build an if node and its projections. 843 // If test is true we take the slow path, which we assume is uncommon. 844 if (_gvn.type(test) == TypeInt::ZERO) { 845 // The slow branch is never taken. No need to build this guard. 846 return NULL; 847 } 848 849 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 850 851 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 852 if (if_slow == top()) { 853 // The slow branch is never taken. No need to build this guard. 854 return NULL; 855 } 856 857 if (region != NULL) 858 region->add_req(if_slow); 859 860 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 861 set_control(if_fast); 862 863 return if_slow; 864 } 865 866 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 867 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 868 } 869 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 870 return generate_guard(test, region, PROB_FAIR); 871 } 872 873 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 874 Node* *pos_index) { 875 if (stopped()) 876 return NULL; // already stopped 877 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 878 return NULL; // index is already adequately typed 879 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 880 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 881 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 882 if (is_neg != NULL && pos_index != NULL) { 883 // Emulate effect of Parse::adjust_map_after_if. 884 Node* ccast = new CastIINode(index, TypeInt::POS); 885 ccast->set_req(0, control()); 886 (*pos_index) = _gvn.transform(ccast); 887 } 888 return is_neg; 889 } 890 891 // Make sure that 'position' is a valid limit index, in [0..length]. 892 // There are two equivalent plans for checking this: 893 // A. (offset + copyLength) unsigned<= arrayLength 894 // B. offset <= (arrayLength - copyLength) 895 // We require that all of the values above, except for the sum and 896 // difference, are already known to be non-negative. 897 // Plan A is robust in the face of overflow, if offset and copyLength 898 // are both hugely positive. 899 // 900 // Plan B is less direct and intuitive, but it does not overflow at 901 // all, since the difference of two non-negatives is always 902 // representable. Whenever Java methods must perform the equivalent 903 // check they generally use Plan B instead of Plan A. 904 // For the moment we use Plan A. 905 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 906 Node* subseq_length, 907 Node* array_length, 908 RegionNode* region) { 909 if (stopped()) 910 return NULL; // already stopped 911 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 912 if (zero_offset && subseq_length->eqv_uncast(array_length)) 913 return NULL; // common case of whole-array copy 914 Node* last = subseq_length; 915 if (!zero_offset) // last += offset 916 last = _gvn.transform(new AddINode(last, offset)); 917 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 918 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 919 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 920 return is_over; 921 } 922 923 // Emit range checks for the given String.value byte array 924 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 925 if (stopped()) { 926 return; // already stopped 927 } 928 RegionNode* bailout = new RegionNode(1); 929 record_for_igvn(bailout); 930 if (char_count) { 931 // Convert char count to byte count 932 count = _gvn.transform(new LShiftINode(count, intcon(1))); 933 } 934 935 // Offset and count must not be negative 936 generate_negative_guard(offset, bailout); 937 generate_negative_guard(count, bailout); 938 // Offset + count must not exceed length of array 939 generate_limit_guard(offset, count, load_array_length(array), bailout); 940 941 if (bailout->req() > 1) { 942 PreserveJVMState pjvms(this); 943 set_control(_gvn.transform(bailout)); 944 uncommon_trap(Deoptimization::Reason_intrinsic, 945 Deoptimization::Action_maybe_recompile); 946 } 947 } 948 949 //--------------------------generate_current_thread-------------------- 950 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 951 ciKlass* thread_klass = env()->Thread_klass(); 952 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 953 Node* thread = _gvn.transform(new ThreadLocalNode()); 954 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 955 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 956 tls_output = thread; 957 return threadObj; 958 } 959 960 961 //------------------------------make_string_method_node------------------------ 962 // Helper method for String intrinsic functions. This version is called with 963 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 964 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 965 // containing the lengths of str1 and str2. 966 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 967 Node* result = NULL; 968 switch (opcode) { 969 case Op_StrIndexOf: 970 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 971 str1_start, cnt1, str2_start, cnt2, ae); 972 break; 973 case Op_StrComp: 974 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 975 str1_start, cnt1, str2_start, cnt2, ae); 976 break; 977 case Op_StrEquals: 978 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 979 // Use the constant length if there is one because optimized match rule may exist. 980 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 981 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 982 break; 983 default: 984 ShouldNotReachHere(); 985 return NULL; 986 } 987 988 // All these intrinsics have checks. 989 C->set_has_split_ifs(true); // Has chance for split-if optimization 990 991 return _gvn.transform(result); 992 } 993 994 //------------------------------inline_string_compareTo------------------------ 995 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 996 Node* arg1 = argument(0); 997 Node* arg2 = argument(1); 998 999 // Get start addr and length of first argument 1000 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1001 Node* arg1_cnt = load_array_length(arg1); 1002 1003 // Get start addr and length of second argument 1004 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1005 Node* arg2_cnt = load_array_length(arg2); 1006 1007 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1008 set_result(result); 1009 return true; 1010 } 1011 1012 //------------------------------inline_string_equals------------------------ 1013 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1014 Node* arg1 = argument(0); 1015 Node* arg2 = argument(1); 1016 1017 // paths (plus control) merge 1018 RegionNode* region = new RegionNode(3); 1019 Node* phi = new PhiNode(region, TypeInt::BOOL); 1020 1021 if (!stopped()) { 1022 // Get start addr and length of first argument 1023 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1024 Node* arg1_cnt = load_array_length(arg1); 1025 1026 // Get start addr and length of second argument 1027 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1028 Node* arg2_cnt = load_array_length(arg2); 1029 1030 // Check for arg1_cnt != arg2_cnt 1031 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1032 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1033 Node* if_ne = generate_slow_guard(bol, NULL); 1034 if (if_ne != NULL) { 1035 phi->init_req(2, intcon(0)); 1036 region->init_req(2, if_ne); 1037 } 1038 1039 // Check for count == 0 is done by assembler code for StrEquals. 1040 1041 if (!stopped()) { 1042 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1043 phi->init_req(1, equals); 1044 region->init_req(1, control()); 1045 } 1046 } 1047 1048 // post merge 1049 set_control(_gvn.transform(region)); 1050 record_for_igvn(region); 1051 1052 set_result(_gvn.transform(phi)); 1053 return true; 1054 } 1055 1056 //------------------------------inline_array_equals---------------------------- 1057 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1058 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1059 Node* arg1 = argument(0); 1060 Node* arg2 = argument(1); 1061 1062 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1063 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1064 return true; 1065 } 1066 1067 //------------------------------inline_hasNegatives------------------------------ 1068 bool LibraryCallKit::inline_hasNegatives() { 1069 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1070 return false; 1071 } 1072 1073 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1074 // no receiver since it is static method 1075 Node* ba = argument(0); 1076 Node* offset = argument(1); 1077 Node* len = argument(2); 1078 1079 // Range checks 1080 generate_string_range_check(ba, offset, len, false); 1081 if (stopped()) { 1082 return true; 1083 } 1084 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1085 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1086 set_result(_gvn.transform(result)); 1087 return true; 1088 } 1089 1090 bool LibraryCallKit::inline_objects_checkIndex() { 1091 Node* index = argument(0); 1092 Node* length = argument(1); 1093 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1094 return false; 1095 } 1096 1097 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1098 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1099 1100 { 1101 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1102 uncommon_trap(Deoptimization::Reason_intrinsic, 1103 Deoptimization::Action_make_not_entrant); 1104 } 1105 1106 if (stopped()) { 1107 return false; 1108 } 1109 1110 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1111 BoolTest::mask btest = BoolTest::lt; 1112 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1113 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1114 _gvn.set_type(rc, rc->Value(&_gvn)); 1115 if (!rc_bool->is_Con()) { 1116 record_for_igvn(rc); 1117 } 1118 set_control(_gvn.transform(new IfTrueNode(rc))); 1119 { 1120 PreserveJVMState pjvms(this); 1121 set_control(_gvn.transform(new IfFalseNode(rc))); 1122 uncommon_trap(Deoptimization::Reason_range_check, 1123 Deoptimization::Action_make_not_entrant); 1124 } 1125 1126 if (stopped()) { 1127 return false; 1128 } 1129 1130 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1131 result->set_req(0, control()); 1132 result = _gvn.transform(result); 1133 set_result(result); 1134 replace_in_map(index, result); 1135 return true; 1136 } 1137 1138 //------------------------------inline_string_indexOf------------------------ 1139 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1140 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1141 return false; 1142 } 1143 Node* src = argument(0); 1144 Node* tgt = argument(1); 1145 1146 // Make the merge point 1147 RegionNode* result_rgn = new RegionNode(4); 1148 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1149 1150 // Get start addr and length of source string 1151 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1152 Node* src_count = load_array_length(src); 1153 1154 // Get start addr and length of substring 1155 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1156 Node* tgt_count = load_array_length(tgt); 1157 1158 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1159 // Divide src size by 2 if String is UTF16 encoded 1160 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1161 } 1162 if (ae == StrIntrinsicNode::UU) { 1163 // Divide substring size by 2 if String is UTF16 encoded 1164 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1165 } 1166 1167 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1168 if (result != NULL) { 1169 result_phi->init_req(3, result); 1170 result_rgn->init_req(3, control()); 1171 } 1172 set_control(_gvn.transform(result_rgn)); 1173 record_for_igvn(result_rgn); 1174 set_result(_gvn.transform(result_phi)); 1175 1176 return true; 1177 } 1178 1179 //-----------------------------inline_string_indexOf----------------------- 1180 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1181 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1182 return false; 1183 } 1184 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1185 return false; 1186 } 1187 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1188 Node* src = argument(0); // byte[] 1189 Node* src_count = argument(1); // char count 1190 Node* tgt = argument(2); // byte[] 1191 Node* tgt_count = argument(3); // char count 1192 Node* from_index = argument(4); // char index 1193 1194 // Multiply byte array index by 2 if String is UTF16 encoded 1195 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1196 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1197 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1198 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1199 1200 // Range checks 1201 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1202 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1203 if (stopped()) { 1204 return true; 1205 } 1206 1207 RegionNode* region = new RegionNode(5); 1208 Node* phi = new PhiNode(region, TypeInt::INT); 1209 1210 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1211 if (result != NULL) { 1212 // The result is index relative to from_index if substring was found, -1 otherwise. 1213 // Generate code which will fold into cmove. 1214 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1215 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1216 1217 Node* if_lt = generate_slow_guard(bol, NULL); 1218 if (if_lt != NULL) { 1219 // result == -1 1220 phi->init_req(3, result); 1221 region->init_req(3, if_lt); 1222 } 1223 if (!stopped()) { 1224 result = _gvn.transform(new AddINode(result, from_index)); 1225 phi->init_req(4, result); 1226 region->init_req(4, control()); 1227 } 1228 } 1229 1230 set_control(_gvn.transform(region)); 1231 record_for_igvn(region); 1232 set_result(_gvn.transform(phi)); 1233 1234 return true; 1235 } 1236 1237 // Create StrIndexOfNode with fast path checks 1238 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1239 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1240 // Check for substr count > string count 1241 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1242 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1243 Node* if_gt = generate_slow_guard(bol, NULL); 1244 if (if_gt != NULL) { 1245 phi->init_req(1, intcon(-1)); 1246 region->init_req(1, if_gt); 1247 } 1248 if (!stopped()) { 1249 // Check for substr count == 0 1250 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1251 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1252 Node* if_zero = generate_slow_guard(bol, NULL); 1253 if (if_zero != NULL) { 1254 phi->init_req(2, intcon(0)); 1255 region->init_req(2, if_zero); 1256 } 1257 } 1258 if (!stopped()) { 1259 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1260 } 1261 return NULL; 1262 } 1263 1264 //-----------------------------inline_string_indexOfChar----------------------- 1265 bool LibraryCallKit::inline_string_indexOfChar() { 1266 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1267 return false; 1268 } 1269 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1270 return false; 1271 } 1272 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1273 Node* src = argument(0); // byte[] 1274 Node* tgt = argument(1); // tgt is int ch 1275 Node* from_index = argument(2); 1276 Node* max = argument(3); 1277 1278 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1279 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1280 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1281 1282 // Range checks 1283 generate_string_range_check(src, src_offset, src_count, true); 1284 if (stopped()) { 1285 return true; 1286 } 1287 1288 RegionNode* region = new RegionNode(3); 1289 Node* phi = new PhiNode(region, TypeInt::INT); 1290 1291 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1292 C->set_has_split_ifs(true); // Has chance for split-if optimization 1293 _gvn.transform(result); 1294 1295 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1296 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1297 1298 Node* if_lt = generate_slow_guard(bol, NULL); 1299 if (if_lt != NULL) { 1300 // result == -1 1301 phi->init_req(2, result); 1302 region->init_req(2, if_lt); 1303 } 1304 if (!stopped()) { 1305 result = _gvn.transform(new AddINode(result, from_index)); 1306 phi->init_req(1, result); 1307 region->init_req(1, control()); 1308 } 1309 set_control(_gvn.transform(region)); 1310 record_for_igvn(region); 1311 set_result(_gvn.transform(phi)); 1312 1313 return true; 1314 } 1315 //---------------------------inline_string_copy--------------------- 1316 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1317 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1318 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1319 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1320 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1321 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1322 bool LibraryCallKit::inline_string_copy(bool compress) { 1323 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1324 return false; 1325 } 1326 int nargs = 5; // 2 oops, 3 ints 1327 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1328 1329 Node* src = argument(0); 1330 Node* src_offset = argument(1); 1331 Node* dst = argument(2); 1332 Node* dst_offset = argument(3); 1333 Node* length = argument(4); 1334 1335 // Check for allocation before we add nodes that would confuse 1336 // tightly_coupled_allocation() 1337 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1338 1339 // Figure out the size and type of the elements we will be copying. 1340 const Type* src_type = src->Value(&_gvn); 1341 const Type* dst_type = dst->Value(&_gvn); 1342 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1343 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1344 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1345 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1346 "Unsupported array types for inline_string_copy"); 1347 1348 // Range checks 1349 generate_string_range_check(src, src_offset, length, compress && src_elem == T_BYTE); 1350 generate_string_range_check(dst, dst_offset, length, !compress && dst_elem == T_BYTE); 1351 if (stopped()) { 1352 return true; 1353 } 1354 1355 // Convert char[] offsets to byte[] offsets 1356 if (compress && src_elem == T_BYTE) { 1357 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1358 } else if (!compress && dst_elem == T_BYTE) { 1359 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1360 } 1361 1362 Node* src_start = array_element_address(src, src_offset, src_elem); 1363 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1364 // 'src_start' points to src array + scaled offset 1365 // 'dst_start' points to dst array + scaled offset 1366 Node* count = NULL; 1367 if (compress) { 1368 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1369 } else { 1370 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1371 } 1372 1373 if (alloc != NULL) { 1374 if (alloc->maybe_set_complete(&_gvn)) { 1375 // "You break it, you buy it." 1376 InitializeNode* init = alloc->initialization(); 1377 assert(init->is_complete(), "we just did this"); 1378 init->set_complete_with_arraycopy(); 1379 assert(dst->is_CheckCastPP(), "sanity"); 1380 assert(dst->in(0)->in(0) == init, "dest pinned"); 1381 } 1382 // Do not let stores that initialize this object be reordered with 1383 // a subsequent store that would make this object accessible by 1384 // other threads. 1385 // Record what AllocateNode this StoreStore protects so that 1386 // escape analysis can go from the MemBarStoreStoreNode to the 1387 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1388 // based on the escape status of the AllocateNode. 1389 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1390 } 1391 if (compress) { 1392 set_result(_gvn.transform(count)); 1393 } 1394 return true; 1395 } 1396 1397 #ifdef _LP64 1398 #define XTOP ,top() /*additional argument*/ 1399 #else //_LP64 1400 #define XTOP /*no additional argument*/ 1401 #endif //_LP64 1402 1403 //------------------------inline_string_toBytesU-------------------------- 1404 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1405 bool LibraryCallKit::inline_string_toBytesU() { 1406 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1407 return false; 1408 } 1409 // Get the arguments. 1410 Node* value = argument(0); 1411 Node* offset = argument(1); 1412 Node* length = argument(2); 1413 1414 Node* newcopy = NULL; 1415 1416 // Set the original stack and the reexecute bit for the interpreter to reexecute 1417 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1418 { PreserveReexecuteState preexecs(this); 1419 jvms()->set_should_reexecute(true); 1420 1421 // Check if a null path was taken unconditionally. 1422 value = null_check(value); 1423 1424 RegionNode* bailout = new RegionNode(1); 1425 record_for_igvn(bailout); 1426 1427 // Range checks 1428 generate_negative_guard(offset, bailout); 1429 generate_negative_guard(length, bailout); 1430 generate_limit_guard(offset, length, load_array_length(value), bailout); 1431 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1432 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1433 1434 if (bailout->req() > 1) { 1435 PreserveJVMState pjvms(this); 1436 set_control(_gvn.transform(bailout)); 1437 uncommon_trap(Deoptimization::Reason_intrinsic, 1438 Deoptimization::Action_maybe_recompile); 1439 } 1440 if (stopped()) { 1441 return true; 1442 } 1443 1444 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1445 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1446 newcopy = new_array(klass_node, size, 0); // no arguments to push 1447 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1448 1449 // Calculate starting addresses. 1450 Node* src_start = array_element_address(value, offset, T_CHAR); 1451 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1452 1453 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1454 const TypeInt* toffset = gvn().type(offset)->is_int(); 1455 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1456 1457 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1458 const char* copyfunc_name = "arraycopy"; 1459 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1460 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1461 OptoRuntime::fast_arraycopy_Type(), 1462 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1463 src_start, dst_start, ConvI2X(length) XTOP); 1464 // Do not let reads from the cloned object float above the arraycopy. 1465 if (alloc != NULL) { 1466 if (alloc->maybe_set_complete(&_gvn)) { 1467 // "You break it, you buy it." 1468 InitializeNode* init = alloc->initialization(); 1469 assert(init->is_complete(), "we just did this"); 1470 init->set_complete_with_arraycopy(); 1471 assert(newcopy->is_CheckCastPP(), "sanity"); 1472 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1473 } 1474 // Do not let stores that initialize this object be reordered with 1475 // a subsequent store that would make this object accessible by 1476 // other threads. 1477 // Record what AllocateNode this StoreStore protects so that 1478 // escape analysis can go from the MemBarStoreStoreNode to the 1479 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1480 // based on the escape status of the AllocateNode. 1481 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1482 } else { 1483 insert_mem_bar(Op_MemBarCPUOrder); 1484 } 1485 } // original reexecute is set back here 1486 1487 C->set_has_split_ifs(true); // Has chance for split-if optimization 1488 if (!stopped()) { 1489 set_result(newcopy); 1490 } 1491 return true; 1492 } 1493 1494 //------------------------inline_string_getCharsU-------------------------- 1495 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1496 bool LibraryCallKit::inline_string_getCharsU() { 1497 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1498 return false; 1499 } 1500 1501 // Get the arguments. 1502 Node* src = argument(0); 1503 Node* src_begin = argument(1); 1504 Node* src_end = argument(2); // exclusive offset (i < src_end) 1505 Node* dst = argument(3); 1506 Node* dst_begin = argument(4); 1507 1508 // Check for allocation before we add nodes that would confuse 1509 // tightly_coupled_allocation() 1510 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1511 1512 // Check if a null path was taken unconditionally. 1513 src = null_check(src); 1514 dst = null_check(dst); 1515 if (stopped()) { 1516 return true; 1517 } 1518 1519 // Get length and convert char[] offset to byte[] offset 1520 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1521 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1522 1523 // Range checks 1524 generate_string_range_check(src, src_begin, length, true); 1525 generate_string_range_check(dst, dst_begin, length, false); 1526 if (stopped()) { 1527 return true; 1528 } 1529 1530 if (!stopped()) { 1531 // Calculate starting addresses. 1532 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1533 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1534 1535 // Check if array addresses are aligned to HeapWordSize 1536 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1537 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1538 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1539 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1540 1541 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1542 const char* copyfunc_name = "arraycopy"; 1543 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1544 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1545 OptoRuntime::fast_arraycopy_Type(), 1546 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1547 src_start, dst_start, ConvI2X(length) XTOP); 1548 // Do not let reads from the cloned object float above the arraycopy. 1549 if (alloc != NULL) { 1550 if (alloc->maybe_set_complete(&_gvn)) { 1551 // "You break it, you buy it." 1552 InitializeNode* init = alloc->initialization(); 1553 assert(init->is_complete(), "we just did this"); 1554 init->set_complete_with_arraycopy(); 1555 assert(dst->is_CheckCastPP(), "sanity"); 1556 assert(dst->in(0)->in(0) == init, "dest pinned"); 1557 } 1558 // Do not let stores that initialize this object be reordered with 1559 // a subsequent store that would make this object accessible by 1560 // other threads. 1561 // Record what AllocateNode this StoreStore protects so that 1562 // escape analysis can go from the MemBarStoreStoreNode to the 1563 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1564 // based on the escape status of the AllocateNode. 1565 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1566 } else { 1567 insert_mem_bar(Op_MemBarCPUOrder); 1568 } 1569 } 1570 1571 C->set_has_split_ifs(true); // Has chance for split-if optimization 1572 return true; 1573 } 1574 1575 //----------------------inline_string_char_access---------------------------- 1576 // Store/Load char to/from byte[] array. 1577 // static void StringUTF16.putChar(byte[] val, int index, int c) 1578 // static char StringUTF16.getChar(byte[] val, int index) 1579 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1580 Node* value = argument(0); 1581 Node* index = argument(1); 1582 Node* ch = is_store ? argument(2) : NULL; 1583 1584 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1585 // correctly requires matched array shapes. 1586 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1587 "sanity: byte[] and char[] bases agree"); 1588 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1589 "sanity: byte[] and char[] scales agree"); 1590 1591 Node* adr = array_element_address(value, index, T_CHAR); 1592 if (is_store) { 1593 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1594 false, false, true /* mismatched */); 1595 } else { 1596 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1597 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */); 1598 set_result(ch); 1599 } 1600 return true; 1601 } 1602 1603 //--------------------------round_double_node-------------------------------- 1604 // Round a double node if necessary. 1605 Node* LibraryCallKit::round_double_node(Node* n) { 1606 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1607 n = _gvn.transform(new RoundDoubleNode(0, n)); 1608 return n; 1609 } 1610 1611 //------------------------------inline_math----------------------------------- 1612 // public static double Math.abs(double) 1613 // public static double Math.sqrt(double) 1614 // public static double Math.log(double) 1615 // public static double Math.log10(double) 1616 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1617 Node* arg = round_double_node(argument(0)); 1618 Node* n = NULL; 1619 switch (id) { 1620 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1621 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1622 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1623 default: fatal_unexpected_iid(id); break; 1624 } 1625 set_result(_gvn.transform(n)); 1626 return true; 1627 } 1628 1629 //------------------------------inline_trig---------------------------------- 1630 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1631 // argument reduction which will turn into a fast/slow diamond. 1632 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1633 Node* arg = round_double_node(argument(0)); 1634 Node* n = NULL; 1635 1636 switch (id) { 1637 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1638 default: fatal_unexpected_iid(id); break; 1639 } 1640 n = _gvn.transform(n); 1641 1642 // Rounding required? Check for argument reduction! 1643 if (Matcher::strict_fp_requires_explicit_rounding) { 1644 static const double pi_4 = 0.7853981633974483; 1645 static const double neg_pi_4 = -0.7853981633974483; 1646 // pi/2 in 80-bit extended precision 1647 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1648 // -pi/2 in 80-bit extended precision 1649 // 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}; 1650 // Cutoff value for using this argument reduction technique 1651 //static const double pi_2_minus_epsilon = 1.564660403643354; 1652 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1653 1654 // Pseudocode for sin: 1655 // if (x <= Math.PI / 4.0) { 1656 // if (x >= -Math.PI / 4.0) return fsin(x); 1657 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1658 // } else { 1659 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1660 // } 1661 // return StrictMath.sin(x); 1662 1663 // Pseudocode for cos: 1664 // if (x <= Math.PI / 4.0) { 1665 // if (x >= -Math.PI / 4.0) return fcos(x); 1666 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1667 // } else { 1668 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1669 // } 1670 // return StrictMath.cos(x); 1671 1672 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1673 // requires a special machine instruction to load it. Instead we'll try 1674 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1675 // probably do the math inside the SIN encoding. 1676 1677 // Make the merge point 1678 RegionNode* r = new RegionNode(3); 1679 Node* phi = new PhiNode(r, Type::DOUBLE); 1680 1681 // Flatten arg so we need only 1 test 1682 Node *abs = _gvn.transform(new AbsDNode(arg)); 1683 // Node for PI/4 constant 1684 Node *pi4 = makecon(TypeD::make(pi_4)); 1685 // Check PI/4 : abs(arg) 1686 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1687 // Check: If PI/4 < abs(arg) then go slow 1688 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1689 // Branch either way 1690 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1691 set_control(opt_iff(r,iff)); 1692 1693 // Set fast path result 1694 phi->init_req(2, n); 1695 1696 // Slow path - non-blocking leaf call 1697 Node* call = NULL; 1698 switch (id) { 1699 case vmIntrinsics::_dtan: 1700 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1701 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1702 "Tan", NULL, arg, top()); 1703 break; 1704 } 1705 assert(control()->in(0) == call, ""); 1706 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1707 r->init_req(1, control()); 1708 phi->init_req(1, slow_result); 1709 1710 // Post-merge 1711 set_control(_gvn.transform(r)); 1712 record_for_igvn(r); 1713 n = _gvn.transform(phi); 1714 1715 C->set_has_split_ifs(true); // Has chance for split-if optimization 1716 } 1717 set_result(n); 1718 return true; 1719 } 1720 1721 //------------------------------runtime_math----------------------------- 1722 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1723 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1724 "must be (DD)D or (D)D type"); 1725 1726 // Inputs 1727 Node* a = round_double_node(argument(0)); 1728 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1729 1730 const TypePtr* no_memory_effects = NULL; 1731 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1732 no_memory_effects, 1733 a, top(), b, b ? top() : NULL); 1734 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1735 #ifdef ASSERT 1736 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1737 assert(value_top == top(), "second value must be top"); 1738 #endif 1739 1740 set_result(value); 1741 return true; 1742 } 1743 1744 //------------------------------inline_math_native----------------------------- 1745 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1746 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1747 switch (id) { 1748 // These intrinsics are not properly supported on all hardware 1749 case vmIntrinsics::_dsin: 1750 return StubRoutines::dsin() != NULL ? 1751 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1752 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1753 case vmIntrinsics::_dcos: 1754 return StubRoutines::dcos() != NULL ? 1755 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1756 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1757 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1758 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1759 1760 case vmIntrinsics::_dlog: 1761 return StubRoutines::dlog() != NULL ? 1762 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1763 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1764 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1765 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1766 1767 // These intrinsics are supported on all hardware 1768 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1769 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1770 1771 case vmIntrinsics::_dexp: 1772 return StubRoutines::dexp() != NULL ? 1773 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1774 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1775 case vmIntrinsics::_dpow: 1776 return StubRoutines::dpow() != NULL ? 1777 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1778 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1779 #undef FN_PTR 1780 1781 // These intrinsics are not yet correctly implemented 1782 case vmIntrinsics::_datan2: 1783 return false; 1784 1785 default: 1786 fatal_unexpected_iid(id); 1787 return false; 1788 } 1789 } 1790 1791 static bool is_simple_name(Node* n) { 1792 return (n->req() == 1 // constant 1793 || (n->is_Type() && n->as_Type()->type()->singleton()) 1794 || n->is_Proj() // parameter or return value 1795 || n->is_Phi() // local of some sort 1796 ); 1797 } 1798 1799 //----------------------------inline_notify-----------------------------------* 1800 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1801 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1802 address func; 1803 if (id == vmIntrinsics::_notify) { 1804 func = OptoRuntime::monitor_notify_Java(); 1805 } else { 1806 func = OptoRuntime::monitor_notifyAll_Java(); 1807 } 1808 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1809 make_slow_call_ex(call, env()->Throwable_klass(), false); 1810 return true; 1811 } 1812 1813 1814 //----------------------------inline_min_max----------------------------------- 1815 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1816 set_result(generate_min_max(id, argument(0), argument(1))); 1817 return true; 1818 } 1819 1820 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1821 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1822 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1823 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1824 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1825 1826 { 1827 PreserveJVMState pjvms(this); 1828 PreserveReexecuteState preexecs(this); 1829 jvms()->set_should_reexecute(true); 1830 1831 set_control(slow_path); 1832 set_i_o(i_o()); 1833 1834 uncommon_trap(Deoptimization::Reason_intrinsic, 1835 Deoptimization::Action_none); 1836 } 1837 1838 set_control(fast_path); 1839 set_result(math); 1840 } 1841 1842 template <typename OverflowOp> 1843 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1844 typedef typename OverflowOp::MathOp MathOp; 1845 1846 MathOp* mathOp = new MathOp(arg1, arg2); 1847 Node* operation = _gvn.transform( mathOp ); 1848 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1849 inline_math_mathExact(operation, ofcheck); 1850 return true; 1851 } 1852 1853 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1854 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1855 } 1856 1857 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1858 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1859 } 1860 1861 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1862 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1863 } 1864 1865 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1866 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1867 } 1868 1869 bool LibraryCallKit::inline_math_negateExactI() { 1870 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1871 } 1872 1873 bool LibraryCallKit::inline_math_negateExactL() { 1874 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 1875 } 1876 1877 bool LibraryCallKit::inline_math_multiplyExactI() { 1878 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 1879 } 1880 1881 bool LibraryCallKit::inline_math_multiplyExactL() { 1882 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 1883 } 1884 1885 Node* 1886 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1887 // These are the candidate return value: 1888 Node* xvalue = x0; 1889 Node* yvalue = y0; 1890 1891 if (xvalue == yvalue) { 1892 return xvalue; 1893 } 1894 1895 bool want_max = (id == vmIntrinsics::_max); 1896 1897 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1898 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1899 if (txvalue == NULL || tyvalue == NULL) return top(); 1900 // This is not really necessary, but it is consistent with a 1901 // hypothetical MaxINode::Value method: 1902 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1903 1904 // %%% This folding logic should (ideally) be in a different place. 1905 // Some should be inside IfNode, and there to be a more reliable 1906 // transformation of ?: style patterns into cmoves. We also want 1907 // more powerful optimizations around cmove and min/max. 1908 1909 // Try to find a dominating comparison of these guys. 1910 // It can simplify the index computation for Arrays.copyOf 1911 // and similar uses of System.arraycopy. 1912 // First, compute the normalized version of CmpI(x, y). 1913 int cmp_op = Op_CmpI; 1914 Node* xkey = xvalue; 1915 Node* ykey = yvalue; 1916 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 1917 if (ideal_cmpxy->is_Cmp()) { 1918 // E.g., if we have CmpI(length - offset, count), 1919 // it might idealize to CmpI(length, count + offset) 1920 cmp_op = ideal_cmpxy->Opcode(); 1921 xkey = ideal_cmpxy->in(1); 1922 ykey = ideal_cmpxy->in(2); 1923 } 1924 1925 // Start by locating any relevant comparisons. 1926 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1927 Node* cmpxy = NULL; 1928 Node* cmpyx = NULL; 1929 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1930 Node* cmp = start_from->fast_out(k); 1931 if (cmp->outcnt() > 0 && // must have prior uses 1932 cmp->in(0) == NULL && // must be context-independent 1933 cmp->Opcode() == cmp_op) { // right kind of compare 1934 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1935 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1936 } 1937 } 1938 1939 const int NCMPS = 2; 1940 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1941 int cmpn; 1942 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1943 if (cmps[cmpn] != NULL) break; // find a result 1944 } 1945 if (cmpn < NCMPS) { 1946 // Look for a dominating test that tells us the min and max. 1947 int depth = 0; // Limit search depth for speed 1948 Node* dom = control(); 1949 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1950 if (++depth >= 100) break; 1951 Node* ifproj = dom; 1952 if (!ifproj->is_Proj()) continue; 1953 Node* iff = ifproj->in(0); 1954 if (!iff->is_If()) continue; 1955 Node* bol = iff->in(1); 1956 if (!bol->is_Bool()) continue; 1957 Node* cmp = bol->in(1); 1958 if (cmp == NULL) continue; 1959 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1960 if (cmps[cmpn] == cmp) break; 1961 if (cmpn == NCMPS) continue; 1962 BoolTest::mask btest = bol->as_Bool()->_test._test; 1963 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1964 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1965 // At this point, we know that 'x btest y' is true. 1966 switch (btest) { 1967 case BoolTest::eq: 1968 // They are proven equal, so we can collapse the min/max. 1969 // Either value is the answer. Choose the simpler. 1970 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1971 return yvalue; 1972 return xvalue; 1973 case BoolTest::lt: // x < y 1974 case BoolTest::le: // x <= y 1975 return (want_max ? yvalue : xvalue); 1976 case BoolTest::gt: // x > y 1977 case BoolTest::ge: // x >= y 1978 return (want_max ? xvalue : yvalue); 1979 } 1980 } 1981 } 1982 1983 // We failed to find a dominating test. 1984 // Let's pick a test that might GVN with prior tests. 1985 Node* best_bol = NULL; 1986 BoolTest::mask best_btest = BoolTest::illegal; 1987 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1988 Node* cmp = cmps[cmpn]; 1989 if (cmp == NULL) continue; 1990 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1991 Node* bol = cmp->fast_out(j); 1992 if (!bol->is_Bool()) continue; 1993 BoolTest::mask btest = bol->as_Bool()->_test._test; 1994 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1995 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1996 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1997 best_bol = bol->as_Bool(); 1998 best_btest = btest; 1999 } 2000 } 2001 } 2002 2003 Node* answer_if_true = NULL; 2004 Node* answer_if_false = NULL; 2005 switch (best_btest) { 2006 default: 2007 if (cmpxy == NULL) 2008 cmpxy = ideal_cmpxy; 2009 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2010 // and fall through: 2011 case BoolTest::lt: // x < y 2012 case BoolTest::le: // x <= y 2013 answer_if_true = (want_max ? yvalue : xvalue); 2014 answer_if_false = (want_max ? xvalue : yvalue); 2015 break; 2016 case BoolTest::gt: // x > y 2017 case BoolTest::ge: // x >= y 2018 answer_if_true = (want_max ? xvalue : yvalue); 2019 answer_if_false = (want_max ? yvalue : xvalue); 2020 break; 2021 } 2022 2023 jint hi, lo; 2024 if (want_max) { 2025 // We can sharpen the minimum. 2026 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2027 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2028 } else { 2029 // We can sharpen the maximum. 2030 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2031 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2032 } 2033 2034 // Use a flow-free graph structure, to avoid creating excess control edges 2035 // which could hinder other optimizations. 2036 // Since Math.min/max is often used with arraycopy, we want 2037 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2038 Node* cmov = CMoveNode::make(NULL, best_bol, 2039 answer_if_false, answer_if_true, 2040 TypeInt::make(lo, hi, widen)); 2041 2042 return _gvn.transform(cmov); 2043 2044 /* 2045 // This is not as desirable as it may seem, since Min and Max 2046 // nodes do not have a full set of optimizations. 2047 // And they would interfere, anyway, with 'if' optimizations 2048 // and with CMoveI canonical forms. 2049 switch (id) { 2050 case vmIntrinsics::_min: 2051 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2052 case vmIntrinsics::_max: 2053 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2054 default: 2055 ShouldNotReachHere(); 2056 } 2057 */ 2058 } 2059 2060 inline int 2061 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2062 const TypePtr* base_type = TypePtr::NULL_PTR; 2063 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2064 if (base_type == NULL) { 2065 // Unknown type. 2066 return Type::AnyPtr; 2067 } else if (base_type == TypePtr::NULL_PTR) { 2068 // Since this is a NULL+long form, we have to switch to a rawptr. 2069 base = _gvn.transform(new CastX2PNode(offset)); 2070 offset = MakeConX(0); 2071 return Type::RawPtr; 2072 } else if (base_type->base() == Type::RawPtr) { 2073 return Type::RawPtr; 2074 } else if (base_type->isa_oopptr()) { 2075 // Base is never null => always a heap address. 2076 if (base_type->ptr() == TypePtr::NotNull) { 2077 return Type::OopPtr; 2078 } 2079 // Offset is small => always a heap address. 2080 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2081 if (offset_type != NULL && 2082 base_type->offset() == 0 && // (should always be?) 2083 offset_type->_lo >= 0 && 2084 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2085 return Type::OopPtr; 2086 } 2087 // Otherwise, it might either be oop+off or NULL+addr. 2088 return Type::AnyPtr; 2089 } else { 2090 // No information: 2091 return Type::AnyPtr; 2092 } 2093 } 2094 2095 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2096 int kind = classify_unsafe_addr(base, offset); 2097 if (kind == Type::RawPtr) { 2098 return basic_plus_adr(top(), base, offset); 2099 } else { 2100 return basic_plus_adr(base, offset); 2101 } 2102 } 2103 2104 //--------------------------inline_number_methods----------------------------- 2105 // inline int Integer.numberOfLeadingZeros(int) 2106 // inline int Long.numberOfLeadingZeros(long) 2107 // 2108 // inline int Integer.numberOfTrailingZeros(int) 2109 // inline int Long.numberOfTrailingZeros(long) 2110 // 2111 // inline int Integer.bitCount(int) 2112 // inline int Long.bitCount(long) 2113 // 2114 // inline char Character.reverseBytes(char) 2115 // inline short Short.reverseBytes(short) 2116 // inline int Integer.reverseBytes(int) 2117 // inline long Long.reverseBytes(long) 2118 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2119 Node* arg = argument(0); 2120 Node* n = NULL; 2121 switch (id) { 2122 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2123 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2124 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2125 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2126 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2127 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2128 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2129 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2130 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2131 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2132 default: fatal_unexpected_iid(id); break; 2133 } 2134 set_result(_gvn.transform(n)); 2135 return true; 2136 } 2137 2138 //----------------------------inline_unsafe_access---------------------------- 2139 2140 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2141 2142 // Helper that guards and inserts a pre-barrier. 2143 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2144 Node* pre_val, bool need_mem_bar) { 2145 // We could be accessing the referent field of a reference object. If so, when G1 2146 // is enabled, we need to log the value in the referent field in an SATB buffer. 2147 // This routine performs some compile time filters and generates suitable 2148 // runtime filters that guard the pre-barrier code. 2149 // Also add memory barrier for non volatile load from the referent field 2150 // to prevent commoning of loads across safepoint. 2151 if (!UseG1GC && !need_mem_bar) 2152 return; 2153 2154 // Some compile time checks. 2155 2156 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2157 const TypeX* otype = offset->find_intptr_t_type(); 2158 if (otype != NULL && otype->is_con() && 2159 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2160 // Constant offset but not the reference_offset so just return 2161 return; 2162 } 2163 2164 // We only need to generate the runtime guards for instances. 2165 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2166 if (btype != NULL) { 2167 if (btype->isa_aryptr()) { 2168 // Array type so nothing to do 2169 return; 2170 } 2171 2172 const TypeInstPtr* itype = btype->isa_instptr(); 2173 if (itype != NULL) { 2174 // Can the klass of base_oop be statically determined to be 2175 // _not_ a sub-class of Reference and _not_ Object? 2176 ciKlass* klass = itype->klass(); 2177 if ( klass->is_loaded() && 2178 !klass->is_subtype_of(env()->Reference_klass()) && 2179 !env()->Object_klass()->is_subtype_of(klass)) { 2180 return; 2181 } 2182 } 2183 } 2184 2185 // The compile time filters did not reject base_oop/offset so 2186 // we need to generate the following runtime filters 2187 // 2188 // if (offset == java_lang_ref_Reference::_reference_offset) { 2189 // if (instance_of(base, java.lang.ref.Reference)) { 2190 // pre_barrier(_, pre_val, ...); 2191 // } 2192 // } 2193 2194 float likely = PROB_LIKELY( 0.999); 2195 float unlikely = PROB_UNLIKELY(0.999); 2196 2197 IdealKit ideal(this); 2198 #define __ ideal. 2199 2200 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2201 2202 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2203 // Update graphKit memory and control from IdealKit. 2204 sync_kit(ideal); 2205 2206 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2207 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2208 2209 // Update IdealKit memory and control from graphKit. 2210 __ sync_kit(this); 2211 2212 Node* one = __ ConI(1); 2213 // is_instof == 0 if base_oop == NULL 2214 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2215 2216 // Update graphKit from IdeakKit. 2217 sync_kit(ideal); 2218 2219 // Use the pre-barrier to record the value in the referent field 2220 pre_barrier(false /* do_load */, 2221 __ ctrl(), 2222 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2223 pre_val /* pre_val */, 2224 T_OBJECT); 2225 if (need_mem_bar) { 2226 // Add memory barrier to prevent commoning reads from this field 2227 // across safepoint since GC can change its value. 2228 insert_mem_bar(Op_MemBarCPUOrder); 2229 } 2230 // Update IdealKit from graphKit. 2231 __ sync_kit(this); 2232 2233 } __ end_if(); // _ref_type != ref_none 2234 } __ end_if(); // offset == referent_offset 2235 2236 // Final sync IdealKit and GraphKit. 2237 final_sync(ideal); 2238 #undef __ 2239 } 2240 2241 2242 // Interpret Unsafe.fieldOffset cookies correctly: 2243 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2244 2245 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2246 // Attempt to infer a sharper value type from the offset and base type. 2247 ciKlass* sharpened_klass = NULL; 2248 2249 // See if it is an instance field, with an object type. 2250 if (alias_type->field() != NULL) { 2251 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2252 if (alias_type->field()->type()->is_klass()) { 2253 sharpened_klass = alias_type->field()->type()->as_klass(); 2254 } 2255 } 2256 2257 // See if it is a narrow oop array. 2258 if (adr_type->isa_aryptr()) { 2259 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2260 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2261 if (elem_type != NULL) { 2262 sharpened_klass = elem_type->klass(); 2263 } 2264 } 2265 } 2266 2267 // The sharpened class might be unloaded if there is no class loader 2268 // contraint in place. 2269 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2270 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2271 2272 #ifndef PRODUCT 2273 if (C->print_intrinsics() || C->print_inlining()) { 2274 tty->print(" from base type: "); adr_type->dump(); 2275 tty->print(" sharpened value: "); tjp->dump(); 2276 } 2277 #endif 2278 // Sharpen the value type. 2279 return tjp; 2280 } 2281 return NULL; 2282 } 2283 2284 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) { 2285 if (callee()->is_static()) return false; // caller must have the capability! 2286 2287 #ifndef PRODUCT 2288 { 2289 ResourceMark rm; 2290 // Check the signatures. 2291 ciSignature* sig = callee()->signature(); 2292 #ifdef ASSERT 2293 if (!is_store) { 2294 // Object getObject(Object base, int/long offset), etc. 2295 BasicType rtype = sig->return_type()->basic_type(); 2296 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2297 rtype = T_ADDRESS; // it is really a C void* 2298 assert(rtype == type, "getter must return the expected value"); 2299 if (!is_native_ptr) { 2300 assert(sig->count() == 2, "oop getter has 2 arguments"); 2301 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2302 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2303 } else { 2304 assert(sig->count() == 1, "native getter has 1 argument"); 2305 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2306 } 2307 } else { 2308 // void putObject(Object base, int/long offset, Object x), etc. 2309 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2310 if (!is_native_ptr) { 2311 assert(sig->count() == 3, "oop putter has 3 arguments"); 2312 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2313 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2314 } else { 2315 assert(sig->count() == 2, "native putter has 2 arguments"); 2316 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2317 } 2318 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2319 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2320 vtype = T_ADDRESS; // it is really a C void* 2321 assert(vtype == type, "putter must accept the expected value"); 2322 } 2323 #endif // ASSERT 2324 } 2325 #endif //PRODUCT 2326 2327 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2328 2329 Node* receiver = argument(0); // type: oop 2330 2331 // Build address expression. 2332 Node* adr; 2333 Node* heap_base_oop = top(); 2334 Node* offset = top(); 2335 Node* val; 2336 2337 if (!is_native_ptr) { 2338 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2339 Node* base = argument(1); // type: oop 2340 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2341 offset = argument(2); // type: long 2342 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2343 // to be plain byte offsets, which are also the same as those accepted 2344 // by oopDesc::field_base. 2345 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2346 "fieldOffset must be byte-scaled"); 2347 // 32-bit machines ignore the high half! 2348 offset = ConvL2X(offset); 2349 adr = make_unsafe_address(base, offset); 2350 heap_base_oop = base; 2351 val = is_store ? argument(4) : NULL; 2352 } else { 2353 Node* ptr = argument(1); // type: long 2354 ptr = ConvL2X(ptr); // adjust Java long to machine word 2355 adr = make_unsafe_address(NULL, ptr); 2356 val = is_store ? argument(3) : NULL; 2357 } 2358 2359 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2360 2361 // First guess at the value type. 2362 const Type *value_type = Type::get_const_basic_type(type); 2363 2364 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2365 // there was not enough information to nail it down. 2366 Compile::AliasType* alias_type = C->alias_type(adr_type); 2367 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2368 2369 // We will need memory barriers unless we can determine a unique 2370 // alias category for this reference. (Note: If for some reason 2371 // the barriers get omitted and the unsafe reference begins to "pollute" 2372 // the alias analysis of the rest of the graph, either Compile::can_alias 2373 // or Compile::must_alias will throw a diagnostic assert.) 2374 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2375 2376 // If we are reading the value of the referent field of a Reference 2377 // object (either by using Unsafe directly or through reflection) 2378 // then, if G1 is enabled, we need to record the referent in an 2379 // SATB log buffer using the pre-barrier mechanism. 2380 // Also we need to add memory barrier to prevent commoning reads 2381 // from this field across safepoint since GC can change its value. 2382 bool need_read_barrier = !is_native_ptr && !is_store && 2383 offset != top() && heap_base_oop != top(); 2384 2385 if (!is_store && type == T_OBJECT) { 2386 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2387 if (tjp != NULL) { 2388 value_type = tjp; 2389 } 2390 } 2391 2392 receiver = null_check(receiver); 2393 if (stopped()) { 2394 return true; 2395 } 2396 // Heap pointers get a null-check from the interpreter, 2397 // as a courtesy. However, this is not guaranteed by Unsafe, 2398 // and it is not possible to fully distinguish unintended nulls 2399 // from intended ones in this API. 2400 2401 if (is_volatile) { 2402 // We need to emit leading and trailing CPU membars (see below) in 2403 // addition to memory membars when is_volatile. This is a little 2404 // too strong, but avoids the need to insert per-alias-type 2405 // volatile membars (for stores; compare Parse::do_put_xxx), which 2406 // we cannot do effectively here because we probably only have a 2407 // rough approximation of type. 2408 need_mem_bar = true; 2409 // For Stores, place a memory ordering barrier now. 2410 if (is_store) { 2411 insert_mem_bar(Op_MemBarRelease); 2412 } else { 2413 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2414 insert_mem_bar(Op_MemBarVolatile); 2415 } 2416 } 2417 } 2418 2419 // Memory barrier to prevent normal and 'unsafe' accesses from 2420 // bypassing each other. Happens after null checks, so the 2421 // exception paths do not take memory state from the memory barrier, 2422 // so there's no problems making a strong assert about mixing users 2423 // of safe & unsafe memory. 2424 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2425 2426 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM || 2427 alias_type->field() != NULL || alias_type->element() != NULL, "field, array element or unknown"); 2428 bool mismatched = false; 2429 if (alias_type->element() != NULL || alias_type->field() != NULL) { 2430 BasicType bt; 2431 if (alias_type->element() != NULL) { 2432 const Type* element = alias_type->element(); 2433 bt = element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 2434 } else { 2435 bt = alias_type->field()->type()->basic_type(); 2436 } 2437 if (bt == T_ARRAY) { 2438 // accessing an array field with getObject is not a mismatch 2439 bt = T_OBJECT; 2440 } 2441 if (bt != type) { 2442 mismatched = true; 2443 } 2444 } 2445 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2446 2447 if (!is_store) { 2448 Node* p = NULL; 2449 // Try to constant fold a load from a constant field 2450 ciField* field = alias_type->field(); 2451 if (heap_base_oop != top() && 2452 field != NULL && field->is_constant() && field->layout_type() == type) { 2453 // final or stable field 2454 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop); 2455 if (con_type != NULL) { 2456 p = makecon(con_type); 2457 } 2458 } 2459 if (p == NULL) { 2460 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2461 // To be valid, unsafe loads may depend on other conditions than 2462 // the one that guards them: pin the Load node 2463 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched); 2464 // load value 2465 switch (type) { 2466 case T_BOOLEAN: 2467 case T_CHAR: 2468 case T_BYTE: 2469 case T_SHORT: 2470 case T_INT: 2471 case T_LONG: 2472 case T_FLOAT: 2473 case T_DOUBLE: 2474 break; 2475 case T_OBJECT: 2476 if (need_read_barrier) { 2477 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2478 } 2479 break; 2480 case T_ADDRESS: 2481 // Cast to an int type. 2482 p = _gvn.transform(new CastP2XNode(NULL, p)); 2483 p = ConvX2UL(p); 2484 break; 2485 default: 2486 fatal("unexpected type %d: %s", type, type2name(type)); 2487 break; 2488 } 2489 } 2490 // The load node has the control of the preceding MemBarCPUOrder. All 2491 // following nodes will have the control of the MemBarCPUOrder inserted at 2492 // the end of this method. So, pushing the load onto the stack at a later 2493 // point is fine. 2494 set_result(p); 2495 } else { 2496 // place effect of store into memory 2497 switch (type) { 2498 case T_DOUBLE: 2499 val = dstore_rounding(val); 2500 break; 2501 case T_ADDRESS: 2502 // Repackage the long as a pointer. 2503 val = ConvL2X(val); 2504 val = _gvn.transform(new CastX2PNode(val)); 2505 break; 2506 } 2507 2508 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2509 if (type != T_OBJECT ) { 2510 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched); 2511 } else { 2512 // Possibly an oop being stored to Java heap or native memory 2513 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2514 // oop to Java heap. 2515 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2516 } else { 2517 // We can't tell at compile time if we are storing in the Java heap or outside 2518 // of it. So we need to emit code to conditionally do the proper type of 2519 // store. 2520 2521 IdealKit ideal(this); 2522 #define __ ideal. 2523 // QQQ who knows what probability is here?? 2524 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2525 // Sync IdealKit and graphKit. 2526 sync_kit(ideal); 2527 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2528 // Update IdealKit memory. 2529 __ sync_kit(this); 2530 } __ else_(); { 2531 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile, mismatched); 2532 } __ end_if(); 2533 // Final sync IdealKit and GraphKit. 2534 final_sync(ideal); 2535 #undef __ 2536 } 2537 } 2538 } 2539 2540 if (is_volatile) { 2541 if (!is_store) { 2542 insert_mem_bar(Op_MemBarAcquire); 2543 } else { 2544 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2545 insert_mem_bar(Op_MemBarVolatile); 2546 } 2547 } 2548 } 2549 2550 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2551 2552 return true; 2553 } 2554 2555 //----------------------------inline_unsafe_load_store---------------------------- 2556 // This method serves a couple of different customers (depending on LoadStoreKind): 2557 // 2558 // LS_cmpxchg: 2559 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2560 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2561 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2562 // 2563 // LS_xadd: 2564 // public int getAndAddInt( Object o, long offset, int delta) 2565 // public long getAndAddLong(Object o, long offset, long delta) 2566 // 2567 // LS_xchg: 2568 // int getAndSet(Object o, long offset, int newValue) 2569 // long getAndSet(Object o, long offset, long newValue) 2570 // Object getAndSet(Object o, long offset, Object newValue) 2571 // 2572 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2573 // This basic scheme here is the same as inline_unsafe_access, but 2574 // differs in enough details that combining them would make the code 2575 // overly confusing. (This is a true fact! I originally combined 2576 // them, but even I was confused by it!) As much code/comments as 2577 // possible are retained from inline_unsafe_access though to make 2578 // the correspondences clearer. - dl 2579 2580 if (callee()->is_static()) return false; // caller must have the capability! 2581 2582 #ifndef PRODUCT 2583 BasicType rtype; 2584 { 2585 ResourceMark rm; 2586 // Check the signatures. 2587 ciSignature* sig = callee()->signature(); 2588 rtype = sig->return_type()->basic_type(); 2589 if (kind == LS_xadd || kind == LS_xchg) { 2590 // Check the signatures. 2591 #ifdef ASSERT 2592 assert(rtype == type, "get and set must return the expected type"); 2593 assert(sig->count() == 3, "get and set has 3 arguments"); 2594 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2595 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2596 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2597 #endif // ASSERT 2598 } else if (kind == LS_cmpxchg) { 2599 // Check the signatures. 2600 #ifdef ASSERT 2601 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2602 assert(sig->count() == 4, "CAS has 4 arguments"); 2603 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2604 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2605 #endif // ASSERT 2606 } else { 2607 ShouldNotReachHere(); 2608 } 2609 } 2610 #endif //PRODUCT 2611 2612 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2613 2614 // Get arguments: 2615 Node* receiver = NULL; 2616 Node* base = NULL; 2617 Node* offset = NULL; 2618 Node* oldval = NULL; 2619 Node* newval = NULL; 2620 if (kind == LS_cmpxchg) { 2621 const bool two_slot_type = type2size[type] == 2; 2622 receiver = argument(0); // type: oop 2623 base = argument(1); // type: oop 2624 offset = argument(2); // type: long 2625 oldval = argument(4); // type: oop, int, or long 2626 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2627 } else if (kind == LS_xadd || kind == LS_xchg){ 2628 receiver = argument(0); // type: oop 2629 base = argument(1); // type: oop 2630 offset = argument(2); // type: long 2631 oldval = NULL; 2632 newval = argument(4); // type: oop, int, or long 2633 } 2634 2635 // Null check receiver. 2636 receiver = null_check(receiver); 2637 if (stopped()) { 2638 return true; 2639 } 2640 2641 // Build field offset expression. 2642 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2643 // to be plain byte offsets, which are also the same as those accepted 2644 // by oopDesc::field_base. 2645 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2646 // 32-bit machines ignore the high half of long offsets 2647 offset = ConvL2X(offset); 2648 Node* adr = make_unsafe_address(base, offset); 2649 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2650 2651 // For CAS, unlike inline_unsafe_access, there seems no point in 2652 // trying to refine types. Just use the coarse types here. 2653 const Type *value_type = Type::get_const_basic_type(type); 2654 Compile::AliasType* alias_type = C->alias_type(adr_type); 2655 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2656 2657 if (kind == LS_xchg && type == T_OBJECT) { 2658 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2659 if (tjp != NULL) { 2660 value_type = tjp; 2661 } 2662 } 2663 2664 int alias_idx = C->get_alias_index(adr_type); 2665 2666 // Memory-model-wise, a LoadStore acts like a little synchronized 2667 // block, so needs barriers on each side. These don't translate 2668 // into actual barriers on most machines, but we still need rest of 2669 // compiler to respect ordering. 2670 2671 insert_mem_bar(Op_MemBarRelease); 2672 insert_mem_bar(Op_MemBarCPUOrder); 2673 2674 // 4984716: MemBars must be inserted before this 2675 // memory node in order to avoid a false 2676 // dependency which will confuse the scheduler. 2677 Node *mem = memory(alias_idx); 2678 2679 // For now, we handle only those cases that actually exist: ints, 2680 // longs, and Object. Adding others should be straightforward. 2681 Node* load_store = NULL; 2682 switch(type) { 2683 case T_INT: 2684 if (kind == LS_xadd) { 2685 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2686 } else if (kind == LS_xchg) { 2687 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2688 } else if (kind == LS_cmpxchg) { 2689 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2690 } else { 2691 ShouldNotReachHere(); 2692 } 2693 break; 2694 case T_LONG: 2695 if (kind == LS_xadd) { 2696 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2697 } else if (kind == LS_xchg) { 2698 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2699 } else if (kind == LS_cmpxchg) { 2700 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2701 } else { 2702 ShouldNotReachHere(); 2703 } 2704 break; 2705 case T_OBJECT: 2706 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2707 // could be delayed during Parse (for example, in adjust_map_after_if()). 2708 // Execute transformation here to avoid barrier generation in such case. 2709 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2710 newval = _gvn.makecon(TypePtr::NULL_PTR); 2711 2712 // Reference stores need a store barrier. 2713 if (kind == LS_xchg) { 2714 // If pre-barrier must execute before the oop store, old value will require do_load here. 2715 if (!can_move_pre_barrier()) { 2716 pre_barrier(true /* do_load*/, 2717 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2718 NULL /* pre_val*/, 2719 T_OBJECT); 2720 } // Else move pre_barrier to use load_store value, see below. 2721 } else if (kind == LS_cmpxchg) { 2722 // Same as for newval above: 2723 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2724 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2725 } 2726 // The only known value which might get overwritten is oldval. 2727 pre_barrier(false /* do_load */, 2728 control(), NULL, NULL, max_juint, NULL, NULL, 2729 oldval /* pre_val */, 2730 T_OBJECT); 2731 } else { 2732 ShouldNotReachHere(); 2733 } 2734 2735 #ifdef _LP64 2736 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2737 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2738 if (kind == LS_xchg) { 2739 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2740 newval_enc, adr_type, value_type->make_narrowoop())); 2741 } else { 2742 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2743 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2744 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2745 newval_enc, oldval_enc)); 2746 } 2747 } else 2748 #endif 2749 { 2750 if (kind == LS_xchg) { 2751 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2752 } else { 2753 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2754 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2755 } 2756 } 2757 if (kind == LS_cmpxchg) { 2758 // Emit the post barrier only when the actual store happened. 2759 // This makes sense to check only for compareAndSet that can fail to set the value. 2760 // CAS success path is marked more likely since we anticipate this is a performance 2761 // critical path, while CAS failure path can use the penalty for going through unlikely 2762 // path as backoff. Which is still better than doing a store barrier there. 2763 IdealKit ideal(this); 2764 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); { 2765 sync_kit(ideal); 2766 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2767 ideal.sync_kit(this); 2768 } ideal.end_if(); 2769 final_sync(ideal); 2770 } else { 2771 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2772 } 2773 break; 2774 default: 2775 fatal("unexpected type %d: %s", type, type2name(type)); 2776 break; 2777 } 2778 2779 // SCMemProjNodes represent the memory state of a LoadStore. Their 2780 // main role is to prevent LoadStore nodes from being optimized away 2781 // when their results aren't used. 2782 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2783 set_memory(proj, alias_idx); 2784 2785 if (type == T_OBJECT && kind == LS_xchg) { 2786 #ifdef _LP64 2787 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2788 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2789 } 2790 #endif 2791 if (can_move_pre_barrier()) { 2792 // Don't need to load pre_val. The old value is returned by load_store. 2793 // The pre_barrier can execute after the xchg as long as no safepoint 2794 // gets inserted between them. 2795 pre_barrier(false /* do_load */, 2796 control(), NULL, NULL, max_juint, NULL, NULL, 2797 load_store /* pre_val */, 2798 T_OBJECT); 2799 } 2800 } 2801 2802 // Add the trailing membar surrounding the access 2803 insert_mem_bar(Op_MemBarCPUOrder); 2804 insert_mem_bar(Op_MemBarAcquire); 2805 2806 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2807 set_result(load_store); 2808 return true; 2809 } 2810 2811 //----------------------------inline_unsafe_ordered_store---------------------- 2812 // public native void Unsafe.putOrderedObject(Object o, long offset, Object x); 2813 // public native void Unsafe.putOrderedInt(Object o, long offset, int x); 2814 // public native void Unsafe.putOrderedLong(Object o, long offset, long x); 2815 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2816 // This is another variant of inline_unsafe_access, differing in 2817 // that it always issues store-store ("release") barrier and ensures 2818 // store-atomicity (which only matters for "long"). 2819 2820 if (callee()->is_static()) return false; // caller must have the capability! 2821 2822 #ifndef PRODUCT 2823 { 2824 ResourceMark rm; 2825 // Check the signatures. 2826 ciSignature* sig = callee()->signature(); 2827 #ifdef ASSERT 2828 BasicType rtype = sig->return_type()->basic_type(); 2829 assert(rtype == T_VOID, "must return void"); 2830 assert(sig->count() == 3, "has 3 arguments"); 2831 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2832 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2833 #endif // ASSERT 2834 } 2835 #endif //PRODUCT 2836 2837 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2838 2839 // Get arguments: 2840 Node* receiver = argument(0); // type: oop 2841 Node* base = argument(1); // type: oop 2842 Node* offset = argument(2); // type: long 2843 Node* val = argument(4); // type: oop, int, or long 2844 2845 // Null check receiver. 2846 receiver = null_check(receiver); 2847 if (stopped()) { 2848 return true; 2849 } 2850 2851 // Build field offset expression. 2852 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2853 // 32-bit machines ignore the high half of long offsets 2854 offset = ConvL2X(offset); 2855 Node* adr = make_unsafe_address(base, offset); 2856 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2857 const Type *value_type = Type::get_const_basic_type(type); 2858 Compile::AliasType* alias_type = C->alias_type(adr_type); 2859 2860 insert_mem_bar(Op_MemBarRelease); 2861 insert_mem_bar(Op_MemBarCPUOrder); 2862 // Ensure that the store is atomic for longs: 2863 const bool require_atomic_access = true; 2864 Node* store; 2865 if (type == T_OBJECT) // reference stores need a store barrier. 2866 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 2867 else { 2868 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 2869 } 2870 insert_mem_bar(Op_MemBarCPUOrder); 2871 return true; 2872 } 2873 2874 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2875 // Regardless of form, don't allow previous ld/st to move down, 2876 // then issue acquire, release, or volatile mem_bar. 2877 insert_mem_bar(Op_MemBarCPUOrder); 2878 switch(id) { 2879 case vmIntrinsics::_loadFence: 2880 insert_mem_bar(Op_LoadFence); 2881 return true; 2882 case vmIntrinsics::_storeFence: 2883 insert_mem_bar(Op_StoreFence); 2884 return true; 2885 case vmIntrinsics::_fullFence: 2886 insert_mem_bar(Op_MemBarVolatile); 2887 return true; 2888 default: 2889 fatal_unexpected_iid(id); 2890 return false; 2891 } 2892 } 2893 2894 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 2895 if (!kls->is_Con()) { 2896 return true; 2897 } 2898 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 2899 if (klsptr == NULL) { 2900 return true; 2901 } 2902 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 2903 // don't need a guard for a klass that is already initialized 2904 return !ik->is_initialized(); 2905 } 2906 2907 //----------------------------inline_unsafe_allocate--------------------------- 2908 // public native Object Unsafe.allocateInstance(Class<?> cls); 2909 bool LibraryCallKit::inline_unsafe_allocate() { 2910 if (callee()->is_static()) return false; // caller must have the capability! 2911 2912 null_check_receiver(); // null-check, then ignore 2913 Node* cls = null_check(argument(1)); 2914 if (stopped()) return true; 2915 2916 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2917 kls = null_check(kls); 2918 if (stopped()) return true; // argument was like int.class 2919 2920 Node* test = NULL; 2921 if (LibraryCallKit::klass_needs_init_guard(kls)) { 2922 // Note: The argument might still be an illegal value like 2923 // Serializable.class or Object[].class. The runtime will handle it. 2924 // But we must make an explicit check for initialization. 2925 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 2926 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 2927 // can generate code to load it as unsigned byte. 2928 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 2929 Node* bits = intcon(InstanceKlass::fully_initialized); 2930 test = _gvn.transform(new SubINode(inst, bits)); 2931 // The 'test' is non-zero if we need to take a slow path. 2932 } 2933 2934 Node* obj = new_instance(kls, test); 2935 set_result(obj); 2936 return true; 2937 } 2938 2939 #ifdef TRACE_HAVE_INTRINSICS 2940 /* 2941 * oop -> myklass 2942 * myklass->trace_id |= USED 2943 * return myklass->trace_id & ~0x3 2944 */ 2945 bool LibraryCallKit::inline_native_classID() { 2946 null_check_receiver(); // null-check, then ignore 2947 Node* cls = null_check(argument(1), T_OBJECT); 2948 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2949 kls = null_check(kls, T_OBJECT); 2950 ByteSize offset = TRACE_ID_OFFSET; 2951 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 2952 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 2953 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 2954 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 2955 Node* clsused = longcon(0x01l); // set the class bit 2956 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 2957 2958 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 2959 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 2960 set_result(andl); 2961 return true; 2962 } 2963 2964 bool LibraryCallKit::inline_native_threadID() { 2965 Node* tls_ptr = NULL; 2966 Node* cur_thr = generate_current_thread(tls_ptr); 2967 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2968 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 2969 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 2970 2971 Node* threadid = NULL; 2972 size_t thread_id_size = OSThread::thread_id_size(); 2973 if (thread_id_size == (size_t) BytesPerLong) { 2974 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 2975 } else if (thread_id_size == (size_t) BytesPerInt) { 2976 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 2977 } else { 2978 ShouldNotReachHere(); 2979 } 2980 set_result(threadid); 2981 return true; 2982 } 2983 #endif 2984 2985 //------------------------inline_native_time_funcs-------------- 2986 // inline code for System.currentTimeMillis() and System.nanoTime() 2987 // these have the same type and signature 2988 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 2989 const TypeFunc* tf = OptoRuntime::void_long_Type(); 2990 const TypePtr* no_memory_effects = NULL; 2991 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2992 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 2993 #ifdef ASSERT 2994 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 2995 assert(value_top == top(), "second value must be top"); 2996 #endif 2997 set_result(value); 2998 return true; 2999 } 3000 3001 //------------------------inline_native_currentThread------------------ 3002 bool LibraryCallKit::inline_native_currentThread() { 3003 Node* junk = NULL; 3004 set_result(generate_current_thread(junk)); 3005 return true; 3006 } 3007 3008 //------------------------inline_native_isInterrupted------------------ 3009 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3010 bool LibraryCallKit::inline_native_isInterrupted() { 3011 // Add a fast path to t.isInterrupted(clear_int): 3012 // (t == Thread.current() && 3013 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3014 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3015 // So, in the common case that the interrupt bit is false, 3016 // we avoid making a call into the VM. Even if the interrupt bit 3017 // is true, if the clear_int argument is false, we avoid the VM call. 3018 // However, if the receiver is not currentThread, we must call the VM, 3019 // because there must be some locking done around the operation. 3020 3021 // We only go to the fast case code if we pass two guards. 3022 // Paths which do not pass are accumulated in the slow_region. 3023 3024 enum { 3025 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3026 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3027 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3028 PATH_LIMIT 3029 }; 3030 3031 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3032 // out of the function. 3033 insert_mem_bar(Op_MemBarCPUOrder); 3034 3035 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3036 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3037 3038 RegionNode* slow_region = new RegionNode(1); 3039 record_for_igvn(slow_region); 3040 3041 // (a) Receiving thread must be the current thread. 3042 Node* rec_thr = argument(0); 3043 Node* tls_ptr = NULL; 3044 Node* cur_thr = generate_current_thread(tls_ptr); 3045 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3046 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3047 3048 generate_slow_guard(bol_thr, slow_region); 3049 3050 // (b) Interrupt bit on TLS must be false. 3051 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3052 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3053 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3054 3055 // Set the control input on the field _interrupted read to prevent it floating up. 3056 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3057 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3058 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3059 3060 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3061 3062 // First fast path: if (!TLS._interrupted) return false; 3063 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3064 result_rgn->init_req(no_int_result_path, false_bit); 3065 result_val->init_req(no_int_result_path, intcon(0)); 3066 3067 // drop through to next case 3068 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3069 3070 #ifndef TARGET_OS_FAMILY_windows 3071 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3072 Node* clr_arg = argument(1); 3073 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3074 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3075 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3076 3077 // Second fast path: ... else if (!clear_int) return true; 3078 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3079 result_rgn->init_req(no_clear_result_path, false_arg); 3080 result_val->init_req(no_clear_result_path, intcon(1)); 3081 3082 // drop through to next case 3083 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3084 #else 3085 // To return true on Windows you must read the _interrupted field 3086 // and check the event state i.e. take the slow path. 3087 #endif // TARGET_OS_FAMILY_windows 3088 3089 // (d) Otherwise, go to the slow path. 3090 slow_region->add_req(control()); 3091 set_control( _gvn.transform(slow_region)); 3092 3093 if (stopped()) { 3094 // There is no slow path. 3095 result_rgn->init_req(slow_result_path, top()); 3096 result_val->init_req(slow_result_path, top()); 3097 } else { 3098 // non-virtual because it is a private non-static 3099 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3100 3101 Node* slow_val = set_results_for_java_call(slow_call); 3102 // this->control() comes from set_results_for_java_call 3103 3104 Node* fast_io = slow_call->in(TypeFunc::I_O); 3105 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3106 3107 // These two phis are pre-filled with copies of of the fast IO and Memory 3108 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3109 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3110 3111 result_rgn->init_req(slow_result_path, control()); 3112 result_io ->init_req(slow_result_path, i_o()); 3113 result_mem->init_req(slow_result_path, reset_memory()); 3114 result_val->init_req(slow_result_path, slow_val); 3115 3116 set_all_memory(_gvn.transform(result_mem)); 3117 set_i_o( _gvn.transform(result_io)); 3118 } 3119 3120 C->set_has_split_ifs(true); // Has chance for split-if optimization 3121 set_result(result_rgn, result_val); 3122 return true; 3123 } 3124 3125 //---------------------------load_mirror_from_klass---------------------------- 3126 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3127 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3128 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3129 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3130 } 3131 3132 //-----------------------load_klass_from_mirror_common------------------------- 3133 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3134 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3135 // and branch to the given path on the region. 3136 // If never_see_null, take an uncommon trap on null, so we can optimistically 3137 // compile for the non-null case. 3138 // If the region is NULL, force never_see_null = true. 3139 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3140 bool never_see_null, 3141 RegionNode* region, 3142 int null_path, 3143 int offset) { 3144 if (region == NULL) never_see_null = true; 3145 Node* p = basic_plus_adr(mirror, offset); 3146 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3147 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3148 Node* null_ctl = top(); 3149 kls = null_check_oop(kls, &null_ctl, never_see_null); 3150 if (region != NULL) { 3151 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3152 region->init_req(null_path, null_ctl); 3153 } else { 3154 assert(null_ctl == top(), "no loose ends"); 3155 } 3156 return kls; 3157 } 3158 3159 //--------------------(inline_native_Class_query helpers)--------------------- 3160 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3161 // Fall through if (mods & mask) == bits, take the guard otherwise. 3162 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3163 // Branch around if the given klass has the given modifier bit set. 3164 // Like generate_guard, adds a new path onto the region. 3165 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3166 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3167 Node* mask = intcon(modifier_mask); 3168 Node* bits = intcon(modifier_bits); 3169 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3170 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3171 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3172 return generate_fair_guard(bol, region); 3173 } 3174 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3175 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3176 } 3177 3178 //-------------------------inline_native_Class_query------------------- 3179 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3180 const Type* return_type = TypeInt::BOOL; 3181 Node* prim_return_value = top(); // what happens if it's a primitive class? 3182 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3183 bool expect_prim = false; // most of these guys expect to work on refs 3184 3185 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3186 3187 Node* mirror = argument(0); 3188 Node* obj = top(); 3189 3190 switch (id) { 3191 case vmIntrinsics::_isInstance: 3192 // nothing is an instance of a primitive type 3193 prim_return_value = intcon(0); 3194 obj = argument(1); 3195 break; 3196 case vmIntrinsics::_getModifiers: 3197 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3198 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3199 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3200 break; 3201 case vmIntrinsics::_isInterface: 3202 prim_return_value = intcon(0); 3203 break; 3204 case vmIntrinsics::_isArray: 3205 prim_return_value = intcon(0); 3206 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3207 break; 3208 case vmIntrinsics::_isPrimitive: 3209 prim_return_value = intcon(1); 3210 expect_prim = true; // obviously 3211 break; 3212 case vmIntrinsics::_getSuperclass: 3213 prim_return_value = null(); 3214 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3215 break; 3216 case vmIntrinsics::_getClassAccessFlags: 3217 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3218 return_type = TypeInt::INT; // not bool! 6297094 3219 break; 3220 default: 3221 fatal_unexpected_iid(id); 3222 break; 3223 } 3224 3225 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3226 if (mirror_con == NULL) return false; // cannot happen? 3227 3228 #ifndef PRODUCT 3229 if (C->print_intrinsics() || C->print_inlining()) { 3230 ciType* k = mirror_con->java_mirror_type(); 3231 if (k) { 3232 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3233 k->print_name(); 3234 tty->cr(); 3235 } 3236 } 3237 #endif 3238 3239 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3240 RegionNode* region = new RegionNode(PATH_LIMIT); 3241 record_for_igvn(region); 3242 PhiNode* phi = new PhiNode(region, return_type); 3243 3244 // The mirror will never be null of Reflection.getClassAccessFlags, however 3245 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3246 // if it is. See bug 4774291. 3247 3248 // For Reflection.getClassAccessFlags(), the null check occurs in 3249 // the wrong place; see inline_unsafe_access(), above, for a similar 3250 // situation. 3251 mirror = null_check(mirror); 3252 // If mirror or obj is dead, only null-path is taken. 3253 if (stopped()) return true; 3254 3255 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3256 3257 // Now load the mirror's klass metaobject, and null-check it. 3258 // Side-effects region with the control path if the klass is null. 3259 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3260 // If kls is null, we have a primitive mirror. 3261 phi->init_req(_prim_path, prim_return_value); 3262 if (stopped()) { set_result(region, phi); return true; } 3263 bool safe_for_replace = (region->in(_prim_path) == top()); 3264 3265 Node* p; // handy temp 3266 Node* null_ctl; 3267 3268 // Now that we have the non-null klass, we can perform the real query. 3269 // For constant classes, the query will constant-fold in LoadNode::Value. 3270 Node* query_value = top(); 3271 switch (id) { 3272 case vmIntrinsics::_isInstance: 3273 // nothing is an instance of a primitive type 3274 query_value = gen_instanceof(obj, kls, safe_for_replace); 3275 break; 3276 3277 case vmIntrinsics::_getModifiers: 3278 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3279 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3280 break; 3281 3282 case vmIntrinsics::_isInterface: 3283 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3284 if (generate_interface_guard(kls, region) != NULL) 3285 // A guard was added. If the guard is taken, it was an interface. 3286 phi->add_req(intcon(1)); 3287 // If we fall through, it's a plain class. 3288 query_value = intcon(0); 3289 break; 3290 3291 case vmIntrinsics::_isArray: 3292 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3293 if (generate_array_guard(kls, region) != NULL) 3294 // A guard was added. If the guard is taken, it was an array. 3295 phi->add_req(intcon(1)); 3296 // If we fall through, it's a plain class. 3297 query_value = intcon(0); 3298 break; 3299 3300 case vmIntrinsics::_isPrimitive: 3301 query_value = intcon(0); // "normal" path produces false 3302 break; 3303 3304 case vmIntrinsics::_getSuperclass: 3305 // The rules here are somewhat unfortunate, but we can still do better 3306 // with random logic than with a JNI call. 3307 // Interfaces store null or Object as _super, but must report null. 3308 // Arrays store an intermediate super as _super, but must report Object. 3309 // Other types can report the actual _super. 3310 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3311 if (generate_interface_guard(kls, region) != NULL) 3312 // A guard was added. If the guard is taken, it was an interface. 3313 phi->add_req(null()); 3314 if (generate_array_guard(kls, region) != NULL) 3315 // A guard was added. If the guard is taken, it was an array. 3316 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3317 // If we fall through, it's a plain class. Get its _super. 3318 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3319 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3320 null_ctl = top(); 3321 kls = null_check_oop(kls, &null_ctl); 3322 if (null_ctl != top()) { 3323 // If the guard is taken, Object.superClass is null (both klass and mirror). 3324 region->add_req(null_ctl); 3325 phi ->add_req(null()); 3326 } 3327 if (!stopped()) { 3328 query_value = load_mirror_from_klass(kls); 3329 } 3330 break; 3331 3332 case vmIntrinsics::_getClassAccessFlags: 3333 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3334 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3335 break; 3336 3337 default: 3338 fatal_unexpected_iid(id); 3339 break; 3340 } 3341 3342 // Fall-through is the normal case of a query to a real class. 3343 phi->init_req(1, query_value); 3344 region->init_req(1, control()); 3345 3346 C->set_has_split_ifs(true); // Has chance for split-if optimization 3347 set_result(region, phi); 3348 return true; 3349 } 3350 3351 //-------------------------inline_Class_cast------------------- 3352 bool LibraryCallKit::inline_Class_cast() { 3353 Node* mirror = argument(0); // Class 3354 Node* obj = argument(1); 3355 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3356 if (mirror_con == NULL) { 3357 return false; // dead path (mirror->is_top()). 3358 } 3359 if (obj == NULL || obj->is_top()) { 3360 return false; // dead path 3361 } 3362 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3363 3364 // First, see if Class.cast() can be folded statically. 3365 // java_mirror_type() returns non-null for compile-time Class constants. 3366 ciType* tm = mirror_con->java_mirror_type(); 3367 if (tm != NULL && tm->is_klass() && 3368 tp != NULL && tp->klass() != NULL) { 3369 if (!tp->klass()->is_loaded()) { 3370 // Don't use intrinsic when class is not loaded. 3371 return false; 3372 } else { 3373 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3374 if (static_res == Compile::SSC_always_true) { 3375 // isInstance() is true - fold the code. 3376 set_result(obj); 3377 return true; 3378 } else if (static_res == Compile::SSC_always_false) { 3379 // Don't use intrinsic, have to throw ClassCastException. 3380 // If the reference is null, the non-intrinsic bytecode will 3381 // be optimized appropriately. 3382 return false; 3383 } 3384 } 3385 } 3386 3387 // Bailout intrinsic and do normal inlining if exception path is frequent. 3388 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3389 return false; 3390 } 3391 3392 // Generate dynamic checks. 3393 // Class.cast() is java implementation of _checkcast bytecode. 3394 // Do checkcast (Parse::do_checkcast()) optimizations here. 3395 3396 mirror = null_check(mirror); 3397 // If mirror is dead, only null-path is taken. 3398 if (stopped()) { 3399 return true; 3400 } 3401 3402 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3403 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3404 RegionNode* region = new RegionNode(PATH_LIMIT); 3405 record_for_igvn(region); 3406 3407 // Now load the mirror's klass metaobject, and null-check it. 3408 // If kls is null, we have a primitive mirror and 3409 // nothing is an instance of a primitive type. 3410 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3411 3412 Node* res = top(); 3413 if (!stopped()) { 3414 Node* bad_type_ctrl = top(); 3415 // Do checkcast optimizations. 3416 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3417 region->init_req(_bad_type_path, bad_type_ctrl); 3418 } 3419 if (region->in(_prim_path) != top() || 3420 region->in(_bad_type_path) != top()) { 3421 // Let Interpreter throw ClassCastException. 3422 PreserveJVMState pjvms(this); 3423 set_control(_gvn.transform(region)); 3424 uncommon_trap(Deoptimization::Reason_intrinsic, 3425 Deoptimization::Action_maybe_recompile); 3426 } 3427 if (!stopped()) { 3428 set_result(res); 3429 } 3430 return true; 3431 } 3432 3433 3434 //--------------------------inline_native_subtype_check------------------------ 3435 // This intrinsic takes the JNI calls out of the heart of 3436 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3437 bool LibraryCallKit::inline_native_subtype_check() { 3438 // Pull both arguments off the stack. 3439 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3440 args[0] = argument(0); 3441 args[1] = argument(1); 3442 Node* klasses[2]; // corresponding Klasses: superk, subk 3443 klasses[0] = klasses[1] = top(); 3444 3445 enum { 3446 // A full decision tree on {superc is prim, subc is prim}: 3447 _prim_0_path = 1, // {P,N} => false 3448 // {P,P} & superc!=subc => false 3449 _prim_same_path, // {P,P} & superc==subc => true 3450 _prim_1_path, // {N,P} => false 3451 _ref_subtype_path, // {N,N} & subtype check wins => true 3452 _both_ref_path, // {N,N} & subtype check loses => false 3453 PATH_LIMIT 3454 }; 3455 3456 RegionNode* region = new RegionNode(PATH_LIMIT); 3457 Node* phi = new PhiNode(region, TypeInt::BOOL); 3458 record_for_igvn(region); 3459 3460 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3461 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3462 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3463 3464 // First null-check both mirrors and load each mirror's klass metaobject. 3465 int which_arg; 3466 for (which_arg = 0; which_arg <= 1; which_arg++) { 3467 Node* arg = args[which_arg]; 3468 arg = null_check(arg); 3469 if (stopped()) break; 3470 args[which_arg] = arg; 3471 3472 Node* p = basic_plus_adr(arg, class_klass_offset); 3473 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3474 klasses[which_arg] = _gvn.transform(kls); 3475 } 3476 3477 // Having loaded both klasses, test each for null. 3478 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3479 for (which_arg = 0; which_arg <= 1; which_arg++) { 3480 Node* kls = klasses[which_arg]; 3481 Node* null_ctl = top(); 3482 kls = null_check_oop(kls, &null_ctl, never_see_null); 3483 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3484 region->init_req(prim_path, null_ctl); 3485 if (stopped()) break; 3486 klasses[which_arg] = kls; 3487 } 3488 3489 if (!stopped()) { 3490 // now we have two reference types, in klasses[0..1] 3491 Node* subk = klasses[1]; // the argument to isAssignableFrom 3492 Node* superk = klasses[0]; // the receiver 3493 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3494 // now we have a successful reference subtype check 3495 region->set_req(_ref_subtype_path, control()); 3496 } 3497 3498 // If both operands are primitive (both klasses null), then 3499 // we must return true when they are identical primitives. 3500 // It is convenient to test this after the first null klass check. 3501 set_control(region->in(_prim_0_path)); // go back to first null check 3502 if (!stopped()) { 3503 // Since superc is primitive, make a guard for the superc==subc case. 3504 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3505 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3506 generate_guard(bol_eq, region, PROB_FAIR); 3507 if (region->req() == PATH_LIMIT+1) { 3508 // A guard was added. If the added guard is taken, superc==subc. 3509 region->swap_edges(PATH_LIMIT, _prim_same_path); 3510 region->del_req(PATH_LIMIT); 3511 } 3512 region->set_req(_prim_0_path, control()); // Not equal after all. 3513 } 3514 3515 // these are the only paths that produce 'true': 3516 phi->set_req(_prim_same_path, intcon(1)); 3517 phi->set_req(_ref_subtype_path, intcon(1)); 3518 3519 // pull together the cases: 3520 assert(region->req() == PATH_LIMIT, "sane region"); 3521 for (uint i = 1; i < region->req(); i++) { 3522 Node* ctl = region->in(i); 3523 if (ctl == NULL || ctl == top()) { 3524 region->set_req(i, top()); 3525 phi ->set_req(i, top()); 3526 } else if (phi->in(i) == NULL) { 3527 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3528 } 3529 } 3530 3531 set_control(_gvn.transform(region)); 3532 set_result(_gvn.transform(phi)); 3533 return true; 3534 } 3535 3536 //---------------------generate_array_guard_common------------------------ 3537 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3538 bool obj_array, bool not_array) { 3539 3540 if (stopped()) { 3541 return NULL; 3542 } 3543 3544 // If obj_array/non_array==false/false: 3545 // Branch around if the given klass is in fact an array (either obj or prim). 3546 // If obj_array/non_array==false/true: 3547 // Branch around if the given klass is not an array klass of any kind. 3548 // If obj_array/non_array==true/true: 3549 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3550 // If obj_array/non_array==true/false: 3551 // Branch around if the kls is an oop array (Object[] or subtype) 3552 // 3553 // Like generate_guard, adds a new path onto the region. 3554 jint layout_con = 0; 3555 Node* layout_val = get_layout_helper(kls, layout_con); 3556 if (layout_val == NULL) { 3557 bool query = (obj_array 3558 ? Klass::layout_helper_is_objArray(layout_con) 3559 : Klass::layout_helper_is_array(layout_con)); 3560 if (query == not_array) { 3561 return NULL; // never a branch 3562 } else { // always a branch 3563 Node* always_branch = control(); 3564 if (region != NULL) 3565 region->add_req(always_branch); 3566 set_control(top()); 3567 return always_branch; 3568 } 3569 } 3570 // Now test the correct condition. 3571 jint nval = (obj_array 3572 ? ((jint)Klass::_lh_array_tag_type_value 3573 << Klass::_lh_array_tag_shift) 3574 : Klass::_lh_neutral_value); 3575 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3576 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3577 // invert the test if we are looking for a non-array 3578 if (not_array) btest = BoolTest(btest).negate(); 3579 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3580 return generate_fair_guard(bol, region); 3581 } 3582 3583 3584 //-----------------------inline_native_newArray-------------------------- 3585 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3586 bool LibraryCallKit::inline_native_newArray() { 3587 Node* mirror = argument(0); 3588 Node* count_val = argument(1); 3589 3590 mirror = null_check(mirror); 3591 // If mirror or obj is dead, only null-path is taken. 3592 if (stopped()) return true; 3593 3594 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3595 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3596 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3597 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3598 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3599 3600 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3601 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3602 result_reg, _slow_path); 3603 Node* normal_ctl = control(); 3604 Node* no_array_ctl = result_reg->in(_slow_path); 3605 3606 // Generate code for the slow case. We make a call to newArray(). 3607 set_control(no_array_ctl); 3608 if (!stopped()) { 3609 // Either the input type is void.class, or else the 3610 // array klass has not yet been cached. Either the 3611 // ensuing call will throw an exception, or else it 3612 // will cache the array klass for next time. 3613 PreserveJVMState pjvms(this); 3614 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3615 Node* slow_result = set_results_for_java_call(slow_call); 3616 // this->control() comes from set_results_for_java_call 3617 result_reg->set_req(_slow_path, control()); 3618 result_val->set_req(_slow_path, slow_result); 3619 result_io ->set_req(_slow_path, i_o()); 3620 result_mem->set_req(_slow_path, reset_memory()); 3621 } 3622 3623 set_control(normal_ctl); 3624 if (!stopped()) { 3625 // Normal case: The array type has been cached in the java.lang.Class. 3626 // The following call works fine even if the array type is polymorphic. 3627 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3628 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3629 result_reg->init_req(_normal_path, control()); 3630 result_val->init_req(_normal_path, obj); 3631 result_io ->init_req(_normal_path, i_o()); 3632 result_mem->init_req(_normal_path, reset_memory()); 3633 } 3634 3635 // Return the combined state. 3636 set_i_o( _gvn.transform(result_io) ); 3637 set_all_memory( _gvn.transform(result_mem)); 3638 3639 C->set_has_split_ifs(true); // Has chance for split-if optimization 3640 set_result(result_reg, result_val); 3641 return true; 3642 } 3643 3644 //----------------------inline_native_getLength-------------------------- 3645 // public static native int java.lang.reflect.Array.getLength(Object array); 3646 bool LibraryCallKit::inline_native_getLength() { 3647 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3648 3649 Node* array = null_check(argument(0)); 3650 // If array is dead, only null-path is taken. 3651 if (stopped()) return true; 3652 3653 // Deoptimize if it is a non-array. 3654 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3655 3656 if (non_array != NULL) { 3657 PreserveJVMState pjvms(this); 3658 set_control(non_array); 3659 uncommon_trap(Deoptimization::Reason_intrinsic, 3660 Deoptimization::Action_maybe_recompile); 3661 } 3662 3663 // If control is dead, only non-array-path is taken. 3664 if (stopped()) return true; 3665 3666 // The works fine even if the array type is polymorphic. 3667 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3668 Node* result = load_array_length(array); 3669 3670 C->set_has_split_ifs(true); // Has chance for split-if optimization 3671 set_result(result); 3672 return true; 3673 } 3674 3675 //------------------------inline_array_copyOf---------------------------- 3676 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3677 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3678 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3679 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3680 3681 // Get the arguments. 3682 Node* original = argument(0); 3683 Node* start = is_copyOfRange? argument(1): intcon(0); 3684 Node* end = is_copyOfRange? argument(2): argument(1); 3685 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3686 3687 Node* newcopy = NULL; 3688 3689 // Set the original stack and the reexecute bit for the interpreter to reexecute 3690 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3691 { PreserveReexecuteState preexecs(this); 3692 jvms()->set_should_reexecute(true); 3693 3694 array_type_mirror = null_check(array_type_mirror); 3695 original = null_check(original); 3696 3697 // Check if a null path was taken unconditionally. 3698 if (stopped()) return true; 3699 3700 Node* orig_length = load_array_length(original); 3701 3702 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3703 klass_node = null_check(klass_node); 3704 3705 RegionNode* bailout = new RegionNode(1); 3706 record_for_igvn(bailout); 3707 3708 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3709 // Bail out if that is so. 3710 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3711 if (not_objArray != NULL) { 3712 // Improve the klass node's type from the new optimistic assumption: 3713 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3714 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3715 Node* cast = new CastPPNode(klass_node, akls); 3716 cast->init_req(0, control()); 3717 klass_node = _gvn.transform(cast); 3718 } 3719 3720 // Bail out if either start or end is negative. 3721 generate_negative_guard(start, bailout, &start); 3722 generate_negative_guard(end, bailout, &end); 3723 3724 Node* length = end; 3725 if (_gvn.type(start) != TypeInt::ZERO) { 3726 length = _gvn.transform(new SubINode(end, start)); 3727 } 3728 3729 // Bail out if length is negative. 3730 // Without this the new_array would throw 3731 // NegativeArraySizeException but IllegalArgumentException is what 3732 // should be thrown 3733 generate_negative_guard(length, bailout, &length); 3734 3735 if (bailout->req() > 1) { 3736 PreserveJVMState pjvms(this); 3737 set_control(_gvn.transform(bailout)); 3738 uncommon_trap(Deoptimization::Reason_intrinsic, 3739 Deoptimization::Action_maybe_recompile); 3740 } 3741 3742 if (!stopped()) { 3743 // How many elements will we copy from the original? 3744 // The answer is MinI(orig_length - start, length). 3745 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3746 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3747 3748 // Generate a direct call to the right arraycopy function(s). 3749 // We know the copy is disjoint but we might not know if the 3750 // oop stores need checking. 3751 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3752 // This will fail a store-check if x contains any non-nulls. 3753 3754 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3755 // loads/stores but it is legal only if we're sure the 3756 // Arrays.copyOf would succeed. So we need all input arguments 3757 // to the copyOf to be validated, including that the copy to the 3758 // new array won't trigger an ArrayStoreException. That subtype 3759 // check can be optimized if we know something on the type of 3760 // the input array from type speculation. 3761 if (_gvn.type(klass_node)->singleton()) { 3762 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3763 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3764 3765 int test = C->static_subtype_check(superk, subk); 3766 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3767 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3768 if (t_original->speculative_type() != NULL) { 3769 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3770 } 3771 } 3772 } 3773 3774 bool validated = false; 3775 // Reason_class_check rather than Reason_intrinsic because we 3776 // want to intrinsify even if this traps. 3777 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3778 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3779 klass_node); 3780 3781 if (not_subtype_ctrl != top()) { 3782 PreserveJVMState pjvms(this); 3783 set_control(not_subtype_ctrl); 3784 uncommon_trap(Deoptimization::Reason_class_check, 3785 Deoptimization::Action_make_not_entrant); 3786 assert(stopped(), "Should be stopped"); 3787 } 3788 validated = true; 3789 } 3790 3791 if (!stopped()) { 3792 newcopy = new_array(klass_node, length, 0); // no arguments to push 3793 3794 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3795 load_object_klass(original), klass_node); 3796 if (!is_copyOfRange) { 3797 ac->set_copyof(validated); 3798 } else { 3799 ac->set_copyofrange(validated); 3800 } 3801 Node* n = _gvn.transform(ac); 3802 if (n == ac) { 3803 ac->connect_outputs(this); 3804 } else { 3805 assert(validated, "shouldn't transform if all arguments not validated"); 3806 set_all_memory(n); 3807 } 3808 } 3809 } 3810 } // original reexecute is set back here 3811 3812 C->set_has_split_ifs(true); // Has chance for split-if optimization 3813 if (!stopped()) { 3814 set_result(newcopy); 3815 } 3816 return true; 3817 } 3818 3819 3820 //----------------------generate_virtual_guard--------------------------- 3821 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3822 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3823 RegionNode* slow_region) { 3824 ciMethod* method = callee(); 3825 int vtable_index = method->vtable_index(); 3826 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3827 "bad index %d", vtable_index); 3828 // Get the Method* out of the appropriate vtable entry. 3829 int entry_offset = (InstanceKlass::vtable_start_offset() + 3830 vtable_index*vtableEntry::size()) * wordSize + 3831 vtableEntry::method_offset_in_bytes(); 3832 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3833 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3834 3835 // Compare the target method with the expected method (e.g., Object.hashCode). 3836 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3837 3838 Node* native_call = makecon(native_call_addr); 3839 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3840 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3841 3842 return generate_slow_guard(test_native, slow_region); 3843 } 3844 3845 //-----------------------generate_method_call---------------------------- 3846 // Use generate_method_call to make a slow-call to the real 3847 // method if the fast path fails. An alternative would be to 3848 // use a stub like OptoRuntime::slow_arraycopy_Java. 3849 // This only works for expanding the current library call, 3850 // not another intrinsic. (E.g., don't use this for making an 3851 // arraycopy call inside of the copyOf intrinsic.) 3852 CallJavaNode* 3853 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3854 // When compiling the intrinsic method itself, do not use this technique. 3855 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3856 3857 ciMethod* method = callee(); 3858 // ensure the JVMS we have will be correct for this call 3859 guarantee(method_id == method->intrinsic_id(), "must match"); 3860 3861 const TypeFunc* tf = TypeFunc::make(method); 3862 CallJavaNode* slow_call; 3863 if (is_static) { 3864 assert(!is_virtual, ""); 3865 slow_call = new CallStaticJavaNode(C, tf, 3866 SharedRuntime::get_resolve_static_call_stub(), 3867 method, bci()); 3868 } else if (is_virtual) { 3869 null_check_receiver(); 3870 int vtable_index = Method::invalid_vtable_index; 3871 if (UseInlineCaches) { 3872 // Suppress the vtable call 3873 } else { 3874 // hashCode and clone are not a miranda methods, 3875 // so the vtable index is fixed. 3876 // No need to use the linkResolver to get it. 3877 vtable_index = method->vtable_index(); 3878 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3879 "bad index %d", vtable_index); 3880 } 3881 slow_call = new CallDynamicJavaNode(tf, 3882 SharedRuntime::get_resolve_virtual_call_stub(), 3883 method, vtable_index, bci()); 3884 } else { // neither virtual nor static: opt_virtual 3885 null_check_receiver(); 3886 slow_call = new CallStaticJavaNode(C, tf, 3887 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3888 method, bci()); 3889 slow_call->set_optimized_virtual(true); 3890 } 3891 set_arguments_for_java_call(slow_call); 3892 set_edges_for_java_call(slow_call); 3893 return slow_call; 3894 } 3895 3896 3897 /** 3898 * Build special case code for calls to hashCode on an object. This call may 3899 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 3900 * slightly different code. 3901 */ 3902 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3903 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3904 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3905 3906 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3907 3908 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3909 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 3910 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3911 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3912 Node* obj = NULL; 3913 if (!is_static) { 3914 // Check for hashing null object 3915 obj = null_check_receiver(); 3916 if (stopped()) return true; // unconditionally null 3917 result_reg->init_req(_null_path, top()); 3918 result_val->init_req(_null_path, top()); 3919 } else { 3920 // Do a null check, and return zero if null. 3921 // System.identityHashCode(null) == 0 3922 obj = argument(0); 3923 Node* null_ctl = top(); 3924 obj = null_check_oop(obj, &null_ctl); 3925 result_reg->init_req(_null_path, null_ctl); 3926 result_val->init_req(_null_path, _gvn.intcon(0)); 3927 } 3928 3929 // Unconditionally null? Then return right away. 3930 if (stopped()) { 3931 set_control( result_reg->in(_null_path)); 3932 if (!stopped()) 3933 set_result(result_val->in(_null_path)); 3934 return true; 3935 } 3936 3937 // We only go to the fast case code if we pass a number of guards. The 3938 // paths which do not pass are accumulated in the slow_region. 3939 RegionNode* slow_region = new RegionNode(1); 3940 record_for_igvn(slow_region); 3941 3942 // If this is a virtual call, we generate a funny guard. We pull out 3943 // the vtable entry corresponding to hashCode() from the target object. 3944 // If the target method which we are calling happens to be the native 3945 // Object hashCode() method, we pass the guard. We do not need this 3946 // guard for non-virtual calls -- the caller is known to be the native 3947 // Object hashCode(). 3948 if (is_virtual) { 3949 // After null check, get the object's klass. 3950 Node* obj_klass = load_object_klass(obj); 3951 generate_virtual_guard(obj_klass, slow_region); 3952 } 3953 3954 // Get the header out of the object, use LoadMarkNode when available 3955 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3956 // The control of the load must be NULL. Otherwise, the load can move before 3957 // the null check after castPP removal. 3958 Node* no_ctrl = NULL; 3959 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3960 3961 // Test the header to see if it is unlocked. 3962 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3963 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 3964 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3965 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 3966 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 3967 3968 generate_slow_guard(test_unlocked, slow_region); 3969 3970 // Get the hash value and check to see that it has been properly assigned. 3971 // We depend on hash_mask being at most 32 bits and avoid the use of 3972 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3973 // vm: see markOop.hpp. 3974 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3975 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3976 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 3977 // This hack lets the hash bits live anywhere in the mark object now, as long 3978 // as the shift drops the relevant bits into the low 32 bits. Note that 3979 // Java spec says that HashCode is an int so there's no point in capturing 3980 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3981 hshifted_header = ConvX2I(hshifted_header); 3982 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 3983 3984 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3985 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 3986 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 3987 3988 generate_slow_guard(test_assigned, slow_region); 3989 3990 Node* init_mem = reset_memory(); 3991 // fill in the rest of the null path: 3992 result_io ->init_req(_null_path, i_o()); 3993 result_mem->init_req(_null_path, init_mem); 3994 3995 result_val->init_req(_fast_path, hash_val); 3996 result_reg->init_req(_fast_path, control()); 3997 result_io ->init_req(_fast_path, i_o()); 3998 result_mem->init_req(_fast_path, init_mem); 3999 4000 // Generate code for the slow case. We make a call to hashCode(). 4001 set_control(_gvn.transform(slow_region)); 4002 if (!stopped()) { 4003 // No need for PreserveJVMState, because we're using up the present state. 4004 set_all_memory(init_mem); 4005 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4006 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4007 Node* slow_result = set_results_for_java_call(slow_call); 4008 // this->control() comes from set_results_for_java_call 4009 result_reg->init_req(_slow_path, control()); 4010 result_val->init_req(_slow_path, slow_result); 4011 result_io ->set_req(_slow_path, i_o()); 4012 result_mem ->set_req(_slow_path, reset_memory()); 4013 } 4014 4015 // Return the combined state. 4016 set_i_o( _gvn.transform(result_io) ); 4017 set_all_memory( _gvn.transform(result_mem)); 4018 4019 set_result(result_reg, result_val); 4020 return true; 4021 } 4022 4023 //---------------------------inline_native_getClass---------------------------- 4024 // public final native Class<?> java.lang.Object.getClass(); 4025 // 4026 // Build special case code for calls to getClass on an object. 4027 bool LibraryCallKit::inline_native_getClass() { 4028 Node* obj = null_check_receiver(); 4029 if (stopped()) return true; 4030 set_result(load_mirror_from_klass(load_object_klass(obj))); 4031 return true; 4032 } 4033 4034 //-----------------inline_native_Reflection_getCallerClass--------------------- 4035 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4036 // 4037 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4038 // 4039 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4040 // in that it must skip particular security frames and checks for 4041 // caller sensitive methods. 4042 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4043 #ifndef PRODUCT 4044 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4045 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4046 } 4047 #endif 4048 4049 if (!jvms()->has_method()) { 4050 #ifndef PRODUCT 4051 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4052 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4053 } 4054 #endif 4055 return false; 4056 } 4057 4058 // Walk back up the JVM state to find the caller at the required 4059 // depth. 4060 JVMState* caller_jvms = jvms(); 4061 4062 // Cf. JVM_GetCallerClass 4063 // NOTE: Start the loop at depth 1 because the current JVM state does 4064 // not include the Reflection.getCallerClass() frame. 4065 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4066 ciMethod* m = caller_jvms->method(); 4067 switch (n) { 4068 case 0: 4069 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4070 break; 4071 case 1: 4072 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4073 if (!m->caller_sensitive()) { 4074 #ifndef PRODUCT 4075 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4076 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4077 } 4078 #endif 4079 return false; // bail-out; let JVM_GetCallerClass do the work 4080 } 4081 break; 4082 default: 4083 if (!m->is_ignored_by_security_stack_walk()) { 4084 // We have reached the desired frame; return the holder class. 4085 // Acquire method holder as java.lang.Class and push as constant. 4086 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4087 ciInstance* caller_mirror = caller_klass->java_mirror(); 4088 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4089 4090 #ifndef PRODUCT 4091 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4092 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()); 4093 tty->print_cr(" JVM state at this point:"); 4094 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4095 ciMethod* m = jvms()->of_depth(i)->method(); 4096 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4097 } 4098 } 4099 #endif 4100 return true; 4101 } 4102 break; 4103 } 4104 } 4105 4106 #ifndef PRODUCT 4107 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4108 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4109 tty->print_cr(" JVM state at this point:"); 4110 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4111 ciMethod* m = jvms()->of_depth(i)->method(); 4112 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4113 } 4114 } 4115 #endif 4116 4117 return false; // bail-out; let JVM_GetCallerClass do the work 4118 } 4119 4120 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4121 Node* arg = argument(0); 4122 Node* result = NULL; 4123 4124 switch (id) { 4125 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4126 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4127 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4128 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4129 4130 case vmIntrinsics::_doubleToLongBits: { 4131 // two paths (plus control) merge in a wood 4132 RegionNode *r = new RegionNode(3); 4133 Node *phi = new PhiNode(r, TypeLong::LONG); 4134 4135 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4136 // Build the boolean node 4137 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4138 4139 // Branch either way. 4140 // NaN case is less traveled, which makes all the difference. 4141 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4142 Node *opt_isnan = _gvn.transform(ifisnan); 4143 assert( opt_isnan->is_If(), "Expect an IfNode"); 4144 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4145 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4146 4147 set_control(iftrue); 4148 4149 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4150 Node *slow_result = longcon(nan_bits); // return NaN 4151 phi->init_req(1, _gvn.transform( slow_result )); 4152 r->init_req(1, iftrue); 4153 4154 // Else fall through 4155 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4156 set_control(iffalse); 4157 4158 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4159 r->init_req(2, iffalse); 4160 4161 // Post merge 4162 set_control(_gvn.transform(r)); 4163 record_for_igvn(r); 4164 4165 C->set_has_split_ifs(true); // Has chance for split-if optimization 4166 result = phi; 4167 assert(result->bottom_type()->isa_long(), "must be"); 4168 break; 4169 } 4170 4171 case vmIntrinsics::_floatToIntBits: { 4172 // two paths (plus control) merge in a wood 4173 RegionNode *r = new RegionNode(3); 4174 Node *phi = new PhiNode(r, TypeInt::INT); 4175 4176 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4177 // Build the boolean node 4178 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4179 4180 // Branch either way. 4181 // NaN case is less traveled, which makes all the difference. 4182 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4183 Node *opt_isnan = _gvn.transform(ifisnan); 4184 assert( opt_isnan->is_If(), "Expect an IfNode"); 4185 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4186 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4187 4188 set_control(iftrue); 4189 4190 static const jint nan_bits = 0x7fc00000; 4191 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4192 phi->init_req(1, _gvn.transform( slow_result )); 4193 r->init_req(1, iftrue); 4194 4195 // Else fall through 4196 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4197 set_control(iffalse); 4198 4199 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4200 r->init_req(2, iffalse); 4201 4202 // Post merge 4203 set_control(_gvn.transform(r)); 4204 record_for_igvn(r); 4205 4206 C->set_has_split_ifs(true); // Has chance for split-if optimization 4207 result = phi; 4208 assert(result->bottom_type()->isa_int(), "must be"); 4209 break; 4210 } 4211 4212 default: 4213 fatal_unexpected_iid(id); 4214 break; 4215 } 4216 set_result(_gvn.transform(result)); 4217 return true; 4218 } 4219 4220 //----------------------inline_unsafe_copyMemory------------------------- 4221 // public native void Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4222 bool LibraryCallKit::inline_unsafe_copyMemory() { 4223 if (callee()->is_static()) return false; // caller must have the capability! 4224 null_check_receiver(); // null-check receiver 4225 if (stopped()) return true; 4226 4227 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4228 4229 Node* src_ptr = argument(1); // type: oop 4230 Node* src_off = ConvL2X(argument(2)); // type: long 4231 Node* dst_ptr = argument(4); // type: oop 4232 Node* dst_off = ConvL2X(argument(5)); // type: long 4233 Node* size = ConvL2X(argument(7)); // type: long 4234 4235 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4236 "fieldOffset must be byte-scaled"); 4237 4238 Node* src = make_unsafe_address(src_ptr, src_off); 4239 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4240 4241 // Conservatively insert a memory barrier on all memory slices. 4242 // Do not let writes of the copy source or destination float below the copy. 4243 insert_mem_bar(Op_MemBarCPUOrder); 4244 4245 // Call it. Note that the length argument is not scaled. 4246 make_runtime_call(RC_LEAF|RC_NO_FP, 4247 OptoRuntime::fast_arraycopy_Type(), 4248 StubRoutines::unsafe_arraycopy(), 4249 "unsafe_arraycopy", 4250 TypeRawPtr::BOTTOM, 4251 src, dst, size XTOP); 4252 4253 // Do not let reads of the copy destination float above the copy. 4254 insert_mem_bar(Op_MemBarCPUOrder); 4255 4256 return true; 4257 } 4258 4259 //------------------------clone_coping----------------------------------- 4260 // Helper function for inline_native_clone. 4261 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4262 assert(obj_size != NULL, ""); 4263 Node* raw_obj = alloc_obj->in(1); 4264 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4265 4266 AllocateNode* alloc = NULL; 4267 if (ReduceBulkZeroing) { 4268 // We will be completely responsible for initializing this object - 4269 // mark Initialize node as complete. 4270 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4271 // The object was just allocated - there should be no any stores! 4272 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4273 // Mark as complete_with_arraycopy so that on AllocateNode 4274 // expansion, we know this AllocateNode is initialized by an array 4275 // copy and a StoreStore barrier exists after the array copy. 4276 alloc->initialization()->set_complete_with_arraycopy(); 4277 } 4278 4279 // Copy the fastest available way. 4280 // TODO: generate fields copies for small objects instead. 4281 Node* src = obj; 4282 Node* dest = alloc_obj; 4283 Node* size = _gvn.transform(obj_size); 4284 4285 // Exclude the header but include array length to copy by 8 bytes words. 4286 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4287 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4288 instanceOopDesc::base_offset_in_bytes(); 4289 // base_off: 4290 // 8 - 32-bit VM 4291 // 12 - 64-bit VM, compressed klass 4292 // 16 - 64-bit VM, normal klass 4293 if (base_off % BytesPerLong != 0) { 4294 assert(UseCompressedClassPointers, ""); 4295 if (is_array) { 4296 // Exclude length to copy by 8 bytes words. 4297 base_off += sizeof(int); 4298 } else { 4299 // Include klass to copy by 8 bytes words. 4300 base_off = instanceOopDesc::klass_offset_in_bytes(); 4301 } 4302 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4303 } 4304 src = basic_plus_adr(src, base_off); 4305 dest = basic_plus_adr(dest, base_off); 4306 4307 // Compute the length also, if needed: 4308 Node* countx = size; 4309 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4310 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4311 4312 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4313 4314 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4315 ac->set_clonebasic(); 4316 Node* n = _gvn.transform(ac); 4317 if (n == ac) { 4318 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4319 } else { 4320 set_all_memory(n); 4321 } 4322 4323 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4324 if (card_mark) { 4325 assert(!is_array, ""); 4326 // Put in store barrier for any and all oops we are sticking 4327 // into this object. (We could avoid this if we could prove 4328 // that the object type contains no oop fields at all.) 4329 Node* no_particular_value = NULL; 4330 Node* no_particular_field = NULL; 4331 int raw_adr_idx = Compile::AliasIdxRaw; 4332 post_barrier(control(), 4333 memory(raw_adr_type), 4334 alloc_obj, 4335 no_particular_field, 4336 raw_adr_idx, 4337 no_particular_value, 4338 T_OBJECT, 4339 false); 4340 } 4341 4342 // Do not let reads from the cloned object float above the arraycopy. 4343 if (alloc != NULL) { 4344 // Do not let stores that initialize this object be reordered with 4345 // a subsequent store that would make this object accessible by 4346 // other threads. 4347 // Record what AllocateNode this StoreStore protects so that 4348 // escape analysis can go from the MemBarStoreStoreNode to the 4349 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4350 // based on the escape status of the AllocateNode. 4351 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4352 } else { 4353 insert_mem_bar(Op_MemBarCPUOrder); 4354 } 4355 } 4356 4357 //------------------------inline_native_clone---------------------------- 4358 // protected native Object java.lang.Object.clone(); 4359 // 4360 // Here are the simple edge cases: 4361 // null receiver => normal trap 4362 // virtual and clone was overridden => slow path to out-of-line clone 4363 // not cloneable or finalizer => slow path to out-of-line Object.clone 4364 // 4365 // The general case has two steps, allocation and copying. 4366 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4367 // 4368 // Copying also has two cases, oop arrays and everything else. 4369 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4370 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4371 // 4372 // These steps fold up nicely if and when the cloned object's klass 4373 // can be sharply typed as an object array, a type array, or an instance. 4374 // 4375 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4376 PhiNode* result_val; 4377 4378 // Set the reexecute bit for the interpreter to reexecute 4379 // the bytecode that invokes Object.clone if deoptimization happens. 4380 { PreserveReexecuteState preexecs(this); 4381 jvms()->set_should_reexecute(true); 4382 4383 Node* obj = null_check_receiver(); 4384 if (stopped()) return true; 4385 4386 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4387 4388 // If we are going to clone an instance, we need its exact type to 4389 // know the number and types of fields to convert the clone to 4390 // loads/stores. Maybe a speculative type can help us. 4391 if (!obj_type->klass_is_exact() && 4392 obj_type->speculative_type() != NULL && 4393 obj_type->speculative_type()->is_instance_klass()) { 4394 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4395 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4396 !spec_ik->has_injected_fields()) { 4397 ciKlass* k = obj_type->klass(); 4398 if (!k->is_instance_klass() || 4399 k->as_instance_klass()->is_interface() || 4400 k->as_instance_klass()->has_subklass()) { 4401 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4402 } 4403 } 4404 } 4405 4406 Node* obj_klass = load_object_klass(obj); 4407 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4408 const TypeOopPtr* toop = ((tklass != NULL) 4409 ? tklass->as_instance_type() 4410 : TypeInstPtr::NOTNULL); 4411 4412 // Conservatively insert a memory barrier on all memory slices. 4413 // Do not let writes into the original float below the clone. 4414 insert_mem_bar(Op_MemBarCPUOrder); 4415 4416 // paths into result_reg: 4417 enum { 4418 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4419 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4420 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4421 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4422 PATH_LIMIT 4423 }; 4424 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4425 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4426 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4427 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4428 record_for_igvn(result_reg); 4429 4430 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4431 int raw_adr_idx = Compile::AliasIdxRaw; 4432 4433 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4434 if (array_ctl != NULL) { 4435 // It's an array. 4436 PreserveJVMState pjvms(this); 4437 set_control(array_ctl); 4438 Node* obj_length = load_array_length(obj); 4439 Node* obj_size = NULL; 4440 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4441 4442 if (!use_ReduceInitialCardMarks()) { 4443 // If it is an oop array, it requires very special treatment, 4444 // because card marking is required on each card of the array. 4445 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4446 if (is_obja != NULL) { 4447 PreserveJVMState pjvms2(this); 4448 set_control(is_obja); 4449 // Generate a direct call to the right arraycopy function(s). 4450 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4451 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4452 ac->set_cloneoop(); 4453 Node* n = _gvn.transform(ac); 4454 assert(n == ac, "cannot disappear"); 4455 ac->connect_outputs(this); 4456 4457 result_reg->init_req(_objArray_path, control()); 4458 result_val->init_req(_objArray_path, alloc_obj); 4459 result_i_o ->set_req(_objArray_path, i_o()); 4460 result_mem ->set_req(_objArray_path, reset_memory()); 4461 } 4462 } 4463 // Otherwise, there are no card marks to worry about. 4464 // (We can dispense with card marks if we know the allocation 4465 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4466 // causes the non-eden paths to take compensating steps to 4467 // simulate a fresh allocation, so that no further 4468 // card marks are required in compiled code to initialize 4469 // the object.) 4470 4471 if (!stopped()) { 4472 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4473 4474 // Present the results of the copy. 4475 result_reg->init_req(_array_path, control()); 4476 result_val->init_req(_array_path, alloc_obj); 4477 result_i_o ->set_req(_array_path, i_o()); 4478 result_mem ->set_req(_array_path, reset_memory()); 4479 } 4480 } 4481 4482 // We only go to the instance fast case code if we pass a number of guards. 4483 // The paths which do not pass are accumulated in the slow_region. 4484 RegionNode* slow_region = new RegionNode(1); 4485 record_for_igvn(slow_region); 4486 if (!stopped()) { 4487 // It's an instance (we did array above). Make the slow-path tests. 4488 // If this is a virtual call, we generate a funny guard. We grab 4489 // the vtable entry corresponding to clone() from the target object. 4490 // If the target method which we are calling happens to be the 4491 // Object clone() method, we pass the guard. We do not need this 4492 // guard for non-virtual calls; the caller is known to be the native 4493 // Object clone(). 4494 if (is_virtual) { 4495 generate_virtual_guard(obj_klass, slow_region); 4496 } 4497 4498 // The object must be easily cloneable and must not have a finalizer. 4499 // Both of these conditions may be checked in a single test. 4500 // We could optimize the test further, but we don't care. 4501 generate_access_flags_guard(obj_klass, 4502 // Test both conditions: 4503 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4504 // Must be cloneable but not finalizer: 4505 JVM_ACC_IS_CLONEABLE_FAST, 4506 slow_region); 4507 } 4508 4509 if (!stopped()) { 4510 // It's an instance, and it passed the slow-path tests. 4511 PreserveJVMState pjvms(this); 4512 Node* obj_size = NULL; 4513 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4514 // is reexecuted if deoptimization occurs and there could be problems when merging 4515 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4516 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4517 4518 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4519 4520 // Present the results of the slow call. 4521 result_reg->init_req(_instance_path, control()); 4522 result_val->init_req(_instance_path, alloc_obj); 4523 result_i_o ->set_req(_instance_path, i_o()); 4524 result_mem ->set_req(_instance_path, reset_memory()); 4525 } 4526 4527 // Generate code for the slow case. We make a call to clone(). 4528 set_control(_gvn.transform(slow_region)); 4529 if (!stopped()) { 4530 PreserveJVMState pjvms(this); 4531 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4532 Node* slow_result = set_results_for_java_call(slow_call); 4533 // this->control() comes from set_results_for_java_call 4534 result_reg->init_req(_slow_path, control()); 4535 result_val->init_req(_slow_path, slow_result); 4536 result_i_o ->set_req(_slow_path, i_o()); 4537 result_mem ->set_req(_slow_path, reset_memory()); 4538 } 4539 4540 // Return the combined state. 4541 set_control( _gvn.transform(result_reg)); 4542 set_i_o( _gvn.transform(result_i_o)); 4543 set_all_memory( _gvn.transform(result_mem)); 4544 } // original reexecute is set back here 4545 4546 set_result(_gvn.transform(result_val)); 4547 return true; 4548 } 4549 4550 // If we have a tighly coupled allocation, the arraycopy may take care 4551 // of the array initialization. If one of the guards we insert between 4552 // the allocation and the arraycopy causes a deoptimization, an 4553 // unitialized array will escape the compiled method. To prevent that 4554 // we set the JVM state for uncommon traps between the allocation and 4555 // the arraycopy to the state before the allocation so, in case of 4556 // deoptimization, we'll reexecute the allocation and the 4557 // initialization. 4558 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, 4559 int& saved_reexecute_sp) { 4560 // Make sure there's no store between the allocation and the 4561 // arraycopy otherwise visible side effects could be rexecuted 4562 // in case of deoptimization and cause incorrect execution. 4563 bool no_interfering_store = true; 4564 Node* mem = alloc->in(TypeFunc::Memory); 4565 if (mem->is_MergeMem()) { 4566 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4567 Node* n = mms.memory(); 4568 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4569 assert(n->is_Store(), "what else?"); 4570 no_interfering_store = false; 4571 break; 4572 } 4573 } 4574 } else { 4575 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4576 Node* n = mms.memory(); 4577 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4578 assert(n->is_Store(), "what else?"); 4579 no_interfering_store = false; 4580 break; 4581 } 4582 } 4583 } 4584 4585 if (no_interfering_store) { 4586 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4587 uint size = alloc->req(); 4588 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4589 old_jvms->set_map(sfpt); 4590 for (uint i = 0; i < size; i++) { 4591 sfpt->init_req(i, alloc->in(i)); 4592 } 4593 // re-push array length for deoptimization 4594 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4595 old_jvms->set_sp(old_jvms->sp()+1); 4596 old_jvms->set_monoff(old_jvms->monoff()+1); 4597 old_jvms->set_scloff(old_jvms->scloff()+1); 4598 old_jvms->set_endoff(old_jvms->endoff()+1); 4599 old_jvms->set_should_reexecute(true); 4600 4601 sfpt->set_i_o(map()->i_o()); 4602 sfpt->set_memory(map()->memory()); 4603 sfpt->set_control(map()->control()); 4604 4605 JVMState* saved_jvms = jvms(); 4606 saved_reexecute_sp = _reexecute_sp; 4607 4608 set_jvms(sfpt->jvms()); 4609 _reexecute_sp = jvms()->sp(); 4610 4611 return saved_jvms; 4612 } 4613 return NULL; 4614 } 4615 4616 // Because of zeroing elimination, a trap between an arraycopy and its 4617 // tightly coupled allocation would leave an uninitialized array in 4618 // the heap that GCs wouldn't expect. Moving the allocation after the 4619 // traps so we don't allocate the array if we deoptimize solves that 4620 // problem. This is possible because tightly_coupled_allocation() 4621 // guarantees there's no observer of the allocated array at this point 4622 // and the control flow is simple enough. In case of a deoptimization, 4623 // we restart execution before the allocation, allocating a new array. 4624 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, 4625 Node* dest, 4626 JVMState* saved_jvms, 4627 int saved_reexecute_sp) { 4628 if (saved_jvms != NULL && !stopped()) { 4629 assert(alloc != NULL, "only with a tightly coupled allocation"); 4630 // restore JVM state to the state at the arraycopy 4631 saved_jvms->map()->set_control(map()->control()); 4632 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4633 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4634 // If we've improved the types of some nodes (null check) while 4635 // emitting the guards, propagate them to the current state 4636 map()->replaced_nodes().apply(saved_jvms->map()); 4637 set_jvms(saved_jvms); 4638 _reexecute_sp = saved_reexecute_sp; 4639 4640 // Remove the allocation from above the guards 4641 CallProjections callprojs; 4642 alloc->extract_projections(&callprojs, true); 4643 InitializeNode* init = alloc->initialization(); 4644 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4645 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4646 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4647 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4648 4649 // move the allocation here (after the guards) 4650 _gvn.hash_delete(alloc); 4651 alloc->set_req(TypeFunc::Control, control()); 4652 alloc->set_req(TypeFunc::I_O, i_o()); 4653 Node *mem = reset_memory(); 4654 set_all_memory(mem); 4655 alloc->set_req(TypeFunc::Memory, mem); 4656 set_control(init->proj_out(TypeFunc::Control)); 4657 set_i_o(callprojs.fallthrough_ioproj); 4658 4659 // Update memory as done in GraphKit::set_output_for_allocation() 4660 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4661 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4662 if (ary_type->isa_aryptr() && length_type != NULL) { 4663 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4664 } 4665 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4666 int elemidx = C->get_alias_index(telemref); 4667 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4668 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4669 4670 Node* allocx = _gvn.transform(alloc); 4671 assert(allocx == alloc, "where has the allocation gone?"); 4672 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4673 4674 _gvn.hash_delete(dest); 4675 dest->set_req(0, control()); 4676 Node* destx = _gvn.transform(dest); 4677 assert(destx == dest, "where has the allocation result gone?"); 4678 } 4679 } 4680 4681 4682 //------------------------------inline_arraycopy----------------------- 4683 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4684 // Object dest, int destPos, 4685 // int length); 4686 bool LibraryCallKit::inline_arraycopy() { 4687 // Get the arguments. 4688 Node* src = argument(0); // type: oop 4689 Node* src_offset = argument(1); // type: int 4690 Node* dest = argument(2); // type: oop 4691 Node* dest_offset = argument(3); // type: int 4692 Node* length = argument(4); // type: int 4693 4694 4695 // Check for allocation before we add nodes that would confuse 4696 // tightly_coupled_allocation() 4697 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4698 4699 // If the allocation is moved below the guards then we deopt before 4700 // the allocation and traps are recorded (and must be checked) at 4701 // the method/bci of the allocation. Otherwise, we record/check for 4702 // traps at the arraycopy call. 4703 ciMethod* alloc_method = NULL; 4704 int alloc_bci = -1; 4705 if (alloc != NULL) { 4706 alloc_method = alloc->jvms()->method(); 4707 alloc_bci = alloc->jvms()->bci(); 4708 } 4709 4710 int saved_reexecute_sp = -1; 4711 JVMState* saved_jvms = NULL; 4712 // In case of a tightly coupled allocation, the allocation needs to 4713 // be moved after the null check for src below so in case we deopt 4714 // at the null check, the array is not left uninitialized (because 4715 // of zeroing elimination). 4716 // If guards are emitted, the allocation is moved all the way after 4717 // the guards. If guards are not emitted (because some guards 4718 // trapped too much in the past), the allocation is moved between 4719 // the null check for src and the array copy code. Once expanded the 4720 // array copy code doesn't include safepoints before the copy itself 4721 // so there's no risk that an uninitialized array escapes the 4722 // method. 4723 if (alloc != NULL && !C->too_many_traps(alloc_method, alloc_bci, Deoptimization::Reason_null_check) && 4724 !C->too_many_traps(alloc_method, alloc_bci, Deoptimization::Reason_speculate_null_check)) { 4725 saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4726 } 4727 4728 // The following tests must be performed 4729 // (1) src and dest are arrays. 4730 // (2) src and dest arrays must have elements of the same BasicType 4731 // (3) src and dest must not be null. 4732 // (4) src_offset must not be negative. 4733 // (5) dest_offset must not be negative. 4734 // (6) length must not be negative. 4735 // (7) src_offset + length must not exceed length of src. 4736 // (8) dest_offset + length must not exceed length of dest. 4737 // (9) each element of an oop array must be assignable 4738 4739 // (3) src and dest must not be null. 4740 // always do this here because we need the JVM state for uncommon traps 4741 Node* null_ctl = top(); 4742 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true, !_gvn.type(src)->speculative_maybe_null()) : null_check(src, T_ARRAY); 4743 assert(null_ctl->is_top(), "no null control here"); 4744 dest = null_check(dest, T_ARRAY); 4745 4746 // We emitted a null check for src. 4747 // In case of a tightly coupled allocation: 4748 // - the null check for dest is optimized out 4749 // - if saved_jvms is not NULL, we decided to move the allocation 4750 // below the src null check and are free to move the allocation 4751 // below the guards (if we emit them) as well. 4752 // - if saved_jvms is NULL, the src null check could be a witness of 4753 // the allocation and so prevent zeroing elimination. We need to 4754 // call tightly_coupled_allocation() again to check if it's the case 4755 // or not. If we still have a tightly coupled allocation, then it 4756 // can be moved below the guards (if we emit them) and we call 4757 // arraycopy_restore_alloc_state() to prepare the move of the 4758 // allocation. 4759 if (saved_jvms == NULL && alloc != NULL) { 4760 alloc = tightly_coupled_allocation(dest, NULL); 4761 if (alloc != NULL && !C->too_many_traps(alloc_method, alloc_bci, Deoptimization::Reason_intrinsic)) { 4762 saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4763 } 4764 } 4765 4766 // See arraycopy_restore_alloc_state() comment 4767 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4768 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4769 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards 4770 bool can_emit_guards = (alloc == NULL || (saved_jvms != NULL && !C->too_many_traps(alloc_method, alloc_bci, Deoptimization::Reason_intrinsic))); 4771 4772 bool validated = false; 4773 4774 const Type* src_type = _gvn.type(src); 4775 const Type* dest_type = _gvn.type(dest); 4776 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4777 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4778 4779 // Do we have the type of src? 4780 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4781 // Do we have the type of dest? 4782 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4783 // Is the type for src from speculation? 4784 bool src_spec = false; 4785 // Is the type for dest from speculation? 4786 bool dest_spec = false; 4787 4788 if ((!has_src || !has_dest) && can_emit_guards) { 4789 // We don't have sufficient type information, let's see if 4790 // speculative types can help. We need to have types for both src 4791 // and dest so that it pays off. 4792 4793 // Do we already have or could we have type information for src 4794 bool could_have_src = has_src; 4795 // Do we already have or could we have type information for dest 4796 bool could_have_dest = has_dest; 4797 4798 ciKlass* src_k = NULL; 4799 if (!has_src) { 4800 src_k = src_type->speculative_type_not_null(); 4801 if (src_k != NULL && src_k->is_array_klass()) { 4802 could_have_src = true; 4803 } 4804 } 4805 4806 ciKlass* dest_k = NULL; 4807 if (!has_dest) { 4808 dest_k = dest_type->speculative_type_not_null(); 4809 if (dest_k != NULL && dest_k->is_array_klass()) { 4810 could_have_dest = true; 4811 } 4812 } 4813 4814 if (could_have_src && could_have_dest) { 4815 // This is going to pay off so emit the required guards 4816 if (!has_src) { 4817 src = maybe_cast_profiled_obj(src, src_k, true); 4818 src_type = _gvn.type(src); 4819 top_src = src_type->isa_aryptr(); 4820 has_src = (top_src != NULL && top_src->klass() != NULL); 4821 src_spec = true; 4822 } 4823 if (!has_dest) { 4824 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4825 dest_type = _gvn.type(dest); 4826 top_dest = dest_type->isa_aryptr(); 4827 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4828 dest_spec = true; 4829 } 4830 } 4831 } 4832 4833 if (has_src && has_dest && can_emit_guards) { 4834 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4835 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4836 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4837 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4838 4839 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4840 // If both arrays are object arrays then having the exact types 4841 // for both will remove the need for a subtype check at runtime 4842 // before the call and may make it possible to pick a faster copy 4843 // routine (without a subtype check on every element) 4844 // Do we have the exact type of src? 4845 bool could_have_src = src_spec; 4846 // Do we have the exact type of dest? 4847 bool could_have_dest = dest_spec; 4848 ciKlass* src_k = top_src->klass(); 4849 ciKlass* dest_k = top_dest->klass(); 4850 if (!src_spec) { 4851 src_k = src_type->speculative_type_not_null(); 4852 if (src_k != NULL && src_k->is_array_klass()) { 4853 could_have_src = true; 4854 } 4855 } 4856 if (!dest_spec) { 4857 dest_k = dest_type->speculative_type_not_null(); 4858 if (dest_k != NULL && dest_k->is_array_klass()) { 4859 could_have_dest = true; 4860 } 4861 } 4862 if (could_have_src && could_have_dest) { 4863 // If we can have both exact types, emit the missing guards 4864 if (could_have_src && !src_spec) { 4865 src = maybe_cast_profiled_obj(src, src_k, true); 4866 } 4867 if (could_have_dest && !dest_spec) { 4868 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4869 } 4870 } 4871 } 4872 } 4873 4874 if ((alloc != NULL || !C->too_many_traps(method(), bci(), Deoptimization::Reason_intrinsic)) && 4875 can_emit_guards && 4876 !src->is_top() && !dest->is_top()) { 4877 // validate arguments: enables transformation the ArrayCopyNode 4878 validated = true; 4879 4880 RegionNode* slow_region = new RegionNode(1); 4881 record_for_igvn(slow_region); 4882 4883 // (1) src and dest are arrays. 4884 generate_non_array_guard(load_object_klass(src), slow_region); 4885 generate_non_array_guard(load_object_klass(dest), slow_region); 4886 4887 // (2) src and dest arrays must have elements of the same BasicType 4888 // done at macro expansion or at Ideal transformation time 4889 4890 // (4) src_offset must not be negative. 4891 generate_negative_guard(src_offset, slow_region); 4892 4893 // (5) dest_offset must not be negative. 4894 generate_negative_guard(dest_offset, slow_region); 4895 4896 // (7) src_offset + length must not exceed length of src. 4897 generate_limit_guard(src_offset, length, 4898 load_array_length(src), 4899 slow_region); 4900 4901 // (8) dest_offset + length must not exceed length of dest. 4902 generate_limit_guard(dest_offset, length, 4903 load_array_length(dest), 4904 slow_region); 4905 4906 // (9) each element of an oop array must be assignable 4907 Node* src_klass = load_object_klass(src); 4908 Node* dest_klass = load_object_klass(dest); 4909 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4910 4911 if (not_subtype_ctrl != top()) { 4912 PreserveJVMState pjvms(this); 4913 set_control(not_subtype_ctrl); 4914 uncommon_trap(Deoptimization::Reason_intrinsic, 4915 Deoptimization::Action_make_not_entrant); 4916 assert(stopped(), "Should be stopped"); 4917 } 4918 { 4919 PreserveJVMState pjvms(this); 4920 set_control(_gvn.transform(slow_region)); 4921 uncommon_trap(Deoptimization::Reason_intrinsic, 4922 Deoptimization::Action_make_not_entrant); 4923 assert(stopped(), "Should be stopped"); 4924 } 4925 } 4926 4927 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 4928 4929 if (stopped()) { 4930 return true; 4931 } 4932 4933 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 4934 // Create LoadRange and LoadKlass nodes for use during macro expansion here 4935 // so the compiler has a chance to eliminate them: during macro expansion, 4936 // we have to set their control (CastPP nodes are eliminated). 4937 load_object_klass(src), load_object_klass(dest), 4938 load_array_length(src), load_array_length(dest)); 4939 4940 ac->set_arraycopy(validated); 4941 4942 Node* n = _gvn.transform(ac); 4943 if (n == ac) { 4944 ac->connect_outputs(this); 4945 } else { 4946 assert(validated, "shouldn't transform if all arguments not validated"); 4947 set_all_memory(n); 4948 } 4949 4950 return true; 4951 } 4952 4953 4954 // Helper function which determines if an arraycopy immediately follows 4955 // an allocation, with no intervening tests or other escapes for the object. 4956 AllocateArrayNode* 4957 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4958 RegionNode* slow_region) { 4959 if (stopped()) return NULL; // no fast path 4960 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4961 4962 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4963 if (alloc == NULL) return NULL; 4964 4965 Node* rawmem = memory(Compile::AliasIdxRaw); 4966 // Is the allocation's memory state untouched? 4967 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4968 // Bail out if there have been raw-memory effects since the allocation. 4969 // (Example: There might have been a call or safepoint.) 4970 return NULL; 4971 } 4972 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4973 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4974 return NULL; 4975 } 4976 4977 // There must be no unexpected observers of this allocation. 4978 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4979 Node* obs = ptr->fast_out(i); 4980 if (obs != this->map()) { 4981 return NULL; 4982 } 4983 } 4984 4985 // This arraycopy must unconditionally follow the allocation of the ptr. 4986 Node* alloc_ctl = ptr->in(0); 4987 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4988 4989 Node* ctl = control(); 4990 while (ctl != alloc_ctl) { 4991 // There may be guards which feed into the slow_region. 4992 // Any other control flow means that we might not get a chance 4993 // to finish initializing the allocated object. 4994 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4995 IfNode* iff = ctl->in(0)->as_If(); 4996 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 4997 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4998 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4999 ctl = iff->in(0); // This test feeds the known slow_region. 5000 continue; 5001 } 5002 // One more try: Various low-level checks bottom out in 5003 // uncommon traps. If the debug-info of the trap omits 5004 // any reference to the allocation, as we've already 5005 // observed, then there can be no objection to the trap. 5006 bool found_trap = false; 5007 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5008 Node* obs = not_ctl->fast_out(j); 5009 if (obs->in(0) == not_ctl && obs->is_Call() && 5010 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5011 found_trap = true; break; 5012 } 5013 } 5014 if (found_trap) { 5015 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5016 continue; 5017 } 5018 } 5019 return NULL; 5020 } 5021 5022 // If we get this far, we have an allocation which immediately 5023 // precedes the arraycopy, and we can take over zeroing the new object. 5024 // The arraycopy will finish the initialization, and provide 5025 // a new control state to which we will anchor the destination pointer. 5026 5027 return alloc; 5028 } 5029 5030 //-------------inline_encodeISOArray----------------------------------- 5031 // encode char[] to byte[] in ISO_8859_1 5032 bool LibraryCallKit::inline_encodeISOArray() { 5033 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5034 // no receiver since it is static method 5035 Node *src = argument(0); 5036 Node *src_offset = argument(1); 5037 Node *dst = argument(2); 5038 Node *dst_offset = argument(3); 5039 Node *length = argument(4); 5040 5041 const Type* src_type = src->Value(&_gvn); 5042 const Type* dst_type = dst->Value(&_gvn); 5043 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5044 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5045 if (top_src == NULL || top_src->klass() == NULL || 5046 top_dest == NULL || top_dest->klass() == NULL) { 5047 // failed array check 5048 return false; 5049 } 5050 5051 // Figure out the size and type of the elements we will be copying. 5052 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5053 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5054 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5055 return false; 5056 } 5057 5058 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5059 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5060 // 'src_start' points to src array + scaled offset 5061 // 'dst_start' points to dst array + scaled offset 5062 5063 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5064 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5065 enc = _gvn.transform(enc); 5066 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5067 set_memory(res_mem, mtype); 5068 set_result(enc); 5069 return true; 5070 } 5071 5072 //-------------inline_multiplyToLen----------------------------------- 5073 bool LibraryCallKit::inline_multiplyToLen() { 5074 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5075 5076 address stubAddr = StubRoutines::multiplyToLen(); 5077 if (stubAddr == NULL) { 5078 return false; // Intrinsic's stub is not implemented on this platform 5079 } 5080 const char* stubName = "multiplyToLen"; 5081 5082 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5083 5084 // no receiver because it is a static method 5085 Node* x = argument(0); 5086 Node* xlen = argument(1); 5087 Node* y = argument(2); 5088 Node* ylen = argument(3); 5089 Node* z = argument(4); 5090 5091 const Type* x_type = x->Value(&_gvn); 5092 const Type* y_type = y->Value(&_gvn); 5093 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5094 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5095 if (top_x == NULL || top_x->klass() == NULL || 5096 top_y == NULL || top_y->klass() == NULL) { 5097 // failed array check 5098 return false; 5099 } 5100 5101 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5102 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5103 if (x_elem != T_INT || y_elem != T_INT) { 5104 return false; 5105 } 5106 5107 // Set the original stack and the reexecute bit for the interpreter to reexecute 5108 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5109 // on the return from z array allocation in runtime. 5110 { PreserveReexecuteState preexecs(this); 5111 jvms()->set_should_reexecute(true); 5112 5113 Node* x_start = array_element_address(x, intcon(0), x_elem); 5114 Node* y_start = array_element_address(y, intcon(0), y_elem); 5115 // 'x_start' points to x array + scaled xlen 5116 // 'y_start' points to y array + scaled ylen 5117 5118 // Allocate the result array 5119 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5120 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5121 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5122 5123 IdealKit ideal(this); 5124 5125 #define __ ideal. 5126 Node* one = __ ConI(1); 5127 Node* zero = __ ConI(0); 5128 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5129 __ set(need_alloc, zero); 5130 __ set(z_alloc, z); 5131 __ if_then(z, BoolTest::eq, null()); { 5132 __ increment (need_alloc, one); 5133 } __ else_(); { 5134 // Update graphKit memory and control from IdealKit. 5135 sync_kit(ideal); 5136 Node* zlen_arg = load_array_length(z); 5137 // Update IdealKit memory and control from graphKit. 5138 __ sync_kit(this); 5139 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5140 __ increment (need_alloc, one); 5141 } __ end_if(); 5142 } __ end_if(); 5143 5144 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5145 // Update graphKit memory and control from IdealKit. 5146 sync_kit(ideal); 5147 Node * narr = new_array(klass_node, zlen, 1); 5148 // Update IdealKit memory and control from graphKit. 5149 __ sync_kit(this); 5150 __ set(z_alloc, narr); 5151 } __ end_if(); 5152 5153 sync_kit(ideal); 5154 z = __ value(z_alloc); 5155 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5156 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5157 // Final sync IdealKit and GraphKit. 5158 final_sync(ideal); 5159 #undef __ 5160 5161 Node* z_start = array_element_address(z, intcon(0), T_INT); 5162 5163 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5164 OptoRuntime::multiplyToLen_Type(), 5165 stubAddr, stubName, TypePtr::BOTTOM, 5166 x_start, xlen, y_start, ylen, z_start, zlen); 5167 } // original reexecute is set back here 5168 5169 C->set_has_split_ifs(true); // Has chance for split-if optimization 5170 set_result(z); 5171 return true; 5172 } 5173 5174 //-------------inline_squareToLen------------------------------------ 5175 bool LibraryCallKit::inline_squareToLen() { 5176 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5177 5178 address stubAddr = StubRoutines::squareToLen(); 5179 if (stubAddr == NULL) { 5180 return false; // Intrinsic's stub is not implemented on this platform 5181 } 5182 const char* stubName = "squareToLen"; 5183 5184 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5185 5186 Node* x = argument(0); 5187 Node* len = argument(1); 5188 Node* z = argument(2); 5189 Node* zlen = argument(3); 5190 5191 const Type* x_type = x->Value(&_gvn); 5192 const Type* z_type = z->Value(&_gvn); 5193 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5194 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5195 if (top_x == NULL || top_x->klass() == NULL || 5196 top_z == NULL || top_z->klass() == NULL) { 5197 // failed array check 5198 return false; 5199 } 5200 5201 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5202 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5203 if (x_elem != T_INT || z_elem != T_INT) { 5204 return false; 5205 } 5206 5207 5208 Node* x_start = array_element_address(x, intcon(0), x_elem); 5209 Node* z_start = array_element_address(z, intcon(0), z_elem); 5210 5211 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5212 OptoRuntime::squareToLen_Type(), 5213 stubAddr, stubName, TypePtr::BOTTOM, 5214 x_start, len, z_start, zlen); 5215 5216 set_result(z); 5217 return true; 5218 } 5219 5220 //-------------inline_mulAdd------------------------------------------ 5221 bool LibraryCallKit::inline_mulAdd() { 5222 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5223 5224 address stubAddr = StubRoutines::mulAdd(); 5225 if (stubAddr == NULL) { 5226 return false; // Intrinsic's stub is not implemented on this platform 5227 } 5228 const char* stubName = "mulAdd"; 5229 5230 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5231 5232 Node* out = argument(0); 5233 Node* in = argument(1); 5234 Node* offset = argument(2); 5235 Node* len = argument(3); 5236 Node* k = argument(4); 5237 5238 const Type* out_type = out->Value(&_gvn); 5239 const Type* in_type = in->Value(&_gvn); 5240 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5241 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5242 if (top_out == NULL || top_out->klass() == NULL || 5243 top_in == NULL || top_in->klass() == NULL) { 5244 // failed array check 5245 return false; 5246 } 5247 5248 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5249 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5250 if (out_elem != T_INT || in_elem != T_INT) { 5251 return false; 5252 } 5253 5254 Node* outlen = load_array_length(out); 5255 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5256 Node* out_start = array_element_address(out, intcon(0), out_elem); 5257 Node* in_start = array_element_address(in, intcon(0), in_elem); 5258 5259 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5260 OptoRuntime::mulAdd_Type(), 5261 stubAddr, stubName, TypePtr::BOTTOM, 5262 out_start,in_start, new_offset, len, k); 5263 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5264 set_result(result); 5265 return true; 5266 } 5267 5268 //-------------inline_montgomeryMultiply----------------------------------- 5269 bool LibraryCallKit::inline_montgomeryMultiply() { 5270 address stubAddr = StubRoutines::montgomeryMultiply(); 5271 if (stubAddr == NULL) { 5272 return false; // Intrinsic's stub is not implemented on this platform 5273 } 5274 5275 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5276 const char* stubName = "montgomery_square"; 5277 5278 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5279 5280 Node* a = argument(0); 5281 Node* b = argument(1); 5282 Node* n = argument(2); 5283 Node* len = argument(3); 5284 Node* inv = argument(4); 5285 Node* m = argument(6); 5286 5287 const Type* a_type = a->Value(&_gvn); 5288 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5289 const Type* b_type = b->Value(&_gvn); 5290 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5291 const Type* n_type = a->Value(&_gvn); 5292 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5293 const Type* m_type = a->Value(&_gvn); 5294 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5295 if (top_a == NULL || top_a->klass() == NULL || 5296 top_b == NULL || top_b->klass() == NULL || 5297 top_n == NULL || top_n->klass() == NULL || 5298 top_m == NULL || top_m->klass() == NULL) { 5299 // failed array check 5300 return false; 5301 } 5302 5303 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5304 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5305 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5306 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5307 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5308 return false; 5309 } 5310 5311 // Make the call 5312 { 5313 Node* a_start = array_element_address(a, intcon(0), a_elem); 5314 Node* b_start = array_element_address(b, intcon(0), b_elem); 5315 Node* n_start = array_element_address(n, intcon(0), n_elem); 5316 Node* m_start = array_element_address(m, intcon(0), m_elem); 5317 5318 Node* call = make_runtime_call(RC_LEAF, 5319 OptoRuntime::montgomeryMultiply_Type(), 5320 stubAddr, stubName, TypePtr::BOTTOM, 5321 a_start, b_start, n_start, len, inv, top(), 5322 m_start); 5323 set_result(m); 5324 } 5325 5326 return true; 5327 } 5328 5329 bool LibraryCallKit::inline_montgomerySquare() { 5330 address stubAddr = StubRoutines::montgomerySquare(); 5331 if (stubAddr == NULL) { 5332 return false; // Intrinsic's stub is not implemented on this platform 5333 } 5334 5335 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5336 const char* stubName = "montgomery_square"; 5337 5338 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5339 5340 Node* a = argument(0); 5341 Node* n = argument(1); 5342 Node* len = argument(2); 5343 Node* inv = argument(3); 5344 Node* m = argument(5); 5345 5346 const Type* a_type = a->Value(&_gvn); 5347 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5348 const Type* n_type = a->Value(&_gvn); 5349 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5350 const Type* m_type = a->Value(&_gvn); 5351 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5352 if (top_a == NULL || top_a->klass() == NULL || 5353 top_n == NULL || top_n->klass() == NULL || 5354 top_m == NULL || top_m->klass() == NULL) { 5355 // failed array check 5356 return false; 5357 } 5358 5359 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5360 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5361 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5362 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5363 return false; 5364 } 5365 5366 // Make the call 5367 { 5368 Node* a_start = array_element_address(a, intcon(0), a_elem); 5369 Node* n_start = array_element_address(n, intcon(0), n_elem); 5370 Node* m_start = array_element_address(m, intcon(0), m_elem); 5371 5372 Node* call = make_runtime_call(RC_LEAF, 5373 OptoRuntime::montgomerySquare_Type(), 5374 stubAddr, stubName, TypePtr::BOTTOM, 5375 a_start, n_start, len, inv, top(), 5376 m_start); 5377 set_result(m); 5378 } 5379 5380 return true; 5381 } 5382 5383 //-------------inline_vectorizedMismatch------------------------------ 5384 bool LibraryCallKit::inline_vectorizedMismatch() { 5385 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5386 5387 address stubAddr = StubRoutines::vectorizedMismatch(); 5388 if (stubAddr == NULL) { 5389 return false; // Intrinsic's stub is not implemented on this platform 5390 } 5391 const char* stubName = "vectorizedMismatch"; 5392 int size_l = callee()->signature()->size(); 5393 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5394 5395 Node* obja = argument(0); 5396 Node* aoffset = argument(1); 5397 Node* objb = argument(3); 5398 Node* boffset = argument(4); 5399 Node* length = argument(6); 5400 Node* scale = argument(7); 5401 5402 const Type* a_type = obja->Value(&_gvn); 5403 const Type* b_type = objb->Value(&_gvn); 5404 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5405 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5406 if (top_a == NULL || top_a->klass() == NULL || 5407 top_b == NULL || top_b->klass() == NULL) { 5408 // failed array check 5409 return false; 5410 } 5411 5412 Node* call; 5413 jvms()->set_should_reexecute(true); 5414 5415 Node* obja_adr = make_unsafe_address(obja, aoffset); 5416 Node* objb_adr = make_unsafe_address(objb, boffset); 5417 5418 call = make_runtime_call(RC_LEAF, 5419 OptoRuntime::vectorizedMismatch_Type(), 5420 stubAddr, stubName, TypePtr::BOTTOM, 5421 obja_adr, objb_adr, length, scale); 5422 5423 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5424 set_result(result); 5425 return true; 5426 } 5427 5428 /** 5429 * Calculate CRC32 for byte. 5430 * int java.util.zip.CRC32.update(int crc, int b) 5431 */ 5432 bool LibraryCallKit::inline_updateCRC32() { 5433 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5434 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5435 // no receiver since it is static method 5436 Node* crc = argument(0); // type: int 5437 Node* b = argument(1); // type: int 5438 5439 /* 5440 * int c = ~ crc; 5441 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5442 * b = b ^ (c >>> 8); 5443 * crc = ~b; 5444 */ 5445 5446 Node* M1 = intcon(-1); 5447 crc = _gvn.transform(new XorINode(crc, M1)); 5448 Node* result = _gvn.transform(new XorINode(crc, b)); 5449 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5450 5451 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5452 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5453 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5454 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5455 5456 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5457 result = _gvn.transform(new XorINode(crc, result)); 5458 result = _gvn.transform(new XorINode(result, M1)); 5459 set_result(result); 5460 return true; 5461 } 5462 5463 /** 5464 * Calculate CRC32 for byte[] array. 5465 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5466 */ 5467 bool LibraryCallKit::inline_updateBytesCRC32() { 5468 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5469 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5470 // no receiver since it is static method 5471 Node* crc = argument(0); // type: int 5472 Node* src = argument(1); // type: oop 5473 Node* offset = argument(2); // type: int 5474 Node* length = argument(3); // type: int 5475 5476 const Type* src_type = src->Value(&_gvn); 5477 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5478 if (top_src == NULL || top_src->klass() == NULL) { 5479 // failed array check 5480 return false; 5481 } 5482 5483 // Figure out the size and type of the elements we will be copying. 5484 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5485 if (src_elem != T_BYTE) { 5486 return false; 5487 } 5488 5489 // 'src_start' points to src array + scaled offset 5490 Node* src_start = array_element_address(src, offset, src_elem); 5491 5492 // We assume that range check is done by caller. 5493 // TODO: generate range check (offset+length < src.length) in debug VM. 5494 5495 // Call the stub. 5496 address stubAddr = StubRoutines::updateBytesCRC32(); 5497 const char *stubName = "updateBytesCRC32"; 5498 5499 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5500 stubAddr, stubName, TypePtr::BOTTOM, 5501 crc, src_start, length); 5502 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5503 set_result(result); 5504 return true; 5505 } 5506 5507 /** 5508 * Calculate CRC32 for ByteBuffer. 5509 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5510 */ 5511 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5512 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5513 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5514 // no receiver since it is static method 5515 Node* crc = argument(0); // type: int 5516 Node* src = argument(1); // type: long 5517 Node* offset = argument(3); // type: int 5518 Node* length = argument(4); // type: int 5519 5520 src = ConvL2X(src); // adjust Java long to machine word 5521 Node* base = _gvn.transform(new CastX2PNode(src)); 5522 offset = ConvI2X(offset); 5523 5524 // 'src_start' points to src array + scaled offset 5525 Node* src_start = basic_plus_adr(top(), base, offset); 5526 5527 // Call the stub. 5528 address stubAddr = StubRoutines::updateBytesCRC32(); 5529 const char *stubName = "updateBytesCRC32"; 5530 5531 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5532 stubAddr, stubName, TypePtr::BOTTOM, 5533 crc, src_start, length); 5534 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5535 set_result(result); 5536 return true; 5537 } 5538 5539 //------------------------------get_table_from_crc32c_class----------------------- 5540 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5541 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5542 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5543 5544 return table; 5545 } 5546 5547 //------------------------------inline_updateBytesCRC32C----------------------- 5548 // 5549 // Calculate CRC32C for byte[] array. 5550 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5551 // 5552 bool LibraryCallKit::inline_updateBytesCRC32C() { 5553 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5554 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5555 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5556 // no receiver since it is a static method 5557 Node* crc = argument(0); // type: int 5558 Node* src = argument(1); // type: oop 5559 Node* offset = argument(2); // type: int 5560 Node* end = argument(3); // type: int 5561 5562 Node* length = _gvn.transform(new SubINode(end, offset)); 5563 5564 const Type* src_type = src->Value(&_gvn); 5565 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5566 if (top_src == NULL || top_src->klass() == NULL) { 5567 // failed array check 5568 return false; 5569 } 5570 5571 // Figure out the size and type of the elements we will be copying. 5572 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5573 if (src_elem != T_BYTE) { 5574 return false; 5575 } 5576 5577 // 'src_start' points to src array + scaled offset 5578 Node* src_start = array_element_address(src, offset, src_elem); 5579 5580 // static final int[] byteTable in class CRC32C 5581 Node* table = get_table_from_crc32c_class(callee()->holder()); 5582 Node* table_start = array_element_address(table, intcon(0), T_INT); 5583 5584 // We assume that range check is done by caller. 5585 // TODO: generate range check (offset+length < src.length) in debug VM. 5586 5587 // Call the stub. 5588 address stubAddr = StubRoutines::updateBytesCRC32C(); 5589 const char *stubName = "updateBytesCRC32C"; 5590 5591 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5592 stubAddr, stubName, TypePtr::BOTTOM, 5593 crc, src_start, length, table_start); 5594 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5595 set_result(result); 5596 return true; 5597 } 5598 5599 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5600 // 5601 // Calculate CRC32C for DirectByteBuffer. 5602 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5603 // 5604 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5605 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5606 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5607 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5608 // no receiver since it is a static method 5609 Node* crc = argument(0); // type: int 5610 Node* src = argument(1); // type: long 5611 Node* offset = argument(3); // type: int 5612 Node* end = argument(4); // type: int 5613 5614 Node* length = _gvn.transform(new SubINode(end, offset)); 5615 5616 src = ConvL2X(src); // adjust Java long to machine word 5617 Node* base = _gvn.transform(new CastX2PNode(src)); 5618 offset = ConvI2X(offset); 5619 5620 // 'src_start' points to src array + scaled offset 5621 Node* src_start = basic_plus_adr(top(), base, offset); 5622 5623 // static final int[] byteTable in class CRC32C 5624 Node* table = get_table_from_crc32c_class(callee()->holder()); 5625 Node* table_start = array_element_address(table, intcon(0), T_INT); 5626 5627 // Call the stub. 5628 address stubAddr = StubRoutines::updateBytesCRC32C(); 5629 const char *stubName = "updateBytesCRC32C"; 5630 5631 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5632 stubAddr, stubName, TypePtr::BOTTOM, 5633 crc, src_start, length, table_start); 5634 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5635 set_result(result); 5636 return true; 5637 } 5638 5639 //------------------------------inline_updateBytesAdler32---------------------- 5640 // 5641 // Calculate Adler32 checksum for byte[] array. 5642 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5643 // 5644 bool LibraryCallKit::inline_updateBytesAdler32() { 5645 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5646 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5647 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5648 // no receiver since it is static method 5649 Node* crc = argument(0); // type: int 5650 Node* src = argument(1); // type: oop 5651 Node* offset = argument(2); // type: int 5652 Node* length = argument(3); // type: int 5653 5654 const Type* src_type = src->Value(&_gvn); 5655 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5656 if (top_src == NULL || top_src->klass() == NULL) { 5657 // failed array check 5658 return false; 5659 } 5660 5661 // Figure out the size and type of the elements we will be copying. 5662 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5663 if (src_elem != T_BYTE) { 5664 return false; 5665 } 5666 5667 // 'src_start' points to src array + scaled offset 5668 Node* src_start = array_element_address(src, offset, src_elem); 5669 5670 // We assume that range check is done by caller. 5671 // TODO: generate range check (offset+length < src.length) in debug VM. 5672 5673 // Call the stub. 5674 address stubAddr = StubRoutines::updateBytesAdler32(); 5675 const char *stubName = "updateBytesAdler32"; 5676 5677 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5678 stubAddr, stubName, TypePtr::BOTTOM, 5679 crc, src_start, length); 5680 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5681 set_result(result); 5682 return true; 5683 } 5684 5685 //------------------------------inline_updateByteBufferAdler32--------------- 5686 // 5687 // Calculate Adler32 checksum for DirectByteBuffer. 5688 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5689 // 5690 bool LibraryCallKit::inline_updateByteBufferAdler32() { 5691 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5692 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5693 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5694 // no receiver since it is static method 5695 Node* crc = argument(0); // type: int 5696 Node* src = argument(1); // type: long 5697 Node* offset = argument(3); // type: int 5698 Node* length = argument(4); // type: int 5699 5700 src = ConvL2X(src); // adjust Java long to machine word 5701 Node* base = _gvn.transform(new CastX2PNode(src)); 5702 offset = ConvI2X(offset); 5703 5704 // 'src_start' points to src array + scaled offset 5705 Node* src_start = basic_plus_adr(top(), base, offset); 5706 5707 // Call the stub. 5708 address stubAddr = StubRoutines::updateBytesAdler32(); 5709 const char *stubName = "updateBytesAdler32"; 5710 5711 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5712 stubAddr, stubName, TypePtr::BOTTOM, 5713 crc, src_start, length); 5714 5715 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5716 set_result(result); 5717 return true; 5718 } 5719 5720 //----------------------------inline_reference_get---------------------------- 5721 // public T java.lang.ref.Reference.get(); 5722 bool LibraryCallKit::inline_reference_get() { 5723 const int referent_offset = java_lang_ref_Reference::referent_offset; 5724 guarantee(referent_offset > 0, "should have already been set"); 5725 5726 // Get the argument: 5727 Node* reference_obj = null_check_receiver(); 5728 if (stopped()) return true; 5729 5730 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5731 5732 ciInstanceKlass* klass = env()->Object_klass(); 5733 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5734 5735 Node* no_ctrl = NULL; 5736 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5737 5738 // Use the pre-barrier to record the value in the referent field 5739 pre_barrier(false /* do_load */, 5740 control(), 5741 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5742 result /* pre_val */, 5743 T_OBJECT); 5744 5745 // Add memory barrier to prevent commoning reads from this field 5746 // across safepoint since GC can change its value. 5747 insert_mem_bar(Op_MemBarCPUOrder); 5748 5749 set_result(result); 5750 return true; 5751 } 5752 5753 5754 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5755 bool is_exact=true, bool is_static=false, 5756 ciInstanceKlass * fromKls=NULL) { 5757 if (fromKls == NULL) { 5758 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5759 assert(tinst != NULL, "obj is null"); 5760 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5761 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5762 fromKls = tinst->klass()->as_instance_klass(); 5763 } else { 5764 assert(is_static, "only for static field access"); 5765 } 5766 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5767 ciSymbol::make(fieldTypeString), 5768 is_static); 5769 5770 assert (field != NULL, "undefined field"); 5771 if (field == NULL) return (Node *) NULL; 5772 5773 if (is_static) { 5774 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5775 fromObj = makecon(tip); 5776 } 5777 5778 // Next code copied from Parse::do_get_xxx(): 5779 5780 // Compute address and memory type. 5781 int offset = field->offset_in_bytes(); 5782 bool is_vol = field->is_volatile(); 5783 ciType* field_klass = field->type(); 5784 assert(field_klass->is_loaded(), "should be loaded"); 5785 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5786 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5787 BasicType bt = field->layout_type(); 5788 5789 // Build the resultant type of the load 5790 const Type *type; 5791 if (bt == T_OBJECT) { 5792 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5793 } else { 5794 type = Type::get_const_basic_type(bt); 5795 } 5796 5797 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5798 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5799 } 5800 // Build the load. 5801 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5802 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5803 // If reference is volatile, prevent following memory ops from 5804 // floating up past the volatile read. Also prevents commoning 5805 // another volatile read. 5806 if (is_vol) { 5807 // Memory barrier includes bogus read of value to force load BEFORE membar 5808 insert_mem_bar(Op_MemBarAcquire, loadedField); 5809 } 5810 return loadedField; 5811 } 5812 5813 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5814 bool is_exact = true, bool is_static = false, 5815 ciInstanceKlass * fromKls = NULL) { 5816 if (fromKls == NULL) { 5817 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5818 assert(tinst != NULL, "obj is null"); 5819 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5820 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5821 fromKls = tinst->klass()->as_instance_klass(); 5822 } 5823 else { 5824 assert(is_static, "only for static field access"); 5825 } 5826 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5827 ciSymbol::make(fieldTypeString), 5828 is_static); 5829 5830 assert(field != NULL, "undefined field"); 5831 assert(!field->is_volatile(), "not defined for volatile fields"); 5832 5833 if (is_static) { 5834 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5835 fromObj = makecon(tip); 5836 } 5837 5838 // Next code copied from Parse::do_get_xxx(): 5839 5840 // Compute address and memory type. 5841 int offset = field->offset_in_bytes(); 5842 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5843 5844 return adr; 5845 } 5846 5847 //------------------------------inline_aescrypt_Block----------------------- 5848 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5849 address stubAddr = NULL; 5850 const char *stubName; 5851 assert(UseAES, "need AES instruction support"); 5852 5853 switch(id) { 5854 case vmIntrinsics::_aescrypt_encryptBlock: 5855 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5856 stubName = "aescrypt_encryptBlock"; 5857 break; 5858 case vmIntrinsics::_aescrypt_decryptBlock: 5859 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5860 stubName = "aescrypt_decryptBlock"; 5861 break; 5862 } 5863 if (stubAddr == NULL) return false; 5864 5865 Node* aescrypt_object = argument(0); 5866 Node* src = argument(1); 5867 Node* src_offset = argument(2); 5868 Node* dest = argument(3); 5869 Node* dest_offset = argument(4); 5870 5871 // (1) src and dest are arrays. 5872 const Type* src_type = src->Value(&_gvn); 5873 const Type* dest_type = dest->Value(&_gvn); 5874 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5875 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5876 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5877 5878 // for the quick and dirty code we will skip all the checks. 5879 // we are just trying to get the call to be generated. 5880 Node* src_start = src; 5881 Node* dest_start = dest; 5882 if (src_offset != NULL || dest_offset != NULL) { 5883 assert(src_offset != NULL && dest_offset != NULL, ""); 5884 src_start = array_element_address(src, src_offset, T_BYTE); 5885 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5886 } 5887 5888 // now need to get the start of its expanded key array 5889 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5890 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5891 if (k_start == NULL) return false; 5892 5893 if (Matcher::pass_original_key_for_aes()) { 5894 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5895 // compatibility issues between Java key expansion and SPARC crypto instructions 5896 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5897 if (original_k_start == NULL) return false; 5898 5899 // Call the stub. 5900 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5901 stubAddr, stubName, TypePtr::BOTTOM, 5902 src_start, dest_start, k_start, original_k_start); 5903 } else { 5904 // Call the stub. 5905 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5906 stubAddr, stubName, TypePtr::BOTTOM, 5907 src_start, dest_start, k_start); 5908 } 5909 5910 return true; 5911 } 5912 5913 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5914 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5915 address stubAddr = NULL; 5916 const char *stubName = NULL; 5917 5918 assert(UseAES, "need AES instruction support"); 5919 5920 switch(id) { 5921 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5922 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5923 stubName = "cipherBlockChaining_encryptAESCrypt"; 5924 break; 5925 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5926 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5927 stubName = "cipherBlockChaining_decryptAESCrypt"; 5928 break; 5929 } 5930 if (stubAddr == NULL) return false; 5931 5932 Node* cipherBlockChaining_object = argument(0); 5933 Node* src = argument(1); 5934 Node* src_offset = argument(2); 5935 Node* len = argument(3); 5936 Node* dest = argument(4); 5937 Node* dest_offset = argument(5); 5938 5939 // (1) src and dest are arrays. 5940 const Type* src_type = src->Value(&_gvn); 5941 const Type* dest_type = dest->Value(&_gvn); 5942 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5943 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5944 assert (top_src != NULL && top_src->klass() != NULL 5945 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5946 5947 // checks are the responsibility of the caller 5948 Node* src_start = src; 5949 Node* dest_start = dest; 5950 if (src_offset != NULL || dest_offset != NULL) { 5951 assert(src_offset != NULL && dest_offset != NULL, ""); 5952 src_start = array_element_address(src, src_offset, T_BYTE); 5953 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5954 } 5955 5956 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5957 // (because of the predicated logic executed earlier). 5958 // so we cast it here safely. 5959 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5960 5961 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5962 if (embeddedCipherObj == NULL) return false; 5963 5964 // cast it to what we know it will be at runtime 5965 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5966 assert(tinst != NULL, "CBC obj is null"); 5967 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5968 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5969 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5970 5971 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5972 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5973 const TypeOopPtr* xtype = aklass->as_instance_type(); 5974 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5975 aescrypt_object = _gvn.transform(aescrypt_object); 5976 5977 // we need to get the start of the aescrypt_object's expanded key array 5978 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5979 if (k_start == NULL) return false; 5980 5981 // similarly, get the start address of the r vector 5982 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5983 if (objRvec == NULL) return false; 5984 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5985 5986 Node* cbcCrypt; 5987 if (Matcher::pass_original_key_for_aes()) { 5988 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5989 // compatibility issues between Java key expansion and SPARC crypto instructions 5990 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5991 if (original_k_start == NULL) return false; 5992 5993 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5994 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5995 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5996 stubAddr, stubName, TypePtr::BOTTOM, 5997 src_start, dest_start, k_start, r_start, len, original_k_start); 5998 } else { 5999 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6000 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6001 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6002 stubAddr, stubName, TypePtr::BOTTOM, 6003 src_start, dest_start, k_start, r_start, len); 6004 } 6005 6006 // return cipher length (int) 6007 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6008 set_result(retvalue); 6009 return true; 6010 } 6011 6012 //------------------------------inline_counterMode_AESCrypt----------------------- 6013 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6014 assert(UseAES, "need AES instruction support"); 6015 if (!UseAESCTRIntrinsics) return false; 6016 6017 address stubAddr = NULL; 6018 const char *stubName = NULL; 6019 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6020 stubAddr = StubRoutines::counterMode_AESCrypt(); 6021 stubName = "counterMode_AESCrypt"; 6022 } 6023 if (stubAddr == NULL) return false; 6024 6025 Node* counterMode_object = argument(0); 6026 Node* src = argument(1); 6027 Node* src_offset = argument(2); 6028 Node* len = argument(3); 6029 Node* dest = argument(4); 6030 Node* dest_offset = argument(5); 6031 6032 // (1) src and dest are arrays. 6033 const Type* src_type = src->Value(&_gvn); 6034 const Type* dest_type = dest->Value(&_gvn); 6035 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6036 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6037 assert(top_src != NULL && top_src->klass() != NULL && 6038 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6039 6040 // checks are the responsibility of the caller 6041 Node* src_start = src; 6042 Node* dest_start = dest; 6043 if (src_offset != NULL || dest_offset != NULL) { 6044 assert(src_offset != NULL && dest_offset != NULL, ""); 6045 src_start = array_element_address(src, src_offset, T_BYTE); 6046 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6047 } 6048 6049 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6050 // (because of the predicated logic executed earlier). 6051 // so we cast it here safely. 6052 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6053 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6054 if (embeddedCipherObj == NULL) return false; 6055 // cast it to what we know it will be at runtime 6056 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6057 assert(tinst != NULL, "CTR obj is null"); 6058 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6059 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6060 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6061 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6062 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6063 const TypeOopPtr* xtype = aklass->as_instance_type(); 6064 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6065 aescrypt_object = _gvn.transform(aescrypt_object); 6066 // we need to get the start of the aescrypt_object's expanded key array 6067 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6068 if (k_start == NULL) return false; 6069 // similarly, get the start address of the r vector 6070 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6071 if (obj_counter == NULL) return false; 6072 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6073 6074 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6075 if (saved_encCounter == NULL) return false; 6076 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6077 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6078 6079 Node* ctrCrypt; 6080 if (Matcher::pass_original_key_for_aes()) { 6081 // no SPARC version for AES/CTR intrinsics now. 6082 return false; 6083 } 6084 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6085 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6086 OptoRuntime::counterMode_aescrypt_Type(), 6087 stubAddr, stubName, TypePtr::BOTTOM, 6088 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6089 6090 // return cipher length (int) 6091 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6092 set_result(retvalue); 6093 return true; 6094 } 6095 6096 //------------------------------get_key_start_from_aescrypt_object----------------------- 6097 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6098 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6099 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6100 if (objAESCryptKey == NULL) return (Node *) NULL; 6101 6102 // now have the array, need to get the start address of the K array 6103 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6104 return k_start; 6105 } 6106 6107 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6108 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6109 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6110 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6111 if (objAESCryptKey == NULL) return (Node *) NULL; 6112 6113 // now have the array, need to get the start address of the lastKey array 6114 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6115 return original_k_start; 6116 } 6117 6118 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6119 // Return node representing slow path of predicate check. 6120 // the pseudo code we want to emulate with this predicate is: 6121 // for encryption: 6122 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6123 // for decryption: 6124 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6125 // note cipher==plain is more conservative than the original java code but that's OK 6126 // 6127 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6128 // The receiver was checked for NULL already. 6129 Node* objCBC = argument(0); 6130 6131 // Load embeddedCipher field of CipherBlockChaining object. 6132 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6133 6134 // get AESCrypt klass for instanceOf check 6135 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6136 // will have same classloader as CipherBlockChaining object 6137 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6138 assert(tinst != NULL, "CBCobj is null"); 6139 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6140 6141 // we want to do an instanceof comparison against the AESCrypt class 6142 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6143 if (!klass_AESCrypt->is_loaded()) { 6144 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6145 Node* ctrl = control(); 6146 set_control(top()); // no regular fast path 6147 return ctrl; 6148 } 6149 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6150 6151 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6152 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6153 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6154 6155 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6156 6157 // for encryption, we are done 6158 if (!decrypting) 6159 return instof_false; // even if it is NULL 6160 6161 // for decryption, we need to add a further check to avoid 6162 // taking the intrinsic path when cipher and plain are the same 6163 // see the original java code for why. 6164 RegionNode* region = new RegionNode(3); 6165 region->init_req(1, instof_false); 6166 Node* src = argument(1); 6167 Node* dest = argument(4); 6168 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6169 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6170 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6171 region->init_req(2, src_dest_conjoint); 6172 6173 record_for_igvn(region); 6174 return _gvn.transform(region); 6175 } 6176 6177 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6178 // Return node representing slow path of predicate check. 6179 // the pseudo code we want to emulate with this predicate is: 6180 // for encryption: 6181 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6182 // for decryption: 6183 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6184 // note cipher==plain is more conservative than the original java code but that's OK 6185 // 6186 6187 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6188 // The receiver was checked for NULL already. 6189 Node* objCTR = argument(0); 6190 6191 // Load embeddedCipher field of CipherBlockChaining object. 6192 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6193 6194 // get AESCrypt klass for instanceOf check 6195 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6196 // will have same classloader as CipherBlockChaining object 6197 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6198 assert(tinst != NULL, "CTRobj is null"); 6199 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6200 6201 // we want to do an instanceof comparison against the AESCrypt class 6202 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6203 if (!klass_AESCrypt->is_loaded()) { 6204 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6205 Node* ctrl = control(); 6206 set_control(top()); // no regular fast path 6207 return ctrl; 6208 } 6209 6210 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6211 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6212 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6213 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6214 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6215 6216 return instof_false; // even if it is NULL 6217 } 6218 6219 //------------------------------inline_ghash_processBlocks 6220 bool LibraryCallKit::inline_ghash_processBlocks() { 6221 address stubAddr; 6222 const char *stubName; 6223 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6224 6225 stubAddr = StubRoutines::ghash_processBlocks(); 6226 stubName = "ghash_processBlocks"; 6227 6228 Node* data = argument(0); 6229 Node* offset = argument(1); 6230 Node* len = argument(2); 6231 Node* state = argument(3); 6232 Node* subkeyH = argument(4); 6233 6234 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6235 assert(state_start, "state is NULL"); 6236 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6237 assert(subkeyH_start, "subkeyH is NULL"); 6238 Node* data_start = array_element_address(data, offset, T_BYTE); 6239 assert(data_start, "data is NULL"); 6240 6241 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6242 OptoRuntime::ghash_processBlocks_Type(), 6243 stubAddr, stubName, TypePtr::BOTTOM, 6244 state_start, subkeyH_start, data_start, len); 6245 return true; 6246 } 6247 6248 //------------------------------inline_sha_implCompress----------------------- 6249 // 6250 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6251 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6252 // 6253 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6254 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6255 // 6256 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6257 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6258 // 6259 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6260 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6261 6262 Node* sha_obj = argument(0); 6263 Node* src = argument(1); // type oop 6264 Node* ofs = argument(2); // type int 6265 6266 const Type* src_type = src->Value(&_gvn); 6267 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6268 if (top_src == NULL || top_src->klass() == NULL) { 6269 // failed array check 6270 return false; 6271 } 6272 // Figure out the size and type of the elements we will be copying. 6273 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6274 if (src_elem != T_BYTE) { 6275 return false; 6276 } 6277 // 'src_start' points to src array + offset 6278 Node* src_start = array_element_address(src, ofs, src_elem); 6279 Node* state = NULL; 6280 address stubAddr; 6281 const char *stubName; 6282 6283 switch(id) { 6284 case vmIntrinsics::_sha_implCompress: 6285 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6286 state = get_state_from_sha_object(sha_obj); 6287 stubAddr = StubRoutines::sha1_implCompress(); 6288 stubName = "sha1_implCompress"; 6289 break; 6290 case vmIntrinsics::_sha2_implCompress: 6291 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6292 state = get_state_from_sha_object(sha_obj); 6293 stubAddr = StubRoutines::sha256_implCompress(); 6294 stubName = "sha256_implCompress"; 6295 break; 6296 case vmIntrinsics::_sha5_implCompress: 6297 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6298 state = get_state_from_sha5_object(sha_obj); 6299 stubAddr = StubRoutines::sha512_implCompress(); 6300 stubName = "sha512_implCompress"; 6301 break; 6302 default: 6303 fatal_unexpected_iid(id); 6304 return false; 6305 } 6306 if (state == NULL) return false; 6307 6308 // Call the stub. 6309 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6310 stubAddr, stubName, TypePtr::BOTTOM, 6311 src_start, state); 6312 6313 return true; 6314 } 6315 6316 //------------------------------inline_digestBase_implCompressMB----------------------- 6317 // 6318 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6319 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6320 // 6321 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6322 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6323 "need SHA1/SHA256/SHA512 instruction support"); 6324 assert((uint)predicate < 3, "sanity"); 6325 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6326 6327 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6328 Node* src = argument(1); // byte[] array 6329 Node* ofs = argument(2); // type int 6330 Node* limit = argument(3); // type int 6331 6332 const Type* src_type = src->Value(&_gvn); 6333 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6334 if (top_src == NULL || top_src->klass() == NULL) { 6335 // failed array check 6336 return false; 6337 } 6338 // Figure out the size and type of the elements we will be copying. 6339 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6340 if (src_elem != T_BYTE) { 6341 return false; 6342 } 6343 // 'src_start' points to src array + offset 6344 Node* src_start = array_element_address(src, ofs, src_elem); 6345 6346 const char* klass_SHA_name = NULL; 6347 const char* stub_name = NULL; 6348 address stub_addr = NULL; 6349 bool long_state = false; 6350 6351 switch (predicate) { 6352 case 0: 6353 if (UseSHA1Intrinsics) { 6354 klass_SHA_name = "sun/security/provider/SHA"; 6355 stub_name = "sha1_implCompressMB"; 6356 stub_addr = StubRoutines::sha1_implCompressMB(); 6357 } 6358 break; 6359 case 1: 6360 if (UseSHA256Intrinsics) { 6361 klass_SHA_name = "sun/security/provider/SHA2"; 6362 stub_name = "sha256_implCompressMB"; 6363 stub_addr = StubRoutines::sha256_implCompressMB(); 6364 } 6365 break; 6366 case 2: 6367 if (UseSHA512Intrinsics) { 6368 klass_SHA_name = "sun/security/provider/SHA5"; 6369 stub_name = "sha512_implCompressMB"; 6370 stub_addr = StubRoutines::sha512_implCompressMB(); 6371 long_state = true; 6372 } 6373 break; 6374 default: 6375 fatal("unknown SHA intrinsic predicate: %d", predicate); 6376 } 6377 if (klass_SHA_name != NULL) { 6378 // get DigestBase klass to lookup for SHA klass 6379 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6380 assert(tinst != NULL, "digestBase_obj is not instance???"); 6381 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6382 6383 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6384 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6385 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6386 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6387 } 6388 return false; 6389 } 6390 //------------------------------inline_sha_implCompressMB----------------------- 6391 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6392 bool long_state, address stubAddr, const char *stubName, 6393 Node* src_start, Node* ofs, Node* limit) { 6394 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6395 const TypeOopPtr* xtype = aklass->as_instance_type(); 6396 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6397 sha_obj = _gvn.transform(sha_obj); 6398 6399 Node* state; 6400 if (long_state) { 6401 state = get_state_from_sha5_object(sha_obj); 6402 } else { 6403 state = get_state_from_sha_object(sha_obj); 6404 } 6405 if (state == NULL) return false; 6406 6407 // Call the stub. 6408 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6409 OptoRuntime::digestBase_implCompressMB_Type(), 6410 stubAddr, stubName, TypePtr::BOTTOM, 6411 src_start, state, ofs, limit); 6412 // return ofs (int) 6413 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6414 set_result(result); 6415 6416 return true; 6417 } 6418 6419 //------------------------------get_state_from_sha_object----------------------- 6420 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6421 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6422 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6423 if (sha_state == NULL) return (Node *) NULL; 6424 6425 // now have the array, need to get the start address of the state array 6426 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6427 return state; 6428 } 6429 6430 //------------------------------get_state_from_sha5_object----------------------- 6431 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6432 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6433 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6434 if (sha_state == NULL) return (Node *) NULL; 6435 6436 // now have the array, need to get the start address of the state array 6437 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6438 return state; 6439 } 6440 6441 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6442 // Return node representing slow path of predicate check. 6443 // the pseudo code we want to emulate with this predicate is: 6444 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6445 // 6446 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6447 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6448 "need SHA1/SHA256/SHA512 instruction support"); 6449 assert((uint)predicate < 3, "sanity"); 6450 6451 // The receiver was checked for NULL already. 6452 Node* digestBaseObj = argument(0); 6453 6454 // get DigestBase klass for instanceOf check 6455 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6456 assert(tinst != NULL, "digestBaseObj is null"); 6457 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6458 6459 const char* klass_SHA_name = NULL; 6460 switch (predicate) { 6461 case 0: 6462 if (UseSHA1Intrinsics) { 6463 // we want to do an instanceof comparison against the SHA class 6464 klass_SHA_name = "sun/security/provider/SHA"; 6465 } 6466 break; 6467 case 1: 6468 if (UseSHA256Intrinsics) { 6469 // we want to do an instanceof comparison against the SHA2 class 6470 klass_SHA_name = "sun/security/provider/SHA2"; 6471 } 6472 break; 6473 case 2: 6474 if (UseSHA512Intrinsics) { 6475 // we want to do an instanceof comparison against the SHA5 class 6476 klass_SHA_name = "sun/security/provider/SHA5"; 6477 } 6478 break; 6479 default: 6480 fatal("unknown SHA intrinsic predicate: %d", predicate); 6481 } 6482 6483 ciKlass* klass_SHA = NULL; 6484 if (klass_SHA_name != NULL) { 6485 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6486 } 6487 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6488 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6489 Node* ctrl = control(); 6490 set_control(top()); // no intrinsic path 6491 return ctrl; 6492 } 6493 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6494 6495 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6496 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6497 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6498 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6499 6500 return instof_false; // even if it is NULL 6501 } 6502 6503 bool LibraryCallKit::inline_profileBoolean() { 6504 Node* counts = argument(1); 6505 const TypeAryPtr* ary = NULL; 6506 ciArray* aobj = NULL; 6507 if (counts->is_Con() 6508 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6509 && (aobj = ary->const_oop()->as_array()) != NULL 6510 && (aobj->length() == 2)) { 6511 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6512 jint false_cnt = aobj->element_value(0).as_int(); 6513 jint true_cnt = aobj->element_value(1).as_int(); 6514 6515 if (C->log() != NULL) { 6516 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6517 false_cnt, true_cnt); 6518 } 6519 6520 if (false_cnt + true_cnt == 0) { 6521 // According to profile, never executed. 6522 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6523 Deoptimization::Action_reinterpret); 6524 return true; 6525 } 6526 6527 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6528 // is a number of each value occurrences. 6529 Node* result = argument(0); 6530 if (false_cnt == 0 || true_cnt == 0) { 6531 // According to profile, one value has been never seen. 6532 int expected_val = (false_cnt == 0) ? 1 : 0; 6533 6534 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6535 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6536 6537 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6538 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6539 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6540 6541 { // Slow path: uncommon trap for never seen value and then reexecute 6542 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6543 // the value has been seen at least once. 6544 PreserveJVMState pjvms(this); 6545 PreserveReexecuteState preexecs(this); 6546 jvms()->set_should_reexecute(true); 6547 6548 set_control(slow_path); 6549 set_i_o(i_o()); 6550 6551 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6552 Deoptimization::Action_reinterpret); 6553 } 6554 // The guard for never seen value enables sharpening of the result and 6555 // returning a constant. It allows to eliminate branches on the same value 6556 // later on. 6557 set_control(fast_path); 6558 result = intcon(expected_val); 6559 } 6560 // Stop profiling. 6561 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6562 // By replacing method body with profile data (represented as ProfileBooleanNode 6563 // on IR level) we effectively disable profiling. 6564 // It enables full speed execution once optimized code is generated. 6565 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6566 C->record_for_igvn(profile); 6567 set_result(profile); 6568 return true; 6569 } else { 6570 // Continue profiling. 6571 // Profile data isn't available at the moment. So, execute method's bytecode version. 6572 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6573 // is compiled and counters aren't available since corresponding MethodHandle 6574 // isn't a compile-time constant. 6575 return false; 6576 } 6577 } 6578 6579 bool LibraryCallKit::inline_isCompileConstant() { 6580 Node* n = argument(0); 6581 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6582 return true; 6583 } 6584 6585 bool LibraryCallKit::inline_deoptimize() { 6586 assert(WhiteBoxAPI, ""); 6587 PreserveReexecuteState preexecs(this); 6588 jvms()->set_should_reexecute(false); 6589 uncommon_trap(Deoptimization::Reason_intrinsic, 6590 Deoptimization::Action_none); 6591 return true; 6592 }