1 /* 2 * Copyright (c) 1999, 2018, 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 "ci/ciUtilities.inline.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "classfile/vmSymbols.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/compileLog.hpp" 32 #include "gc/shared/barrierSet.hpp" 33 #include "jfr/support/jfrIntrinsics.hpp" 34 #include "memory/resourceArea.hpp" 35 #include "oops/objArrayKlass.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/arraycopynode.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/castnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/convertnode.hpp" 43 #include "opto/countbitsnode.hpp" 44 #include "opto/intrinsicnode.hpp" 45 #include "opto/idealKit.hpp" 46 #include "opto/mathexactnode.hpp" 47 #include "opto/movenode.hpp" 48 #include "opto/mulnode.hpp" 49 #include "opto/narrowptrnode.hpp" 50 #include "opto/opaquenode.hpp" 51 #include "opto/parse.hpp" 52 #include "opto/runtime.hpp" 53 #include "opto/rootnode.hpp" 54 #include "opto/subnode.hpp" 55 #include "opto/valuetypenode.hpp" 56 #include "prims/nativeLookup.hpp" 57 #include "prims/unsafe.hpp" 58 #include "runtime/objectMonitor.hpp" 59 #include "runtime/sharedRuntime.hpp" 60 #include "utilities/macros.hpp" 61 62 63 class LibraryIntrinsic : public InlineCallGenerator { 64 // Extend the set of intrinsics known to the runtime: 65 public: 66 private: 67 bool _is_virtual; 68 bool _does_virtual_dispatch; 69 int8_t _predicates_count; // Intrinsic is predicated by several conditions 70 int8_t _last_predicate; // Last generated predicate 71 vmIntrinsics::ID _intrinsic_id; 72 73 public: 74 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 75 : InlineCallGenerator(m), 76 _is_virtual(is_virtual), 77 _does_virtual_dispatch(does_virtual_dispatch), 78 _predicates_count((int8_t)predicates_count), 79 _last_predicate((int8_t)-1), 80 _intrinsic_id(id) 81 { 82 } 83 virtual bool is_intrinsic() const { return true; } 84 virtual bool is_virtual() const { return _is_virtual; } 85 virtual bool is_predicated() const { return _predicates_count > 0; } 86 virtual int predicates_count() const { return _predicates_count; } 87 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 88 virtual JVMState* generate(JVMState* jvms); 89 virtual Node* generate_predicate(JVMState* jvms, int predicate); 90 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 91 }; 92 93 94 // Local helper class for LibraryIntrinsic: 95 class LibraryCallKit : public GraphKit { 96 private: 97 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 98 Node* _result; // the result node, if any 99 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 100 101 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type); 102 103 public: 104 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 105 : GraphKit(jvms), 106 _intrinsic(intrinsic), 107 _result(NULL) 108 { 109 // Check if this is a root compile. In that case we don't have a caller. 110 if (!jvms->has_method()) { 111 _reexecute_sp = sp(); 112 } else { 113 // Find out how many arguments the interpreter needs when deoptimizing 114 // and save the stack pointer value so it can used by uncommon_trap. 115 // We find the argument count by looking at the declared signature. 116 bool ignored_will_link; 117 ciSignature* declared_signature = NULL; 118 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 119 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 120 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 121 } 122 } 123 124 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 125 126 ciMethod* caller() const { return jvms()->method(); } 127 int bci() const { return jvms()->bci(); } 128 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 129 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 130 ciMethod* callee() const { return _intrinsic->method(); } 131 132 bool try_to_inline(int predicate); 133 Node* try_to_predicate(int predicate); 134 135 void push_result() { 136 // Push the result onto the stack. 137 if (!stopped() && result() != NULL) { 138 BasicType bt = result()->bottom_type()->basic_type(); 139 push_node(bt, result()); 140 } 141 } 142 143 private: 144 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 145 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 146 } 147 148 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 149 void set_result(RegionNode* region, PhiNode* value); 150 Node* result() { return _result; } 151 152 virtual int reexecute_sp() { return _reexecute_sp; } 153 154 // Helper functions to inline natives 155 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 156 Node* generate_slow_guard(Node* test, RegionNode* region); 157 Node* generate_fair_guard(Node* test, RegionNode* region); 158 Node* generate_negative_guard(Node* index, RegionNode* region, 159 // resulting CastII of index: 160 Node* *pos_index = NULL); 161 Node* generate_limit_guard(Node* offset, Node* subseq_length, 162 Node* array_length, 163 RegionNode* region); 164 void generate_string_range_check(Node* array, Node* offset, 165 Node* length, bool char_count); 166 Node* generate_current_thread(Node* &tls_output); 167 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 168 RegionNode* region, int null_path, 169 int offset); 170 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 171 RegionNode* region, int null_path) { 172 int offset = java_lang_Class::klass_offset_in_bytes(); 173 return load_klass_from_mirror_common(mirror, never_see_null, 174 region, null_path, 175 offset); 176 } 177 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 178 RegionNode* region, int null_path) { 179 int offset = java_lang_Class::array_klass_offset_in_bytes(); 180 return load_klass_from_mirror_common(mirror, never_see_null, 181 region, null_path, 182 offset); 183 } 184 Node* generate_access_flags_guard(Node* kls, 185 int modifier_mask, int modifier_bits, 186 RegionNode* region); 187 Node* generate_interface_guard(Node* kls, RegionNode* region); 188 Node* generate_value_guard(Node* kls, RegionNode* region); 189 190 enum ArrayKind { 191 AnyArray, 192 NonArray, 193 ObjectArray, 194 NonObjectArray, 195 TypeArray, 196 ValueArray 197 }; 198 199 Node* generate_array_guard(Node* kls, RegionNode* region) { 200 return generate_array_guard_common(kls, region, AnyArray); 201 } 202 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 203 return generate_array_guard_common(kls, region, NonArray); 204 } 205 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 206 return generate_array_guard_common(kls, region, ObjectArray); 207 } 208 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 209 return generate_array_guard_common(kls, region, NonObjectArray); 210 } 211 Node* generate_typeArray_guard(Node* kls, RegionNode* region) { 212 return generate_array_guard_common(kls, region, TypeArray); 213 } 214 Node* generate_valueArray_guard(Node* kls, RegionNode* region) { 215 return generate_array_guard_common(kls, region, ValueArray); 216 } 217 Node* generate_array_guard_common(Node* kls, RegionNode* region, ArrayKind kind); 218 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 219 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 220 bool is_virtual = false, bool is_static = false); 221 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 222 return generate_method_call(method_id, false, true); 223 } 224 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 225 return generate_method_call(method_id, true, false); 226 } 227 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 228 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 229 230 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 231 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 232 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 233 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 234 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 235 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); 236 bool inline_string_indexOfChar(); 237 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 238 bool inline_string_toBytesU(); 239 bool inline_string_getCharsU(); 240 bool inline_string_copy(bool compress); 241 bool inline_string_char_access(bool is_store); 242 Node* round_double_node(Node* n); 243 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 244 bool inline_math_native(vmIntrinsics::ID id); 245 bool inline_math(vmIntrinsics::ID id); 246 template <typename OverflowOp> 247 bool inline_math_overflow(Node* arg1, Node* arg2); 248 void inline_math_mathExact(Node* math, Node* test); 249 bool inline_math_addExactI(bool is_increment); 250 bool inline_math_addExactL(bool is_increment); 251 bool inline_math_multiplyExactI(); 252 bool inline_math_multiplyExactL(); 253 bool inline_math_multiplyHigh(); 254 bool inline_math_negateExactI(); 255 bool inline_math_negateExactL(); 256 bool inline_math_subtractExactI(bool is_decrement); 257 bool inline_math_subtractExactL(bool is_decrement); 258 bool inline_min_max(vmIntrinsics::ID id); 259 bool inline_notify(vmIntrinsics::ID id); 260 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 261 // This returns Type::AnyPtr, RawPtr, or OopPtr. 262 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type); 263 Node* make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type = T_ILLEGAL, bool can_cast = false); 264 265 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind; 266 DecoratorSet mo_decorator_for_access_kind(AccessKind kind); 267 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned); 268 static bool klass_needs_init_guard(Node* kls); 269 bool inline_unsafe_allocate(); 270 bool inline_unsafe_newArray(bool uninitialized); 271 bool inline_unsafe_copyMemory(); 272 bool inline_native_currentThread(); 273 274 bool inline_native_time_funcs(address method, const char* funcName); 275 #ifdef JFR_HAVE_INTRINSICS 276 bool inline_native_classID(); 277 bool inline_native_getEventWriter(); 278 #endif 279 bool inline_native_isInterrupted(); 280 bool inline_native_Class_query(vmIntrinsics::ID id); 281 bool inline_native_subtype_check(); 282 bool inline_native_getLength(); 283 bool inline_array_copyOf(bool is_copyOfRange); 284 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 285 bool inline_preconditions_checkIndex(); 286 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array); 287 bool inline_native_clone(bool is_virtual); 288 bool inline_native_Reflection_getCallerClass(); 289 // Helper function for inlining native object hash method 290 bool inline_native_hashcode(bool is_virtual, bool is_static); 291 bool inline_native_getClass(); 292 293 // Helper functions for inlining arraycopy 294 bool inline_arraycopy(); 295 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 296 RegionNode* slow_region); 297 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 298 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp, 299 uint new_idx); 300 301 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind; 302 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind); 303 bool inline_unsafe_fence(vmIntrinsics::ID id); 304 bool inline_onspinwait(); 305 bool inline_fp_conversions(vmIntrinsics::ID id); 306 bool inline_number_methods(vmIntrinsics::ID id); 307 bool inline_reference_get(); 308 bool inline_Class_cast(); 309 bool inline_aescrypt_Block(vmIntrinsics::ID id); 310 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 311 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 312 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 313 Node* inline_counterMode_AESCrypt_predicate(); 314 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 315 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 316 bool inline_ghash_processBlocks(); 317 bool inline_base64_encodeBlock(); 318 bool inline_sha_implCompress(vmIntrinsics::ID id); 319 bool inline_digestBase_implCompressMB(int predicate); 320 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 321 bool long_state, address stubAddr, const char *stubName, 322 Node* src_start, Node* ofs, Node* limit); 323 Node* get_state_from_sha_object(Node *sha_object); 324 Node* get_state_from_sha5_object(Node *sha_object); 325 Node* inline_digestBase_implCompressMB_predicate(int predicate); 326 bool inline_encodeISOArray(); 327 bool inline_updateCRC32(); 328 bool inline_updateBytesCRC32(); 329 bool inline_updateByteBufferCRC32(); 330 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 331 bool inline_updateBytesCRC32C(); 332 bool inline_updateDirectByteBufferCRC32C(); 333 bool inline_updateBytesAdler32(); 334 bool inline_updateByteBufferAdler32(); 335 bool inline_multiplyToLen(); 336 bool inline_hasNegatives(); 337 bool inline_squareToLen(); 338 bool inline_mulAdd(); 339 bool inline_montgomeryMultiply(); 340 bool inline_montgomerySquare(); 341 bool inline_vectorizedMismatch(); 342 bool inline_fma(vmIntrinsics::ID id); 343 bool inline_character_compare(vmIntrinsics::ID id); 344 bool inline_fp_min_max(vmIntrinsics::ID id); 345 346 bool inline_profileBoolean(); 347 bool inline_isCompileConstant(); 348 void clear_upper_avx() { 349 #ifdef X86 350 if (UseAVX >= 2) { 351 C->set_clear_upper_avx(true); 352 } 353 #endif 354 } 355 }; 356 357 //---------------------------make_vm_intrinsic---------------------------- 358 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 359 vmIntrinsics::ID id = m->intrinsic_id(); 360 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 361 362 if (!m->is_loaded()) { 363 // Do not attempt to inline unloaded methods. 364 return NULL; 365 } 366 367 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 368 bool is_available = false; 369 370 { 371 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 372 // the compiler must transition to '_thread_in_vm' state because both 373 // methods access VM-internal data. 374 VM_ENTRY_MARK; 375 methodHandle mh(THREAD, m->get_Method()); 376 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) && 377 !C->directive()->is_intrinsic_disabled(mh) && 378 !vmIntrinsics::is_disabled_by_flags(mh); 379 380 } 381 382 if (is_available) { 383 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 384 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 385 return new LibraryIntrinsic(m, is_virtual, 386 vmIntrinsics::predicates_needed(id), 387 vmIntrinsics::does_virtual_dispatch(id), 388 (vmIntrinsics::ID) id); 389 } else { 390 return NULL; 391 } 392 } 393 394 //----------------------register_library_intrinsics----------------------- 395 // Initialize this file's data structures, for each Compile instance. 396 void Compile::register_library_intrinsics() { 397 // Nothing to do here. 398 } 399 400 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 401 LibraryCallKit kit(jvms, this); 402 Compile* C = kit.C; 403 int nodes = C->unique(); 404 #ifndef PRODUCT 405 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 406 char buf[1000]; 407 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 408 tty->print_cr("Intrinsic %s", str); 409 } 410 #endif 411 ciMethod* callee = kit.callee(); 412 const int bci = kit.bci(); 413 414 // Try to inline the intrinsic. 415 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 416 kit.try_to_inline(_last_predicate)) { 417 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)" 418 : "(intrinsic)"; 419 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 420 if (C->print_intrinsics() || C->print_inlining()) { 421 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 422 } 423 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 424 if (C->log()) { 425 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 426 vmIntrinsics::name_at(intrinsic_id()), 427 (is_virtual() ? " virtual='1'" : ""), 428 C->unique() - nodes); 429 } 430 // Push the result from the inlined method onto the stack. 431 kit.push_result(); 432 C->print_inlining_update(this); 433 return kit.transfer_exceptions_into_jvms(); 434 } 435 436 // The intrinsic bailed out 437 if (jvms->has_method()) { 438 // Not a root compile. 439 const char* msg; 440 if (callee->intrinsic_candidate()) { 441 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 442 } else { 443 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 444 : "failed to inline (intrinsic), method not annotated"; 445 } 446 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg); 447 if (C->print_intrinsics() || C->print_inlining()) { 448 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 449 } 450 } else { 451 // Root compile 452 ResourceMark rm; 453 stringStream msg_stream; 454 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 455 vmIntrinsics::name_at(intrinsic_id()), 456 is_virtual() ? " (virtual)" : "", bci); 457 const char *msg = msg_stream.as_string(); 458 log_debug(jit, inlining)("%s", msg); 459 if (C->print_intrinsics() || C->print_inlining()) { 460 tty->print("%s", msg); 461 } 462 } 463 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 464 C->print_inlining_update(this); 465 return NULL; 466 } 467 468 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 469 LibraryCallKit kit(jvms, this); 470 Compile* C = kit.C; 471 int nodes = C->unique(); 472 _last_predicate = predicate; 473 #ifndef PRODUCT 474 assert(is_predicated() && predicate < predicates_count(), "sanity"); 475 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 476 char buf[1000]; 477 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 478 tty->print_cr("Predicate for intrinsic %s", str); 479 } 480 #endif 481 ciMethod* callee = kit.callee(); 482 const int bci = kit.bci(); 483 484 Node* slow_ctl = kit.try_to_predicate(predicate); 485 if (!kit.failing()) { 486 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)" 487 : "(intrinsic, predicate)"; 488 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 489 if (C->print_intrinsics() || C->print_inlining()) { 490 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 491 } 492 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 493 if (C->log()) { 494 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 495 vmIntrinsics::name_at(intrinsic_id()), 496 (is_virtual() ? " virtual='1'" : ""), 497 C->unique() - nodes); 498 } 499 return slow_ctl; // Could be NULL if the check folds. 500 } 501 502 // The intrinsic bailed out 503 if (jvms->has_method()) { 504 // Not a root compile. 505 const char* msg = "failed to generate predicate for intrinsic"; 506 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg); 507 if (C->print_intrinsics() || C->print_inlining()) { 508 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 509 } 510 } else { 511 // Root compile 512 ResourceMark rm; 513 stringStream msg_stream; 514 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 515 vmIntrinsics::name_at(intrinsic_id()), 516 is_virtual() ? " (virtual)" : "", bci); 517 const char *msg = msg_stream.as_string(); 518 log_debug(jit, inlining)("%s", msg); 519 if (C->print_intrinsics() || C->print_inlining()) { 520 C->print_inlining_stream()->print("%s", msg); 521 } 522 } 523 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 524 return NULL; 525 } 526 527 bool LibraryCallKit::try_to_inline(int predicate) { 528 // Handle symbolic names for otherwise undistinguished boolean switches: 529 const bool is_store = true; 530 const bool is_compress = true; 531 const bool is_static = true; 532 const bool is_volatile = true; 533 534 if (!jvms()->has_method()) { 535 // Root JVMState has a null method. 536 assert(map()->memory()->Opcode() == Op_Parm, ""); 537 // Insert the memory aliasing node 538 set_all_memory(reset_memory()); 539 } 540 assert(merged_memory(), ""); 541 542 543 switch (intrinsic_id()) { 544 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 545 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 546 case vmIntrinsics::_getClass: return inline_native_getClass(); 547 548 case vmIntrinsics::_dsin: 549 case vmIntrinsics::_dcos: 550 case vmIntrinsics::_dtan: 551 case vmIntrinsics::_dabs: 552 case vmIntrinsics::_datan2: 553 case vmIntrinsics::_dsqrt: 554 case vmIntrinsics::_dexp: 555 case vmIntrinsics::_dlog: 556 case vmIntrinsics::_dlog10: 557 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 558 559 case vmIntrinsics::_min: 560 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 561 562 case vmIntrinsics::_notify: 563 case vmIntrinsics::_notifyAll: 564 return inline_notify(intrinsic_id()); 565 566 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 567 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 568 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 569 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 570 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 571 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 572 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 573 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 574 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh(); 575 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 576 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 577 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 578 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 579 580 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 581 582 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 583 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 584 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 585 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 586 587 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 588 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 589 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 590 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 591 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 592 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 593 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 594 595 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 596 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 597 598 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 599 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 600 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 601 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 602 603 case vmIntrinsics::_compressStringC: 604 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 605 case vmIntrinsics::_inflateStringC: 606 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 607 608 case vmIntrinsics::_getReference: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 609 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 610 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 611 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 612 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 613 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 614 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 615 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 616 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 617 618 case vmIntrinsics::_putReference: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 619 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 620 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 621 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 622 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 623 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 624 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 625 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 626 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 627 628 case vmIntrinsics::_getReferenceVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 629 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 630 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 631 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 632 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 633 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 634 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 635 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 636 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 637 638 case vmIntrinsics::_putReferenceVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 639 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 640 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 641 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 642 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 643 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 644 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 645 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 646 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 647 648 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 649 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 650 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 651 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 652 653 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 654 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 655 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 656 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 657 658 case vmIntrinsics::_getReferenceAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 659 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 660 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 661 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 662 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 663 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 664 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 665 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 666 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 667 668 case vmIntrinsics::_putReferenceRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 669 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 670 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 671 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 672 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 673 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 674 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 675 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 676 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 677 678 case vmIntrinsics::_getReferenceOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 679 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 680 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 681 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 682 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 683 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 684 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 685 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 686 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 687 688 case vmIntrinsics::_putReferenceOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 689 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 690 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 691 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 692 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 693 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 694 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 695 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 696 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 697 698 case vmIntrinsics::_compareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 699 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); 700 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); 701 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 702 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 703 704 case vmIntrinsics::_weakCompareAndSetReferencePlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 705 case vmIntrinsics::_weakCompareAndSetReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 706 case vmIntrinsics::_weakCompareAndSetReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 707 case vmIntrinsics::_weakCompareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 708 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); 709 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); 710 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); 711 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); 712 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); 713 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); 714 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); 715 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); 716 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 717 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 718 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 719 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 720 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 721 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 722 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 723 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 724 725 case vmIntrinsics::_compareAndExchangeReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 726 case vmIntrinsics::_compareAndExchangeReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 727 case vmIntrinsics::_compareAndExchangeReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 728 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); 729 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); 730 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); 731 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); 732 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); 733 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); 734 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 735 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 736 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 737 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 738 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 739 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 740 741 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); 742 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); 743 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 744 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 745 746 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); 747 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); 748 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 749 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 750 case vmIntrinsics::_getAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 751 752 case vmIntrinsics::_loadFence: 753 case vmIntrinsics::_storeFence: 754 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 755 756 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 757 758 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 759 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 760 761 #ifdef JFR_HAVE_INTRINSICS 762 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime"); 763 case vmIntrinsics::_getClassId: return inline_native_classID(); 764 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter(); 765 #endif 766 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 767 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 768 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 769 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 770 case vmIntrinsics::_getLength: return inline_native_getLength(); 771 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 772 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 773 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 774 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 775 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 776 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 777 778 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 779 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 780 781 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 782 783 case vmIntrinsics::_isInstance: 784 case vmIntrinsics::_getModifiers: 785 case vmIntrinsics::_isInterface: 786 case vmIntrinsics::_isArray: 787 case vmIntrinsics::_isPrimitive: 788 case vmIntrinsics::_getSuperclass: 789 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 790 791 case vmIntrinsics::_floatToRawIntBits: 792 case vmIntrinsics::_floatToIntBits: 793 case vmIntrinsics::_intBitsToFloat: 794 case vmIntrinsics::_doubleToRawLongBits: 795 case vmIntrinsics::_doubleToLongBits: 796 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 797 798 case vmIntrinsics::_numberOfLeadingZeros_i: 799 case vmIntrinsics::_numberOfLeadingZeros_l: 800 case vmIntrinsics::_numberOfTrailingZeros_i: 801 case vmIntrinsics::_numberOfTrailingZeros_l: 802 case vmIntrinsics::_bitCount_i: 803 case vmIntrinsics::_bitCount_l: 804 case vmIntrinsics::_reverseBytes_i: 805 case vmIntrinsics::_reverseBytes_l: 806 case vmIntrinsics::_reverseBytes_s: 807 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 808 809 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 810 811 case vmIntrinsics::_Reference_get: return inline_reference_get(); 812 813 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 814 815 case vmIntrinsics::_aescrypt_encryptBlock: 816 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 817 818 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 819 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 820 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 821 822 case vmIntrinsics::_counterMode_AESCrypt: 823 return inline_counterMode_AESCrypt(intrinsic_id()); 824 825 case vmIntrinsics::_sha_implCompress: 826 case vmIntrinsics::_sha2_implCompress: 827 case vmIntrinsics::_sha5_implCompress: 828 return inline_sha_implCompress(intrinsic_id()); 829 830 case vmIntrinsics::_digestBase_implCompressMB: 831 return inline_digestBase_implCompressMB(predicate); 832 833 case vmIntrinsics::_multiplyToLen: 834 return inline_multiplyToLen(); 835 836 case vmIntrinsics::_squareToLen: 837 return inline_squareToLen(); 838 839 case vmIntrinsics::_mulAdd: 840 return inline_mulAdd(); 841 842 case vmIntrinsics::_montgomeryMultiply: 843 return inline_montgomeryMultiply(); 844 case vmIntrinsics::_montgomerySquare: 845 return inline_montgomerySquare(); 846 847 case vmIntrinsics::_vectorizedMismatch: 848 return inline_vectorizedMismatch(); 849 850 case vmIntrinsics::_ghash_processBlocks: 851 return inline_ghash_processBlocks(); 852 case vmIntrinsics::_base64_encodeBlock: 853 return inline_base64_encodeBlock(); 854 855 case vmIntrinsics::_encodeISOArray: 856 case vmIntrinsics::_encodeByteISOArray: 857 return inline_encodeISOArray(); 858 859 case vmIntrinsics::_updateCRC32: 860 return inline_updateCRC32(); 861 case vmIntrinsics::_updateBytesCRC32: 862 return inline_updateBytesCRC32(); 863 case vmIntrinsics::_updateByteBufferCRC32: 864 return inline_updateByteBufferCRC32(); 865 866 case vmIntrinsics::_updateBytesCRC32C: 867 return inline_updateBytesCRC32C(); 868 case vmIntrinsics::_updateDirectByteBufferCRC32C: 869 return inline_updateDirectByteBufferCRC32C(); 870 871 case vmIntrinsics::_updateBytesAdler32: 872 return inline_updateBytesAdler32(); 873 case vmIntrinsics::_updateByteBufferAdler32: 874 return inline_updateByteBufferAdler32(); 875 876 case vmIntrinsics::_profileBoolean: 877 return inline_profileBoolean(); 878 case vmIntrinsics::_isCompileConstant: 879 return inline_isCompileConstant(); 880 881 case vmIntrinsics::_hasNegatives: 882 return inline_hasNegatives(); 883 884 case vmIntrinsics::_fmaD: 885 case vmIntrinsics::_fmaF: 886 return inline_fma(intrinsic_id()); 887 888 case vmIntrinsics::_isDigit: 889 case vmIntrinsics::_isLowerCase: 890 case vmIntrinsics::_isUpperCase: 891 case vmIntrinsics::_isWhitespace: 892 return inline_character_compare(intrinsic_id()); 893 894 case vmIntrinsics::_maxF: 895 case vmIntrinsics::_minF: 896 case vmIntrinsics::_maxD: 897 case vmIntrinsics::_minD: 898 return inline_fp_min_max(intrinsic_id()); 899 900 default: 901 // If you get here, it may be that someone has added a new intrinsic 902 // to the list in vmSymbols.hpp without implementing it here. 903 #ifndef PRODUCT 904 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 905 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 906 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 907 } 908 #endif 909 return false; 910 } 911 } 912 913 Node* LibraryCallKit::try_to_predicate(int predicate) { 914 if (!jvms()->has_method()) { 915 // Root JVMState has a null method. 916 assert(map()->memory()->Opcode() == Op_Parm, ""); 917 // Insert the memory aliasing node 918 set_all_memory(reset_memory()); 919 } 920 assert(merged_memory(), ""); 921 922 switch (intrinsic_id()) { 923 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 924 return inline_cipherBlockChaining_AESCrypt_predicate(false); 925 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 926 return inline_cipherBlockChaining_AESCrypt_predicate(true); 927 case vmIntrinsics::_counterMode_AESCrypt: 928 return inline_counterMode_AESCrypt_predicate(); 929 case vmIntrinsics::_digestBase_implCompressMB: 930 return inline_digestBase_implCompressMB_predicate(predicate); 931 932 default: 933 // If you get here, it may be that someone has added a new intrinsic 934 // to the list in vmSymbols.hpp without implementing it here. 935 #ifndef PRODUCT 936 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 937 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 938 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 939 } 940 #endif 941 Node* slow_ctl = control(); 942 set_control(top()); // No fast path instrinsic 943 return slow_ctl; 944 } 945 } 946 947 //------------------------------set_result------------------------------- 948 // Helper function for finishing intrinsics. 949 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 950 record_for_igvn(region); 951 set_control(_gvn.transform(region)); 952 set_result( _gvn.transform(value)); 953 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 954 } 955 956 //------------------------------generate_guard--------------------------- 957 // Helper function for generating guarded fast-slow graph structures. 958 // The given 'test', if true, guards a slow path. If the test fails 959 // then a fast path can be taken. (We generally hope it fails.) 960 // In all cases, GraphKit::control() is updated to the fast path. 961 // The returned value represents the control for the slow path. 962 // The return value is never 'top'; it is either a valid control 963 // or NULL if it is obvious that the slow path can never be taken. 964 // Also, if region and the slow control are not NULL, the slow edge 965 // is appended to the region. 966 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 967 if (stopped()) { 968 // Already short circuited. 969 return NULL; 970 } 971 972 // Build an if node and its projections. 973 // If test is true we take the slow path, which we assume is uncommon. 974 if (_gvn.type(test) == TypeInt::ZERO) { 975 // The slow branch is never taken. No need to build this guard. 976 return NULL; 977 } 978 979 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 980 981 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 982 if (if_slow == top()) { 983 // The slow branch is never taken. No need to build this guard. 984 return NULL; 985 } 986 987 if (region != NULL) 988 region->add_req(if_slow); 989 990 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 991 set_control(if_fast); 992 993 return if_slow; 994 } 995 996 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 997 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 998 } 999 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1000 return generate_guard(test, region, PROB_FAIR); 1001 } 1002 1003 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1004 Node* *pos_index) { 1005 if (stopped()) 1006 return NULL; // already stopped 1007 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1008 return NULL; // index is already adequately typed 1009 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1010 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1011 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1012 if (is_neg != NULL && pos_index != NULL) { 1013 // Emulate effect of Parse::adjust_map_after_if. 1014 Node* ccast = new CastIINode(index, TypeInt::POS); 1015 ccast->set_req(0, control()); 1016 (*pos_index) = _gvn.transform(ccast); 1017 } 1018 return is_neg; 1019 } 1020 1021 // Make sure that 'position' is a valid limit index, in [0..length]. 1022 // There are two equivalent plans for checking this: 1023 // A. (offset + copyLength) unsigned<= arrayLength 1024 // B. offset <= (arrayLength - copyLength) 1025 // We require that all of the values above, except for the sum and 1026 // difference, are already known to be non-negative. 1027 // Plan A is robust in the face of overflow, if offset and copyLength 1028 // are both hugely positive. 1029 // 1030 // Plan B is less direct and intuitive, but it does not overflow at 1031 // all, since the difference of two non-negatives is always 1032 // representable. Whenever Java methods must perform the equivalent 1033 // check they generally use Plan B instead of Plan A. 1034 // For the moment we use Plan A. 1035 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1036 Node* subseq_length, 1037 Node* array_length, 1038 RegionNode* region) { 1039 if (stopped()) 1040 return NULL; // already stopped 1041 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1042 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1043 return NULL; // common case of whole-array copy 1044 Node* last = subseq_length; 1045 if (!zero_offset) // last += offset 1046 last = _gvn.transform(new AddINode(last, offset)); 1047 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1048 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1049 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1050 return is_over; 1051 } 1052 1053 // Emit range checks for the given String.value byte array 1054 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 1055 if (stopped()) { 1056 return; // already stopped 1057 } 1058 RegionNode* bailout = new RegionNode(1); 1059 record_for_igvn(bailout); 1060 if (char_count) { 1061 // Convert char count to byte count 1062 count = _gvn.transform(new LShiftINode(count, intcon(1))); 1063 } 1064 1065 // Offset and count must not be negative 1066 generate_negative_guard(offset, bailout); 1067 generate_negative_guard(count, bailout); 1068 // Offset + count must not exceed length of array 1069 generate_limit_guard(offset, count, load_array_length(array), bailout); 1070 1071 if (bailout->req() > 1) { 1072 PreserveJVMState pjvms(this); 1073 set_control(_gvn.transform(bailout)); 1074 uncommon_trap(Deoptimization::Reason_intrinsic, 1075 Deoptimization::Action_maybe_recompile); 1076 } 1077 } 1078 1079 //--------------------------generate_current_thread-------------------- 1080 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1081 ciKlass* thread_klass = env()->Thread_klass(); 1082 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1083 Node* thread = _gvn.transform(new ThreadLocalNode()); 1084 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1085 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1086 tls_output = thread; 1087 return threadObj; 1088 } 1089 1090 1091 //------------------------------make_string_method_node------------------------ 1092 // Helper method for String intrinsic functions. This version is called with 1093 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 1094 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 1095 // containing the lengths of str1 and str2. 1096 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 1097 Node* result = NULL; 1098 switch (opcode) { 1099 case Op_StrIndexOf: 1100 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 1101 str1_start, cnt1, str2_start, cnt2, ae); 1102 break; 1103 case Op_StrComp: 1104 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 1105 str1_start, cnt1, str2_start, cnt2, ae); 1106 break; 1107 case Op_StrEquals: 1108 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 1109 // Use the constant length if there is one because optimized match rule may exist. 1110 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 1111 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 1112 break; 1113 default: 1114 ShouldNotReachHere(); 1115 return NULL; 1116 } 1117 1118 // All these intrinsics have checks. 1119 C->set_has_split_ifs(true); // Has chance for split-if optimization 1120 clear_upper_avx(); 1121 1122 return _gvn.transform(result); 1123 } 1124 1125 //------------------------------inline_string_compareTo------------------------ 1126 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 1127 Node* arg1 = argument(0); 1128 Node* arg2 = argument(1); 1129 1130 arg1 = must_be_not_null(arg1, true); 1131 arg2 = must_be_not_null(arg2, true); 1132 1133 arg1 = access_resolve(arg1, ACCESS_READ); 1134 arg2 = access_resolve(arg2, ACCESS_READ); 1135 1136 // Get start addr and length of first argument 1137 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1138 Node* arg1_cnt = load_array_length(arg1); 1139 1140 // Get start addr and length of second argument 1141 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1142 Node* arg2_cnt = load_array_length(arg2); 1143 1144 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1145 set_result(result); 1146 return true; 1147 } 1148 1149 //------------------------------inline_string_equals------------------------ 1150 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1151 Node* arg1 = argument(0); 1152 Node* arg2 = argument(1); 1153 1154 // paths (plus control) merge 1155 RegionNode* region = new RegionNode(3); 1156 Node* phi = new PhiNode(region, TypeInt::BOOL); 1157 1158 if (!stopped()) { 1159 1160 arg1 = must_be_not_null(arg1, true); 1161 arg2 = must_be_not_null(arg2, true); 1162 1163 arg1 = access_resolve(arg1, ACCESS_READ); 1164 arg2 = access_resolve(arg2, ACCESS_READ); 1165 1166 // Get start addr and length of first argument 1167 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1168 Node* arg1_cnt = load_array_length(arg1); 1169 1170 // Get start addr and length of second argument 1171 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1172 Node* arg2_cnt = load_array_length(arg2); 1173 1174 // Check for arg1_cnt != arg2_cnt 1175 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1176 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1177 Node* if_ne = generate_slow_guard(bol, NULL); 1178 if (if_ne != NULL) { 1179 phi->init_req(2, intcon(0)); 1180 region->init_req(2, if_ne); 1181 } 1182 1183 // Check for count == 0 is done by assembler code for StrEquals. 1184 1185 if (!stopped()) { 1186 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1187 phi->init_req(1, equals); 1188 region->init_req(1, control()); 1189 } 1190 } 1191 1192 // post merge 1193 set_control(_gvn.transform(region)); 1194 record_for_igvn(region); 1195 1196 set_result(_gvn.transform(phi)); 1197 return true; 1198 } 1199 1200 //------------------------------inline_array_equals---------------------------- 1201 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1202 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1203 Node* arg1 = argument(0); 1204 Node* arg2 = argument(1); 1205 1206 arg1 = access_resolve(arg1, ACCESS_READ); 1207 arg2 = access_resolve(arg2, ACCESS_READ); 1208 1209 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1210 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1211 clear_upper_avx(); 1212 1213 return true; 1214 } 1215 1216 //------------------------------inline_hasNegatives------------------------------ 1217 bool LibraryCallKit::inline_hasNegatives() { 1218 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1219 return false; 1220 } 1221 1222 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1223 // no receiver since it is static method 1224 Node* ba = argument(0); 1225 Node* offset = argument(1); 1226 Node* len = argument(2); 1227 1228 ba = must_be_not_null(ba, true); 1229 1230 // Range checks 1231 generate_string_range_check(ba, offset, len, false); 1232 if (stopped()) { 1233 return true; 1234 } 1235 ba = access_resolve(ba, ACCESS_READ); 1236 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1237 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1238 set_result(_gvn.transform(result)); 1239 return true; 1240 } 1241 1242 bool LibraryCallKit::inline_preconditions_checkIndex() { 1243 Node* index = argument(0); 1244 Node* length = argument(1); 1245 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1246 return false; 1247 } 1248 1249 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1250 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1251 1252 { 1253 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1254 uncommon_trap(Deoptimization::Reason_intrinsic, 1255 Deoptimization::Action_make_not_entrant); 1256 } 1257 1258 if (stopped()) { 1259 return false; 1260 } 1261 1262 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1263 BoolTest::mask btest = BoolTest::lt; 1264 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1265 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1266 _gvn.set_type(rc, rc->Value(&_gvn)); 1267 if (!rc_bool->is_Con()) { 1268 record_for_igvn(rc); 1269 } 1270 set_control(_gvn.transform(new IfTrueNode(rc))); 1271 { 1272 PreserveJVMState pjvms(this); 1273 set_control(_gvn.transform(new IfFalseNode(rc))); 1274 uncommon_trap(Deoptimization::Reason_range_check, 1275 Deoptimization::Action_make_not_entrant); 1276 } 1277 1278 if (stopped()) { 1279 return false; 1280 } 1281 1282 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1283 result->set_req(0, control()); 1284 result = _gvn.transform(result); 1285 set_result(result); 1286 replace_in_map(index, result); 1287 clear_upper_avx(); 1288 return true; 1289 } 1290 1291 //------------------------------inline_string_indexOf------------------------ 1292 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1293 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1294 return false; 1295 } 1296 Node* src = argument(0); 1297 Node* tgt = argument(1); 1298 1299 // Make the merge point 1300 RegionNode* result_rgn = new RegionNode(4); 1301 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1302 1303 src = must_be_not_null(src, true); 1304 tgt = must_be_not_null(tgt, true); 1305 1306 src = access_resolve(src, ACCESS_READ); 1307 tgt = access_resolve(tgt, ACCESS_READ); 1308 1309 // Get start addr and length of source string 1310 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1311 Node* src_count = load_array_length(src); 1312 1313 // Get start addr and length of substring 1314 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1315 Node* tgt_count = load_array_length(tgt); 1316 1317 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1318 // Divide src size by 2 if String is UTF16 encoded 1319 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1320 } 1321 if (ae == StrIntrinsicNode::UU) { 1322 // Divide substring size by 2 if String is UTF16 encoded 1323 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1324 } 1325 1326 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1327 if (result != NULL) { 1328 result_phi->init_req(3, result); 1329 result_rgn->init_req(3, control()); 1330 } 1331 set_control(_gvn.transform(result_rgn)); 1332 record_for_igvn(result_rgn); 1333 set_result(_gvn.transform(result_phi)); 1334 1335 return true; 1336 } 1337 1338 //-----------------------------inline_string_indexOf----------------------- 1339 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1340 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1341 return false; 1342 } 1343 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1344 return false; 1345 } 1346 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1347 Node* src = argument(0); // byte[] 1348 Node* src_count = argument(1); // char count 1349 Node* tgt = argument(2); // byte[] 1350 Node* tgt_count = argument(3); // char count 1351 Node* from_index = argument(4); // char index 1352 1353 src = must_be_not_null(src, true); 1354 tgt = must_be_not_null(tgt, true); 1355 1356 src = access_resolve(src, ACCESS_READ); 1357 tgt = access_resolve(tgt, ACCESS_READ); 1358 1359 // Multiply byte array index by 2 if String is UTF16 encoded 1360 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1361 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1362 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1363 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1364 1365 // Range checks 1366 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1367 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1368 if (stopped()) { 1369 return true; 1370 } 1371 1372 RegionNode* region = new RegionNode(5); 1373 Node* phi = new PhiNode(region, TypeInt::INT); 1374 1375 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1376 if (result != NULL) { 1377 // The result is index relative to from_index if substring was found, -1 otherwise. 1378 // Generate code which will fold into cmove. 1379 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1380 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1381 1382 Node* if_lt = generate_slow_guard(bol, NULL); 1383 if (if_lt != NULL) { 1384 // result == -1 1385 phi->init_req(3, result); 1386 region->init_req(3, if_lt); 1387 } 1388 if (!stopped()) { 1389 result = _gvn.transform(new AddINode(result, from_index)); 1390 phi->init_req(4, result); 1391 region->init_req(4, control()); 1392 } 1393 } 1394 1395 set_control(_gvn.transform(region)); 1396 record_for_igvn(region); 1397 set_result(_gvn.transform(phi)); 1398 clear_upper_avx(); 1399 1400 return true; 1401 } 1402 1403 // Create StrIndexOfNode with fast path checks 1404 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1405 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1406 // Check for substr count > string count 1407 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1408 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1409 Node* if_gt = generate_slow_guard(bol, NULL); 1410 if (if_gt != NULL) { 1411 phi->init_req(1, intcon(-1)); 1412 region->init_req(1, if_gt); 1413 } 1414 if (!stopped()) { 1415 // Check for substr count == 0 1416 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1417 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1418 Node* if_zero = generate_slow_guard(bol, NULL); 1419 if (if_zero != NULL) { 1420 phi->init_req(2, intcon(0)); 1421 region->init_req(2, if_zero); 1422 } 1423 } 1424 if (!stopped()) { 1425 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1426 } 1427 return NULL; 1428 } 1429 1430 //-----------------------------inline_string_indexOfChar----------------------- 1431 bool LibraryCallKit::inline_string_indexOfChar() { 1432 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1433 return false; 1434 } 1435 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1436 return false; 1437 } 1438 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1439 Node* src = argument(0); // byte[] 1440 Node* tgt = argument(1); // tgt is int ch 1441 Node* from_index = argument(2); 1442 Node* max = argument(3); 1443 1444 src = must_be_not_null(src, true); 1445 src = access_resolve(src, ACCESS_READ); 1446 1447 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1448 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1449 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1450 1451 // Range checks 1452 generate_string_range_check(src, src_offset, src_count, true); 1453 if (stopped()) { 1454 return true; 1455 } 1456 1457 RegionNode* region = new RegionNode(3); 1458 Node* phi = new PhiNode(region, TypeInt::INT); 1459 1460 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1461 C->set_has_split_ifs(true); // Has chance for split-if optimization 1462 _gvn.transform(result); 1463 1464 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1465 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1466 1467 Node* if_lt = generate_slow_guard(bol, NULL); 1468 if (if_lt != NULL) { 1469 // result == -1 1470 phi->init_req(2, result); 1471 region->init_req(2, if_lt); 1472 } 1473 if (!stopped()) { 1474 result = _gvn.transform(new AddINode(result, from_index)); 1475 phi->init_req(1, result); 1476 region->init_req(1, control()); 1477 } 1478 set_control(_gvn.transform(region)); 1479 record_for_igvn(region); 1480 set_result(_gvn.transform(phi)); 1481 1482 return true; 1483 } 1484 //---------------------------inline_string_copy--------------------- 1485 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1486 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1487 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1488 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1489 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1490 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1491 bool LibraryCallKit::inline_string_copy(bool compress) { 1492 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1493 return false; 1494 } 1495 int nargs = 5; // 2 oops, 3 ints 1496 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1497 1498 Node* src = argument(0); 1499 Node* src_offset = argument(1); 1500 Node* dst = argument(2); 1501 Node* dst_offset = argument(3); 1502 Node* length = argument(4); 1503 1504 // Check for allocation before we add nodes that would confuse 1505 // tightly_coupled_allocation() 1506 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1507 1508 // Figure out the size and type of the elements we will be copying. 1509 const Type* src_type = src->Value(&_gvn); 1510 const Type* dst_type = dst->Value(&_gvn); 1511 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1512 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1513 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1514 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1515 "Unsupported array types for inline_string_copy"); 1516 1517 src = must_be_not_null(src, true); 1518 dst = must_be_not_null(dst, true); 1519 1520 // Convert char[] offsets to byte[] offsets 1521 bool convert_src = (compress && src_elem == T_BYTE); 1522 bool convert_dst = (!compress && dst_elem == T_BYTE); 1523 if (convert_src) { 1524 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1525 } else if (convert_dst) { 1526 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1527 } 1528 1529 // Range checks 1530 generate_string_range_check(src, src_offset, length, convert_src); 1531 generate_string_range_check(dst, dst_offset, length, convert_dst); 1532 if (stopped()) { 1533 return true; 1534 } 1535 1536 src = access_resolve(src, ACCESS_READ); 1537 dst = access_resolve(dst, ACCESS_WRITE); 1538 1539 Node* src_start = array_element_address(src, src_offset, src_elem); 1540 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1541 // 'src_start' points to src array + scaled offset 1542 // 'dst_start' points to dst array + scaled offset 1543 Node* count = NULL; 1544 if (compress) { 1545 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1546 } else { 1547 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1548 } 1549 1550 if (alloc != NULL) { 1551 if (alloc->maybe_set_complete(&_gvn)) { 1552 // "You break it, you buy it." 1553 InitializeNode* init = alloc->initialization(); 1554 assert(init->is_complete(), "we just did this"); 1555 init->set_complete_with_arraycopy(); 1556 assert(dst->is_CheckCastPP(), "sanity"); 1557 assert(dst->in(0)->in(0) == init, "dest pinned"); 1558 } 1559 // Do not let stores that initialize this object be reordered with 1560 // a subsequent store that would make this object accessible by 1561 // other threads. 1562 // Record what AllocateNode this StoreStore protects so that 1563 // escape analysis can go from the MemBarStoreStoreNode to the 1564 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1565 // based on the escape status of the AllocateNode. 1566 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1567 } 1568 if (compress) { 1569 set_result(_gvn.transform(count)); 1570 } 1571 clear_upper_avx(); 1572 1573 return true; 1574 } 1575 1576 #ifdef _LP64 1577 #define XTOP ,top() /*additional argument*/ 1578 #else //_LP64 1579 #define XTOP /*no additional argument*/ 1580 #endif //_LP64 1581 1582 //------------------------inline_string_toBytesU-------------------------- 1583 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1584 bool LibraryCallKit::inline_string_toBytesU() { 1585 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1586 return false; 1587 } 1588 // Get the arguments. 1589 Node* value = argument(0); 1590 Node* offset = argument(1); 1591 Node* length = argument(2); 1592 1593 Node* newcopy = NULL; 1594 1595 // Set the original stack and the reexecute bit for the interpreter to reexecute 1596 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1597 { PreserveReexecuteState preexecs(this); 1598 jvms()->set_should_reexecute(true); 1599 1600 // Check if a null path was taken unconditionally. 1601 value = null_check(value); 1602 1603 RegionNode* bailout = new RegionNode(1); 1604 record_for_igvn(bailout); 1605 1606 // Range checks 1607 generate_negative_guard(offset, bailout); 1608 generate_negative_guard(length, bailout); 1609 generate_limit_guard(offset, length, load_array_length(value), bailout); 1610 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1611 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1612 1613 if (bailout->req() > 1) { 1614 PreserveJVMState pjvms(this); 1615 set_control(_gvn.transform(bailout)); 1616 uncommon_trap(Deoptimization::Reason_intrinsic, 1617 Deoptimization::Action_maybe_recompile); 1618 } 1619 if (stopped()) { 1620 return true; 1621 } 1622 1623 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1624 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1625 newcopy = new_array(klass_node, size, 0); // no arguments to push 1626 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1627 1628 // Calculate starting addresses. 1629 value = access_resolve(value, ACCESS_READ); 1630 Node* src_start = array_element_address(value, offset, T_CHAR); 1631 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1632 1633 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1634 const TypeInt* toffset = gvn().type(offset)->is_int(); 1635 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1636 1637 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1638 const char* copyfunc_name = "arraycopy"; 1639 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1640 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1641 OptoRuntime::fast_arraycopy_Type(), 1642 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1643 src_start, dst_start, ConvI2X(length) XTOP); 1644 // Do not let reads from the cloned object float above the arraycopy. 1645 if (alloc != NULL) { 1646 if (alloc->maybe_set_complete(&_gvn)) { 1647 // "You break it, you buy it." 1648 InitializeNode* init = alloc->initialization(); 1649 assert(init->is_complete(), "we just did this"); 1650 init->set_complete_with_arraycopy(); 1651 assert(newcopy->is_CheckCastPP(), "sanity"); 1652 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1653 } 1654 // Do not let stores that initialize this object be reordered with 1655 // a subsequent store that would make this object accessible by 1656 // other threads. 1657 // Record what AllocateNode this StoreStore protects so that 1658 // escape analysis can go from the MemBarStoreStoreNode to the 1659 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1660 // based on the escape status of the AllocateNode. 1661 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1662 } else { 1663 insert_mem_bar(Op_MemBarCPUOrder); 1664 } 1665 } // original reexecute is set back here 1666 1667 C->set_has_split_ifs(true); // Has chance for split-if optimization 1668 if (!stopped()) { 1669 set_result(newcopy); 1670 } 1671 clear_upper_avx(); 1672 1673 return true; 1674 } 1675 1676 //------------------------inline_string_getCharsU-------------------------- 1677 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1678 bool LibraryCallKit::inline_string_getCharsU() { 1679 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1680 return false; 1681 } 1682 1683 // Get the arguments. 1684 Node* src = argument(0); 1685 Node* src_begin = argument(1); 1686 Node* src_end = argument(2); // exclusive offset (i < src_end) 1687 Node* dst = argument(3); 1688 Node* dst_begin = argument(4); 1689 1690 // Check for allocation before we add nodes that would confuse 1691 // tightly_coupled_allocation() 1692 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1693 1694 // Check if a null path was taken unconditionally. 1695 src = null_check(src); 1696 dst = null_check(dst); 1697 if (stopped()) { 1698 return true; 1699 } 1700 1701 // Get length and convert char[] offset to byte[] offset 1702 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1703 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1704 1705 // Range checks 1706 generate_string_range_check(src, src_begin, length, true); 1707 generate_string_range_check(dst, dst_begin, length, false); 1708 if (stopped()) { 1709 return true; 1710 } 1711 1712 if (!stopped()) { 1713 src = access_resolve(src, ACCESS_READ); 1714 dst = access_resolve(dst, ACCESS_WRITE); 1715 1716 // Calculate starting addresses. 1717 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1718 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1719 1720 // Check if array addresses are aligned to HeapWordSize 1721 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1722 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1723 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1724 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1725 1726 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1727 const char* copyfunc_name = "arraycopy"; 1728 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1729 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1730 OptoRuntime::fast_arraycopy_Type(), 1731 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1732 src_start, dst_start, ConvI2X(length) XTOP); 1733 // Do not let reads from the cloned object float above the arraycopy. 1734 if (alloc != NULL) { 1735 if (alloc->maybe_set_complete(&_gvn)) { 1736 // "You break it, you buy it." 1737 InitializeNode* init = alloc->initialization(); 1738 assert(init->is_complete(), "we just did this"); 1739 init->set_complete_with_arraycopy(); 1740 assert(dst->is_CheckCastPP(), "sanity"); 1741 assert(dst->in(0)->in(0) == init, "dest pinned"); 1742 } 1743 // Do not let stores that initialize this object be reordered with 1744 // a subsequent store that would make this object accessible by 1745 // other threads. 1746 // Record what AllocateNode this StoreStore protects so that 1747 // escape analysis can go from the MemBarStoreStoreNode to the 1748 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1749 // based on the escape status of the AllocateNode. 1750 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1751 } else { 1752 insert_mem_bar(Op_MemBarCPUOrder); 1753 } 1754 } 1755 1756 C->set_has_split_ifs(true); // Has chance for split-if optimization 1757 return true; 1758 } 1759 1760 //----------------------inline_string_char_access---------------------------- 1761 // Store/Load char to/from byte[] array. 1762 // static void StringUTF16.putChar(byte[] val, int index, int c) 1763 // static char StringUTF16.getChar(byte[] val, int index) 1764 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1765 Node* value = argument(0); 1766 Node* index = argument(1); 1767 Node* ch = is_store ? argument(2) : NULL; 1768 1769 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1770 // correctly requires matched array shapes. 1771 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1772 "sanity: byte[] and char[] bases agree"); 1773 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1774 "sanity: byte[] and char[] scales agree"); 1775 1776 // Bail when getChar over constants is requested: constant folding would 1777 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1778 // Java method would constant fold nicely instead. 1779 if (!is_store && value->is_Con() && index->is_Con()) { 1780 return false; 1781 } 1782 1783 value = must_be_not_null(value, true); 1784 value = access_resolve(value, is_store ? ACCESS_WRITE : ACCESS_READ); 1785 1786 Node* adr = array_element_address(value, index, T_CHAR); 1787 if (adr->is_top()) { 1788 return false; 1789 } 1790 if (is_store) { 1791 access_store_at(value, adr, TypeAryPtr::BYTES, ch, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED); 1792 } else { 1793 ch = access_load_at(value, adr, TypeAryPtr::BYTES, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED | C2_CONTROL_DEPENDENT_LOAD); 1794 set_result(ch); 1795 } 1796 return true; 1797 } 1798 1799 //--------------------------round_double_node-------------------------------- 1800 // Round a double node if necessary. 1801 Node* LibraryCallKit::round_double_node(Node* n) { 1802 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1803 n = _gvn.transform(new RoundDoubleNode(0, n)); 1804 return n; 1805 } 1806 1807 //------------------------------inline_math----------------------------------- 1808 // public static double Math.abs(double) 1809 // public static double Math.sqrt(double) 1810 // public static double Math.log(double) 1811 // public static double Math.log10(double) 1812 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1813 Node* arg = round_double_node(argument(0)); 1814 Node* n = NULL; 1815 switch (id) { 1816 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1817 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1818 default: fatal_unexpected_iid(id); break; 1819 } 1820 set_result(_gvn.transform(n)); 1821 return true; 1822 } 1823 1824 //------------------------------runtime_math----------------------------- 1825 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1826 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1827 "must be (DD)D or (D)D type"); 1828 1829 // Inputs 1830 Node* a = round_double_node(argument(0)); 1831 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1832 1833 const TypePtr* no_memory_effects = NULL; 1834 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1835 no_memory_effects, 1836 a, top(), b, b ? top() : NULL); 1837 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1838 #ifdef ASSERT 1839 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1840 assert(value_top == top(), "second value must be top"); 1841 #endif 1842 1843 set_result(value); 1844 return true; 1845 } 1846 1847 //------------------------------inline_math_native----------------------------- 1848 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1849 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1850 switch (id) { 1851 // These intrinsics are not properly supported on all hardware 1852 case vmIntrinsics::_dsin: 1853 return StubRoutines::dsin() != NULL ? 1854 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1855 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1856 case vmIntrinsics::_dcos: 1857 return StubRoutines::dcos() != NULL ? 1858 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1859 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1860 case vmIntrinsics::_dtan: 1861 return StubRoutines::dtan() != NULL ? 1862 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1863 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1864 case vmIntrinsics::_dlog: 1865 return StubRoutines::dlog() != NULL ? 1866 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1867 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1868 case vmIntrinsics::_dlog10: 1869 return StubRoutines::dlog10() != NULL ? 1870 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1871 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1872 1873 // These intrinsics are supported on all hardware 1874 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1875 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1876 1877 case vmIntrinsics::_dexp: 1878 return StubRoutines::dexp() != NULL ? 1879 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1880 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1881 case vmIntrinsics::_dpow: { 1882 Node* exp = round_double_node(argument(2)); 1883 const TypeD* d = _gvn.type(exp)->isa_double_constant(); 1884 if (d != NULL && d->getd() == 2.0) { 1885 // Special case: pow(x, 2.0) => x * x 1886 Node* base = round_double_node(argument(0)); 1887 set_result(_gvn.transform(new MulDNode(base, base))); 1888 return true; 1889 } 1890 return StubRoutines::dpow() != NULL ? 1891 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1892 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1893 } 1894 #undef FN_PTR 1895 1896 // These intrinsics are not yet correctly implemented 1897 case vmIntrinsics::_datan2: 1898 return false; 1899 1900 default: 1901 fatal_unexpected_iid(id); 1902 return false; 1903 } 1904 } 1905 1906 static bool is_simple_name(Node* n) { 1907 return (n->req() == 1 // constant 1908 || (n->is_Type() && n->as_Type()->type()->singleton()) 1909 || n->is_Proj() // parameter or return value 1910 || n->is_Phi() // local of some sort 1911 ); 1912 } 1913 1914 //----------------------------inline_notify-----------------------------------* 1915 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1916 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1917 address func; 1918 if (id == vmIntrinsics::_notify) { 1919 func = OptoRuntime::monitor_notify_Java(); 1920 } else { 1921 func = OptoRuntime::monitor_notifyAll_Java(); 1922 } 1923 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1924 make_slow_call_ex(call, env()->Throwable_klass(), false); 1925 return true; 1926 } 1927 1928 1929 //----------------------------inline_min_max----------------------------------- 1930 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1931 set_result(generate_min_max(id, argument(0), argument(1))); 1932 return true; 1933 } 1934 1935 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1936 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1937 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1938 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1939 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1940 1941 { 1942 PreserveJVMState pjvms(this); 1943 PreserveReexecuteState preexecs(this); 1944 jvms()->set_should_reexecute(true); 1945 1946 set_control(slow_path); 1947 set_i_o(i_o()); 1948 1949 uncommon_trap(Deoptimization::Reason_intrinsic, 1950 Deoptimization::Action_none); 1951 } 1952 1953 set_control(fast_path); 1954 set_result(math); 1955 } 1956 1957 template <typename OverflowOp> 1958 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1959 typedef typename OverflowOp::MathOp MathOp; 1960 1961 MathOp* mathOp = new MathOp(arg1, arg2); 1962 Node* operation = _gvn.transform( mathOp ); 1963 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1964 inline_math_mathExact(operation, ofcheck); 1965 return true; 1966 } 1967 1968 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1969 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1970 } 1971 1972 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1973 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1974 } 1975 1976 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1977 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1978 } 1979 1980 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1981 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1982 } 1983 1984 bool LibraryCallKit::inline_math_negateExactI() { 1985 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1986 } 1987 1988 bool LibraryCallKit::inline_math_negateExactL() { 1989 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 1990 } 1991 1992 bool LibraryCallKit::inline_math_multiplyExactI() { 1993 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 1994 } 1995 1996 bool LibraryCallKit::inline_math_multiplyExactL() { 1997 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 1998 } 1999 2000 bool LibraryCallKit::inline_math_multiplyHigh() { 2001 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2)))); 2002 return true; 2003 } 2004 2005 Node* 2006 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2007 // These are the candidate return value: 2008 Node* xvalue = x0; 2009 Node* yvalue = y0; 2010 2011 if (xvalue == yvalue) { 2012 return xvalue; 2013 } 2014 2015 bool want_max = (id == vmIntrinsics::_max); 2016 2017 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2018 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2019 if (txvalue == NULL || tyvalue == NULL) return top(); 2020 // This is not really necessary, but it is consistent with a 2021 // hypothetical MaxINode::Value method: 2022 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2023 2024 // %%% This folding logic should (ideally) be in a different place. 2025 // Some should be inside IfNode, and there to be a more reliable 2026 // transformation of ?: style patterns into cmoves. We also want 2027 // more powerful optimizations around cmove and min/max. 2028 2029 // Try to find a dominating comparison of these guys. 2030 // It can simplify the index computation for Arrays.copyOf 2031 // and similar uses of System.arraycopy. 2032 // First, compute the normalized version of CmpI(x, y). 2033 int cmp_op = Op_CmpI; 2034 Node* xkey = xvalue; 2035 Node* ykey = yvalue; 2036 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2037 if (ideal_cmpxy->is_Cmp()) { 2038 // E.g., if we have CmpI(length - offset, count), 2039 // it might idealize to CmpI(length, count + offset) 2040 cmp_op = ideal_cmpxy->Opcode(); 2041 xkey = ideal_cmpxy->in(1); 2042 ykey = ideal_cmpxy->in(2); 2043 } 2044 2045 // Start by locating any relevant comparisons. 2046 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2047 Node* cmpxy = NULL; 2048 Node* cmpyx = NULL; 2049 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2050 Node* cmp = start_from->fast_out(k); 2051 if (cmp->outcnt() > 0 && // must have prior uses 2052 cmp->in(0) == NULL && // must be context-independent 2053 cmp->Opcode() == cmp_op) { // right kind of compare 2054 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2055 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2056 } 2057 } 2058 2059 const int NCMPS = 2; 2060 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2061 int cmpn; 2062 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2063 if (cmps[cmpn] != NULL) break; // find a result 2064 } 2065 if (cmpn < NCMPS) { 2066 // Look for a dominating test that tells us the min and max. 2067 int depth = 0; // Limit search depth for speed 2068 Node* dom = control(); 2069 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2070 if (++depth >= 100) break; 2071 Node* ifproj = dom; 2072 if (!ifproj->is_Proj()) continue; 2073 Node* iff = ifproj->in(0); 2074 if (!iff->is_If()) continue; 2075 Node* bol = iff->in(1); 2076 if (!bol->is_Bool()) continue; 2077 Node* cmp = bol->in(1); 2078 if (cmp == NULL) continue; 2079 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2080 if (cmps[cmpn] == cmp) break; 2081 if (cmpn == NCMPS) continue; 2082 BoolTest::mask btest = bol->as_Bool()->_test._test; 2083 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2084 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2085 // At this point, we know that 'x btest y' is true. 2086 switch (btest) { 2087 case BoolTest::eq: 2088 // They are proven equal, so we can collapse the min/max. 2089 // Either value is the answer. Choose the simpler. 2090 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2091 return yvalue; 2092 return xvalue; 2093 case BoolTest::lt: // x < y 2094 case BoolTest::le: // x <= y 2095 return (want_max ? yvalue : xvalue); 2096 case BoolTest::gt: // x > y 2097 case BoolTest::ge: // x >= y 2098 return (want_max ? xvalue : yvalue); 2099 default: 2100 break; 2101 } 2102 } 2103 } 2104 2105 // We failed to find a dominating test. 2106 // Let's pick a test that might GVN with prior tests. 2107 Node* best_bol = NULL; 2108 BoolTest::mask best_btest = BoolTest::illegal; 2109 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2110 Node* cmp = cmps[cmpn]; 2111 if (cmp == NULL) continue; 2112 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2113 Node* bol = cmp->fast_out(j); 2114 if (!bol->is_Bool()) continue; 2115 BoolTest::mask btest = bol->as_Bool()->_test._test; 2116 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2117 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2118 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2119 best_bol = bol->as_Bool(); 2120 best_btest = btest; 2121 } 2122 } 2123 } 2124 2125 Node* answer_if_true = NULL; 2126 Node* answer_if_false = NULL; 2127 switch (best_btest) { 2128 default: 2129 if (cmpxy == NULL) 2130 cmpxy = ideal_cmpxy; 2131 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2132 // and fall through: 2133 case BoolTest::lt: // x < y 2134 case BoolTest::le: // x <= y 2135 answer_if_true = (want_max ? yvalue : xvalue); 2136 answer_if_false = (want_max ? xvalue : yvalue); 2137 break; 2138 case BoolTest::gt: // x > y 2139 case BoolTest::ge: // x >= y 2140 answer_if_true = (want_max ? xvalue : yvalue); 2141 answer_if_false = (want_max ? yvalue : xvalue); 2142 break; 2143 } 2144 2145 jint hi, lo; 2146 if (want_max) { 2147 // We can sharpen the minimum. 2148 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2149 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2150 } else { 2151 // We can sharpen the maximum. 2152 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2153 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2154 } 2155 2156 // Use a flow-free graph structure, to avoid creating excess control edges 2157 // which could hinder other optimizations. 2158 // Since Math.min/max is often used with arraycopy, we want 2159 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2160 Node* cmov = CMoveNode::make(NULL, best_bol, 2161 answer_if_false, answer_if_true, 2162 TypeInt::make(lo, hi, widen)); 2163 2164 return _gvn.transform(cmov); 2165 2166 /* 2167 // This is not as desirable as it may seem, since Min and Max 2168 // nodes do not have a full set of optimizations. 2169 // And they would interfere, anyway, with 'if' optimizations 2170 // and with CMoveI canonical forms. 2171 switch (id) { 2172 case vmIntrinsics::_min: 2173 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2174 case vmIntrinsics::_max: 2175 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2176 default: 2177 ShouldNotReachHere(); 2178 } 2179 */ 2180 } 2181 2182 inline int 2183 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) { 2184 const TypePtr* base_type = TypePtr::NULL_PTR; 2185 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2186 if (base_type == NULL) { 2187 // Unknown type. 2188 return Type::AnyPtr; 2189 } else if (base_type == TypePtr::NULL_PTR) { 2190 // Since this is a NULL+long form, we have to switch to a rawptr. 2191 base = _gvn.transform(new CastX2PNode(offset)); 2192 offset = MakeConX(0); 2193 return Type::RawPtr; 2194 } else if (base_type->base() == Type::RawPtr) { 2195 return Type::RawPtr; 2196 } else if (base_type->isa_oopptr()) { 2197 // Base is never null => always a heap address. 2198 if (!TypePtr::NULL_PTR->higher_equal(base_type)) { 2199 return Type::OopPtr; 2200 } 2201 // Offset is small => always a heap address. 2202 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2203 if (offset_type != NULL && 2204 base_type->offset() == 0 && // (should always be?) 2205 offset_type->_lo >= 0 && 2206 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2207 return Type::OopPtr; 2208 } else if (type == T_OBJECT) { 2209 // off heap access to an oop doesn't make any sense. Has to be on 2210 // heap. 2211 return Type::OopPtr; 2212 } 2213 // Otherwise, it might either be oop+off or NULL+addr. 2214 return Type::AnyPtr; 2215 } else { 2216 // No information: 2217 return Type::AnyPtr; 2218 } 2219 } 2220 2221 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type, bool can_cast) { 2222 Node* uncasted_base = base; 2223 int kind = classify_unsafe_addr(uncasted_base, offset, type); 2224 if (kind == Type::RawPtr) { 2225 return basic_plus_adr(top(), uncasted_base, offset); 2226 } else if (kind == Type::AnyPtr) { 2227 assert(base == uncasted_base, "unexpected base change"); 2228 if (can_cast) { 2229 if (!_gvn.type(base)->speculative_maybe_null() && 2230 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2231 // According to profiling, this access is always on 2232 // heap. Casting the base to not null and thus avoiding membars 2233 // around the access should allow better optimizations 2234 Node* null_ctl = top(); 2235 base = null_check_oop(base, &null_ctl, true, true, true); 2236 assert(null_ctl->is_top(), "no null control here"); 2237 return basic_plus_adr(base, offset); 2238 } else if (_gvn.type(base)->speculative_always_null() && 2239 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2240 // According to profiling, this access is always off 2241 // heap. 2242 base = null_assert(base); 2243 Node* raw_base = _gvn.transform(new CastX2PNode(offset)); 2244 offset = MakeConX(0); 2245 return basic_plus_adr(top(), raw_base, offset); 2246 } 2247 } 2248 // We don't know if it's an on heap or off heap access. Fall back 2249 // to raw memory access. 2250 base = access_resolve(base, decorators); 2251 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM)); 2252 return basic_plus_adr(top(), raw, offset); 2253 } else { 2254 assert(base == uncasted_base, "unexpected base change"); 2255 // We know it's an on heap access so base can't be null 2256 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) { 2257 base = must_be_not_null(base, true); 2258 } 2259 return basic_plus_adr(base, offset); 2260 } 2261 } 2262 2263 //--------------------------inline_number_methods----------------------------- 2264 // inline int Integer.numberOfLeadingZeros(int) 2265 // inline int Long.numberOfLeadingZeros(long) 2266 // 2267 // inline int Integer.numberOfTrailingZeros(int) 2268 // inline int Long.numberOfTrailingZeros(long) 2269 // 2270 // inline int Integer.bitCount(int) 2271 // inline int Long.bitCount(long) 2272 // 2273 // inline char Character.reverseBytes(char) 2274 // inline short Short.reverseBytes(short) 2275 // inline int Integer.reverseBytes(int) 2276 // inline long Long.reverseBytes(long) 2277 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2278 Node* arg = argument(0); 2279 Node* n = NULL; 2280 switch (id) { 2281 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2282 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2283 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2284 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2285 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2286 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2287 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2288 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2289 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2290 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2291 default: fatal_unexpected_iid(id); break; 2292 } 2293 set_result(_gvn.transform(n)); 2294 return true; 2295 } 2296 2297 //----------------------------inline_unsafe_access---------------------------- 2298 2299 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2300 // Attempt to infer a sharper value type from the offset and base type. 2301 ciKlass* sharpened_klass = NULL; 2302 2303 // See if it is an instance field, with an object type. 2304 if (alias_type->field() != NULL) { 2305 if (alias_type->field()->type()->is_klass()) { 2306 sharpened_klass = alias_type->field()->type()->as_klass(); 2307 } 2308 } 2309 2310 // See if it is a narrow oop array. 2311 if (adr_type->isa_aryptr()) { 2312 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2313 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2314 if (elem_type != NULL) { 2315 sharpened_klass = elem_type->klass(); 2316 } 2317 } 2318 } 2319 2320 // The sharpened class might be unloaded if there is no class loader 2321 // contraint in place. 2322 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2323 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2324 2325 #ifndef PRODUCT 2326 if (C->print_intrinsics() || C->print_inlining()) { 2327 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2328 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2329 } 2330 #endif 2331 // Sharpen the value type. 2332 return tjp; 2333 } 2334 return NULL; 2335 } 2336 2337 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) { 2338 switch (kind) { 2339 case Relaxed: 2340 return MO_UNORDERED; 2341 case Opaque: 2342 return MO_RELAXED; 2343 case Acquire: 2344 return MO_ACQUIRE; 2345 case Release: 2346 return MO_RELEASE; 2347 case Volatile: 2348 return MO_SEQ_CST; 2349 default: 2350 ShouldNotReachHere(); 2351 return 0; 2352 } 2353 } 2354 2355 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2356 if (callee()->is_static()) return false; // caller must have the capability! 2357 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2358 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2359 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2360 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2361 2362 if (type == T_OBJECT || type == T_ARRAY) { 2363 decorators |= ON_UNKNOWN_OOP_REF; 2364 } 2365 2366 if (unaligned) { 2367 decorators |= C2_UNALIGNED; 2368 } 2369 2370 #ifndef PRODUCT 2371 { 2372 ResourceMark rm; 2373 // Check the signatures. 2374 ciSignature* sig = callee()->signature(); 2375 #ifdef ASSERT 2376 if (!is_store) { 2377 // Object getReference(Object base, int/long offset), etc. 2378 BasicType rtype = sig->return_type()->basic_type(); 2379 assert(rtype == type, "getter must return the expected value"); 2380 assert(sig->count() == 2, "oop getter has 2 arguments"); 2381 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2382 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2383 } else { 2384 // void putReference(Object base, int/long offset, Object x), etc. 2385 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2386 assert(sig->count() == 3, "oop putter has 3 arguments"); 2387 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2388 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2389 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2390 assert(vtype == type, "putter must accept the expected value"); 2391 } 2392 #endif // ASSERT 2393 } 2394 #endif //PRODUCT 2395 2396 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2397 2398 Node* receiver = argument(0); // type: oop 2399 2400 // Build address expression. 2401 Node* adr; 2402 Node* heap_base_oop = top(); 2403 Node* offset = top(); 2404 Node* val; 2405 2406 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2407 Node* base = argument(1); // type: oop 2408 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2409 offset = argument(2); // type: long 2410 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2411 // to be plain byte offsets, which are also the same as those accepted 2412 // by oopDesc::field_addr. 2413 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2414 "fieldOffset must be byte-scaled"); 2415 2416 if (base->is_ValueType()) { 2417 if (is_store) { 2418 return false; 2419 } 2420 2421 ValueTypeNode* vt = base->as_ValueType(); 2422 if (offset->is_Con()) { 2423 long off = find_long_con(offset, 0); 2424 ciValueKlass* vk = _gvn.type(vt)->is_valuetype()->value_klass(); 2425 if ((long)(int)off != off || !vk->contains_field_offset(off)) { 2426 return false; 2427 } 2428 2429 receiver = null_check(receiver); 2430 if (stopped()) { 2431 return true; 2432 } 2433 2434 set_result(vt->field_value_by_offset((int)off, true)); 2435 return true; 2436 } else { 2437 receiver = null_check(receiver); 2438 if (stopped()) { 2439 return true; 2440 } 2441 vt = vt->allocate(this)->as_ValueType(); 2442 base = vt->get_oop(); 2443 } 2444 } 2445 2446 // 32-bit machines ignore the high half! 2447 offset = ConvL2X(offset); 2448 adr = make_unsafe_address(base, offset, is_store ? ACCESS_WRITE : ACCESS_READ, type, kind == Relaxed); 2449 2450 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) { 2451 heap_base_oop = base; 2452 } else if (type == T_OBJECT) { 2453 return false; // off-heap oop accesses are not supported 2454 } 2455 2456 // Can base be NULL? Otherwise, always on-heap access. 2457 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop)); 2458 2459 if (!can_access_non_heap) { 2460 decorators |= IN_HEAP; 2461 } 2462 2463 val = is_store ? argument(4) : NULL; 2464 2465 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2466 2467 // Try to categorize the address. 2468 Compile::AliasType* alias_type = C->alias_type(adr_type); 2469 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2470 2471 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2472 alias_type->adr_type() == TypeAryPtr::RANGE) { 2473 return false; // not supported 2474 } 2475 2476 bool mismatched = false; 2477 BasicType bt = alias_type->basic_type(); 2478 if (bt != T_ILLEGAL) { 2479 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2480 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2481 // Alias type doesn't differentiate between byte[] and boolean[]). 2482 // Use address type to get the element type. 2483 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2484 } 2485 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2486 // accessing an array field with getReference is not a mismatch 2487 bt = T_OBJECT; 2488 } 2489 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2490 // Don't intrinsify mismatched object accesses 2491 return false; 2492 } 2493 mismatched = (bt != type); 2494 } else if (alias_type->adr_type()->isa_oopptr()) { 2495 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched 2496 } 2497 2498 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2499 2500 if (mismatched) { 2501 decorators |= C2_MISMATCHED; 2502 } 2503 2504 // First guess at the value type. 2505 const Type *value_type = Type::get_const_basic_type(type); 2506 2507 // Figure out the memory ordering. 2508 decorators |= mo_decorator_for_access_kind(kind); 2509 2510 if (!is_store && type == T_OBJECT) { 2511 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2512 if (tjp != NULL) { 2513 value_type = tjp; 2514 } 2515 } 2516 2517 receiver = null_check(receiver); 2518 if (stopped()) { 2519 return true; 2520 } 2521 // Heap pointers get a null-check from the interpreter, 2522 // as a courtesy. However, this is not guaranteed by Unsafe, 2523 // and it is not possible to fully distinguish unintended nulls 2524 // from intended ones in this API. 2525 2526 if (!is_store) { 2527 Node* p = NULL; 2528 // Try to constant fold a load from a constant field 2529 ciField* field = alias_type->field(); 2530 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { 2531 // final or stable field 2532 p = make_constant_from_field(field, heap_base_oop); 2533 } 2534 2535 if (p == NULL) { // Could not constant fold the load 2536 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators); 2537 // Normalize the value returned by getBoolean in the following cases 2538 if (type == T_BOOLEAN && 2539 (mismatched || 2540 heap_base_oop == top() || // - heap_base_oop is NULL or 2541 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL 2542 // and the unsafe access is made to large offset 2543 // (i.e., larger than the maximum offset necessary for any 2544 // field access) 2545 ) { 2546 IdealKit ideal = IdealKit(this); 2547 #define __ ideal. 2548 IdealVariable normalized_result(ideal); 2549 __ declarations_done(); 2550 __ set(normalized_result, p); 2551 __ if_then(p, BoolTest::ne, ideal.ConI(0)); 2552 __ set(normalized_result, ideal.ConI(1)); 2553 ideal.end_if(); 2554 final_sync(ideal); 2555 p = __ value(normalized_result); 2556 #undef __ 2557 } 2558 } 2559 if (type == T_ADDRESS) { 2560 p = gvn().transform(new CastP2XNode(NULL, p)); 2561 p = ConvX2UL(p); 2562 } 2563 if (field != NULL && field->is_flattenable()) { 2564 // Load a non-flattened but flattenable value type from memory 2565 assert(!field->is_flattened(), "unsafe value type load from flattened field"); 2566 if (value_type->value_klass()->is_scalarizable()) { 2567 p = ValueTypeNode::make_from_oop(this, p, value_type->value_klass()); 2568 } else { 2569 p = null2default(p, value_type->value_klass()); 2570 } 2571 } 2572 // The load node has the control of the preceding MemBarCPUOrder. All 2573 // following nodes will have the control of the MemBarCPUOrder inserted at 2574 // the end of this method. So, pushing the load onto the stack at a later 2575 // point is fine. 2576 set_result(p); 2577 } else { 2578 if (bt == T_ADDRESS) { 2579 // Repackage the long as a pointer. 2580 val = ConvL2X(val); 2581 val = gvn().transform(new CastX2PNode(val)); 2582 } 2583 access_store_at(heap_base_oop, adr, adr_type, val, value_type, type, decorators); 2584 } 2585 2586 return true; 2587 } 2588 2589 //----------------------------inline_unsafe_load_store---------------------------- 2590 // This method serves a couple of different customers (depending on LoadStoreKind): 2591 // 2592 // LS_cmp_swap: 2593 // 2594 // boolean compareAndSetReference(Object o, long offset, Object expected, Object x); 2595 // boolean compareAndSetInt( Object o, long offset, int expected, int x); 2596 // boolean compareAndSetLong( Object o, long offset, long expected, long x); 2597 // 2598 // LS_cmp_swap_weak: 2599 // 2600 // boolean weakCompareAndSetReference( Object o, long offset, Object expected, Object x); 2601 // boolean weakCompareAndSetReferencePlain( Object o, long offset, Object expected, Object x); 2602 // boolean weakCompareAndSetReferenceAcquire(Object o, long offset, Object expected, Object x); 2603 // boolean weakCompareAndSetReferenceRelease(Object o, long offset, Object expected, Object x); 2604 // 2605 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x); 2606 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x); 2607 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x); 2608 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x); 2609 // 2610 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x); 2611 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x); 2612 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x); 2613 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x); 2614 // 2615 // LS_cmp_exchange: 2616 // 2617 // Object compareAndExchangeReferenceVolatile(Object o, long offset, Object expected, Object x); 2618 // Object compareAndExchangeReferenceAcquire( Object o, long offset, Object expected, Object x); 2619 // Object compareAndExchangeReferenceRelease( Object o, long offset, Object expected, Object x); 2620 // 2621 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2622 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2623 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2624 // 2625 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2626 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2627 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2628 // 2629 // LS_get_add: 2630 // 2631 // int getAndAddInt( Object o, long offset, int delta) 2632 // long getAndAddLong(Object o, long offset, long delta) 2633 // 2634 // LS_get_set: 2635 // 2636 // int getAndSet(Object o, long offset, int newValue) 2637 // long getAndSet(Object o, long offset, long newValue) 2638 // Object getAndSet(Object o, long offset, Object newValue) 2639 // 2640 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2641 // This basic scheme here is the same as inline_unsafe_access, but 2642 // differs in enough details that combining them would make the code 2643 // overly confusing. (This is a true fact! I originally combined 2644 // them, but even I was confused by it!) As much code/comments as 2645 // possible are retained from inline_unsafe_access though to make 2646 // the correspondences clearer. - dl 2647 2648 if (callee()->is_static()) return false; // caller must have the capability! 2649 2650 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2651 decorators |= mo_decorator_for_access_kind(access_kind); 2652 2653 #ifndef PRODUCT 2654 BasicType rtype; 2655 { 2656 ResourceMark rm; 2657 // Check the signatures. 2658 ciSignature* sig = callee()->signature(); 2659 rtype = sig->return_type()->basic_type(); 2660 switch(kind) { 2661 case LS_get_add: 2662 case LS_get_set: { 2663 // Check the signatures. 2664 #ifdef ASSERT 2665 assert(rtype == type, "get and set must return the expected type"); 2666 assert(sig->count() == 3, "get and set has 3 arguments"); 2667 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2668 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2669 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2670 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2671 #endif // ASSERT 2672 break; 2673 } 2674 case LS_cmp_swap: 2675 case LS_cmp_swap_weak: { 2676 // Check the signatures. 2677 #ifdef ASSERT 2678 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2679 assert(sig->count() == 4, "CAS has 4 arguments"); 2680 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2681 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2682 #endif // ASSERT 2683 break; 2684 } 2685 case LS_cmp_exchange: { 2686 // Check the signatures. 2687 #ifdef ASSERT 2688 assert(rtype == type, "CAS must return the expected type"); 2689 assert(sig->count() == 4, "CAS has 4 arguments"); 2690 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2691 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2692 #endif // ASSERT 2693 break; 2694 } 2695 default: 2696 ShouldNotReachHere(); 2697 } 2698 } 2699 #endif //PRODUCT 2700 2701 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2702 2703 // Get arguments: 2704 Node* receiver = NULL; 2705 Node* base = NULL; 2706 Node* offset = NULL; 2707 Node* oldval = NULL; 2708 Node* newval = NULL; 2709 switch(kind) { 2710 case LS_cmp_swap: 2711 case LS_cmp_swap_weak: 2712 case LS_cmp_exchange: { 2713 const bool two_slot_type = type2size[type] == 2; 2714 receiver = argument(0); // type: oop 2715 base = argument(1); // type: oop 2716 offset = argument(2); // type: long 2717 oldval = argument(4); // type: oop, int, or long 2718 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2719 break; 2720 } 2721 case LS_get_add: 2722 case LS_get_set: { 2723 receiver = argument(0); // type: oop 2724 base = argument(1); // type: oop 2725 offset = argument(2); // type: long 2726 oldval = NULL; 2727 newval = argument(4); // type: oop, int, or long 2728 break; 2729 } 2730 default: 2731 ShouldNotReachHere(); 2732 } 2733 2734 // Build field offset expression. 2735 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2736 // to be plain byte offsets, which are also the same as those accepted 2737 // by oopDesc::field_addr. 2738 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2739 // 32-bit machines ignore the high half of long offsets 2740 offset = ConvL2X(offset); 2741 Node* adr = make_unsafe_address(base, offset, ACCESS_WRITE | ACCESS_READ, type, false); 2742 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2743 2744 Compile::AliasType* alias_type = C->alias_type(adr_type); 2745 BasicType bt = alias_type->basic_type(); 2746 if (bt != T_ILLEGAL && 2747 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2748 // Don't intrinsify mismatched object accesses. 2749 return false; 2750 } 2751 2752 // For CAS, unlike inline_unsafe_access, there seems no point in 2753 // trying to refine types. Just use the coarse types here. 2754 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2755 const Type *value_type = Type::get_const_basic_type(type); 2756 2757 switch (kind) { 2758 case LS_get_set: 2759 case LS_cmp_exchange: { 2760 if (type == T_OBJECT) { 2761 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2762 if (tjp != NULL) { 2763 value_type = tjp; 2764 } 2765 } 2766 break; 2767 } 2768 case LS_cmp_swap: 2769 case LS_cmp_swap_weak: 2770 case LS_get_add: 2771 break; 2772 default: 2773 ShouldNotReachHere(); 2774 } 2775 2776 // Null check receiver. 2777 receiver = null_check(receiver); 2778 if (stopped()) { 2779 return true; 2780 } 2781 2782 int alias_idx = C->get_alias_index(adr_type); 2783 2784 if (type == T_OBJECT || type == T_ARRAY) { 2785 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF; 2786 2787 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2788 // could be delayed during Parse (for example, in adjust_map_after_if()). 2789 // Execute transformation here to avoid barrier generation in such case. 2790 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2791 newval = _gvn.makecon(TypePtr::NULL_PTR); 2792 2793 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) { 2794 // Refine the value to a null constant, when it is known to be null 2795 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2796 } 2797 } 2798 2799 Node* result = NULL; 2800 switch (kind) { 2801 case LS_cmp_exchange: { 2802 result = access_atomic_cmpxchg_val_at(base, adr, adr_type, alias_idx, 2803 oldval, newval, value_type, type, decorators); 2804 break; 2805 } 2806 case LS_cmp_swap_weak: 2807 decorators |= C2_WEAK_CMPXCHG; 2808 case LS_cmp_swap: { 2809 result = access_atomic_cmpxchg_bool_at(base, adr, adr_type, alias_idx, 2810 oldval, newval, value_type, type, decorators); 2811 break; 2812 } 2813 case LS_get_set: { 2814 result = access_atomic_xchg_at(base, adr, adr_type, alias_idx, 2815 newval, value_type, type, decorators); 2816 break; 2817 } 2818 case LS_get_add: { 2819 result = access_atomic_add_at(base, adr, adr_type, alias_idx, 2820 newval, value_type, type, decorators); 2821 break; 2822 } 2823 default: 2824 ShouldNotReachHere(); 2825 } 2826 2827 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2828 set_result(result); 2829 return true; 2830 } 2831 2832 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2833 // Regardless of form, don't allow previous ld/st to move down, 2834 // then issue acquire, release, or volatile mem_bar. 2835 insert_mem_bar(Op_MemBarCPUOrder); 2836 switch(id) { 2837 case vmIntrinsics::_loadFence: 2838 insert_mem_bar(Op_LoadFence); 2839 return true; 2840 case vmIntrinsics::_storeFence: 2841 insert_mem_bar(Op_StoreFence); 2842 return true; 2843 case vmIntrinsics::_fullFence: 2844 insert_mem_bar(Op_MemBarVolatile); 2845 return true; 2846 default: 2847 fatal_unexpected_iid(id); 2848 return false; 2849 } 2850 } 2851 2852 bool LibraryCallKit::inline_onspinwait() { 2853 insert_mem_bar(Op_OnSpinWait); 2854 return true; 2855 } 2856 2857 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 2858 if (!kls->is_Con()) { 2859 return true; 2860 } 2861 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 2862 if (klsptr == NULL) { 2863 return true; 2864 } 2865 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 2866 // don't need a guard for a klass that is already initialized 2867 return !ik->is_initialized(); 2868 } 2869 2870 //----------------------------inline_unsafe_allocate--------------------------- 2871 // public native Object Unsafe.allocateInstance(Class<?> cls); 2872 bool LibraryCallKit::inline_unsafe_allocate() { 2873 if (callee()->is_static()) return false; // caller must have the capability! 2874 2875 null_check_receiver(); // null-check, then ignore 2876 Node* cls = null_check(argument(1)); 2877 if (stopped()) return true; 2878 2879 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2880 kls = null_check(kls); 2881 if (stopped()) return true; // argument was like int.class 2882 2883 Node* test = NULL; 2884 if (LibraryCallKit::klass_needs_init_guard(kls)) { 2885 // Note: The argument might still be an illegal value like 2886 // Serializable.class or Object[].class. The runtime will handle it. 2887 // But we must make an explicit check for initialization. 2888 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 2889 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 2890 // can generate code to load it as unsigned byte. 2891 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 2892 Node* bits = intcon(InstanceKlass::fully_initialized); 2893 test = _gvn.transform(new SubINode(inst, bits)); 2894 // The 'test' is non-zero if we need to take a slow path. 2895 } 2896 2897 Node* obj = new_instance(kls, test); 2898 set_result(obj); 2899 return true; 2900 } 2901 2902 //------------------------inline_native_time_funcs-------------- 2903 // inline code for System.currentTimeMillis() and System.nanoTime() 2904 // these have the same type and signature 2905 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 2906 const TypeFunc* tf = OptoRuntime::void_long_Type(); 2907 const TypePtr* no_memory_effects = NULL; 2908 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2909 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 2910 #ifdef ASSERT 2911 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 2912 assert(value_top == top(), "second value must be top"); 2913 #endif 2914 set_result(value); 2915 return true; 2916 } 2917 2918 #ifdef JFR_HAVE_INTRINSICS 2919 2920 /* 2921 * oop -> myklass 2922 * myklass->trace_id |= USED 2923 * return myklass->trace_id & ~0x3 2924 */ 2925 bool LibraryCallKit::inline_native_classID() { 2926 Node* cls = null_check(argument(0), T_OBJECT); 2927 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2928 kls = null_check(kls, T_OBJECT); 2929 2930 ByteSize offset = KLASS_TRACE_ID_OFFSET; 2931 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 2932 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 2933 2934 Node* clsused = longcon(0x01l); // set the class bit 2935 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 2936 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 2937 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 2938 2939 #ifdef TRACE_ID_META_BITS 2940 Node* mbits = longcon(~TRACE_ID_META_BITS); 2941 tvalue = _gvn.transform(new AndLNode(tvalue, mbits)); 2942 #endif 2943 #ifdef TRACE_ID_SHIFT 2944 Node* cbits = intcon(TRACE_ID_SHIFT); 2945 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits)); 2946 #endif 2947 2948 set_result(tvalue); 2949 return true; 2950 2951 } 2952 2953 bool LibraryCallKit::inline_native_getEventWriter() { 2954 Node* tls_ptr = _gvn.transform(new ThreadLocalNode()); 2955 2956 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr, 2957 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)); 2958 2959 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered); 2960 2961 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) ); 2962 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) ); 2963 2964 IfNode* iff_jobj_null = 2965 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN); 2966 2967 enum { _normal_path = 1, 2968 _null_path = 2, 2969 PATH_LIMIT }; 2970 2971 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 2972 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM); 2973 2974 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null)); 2975 result_rgn->init_req(_null_path, jobj_is_null); 2976 result_val->init_req(_null_path, null()); 2977 2978 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null)); 2979 set_control(jobj_is_not_null); 2980 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT, 2981 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD); 2982 result_rgn->init_req(_normal_path, control()); 2983 result_val->init_req(_normal_path, res); 2984 2985 set_result(result_rgn, result_val); 2986 2987 return true; 2988 } 2989 2990 #endif // JFR_HAVE_INTRINSICS 2991 2992 //------------------------inline_native_currentThread------------------ 2993 bool LibraryCallKit::inline_native_currentThread() { 2994 Node* junk = NULL; 2995 set_result(generate_current_thread(junk)); 2996 return true; 2997 } 2998 2999 //------------------------inline_native_isInterrupted------------------ 3000 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3001 bool LibraryCallKit::inline_native_isInterrupted() { 3002 // Add a fast path to t.isInterrupted(clear_int): 3003 // (t == Thread.current() && 3004 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3005 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3006 // So, in the common case that the interrupt bit is false, 3007 // we avoid making a call into the VM. Even if the interrupt bit 3008 // is true, if the clear_int argument is false, we avoid the VM call. 3009 // However, if the receiver is not currentThread, we must call the VM, 3010 // because there must be some locking done around the operation. 3011 3012 // We only go to the fast case code if we pass two guards. 3013 // Paths which do not pass are accumulated in the slow_region. 3014 3015 enum { 3016 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3017 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3018 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3019 PATH_LIMIT 3020 }; 3021 3022 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3023 // out of the function. 3024 insert_mem_bar(Op_MemBarCPUOrder); 3025 3026 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3027 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3028 3029 RegionNode* slow_region = new RegionNode(1); 3030 record_for_igvn(slow_region); 3031 3032 // (a) Receiving thread must be the current thread. 3033 Node* rec_thr = argument(0); 3034 Node* tls_ptr = NULL; 3035 Node* cur_thr = generate_current_thread(tls_ptr); 3036 3037 // Resolve oops to stable for CmpP below. 3038 cur_thr = access_resolve(cur_thr, 0); 3039 rec_thr = access_resolve(rec_thr, 0); 3040 3041 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3042 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3043 3044 generate_slow_guard(bol_thr, slow_region); 3045 3046 // (b) Interrupt bit on TLS must be false. 3047 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3048 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3049 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3050 3051 // Set the control input on the field _interrupted read to prevent it floating up. 3052 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3053 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3054 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3055 3056 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3057 3058 // First fast path: if (!TLS._interrupted) return false; 3059 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3060 result_rgn->init_req(no_int_result_path, false_bit); 3061 result_val->init_req(no_int_result_path, intcon(0)); 3062 3063 // drop through to next case 3064 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3065 3066 #ifndef _WINDOWS 3067 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3068 Node* clr_arg = argument(1); 3069 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3070 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3071 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3072 3073 // Second fast path: ... else if (!clear_int) return true; 3074 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3075 result_rgn->init_req(no_clear_result_path, false_arg); 3076 result_val->init_req(no_clear_result_path, intcon(1)); 3077 3078 // drop through to next case 3079 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3080 #else 3081 // To return true on Windows you must read the _interrupted field 3082 // and check the event state i.e. take the slow path. 3083 #endif // _WINDOWS 3084 3085 // (d) Otherwise, go to the slow path. 3086 slow_region->add_req(control()); 3087 set_control( _gvn.transform(slow_region)); 3088 3089 if (stopped()) { 3090 // There is no slow path. 3091 result_rgn->init_req(slow_result_path, top()); 3092 result_val->init_req(slow_result_path, top()); 3093 } else { 3094 // non-virtual because it is a private non-static 3095 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3096 3097 Node* slow_val = set_results_for_java_call(slow_call); 3098 // this->control() comes from set_results_for_java_call 3099 3100 Node* fast_io = slow_call->in(TypeFunc::I_O); 3101 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3102 3103 // These two phis are pre-filled with copies of of the fast IO and Memory 3104 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3105 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3106 3107 result_rgn->init_req(slow_result_path, control()); 3108 result_io ->init_req(slow_result_path, i_o()); 3109 result_mem->init_req(slow_result_path, reset_memory()); 3110 result_val->init_req(slow_result_path, slow_val); 3111 3112 set_all_memory(_gvn.transform(result_mem)); 3113 set_i_o( _gvn.transform(result_io)); 3114 } 3115 3116 C->set_has_split_ifs(true); // Has chance for split-if optimization 3117 set_result(result_rgn, result_val); 3118 return true; 3119 } 3120 3121 //-----------------------load_klass_from_mirror_common------------------------- 3122 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3123 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3124 // and branch to the given path on the region. 3125 // If never_see_null, take an uncommon trap on null, so we can optimistically 3126 // compile for the non-null case. 3127 // If the region is NULL, force never_see_null = true. 3128 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3129 bool never_see_null, 3130 RegionNode* region, 3131 int null_path, 3132 int offset) { 3133 if (region == NULL) never_see_null = true; 3134 Node* p = basic_plus_adr(mirror, offset); 3135 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3136 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3137 Node* null_ctl = top(); 3138 kls = null_check_oop(kls, &null_ctl, never_see_null); 3139 if (region != NULL) { 3140 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3141 region->init_req(null_path, null_ctl); 3142 } else { 3143 assert(null_ctl == top(), "no loose ends"); 3144 } 3145 return kls; 3146 } 3147 3148 //--------------------(inline_native_Class_query helpers)--------------------- 3149 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3150 // Fall through if (mods & mask) == bits, take the guard otherwise. 3151 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3152 // Branch around if the given klass has the given modifier bit set. 3153 // Like generate_guard, adds a new path onto the region. 3154 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3155 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3156 Node* mask = intcon(modifier_mask); 3157 Node* bits = intcon(modifier_bits); 3158 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3159 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3160 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3161 return generate_fair_guard(bol, region); 3162 } 3163 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3164 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3165 } 3166 3167 Node* LibraryCallKit::generate_value_guard(Node* kls, RegionNode* region) { 3168 return generate_access_flags_guard(kls, JVM_ACC_VALUE, 0, region); 3169 } 3170 3171 //-------------------------inline_native_Class_query------------------- 3172 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3173 const Type* return_type = TypeInt::BOOL; 3174 Node* prim_return_value = top(); // what happens if it's a primitive class? 3175 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3176 bool expect_prim = false; // most of these guys expect to work on refs 3177 3178 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3179 3180 Node* mirror = argument(0); 3181 Node* obj = top(); 3182 3183 switch (id) { 3184 case vmIntrinsics::_isInstance: 3185 // nothing is an instance of a primitive type 3186 prim_return_value = intcon(0); 3187 obj = argument(1); 3188 break; 3189 case vmIntrinsics::_getModifiers: 3190 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3191 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3192 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3193 break; 3194 case vmIntrinsics::_isInterface: 3195 prim_return_value = intcon(0); 3196 break; 3197 case vmIntrinsics::_isArray: 3198 prim_return_value = intcon(0); 3199 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3200 break; 3201 case vmIntrinsics::_isPrimitive: 3202 prim_return_value = intcon(1); 3203 expect_prim = true; // obviously 3204 break; 3205 case vmIntrinsics::_getSuperclass: 3206 prim_return_value = null(); 3207 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3208 break; 3209 case vmIntrinsics::_getClassAccessFlags: 3210 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3211 return_type = TypeInt::INT; // not bool! 6297094 3212 break; 3213 default: 3214 fatal_unexpected_iid(id); 3215 break; 3216 } 3217 3218 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3219 if (mirror_con == NULL) return false; // cannot happen? 3220 3221 #ifndef PRODUCT 3222 if (C->print_intrinsics() || C->print_inlining()) { 3223 ciType* k = mirror_con->java_mirror_type(); 3224 if (k) { 3225 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3226 k->print_name(); 3227 tty->cr(); 3228 } 3229 } 3230 #endif 3231 3232 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3233 RegionNode* region = new RegionNode(PATH_LIMIT); 3234 record_for_igvn(region); 3235 PhiNode* phi = new PhiNode(region, return_type); 3236 3237 // The mirror will never be null of Reflection.getClassAccessFlags, however 3238 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3239 // if it is. See bug 4774291. 3240 3241 // For Reflection.getClassAccessFlags(), the null check occurs in 3242 // the wrong place; see inline_unsafe_access(), above, for a similar 3243 // situation. 3244 mirror = null_check(mirror); 3245 // If mirror or obj is dead, only null-path is taken. 3246 if (stopped()) return true; 3247 3248 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3249 3250 // Now load the mirror's klass metaobject, and null-check it. 3251 // Side-effects region with the control path if the klass is null. 3252 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3253 // If kls is null, we have a primitive mirror. 3254 phi->init_req(_prim_path, prim_return_value); 3255 if (stopped()) { set_result(region, phi); return true; } 3256 bool safe_for_replace = (region->in(_prim_path) == top()); 3257 3258 Node* p; // handy temp 3259 Node* null_ctl; 3260 3261 // Now that we have the non-null klass, we can perform the real query. 3262 // For constant classes, the query will constant-fold in LoadNode::Value. 3263 Node* query_value = top(); 3264 switch (id) { 3265 case vmIntrinsics::_isInstance: 3266 // nothing is an instance of a primitive type 3267 query_value = gen_instanceof(obj, kls, safe_for_replace); 3268 break; 3269 3270 case vmIntrinsics::_getModifiers: 3271 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3272 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3273 break; 3274 3275 case vmIntrinsics::_isInterface: 3276 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3277 if (generate_interface_guard(kls, region) != NULL) 3278 // A guard was added. If the guard is taken, it was an interface. 3279 phi->add_req(intcon(1)); 3280 // If we fall through, it's a plain class. 3281 query_value = intcon(0); 3282 break; 3283 3284 case vmIntrinsics::_isArray: 3285 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3286 if (generate_array_guard(kls, region) != NULL) 3287 // A guard was added. If the guard is taken, it was an array. 3288 phi->add_req(intcon(1)); 3289 // If we fall through, it's a plain class. 3290 query_value = intcon(0); 3291 break; 3292 3293 case vmIntrinsics::_isPrimitive: 3294 query_value = intcon(0); // "normal" path produces false 3295 break; 3296 3297 case vmIntrinsics::_getSuperclass: 3298 // The rules here are somewhat unfortunate, but we can still do better 3299 // with random logic than with a JNI call. 3300 // Interfaces store null or Object as _super, but must report null. 3301 // Arrays store an intermediate super as _super, but must report Object. 3302 // Other types can report the actual _super. 3303 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3304 if (generate_interface_guard(kls, region) != NULL) 3305 // A guard was added. If the guard is taken, it was an interface. 3306 phi->add_req(null()); 3307 if (generate_array_guard(kls, region) != NULL) 3308 // A guard was added. If the guard is taken, it was an array. 3309 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3310 // If we fall through, it's a plain class. Get its _super. 3311 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3312 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3313 null_ctl = top(); 3314 kls = null_check_oop(kls, &null_ctl); 3315 if (null_ctl != top()) { 3316 // If the guard is taken, Object.superClass is null (both klass and mirror). 3317 region->add_req(null_ctl); 3318 phi ->add_req(null()); 3319 } 3320 if (!stopped()) { 3321 query_value = load_mirror_from_klass(kls); 3322 } 3323 break; 3324 3325 case vmIntrinsics::_getClassAccessFlags: 3326 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3327 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3328 break; 3329 3330 default: 3331 fatal_unexpected_iid(id); 3332 break; 3333 } 3334 3335 // Fall-through is the normal case of a query to a real class. 3336 phi->init_req(1, query_value); 3337 region->init_req(1, control()); 3338 3339 C->set_has_split_ifs(true); // Has chance for split-if optimization 3340 set_result(region, phi); 3341 return true; 3342 } 3343 3344 //-------------------------inline_Class_cast------------------- 3345 bool LibraryCallKit::inline_Class_cast() { 3346 Node* mirror = argument(0); // Class 3347 Node* obj = argument(1); 3348 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3349 if (mirror_con == NULL) { 3350 return false; // dead path (mirror->is_top()). 3351 } 3352 if (obj == NULL || obj->is_top()) { 3353 return false; // dead path 3354 } 3355 3356 ciKlass* obj_klass = NULL; 3357 if (obj->is_ValueType()) { 3358 const TypeValueType* tvt = _gvn.type(obj)->is_valuetype(); 3359 obj_klass = tvt->value_klass(); 3360 } else { 3361 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3362 if (tp != NULL) { 3363 obj_klass = tp->klass(); 3364 } 3365 } 3366 3367 // First, see if Class.cast() can be folded statically. 3368 // java_mirror_type() returns non-null for compile-time Class constants. 3369 ciType* tm = mirror_con->java_mirror_type(); 3370 if (tm != NULL && tm->is_klass() && 3371 obj_klass != NULL) { 3372 if (!obj_klass->is_loaded()) { 3373 // Don't use intrinsic when class is not loaded. 3374 return false; 3375 } else { 3376 int static_res = C->static_subtype_check(tm->as_klass(), obj_klass); 3377 if (static_res == Compile::SSC_always_true) { 3378 // isInstance() is true - fold the code. 3379 set_result(obj); 3380 return true; 3381 } else if (static_res == Compile::SSC_always_false) { 3382 // Don't use intrinsic, have to throw ClassCastException. 3383 // If the reference is null, the non-intrinsic bytecode will 3384 // be optimized appropriately. 3385 return false; 3386 } 3387 } 3388 } 3389 3390 // Bailout intrinsic and do normal inlining if exception path is frequent. 3391 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3392 return false; 3393 } 3394 3395 // Generate dynamic checks. 3396 // Class.cast() is java implementation of _checkcast bytecode. 3397 // Do checkcast (Parse::do_checkcast()) optimizations here. 3398 3399 mirror = null_check(mirror); 3400 // If mirror is dead, only null-path is taken. 3401 if (stopped()) { 3402 return true; 3403 } 3404 3405 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3406 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3407 RegionNode* region = new RegionNode(PATH_LIMIT); 3408 record_for_igvn(region); 3409 3410 // Now load the mirror's klass metaobject, and null-check it. 3411 // If kls is null, we have a primitive mirror and 3412 // nothing is an instance of a primitive type. 3413 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3414 3415 Node* res = top(); 3416 if (!stopped()) { 3417 Node* bad_type_ctrl = top(); 3418 // Do checkcast optimizations. 3419 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3420 region->init_req(_bad_type_path, bad_type_ctrl); 3421 } 3422 if (region->in(_prim_path) != top() || 3423 region->in(_bad_type_path) != top()) { 3424 // Let Interpreter throw ClassCastException. 3425 PreserveJVMState pjvms(this); 3426 set_control(_gvn.transform(region)); 3427 uncommon_trap(Deoptimization::Reason_intrinsic, 3428 Deoptimization::Action_maybe_recompile); 3429 } 3430 if (!stopped()) { 3431 set_result(res); 3432 } 3433 return true; 3434 } 3435 3436 3437 //--------------------------inline_native_subtype_check------------------------ 3438 // This intrinsic takes the JNI calls out of the heart of 3439 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3440 bool LibraryCallKit::inline_native_subtype_check() { 3441 // Pull both arguments off the stack. 3442 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3443 args[0] = argument(0); 3444 args[1] = argument(1); 3445 Node* klasses[2]; // corresponding Klasses: superk, subk 3446 klasses[0] = klasses[1] = top(); 3447 3448 enum { 3449 // A full decision tree on {superc is prim, subc is prim}: 3450 _prim_0_path = 1, // {P,N} => false 3451 // {P,P} & superc!=subc => false 3452 _prim_same_path, // {P,P} & superc==subc => true 3453 _prim_1_path, // {N,P} => false 3454 _ref_subtype_path, // {N,N} & subtype check wins => true 3455 _both_ref_path, // {N,N} & subtype check loses => false 3456 PATH_LIMIT 3457 }; 3458 3459 RegionNode* region = new RegionNode(PATH_LIMIT); 3460 Node* phi = new PhiNode(region, TypeInt::BOOL); 3461 record_for_igvn(region); 3462 3463 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3464 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3465 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3466 3467 // First null-check both mirrors and load each mirror's klass metaobject. 3468 int which_arg; 3469 for (which_arg = 0; which_arg <= 1; which_arg++) { 3470 Node* arg = args[which_arg]; 3471 arg = null_check(arg); 3472 if (stopped()) break; 3473 args[which_arg] = arg; 3474 3475 Node* p = basic_plus_adr(arg, class_klass_offset); 3476 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3477 klasses[which_arg] = _gvn.transform(kls); 3478 } 3479 3480 // Resolve oops to stable for CmpP below. 3481 args[0] = access_resolve(args[0], 0); 3482 args[1] = access_resolve(args[1], 0); 3483 3484 // Having loaded both klasses, test each for null. 3485 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3486 for (which_arg = 0; which_arg <= 1; which_arg++) { 3487 Node* kls = klasses[which_arg]; 3488 Node* null_ctl = top(); 3489 kls = null_check_oop(kls, &null_ctl, never_see_null); 3490 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3491 region->init_req(prim_path, null_ctl); 3492 if (stopped()) break; 3493 klasses[which_arg] = kls; 3494 } 3495 3496 if (!stopped()) { 3497 // now we have two reference types, in klasses[0..1] 3498 Node* subk = klasses[1]; // the argument to isAssignableFrom 3499 Node* superk = klasses[0]; // the receiver 3500 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3501 // now we have a successful reference subtype check 3502 region->set_req(_ref_subtype_path, control()); 3503 } 3504 3505 // If both operands are primitive (both klasses null), then 3506 // we must return true when they are identical primitives. 3507 // It is convenient to test this after the first null klass check. 3508 set_control(region->in(_prim_0_path)); // go back to first null check 3509 if (!stopped()) { 3510 // Since superc is primitive, make a guard for the superc==subc case. 3511 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3512 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3513 generate_guard(bol_eq, region, PROB_FAIR); 3514 if (region->req() == PATH_LIMIT+1) { 3515 // A guard was added. If the added guard is taken, superc==subc. 3516 region->swap_edges(PATH_LIMIT, _prim_same_path); 3517 region->del_req(PATH_LIMIT); 3518 } 3519 region->set_req(_prim_0_path, control()); // Not equal after all. 3520 } 3521 3522 // these are the only paths that produce 'true': 3523 phi->set_req(_prim_same_path, intcon(1)); 3524 phi->set_req(_ref_subtype_path, intcon(1)); 3525 3526 // pull together the cases: 3527 assert(region->req() == PATH_LIMIT, "sane region"); 3528 for (uint i = 1; i < region->req(); i++) { 3529 Node* ctl = region->in(i); 3530 if (ctl == NULL || ctl == top()) { 3531 region->set_req(i, top()); 3532 phi ->set_req(i, top()); 3533 } else if (phi->in(i) == NULL) { 3534 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3535 } 3536 } 3537 3538 set_control(_gvn.transform(region)); 3539 set_result(_gvn.transform(phi)); 3540 return true; 3541 } 3542 3543 //---------------------generate_array_guard_common------------------------ 3544 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, ArrayKind kind) { 3545 3546 if (stopped()) { 3547 return NULL; 3548 } 3549 3550 // Like generate_guard, adds a new path onto the region. 3551 jint layout_con = 0; 3552 Node* layout_val = get_layout_helper(kls, layout_con); 3553 if (layout_val == NULL) { 3554 bool query = 0; 3555 switch(kind) { 3556 case ObjectArray: query = Klass::layout_helper_is_objArray(layout_con); break; 3557 case NonObjectArray: query = !Klass::layout_helper_is_objArray(layout_con); break; 3558 case TypeArray: query = Klass::layout_helper_is_typeArray(layout_con); break; 3559 case ValueArray: query = Klass::layout_helper_is_valueArray(layout_con); break; 3560 case AnyArray: query = Klass::layout_helper_is_array(layout_con); break; 3561 case NonArray: query = !Klass::layout_helper_is_array(layout_con); break; 3562 default: 3563 ShouldNotReachHere(); 3564 } 3565 if (!query) { 3566 return NULL; // never a branch 3567 } else { // always a branch 3568 Node* always_branch = control(); 3569 if (region != NULL) 3570 region->add_req(always_branch); 3571 set_control(top()); 3572 return always_branch; 3573 } 3574 } 3575 unsigned int value = 0; 3576 BoolTest::mask btest = BoolTest::illegal; 3577 switch(kind) { 3578 case ObjectArray: 3579 case NonObjectArray: { 3580 value = Klass::_lh_array_tag_obj_value; 3581 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3582 btest = kind == ObjectArray ? BoolTest::eq : BoolTest::ne; 3583 break; 3584 } 3585 case TypeArray: { 3586 value = Klass::_lh_array_tag_type_value; 3587 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3588 btest = BoolTest::eq; 3589 break; 3590 } 3591 case ValueArray: { 3592 value = Klass::_lh_array_tag_vt_value; 3593 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3594 btest = BoolTest::eq; 3595 break; 3596 } 3597 case AnyArray: value = Klass::_lh_neutral_value; btest = BoolTest::lt; break; 3598 case NonArray: value = Klass::_lh_neutral_value; btest = BoolTest::gt; break; 3599 default: 3600 ShouldNotReachHere(); 3601 } 3602 // Now test the correct condition. 3603 jint nval = (jint)value; 3604 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3605 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3606 return generate_fair_guard(bol, region); 3607 } 3608 3609 3610 //-----------------------inline_native_newArray-------------------------- 3611 // private static native Object java.lang.reflect.Array.newArray(Class<?> componentType, int length); 3612 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 3613 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 3614 Node* mirror; 3615 Node* count_val; 3616 if (uninitialized) { 3617 mirror = argument(1); 3618 count_val = argument(2); 3619 } else { 3620 mirror = argument(0); 3621 count_val = argument(1); 3622 } 3623 3624 mirror = null_check(mirror); 3625 // If mirror or obj is dead, only null-path is taken. 3626 if (stopped()) return true; 3627 3628 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3629 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3630 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3631 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3632 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3633 3634 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3635 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3636 result_reg, _slow_path); 3637 Node* normal_ctl = control(); 3638 Node* no_array_ctl = result_reg->in(_slow_path); 3639 3640 // Generate code for the slow case. We make a call to newArray(). 3641 set_control(no_array_ctl); 3642 if (!stopped()) { 3643 // Either the input type is void.class, or else the 3644 // array klass has not yet been cached. Either the 3645 // ensuing call will throw an exception, or else it 3646 // will cache the array klass for next time. 3647 PreserveJVMState pjvms(this); 3648 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3649 Node* slow_result = set_results_for_java_call(slow_call); 3650 // this->control() comes from set_results_for_java_call 3651 result_reg->set_req(_slow_path, control()); 3652 result_val->set_req(_slow_path, slow_result); 3653 result_io ->set_req(_slow_path, i_o()); 3654 result_mem->set_req(_slow_path, reset_memory()); 3655 } 3656 3657 set_control(normal_ctl); 3658 if (!stopped()) { 3659 // Normal case: The array type has been cached in the java.lang.Class. 3660 // The following call works fine even if the array type is polymorphic. 3661 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3662 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3663 result_reg->init_req(_normal_path, control()); 3664 result_val->init_req(_normal_path, obj); 3665 result_io ->init_req(_normal_path, i_o()); 3666 result_mem->init_req(_normal_path, reset_memory()); 3667 3668 if (uninitialized) { 3669 // Mark the allocation so that zeroing is skipped 3670 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 3671 alloc->maybe_set_complete(&_gvn); 3672 } 3673 } 3674 3675 // Return the combined state. 3676 set_i_o( _gvn.transform(result_io) ); 3677 set_all_memory( _gvn.transform(result_mem)); 3678 3679 C->set_has_split_ifs(true); // Has chance for split-if optimization 3680 set_result(result_reg, result_val); 3681 return true; 3682 } 3683 3684 //----------------------inline_native_getLength-------------------------- 3685 // public static native int java.lang.reflect.Array.getLength(Object array); 3686 bool LibraryCallKit::inline_native_getLength() { 3687 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3688 3689 Node* array = null_check(argument(0)); 3690 // If array is dead, only null-path is taken. 3691 if (stopped()) return true; 3692 3693 // Deoptimize if it is a non-array. 3694 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3695 3696 if (non_array != NULL) { 3697 PreserveJVMState pjvms(this); 3698 set_control(non_array); 3699 uncommon_trap(Deoptimization::Reason_intrinsic, 3700 Deoptimization::Action_maybe_recompile); 3701 } 3702 3703 // If control is dead, only non-array-path is taken. 3704 if (stopped()) return true; 3705 3706 // The works fine even if the array type is polymorphic. 3707 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3708 Node* result = load_array_length(array); 3709 3710 C->set_has_split_ifs(true); // Has chance for split-if optimization 3711 set_result(result); 3712 return true; 3713 } 3714 3715 //------------------------inline_array_copyOf---------------------------- 3716 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3717 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3718 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3719 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3720 3721 // Get the arguments. 3722 Node* original = argument(0); 3723 Node* start = is_copyOfRange? argument(1): intcon(0); 3724 Node* end = is_copyOfRange? argument(2): argument(1); 3725 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3726 3727 const TypeAryPtr* original_t = _gvn.type(original)->isa_aryptr(); 3728 const TypeInstPtr* mirror_t = _gvn.type(array_type_mirror)->isa_instptr(); 3729 if (EnableValhalla && ValueArrayFlatten && 3730 (original_t == NULL || mirror_t == NULL || 3731 (mirror_t->java_mirror_type() == NULL && 3732 (original_t->elem()->isa_valuetype() || 3733 (original_t->elem()->make_oopptr() != NULL && 3734 original_t->elem()->make_oopptr()->can_be_value_type()))))) { 3735 // We need to know statically if the copy is to a flattened array 3736 // or not but can't tell. 3737 return false; 3738 } 3739 3740 Node* newcopy = NULL; 3741 3742 // Set the original stack and the reexecute bit for the interpreter to reexecute 3743 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3744 { PreserveReexecuteState preexecs(this); 3745 jvms()->set_should_reexecute(true); 3746 3747 array_type_mirror = null_check(array_type_mirror); 3748 original = null_check(original); 3749 3750 // Check if a null path was taken unconditionally. 3751 if (stopped()) return true; 3752 3753 Node* orig_length = load_array_length(original); 3754 3755 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3756 klass_node = null_check(klass_node); 3757 3758 RegionNode* bailout = new RegionNode(1); 3759 record_for_igvn(bailout); 3760 3761 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3762 // Bail out if that is so. 3763 // Value type array may have object field that would require a 3764 // write barrier. Conservatively, go to slow path. 3765 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 3766 Node* not_objArray = !bs->array_copy_requires_gc_barriers(false, T_OBJECT, false, BarrierSetC2::Parsing) ? 3767 generate_typeArray_guard(klass_node, bailout) : generate_non_objArray_guard(klass_node, bailout); 3768 if (not_objArray != NULL) { 3769 // Improve the klass node's type from the new optimistic assumption: 3770 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3771 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0)); 3772 Node* cast = new CastPPNode(klass_node, akls); 3773 cast->init_req(0, control()); 3774 klass_node = _gvn.transform(cast); 3775 } 3776 3777 Node* original_kls = load_object_klass(original); 3778 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3779 // loads/stores but it is legal only if we're sure the 3780 // Arrays.copyOf would succeed. So we need all input arguments 3781 // to the copyOf to be validated, including that the copy to the 3782 // new array won't trigger an ArrayStoreException. That subtype 3783 // check can be optimized if we know something on the type of 3784 // the input array from type speculation. 3785 if (_gvn.type(klass_node)->singleton() && !stopped()) { 3786 ciKlass* subk = _gvn.type(original_kls)->is_klassptr()->klass(); 3787 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3788 3789 int test = C->static_subtype_check(superk, subk); 3790 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3791 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3792 if (t_original->speculative_type() != NULL) { 3793 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3794 original_kls = load_object_klass(original); 3795 } 3796 } 3797 } 3798 3799 if (EnableValhalla) { 3800 // Either both or neither new array klass and original array 3801 // klass must be flattened 3802 Node* flattened_klass = generate_valueArray_guard(klass_node, NULL); 3803 generate_valueArray_guard(original_kls, bailout); 3804 if (flattened_klass != NULL) { 3805 RegionNode* r = new RegionNode(2); 3806 record_for_igvn(r); 3807 r->init_req(1, control()); 3808 set_control(flattened_klass); 3809 generate_valueArray_guard(original_kls, r); 3810 bailout->add_req(control()); 3811 set_control(_gvn.transform(r)); 3812 } 3813 } 3814 3815 // Bail out if either start or end is negative. 3816 generate_negative_guard(start, bailout, &start); 3817 generate_negative_guard(end, bailout, &end); 3818 3819 Node* length = end; 3820 if (_gvn.type(start) != TypeInt::ZERO) { 3821 length = _gvn.transform(new SubINode(end, start)); 3822 } 3823 3824 // Bail out if length is negative. 3825 // Without this the new_array would throw 3826 // NegativeArraySizeException but IllegalArgumentException is what 3827 // should be thrown 3828 generate_negative_guard(length, bailout, &length); 3829 3830 if (bailout->req() > 1) { 3831 PreserveJVMState pjvms(this); 3832 set_control(_gvn.transform(bailout)); 3833 uncommon_trap(Deoptimization::Reason_intrinsic, 3834 Deoptimization::Action_maybe_recompile); 3835 } 3836 3837 if (!stopped()) { 3838 // How many elements will we copy from the original? 3839 // The answer is MinI(orig_length - start, length). 3840 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3841 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3842 3843 original = access_resolve(original, ACCESS_READ); 3844 3845 // Generate a direct call to the right arraycopy function(s). 3846 // We know the copy is disjoint but we might not know if the 3847 // oop stores need checking. 3848 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3849 // This will fail a store-check if x contains any non-nulls. 3850 3851 bool validated = false; 3852 // Reason_class_check rather than Reason_intrinsic because we 3853 // want to intrinsify even if this traps. 3854 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3855 Node* not_subtype_ctrl = gen_subtype_check(original_kls, 3856 klass_node); 3857 3858 if (not_subtype_ctrl != top()) { 3859 PreserveJVMState pjvms(this); 3860 set_control(not_subtype_ctrl); 3861 uncommon_trap(Deoptimization::Reason_class_check, 3862 Deoptimization::Action_make_not_entrant); 3863 assert(stopped(), "Should be stopped"); 3864 } 3865 validated = true; 3866 } 3867 3868 if (!stopped()) { 3869 newcopy = new_array(klass_node, length, 0); // no arguments to push 3870 3871 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false, 3872 original_kls, klass_node); 3873 if (!is_copyOfRange) { 3874 ac->set_copyof(validated); 3875 } else { 3876 ac->set_copyofrange(validated); 3877 } 3878 Node* n = _gvn.transform(ac); 3879 if (n == ac) { 3880 ac->connect_outputs(this); 3881 } else { 3882 assert(validated, "shouldn't transform if all arguments not validated"); 3883 set_all_memory(n); 3884 } 3885 } 3886 } 3887 } // original reexecute is set back here 3888 3889 C->set_has_split_ifs(true); // Has chance for split-if optimization 3890 if (!stopped()) { 3891 set_result(newcopy); 3892 } 3893 return true; 3894 } 3895 3896 3897 //----------------------generate_virtual_guard--------------------------- 3898 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3899 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3900 RegionNode* slow_region) { 3901 ciMethod* method = callee(); 3902 int vtable_index = method->vtable_index(); 3903 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3904 "bad index %d", vtable_index); 3905 // Get the Method* out of the appropriate vtable entry. 3906 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 3907 vtable_index*vtableEntry::size_in_bytes() + 3908 vtableEntry::method_offset_in_bytes(); 3909 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3910 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3911 3912 // Compare the target method with the expected method (e.g., Object.hashCode). 3913 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3914 3915 Node* native_call = makecon(native_call_addr); 3916 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3917 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3918 3919 return generate_slow_guard(test_native, slow_region); 3920 } 3921 3922 //-----------------------generate_method_call---------------------------- 3923 // Use generate_method_call to make a slow-call to the real 3924 // method if the fast path fails. An alternative would be to 3925 // use a stub like OptoRuntime::slow_arraycopy_Java. 3926 // This only works for expanding the current library call, 3927 // not another intrinsic. (E.g., don't use this for making an 3928 // arraycopy call inside of the copyOf intrinsic.) 3929 CallJavaNode* 3930 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3931 // When compiling the intrinsic method itself, do not use this technique. 3932 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3933 3934 ciMethod* method = callee(); 3935 // ensure the JVMS we have will be correct for this call 3936 guarantee(method_id == method->intrinsic_id(), "must match"); 3937 3938 const TypeFunc* tf = TypeFunc::make(method); 3939 CallJavaNode* slow_call; 3940 if (is_static) { 3941 assert(!is_virtual, ""); 3942 slow_call = new CallStaticJavaNode(C, tf, 3943 SharedRuntime::get_resolve_static_call_stub(), 3944 method, bci()); 3945 } else if (is_virtual) { 3946 null_check_receiver(); 3947 int vtable_index = Method::invalid_vtable_index; 3948 if (UseInlineCaches) { 3949 // Suppress the vtable call 3950 } else { 3951 // hashCode and clone are not a miranda methods, 3952 // so the vtable index is fixed. 3953 // No need to use the linkResolver to get it. 3954 vtable_index = method->vtable_index(); 3955 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3956 "bad index %d", vtable_index); 3957 } 3958 slow_call = new CallDynamicJavaNode(tf, 3959 SharedRuntime::get_resolve_virtual_call_stub(), 3960 method, vtable_index, bci()); 3961 } else { // neither virtual nor static: opt_virtual 3962 null_check_receiver(); 3963 slow_call = new CallStaticJavaNode(C, tf, 3964 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3965 method, bci()); 3966 slow_call->set_optimized_virtual(true); 3967 } 3968 set_arguments_for_java_call(slow_call); 3969 set_edges_for_java_call(slow_call); 3970 return slow_call; 3971 } 3972 3973 3974 /** 3975 * Build special case code for calls to hashCode on an object. This call may 3976 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 3977 * slightly different code. 3978 */ 3979 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3980 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3981 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3982 3983 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3984 3985 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3986 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 3987 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3988 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3989 Node* obj = argument(0); 3990 3991 if (obj->is_ValueType() || gvn().type(obj)->is_valuetypeptr()) { 3992 return false; 3993 } 3994 3995 if (!is_static) { 3996 // Check for hashing null object 3997 obj = null_check_receiver(); 3998 if (stopped()) return true; // unconditionally null 3999 result_reg->init_req(_null_path, top()); 4000 result_val->init_req(_null_path, top()); 4001 } else { 4002 // Do a null check, and return zero if null. 4003 // System.identityHashCode(null) == 0 4004 Node* null_ctl = top(); 4005 obj = null_check_oop(obj, &null_ctl); 4006 result_reg->init_req(_null_path, null_ctl); 4007 result_val->init_req(_null_path, _gvn.intcon(0)); 4008 } 4009 4010 // Unconditionally null? Then return right away. 4011 if (stopped()) { 4012 set_control( result_reg->in(_null_path)); 4013 if (!stopped()) 4014 set_result(result_val->in(_null_path)); 4015 return true; 4016 } 4017 4018 // We only go to the fast case code if we pass a number of guards. The 4019 // paths which do not pass are accumulated in the slow_region. 4020 RegionNode* slow_region = new RegionNode(1); 4021 record_for_igvn(slow_region); 4022 4023 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4024 assert(!obj_type->isa_valuetype() || !obj_type->is_valuetypeptr(), "no value type here"); 4025 if (is_static && obj_type->can_be_value_type()) { 4026 Node* obj_klass = load_object_klass(obj); 4027 generate_value_guard(obj_klass, slow_region); 4028 } 4029 4030 // If this is a virtual call, we generate a funny guard. We pull out 4031 // the vtable entry corresponding to hashCode() from the target object. 4032 // If the target method which we are calling happens to be the native 4033 // Object hashCode() method, we pass the guard. We do not need this 4034 // guard for non-virtual calls -- the caller is known to be the native 4035 // Object hashCode(). 4036 if (is_virtual) { 4037 // After null check, get the object's klass. 4038 Node* obj_klass = load_object_klass(obj); 4039 generate_virtual_guard(obj_klass, slow_region); 4040 } 4041 4042 // Get the header out of the object, use LoadMarkNode when available 4043 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4044 // The control of the load must be NULL. Otherwise, the load can move before 4045 // the null check after castPP removal. 4046 Node* no_ctrl = NULL; 4047 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4048 4049 // Test the header to see if it is unlocked. 4050 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4051 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4052 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4053 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4054 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4055 4056 generate_slow_guard(test_unlocked, slow_region); 4057 4058 // Get the hash value and check to see that it has been properly assigned. 4059 // We depend on hash_mask being at most 32 bits and avoid the use of 4060 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4061 // vm: see markOop.hpp. 4062 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4063 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4064 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4065 // This hack lets the hash bits live anywhere in the mark object now, as long 4066 // as the shift drops the relevant bits into the low 32 bits. Note that 4067 // Java spec says that HashCode is an int so there's no point in capturing 4068 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4069 hshifted_header = ConvX2I(hshifted_header); 4070 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4071 4072 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4073 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4074 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4075 4076 generate_slow_guard(test_assigned, slow_region); 4077 4078 Node* init_mem = reset_memory(); 4079 // fill in the rest of the null path: 4080 result_io ->init_req(_null_path, i_o()); 4081 result_mem->init_req(_null_path, init_mem); 4082 4083 result_val->init_req(_fast_path, hash_val); 4084 result_reg->init_req(_fast_path, control()); 4085 result_io ->init_req(_fast_path, i_o()); 4086 result_mem->init_req(_fast_path, init_mem); 4087 4088 // Generate code for the slow case. We make a call to hashCode(). 4089 set_control(_gvn.transform(slow_region)); 4090 if (!stopped()) { 4091 // No need for PreserveJVMState, because we're using up the present state. 4092 set_all_memory(init_mem); 4093 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4094 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4095 Node* slow_result = set_results_for_java_call(slow_call); 4096 // this->control() comes from set_results_for_java_call 4097 result_reg->init_req(_slow_path, control()); 4098 result_val->init_req(_slow_path, slow_result); 4099 result_io ->set_req(_slow_path, i_o()); 4100 result_mem ->set_req(_slow_path, reset_memory()); 4101 } 4102 4103 // Return the combined state. 4104 set_i_o( _gvn.transform(result_io) ); 4105 set_all_memory( _gvn.transform(result_mem)); 4106 4107 set_result(result_reg, result_val); 4108 return true; 4109 } 4110 4111 //---------------------------inline_native_getClass---------------------------- 4112 // public final native Class<?> java.lang.Object.getClass(); 4113 // 4114 // Build special case code for calls to getClass on an object. 4115 bool LibraryCallKit::inline_native_getClass() { 4116 Node* obj = argument(0); 4117 if (obj->is_ValueType()) { 4118 ciKlass* vk = _gvn.type(obj)->is_valuetype()->value_klass(); 4119 set_result(makecon(TypeInstPtr::make(vk->java_mirror()))); 4120 return true; 4121 } 4122 obj = null_check_receiver(); 4123 if (stopped()) return true; 4124 set_result(load_mirror_from_klass(load_object_klass(obj))); 4125 return true; 4126 } 4127 4128 //-----------------inline_native_Reflection_getCallerClass--------------------- 4129 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4130 // 4131 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4132 // 4133 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4134 // in that it must skip particular security frames and checks for 4135 // caller sensitive methods. 4136 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4137 #ifndef PRODUCT 4138 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4139 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4140 } 4141 #endif 4142 4143 if (!jvms()->has_method()) { 4144 #ifndef PRODUCT 4145 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4146 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4147 } 4148 #endif 4149 return false; 4150 } 4151 4152 // Walk back up the JVM state to find the caller at the required 4153 // depth. 4154 JVMState* caller_jvms = jvms(); 4155 4156 // Cf. JVM_GetCallerClass 4157 // NOTE: Start the loop at depth 1 because the current JVM state does 4158 // not include the Reflection.getCallerClass() frame. 4159 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4160 ciMethod* m = caller_jvms->method(); 4161 switch (n) { 4162 case 0: 4163 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4164 break; 4165 case 1: 4166 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4167 if (!m->caller_sensitive()) { 4168 #ifndef PRODUCT 4169 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4170 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4171 } 4172 #endif 4173 return false; // bail-out; let JVM_GetCallerClass do the work 4174 } 4175 break; 4176 default: 4177 if (!m->is_ignored_by_security_stack_walk()) { 4178 // We have reached the desired frame; return the holder class. 4179 // Acquire method holder as java.lang.Class and push as constant. 4180 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4181 ciInstance* caller_mirror = caller_klass->java_mirror(); 4182 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4183 4184 #ifndef PRODUCT 4185 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4186 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()); 4187 tty->print_cr(" JVM state at this point:"); 4188 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4189 ciMethod* m = jvms()->of_depth(i)->method(); 4190 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4191 } 4192 } 4193 #endif 4194 return true; 4195 } 4196 break; 4197 } 4198 } 4199 4200 #ifndef PRODUCT 4201 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4202 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4203 tty->print_cr(" JVM state at this point:"); 4204 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4205 ciMethod* m = jvms()->of_depth(i)->method(); 4206 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4207 } 4208 } 4209 #endif 4210 4211 return false; // bail-out; let JVM_GetCallerClass do the work 4212 } 4213 4214 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4215 Node* arg = argument(0); 4216 Node* result = NULL; 4217 4218 switch (id) { 4219 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4220 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4221 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4222 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4223 4224 case vmIntrinsics::_doubleToLongBits: { 4225 // two paths (plus control) merge in a wood 4226 RegionNode *r = new RegionNode(3); 4227 Node *phi = new PhiNode(r, TypeLong::LONG); 4228 4229 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4230 // Build the boolean node 4231 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4232 4233 // Branch either way. 4234 // NaN case is less traveled, which makes all the difference. 4235 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4236 Node *opt_isnan = _gvn.transform(ifisnan); 4237 assert( opt_isnan->is_If(), "Expect an IfNode"); 4238 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4239 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4240 4241 set_control(iftrue); 4242 4243 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4244 Node *slow_result = longcon(nan_bits); // return NaN 4245 phi->init_req(1, _gvn.transform( slow_result )); 4246 r->init_req(1, iftrue); 4247 4248 // Else fall through 4249 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4250 set_control(iffalse); 4251 4252 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4253 r->init_req(2, iffalse); 4254 4255 // Post merge 4256 set_control(_gvn.transform(r)); 4257 record_for_igvn(r); 4258 4259 C->set_has_split_ifs(true); // Has chance for split-if optimization 4260 result = phi; 4261 assert(result->bottom_type()->isa_long(), "must be"); 4262 break; 4263 } 4264 4265 case vmIntrinsics::_floatToIntBits: { 4266 // two paths (plus control) merge in a wood 4267 RegionNode *r = new RegionNode(3); 4268 Node *phi = new PhiNode(r, TypeInt::INT); 4269 4270 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4271 // Build the boolean node 4272 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4273 4274 // Branch either way. 4275 // NaN case is less traveled, which makes all the difference. 4276 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4277 Node *opt_isnan = _gvn.transform(ifisnan); 4278 assert( opt_isnan->is_If(), "Expect an IfNode"); 4279 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4280 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4281 4282 set_control(iftrue); 4283 4284 static const jint nan_bits = 0x7fc00000; 4285 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4286 phi->init_req(1, _gvn.transform( slow_result )); 4287 r->init_req(1, iftrue); 4288 4289 // Else fall through 4290 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4291 set_control(iffalse); 4292 4293 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4294 r->init_req(2, iffalse); 4295 4296 // Post merge 4297 set_control(_gvn.transform(r)); 4298 record_for_igvn(r); 4299 4300 C->set_has_split_ifs(true); // Has chance for split-if optimization 4301 result = phi; 4302 assert(result->bottom_type()->isa_int(), "must be"); 4303 break; 4304 } 4305 4306 default: 4307 fatal_unexpected_iid(id); 4308 break; 4309 } 4310 set_result(_gvn.transform(result)); 4311 return true; 4312 } 4313 4314 //----------------------inline_unsafe_copyMemory------------------------- 4315 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4316 bool LibraryCallKit::inline_unsafe_copyMemory() { 4317 if (callee()->is_static()) return false; // caller must have the capability! 4318 null_check_receiver(); // null-check receiver 4319 if (stopped()) return true; 4320 4321 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4322 4323 Node* src_ptr = argument(1); // type: oop 4324 Node* src_off = ConvL2X(argument(2)); // type: long 4325 Node* dst_ptr = argument(4); // type: oop 4326 Node* dst_off = ConvL2X(argument(5)); // type: long 4327 Node* size = ConvL2X(argument(7)); // type: long 4328 4329 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4330 "fieldOffset must be byte-scaled"); 4331 4332 src_ptr = access_resolve(src_ptr, ACCESS_READ); 4333 dst_ptr = access_resolve(dst_ptr, ACCESS_WRITE); 4334 Node* src = make_unsafe_address(src_ptr, src_off, ACCESS_READ); 4335 Node* dst = make_unsafe_address(dst_ptr, dst_off, ACCESS_WRITE); 4336 4337 // Conservatively insert a memory barrier on all memory slices. 4338 // Do not let writes of the copy source or destination float below the copy. 4339 insert_mem_bar(Op_MemBarCPUOrder); 4340 4341 // Call it. Note that the length argument is not scaled. 4342 make_runtime_call(RC_LEAF|RC_NO_FP, 4343 OptoRuntime::fast_arraycopy_Type(), 4344 StubRoutines::unsafe_arraycopy(), 4345 "unsafe_arraycopy", 4346 TypeRawPtr::BOTTOM, 4347 src, dst, size XTOP); 4348 4349 // Do not let reads of the copy destination float above the copy. 4350 insert_mem_bar(Op_MemBarCPUOrder); 4351 4352 return true; 4353 } 4354 4355 //------------------------clone_coping----------------------------------- 4356 // Helper function for inline_native_clone. 4357 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) { 4358 assert(obj_size != NULL, ""); 4359 Node* raw_obj = alloc_obj->in(1); 4360 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4361 4362 AllocateNode* alloc = NULL; 4363 if (ReduceBulkZeroing) { 4364 // We will be completely responsible for initializing this object - 4365 // mark Initialize node as complete. 4366 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4367 // The object was just allocated - there should be no any stores! 4368 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4369 // Mark as complete_with_arraycopy so that on AllocateNode 4370 // expansion, we know this AllocateNode is initialized by an array 4371 // copy and a StoreStore barrier exists after the array copy. 4372 alloc->initialization()->set_complete_with_arraycopy(); 4373 } 4374 4375 // Copy the fastest available way. 4376 // TODO: generate fields copies for small objects instead. 4377 Node* size = _gvn.transform(obj_size); 4378 4379 // Exclude the header but include array length to copy by 8 bytes words. 4380 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4381 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4382 instanceOopDesc::base_offset_in_bytes(); 4383 // base_off: 4384 // 8 - 32-bit VM 4385 // 12 - 64-bit VM, compressed klass 4386 // 16 - 64-bit VM, normal klass 4387 if (base_off % BytesPerLong != 0) { 4388 assert(UseCompressedClassPointers, ""); 4389 if (is_array) { 4390 // Exclude length to copy by 8 bytes words. 4391 base_off += sizeof(int); 4392 } else { 4393 // Include klass to copy by 8 bytes words. 4394 base_off = instanceOopDesc::klass_offset_in_bytes(); 4395 } 4396 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4397 } 4398 Node* src_base = basic_plus_adr(obj, base_off); 4399 Node* dst_base = basic_plus_adr(alloc_obj, base_off); 4400 4401 // Compute the length also, if needed: 4402 Node* countx = size; 4403 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4404 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong))); 4405 4406 access_clone(src_base, dst_base, countx, is_array); 4407 4408 // Do not let reads from the cloned object float above the arraycopy. 4409 if (alloc != NULL) { 4410 // Do not let stores that initialize this object be reordered with 4411 // a subsequent store that would make this object accessible by 4412 // other threads. 4413 // Record what AllocateNode this StoreStore protects so that 4414 // escape analysis can go from the MemBarStoreStoreNode to the 4415 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4416 // based on the escape status of the AllocateNode. 4417 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 4418 } else { 4419 insert_mem_bar(Op_MemBarCPUOrder); 4420 } 4421 } 4422 4423 //------------------------inline_native_clone---------------------------- 4424 // protected native Object java.lang.Object.clone(); 4425 // 4426 // Here are the simple edge cases: 4427 // null receiver => normal trap 4428 // virtual and clone was overridden => slow path to out-of-line clone 4429 // not cloneable or finalizer => slow path to out-of-line Object.clone 4430 // 4431 // The general case has two steps, allocation and copying. 4432 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4433 // 4434 // Copying also has two cases, oop arrays and everything else. 4435 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4436 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4437 // 4438 // These steps fold up nicely if and when the cloned object's klass 4439 // can be sharply typed as an object array, a type array, or an instance. 4440 // 4441 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4442 PhiNode* result_val; 4443 4444 // Set the reexecute bit for the interpreter to reexecute 4445 // the bytecode that invokes Object.clone if deoptimization happens. 4446 { PreserveReexecuteState preexecs(this); 4447 jvms()->set_should_reexecute(true); 4448 4449 Node* obj = argument(0); 4450 if (obj->is_ValueType()) { 4451 return false; 4452 } 4453 4454 obj = null_check_receiver(); 4455 if (stopped()) return true; 4456 4457 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4458 4459 // If we are going to clone an instance, we need its exact type to 4460 // know the number and types of fields to convert the clone to 4461 // loads/stores. Maybe a speculative type can help us. 4462 if (!obj_type->klass_is_exact() && 4463 obj_type->speculative_type() != NULL && 4464 obj_type->speculative_type()->is_instance_klass() && 4465 !obj_type->speculative_type()->is_valuetype()) { 4466 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4467 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4468 !spec_ik->has_injected_fields()) { 4469 ciKlass* k = obj_type->klass(); 4470 if (!k->is_instance_klass() || 4471 k->as_instance_klass()->is_interface() || 4472 k->as_instance_klass()->has_subklass()) { 4473 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4474 } 4475 } 4476 } 4477 4478 Node* obj_klass = load_object_klass(obj); 4479 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4480 const TypeOopPtr* toop = ((tklass != NULL) 4481 ? tklass->as_instance_type() 4482 : TypeInstPtr::NOTNULL); 4483 4484 // Conservatively insert a memory barrier on all memory slices. 4485 // Do not let writes into the original float below the clone. 4486 insert_mem_bar(Op_MemBarCPUOrder); 4487 4488 // paths into result_reg: 4489 enum { 4490 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4491 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4492 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4493 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4494 PATH_LIMIT 4495 }; 4496 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4497 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4498 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4499 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4500 record_for_igvn(result_reg); 4501 4502 // We only go to the fast case code if we pass a number of guards. 4503 // The paths which do not pass are accumulated in the slow_region. 4504 RegionNode* slow_region = new RegionNode(1); 4505 record_for_igvn(slow_region); 4506 4507 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4508 if (array_ctl != NULL) { 4509 // It's an array. 4510 PreserveJVMState pjvms(this); 4511 set_control(array_ctl); 4512 4513 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4514 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) { 4515 // Value type array may have object field that would require a 4516 // write barrier. Conservatively, go to slow path. 4517 generate_valueArray_guard(obj_klass, slow_region); 4518 } 4519 4520 if (!stopped()) { 4521 Node* obj_length = load_array_length(obj); 4522 Node* obj_size = NULL; 4523 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4524 4525 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4526 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) { 4527 // If it is an oop array, it requires very special treatment, 4528 // because gc barriers are required when accessing the array. 4529 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4530 if (is_obja != NULL) { 4531 PreserveJVMState pjvms2(this); 4532 set_control(is_obja); 4533 // Generate a direct call to the right arraycopy function(s). 4534 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4535 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false); 4536 ac->set_cloneoop(); 4537 Node* n = _gvn.transform(ac); 4538 assert(n == ac, "cannot disappear"); 4539 ac->connect_outputs(this); 4540 4541 result_reg->init_req(_objArray_path, control()); 4542 result_val->init_req(_objArray_path, alloc_obj); 4543 result_i_o ->set_req(_objArray_path, i_o()); 4544 result_mem ->set_req(_objArray_path, reset_memory()); 4545 } 4546 } 4547 4548 // Otherwise, there are no barriers to worry about. 4549 // (We can dispense with card marks if we know the allocation 4550 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4551 // causes the non-eden paths to take compensating steps to 4552 // simulate a fresh allocation, so that no further 4553 // card marks are required in compiled code to initialize 4554 // the object.) 4555 4556 if (!stopped()) { 4557 copy_to_clone(obj, alloc_obj, obj_size, true); 4558 4559 // Present the results of the copy. 4560 result_reg->init_req(_array_path, control()); 4561 result_val->init_req(_array_path, alloc_obj); 4562 result_i_o ->set_req(_array_path, i_o()); 4563 result_mem ->set_req(_array_path, reset_memory()); 4564 } 4565 } 4566 } 4567 4568 if (!stopped()) { 4569 // It's an instance (we did array above). Make the slow-path tests. 4570 // If this is a virtual call, we generate a funny guard. We grab 4571 // the vtable entry corresponding to clone() from the target object. 4572 // If the target method which we are calling happens to be the 4573 // Object clone() method, we pass the guard. We do not need this 4574 // guard for non-virtual calls; the caller is known to be the native 4575 // Object clone(). 4576 if (is_virtual) { 4577 generate_virtual_guard(obj_klass, slow_region); 4578 } 4579 4580 // The object must be easily cloneable and must not have a finalizer. 4581 // Both of these conditions may be checked in a single test. 4582 // We could optimize the test further, but we don't care. 4583 generate_access_flags_guard(obj_klass, 4584 // Test both conditions: 4585 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4586 // Must be cloneable but not finalizer: 4587 JVM_ACC_IS_CLONEABLE_FAST, 4588 slow_region); 4589 } 4590 4591 if (!stopped()) { 4592 // It's an instance, and it passed the slow-path tests. 4593 PreserveJVMState pjvms(this); 4594 Node* obj_size = NULL; 4595 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4596 // is reexecuted if deoptimization occurs and there could be problems when merging 4597 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4598 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4599 4600 copy_to_clone(obj, alloc_obj, obj_size, false); 4601 4602 // Present the results of the slow call. 4603 result_reg->init_req(_instance_path, control()); 4604 result_val->init_req(_instance_path, alloc_obj); 4605 result_i_o ->set_req(_instance_path, i_o()); 4606 result_mem ->set_req(_instance_path, reset_memory()); 4607 } 4608 4609 // Generate code for the slow case. We make a call to clone(). 4610 set_control(_gvn.transform(slow_region)); 4611 if (!stopped()) { 4612 PreserveJVMState pjvms(this); 4613 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4614 // We need to deoptimize on exception (see comment above) 4615 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true); 4616 // this->control() comes from set_results_for_java_call 4617 result_reg->init_req(_slow_path, control()); 4618 result_val->init_req(_slow_path, slow_result); 4619 result_i_o ->set_req(_slow_path, i_o()); 4620 result_mem ->set_req(_slow_path, reset_memory()); 4621 } 4622 4623 // Return the combined state. 4624 set_control( _gvn.transform(result_reg)); 4625 set_i_o( _gvn.transform(result_i_o)); 4626 set_all_memory( _gvn.transform(result_mem)); 4627 } // original reexecute is set back here 4628 4629 set_result(_gvn.transform(result_val)); 4630 return true; 4631 } 4632 4633 // If we have a tighly coupled allocation, the arraycopy may take care 4634 // of the array initialization. If one of the guards we insert between 4635 // the allocation and the arraycopy causes a deoptimization, an 4636 // unitialized array will escape the compiled method. To prevent that 4637 // we set the JVM state for uncommon traps between the allocation and 4638 // the arraycopy to the state before the allocation so, in case of 4639 // deoptimization, we'll reexecute the allocation and the 4640 // initialization. 4641 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4642 if (alloc != NULL) { 4643 ciMethod* trap_method = alloc->jvms()->method(); 4644 int trap_bci = alloc->jvms()->bci(); 4645 4646 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4647 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4648 // Make sure there's no store between the allocation and the 4649 // arraycopy otherwise visible side effects could be rexecuted 4650 // in case of deoptimization and cause incorrect execution. 4651 bool no_interfering_store = true; 4652 Node* mem = alloc->in(TypeFunc::Memory); 4653 if (mem->is_MergeMem()) { 4654 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4655 Node* n = mms.memory(); 4656 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4657 assert(n->is_Store() || n->Opcode() == Op_ShenandoahWBMemProj, "what else?"); 4658 no_interfering_store = false; 4659 break; 4660 } 4661 } 4662 } else { 4663 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4664 Node* n = mms.memory(); 4665 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4666 assert(n->is_Store() || n->Opcode() == Op_ShenandoahWBMemProj, "what else?"); 4667 no_interfering_store = false; 4668 break; 4669 } 4670 } 4671 } 4672 4673 if (no_interfering_store) { 4674 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4675 uint size = alloc->req(); 4676 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4677 old_jvms->set_map(sfpt); 4678 for (uint i = 0; i < size; i++) { 4679 sfpt->init_req(i, alloc->in(i)); 4680 } 4681 // re-push array length for deoptimization 4682 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4683 old_jvms->set_sp(old_jvms->sp()+1); 4684 old_jvms->set_monoff(old_jvms->monoff()+1); 4685 old_jvms->set_scloff(old_jvms->scloff()+1); 4686 old_jvms->set_endoff(old_jvms->endoff()+1); 4687 old_jvms->set_should_reexecute(true); 4688 4689 sfpt->set_i_o(map()->i_o()); 4690 sfpt->set_memory(map()->memory()); 4691 sfpt->set_control(map()->control()); 4692 4693 JVMState* saved_jvms = jvms(); 4694 saved_reexecute_sp = _reexecute_sp; 4695 4696 set_jvms(sfpt->jvms()); 4697 _reexecute_sp = jvms()->sp(); 4698 4699 return saved_jvms; 4700 } 4701 } 4702 } 4703 return NULL; 4704 } 4705 4706 // In case of a deoptimization, we restart execution at the 4707 // allocation, allocating a new array. We would leave an uninitialized 4708 // array in the heap that GCs wouldn't expect. Move the allocation 4709 // after the traps so we don't allocate the array if we 4710 // deoptimize. This is possible because tightly_coupled_allocation() 4711 // guarantees there's no observer of the allocated array at this point 4712 // and the control flow is simple enough. 4713 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, 4714 int saved_reexecute_sp, uint new_idx) { 4715 if (saved_jvms != NULL && !stopped()) { 4716 assert(alloc != NULL, "only with a tightly coupled allocation"); 4717 // restore JVM state to the state at the arraycopy 4718 saved_jvms->map()->set_control(map()->control()); 4719 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4720 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4721 // If we've improved the types of some nodes (null check) while 4722 // emitting the guards, propagate them to the current state 4723 map()->replaced_nodes().apply(saved_jvms->map(), new_idx); 4724 set_jvms(saved_jvms); 4725 _reexecute_sp = saved_reexecute_sp; 4726 4727 // Remove the allocation from above the guards 4728 CallProjections* callprojs = alloc->extract_projections(true); 4729 InitializeNode* init = alloc->initialization(); 4730 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4731 C->gvn_replace_by(callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4732 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4733 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4734 4735 // move the allocation here (after the guards) 4736 _gvn.hash_delete(alloc); 4737 alloc->set_req(TypeFunc::Control, control()); 4738 alloc->set_req(TypeFunc::I_O, i_o()); 4739 Node *mem = reset_memory(); 4740 set_all_memory(mem); 4741 alloc->set_req(TypeFunc::Memory, mem); 4742 set_control(init->proj_out_or_null(TypeFunc::Control)); 4743 set_i_o(callprojs->fallthrough_ioproj); 4744 4745 // Update memory as done in GraphKit::set_output_for_allocation() 4746 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4747 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4748 if (ary_type->isa_aryptr() && length_type != NULL) { 4749 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4750 } 4751 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4752 int elemidx = C->get_alias_index(telemref); 4753 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw); 4754 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx); 4755 4756 Node* allocx = _gvn.transform(alloc); 4757 assert(allocx == alloc, "where has the allocation gone?"); 4758 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4759 4760 _gvn.hash_delete(dest); 4761 dest->set_req(0, control()); 4762 Node* destx = _gvn.transform(dest); 4763 assert(destx == dest, "where has the allocation result gone?"); 4764 } 4765 } 4766 4767 4768 //------------------------------inline_arraycopy----------------------- 4769 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4770 // Object dest, int destPos, 4771 // int length); 4772 bool LibraryCallKit::inline_arraycopy() { 4773 // Get the arguments. 4774 Node* src = argument(0); // type: oop 4775 Node* src_offset = argument(1); // type: int 4776 Node* dest = argument(2); // type: oop 4777 Node* dest_offset = argument(3); // type: int 4778 Node* length = argument(4); // type: int 4779 4780 uint new_idx = C->unique(); 4781 4782 // Check for allocation before we add nodes that would confuse 4783 // tightly_coupled_allocation() 4784 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4785 4786 int saved_reexecute_sp = -1; 4787 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4788 // See arraycopy_restore_alloc_state() comment 4789 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4790 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4791 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards 4792 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4793 4794 // The following tests must be performed 4795 // (1) src and dest are arrays. 4796 // (2) src and dest arrays must have elements of the same BasicType 4797 // (3) src and dest must not be null. 4798 // (4) src_offset must not be negative. 4799 // (5) dest_offset must not be negative. 4800 // (6) length must not be negative. 4801 // (7) src_offset + length must not exceed length of src. 4802 // (8) dest_offset + length must not exceed length of dest. 4803 // (9) each element of an oop array must be assignable 4804 4805 // (3) src and dest must not be null. 4806 // always do this here because we need the JVM state for uncommon traps 4807 Node* null_ctl = top(); 4808 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4809 assert(null_ctl->is_top(), "no null control here"); 4810 dest = null_check(dest, T_ARRAY); 4811 4812 if (!can_emit_guards) { 4813 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4814 // guards but the arraycopy node could still take advantage of a 4815 // tightly allocated allocation. tightly_coupled_allocation() is 4816 // called again to make sure it takes the null check above into 4817 // account: the null check is mandatory and if it caused an 4818 // uncommon trap to be emitted then the allocation can't be 4819 // considered tightly coupled in this context. 4820 alloc = tightly_coupled_allocation(dest, NULL); 4821 } 4822 4823 bool validated = false; 4824 4825 const Type* src_type = _gvn.type(src); 4826 const Type* dest_type = _gvn.type(dest); 4827 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4828 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4829 4830 // Do we have the type of src? 4831 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4832 // Do we have the type of dest? 4833 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4834 // Is the type for src from speculation? 4835 bool src_spec = false; 4836 // Is the type for dest from speculation? 4837 bool dest_spec = false; 4838 4839 if ((!has_src || !has_dest) && can_emit_guards) { 4840 // We don't have sufficient type information, let's see if 4841 // speculative types can help. We need to have types for both src 4842 // and dest so that it pays off. 4843 4844 // Do we already have or could we have type information for src 4845 bool could_have_src = has_src; 4846 // Do we already have or could we have type information for dest 4847 bool could_have_dest = has_dest; 4848 4849 ciKlass* src_k = NULL; 4850 if (!has_src) { 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 4857 ciKlass* dest_k = NULL; 4858 if (!has_dest) { 4859 dest_k = dest_type->speculative_type_not_null(); 4860 if (dest_k != NULL && dest_k->is_array_klass()) { 4861 could_have_dest = true; 4862 } 4863 } 4864 4865 if (could_have_src && could_have_dest) { 4866 // This is going to pay off so emit the required guards 4867 if (!has_src) { 4868 src = maybe_cast_profiled_obj(src, src_k, true); 4869 src_type = _gvn.type(src); 4870 top_src = src_type->isa_aryptr(); 4871 has_src = (top_src != NULL && top_src->klass() != NULL); 4872 src_spec = true; 4873 } 4874 if (!has_dest) { 4875 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4876 dest_type = _gvn.type(dest); 4877 top_dest = dest_type->isa_aryptr(); 4878 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4879 dest_spec = true; 4880 } 4881 } 4882 } 4883 4884 if (has_src && has_dest && can_emit_guards) { 4885 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4886 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4887 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4888 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4889 4890 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4891 // If both arrays are object arrays then having the exact types 4892 // for both will remove the need for a subtype check at runtime 4893 // before the call and may make it possible to pick a faster copy 4894 // routine (without a subtype check on every element) 4895 // Do we have the exact type of src? 4896 bool could_have_src = src_spec; 4897 // Do we have the exact type of dest? 4898 bool could_have_dest = dest_spec; 4899 ciKlass* src_k = top_src->klass(); 4900 ciKlass* dest_k = top_dest->klass(); 4901 if (!src_spec) { 4902 src_k = src_type->speculative_type_not_null(); 4903 if (src_k != NULL && src_k->is_array_klass()) { 4904 could_have_src = true; 4905 } 4906 } 4907 if (!dest_spec) { 4908 dest_k = dest_type->speculative_type_not_null(); 4909 if (dest_k != NULL && dest_k->is_array_klass()) { 4910 could_have_dest = true; 4911 } 4912 } 4913 if (could_have_src && could_have_dest) { 4914 // If we can have both exact types, emit the missing guards 4915 if (could_have_src && !src_spec) { 4916 src = maybe_cast_profiled_obj(src, src_k, true); 4917 } 4918 if (could_have_dest && !dest_spec) { 4919 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4920 } 4921 } 4922 } 4923 } 4924 4925 ciMethod* trap_method = method(); 4926 int trap_bci = bci(); 4927 if (saved_jvms != NULL) { 4928 trap_method = alloc->jvms()->method(); 4929 trap_bci = alloc->jvms()->bci(); 4930 } 4931 4932 bool negative_length_guard_generated = false; 4933 4934 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 4935 can_emit_guards && 4936 !src->is_top() && !dest->is_top()) { 4937 // validate arguments: enables transformation the ArrayCopyNode 4938 validated = true; 4939 4940 RegionNode* slow_region = new RegionNode(1); 4941 record_for_igvn(slow_region); 4942 4943 // (1) src and dest are arrays. 4944 generate_non_array_guard(load_object_klass(src), slow_region); 4945 generate_non_array_guard(load_object_klass(dest), slow_region); 4946 4947 // (2) src and dest arrays must have elements of the same BasicType 4948 // done at macro expansion or at Ideal transformation time 4949 4950 // (4) src_offset must not be negative. 4951 generate_negative_guard(src_offset, slow_region); 4952 4953 // (5) dest_offset must not be negative. 4954 generate_negative_guard(dest_offset, slow_region); 4955 4956 // (7) src_offset + length must not exceed length of src. 4957 generate_limit_guard(src_offset, length, 4958 load_array_length(src), 4959 slow_region); 4960 4961 // (8) dest_offset + length must not exceed length of dest. 4962 generate_limit_guard(dest_offset, length, 4963 load_array_length(dest), 4964 slow_region); 4965 4966 // (6) length must not be negative. 4967 // This is also checked in generate_arraycopy() during macro expansion, but 4968 // we also have to check it here for the case where the ArrayCopyNode will 4969 // be eliminated by Escape Analysis. 4970 if (EliminateAllocations) { 4971 generate_negative_guard(length, slow_region); 4972 negative_length_guard_generated = true; 4973 } 4974 4975 // (9) each element of an oop array must be assignable 4976 Node* src_klass = load_object_klass(src); 4977 Node* dest_klass = load_object_klass(dest); 4978 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4979 4980 if (not_subtype_ctrl != top()) { 4981 PreserveJVMState pjvms(this); 4982 set_control(not_subtype_ctrl); 4983 uncommon_trap(Deoptimization::Reason_intrinsic, 4984 Deoptimization::Action_make_not_entrant); 4985 assert(stopped(), "Should be stopped"); 4986 } 4987 4988 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr(); 4989 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass()); 4990 src = _gvn.transform(new CheckCastPPNode(control(), src, toop)); 4991 4992 src_type = _gvn.type(src); 4993 top_src = src_type->isa_aryptr(); 4994 4995 if (top_dest != NULL && 4996 top_dest->elem()->make_oopptr() != NULL && 4997 top_dest->elem()->make_oopptr()->can_be_value_type()) { 4998 generate_valueArray_guard(load_object_klass(dest), slow_region); 4999 } 5000 5001 if (top_src != NULL && 5002 top_src->elem()->make_oopptr() != NULL && 5003 top_src->elem()->make_oopptr()->can_be_value_type()) { 5004 generate_valueArray_guard(load_object_klass(src), slow_region); 5005 } 5006 5007 { 5008 PreserveJVMState pjvms(this); 5009 set_control(_gvn.transform(slow_region)); 5010 uncommon_trap(Deoptimization::Reason_intrinsic, 5011 Deoptimization::Action_make_not_entrant); 5012 assert(stopped(), "Should be stopped"); 5013 } 5014 } 5015 5016 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx); 5017 5018 if (stopped()) { 5019 return true; 5020 } 5021 5022 Node* new_src = access_resolve(src, ACCESS_READ); 5023 Node* new_dest = access_resolve(dest, ACCESS_WRITE); 5024 5025 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, new_src, src_offset, new_dest, dest_offset, length, alloc != NULL, negative_length_guard_generated, 5026 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5027 // so the compiler has a chance to eliminate them: during macro expansion, 5028 // we have to set their control (CastPP nodes are eliminated). 5029 load_object_klass(src), load_object_klass(dest), 5030 load_array_length(src), load_array_length(dest)); 5031 5032 ac->set_arraycopy(validated); 5033 5034 Node* n = _gvn.transform(ac); 5035 if (n == ac) { 5036 ac->connect_outputs(this); 5037 } else { 5038 assert(validated, "shouldn't transform if all arguments not validated"); 5039 set_all_memory(n); 5040 } 5041 clear_upper_avx(); 5042 5043 5044 return true; 5045 } 5046 5047 5048 // Helper function which determines if an arraycopy immediately follows 5049 // an allocation, with no intervening tests or other escapes for the object. 5050 AllocateArrayNode* 5051 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5052 RegionNode* slow_region) { 5053 if (stopped()) return NULL; // no fast path 5054 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5055 5056 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5057 if (alloc == NULL) return NULL; 5058 5059 Node* rawmem = memory(Compile::AliasIdxRaw); 5060 // Is the allocation's memory state untouched? 5061 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5062 // Bail out if there have been raw-memory effects since the allocation. 5063 // (Example: There might have been a call or safepoint.) 5064 return NULL; 5065 } 5066 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5067 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5068 return NULL; 5069 } 5070 5071 // There must be no unexpected observers of this allocation. 5072 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5073 Node* obs = ptr->fast_out(i); 5074 if (obs != this->map()) { 5075 return NULL; 5076 } 5077 } 5078 5079 // This arraycopy must unconditionally follow the allocation of the ptr. 5080 Node* alloc_ctl = ptr->in(0); 5081 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5082 5083 Node* ctl = control(); 5084 while (ctl != alloc_ctl) { 5085 // There may be guards which feed into the slow_region. 5086 // Any other control flow means that we might not get a chance 5087 // to finish initializing the allocated object. 5088 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5089 IfNode* iff = ctl->in(0)->as_If(); 5090 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con); 5091 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5092 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5093 ctl = iff->in(0); // This test feeds the known slow_region. 5094 continue; 5095 } 5096 // One more try: Various low-level checks bottom out in 5097 // uncommon traps. If the debug-info of the trap omits 5098 // any reference to the allocation, as we've already 5099 // observed, then there can be no objection to the trap. 5100 bool found_trap = false; 5101 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5102 Node* obs = not_ctl->fast_out(j); 5103 if (obs->in(0) == not_ctl && obs->is_Call() && 5104 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5105 found_trap = true; break; 5106 } 5107 } 5108 if (found_trap) { 5109 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5110 continue; 5111 } 5112 } 5113 return NULL; 5114 } 5115 5116 // If we get this far, we have an allocation which immediately 5117 // precedes the arraycopy, and we can take over zeroing the new object. 5118 // The arraycopy will finish the initialization, and provide 5119 // a new control state to which we will anchor the destination pointer. 5120 5121 return alloc; 5122 } 5123 5124 //-------------inline_encodeISOArray----------------------------------- 5125 // encode char[] to byte[] in ISO_8859_1 5126 bool LibraryCallKit::inline_encodeISOArray() { 5127 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5128 // no receiver since it is static method 5129 Node *src = argument(0); 5130 Node *src_offset = argument(1); 5131 Node *dst = argument(2); 5132 Node *dst_offset = argument(3); 5133 Node *length = argument(4); 5134 5135 src = must_be_not_null(src, true); 5136 dst = must_be_not_null(dst, true); 5137 5138 src = access_resolve(src, ACCESS_READ); 5139 dst = access_resolve(dst, ACCESS_WRITE); 5140 5141 const Type* src_type = src->Value(&_gvn); 5142 const Type* dst_type = dst->Value(&_gvn); 5143 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5144 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5145 if (top_src == NULL || top_src->klass() == NULL || 5146 top_dest == NULL || top_dest->klass() == NULL) { 5147 // failed array check 5148 return false; 5149 } 5150 5151 // Figure out the size and type of the elements we will be copying. 5152 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5153 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5154 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5155 return false; 5156 } 5157 5158 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5159 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5160 // 'src_start' points to src array + scaled offset 5161 // 'dst_start' points to dst array + scaled offset 5162 5163 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5164 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5165 enc = _gvn.transform(enc); 5166 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5167 set_memory(res_mem, mtype); 5168 set_result(enc); 5169 clear_upper_avx(); 5170 5171 return true; 5172 } 5173 5174 //-------------inline_multiplyToLen----------------------------------- 5175 bool LibraryCallKit::inline_multiplyToLen() { 5176 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5177 5178 address stubAddr = StubRoutines::multiplyToLen(); 5179 if (stubAddr == NULL) { 5180 return false; // Intrinsic's stub is not implemented on this platform 5181 } 5182 const char* stubName = "multiplyToLen"; 5183 5184 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5185 5186 // no receiver because it is a static method 5187 Node* x = argument(0); 5188 Node* xlen = argument(1); 5189 Node* y = argument(2); 5190 Node* ylen = argument(3); 5191 Node* z = argument(4); 5192 5193 x = must_be_not_null(x, true); 5194 y = must_be_not_null(y, true); 5195 5196 x = access_resolve(x, ACCESS_READ); 5197 y = access_resolve(y, ACCESS_READ); 5198 z = access_resolve(z, ACCESS_WRITE); 5199 5200 const Type* x_type = x->Value(&_gvn); 5201 const Type* y_type = y->Value(&_gvn); 5202 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5203 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5204 if (top_x == NULL || top_x->klass() == NULL || 5205 top_y == NULL || top_y->klass() == NULL) { 5206 // failed array check 5207 return false; 5208 } 5209 5210 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5211 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5212 if (x_elem != T_INT || y_elem != T_INT) { 5213 return false; 5214 } 5215 5216 // Set the original stack and the reexecute bit for the interpreter to reexecute 5217 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5218 // on the return from z array allocation in runtime. 5219 { PreserveReexecuteState preexecs(this); 5220 jvms()->set_should_reexecute(true); 5221 5222 Node* x_start = array_element_address(x, intcon(0), x_elem); 5223 Node* y_start = array_element_address(y, intcon(0), y_elem); 5224 // 'x_start' points to x array + scaled xlen 5225 // 'y_start' points to y array + scaled ylen 5226 5227 // Allocate the result array 5228 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5229 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5230 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5231 5232 IdealKit ideal(this); 5233 5234 #define __ ideal. 5235 Node* one = __ ConI(1); 5236 Node* zero = __ ConI(0); 5237 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5238 __ set(need_alloc, zero); 5239 __ set(z_alloc, z); 5240 __ if_then(z, BoolTest::eq, null()); { 5241 __ increment (need_alloc, one); 5242 } __ else_(); { 5243 // Update graphKit memory and control from IdealKit. 5244 sync_kit(ideal); 5245 Node *cast = new CastPPNode(z, TypePtr::NOTNULL); 5246 cast->init_req(0, control()); 5247 _gvn.set_type(cast, cast->bottom_type()); 5248 C->record_for_igvn(cast); 5249 5250 Node* zlen_arg = load_array_length(cast); 5251 // Update IdealKit memory and control from graphKit. 5252 __ sync_kit(this); 5253 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5254 __ increment (need_alloc, one); 5255 } __ end_if(); 5256 } __ end_if(); 5257 5258 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5259 // Update graphKit memory and control from IdealKit. 5260 sync_kit(ideal); 5261 Node * narr = new_array(klass_node, zlen, 1); 5262 // Update IdealKit memory and control from graphKit. 5263 __ sync_kit(this); 5264 __ set(z_alloc, narr); 5265 } __ end_if(); 5266 5267 sync_kit(ideal); 5268 z = __ value(z_alloc); 5269 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5270 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5271 // Final sync IdealKit and GraphKit. 5272 final_sync(ideal); 5273 #undef __ 5274 5275 Node* z_start = array_element_address(z, intcon(0), T_INT); 5276 5277 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5278 OptoRuntime::multiplyToLen_Type(), 5279 stubAddr, stubName, TypePtr::BOTTOM, 5280 x_start, xlen, y_start, ylen, z_start, zlen); 5281 } // original reexecute is set back here 5282 5283 C->set_has_split_ifs(true); // Has chance for split-if optimization 5284 set_result(z); 5285 return true; 5286 } 5287 5288 //-------------inline_squareToLen------------------------------------ 5289 bool LibraryCallKit::inline_squareToLen() { 5290 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5291 5292 address stubAddr = StubRoutines::squareToLen(); 5293 if (stubAddr == NULL) { 5294 return false; // Intrinsic's stub is not implemented on this platform 5295 } 5296 const char* stubName = "squareToLen"; 5297 5298 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5299 5300 Node* x = argument(0); 5301 Node* len = argument(1); 5302 Node* z = argument(2); 5303 Node* zlen = argument(3); 5304 5305 x = must_be_not_null(x, true); 5306 z = must_be_not_null(z, true); 5307 5308 x = access_resolve(x, ACCESS_READ); 5309 z = access_resolve(z, ACCESS_WRITE); 5310 5311 const Type* x_type = x->Value(&_gvn); 5312 const Type* z_type = z->Value(&_gvn); 5313 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5314 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5315 if (top_x == NULL || top_x->klass() == NULL || 5316 top_z == NULL || top_z->klass() == NULL) { 5317 // failed array check 5318 return false; 5319 } 5320 5321 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5322 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5323 if (x_elem != T_INT || z_elem != T_INT) { 5324 return false; 5325 } 5326 5327 5328 Node* x_start = array_element_address(x, intcon(0), x_elem); 5329 Node* z_start = array_element_address(z, intcon(0), z_elem); 5330 5331 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5332 OptoRuntime::squareToLen_Type(), 5333 stubAddr, stubName, TypePtr::BOTTOM, 5334 x_start, len, z_start, zlen); 5335 5336 set_result(z); 5337 return true; 5338 } 5339 5340 //-------------inline_mulAdd------------------------------------------ 5341 bool LibraryCallKit::inline_mulAdd() { 5342 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5343 5344 address stubAddr = StubRoutines::mulAdd(); 5345 if (stubAddr == NULL) { 5346 return false; // Intrinsic's stub is not implemented on this platform 5347 } 5348 const char* stubName = "mulAdd"; 5349 5350 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5351 5352 Node* out = argument(0); 5353 Node* in = argument(1); 5354 Node* offset = argument(2); 5355 Node* len = argument(3); 5356 Node* k = argument(4); 5357 5358 out = must_be_not_null(out, true); 5359 5360 in = access_resolve(in, ACCESS_READ); 5361 out = access_resolve(out, ACCESS_WRITE); 5362 5363 const Type* out_type = out->Value(&_gvn); 5364 const Type* in_type = in->Value(&_gvn); 5365 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5366 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5367 if (top_out == NULL || top_out->klass() == NULL || 5368 top_in == NULL || top_in->klass() == NULL) { 5369 // failed array check 5370 return false; 5371 } 5372 5373 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5374 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5375 if (out_elem != T_INT || in_elem != T_INT) { 5376 return false; 5377 } 5378 5379 Node* outlen = load_array_length(out); 5380 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5381 Node* out_start = array_element_address(out, intcon(0), out_elem); 5382 Node* in_start = array_element_address(in, intcon(0), in_elem); 5383 5384 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5385 OptoRuntime::mulAdd_Type(), 5386 stubAddr, stubName, TypePtr::BOTTOM, 5387 out_start,in_start, new_offset, len, k); 5388 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5389 set_result(result); 5390 return true; 5391 } 5392 5393 //-------------inline_montgomeryMultiply----------------------------------- 5394 bool LibraryCallKit::inline_montgomeryMultiply() { 5395 address stubAddr = StubRoutines::montgomeryMultiply(); 5396 if (stubAddr == NULL) { 5397 return false; // Intrinsic's stub is not implemented on this platform 5398 } 5399 5400 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5401 const char* stubName = "montgomery_multiply"; 5402 5403 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5404 5405 Node* a = argument(0); 5406 Node* b = argument(1); 5407 Node* n = argument(2); 5408 Node* len = argument(3); 5409 Node* inv = argument(4); 5410 Node* m = argument(6); 5411 5412 a = access_resolve(a, ACCESS_READ); 5413 b = access_resolve(b, ACCESS_READ); 5414 n = access_resolve(n, ACCESS_READ); 5415 m = access_resolve(m, ACCESS_WRITE); 5416 5417 const Type* a_type = a->Value(&_gvn); 5418 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5419 const Type* b_type = b->Value(&_gvn); 5420 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5421 const Type* n_type = a->Value(&_gvn); 5422 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5423 const Type* m_type = a->Value(&_gvn); 5424 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5425 if (top_a == NULL || top_a->klass() == NULL || 5426 top_b == NULL || top_b->klass() == NULL || 5427 top_n == NULL || top_n->klass() == NULL || 5428 top_m == NULL || top_m->klass() == NULL) { 5429 // failed array check 5430 return false; 5431 } 5432 5433 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5434 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5435 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5436 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5437 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5438 return false; 5439 } 5440 5441 // Make the call 5442 { 5443 Node* a_start = array_element_address(a, intcon(0), a_elem); 5444 Node* b_start = array_element_address(b, intcon(0), b_elem); 5445 Node* n_start = array_element_address(n, intcon(0), n_elem); 5446 Node* m_start = array_element_address(m, intcon(0), m_elem); 5447 5448 Node* call = make_runtime_call(RC_LEAF, 5449 OptoRuntime::montgomeryMultiply_Type(), 5450 stubAddr, stubName, TypePtr::BOTTOM, 5451 a_start, b_start, n_start, len, inv, top(), 5452 m_start); 5453 set_result(m); 5454 } 5455 5456 return true; 5457 } 5458 5459 bool LibraryCallKit::inline_montgomerySquare() { 5460 address stubAddr = StubRoutines::montgomerySquare(); 5461 if (stubAddr == NULL) { 5462 return false; // Intrinsic's stub is not implemented on this platform 5463 } 5464 5465 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5466 const char* stubName = "montgomery_square"; 5467 5468 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5469 5470 Node* a = argument(0); 5471 Node* n = argument(1); 5472 Node* len = argument(2); 5473 Node* inv = argument(3); 5474 Node* m = argument(5); 5475 5476 a = access_resolve(a, ACCESS_READ); 5477 n = access_resolve(n, ACCESS_READ); 5478 m = access_resolve(m, ACCESS_WRITE); 5479 5480 const Type* a_type = a->Value(&_gvn); 5481 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5482 const Type* n_type = a->Value(&_gvn); 5483 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5484 const Type* m_type = a->Value(&_gvn); 5485 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5486 if (top_a == NULL || top_a->klass() == NULL || 5487 top_n == NULL || top_n->klass() == NULL || 5488 top_m == NULL || top_m->klass() == NULL) { 5489 // failed array check 5490 return false; 5491 } 5492 5493 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5494 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5495 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5496 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5497 return false; 5498 } 5499 5500 // Make the call 5501 { 5502 Node* a_start = array_element_address(a, intcon(0), a_elem); 5503 Node* n_start = array_element_address(n, intcon(0), n_elem); 5504 Node* m_start = array_element_address(m, intcon(0), m_elem); 5505 5506 Node* call = make_runtime_call(RC_LEAF, 5507 OptoRuntime::montgomerySquare_Type(), 5508 stubAddr, stubName, TypePtr::BOTTOM, 5509 a_start, n_start, len, inv, top(), 5510 m_start); 5511 set_result(m); 5512 } 5513 5514 return true; 5515 } 5516 5517 //-------------inline_vectorizedMismatch------------------------------ 5518 bool LibraryCallKit::inline_vectorizedMismatch() { 5519 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5520 5521 address stubAddr = StubRoutines::vectorizedMismatch(); 5522 if (stubAddr == NULL) { 5523 return false; // Intrinsic's stub is not implemented on this platform 5524 } 5525 const char* stubName = "vectorizedMismatch"; 5526 int size_l = callee()->signature()->size(); 5527 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5528 5529 Node* obja = argument(0); 5530 Node* aoffset = argument(1); 5531 Node* objb = argument(3); 5532 Node* boffset = argument(4); 5533 Node* length = argument(6); 5534 Node* scale = argument(7); 5535 5536 const Type* a_type = obja->Value(&_gvn); 5537 const Type* b_type = objb->Value(&_gvn); 5538 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5539 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5540 if (top_a == NULL || top_a->klass() == NULL || 5541 top_b == NULL || top_b->klass() == NULL) { 5542 // failed array check 5543 return false; 5544 } 5545 5546 Node* call; 5547 jvms()->set_should_reexecute(true); 5548 5549 obja = access_resolve(obja, ACCESS_READ); 5550 objb = access_resolve(objb, ACCESS_READ); 5551 Node* obja_adr = make_unsafe_address(obja, aoffset, ACCESS_READ); 5552 Node* objb_adr = make_unsafe_address(objb, boffset, ACCESS_READ); 5553 5554 call = make_runtime_call(RC_LEAF, 5555 OptoRuntime::vectorizedMismatch_Type(), 5556 stubAddr, stubName, TypePtr::BOTTOM, 5557 obja_adr, objb_adr, length, scale); 5558 5559 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5560 set_result(result); 5561 return true; 5562 } 5563 5564 /** 5565 * Calculate CRC32 for byte. 5566 * int java.util.zip.CRC32.update(int crc, int b) 5567 */ 5568 bool LibraryCallKit::inline_updateCRC32() { 5569 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5570 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5571 // no receiver since it is static method 5572 Node* crc = argument(0); // type: int 5573 Node* b = argument(1); // type: int 5574 5575 /* 5576 * int c = ~ crc; 5577 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5578 * b = b ^ (c >>> 8); 5579 * crc = ~b; 5580 */ 5581 5582 Node* M1 = intcon(-1); 5583 crc = _gvn.transform(new XorINode(crc, M1)); 5584 Node* result = _gvn.transform(new XorINode(crc, b)); 5585 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5586 5587 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5588 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5589 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5590 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5591 5592 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5593 result = _gvn.transform(new XorINode(crc, result)); 5594 result = _gvn.transform(new XorINode(result, M1)); 5595 set_result(result); 5596 return true; 5597 } 5598 5599 /** 5600 * Calculate CRC32 for byte[] array. 5601 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5602 */ 5603 bool LibraryCallKit::inline_updateBytesCRC32() { 5604 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5605 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5606 // no receiver since it is static method 5607 Node* crc = argument(0); // type: int 5608 Node* src = argument(1); // type: oop 5609 Node* offset = argument(2); // type: int 5610 Node* length = argument(3); // type: int 5611 5612 const Type* src_type = src->Value(&_gvn); 5613 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5614 if (top_src == NULL || top_src->klass() == NULL) { 5615 // failed array check 5616 return false; 5617 } 5618 5619 // Figure out the size and type of the elements we will be copying. 5620 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5621 if (src_elem != T_BYTE) { 5622 return false; 5623 } 5624 5625 // 'src_start' points to src array + scaled offset 5626 src = must_be_not_null(src, true); 5627 src = access_resolve(src, ACCESS_READ); 5628 Node* src_start = array_element_address(src, offset, src_elem); 5629 5630 // We assume that range check is done by caller. 5631 // TODO: generate range check (offset+length < src.length) in debug VM. 5632 5633 // Call the stub. 5634 address stubAddr = StubRoutines::updateBytesCRC32(); 5635 const char *stubName = "updateBytesCRC32"; 5636 5637 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5638 stubAddr, stubName, TypePtr::BOTTOM, 5639 crc, src_start, length); 5640 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5641 set_result(result); 5642 return true; 5643 } 5644 5645 /** 5646 * Calculate CRC32 for ByteBuffer. 5647 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5648 */ 5649 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5650 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5651 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5652 // no receiver since it is static method 5653 Node* crc = argument(0); // type: int 5654 Node* src = argument(1); // type: long 5655 Node* offset = argument(3); // type: int 5656 Node* length = argument(4); // type: int 5657 5658 src = ConvL2X(src); // adjust Java long to machine word 5659 Node* base = _gvn.transform(new CastX2PNode(src)); 5660 offset = ConvI2X(offset); 5661 5662 // 'src_start' points to src array + scaled offset 5663 Node* src_start = basic_plus_adr(top(), base, offset); 5664 5665 // Call the stub. 5666 address stubAddr = StubRoutines::updateBytesCRC32(); 5667 const char *stubName = "updateBytesCRC32"; 5668 5669 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5670 stubAddr, stubName, TypePtr::BOTTOM, 5671 crc, src_start, length); 5672 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5673 set_result(result); 5674 return true; 5675 } 5676 5677 //------------------------------get_table_from_crc32c_class----------------------- 5678 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5679 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5680 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5681 5682 return table; 5683 } 5684 5685 //------------------------------inline_updateBytesCRC32C----------------------- 5686 // 5687 // Calculate CRC32C for byte[] array. 5688 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5689 // 5690 bool LibraryCallKit::inline_updateBytesCRC32C() { 5691 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5692 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5693 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5694 // no receiver since it is a static method 5695 Node* crc = argument(0); // type: int 5696 Node* src = argument(1); // type: oop 5697 Node* offset = argument(2); // type: int 5698 Node* end = argument(3); // type: int 5699 5700 Node* length = _gvn.transform(new SubINode(end, offset)); 5701 5702 const Type* src_type = src->Value(&_gvn); 5703 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5704 if (top_src == NULL || top_src->klass() == NULL) { 5705 // failed array check 5706 return false; 5707 } 5708 5709 // Figure out the size and type of the elements we will be copying. 5710 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5711 if (src_elem != T_BYTE) { 5712 return false; 5713 } 5714 5715 // 'src_start' points to src array + scaled offset 5716 src = must_be_not_null(src, true); 5717 src = access_resolve(src, ACCESS_READ); 5718 Node* src_start = array_element_address(src, offset, src_elem); 5719 5720 // static final int[] byteTable in class CRC32C 5721 Node* table = get_table_from_crc32c_class(callee()->holder()); 5722 table = must_be_not_null(table, true); 5723 table = access_resolve(table, ACCESS_READ); 5724 Node* table_start = array_element_address(table, intcon(0), T_INT); 5725 5726 // We assume that range check is done by caller. 5727 // TODO: generate range check (offset+length < src.length) in debug VM. 5728 5729 // Call the stub. 5730 address stubAddr = StubRoutines::updateBytesCRC32C(); 5731 const char *stubName = "updateBytesCRC32C"; 5732 5733 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5734 stubAddr, stubName, TypePtr::BOTTOM, 5735 crc, src_start, length, table_start); 5736 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5737 set_result(result); 5738 return true; 5739 } 5740 5741 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5742 // 5743 // Calculate CRC32C for DirectByteBuffer. 5744 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5745 // 5746 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5747 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5748 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5749 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5750 // no receiver since it is a static method 5751 Node* crc = argument(0); // type: int 5752 Node* src = argument(1); // type: long 5753 Node* offset = argument(3); // type: int 5754 Node* end = argument(4); // type: int 5755 5756 Node* length = _gvn.transform(new SubINode(end, offset)); 5757 5758 src = ConvL2X(src); // adjust Java long to machine word 5759 Node* base = _gvn.transform(new CastX2PNode(src)); 5760 offset = ConvI2X(offset); 5761 5762 // 'src_start' points to src array + scaled offset 5763 Node* src_start = basic_plus_adr(top(), base, offset); 5764 5765 // static final int[] byteTable in class CRC32C 5766 Node* table = get_table_from_crc32c_class(callee()->holder()); 5767 table = must_be_not_null(table, true); 5768 table = access_resolve(table, ACCESS_READ); 5769 Node* table_start = array_element_address(table, intcon(0), T_INT); 5770 5771 // Call the stub. 5772 address stubAddr = StubRoutines::updateBytesCRC32C(); 5773 const char *stubName = "updateBytesCRC32C"; 5774 5775 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5776 stubAddr, stubName, TypePtr::BOTTOM, 5777 crc, src_start, length, table_start); 5778 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5779 set_result(result); 5780 return true; 5781 } 5782 5783 //------------------------------inline_updateBytesAdler32---------------------- 5784 // 5785 // Calculate Adler32 checksum for byte[] array. 5786 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5787 // 5788 bool LibraryCallKit::inline_updateBytesAdler32() { 5789 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5790 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5791 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5792 // no receiver since it is static method 5793 Node* crc = argument(0); // type: int 5794 Node* src = argument(1); // type: oop 5795 Node* offset = argument(2); // type: int 5796 Node* length = argument(3); // type: int 5797 5798 const Type* src_type = src->Value(&_gvn); 5799 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5800 if (top_src == NULL || top_src->klass() == NULL) { 5801 // failed array check 5802 return false; 5803 } 5804 5805 // Figure out the size and type of the elements we will be copying. 5806 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5807 if (src_elem != T_BYTE) { 5808 return false; 5809 } 5810 5811 // 'src_start' points to src array + scaled offset 5812 src = access_resolve(src, ACCESS_READ); 5813 Node* src_start = array_element_address(src, offset, src_elem); 5814 5815 // We assume that range check is done by caller. 5816 // TODO: generate range check (offset+length < src.length) in debug VM. 5817 5818 // Call the stub. 5819 address stubAddr = StubRoutines::updateBytesAdler32(); 5820 const char *stubName = "updateBytesAdler32"; 5821 5822 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5823 stubAddr, stubName, TypePtr::BOTTOM, 5824 crc, src_start, length); 5825 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5826 set_result(result); 5827 return true; 5828 } 5829 5830 //------------------------------inline_updateByteBufferAdler32--------------- 5831 // 5832 // Calculate Adler32 checksum for DirectByteBuffer. 5833 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5834 // 5835 bool LibraryCallKit::inline_updateByteBufferAdler32() { 5836 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5837 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5838 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5839 // no receiver since it is static method 5840 Node* crc = argument(0); // type: int 5841 Node* src = argument(1); // type: long 5842 Node* offset = argument(3); // type: int 5843 Node* length = argument(4); // type: int 5844 5845 src = ConvL2X(src); // adjust Java long to machine word 5846 Node* base = _gvn.transform(new CastX2PNode(src)); 5847 offset = ConvI2X(offset); 5848 5849 // 'src_start' points to src array + scaled offset 5850 Node* src_start = basic_plus_adr(top(), base, offset); 5851 5852 // Call the stub. 5853 address stubAddr = StubRoutines::updateBytesAdler32(); 5854 const char *stubName = "updateBytesAdler32"; 5855 5856 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5857 stubAddr, stubName, TypePtr::BOTTOM, 5858 crc, src_start, length); 5859 5860 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5861 set_result(result); 5862 return true; 5863 } 5864 5865 //----------------------------inline_reference_get---------------------------- 5866 // public T java.lang.ref.Reference.get(); 5867 bool LibraryCallKit::inline_reference_get() { 5868 const int referent_offset = java_lang_ref_Reference::referent_offset; 5869 guarantee(referent_offset > 0, "should have already been set"); 5870 5871 // Get the argument: 5872 Node* reference_obj = null_check_receiver(); 5873 if (stopped()) return true; 5874 5875 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr(); 5876 assert(tinst != NULL, "obj is null"); 5877 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5878 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass(); 5879 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"), 5880 ciSymbol::make("Ljava/lang/Object;"), 5881 false); 5882 assert (field != NULL, "undefined field"); 5883 5884 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5885 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5886 5887 ciInstanceKlass* klass = env()->Object_klass(); 5888 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5889 5890 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF; 5891 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators); 5892 // Add memory barrier to prevent commoning reads from this field 5893 // across safepoint since GC can change its value. 5894 insert_mem_bar(Op_MemBarCPUOrder); 5895 5896 set_result(result); 5897 return true; 5898 } 5899 5900 5901 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5902 bool is_exact=true, bool is_static=false, 5903 ciInstanceKlass * fromKls=NULL) { 5904 if (fromKls == NULL) { 5905 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5906 assert(tinst != NULL, "obj is null"); 5907 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5908 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5909 fromKls = tinst->klass()->as_instance_klass(); 5910 } else { 5911 assert(is_static, "only for static field access"); 5912 } 5913 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5914 ciSymbol::make(fieldTypeString), 5915 is_static); 5916 5917 assert (field != NULL, "undefined field"); 5918 if (field == NULL) return (Node *) NULL; 5919 5920 if (is_static) { 5921 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5922 fromObj = makecon(tip); 5923 } 5924 5925 // Next code copied from Parse::do_get_xxx(): 5926 5927 // Compute address and memory type. 5928 int offset = field->offset_in_bytes(); 5929 bool is_vol = field->is_volatile(); 5930 ciType* field_klass = field->type(); 5931 assert(field_klass->is_loaded(), "should be loaded"); 5932 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5933 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5934 BasicType bt = field->layout_type(); 5935 5936 // Build the resultant type of the load 5937 const Type *type; 5938 if (bt == T_OBJECT) { 5939 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5940 } else { 5941 type = Type::get_const_basic_type(bt); 5942 } 5943 5944 DecoratorSet decorators = IN_HEAP; 5945 5946 if (is_vol) { 5947 decorators |= MO_SEQ_CST; 5948 } 5949 5950 return access_load_at(fromObj, adr, adr_type, type, bt, decorators); 5951 } 5952 5953 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5954 bool is_exact = true, bool is_static = false, 5955 ciInstanceKlass * fromKls = NULL) { 5956 if (fromKls == NULL) { 5957 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5958 assert(tinst != NULL, "obj is null"); 5959 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5960 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5961 fromKls = tinst->klass()->as_instance_klass(); 5962 } 5963 else { 5964 assert(is_static, "only for static field access"); 5965 } 5966 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5967 ciSymbol::make(fieldTypeString), 5968 is_static); 5969 5970 assert(field != NULL, "undefined field"); 5971 assert(!field->is_volatile(), "not defined for volatile fields"); 5972 5973 if (is_static) { 5974 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5975 fromObj = makecon(tip); 5976 } 5977 5978 // Next code copied from Parse::do_get_xxx(): 5979 5980 // Compute address and memory type. 5981 int offset = field->offset_in_bytes(); 5982 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5983 5984 return adr; 5985 } 5986 5987 //------------------------------inline_aescrypt_Block----------------------- 5988 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5989 address stubAddr = NULL; 5990 const char *stubName; 5991 assert(UseAES, "need AES instruction support"); 5992 5993 switch(id) { 5994 case vmIntrinsics::_aescrypt_encryptBlock: 5995 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5996 stubName = "aescrypt_encryptBlock"; 5997 break; 5998 case vmIntrinsics::_aescrypt_decryptBlock: 5999 stubAddr = StubRoutines::aescrypt_decryptBlock(); 6000 stubName = "aescrypt_decryptBlock"; 6001 break; 6002 default: 6003 break; 6004 } 6005 if (stubAddr == NULL) return false; 6006 6007 Node* aescrypt_object = argument(0); 6008 Node* src = argument(1); 6009 Node* src_offset = argument(2); 6010 Node* dest = argument(3); 6011 Node* dest_offset = argument(4); 6012 6013 src = must_be_not_null(src, true); 6014 dest = must_be_not_null(dest, true); 6015 6016 src = access_resolve(src, ACCESS_READ); 6017 dest = access_resolve(dest, ACCESS_WRITE); 6018 6019 // (1) src and dest are arrays. 6020 const Type* src_type = src->Value(&_gvn); 6021 const Type* dest_type = dest->Value(&_gvn); 6022 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6023 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6024 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6025 6026 // for the quick and dirty code we will skip all the checks. 6027 // we are just trying to get the call to be generated. 6028 Node* src_start = src; 6029 Node* dest_start = dest; 6030 if (src_offset != NULL || dest_offset != NULL) { 6031 assert(src_offset != NULL && dest_offset != NULL, ""); 6032 src_start = array_element_address(src, src_offset, T_BYTE); 6033 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6034 } 6035 6036 // now need to get the start of its expanded key array 6037 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6038 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6039 if (k_start == NULL) return false; 6040 6041 if (Matcher::pass_original_key_for_aes()) { 6042 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6043 // compatibility issues between Java key expansion and SPARC crypto instructions 6044 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6045 if (original_k_start == NULL) return false; 6046 6047 // Call the stub. 6048 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6049 stubAddr, stubName, TypePtr::BOTTOM, 6050 src_start, dest_start, k_start, original_k_start); 6051 } else { 6052 // Call the stub. 6053 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6054 stubAddr, stubName, TypePtr::BOTTOM, 6055 src_start, dest_start, k_start); 6056 } 6057 6058 return true; 6059 } 6060 6061 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 6062 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 6063 address stubAddr = NULL; 6064 const char *stubName = NULL; 6065 6066 assert(UseAES, "need AES instruction support"); 6067 6068 switch(id) { 6069 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6070 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6071 stubName = "cipherBlockChaining_encryptAESCrypt"; 6072 break; 6073 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6074 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6075 stubName = "cipherBlockChaining_decryptAESCrypt"; 6076 break; 6077 default: 6078 break; 6079 } 6080 if (stubAddr == NULL) return false; 6081 6082 Node* cipherBlockChaining_object = argument(0); 6083 Node* src = argument(1); 6084 Node* src_offset = argument(2); 6085 Node* len = argument(3); 6086 Node* dest = argument(4); 6087 Node* dest_offset = argument(5); 6088 6089 src = must_be_not_null(src, false); 6090 dest = must_be_not_null(dest, false); 6091 6092 src = access_resolve(src, ACCESS_READ); 6093 dest = access_resolve(dest, ACCESS_WRITE); 6094 6095 // (1) src and dest are arrays. 6096 const Type* src_type = src->Value(&_gvn); 6097 const Type* dest_type = dest->Value(&_gvn); 6098 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6099 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6100 assert (top_src != NULL && top_src->klass() != NULL 6101 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6102 6103 // checks are the responsibility of the caller 6104 Node* src_start = src; 6105 Node* dest_start = dest; 6106 if (src_offset != NULL || dest_offset != NULL) { 6107 assert(src_offset != NULL && dest_offset != NULL, ""); 6108 src_start = array_element_address(src, src_offset, T_BYTE); 6109 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6110 } 6111 6112 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6113 // (because of the predicated logic executed earlier). 6114 // so we cast it here safely. 6115 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6116 6117 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6118 if (embeddedCipherObj == NULL) return false; 6119 6120 // cast it to what we know it will be at runtime 6121 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6122 assert(tinst != NULL, "CBC obj is null"); 6123 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6124 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6125 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6126 6127 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6128 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6129 const TypeOopPtr* xtype = aklass->as_instance_type(); 6130 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6131 aescrypt_object = _gvn.transform(aescrypt_object); 6132 6133 // we need to get the start of the aescrypt_object's expanded key array 6134 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6135 if (k_start == NULL) return false; 6136 6137 // similarly, get the start address of the r vector 6138 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6139 if (objRvec == NULL) return false; 6140 objRvec = access_resolve(objRvec, ACCESS_WRITE); 6141 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6142 6143 Node* cbcCrypt; 6144 if (Matcher::pass_original_key_for_aes()) { 6145 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6146 // compatibility issues between Java key expansion and SPARC crypto instructions 6147 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6148 if (original_k_start == NULL) return false; 6149 6150 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6151 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6152 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6153 stubAddr, stubName, TypePtr::BOTTOM, 6154 src_start, dest_start, k_start, r_start, len, original_k_start); 6155 } else { 6156 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6157 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6158 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6159 stubAddr, stubName, TypePtr::BOTTOM, 6160 src_start, dest_start, k_start, r_start, len); 6161 } 6162 6163 // return cipher length (int) 6164 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6165 set_result(retvalue); 6166 return true; 6167 } 6168 6169 //------------------------------inline_counterMode_AESCrypt----------------------- 6170 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6171 assert(UseAES, "need AES instruction support"); 6172 if (!UseAESCTRIntrinsics) return false; 6173 6174 address stubAddr = NULL; 6175 const char *stubName = NULL; 6176 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6177 stubAddr = StubRoutines::counterMode_AESCrypt(); 6178 stubName = "counterMode_AESCrypt"; 6179 } 6180 if (stubAddr == NULL) return false; 6181 6182 Node* counterMode_object = argument(0); 6183 Node* src = argument(1); 6184 Node* src_offset = argument(2); 6185 Node* len = argument(3); 6186 Node* dest = argument(4); 6187 Node* dest_offset = argument(5); 6188 6189 src = access_resolve(src, ACCESS_READ); 6190 dest = access_resolve(dest, ACCESS_WRITE); 6191 counterMode_object = access_resolve(counterMode_object, ACCESS_WRITE); 6192 6193 // (1) src and dest are arrays. 6194 const Type* src_type = src->Value(&_gvn); 6195 const Type* dest_type = dest->Value(&_gvn); 6196 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6197 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6198 assert(top_src != NULL && top_src->klass() != NULL && 6199 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6200 6201 // checks are the responsibility of the caller 6202 Node* src_start = src; 6203 Node* dest_start = dest; 6204 if (src_offset != NULL || dest_offset != NULL) { 6205 assert(src_offset != NULL && dest_offset != NULL, ""); 6206 src_start = array_element_address(src, src_offset, T_BYTE); 6207 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6208 } 6209 6210 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6211 // (because of the predicated logic executed earlier). 6212 // so we cast it here safely. 6213 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6214 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6215 if (embeddedCipherObj == NULL) return false; 6216 // cast it to what we know it will be at runtime 6217 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6218 assert(tinst != NULL, "CTR obj is null"); 6219 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6220 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6221 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6222 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6223 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6224 const TypeOopPtr* xtype = aklass->as_instance_type(); 6225 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6226 aescrypt_object = _gvn.transform(aescrypt_object); 6227 // we need to get the start of the aescrypt_object's expanded key array 6228 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6229 if (k_start == NULL) return false; 6230 // similarly, get the start address of the r vector 6231 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6232 if (obj_counter == NULL) return false; 6233 obj_counter = access_resolve(obj_counter, ACCESS_WRITE); 6234 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6235 6236 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6237 if (saved_encCounter == NULL) return false; 6238 saved_encCounter = access_resolve(saved_encCounter, ACCESS_WRITE); 6239 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6240 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6241 6242 Node* ctrCrypt; 6243 if (Matcher::pass_original_key_for_aes()) { 6244 // no SPARC version for AES/CTR intrinsics now. 6245 return false; 6246 } 6247 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6248 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6249 OptoRuntime::counterMode_aescrypt_Type(), 6250 stubAddr, stubName, TypePtr::BOTTOM, 6251 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6252 6253 // return cipher length (int) 6254 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6255 set_result(retvalue); 6256 return true; 6257 } 6258 6259 //------------------------------get_key_start_from_aescrypt_object----------------------- 6260 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6261 #if defined(PPC64) || defined(S390) 6262 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6263 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6264 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6265 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6266 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6267 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6268 if (objSessionK == NULL) { 6269 return (Node *) NULL; 6270 } 6271 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6272 #else 6273 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6274 #endif // PPC64 6275 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6276 if (objAESCryptKey == NULL) return (Node *) NULL; 6277 6278 // now have the array, need to get the start address of the K array 6279 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ); 6280 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6281 return k_start; 6282 } 6283 6284 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6285 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6286 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6287 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6288 if (objAESCryptKey == NULL) return (Node *) NULL; 6289 6290 // now have the array, need to get the start address of the lastKey array 6291 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ); 6292 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6293 return original_k_start; 6294 } 6295 6296 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6297 // Return node representing slow path of predicate check. 6298 // the pseudo code we want to emulate with this predicate is: 6299 // for encryption: 6300 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6301 // for decryption: 6302 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6303 // note cipher==plain is more conservative than the original java code but that's OK 6304 // 6305 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6306 // The receiver was checked for NULL already. 6307 Node* objCBC = argument(0); 6308 6309 Node* src = argument(1); 6310 Node* dest = argument(4); 6311 6312 // Load embeddedCipher field of CipherBlockChaining object. 6313 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6314 6315 // get AESCrypt klass for instanceOf check 6316 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6317 // will have same classloader as CipherBlockChaining object 6318 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6319 assert(tinst != NULL, "CBCobj is null"); 6320 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6321 6322 // we want to do an instanceof comparison against the AESCrypt class 6323 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6324 if (!klass_AESCrypt->is_loaded()) { 6325 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6326 Node* ctrl = control(); 6327 set_control(top()); // no regular fast path 6328 return ctrl; 6329 } 6330 6331 src = must_be_not_null(src, true); 6332 dest = must_be_not_null(dest, true); 6333 6334 // Resolve oops to stable for CmpP below. 6335 src = access_resolve(src, 0); 6336 dest = access_resolve(dest, 0); 6337 6338 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6339 6340 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6341 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6342 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6343 6344 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6345 6346 // for encryption, we are done 6347 if (!decrypting) 6348 return instof_false; // even if it is NULL 6349 6350 // for decryption, we need to add a further check to avoid 6351 // taking the intrinsic path when cipher and plain are the same 6352 // see the original java code for why. 6353 RegionNode* region = new RegionNode(3); 6354 region->init_req(1, instof_false); 6355 6356 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6357 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6358 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6359 region->init_req(2, src_dest_conjoint); 6360 6361 record_for_igvn(region); 6362 return _gvn.transform(region); 6363 } 6364 6365 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6366 // Return node representing slow path of predicate check. 6367 // the pseudo code we want to emulate with this predicate is: 6368 // for encryption: 6369 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6370 // for decryption: 6371 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6372 // note cipher==plain is more conservative than the original java code but that's OK 6373 // 6374 6375 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6376 // The receiver was checked for NULL already. 6377 Node* objCTR = argument(0); 6378 6379 // Load embeddedCipher field of CipherBlockChaining object. 6380 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6381 6382 // get AESCrypt klass for instanceOf check 6383 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6384 // will have same classloader as CipherBlockChaining object 6385 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6386 assert(tinst != NULL, "CTRobj is null"); 6387 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6388 6389 // we want to do an instanceof comparison against the AESCrypt class 6390 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6391 if (!klass_AESCrypt->is_loaded()) { 6392 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6393 Node* ctrl = control(); 6394 set_control(top()); // no regular fast path 6395 return ctrl; 6396 } 6397 6398 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6399 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6400 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6401 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6402 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6403 6404 return instof_false; // even if it is NULL 6405 } 6406 6407 //------------------------------inline_ghash_processBlocks 6408 bool LibraryCallKit::inline_ghash_processBlocks() { 6409 address stubAddr; 6410 const char *stubName; 6411 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6412 6413 stubAddr = StubRoutines::ghash_processBlocks(); 6414 stubName = "ghash_processBlocks"; 6415 6416 Node* data = argument(0); 6417 Node* offset = argument(1); 6418 Node* len = argument(2); 6419 Node* state = argument(3); 6420 Node* subkeyH = argument(4); 6421 6422 state = must_be_not_null(state, true); 6423 subkeyH = must_be_not_null(subkeyH, true); 6424 data = must_be_not_null(data, true); 6425 6426 state = access_resolve(state, ACCESS_WRITE); 6427 subkeyH = access_resolve(subkeyH, ACCESS_READ); 6428 data = access_resolve(data, ACCESS_READ); 6429 6430 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6431 assert(state_start, "state is NULL"); 6432 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6433 assert(subkeyH_start, "subkeyH is NULL"); 6434 Node* data_start = array_element_address(data, offset, T_BYTE); 6435 assert(data_start, "data is NULL"); 6436 6437 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6438 OptoRuntime::ghash_processBlocks_Type(), 6439 stubAddr, stubName, TypePtr::BOTTOM, 6440 state_start, subkeyH_start, data_start, len); 6441 return true; 6442 } 6443 6444 bool LibraryCallKit::inline_base64_encodeBlock() { 6445 address stubAddr; 6446 const char *stubName; 6447 assert(UseBASE64Intrinsics, "need Base64 intrinsics support"); 6448 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters"); 6449 stubAddr = StubRoutines::base64_encodeBlock(); 6450 stubName = "encodeBlock"; 6451 6452 if (!stubAddr) return false; 6453 Node* base64obj = argument(0); 6454 Node* src = argument(1); 6455 Node* offset = argument(2); 6456 Node* len = argument(3); 6457 Node* dest = argument(4); 6458 Node* dp = argument(5); 6459 Node* isURL = argument(6); 6460 6461 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 6462 assert(src_start, "source array is NULL"); 6463 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE); 6464 assert(dest_start, "destination array is NULL"); 6465 6466 Node* base64 = make_runtime_call(RC_LEAF, 6467 OptoRuntime::base64_encodeBlock_Type(), 6468 stubAddr, stubName, TypePtr::BOTTOM, 6469 src_start, offset, len, dest_start, dp, isURL); 6470 return true; 6471 } 6472 6473 //------------------------------inline_sha_implCompress----------------------- 6474 // 6475 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6476 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6477 // 6478 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6479 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6480 // 6481 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6482 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6483 // 6484 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6485 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6486 6487 Node* sha_obj = argument(0); 6488 Node* src = argument(1); // type oop 6489 Node* ofs = argument(2); // type int 6490 6491 const Type* src_type = src->Value(&_gvn); 6492 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6493 if (top_src == NULL || top_src->klass() == NULL) { 6494 // failed array check 6495 return false; 6496 } 6497 // Figure out the size and type of the elements we will be copying. 6498 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6499 if (src_elem != T_BYTE) { 6500 return false; 6501 } 6502 // 'src_start' points to src array + offset 6503 src = must_be_not_null(src, true); 6504 src = access_resolve(src, ACCESS_READ); 6505 Node* src_start = array_element_address(src, ofs, src_elem); 6506 Node* state = NULL; 6507 address stubAddr; 6508 const char *stubName; 6509 6510 switch(id) { 6511 case vmIntrinsics::_sha_implCompress: 6512 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6513 state = get_state_from_sha_object(sha_obj); 6514 stubAddr = StubRoutines::sha1_implCompress(); 6515 stubName = "sha1_implCompress"; 6516 break; 6517 case vmIntrinsics::_sha2_implCompress: 6518 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6519 state = get_state_from_sha_object(sha_obj); 6520 stubAddr = StubRoutines::sha256_implCompress(); 6521 stubName = "sha256_implCompress"; 6522 break; 6523 case vmIntrinsics::_sha5_implCompress: 6524 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6525 state = get_state_from_sha5_object(sha_obj); 6526 stubAddr = StubRoutines::sha512_implCompress(); 6527 stubName = "sha512_implCompress"; 6528 break; 6529 default: 6530 fatal_unexpected_iid(id); 6531 return false; 6532 } 6533 if (state == NULL) return false; 6534 6535 // Call the stub. 6536 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6537 stubAddr, stubName, TypePtr::BOTTOM, 6538 src_start, state); 6539 6540 return true; 6541 } 6542 6543 //------------------------------inline_digestBase_implCompressMB----------------------- 6544 // 6545 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6546 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6547 // 6548 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6549 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6550 "need SHA1/SHA256/SHA512 instruction support"); 6551 assert((uint)predicate < 3, "sanity"); 6552 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6553 6554 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6555 Node* src = argument(1); // byte[] array 6556 Node* ofs = argument(2); // type int 6557 Node* limit = argument(3); // type int 6558 6559 const Type* src_type = src->Value(&_gvn); 6560 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6561 if (top_src == NULL || top_src->klass() == NULL) { 6562 // failed array check 6563 return false; 6564 } 6565 // Figure out the size and type of the elements we will be copying. 6566 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6567 if (src_elem != T_BYTE) { 6568 return false; 6569 } 6570 // 'src_start' points to src array + offset 6571 src = must_be_not_null(src, false); 6572 src = access_resolve(src, ACCESS_READ); 6573 Node* src_start = array_element_address(src, ofs, src_elem); 6574 6575 const char* klass_SHA_name = NULL; 6576 const char* stub_name = NULL; 6577 address stub_addr = NULL; 6578 bool long_state = false; 6579 6580 switch (predicate) { 6581 case 0: 6582 if (UseSHA1Intrinsics) { 6583 klass_SHA_name = "sun/security/provider/SHA"; 6584 stub_name = "sha1_implCompressMB"; 6585 stub_addr = StubRoutines::sha1_implCompressMB(); 6586 } 6587 break; 6588 case 1: 6589 if (UseSHA256Intrinsics) { 6590 klass_SHA_name = "sun/security/provider/SHA2"; 6591 stub_name = "sha256_implCompressMB"; 6592 stub_addr = StubRoutines::sha256_implCompressMB(); 6593 } 6594 break; 6595 case 2: 6596 if (UseSHA512Intrinsics) { 6597 klass_SHA_name = "sun/security/provider/SHA5"; 6598 stub_name = "sha512_implCompressMB"; 6599 stub_addr = StubRoutines::sha512_implCompressMB(); 6600 long_state = true; 6601 } 6602 break; 6603 default: 6604 fatal("unknown SHA intrinsic predicate: %d", predicate); 6605 } 6606 if (klass_SHA_name != NULL) { 6607 // get DigestBase klass to lookup for SHA klass 6608 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6609 assert(tinst != NULL, "digestBase_obj is not instance???"); 6610 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6611 6612 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6613 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6614 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6615 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6616 } 6617 return false; 6618 } 6619 //------------------------------inline_sha_implCompressMB----------------------- 6620 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6621 bool long_state, address stubAddr, const char *stubName, 6622 Node* src_start, Node* ofs, Node* limit) { 6623 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6624 const TypeOopPtr* xtype = aklass->as_instance_type(); 6625 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6626 sha_obj = _gvn.transform(sha_obj); 6627 6628 Node* state; 6629 if (long_state) { 6630 state = get_state_from_sha5_object(sha_obj); 6631 } else { 6632 state = get_state_from_sha_object(sha_obj); 6633 } 6634 if (state == NULL) return false; 6635 6636 // Call the stub. 6637 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6638 OptoRuntime::digestBase_implCompressMB_Type(), 6639 stubAddr, stubName, TypePtr::BOTTOM, 6640 src_start, state, ofs, limit); 6641 // return ofs (int) 6642 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6643 set_result(result); 6644 6645 return true; 6646 } 6647 6648 //------------------------------get_state_from_sha_object----------------------- 6649 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6650 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6651 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6652 if (sha_state == NULL) return (Node *) NULL; 6653 6654 // now have the array, need to get the start address of the state array 6655 sha_state = access_resolve(sha_state, ACCESS_WRITE); 6656 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6657 return state; 6658 } 6659 6660 //------------------------------get_state_from_sha5_object----------------------- 6661 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6662 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6663 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6664 if (sha_state == NULL) return (Node *) NULL; 6665 6666 // now have the array, need to get the start address of the state array 6667 sha_state = access_resolve(sha_state, ACCESS_WRITE); 6668 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6669 return state; 6670 } 6671 6672 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6673 // Return node representing slow path of predicate check. 6674 // the pseudo code we want to emulate with this predicate is: 6675 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6676 // 6677 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6678 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6679 "need SHA1/SHA256/SHA512 instruction support"); 6680 assert((uint)predicate < 3, "sanity"); 6681 6682 // The receiver was checked for NULL already. 6683 Node* digestBaseObj = argument(0); 6684 6685 // get DigestBase klass for instanceOf check 6686 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6687 assert(tinst != NULL, "digestBaseObj is null"); 6688 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6689 6690 const char* klass_SHA_name = NULL; 6691 switch (predicate) { 6692 case 0: 6693 if (UseSHA1Intrinsics) { 6694 // we want to do an instanceof comparison against the SHA class 6695 klass_SHA_name = "sun/security/provider/SHA"; 6696 } 6697 break; 6698 case 1: 6699 if (UseSHA256Intrinsics) { 6700 // we want to do an instanceof comparison against the SHA2 class 6701 klass_SHA_name = "sun/security/provider/SHA2"; 6702 } 6703 break; 6704 case 2: 6705 if (UseSHA512Intrinsics) { 6706 // we want to do an instanceof comparison against the SHA5 class 6707 klass_SHA_name = "sun/security/provider/SHA5"; 6708 } 6709 break; 6710 default: 6711 fatal("unknown SHA intrinsic predicate: %d", predicate); 6712 } 6713 6714 ciKlass* klass_SHA = NULL; 6715 if (klass_SHA_name != NULL) { 6716 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6717 } 6718 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6719 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6720 Node* ctrl = control(); 6721 set_control(top()); // no intrinsic path 6722 return ctrl; 6723 } 6724 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6725 6726 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6727 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6728 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6729 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6730 6731 return instof_false; // even if it is NULL 6732 } 6733 6734 //-------------inline_fma----------------------------------- 6735 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) { 6736 Node *a = NULL; 6737 Node *b = NULL; 6738 Node *c = NULL; 6739 Node* result = NULL; 6740 switch (id) { 6741 case vmIntrinsics::_fmaD: 6742 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each."); 6743 // no receiver since it is static method 6744 a = round_double_node(argument(0)); 6745 b = round_double_node(argument(2)); 6746 c = round_double_node(argument(4)); 6747 result = _gvn.transform(new FmaDNode(control(), a, b, c)); 6748 break; 6749 case vmIntrinsics::_fmaF: 6750 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each."); 6751 a = argument(0); 6752 b = argument(1); 6753 c = argument(2); 6754 result = _gvn.transform(new FmaFNode(control(), a, b, c)); 6755 break; 6756 default: 6757 fatal_unexpected_iid(id); break; 6758 } 6759 set_result(result); 6760 return true; 6761 } 6762 6763 bool LibraryCallKit::inline_character_compare(vmIntrinsics::ID id) { 6764 // argument(0) is receiver 6765 Node* codePoint = argument(1); 6766 Node* n = NULL; 6767 6768 switch (id) { 6769 case vmIntrinsics::_isDigit : 6770 n = new DigitNode(control(), codePoint); 6771 break; 6772 case vmIntrinsics::_isLowerCase : 6773 n = new LowerCaseNode(control(), codePoint); 6774 break; 6775 case vmIntrinsics::_isUpperCase : 6776 n = new UpperCaseNode(control(), codePoint); 6777 break; 6778 case vmIntrinsics::_isWhitespace : 6779 n = new WhitespaceNode(control(), codePoint); 6780 break; 6781 default: 6782 fatal_unexpected_iid(id); 6783 } 6784 6785 set_result(_gvn.transform(n)); 6786 return true; 6787 } 6788 6789 //------------------------------inline_fp_min_max------------------------------ 6790 bool LibraryCallKit::inline_fp_min_max(vmIntrinsics::ID id) { 6791 Node *a = NULL; 6792 Node *b = NULL; 6793 Node *n = NULL; 6794 switch (id) { 6795 case vmIntrinsics::_maxF: 6796 case vmIntrinsics::_minF: 6797 assert(callee()->signature()->size() == 2, "minF/maxF has 2 parameters of size 1 each."); 6798 a = argument(0); 6799 b = argument(1); 6800 break; 6801 case vmIntrinsics::_maxD: 6802 case vmIntrinsics::_minD: 6803 assert(callee()->signature()->size() == 4, "minD/maxD has 2 parameters of size 2 each."); 6804 a = round_double_node(argument(0)); 6805 b = round_double_node(argument(2)); 6806 break; 6807 default: 6808 fatal_unexpected_iid(id); 6809 break; 6810 } 6811 if (a->is_Con() || b->is_Con()) { 6812 return false; 6813 } 6814 switch (id) { 6815 case vmIntrinsics::_maxF: n = new MaxFNode(a, b); break; 6816 case vmIntrinsics::_minF: n = new MinFNode(a, b); break; 6817 case vmIntrinsics::_maxD: n = new MaxDNode(a, b); break; 6818 case vmIntrinsics::_minD: n = new MinDNode(a, b); break; 6819 default: fatal_unexpected_iid(id); break; 6820 } 6821 set_result(_gvn.transform(n)); 6822 return true; 6823 } 6824 6825 bool LibraryCallKit::inline_profileBoolean() { 6826 Node* counts = argument(1); 6827 const TypeAryPtr* ary = NULL; 6828 ciArray* aobj = NULL; 6829 if (counts->is_Con() 6830 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6831 && (aobj = ary->const_oop()->as_array()) != NULL 6832 && (aobj->length() == 2)) { 6833 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6834 jint false_cnt = aobj->element_value(0).as_int(); 6835 jint true_cnt = aobj->element_value(1).as_int(); 6836 6837 if (C->log() != NULL) { 6838 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6839 false_cnt, true_cnt); 6840 } 6841 6842 if (false_cnt + true_cnt == 0) { 6843 // According to profile, never executed. 6844 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6845 Deoptimization::Action_reinterpret); 6846 return true; 6847 } 6848 6849 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6850 // is a number of each value occurrences. 6851 Node* result = argument(0); 6852 if (false_cnt == 0 || true_cnt == 0) { 6853 // According to profile, one value has been never seen. 6854 int expected_val = (false_cnt == 0) ? 1 : 0; 6855 6856 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6857 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6858 6859 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6860 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6861 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6862 6863 { // Slow path: uncommon trap for never seen value and then reexecute 6864 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6865 // the value has been seen at least once. 6866 PreserveJVMState pjvms(this); 6867 PreserveReexecuteState preexecs(this); 6868 jvms()->set_should_reexecute(true); 6869 6870 set_control(slow_path); 6871 set_i_o(i_o()); 6872 6873 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6874 Deoptimization::Action_reinterpret); 6875 } 6876 // The guard for never seen value enables sharpening of the result and 6877 // returning a constant. It allows to eliminate branches on the same value 6878 // later on. 6879 set_control(fast_path); 6880 result = intcon(expected_val); 6881 } 6882 // Stop profiling. 6883 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6884 // By replacing method body with profile data (represented as ProfileBooleanNode 6885 // on IR level) we effectively disable profiling. 6886 // It enables full speed execution once optimized code is generated. 6887 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6888 C->record_for_igvn(profile); 6889 set_result(profile); 6890 return true; 6891 } else { 6892 // Continue profiling. 6893 // Profile data isn't available at the moment. So, execute method's bytecode version. 6894 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6895 // is compiled and counters aren't available since corresponding MethodHandle 6896 // isn't a compile-time constant. 6897 return false; 6898 } 6899 } 6900 6901 bool LibraryCallKit::inline_isCompileConstant() { 6902 Node* n = argument(0); 6903 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6904 return true; 6905 }