1 /* 2 * Copyright (c) 1999, 2019, 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_unsafe_make_private_buffer(); 273 bool inline_unsafe_finish_private_buffer(); 274 bool inline_native_currentThread(); 275 276 bool inline_native_time_funcs(address method, const char* funcName); 277 #ifdef JFR_HAVE_INTRINSICS 278 bool inline_native_classID(); 279 bool inline_native_getEventWriter(); 280 #endif 281 bool inline_native_isInterrupted(); 282 bool inline_native_Class_query(vmIntrinsics::ID id); 283 bool inline_value_Class_conversion(vmIntrinsics::ID id); 284 bool inline_native_subtype_check(); 285 bool inline_native_getLength(); 286 bool inline_array_copyOf(bool is_copyOfRange); 287 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 288 bool inline_preconditions_checkIndex(); 289 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array); 290 bool inline_native_clone(bool is_virtual); 291 bool inline_native_Reflection_getCallerClass(); 292 // Helper function for inlining native object hash method 293 bool inline_native_hashcode(bool is_virtual, bool is_static); 294 bool inline_native_getClass(); 295 296 // Helper functions for inlining arraycopy 297 bool inline_arraycopy(); 298 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 299 RegionNode* slow_region); 300 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 301 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp, 302 uint new_idx); 303 304 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind; 305 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind); 306 bool inline_unsafe_fence(vmIntrinsics::ID id); 307 bool inline_onspinwait(); 308 bool inline_fp_conversions(vmIntrinsics::ID id); 309 bool inline_number_methods(vmIntrinsics::ID id); 310 bool inline_reference_get(); 311 bool inline_Class_cast(); 312 bool inline_aescrypt_Block(vmIntrinsics::ID id); 313 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 314 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 315 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 316 Node* inline_counterMode_AESCrypt_predicate(); 317 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 318 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 319 bool inline_ghash_processBlocks(); 320 bool inline_base64_encodeBlock(); 321 bool inline_sha_implCompress(vmIntrinsics::ID id); 322 bool inline_digestBase_implCompressMB(int predicate); 323 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 324 bool long_state, address stubAddr, const char *stubName, 325 Node* src_start, Node* ofs, Node* limit); 326 Node* get_state_from_sha_object(Node *sha_object); 327 Node* get_state_from_sha5_object(Node *sha_object); 328 Node* inline_digestBase_implCompressMB_predicate(int predicate); 329 bool inline_encodeISOArray(); 330 bool inline_updateCRC32(); 331 bool inline_updateBytesCRC32(); 332 bool inline_updateByteBufferCRC32(); 333 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 334 bool inline_updateBytesCRC32C(); 335 bool inline_updateDirectByteBufferCRC32C(); 336 bool inline_updateBytesAdler32(); 337 bool inline_updateByteBufferAdler32(); 338 bool inline_multiplyToLen(); 339 bool inline_hasNegatives(); 340 bool inline_squareToLen(); 341 bool inline_mulAdd(); 342 bool inline_montgomeryMultiply(); 343 bool inline_montgomerySquare(); 344 bool inline_vectorizedMismatch(); 345 bool inline_fma(vmIntrinsics::ID id); 346 bool inline_character_compare(vmIntrinsics::ID id); 347 bool inline_fp_min_max(vmIntrinsics::ID id); 348 349 bool inline_profileBoolean(); 350 bool inline_isCompileConstant(); 351 void clear_upper_avx() { 352 #ifdef X86 353 if (UseAVX >= 2) { 354 C->set_clear_upper_avx(true); 355 } 356 #endif 357 } 358 }; 359 360 //---------------------------make_vm_intrinsic---------------------------- 361 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 362 vmIntrinsics::ID id = m->intrinsic_id(); 363 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 364 365 if (!m->is_loaded()) { 366 // Do not attempt to inline unloaded methods. 367 return NULL; 368 } 369 370 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 371 bool is_available = false; 372 373 { 374 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 375 // the compiler must transition to '_thread_in_vm' state because both 376 // methods access VM-internal data. 377 VM_ENTRY_MARK; 378 methodHandle mh(THREAD, m->get_Method()); 379 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) && 380 !C->directive()->is_intrinsic_disabled(mh) && 381 !vmIntrinsics::is_disabled_by_flags(mh); 382 383 } 384 385 if (is_available) { 386 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 387 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 388 return new LibraryIntrinsic(m, is_virtual, 389 vmIntrinsics::predicates_needed(id), 390 vmIntrinsics::does_virtual_dispatch(id), 391 (vmIntrinsics::ID) id); 392 } else { 393 return NULL; 394 } 395 } 396 397 //----------------------register_library_intrinsics----------------------- 398 // Initialize this file's data structures, for each Compile instance. 399 void Compile::register_library_intrinsics() { 400 // Nothing to do here. 401 } 402 403 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 404 LibraryCallKit kit(jvms, this); 405 Compile* C = kit.C; 406 int nodes = C->unique(); 407 #ifndef PRODUCT 408 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 409 char buf[1000]; 410 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 411 tty->print_cr("Intrinsic %s", str); 412 } 413 #endif 414 ciMethod* callee = kit.callee(); 415 const int bci = kit.bci(); 416 417 // Try to inline the intrinsic. 418 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 419 kit.try_to_inline(_last_predicate)) { 420 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)" 421 : "(intrinsic)"; 422 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 423 if (C->print_intrinsics() || C->print_inlining()) { 424 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 425 } 426 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 427 if (C->log()) { 428 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 429 vmIntrinsics::name_at(intrinsic_id()), 430 (is_virtual() ? " virtual='1'" : ""), 431 C->unique() - nodes); 432 } 433 // Push the result from the inlined method onto the stack. 434 kit.push_result(); 435 C->print_inlining_update(this); 436 return kit.transfer_exceptions_into_jvms(); 437 } 438 439 // The intrinsic bailed out 440 if (jvms->has_method()) { 441 // Not a root compile. 442 const char* msg; 443 if (callee->intrinsic_candidate()) { 444 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 445 } else { 446 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 447 : "failed to inline (intrinsic), method not annotated"; 448 } 449 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg); 450 if (C->print_intrinsics() || C->print_inlining()) { 451 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 452 } 453 } else { 454 // Root compile 455 ResourceMark rm; 456 stringStream msg_stream; 457 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 458 vmIntrinsics::name_at(intrinsic_id()), 459 is_virtual() ? " (virtual)" : "", bci); 460 const char *msg = msg_stream.as_string(); 461 log_debug(jit, inlining)("%s", msg); 462 if (C->print_intrinsics() || C->print_inlining()) { 463 tty->print("%s", msg); 464 } 465 } 466 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 467 C->print_inlining_update(this); 468 return NULL; 469 } 470 471 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 472 LibraryCallKit kit(jvms, this); 473 Compile* C = kit.C; 474 int nodes = C->unique(); 475 _last_predicate = predicate; 476 #ifndef PRODUCT 477 assert(is_predicated() && predicate < predicates_count(), "sanity"); 478 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 479 char buf[1000]; 480 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 481 tty->print_cr("Predicate for intrinsic %s", str); 482 } 483 #endif 484 ciMethod* callee = kit.callee(); 485 const int bci = kit.bci(); 486 487 Node* slow_ctl = kit.try_to_predicate(predicate); 488 if (!kit.failing()) { 489 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)" 490 : "(intrinsic, predicate)"; 491 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 492 if (C->print_intrinsics() || C->print_inlining()) { 493 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 494 } 495 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 496 if (C->log()) { 497 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 498 vmIntrinsics::name_at(intrinsic_id()), 499 (is_virtual() ? " virtual='1'" : ""), 500 C->unique() - nodes); 501 } 502 return slow_ctl; // Could be NULL if the check folds. 503 } 504 505 // The intrinsic bailed out 506 if (jvms->has_method()) { 507 // Not a root compile. 508 const char* msg = "failed to generate predicate for intrinsic"; 509 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg); 510 if (C->print_intrinsics() || C->print_inlining()) { 511 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 512 } 513 } else { 514 // Root compile 515 ResourceMark rm; 516 stringStream msg_stream; 517 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 518 vmIntrinsics::name_at(intrinsic_id()), 519 is_virtual() ? " (virtual)" : "", bci); 520 const char *msg = msg_stream.as_string(); 521 log_debug(jit, inlining)("%s", msg); 522 if (C->print_intrinsics() || C->print_inlining()) { 523 C->print_inlining_stream()->print("%s", msg); 524 } 525 } 526 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 527 return NULL; 528 } 529 530 bool LibraryCallKit::try_to_inline(int predicate) { 531 // Handle symbolic names for otherwise undistinguished boolean switches: 532 const bool is_store = true; 533 const bool is_compress = true; 534 const bool is_static = true; 535 const bool is_volatile = true; 536 537 if (!jvms()->has_method()) { 538 // Root JVMState has a null method. 539 assert(map()->memory()->Opcode() == Op_Parm, ""); 540 // Insert the memory aliasing node 541 set_all_memory(reset_memory()); 542 } 543 assert(merged_memory(), ""); 544 545 546 switch (intrinsic_id()) { 547 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 548 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 549 case vmIntrinsics::_getClass: return inline_native_getClass(); 550 551 case vmIntrinsics::_dsin: 552 case vmIntrinsics::_dcos: 553 case vmIntrinsics::_dtan: 554 case vmIntrinsics::_dabs: 555 case vmIntrinsics::_datan2: 556 case vmIntrinsics::_dsqrt: 557 case vmIntrinsics::_dexp: 558 case vmIntrinsics::_dlog: 559 case vmIntrinsics::_dlog10: 560 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 561 562 case vmIntrinsics::_min: 563 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 564 565 case vmIntrinsics::_notify: 566 case vmIntrinsics::_notifyAll: 567 return inline_notify(intrinsic_id()); 568 569 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 570 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 571 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 572 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 573 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 574 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 575 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 576 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 577 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh(); 578 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 579 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 580 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 581 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 582 583 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 584 585 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 586 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 587 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 588 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 589 590 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 591 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 592 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 593 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 594 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 595 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 596 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 597 598 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 599 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 600 601 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 602 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 603 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 604 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 605 606 case vmIntrinsics::_compressStringC: 607 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 608 case vmIntrinsics::_inflateStringC: 609 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 610 611 case vmIntrinsics::_makePrivateBuffer: return inline_unsafe_make_private_buffer(); 612 case vmIntrinsics::_finishPrivateBuffer: return inline_unsafe_finish_private_buffer(); 613 case vmIntrinsics::_getReference: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 614 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 615 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 616 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 617 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 618 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 619 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 620 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 621 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 622 case vmIntrinsics::_getValue: return inline_unsafe_access(!is_store, T_VALUETYPE,Relaxed, false); 623 624 case vmIntrinsics::_putReference: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 625 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 626 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 627 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 628 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 629 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 630 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 631 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 632 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 633 case vmIntrinsics::_putValue: return inline_unsafe_access( is_store, T_VALUETYPE,Relaxed, false); 634 635 case vmIntrinsics::_getReferenceVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 636 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 637 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 638 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 639 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 640 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 641 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 642 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 643 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 644 645 case vmIntrinsics::_putReferenceVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 646 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 647 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 648 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 649 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 650 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 651 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 652 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 653 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 654 655 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 656 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 657 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 658 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 659 660 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 661 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 662 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 663 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 664 665 case vmIntrinsics::_getReferenceAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 666 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 667 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 668 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 669 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 670 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 671 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 672 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 673 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 674 675 case vmIntrinsics::_putReferenceRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 676 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 677 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 678 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 679 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 680 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 681 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 682 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 683 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 684 685 case vmIntrinsics::_getReferenceOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 686 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 687 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 688 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 689 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 690 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 691 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 692 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 693 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 694 695 case vmIntrinsics::_putReferenceOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 696 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 697 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 698 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 699 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 700 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 701 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 702 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 703 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 704 705 case vmIntrinsics::_compareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 706 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); 707 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); 708 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 709 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 710 711 case vmIntrinsics::_weakCompareAndSetReferencePlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 712 case vmIntrinsics::_weakCompareAndSetReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 713 case vmIntrinsics::_weakCompareAndSetReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 714 case vmIntrinsics::_weakCompareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 715 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); 716 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); 717 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); 718 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); 719 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); 720 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); 721 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); 722 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); 723 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 724 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 725 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 726 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 727 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 728 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 729 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 730 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 731 732 case vmIntrinsics::_compareAndExchangeReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 733 case vmIntrinsics::_compareAndExchangeReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 734 case vmIntrinsics::_compareAndExchangeReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 735 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); 736 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); 737 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); 738 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); 739 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); 740 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); 741 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 742 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 743 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 744 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 745 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 746 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 747 748 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); 749 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); 750 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 751 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 752 753 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); 754 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); 755 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 756 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 757 case vmIntrinsics::_getAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 758 759 case vmIntrinsics::_loadFence: 760 case vmIntrinsics::_storeFence: 761 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 762 763 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 764 765 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 766 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 767 768 #ifdef JFR_HAVE_INTRINSICS 769 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime"); 770 case vmIntrinsics::_getClassId: return inline_native_classID(); 771 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter(); 772 #endif 773 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 774 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 775 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 776 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 777 case vmIntrinsics::_getLength: return inline_native_getLength(); 778 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 779 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 780 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 781 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 782 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 783 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 784 785 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 786 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 787 788 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 789 790 case vmIntrinsics::_isInstance: 791 case vmIntrinsics::_getModifiers: 792 case vmIntrinsics::_isInterface: 793 case vmIntrinsics::_isArray: 794 case vmIntrinsics::_isPrimitive: 795 case vmIntrinsics::_getSuperclass: 796 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 797 798 case vmIntrinsics::_asValueType: 799 case vmIntrinsics::_asBoxType: return inline_value_Class_conversion(intrinsic_id()); 800 801 case vmIntrinsics::_floatToRawIntBits: 802 case vmIntrinsics::_floatToIntBits: 803 case vmIntrinsics::_intBitsToFloat: 804 case vmIntrinsics::_doubleToRawLongBits: 805 case vmIntrinsics::_doubleToLongBits: 806 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 807 808 case vmIntrinsics::_numberOfLeadingZeros_i: 809 case vmIntrinsics::_numberOfLeadingZeros_l: 810 case vmIntrinsics::_numberOfTrailingZeros_i: 811 case vmIntrinsics::_numberOfTrailingZeros_l: 812 case vmIntrinsics::_bitCount_i: 813 case vmIntrinsics::_bitCount_l: 814 case vmIntrinsics::_reverseBytes_i: 815 case vmIntrinsics::_reverseBytes_l: 816 case vmIntrinsics::_reverseBytes_s: 817 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 818 819 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 820 821 case vmIntrinsics::_Reference_get: return inline_reference_get(); 822 823 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 824 825 case vmIntrinsics::_aescrypt_encryptBlock: 826 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 827 828 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 829 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 830 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 831 832 case vmIntrinsics::_counterMode_AESCrypt: 833 return inline_counterMode_AESCrypt(intrinsic_id()); 834 835 case vmIntrinsics::_sha_implCompress: 836 case vmIntrinsics::_sha2_implCompress: 837 case vmIntrinsics::_sha5_implCompress: 838 return inline_sha_implCompress(intrinsic_id()); 839 840 case vmIntrinsics::_digestBase_implCompressMB: 841 return inline_digestBase_implCompressMB(predicate); 842 843 case vmIntrinsics::_multiplyToLen: 844 return inline_multiplyToLen(); 845 846 case vmIntrinsics::_squareToLen: 847 return inline_squareToLen(); 848 849 case vmIntrinsics::_mulAdd: 850 return inline_mulAdd(); 851 852 case vmIntrinsics::_montgomeryMultiply: 853 return inline_montgomeryMultiply(); 854 case vmIntrinsics::_montgomerySquare: 855 return inline_montgomerySquare(); 856 857 case vmIntrinsics::_vectorizedMismatch: 858 return inline_vectorizedMismatch(); 859 860 case vmIntrinsics::_ghash_processBlocks: 861 return inline_ghash_processBlocks(); 862 case vmIntrinsics::_base64_encodeBlock: 863 return inline_base64_encodeBlock(); 864 865 case vmIntrinsics::_encodeISOArray: 866 case vmIntrinsics::_encodeByteISOArray: 867 return inline_encodeISOArray(); 868 869 case vmIntrinsics::_updateCRC32: 870 return inline_updateCRC32(); 871 case vmIntrinsics::_updateBytesCRC32: 872 return inline_updateBytesCRC32(); 873 case vmIntrinsics::_updateByteBufferCRC32: 874 return inline_updateByteBufferCRC32(); 875 876 case vmIntrinsics::_updateBytesCRC32C: 877 return inline_updateBytesCRC32C(); 878 case vmIntrinsics::_updateDirectByteBufferCRC32C: 879 return inline_updateDirectByteBufferCRC32C(); 880 881 case vmIntrinsics::_updateBytesAdler32: 882 return inline_updateBytesAdler32(); 883 case vmIntrinsics::_updateByteBufferAdler32: 884 return inline_updateByteBufferAdler32(); 885 886 case vmIntrinsics::_profileBoolean: 887 return inline_profileBoolean(); 888 case vmIntrinsics::_isCompileConstant: 889 return inline_isCompileConstant(); 890 891 case vmIntrinsics::_hasNegatives: 892 return inline_hasNegatives(); 893 894 case vmIntrinsics::_fmaD: 895 case vmIntrinsics::_fmaF: 896 return inline_fma(intrinsic_id()); 897 898 case vmIntrinsics::_isDigit: 899 case vmIntrinsics::_isLowerCase: 900 case vmIntrinsics::_isUpperCase: 901 case vmIntrinsics::_isWhitespace: 902 return inline_character_compare(intrinsic_id()); 903 904 case vmIntrinsics::_maxF: 905 case vmIntrinsics::_minF: 906 case vmIntrinsics::_maxD: 907 case vmIntrinsics::_minD: 908 return inline_fp_min_max(intrinsic_id()); 909 910 default: 911 // If you get here, it may be that someone has added a new intrinsic 912 // to the list in vmSymbols.hpp without implementing it here. 913 #ifndef PRODUCT 914 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 915 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 916 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 917 } 918 #endif 919 return false; 920 } 921 } 922 923 Node* LibraryCallKit::try_to_predicate(int predicate) { 924 if (!jvms()->has_method()) { 925 // Root JVMState has a null method. 926 assert(map()->memory()->Opcode() == Op_Parm, ""); 927 // Insert the memory aliasing node 928 set_all_memory(reset_memory()); 929 } 930 assert(merged_memory(), ""); 931 932 switch (intrinsic_id()) { 933 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 934 return inline_cipherBlockChaining_AESCrypt_predicate(false); 935 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 936 return inline_cipherBlockChaining_AESCrypt_predicate(true); 937 case vmIntrinsics::_counterMode_AESCrypt: 938 return inline_counterMode_AESCrypt_predicate(); 939 case vmIntrinsics::_digestBase_implCompressMB: 940 return inline_digestBase_implCompressMB_predicate(predicate); 941 942 default: 943 // If you get here, it may be that someone has added a new intrinsic 944 // to the list in vmSymbols.hpp without implementing it here. 945 #ifndef PRODUCT 946 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 947 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 948 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 949 } 950 #endif 951 Node* slow_ctl = control(); 952 set_control(top()); // No fast path instrinsic 953 return slow_ctl; 954 } 955 } 956 957 //------------------------------set_result------------------------------- 958 // Helper function for finishing intrinsics. 959 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 960 record_for_igvn(region); 961 set_control(_gvn.transform(region)); 962 set_result( _gvn.transform(value)); 963 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 964 } 965 966 //------------------------------generate_guard--------------------------- 967 // Helper function for generating guarded fast-slow graph structures. 968 // The given 'test', if true, guards a slow path. If the test fails 969 // then a fast path can be taken. (We generally hope it fails.) 970 // In all cases, GraphKit::control() is updated to the fast path. 971 // The returned value represents the control for the slow path. 972 // The return value is never 'top'; it is either a valid control 973 // or NULL if it is obvious that the slow path can never be taken. 974 // Also, if region and the slow control are not NULL, the slow edge 975 // is appended to the region. 976 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 977 if (stopped()) { 978 // Already short circuited. 979 return NULL; 980 } 981 982 // Build an if node and its projections. 983 // If test is true we take the slow path, which we assume is uncommon. 984 if (_gvn.type(test) == TypeInt::ZERO) { 985 // The slow branch is never taken. No need to build this guard. 986 return NULL; 987 } 988 989 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 990 991 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 992 if (if_slow == top()) { 993 // The slow branch is never taken. No need to build this guard. 994 return NULL; 995 } 996 997 if (region != NULL) 998 region->add_req(if_slow); 999 1000 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1001 set_control(if_fast); 1002 1003 return if_slow; 1004 } 1005 1006 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1007 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1008 } 1009 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1010 return generate_guard(test, region, PROB_FAIR); 1011 } 1012 1013 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1014 Node* *pos_index) { 1015 if (stopped()) 1016 return NULL; // already stopped 1017 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1018 return NULL; // index is already adequately typed 1019 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1020 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1021 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1022 if (is_neg != NULL && pos_index != NULL) { 1023 // Emulate effect of Parse::adjust_map_after_if. 1024 Node* ccast = new CastIINode(index, TypeInt::POS); 1025 ccast->set_req(0, control()); 1026 (*pos_index) = _gvn.transform(ccast); 1027 } 1028 return is_neg; 1029 } 1030 1031 // Make sure that 'position' is a valid limit index, in [0..length]. 1032 // There are two equivalent plans for checking this: 1033 // A. (offset + copyLength) unsigned<= arrayLength 1034 // B. offset <= (arrayLength - copyLength) 1035 // We require that all of the values above, except for the sum and 1036 // difference, are already known to be non-negative. 1037 // Plan A is robust in the face of overflow, if offset and copyLength 1038 // are both hugely positive. 1039 // 1040 // Plan B is less direct and intuitive, but it does not overflow at 1041 // all, since the difference of two non-negatives is always 1042 // representable. Whenever Java methods must perform the equivalent 1043 // check they generally use Plan B instead of Plan A. 1044 // For the moment we use Plan A. 1045 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1046 Node* subseq_length, 1047 Node* array_length, 1048 RegionNode* region) { 1049 if (stopped()) 1050 return NULL; // already stopped 1051 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1052 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1053 return NULL; // common case of whole-array copy 1054 Node* last = subseq_length; 1055 if (!zero_offset) // last += offset 1056 last = _gvn.transform(new AddINode(last, offset)); 1057 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1058 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1059 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1060 return is_over; 1061 } 1062 1063 // Emit range checks for the given String.value byte array 1064 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 1065 if (stopped()) { 1066 return; // already stopped 1067 } 1068 RegionNode* bailout = new RegionNode(1); 1069 record_for_igvn(bailout); 1070 if (char_count) { 1071 // Convert char count to byte count 1072 count = _gvn.transform(new LShiftINode(count, intcon(1))); 1073 } 1074 1075 // Offset and count must not be negative 1076 generate_negative_guard(offset, bailout); 1077 generate_negative_guard(count, bailout); 1078 // Offset + count must not exceed length of array 1079 generate_limit_guard(offset, count, load_array_length(array), bailout); 1080 1081 if (bailout->req() > 1) { 1082 PreserveJVMState pjvms(this); 1083 set_control(_gvn.transform(bailout)); 1084 uncommon_trap(Deoptimization::Reason_intrinsic, 1085 Deoptimization::Action_maybe_recompile); 1086 } 1087 } 1088 1089 //--------------------------generate_current_thread-------------------- 1090 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1091 ciKlass* thread_klass = env()->Thread_klass(); 1092 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1093 Node* thread = _gvn.transform(new ThreadLocalNode()); 1094 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1095 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1096 tls_output = thread; 1097 return threadObj; 1098 } 1099 1100 1101 //------------------------------make_string_method_node------------------------ 1102 // Helper method for String intrinsic functions. This version is called with 1103 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 1104 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 1105 // containing the lengths of str1 and str2. 1106 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 1107 Node* result = NULL; 1108 switch (opcode) { 1109 case Op_StrIndexOf: 1110 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 1111 str1_start, cnt1, str2_start, cnt2, ae); 1112 break; 1113 case Op_StrComp: 1114 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 1115 str1_start, cnt1, str2_start, cnt2, ae); 1116 break; 1117 case Op_StrEquals: 1118 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 1119 // Use the constant length if there is one because optimized match rule may exist. 1120 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 1121 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 1122 break; 1123 default: 1124 ShouldNotReachHere(); 1125 return NULL; 1126 } 1127 1128 // All these intrinsics have checks. 1129 C->set_has_split_ifs(true); // Has chance for split-if optimization 1130 clear_upper_avx(); 1131 1132 return _gvn.transform(result); 1133 } 1134 1135 //------------------------------inline_string_compareTo------------------------ 1136 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 1137 Node* arg1 = argument(0); 1138 Node* arg2 = argument(1); 1139 1140 arg1 = must_be_not_null(arg1, true); 1141 arg2 = must_be_not_null(arg2, true); 1142 1143 arg1 = access_resolve(arg1, ACCESS_READ); 1144 arg2 = access_resolve(arg2, ACCESS_READ); 1145 1146 // Get start addr and length of first argument 1147 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1148 Node* arg1_cnt = load_array_length(arg1); 1149 1150 // Get start addr and length of second argument 1151 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1152 Node* arg2_cnt = load_array_length(arg2); 1153 1154 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1155 set_result(result); 1156 return true; 1157 } 1158 1159 //------------------------------inline_string_equals------------------------ 1160 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1161 Node* arg1 = argument(0); 1162 Node* arg2 = argument(1); 1163 1164 // paths (plus control) merge 1165 RegionNode* region = new RegionNode(3); 1166 Node* phi = new PhiNode(region, TypeInt::BOOL); 1167 1168 if (!stopped()) { 1169 1170 arg1 = must_be_not_null(arg1, true); 1171 arg2 = must_be_not_null(arg2, true); 1172 1173 arg1 = access_resolve(arg1, ACCESS_READ); 1174 arg2 = access_resolve(arg2, ACCESS_READ); 1175 1176 // Get start addr and length of first argument 1177 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1178 Node* arg1_cnt = load_array_length(arg1); 1179 1180 // Get start addr and length of second argument 1181 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1182 Node* arg2_cnt = load_array_length(arg2); 1183 1184 // Check for arg1_cnt != arg2_cnt 1185 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1186 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1187 Node* if_ne = generate_slow_guard(bol, NULL); 1188 if (if_ne != NULL) { 1189 phi->init_req(2, intcon(0)); 1190 region->init_req(2, if_ne); 1191 } 1192 1193 // Check for count == 0 is done by assembler code for StrEquals. 1194 1195 if (!stopped()) { 1196 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1197 phi->init_req(1, equals); 1198 region->init_req(1, control()); 1199 } 1200 } 1201 1202 // post merge 1203 set_control(_gvn.transform(region)); 1204 record_for_igvn(region); 1205 1206 set_result(_gvn.transform(phi)); 1207 return true; 1208 } 1209 1210 //------------------------------inline_array_equals---------------------------- 1211 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1212 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1213 Node* arg1 = argument(0); 1214 Node* arg2 = argument(1); 1215 1216 arg1 = access_resolve(arg1, ACCESS_READ); 1217 arg2 = access_resolve(arg2, ACCESS_READ); 1218 1219 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1220 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1221 clear_upper_avx(); 1222 1223 return true; 1224 } 1225 1226 //------------------------------inline_hasNegatives------------------------------ 1227 bool LibraryCallKit::inline_hasNegatives() { 1228 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1229 return false; 1230 } 1231 1232 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1233 // no receiver since it is static method 1234 Node* ba = argument(0); 1235 Node* offset = argument(1); 1236 Node* len = argument(2); 1237 1238 ba = must_be_not_null(ba, true); 1239 1240 // Range checks 1241 generate_string_range_check(ba, offset, len, false); 1242 if (stopped()) { 1243 return true; 1244 } 1245 ba = access_resolve(ba, ACCESS_READ); 1246 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1247 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1248 set_result(_gvn.transform(result)); 1249 return true; 1250 } 1251 1252 bool LibraryCallKit::inline_preconditions_checkIndex() { 1253 Node* index = argument(0); 1254 Node* length = argument(1); 1255 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1256 return false; 1257 } 1258 1259 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1260 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1261 1262 { 1263 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1264 uncommon_trap(Deoptimization::Reason_intrinsic, 1265 Deoptimization::Action_make_not_entrant); 1266 } 1267 1268 if (stopped()) { 1269 return false; 1270 } 1271 1272 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1273 BoolTest::mask btest = BoolTest::lt; 1274 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1275 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1276 _gvn.set_type(rc, rc->Value(&_gvn)); 1277 if (!rc_bool->is_Con()) { 1278 record_for_igvn(rc); 1279 } 1280 set_control(_gvn.transform(new IfTrueNode(rc))); 1281 { 1282 PreserveJVMState pjvms(this); 1283 set_control(_gvn.transform(new IfFalseNode(rc))); 1284 uncommon_trap(Deoptimization::Reason_range_check, 1285 Deoptimization::Action_make_not_entrant); 1286 } 1287 1288 if (stopped()) { 1289 return false; 1290 } 1291 1292 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1293 result->set_req(0, control()); 1294 result = _gvn.transform(result); 1295 set_result(result); 1296 replace_in_map(index, result); 1297 clear_upper_avx(); 1298 return true; 1299 } 1300 1301 //------------------------------inline_string_indexOf------------------------ 1302 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1303 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1304 return false; 1305 } 1306 Node* src = argument(0); 1307 Node* tgt = argument(1); 1308 1309 // Make the merge point 1310 RegionNode* result_rgn = new RegionNode(4); 1311 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1312 1313 src = must_be_not_null(src, true); 1314 tgt = must_be_not_null(tgt, true); 1315 1316 src = access_resolve(src, ACCESS_READ); 1317 tgt = access_resolve(tgt, ACCESS_READ); 1318 1319 // Get start addr and length of source string 1320 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1321 Node* src_count = load_array_length(src); 1322 1323 // Get start addr and length of substring 1324 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1325 Node* tgt_count = load_array_length(tgt); 1326 1327 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1328 // Divide src size by 2 if String is UTF16 encoded 1329 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1330 } 1331 if (ae == StrIntrinsicNode::UU) { 1332 // Divide substring size by 2 if String is UTF16 encoded 1333 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1334 } 1335 1336 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1337 if (result != NULL) { 1338 result_phi->init_req(3, result); 1339 result_rgn->init_req(3, control()); 1340 } 1341 set_control(_gvn.transform(result_rgn)); 1342 record_for_igvn(result_rgn); 1343 set_result(_gvn.transform(result_phi)); 1344 1345 return true; 1346 } 1347 1348 //-----------------------------inline_string_indexOf----------------------- 1349 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1350 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1351 return false; 1352 } 1353 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1354 return false; 1355 } 1356 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1357 Node* src = argument(0); // byte[] 1358 Node* src_count = argument(1); // char count 1359 Node* tgt = argument(2); // byte[] 1360 Node* tgt_count = argument(3); // char count 1361 Node* from_index = argument(4); // char index 1362 1363 src = must_be_not_null(src, true); 1364 tgt = must_be_not_null(tgt, true); 1365 1366 src = access_resolve(src, ACCESS_READ); 1367 tgt = access_resolve(tgt, ACCESS_READ); 1368 1369 // Multiply byte array index by 2 if String is UTF16 encoded 1370 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1371 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1372 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1373 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1374 1375 // Range checks 1376 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1377 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1378 if (stopped()) { 1379 return true; 1380 } 1381 1382 RegionNode* region = new RegionNode(5); 1383 Node* phi = new PhiNode(region, TypeInt::INT); 1384 1385 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1386 if (result != NULL) { 1387 // The result is index relative to from_index if substring was found, -1 otherwise. 1388 // Generate code which will fold into cmove. 1389 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1390 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1391 1392 Node* if_lt = generate_slow_guard(bol, NULL); 1393 if (if_lt != NULL) { 1394 // result == -1 1395 phi->init_req(3, result); 1396 region->init_req(3, if_lt); 1397 } 1398 if (!stopped()) { 1399 result = _gvn.transform(new AddINode(result, from_index)); 1400 phi->init_req(4, result); 1401 region->init_req(4, control()); 1402 } 1403 } 1404 1405 set_control(_gvn.transform(region)); 1406 record_for_igvn(region); 1407 set_result(_gvn.transform(phi)); 1408 clear_upper_avx(); 1409 1410 return true; 1411 } 1412 1413 // Create StrIndexOfNode with fast path checks 1414 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1415 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1416 // Check for substr count > string count 1417 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1418 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1419 Node* if_gt = generate_slow_guard(bol, NULL); 1420 if (if_gt != NULL) { 1421 phi->init_req(1, intcon(-1)); 1422 region->init_req(1, if_gt); 1423 } 1424 if (!stopped()) { 1425 // Check for substr count == 0 1426 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1427 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1428 Node* if_zero = generate_slow_guard(bol, NULL); 1429 if (if_zero != NULL) { 1430 phi->init_req(2, intcon(0)); 1431 region->init_req(2, if_zero); 1432 } 1433 } 1434 if (!stopped()) { 1435 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1436 } 1437 return NULL; 1438 } 1439 1440 //-----------------------------inline_string_indexOfChar----------------------- 1441 bool LibraryCallKit::inline_string_indexOfChar() { 1442 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1443 return false; 1444 } 1445 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1446 return false; 1447 } 1448 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1449 Node* src = argument(0); // byte[] 1450 Node* tgt = argument(1); // tgt is int ch 1451 Node* from_index = argument(2); 1452 Node* max = argument(3); 1453 1454 src = must_be_not_null(src, true); 1455 src = access_resolve(src, ACCESS_READ); 1456 1457 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1458 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1459 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1460 1461 // Range checks 1462 generate_string_range_check(src, src_offset, src_count, true); 1463 if (stopped()) { 1464 return true; 1465 } 1466 1467 RegionNode* region = new RegionNode(3); 1468 Node* phi = new PhiNode(region, TypeInt::INT); 1469 1470 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1471 C->set_has_split_ifs(true); // Has chance for split-if optimization 1472 _gvn.transform(result); 1473 1474 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1475 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1476 1477 Node* if_lt = generate_slow_guard(bol, NULL); 1478 if (if_lt != NULL) { 1479 // result == -1 1480 phi->init_req(2, result); 1481 region->init_req(2, if_lt); 1482 } 1483 if (!stopped()) { 1484 result = _gvn.transform(new AddINode(result, from_index)); 1485 phi->init_req(1, result); 1486 region->init_req(1, control()); 1487 } 1488 set_control(_gvn.transform(region)); 1489 record_for_igvn(region); 1490 set_result(_gvn.transform(phi)); 1491 1492 return true; 1493 } 1494 //---------------------------inline_string_copy--------------------- 1495 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1496 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1497 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1498 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1499 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1500 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1501 bool LibraryCallKit::inline_string_copy(bool compress) { 1502 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1503 return false; 1504 } 1505 int nargs = 5; // 2 oops, 3 ints 1506 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1507 1508 Node* src = argument(0); 1509 Node* src_offset = argument(1); 1510 Node* dst = argument(2); 1511 Node* dst_offset = argument(3); 1512 Node* length = argument(4); 1513 1514 // Check for allocation before we add nodes that would confuse 1515 // tightly_coupled_allocation() 1516 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1517 1518 // Figure out the size and type of the elements we will be copying. 1519 const Type* src_type = src->Value(&_gvn); 1520 const Type* dst_type = dst->Value(&_gvn); 1521 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1522 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1523 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1524 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1525 "Unsupported array types for inline_string_copy"); 1526 1527 src = must_be_not_null(src, true); 1528 dst = must_be_not_null(dst, true); 1529 1530 // Convert char[] offsets to byte[] offsets 1531 bool convert_src = (compress && src_elem == T_BYTE); 1532 bool convert_dst = (!compress && dst_elem == T_BYTE); 1533 if (convert_src) { 1534 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1535 } else if (convert_dst) { 1536 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1537 } 1538 1539 // Range checks 1540 generate_string_range_check(src, src_offset, length, convert_src); 1541 generate_string_range_check(dst, dst_offset, length, convert_dst); 1542 if (stopped()) { 1543 return true; 1544 } 1545 1546 src = access_resolve(src, ACCESS_READ); 1547 dst = access_resolve(dst, ACCESS_WRITE); 1548 1549 Node* src_start = array_element_address(src, src_offset, src_elem); 1550 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1551 // 'src_start' points to src array + scaled offset 1552 // 'dst_start' points to dst array + scaled offset 1553 Node* count = NULL; 1554 if (compress) { 1555 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1556 } else { 1557 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1558 } 1559 1560 if (alloc != NULL) { 1561 if (alloc->maybe_set_complete(&_gvn)) { 1562 // "You break it, you buy it." 1563 InitializeNode* init = alloc->initialization(); 1564 assert(init->is_complete(), "we just did this"); 1565 init->set_complete_with_arraycopy(); 1566 assert(dst->is_CheckCastPP(), "sanity"); 1567 assert(dst->in(0)->in(0) == init, "dest pinned"); 1568 } 1569 // Do not let stores that initialize this object be reordered with 1570 // a subsequent store that would make this object accessible by 1571 // other threads. 1572 // Record what AllocateNode this StoreStore protects so that 1573 // escape analysis can go from the MemBarStoreStoreNode to the 1574 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1575 // based on the escape status of the AllocateNode. 1576 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1577 } 1578 if (compress) { 1579 set_result(_gvn.transform(count)); 1580 } 1581 clear_upper_avx(); 1582 1583 return true; 1584 } 1585 1586 #ifdef _LP64 1587 #define XTOP ,top() /*additional argument*/ 1588 #else //_LP64 1589 #define XTOP /*no additional argument*/ 1590 #endif //_LP64 1591 1592 //------------------------inline_string_toBytesU-------------------------- 1593 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1594 bool LibraryCallKit::inline_string_toBytesU() { 1595 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1596 return false; 1597 } 1598 // Get the arguments. 1599 Node* value = argument(0); 1600 Node* offset = argument(1); 1601 Node* length = argument(2); 1602 1603 Node* newcopy = NULL; 1604 1605 // Set the original stack and the reexecute bit for the interpreter to reexecute 1606 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1607 { PreserveReexecuteState preexecs(this); 1608 jvms()->set_should_reexecute(true); 1609 1610 // Check if a null path was taken unconditionally. 1611 value = null_check(value); 1612 1613 RegionNode* bailout = new RegionNode(1); 1614 record_for_igvn(bailout); 1615 1616 // Range checks 1617 generate_negative_guard(offset, bailout); 1618 generate_negative_guard(length, bailout); 1619 generate_limit_guard(offset, length, load_array_length(value), bailout); 1620 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1621 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1622 1623 if (bailout->req() > 1) { 1624 PreserveJVMState pjvms(this); 1625 set_control(_gvn.transform(bailout)); 1626 uncommon_trap(Deoptimization::Reason_intrinsic, 1627 Deoptimization::Action_maybe_recompile); 1628 } 1629 if (stopped()) { 1630 return true; 1631 } 1632 1633 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1634 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1635 newcopy = new_array(klass_node, size, 0); // no arguments to push 1636 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1637 1638 // Calculate starting addresses. 1639 value = access_resolve(value, ACCESS_READ); 1640 Node* src_start = array_element_address(value, offset, T_CHAR); 1641 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1642 1643 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1644 const TypeInt* toffset = gvn().type(offset)->is_int(); 1645 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1646 1647 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1648 const char* copyfunc_name = "arraycopy"; 1649 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1650 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1651 OptoRuntime::fast_arraycopy_Type(), 1652 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1653 src_start, dst_start, ConvI2X(length) XTOP); 1654 // Do not let reads from the cloned object float above the arraycopy. 1655 if (alloc != NULL) { 1656 if (alloc->maybe_set_complete(&_gvn)) { 1657 // "You break it, you buy it." 1658 InitializeNode* init = alloc->initialization(); 1659 assert(init->is_complete(), "we just did this"); 1660 init->set_complete_with_arraycopy(); 1661 assert(newcopy->is_CheckCastPP(), "sanity"); 1662 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1663 } 1664 // Do not let stores that initialize this object be reordered with 1665 // a subsequent store that would make this object accessible by 1666 // other threads. 1667 // Record what AllocateNode this StoreStore protects so that 1668 // escape analysis can go from the MemBarStoreStoreNode to the 1669 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1670 // based on the escape status of the AllocateNode. 1671 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1672 } else { 1673 insert_mem_bar(Op_MemBarCPUOrder); 1674 } 1675 } // original reexecute is set back here 1676 1677 C->set_has_split_ifs(true); // Has chance for split-if optimization 1678 if (!stopped()) { 1679 set_result(newcopy); 1680 } 1681 clear_upper_avx(); 1682 1683 return true; 1684 } 1685 1686 //------------------------inline_string_getCharsU-------------------------- 1687 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1688 bool LibraryCallKit::inline_string_getCharsU() { 1689 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1690 return false; 1691 } 1692 1693 // Get the arguments. 1694 Node* src = argument(0); 1695 Node* src_begin = argument(1); 1696 Node* src_end = argument(2); // exclusive offset (i < src_end) 1697 Node* dst = argument(3); 1698 Node* dst_begin = argument(4); 1699 1700 // Check for allocation before we add nodes that would confuse 1701 // tightly_coupled_allocation() 1702 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1703 1704 // Check if a null path was taken unconditionally. 1705 src = null_check(src); 1706 dst = null_check(dst); 1707 if (stopped()) { 1708 return true; 1709 } 1710 1711 // Get length and convert char[] offset to byte[] offset 1712 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1713 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1714 1715 // Range checks 1716 generate_string_range_check(src, src_begin, length, true); 1717 generate_string_range_check(dst, dst_begin, length, false); 1718 if (stopped()) { 1719 return true; 1720 } 1721 1722 if (!stopped()) { 1723 src = access_resolve(src, ACCESS_READ); 1724 dst = access_resolve(dst, ACCESS_WRITE); 1725 1726 // Calculate starting addresses. 1727 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1728 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1729 1730 // Check if array addresses are aligned to HeapWordSize 1731 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1732 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1733 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1734 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1735 1736 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1737 const char* copyfunc_name = "arraycopy"; 1738 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1739 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1740 OptoRuntime::fast_arraycopy_Type(), 1741 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1742 src_start, dst_start, ConvI2X(length) XTOP); 1743 // Do not let reads from the cloned object float above the arraycopy. 1744 if (alloc != NULL) { 1745 if (alloc->maybe_set_complete(&_gvn)) { 1746 // "You break it, you buy it." 1747 InitializeNode* init = alloc->initialization(); 1748 assert(init->is_complete(), "we just did this"); 1749 init->set_complete_with_arraycopy(); 1750 assert(dst->is_CheckCastPP(), "sanity"); 1751 assert(dst->in(0)->in(0) == init, "dest pinned"); 1752 } 1753 // Do not let stores that initialize this object be reordered with 1754 // a subsequent store that would make this object accessible by 1755 // other threads. 1756 // Record what AllocateNode this StoreStore protects so that 1757 // escape analysis can go from the MemBarStoreStoreNode to the 1758 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1759 // based on the escape status of the AllocateNode. 1760 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1761 } else { 1762 insert_mem_bar(Op_MemBarCPUOrder); 1763 } 1764 } 1765 1766 C->set_has_split_ifs(true); // Has chance for split-if optimization 1767 return true; 1768 } 1769 1770 //----------------------inline_string_char_access---------------------------- 1771 // Store/Load char to/from byte[] array. 1772 // static void StringUTF16.putChar(byte[] val, int index, int c) 1773 // static char StringUTF16.getChar(byte[] val, int index) 1774 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1775 Node* value = argument(0); 1776 Node* index = argument(1); 1777 Node* ch = is_store ? argument(2) : NULL; 1778 1779 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1780 // correctly requires matched array shapes. 1781 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1782 "sanity: byte[] and char[] bases agree"); 1783 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1784 "sanity: byte[] and char[] scales agree"); 1785 1786 // Bail when getChar over constants is requested: constant folding would 1787 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1788 // Java method would constant fold nicely instead. 1789 if (!is_store && value->is_Con() && index->is_Con()) { 1790 return false; 1791 } 1792 1793 value = must_be_not_null(value, true); 1794 value = access_resolve(value, is_store ? ACCESS_WRITE : ACCESS_READ); 1795 1796 Node* adr = array_element_address(value, index, T_CHAR); 1797 if (adr->is_top()) { 1798 return false; 1799 } 1800 if (is_store) { 1801 access_store_at(value, adr, TypeAryPtr::BYTES, ch, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED); 1802 } else { 1803 ch = access_load_at(value, adr, TypeAryPtr::BYTES, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED | C2_CONTROL_DEPENDENT_LOAD); 1804 set_result(ch); 1805 } 1806 return true; 1807 } 1808 1809 //--------------------------round_double_node-------------------------------- 1810 // Round a double node if necessary. 1811 Node* LibraryCallKit::round_double_node(Node* n) { 1812 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1813 n = _gvn.transform(new RoundDoubleNode(0, n)); 1814 return n; 1815 } 1816 1817 //------------------------------inline_math----------------------------------- 1818 // public static double Math.abs(double) 1819 // public static double Math.sqrt(double) 1820 // public static double Math.log(double) 1821 // public static double Math.log10(double) 1822 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1823 Node* arg = round_double_node(argument(0)); 1824 Node* n = NULL; 1825 switch (id) { 1826 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1827 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1828 default: fatal_unexpected_iid(id); break; 1829 } 1830 set_result(_gvn.transform(n)); 1831 return true; 1832 } 1833 1834 //------------------------------runtime_math----------------------------- 1835 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1836 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1837 "must be (DD)D or (D)D type"); 1838 1839 // Inputs 1840 Node* a = round_double_node(argument(0)); 1841 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1842 1843 const TypePtr* no_memory_effects = NULL; 1844 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1845 no_memory_effects, 1846 a, top(), b, b ? top() : NULL); 1847 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1848 #ifdef ASSERT 1849 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1850 assert(value_top == top(), "second value must be top"); 1851 #endif 1852 1853 set_result(value); 1854 return true; 1855 } 1856 1857 //------------------------------inline_math_native----------------------------- 1858 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1859 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1860 switch (id) { 1861 // These intrinsics are not properly supported on all hardware 1862 case vmIntrinsics::_dsin: 1863 return StubRoutines::dsin() != NULL ? 1864 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1865 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1866 case vmIntrinsics::_dcos: 1867 return StubRoutines::dcos() != NULL ? 1868 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1869 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1870 case vmIntrinsics::_dtan: 1871 return StubRoutines::dtan() != NULL ? 1872 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1873 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1874 case vmIntrinsics::_dlog: 1875 return StubRoutines::dlog() != NULL ? 1876 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1877 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1878 case vmIntrinsics::_dlog10: 1879 return StubRoutines::dlog10() != NULL ? 1880 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1881 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1882 1883 // These intrinsics are supported on all hardware 1884 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1885 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1886 1887 case vmIntrinsics::_dexp: 1888 return StubRoutines::dexp() != NULL ? 1889 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1890 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1891 case vmIntrinsics::_dpow: { 1892 Node* exp = round_double_node(argument(2)); 1893 const TypeD* d = _gvn.type(exp)->isa_double_constant(); 1894 if (d != NULL && d->getd() == 2.0) { 1895 // Special case: pow(x, 2.0) => x * x 1896 Node* base = round_double_node(argument(0)); 1897 set_result(_gvn.transform(new MulDNode(base, base))); 1898 return true; 1899 } 1900 return StubRoutines::dpow() != NULL ? 1901 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1902 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1903 } 1904 #undef FN_PTR 1905 1906 // These intrinsics are not yet correctly implemented 1907 case vmIntrinsics::_datan2: 1908 return false; 1909 1910 default: 1911 fatal_unexpected_iid(id); 1912 return false; 1913 } 1914 } 1915 1916 static bool is_simple_name(Node* n) { 1917 return (n->req() == 1 // constant 1918 || (n->is_Type() && n->as_Type()->type()->singleton()) 1919 || n->is_Proj() // parameter or return value 1920 || n->is_Phi() // local of some sort 1921 ); 1922 } 1923 1924 //----------------------------inline_notify-----------------------------------* 1925 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1926 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1927 address func; 1928 if (id == vmIntrinsics::_notify) { 1929 func = OptoRuntime::monitor_notify_Java(); 1930 } else { 1931 func = OptoRuntime::monitor_notifyAll_Java(); 1932 } 1933 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1934 make_slow_call_ex(call, env()->Throwable_klass(), false); 1935 return true; 1936 } 1937 1938 1939 //----------------------------inline_min_max----------------------------------- 1940 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1941 set_result(generate_min_max(id, argument(0), argument(1))); 1942 return true; 1943 } 1944 1945 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1946 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1947 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1948 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1949 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1950 1951 { 1952 PreserveJVMState pjvms(this); 1953 PreserveReexecuteState preexecs(this); 1954 jvms()->set_should_reexecute(true); 1955 1956 set_control(slow_path); 1957 set_i_o(i_o()); 1958 1959 uncommon_trap(Deoptimization::Reason_intrinsic, 1960 Deoptimization::Action_none); 1961 } 1962 1963 set_control(fast_path); 1964 set_result(math); 1965 } 1966 1967 template <typename OverflowOp> 1968 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1969 typedef typename OverflowOp::MathOp MathOp; 1970 1971 MathOp* mathOp = new MathOp(arg1, arg2); 1972 Node* operation = _gvn.transform( mathOp ); 1973 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1974 inline_math_mathExact(operation, ofcheck); 1975 return true; 1976 } 1977 1978 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1979 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1980 } 1981 1982 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1983 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1984 } 1985 1986 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1987 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1988 } 1989 1990 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1991 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1992 } 1993 1994 bool LibraryCallKit::inline_math_negateExactI() { 1995 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1996 } 1997 1998 bool LibraryCallKit::inline_math_negateExactL() { 1999 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2000 } 2001 2002 bool LibraryCallKit::inline_math_multiplyExactI() { 2003 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2004 } 2005 2006 bool LibraryCallKit::inline_math_multiplyExactL() { 2007 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2008 } 2009 2010 bool LibraryCallKit::inline_math_multiplyHigh() { 2011 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2)))); 2012 return true; 2013 } 2014 2015 Node* 2016 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2017 // These are the candidate return value: 2018 Node* xvalue = x0; 2019 Node* yvalue = y0; 2020 2021 if (xvalue == yvalue) { 2022 return xvalue; 2023 } 2024 2025 bool want_max = (id == vmIntrinsics::_max); 2026 2027 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2028 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2029 if (txvalue == NULL || tyvalue == NULL) return top(); 2030 // This is not really necessary, but it is consistent with a 2031 // hypothetical MaxINode::Value method: 2032 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2033 2034 // %%% This folding logic should (ideally) be in a different place. 2035 // Some should be inside IfNode, and there to be a more reliable 2036 // transformation of ?: style patterns into cmoves. We also want 2037 // more powerful optimizations around cmove and min/max. 2038 2039 // Try to find a dominating comparison of these guys. 2040 // It can simplify the index computation for Arrays.copyOf 2041 // and similar uses of System.arraycopy. 2042 // First, compute the normalized version of CmpI(x, y). 2043 int cmp_op = Op_CmpI; 2044 Node* xkey = xvalue; 2045 Node* ykey = yvalue; 2046 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2047 if (ideal_cmpxy->is_Cmp()) { 2048 // E.g., if we have CmpI(length - offset, count), 2049 // it might idealize to CmpI(length, count + offset) 2050 cmp_op = ideal_cmpxy->Opcode(); 2051 xkey = ideal_cmpxy->in(1); 2052 ykey = ideal_cmpxy->in(2); 2053 } 2054 2055 // Start by locating any relevant comparisons. 2056 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2057 Node* cmpxy = NULL; 2058 Node* cmpyx = NULL; 2059 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2060 Node* cmp = start_from->fast_out(k); 2061 if (cmp->outcnt() > 0 && // must have prior uses 2062 cmp->in(0) == NULL && // must be context-independent 2063 cmp->Opcode() == cmp_op) { // right kind of compare 2064 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2065 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2066 } 2067 } 2068 2069 const int NCMPS = 2; 2070 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2071 int cmpn; 2072 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2073 if (cmps[cmpn] != NULL) break; // find a result 2074 } 2075 if (cmpn < NCMPS) { 2076 // Look for a dominating test that tells us the min and max. 2077 int depth = 0; // Limit search depth for speed 2078 Node* dom = control(); 2079 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2080 if (++depth >= 100) break; 2081 Node* ifproj = dom; 2082 if (!ifproj->is_Proj()) continue; 2083 Node* iff = ifproj->in(0); 2084 if (!iff->is_If()) continue; 2085 Node* bol = iff->in(1); 2086 if (!bol->is_Bool()) continue; 2087 Node* cmp = bol->in(1); 2088 if (cmp == NULL) continue; 2089 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2090 if (cmps[cmpn] == cmp) break; 2091 if (cmpn == NCMPS) continue; 2092 BoolTest::mask btest = bol->as_Bool()->_test._test; 2093 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2094 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2095 // At this point, we know that 'x btest y' is true. 2096 switch (btest) { 2097 case BoolTest::eq: 2098 // They are proven equal, so we can collapse the min/max. 2099 // Either value is the answer. Choose the simpler. 2100 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2101 return yvalue; 2102 return xvalue; 2103 case BoolTest::lt: // x < y 2104 case BoolTest::le: // x <= y 2105 return (want_max ? yvalue : xvalue); 2106 case BoolTest::gt: // x > y 2107 case BoolTest::ge: // x >= y 2108 return (want_max ? xvalue : yvalue); 2109 default: 2110 break; 2111 } 2112 } 2113 } 2114 2115 // We failed to find a dominating test. 2116 // Let's pick a test that might GVN with prior tests. 2117 Node* best_bol = NULL; 2118 BoolTest::mask best_btest = BoolTest::illegal; 2119 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2120 Node* cmp = cmps[cmpn]; 2121 if (cmp == NULL) continue; 2122 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2123 Node* bol = cmp->fast_out(j); 2124 if (!bol->is_Bool()) continue; 2125 BoolTest::mask btest = bol->as_Bool()->_test._test; 2126 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2127 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2128 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2129 best_bol = bol->as_Bool(); 2130 best_btest = btest; 2131 } 2132 } 2133 } 2134 2135 Node* answer_if_true = NULL; 2136 Node* answer_if_false = NULL; 2137 switch (best_btest) { 2138 default: 2139 if (cmpxy == NULL) 2140 cmpxy = ideal_cmpxy; 2141 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2142 // and fall through: 2143 case BoolTest::lt: // x < y 2144 case BoolTest::le: // x <= y 2145 answer_if_true = (want_max ? yvalue : xvalue); 2146 answer_if_false = (want_max ? xvalue : yvalue); 2147 break; 2148 case BoolTest::gt: // x > y 2149 case BoolTest::ge: // x >= y 2150 answer_if_true = (want_max ? xvalue : yvalue); 2151 answer_if_false = (want_max ? yvalue : xvalue); 2152 break; 2153 } 2154 2155 jint hi, lo; 2156 if (want_max) { 2157 // We can sharpen the minimum. 2158 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2159 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2160 } else { 2161 // We can sharpen the maximum. 2162 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2163 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2164 } 2165 2166 // Use a flow-free graph structure, to avoid creating excess control edges 2167 // which could hinder other optimizations. 2168 // Since Math.min/max is often used with arraycopy, we want 2169 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2170 Node* cmov = CMoveNode::make(NULL, best_bol, 2171 answer_if_false, answer_if_true, 2172 TypeInt::make(lo, hi, widen)); 2173 2174 return _gvn.transform(cmov); 2175 2176 /* 2177 // This is not as desirable as it may seem, since Min and Max 2178 // nodes do not have a full set of optimizations. 2179 // And they would interfere, anyway, with 'if' optimizations 2180 // and with CMoveI canonical forms. 2181 switch (id) { 2182 case vmIntrinsics::_min: 2183 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2184 case vmIntrinsics::_max: 2185 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2186 default: 2187 ShouldNotReachHere(); 2188 } 2189 */ 2190 } 2191 2192 inline int 2193 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) { 2194 const TypePtr* base_type = TypePtr::NULL_PTR; 2195 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2196 if (base_type == NULL) { 2197 // Unknown type. 2198 return Type::AnyPtr; 2199 } else if (base_type == TypePtr::NULL_PTR) { 2200 // Since this is a NULL+long form, we have to switch to a rawptr. 2201 base = _gvn.transform(new CastX2PNode(offset)); 2202 offset = MakeConX(0); 2203 return Type::RawPtr; 2204 } else if (base_type->base() == Type::RawPtr) { 2205 return Type::RawPtr; 2206 } else if (base_type->isa_oopptr()) { 2207 // Base is never null => always a heap address. 2208 if (!TypePtr::NULL_PTR->higher_equal(base_type)) { 2209 return Type::OopPtr; 2210 } 2211 // Offset is small => always a heap address. 2212 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2213 if (offset_type != NULL && 2214 base_type->offset() == 0 && // (should always be?) 2215 offset_type->_lo >= 0 && 2216 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2217 return Type::OopPtr; 2218 } else if (type == T_OBJECT) { 2219 // off heap access to an oop doesn't make any sense. Has to be on 2220 // heap. 2221 return Type::OopPtr; 2222 } 2223 // Otherwise, it might either be oop+off or NULL+addr. 2224 return Type::AnyPtr; 2225 } else { 2226 // No information: 2227 return Type::AnyPtr; 2228 } 2229 } 2230 2231 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type, bool can_cast) { 2232 Node* uncasted_base = base; 2233 int kind = classify_unsafe_addr(uncasted_base, offset, type); 2234 if (kind == Type::RawPtr) { 2235 return basic_plus_adr(top(), uncasted_base, offset); 2236 } else if (kind == Type::AnyPtr) { 2237 assert(base == uncasted_base, "unexpected base change"); 2238 if (can_cast) { 2239 if (!_gvn.type(base)->speculative_maybe_null() && 2240 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2241 // According to profiling, this access is always on 2242 // heap. Casting the base to not null and thus avoiding membars 2243 // around the access should allow better optimizations 2244 Node* null_ctl = top(); 2245 base = null_check_oop(base, &null_ctl, true, true, true); 2246 assert(null_ctl->is_top(), "no null control here"); 2247 return basic_plus_adr(base, offset); 2248 } else if (_gvn.type(base)->speculative_always_null() && 2249 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2250 // According to profiling, this access is always off 2251 // heap. 2252 base = null_assert(base); 2253 Node* raw_base = _gvn.transform(new CastX2PNode(offset)); 2254 offset = MakeConX(0); 2255 return basic_plus_adr(top(), raw_base, offset); 2256 } 2257 } 2258 // We don't know if it's an on heap or off heap access. Fall back 2259 // to raw memory access. 2260 base = access_resolve(base, decorators); 2261 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM)); 2262 return basic_plus_adr(top(), raw, offset); 2263 } else { 2264 assert(base == uncasted_base, "unexpected base change"); 2265 // We know it's an on heap access so base can't be null 2266 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) { 2267 base = must_be_not_null(base, true); 2268 } 2269 return basic_plus_adr(base, offset); 2270 } 2271 } 2272 2273 //--------------------------inline_number_methods----------------------------- 2274 // inline int Integer.numberOfLeadingZeros(int) 2275 // inline int Long.numberOfLeadingZeros(long) 2276 // 2277 // inline int Integer.numberOfTrailingZeros(int) 2278 // inline int Long.numberOfTrailingZeros(long) 2279 // 2280 // inline int Integer.bitCount(int) 2281 // inline int Long.bitCount(long) 2282 // 2283 // inline char Character.reverseBytes(char) 2284 // inline short Short.reverseBytes(short) 2285 // inline int Integer.reverseBytes(int) 2286 // inline long Long.reverseBytes(long) 2287 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2288 Node* arg = argument(0); 2289 Node* n = NULL; 2290 switch (id) { 2291 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2292 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2293 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2294 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2295 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2296 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2297 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2298 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2299 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2300 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2301 default: fatal_unexpected_iid(id); break; 2302 } 2303 set_result(_gvn.transform(n)); 2304 return true; 2305 } 2306 2307 //----------------------------inline_unsafe_access---------------------------- 2308 2309 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2310 // Attempt to infer a sharper value type from the offset and base type. 2311 ciKlass* sharpened_klass = NULL; 2312 2313 // See if it is an instance field, with an object type. 2314 if (alias_type->field() != NULL) { 2315 if (alias_type->field()->type()->is_klass()) { 2316 sharpened_klass = alias_type->field()->type()->as_klass(); 2317 } 2318 } 2319 2320 // See if it is a narrow oop array. 2321 if (adr_type->isa_aryptr()) { 2322 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2323 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2324 if (elem_type != NULL) { 2325 sharpened_klass = elem_type->klass(); 2326 } 2327 } 2328 } 2329 2330 // The sharpened class might be unloaded if there is no class loader 2331 // contraint in place. 2332 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2333 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2334 2335 #ifndef PRODUCT 2336 if (C->print_intrinsics() || C->print_inlining()) { 2337 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2338 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2339 } 2340 #endif 2341 // Sharpen the value type. 2342 return tjp; 2343 } 2344 return NULL; 2345 } 2346 2347 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) { 2348 switch (kind) { 2349 case Relaxed: 2350 return MO_UNORDERED; 2351 case Opaque: 2352 return MO_RELAXED; 2353 case Acquire: 2354 return MO_ACQUIRE; 2355 case Release: 2356 return MO_RELEASE; 2357 case Volatile: 2358 return MO_SEQ_CST; 2359 default: 2360 ShouldNotReachHere(); 2361 return 0; 2362 } 2363 } 2364 2365 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2366 if (callee()->is_static()) return false; // caller must have the capability! 2367 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2368 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2369 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2370 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2371 2372 if (type == T_OBJECT || type == T_ARRAY) { 2373 decorators |= ON_UNKNOWN_OOP_REF; 2374 } 2375 2376 if (unaligned) { 2377 decorators |= C2_UNALIGNED; 2378 } 2379 2380 #ifndef PRODUCT 2381 { 2382 ResourceMark rm; 2383 // Check the signatures. 2384 ciSignature* sig = callee()->signature(); 2385 #ifdef ASSERT 2386 if (!is_store) { 2387 // Object getReference(Object base, int/long offset), etc. 2388 BasicType rtype = sig->return_type()->basic_type(); 2389 assert(rtype == type || (rtype == T_OBJECT && type == T_VALUETYPE), "getter must return the expected value"); 2390 assert(sig->count() == 2 || (type == T_VALUETYPE && sig->count() == 3), "oop getter has 2 or 3 arguments"); 2391 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2392 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2393 } else { 2394 // void putReference(Object base, int/long offset, Object x), etc. 2395 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2396 assert(sig->count() == 3 || (type == T_VALUETYPE && sig->count() == 4), "oop putter has 3 arguments"); 2397 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2398 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2399 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2400 assert(vtype == type || (type == T_VALUETYPE && vtype == T_OBJECT), "putter must accept the expected value"); 2401 } 2402 #endif // ASSERT 2403 } 2404 #endif //PRODUCT 2405 2406 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2407 2408 Node* receiver = argument(0); // type: oop 2409 2410 // Build address expression. 2411 Node* adr; 2412 Node* heap_base_oop = top(); 2413 Node* offset = top(); 2414 Node* val; 2415 2416 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2417 Node* base = argument(1); // type: oop 2418 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2419 offset = argument(2); // type: long 2420 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2421 // to be plain byte offsets, which are also the same as those accepted 2422 // by oopDesc::field_addr. 2423 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2424 "fieldOffset must be byte-scaled"); 2425 2426 ciValueKlass* value_klass = NULL; 2427 if (type == T_VALUETYPE) { 2428 Node* cls = null_check(argument(4)); 2429 if (stopped()) { 2430 return true; 2431 } 2432 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2433 const TypeKlassPtr* kls_t = _gvn.type(kls)->isa_klassptr(); 2434 if (!kls_t->klass_is_exact()) { 2435 return false; 2436 } 2437 ciKlass* klass = kls_t->klass(); 2438 if (!klass->is_valuetype()) { 2439 return false; 2440 } 2441 value_klass = klass->as_value_klass(); 2442 } 2443 2444 receiver = null_check(receiver); 2445 if (stopped()) { 2446 return true; 2447 } 2448 2449 if (base->is_ValueType()) { 2450 ValueTypeNode* vt = base->as_ValueType(); 2451 2452 if (is_store) { 2453 if (!vt->is_allocated(&_gvn) || !_gvn.type(vt)->is_valuetype()->larval()) { 2454 return false; 2455 } 2456 base = vt->get_oop(); 2457 } else { 2458 if (offset->is_Con()) { 2459 long off = find_long_con(offset, 0); 2460 ciValueKlass* vk = _gvn.type(vt)->is_valuetype()->value_klass(); 2461 if ((long)(int)off != off || !vk->contains_field_offset(off)) { 2462 return false; 2463 } 2464 2465 ciField* f = vk->get_non_flattened_field_by_offset((int)off); 2466 2467 if (f != NULL) { 2468 BasicType bt = f->layout_type(); 2469 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2470 bt = T_OBJECT; 2471 } 2472 if (bt == type) { 2473 if (bt != T_VALUETYPE || f->type() == value_klass) { 2474 set_result(vt->field_value_by_offset((int)off, false)); 2475 return true; 2476 } 2477 } 2478 } 2479 } 2480 vt = vt->allocate(this)->as_ValueType(); 2481 base = vt->get_oop(); 2482 } 2483 } 2484 2485 // 32-bit machines ignore the high half! 2486 offset = ConvL2X(offset); 2487 adr = make_unsafe_address(base, offset, is_store ? ACCESS_WRITE : ACCESS_READ, type, kind == Relaxed); 2488 2489 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) { 2490 heap_base_oop = base; 2491 } else if (type == T_OBJECT || (value_klass != NULL && value_klass->has_object_fields())) { 2492 return false; // off-heap oop accesses are not supported 2493 } 2494 2495 // Can base be NULL? Otherwise, always on-heap access. 2496 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base)); 2497 2498 if (!can_access_non_heap) { 2499 decorators |= IN_HEAP; 2500 } 2501 2502 val = is_store ? argument(4 + (type == T_VALUETYPE ? 1 : 0)) : NULL; 2503 2504 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2505 2506 // Try to categorize the address. 2507 Compile::AliasType* alias_type = C->alias_type(adr_type); 2508 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2509 2510 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2511 alias_type->adr_type() == TypeAryPtr::RANGE) { 2512 return false; // not supported 2513 } 2514 2515 bool mismatched = false; 2516 BasicType bt = T_ILLEGAL; 2517 ciField* field = NULL; 2518 if (adr_type->isa_instptr()) { 2519 const TypeInstPtr* instptr = adr_type->is_instptr(); 2520 ciInstanceKlass* k = instptr->klass()->as_instance_klass(); 2521 int off = instptr->offset(); 2522 if (instptr->const_oop() != NULL && 2523 instptr->klass() == ciEnv::current()->Class_klass() && 2524 instptr->offset() >= (instptr->klass()->as_instance_klass()->size_helper() * wordSize)) { 2525 k = instptr->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 2526 field = k->get_field_by_offset(off, true); 2527 } else { 2528 field = k->get_non_flattened_field_by_offset(off); 2529 } 2530 if (field != NULL) { 2531 bt = field->layout_type(); 2532 } 2533 assert(bt == alias_type->basic_type() || bt == T_VALUETYPE, "should match"); 2534 if (field != NULL && bt == T_VALUETYPE && !field->is_flattened()) { 2535 bt = T_OBJECT; 2536 } 2537 } else { 2538 bt = alias_type->basic_type(); 2539 } 2540 2541 if (bt != T_ILLEGAL) { 2542 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2543 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2544 // Alias type doesn't differentiate between byte[] and boolean[]). 2545 // Use address type to get the element type. 2546 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2547 } 2548 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2549 // accessing an array field with getReference is not a mismatch 2550 bt = T_OBJECT; 2551 } 2552 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2553 // Don't intrinsify mismatched object accesses 2554 return false; 2555 } 2556 mismatched = (bt != type); 2557 } else if (alias_type->adr_type()->isa_oopptr()) { 2558 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched 2559 } 2560 2561 if (type == T_VALUETYPE) { 2562 if (adr_type->isa_instptr()) { 2563 if (field == NULL || field->type() != value_klass) { 2564 mismatched = true; 2565 } 2566 } else if (adr_type->isa_aryptr()) { 2567 const Type* elem = adr_type->is_aryptr()->elem(); 2568 if (!elem->isa_valuetype()) { 2569 mismatched = true; 2570 } else if (elem->is_valuetype()->value_klass() != value_klass) { 2571 mismatched = true; 2572 } 2573 } 2574 if (is_store) { 2575 const Type* val_t = _gvn.type(val); 2576 if (!val_t->isa_valuetype() || 2577 val_t->is_valuetype()->value_klass() != value_klass) { 2578 return false; 2579 } 2580 } 2581 } 2582 2583 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2584 2585 if (mismatched) { 2586 decorators |= C2_MISMATCHED; 2587 } 2588 2589 // First guess at the value type. 2590 const Type *value_type = Type::get_const_basic_type(type); 2591 2592 // Figure out the memory ordering. 2593 decorators |= mo_decorator_for_access_kind(kind); 2594 2595 if (!is_store) { 2596 if (type == T_OBJECT) { 2597 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2598 if (tjp != NULL) { 2599 value_type = tjp; 2600 } 2601 } else if (type == T_VALUETYPE) { 2602 value_type = NULL; 2603 } 2604 } 2605 2606 // Heap pointers get a null-check from the interpreter, 2607 // as a courtesy. However, this is not guaranteed by Unsafe, 2608 // and it is not possible to fully distinguish unintended nulls 2609 // from intended ones in this API. 2610 2611 if (!is_store) { 2612 Node* p = NULL; 2613 // Try to constant fold a load from a constant field 2614 2615 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { 2616 // final or stable field 2617 p = make_constant_from_field(field, heap_base_oop); 2618 } 2619 2620 if (p == NULL) { // Could not constant fold the load 2621 if (type == T_VALUETYPE) { 2622 if (adr_type->isa_instptr() && !mismatched) { 2623 ciInstanceKlass* holder = adr_type->is_instptr()->klass()->as_instance_klass(); 2624 int offset = adr_type->is_instptr()->offset(); 2625 p = ValueTypeNode::make_from_flattened(this, value_klass, base, base, holder, offset, decorators); 2626 } else { 2627 p = ValueTypeNode::make_from_flattened(this, value_klass, base, adr, NULL, 0, decorators); 2628 } 2629 } else { 2630 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators); 2631 } 2632 // Normalize the value returned by getBoolean in the following cases 2633 if (type == T_BOOLEAN && 2634 (mismatched || 2635 heap_base_oop == top() || // - heap_base_oop is NULL or 2636 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL 2637 // and the unsafe access is made to large offset 2638 // (i.e., larger than the maximum offset necessary for any 2639 // field access) 2640 ) { 2641 IdealKit ideal = IdealKit(this); 2642 #define __ ideal. 2643 IdealVariable normalized_result(ideal); 2644 __ declarations_done(); 2645 __ set(normalized_result, p); 2646 __ if_then(p, BoolTest::ne, ideal.ConI(0)); 2647 __ set(normalized_result, ideal.ConI(1)); 2648 ideal.end_if(); 2649 final_sync(ideal); 2650 p = __ value(normalized_result); 2651 #undef __ 2652 } 2653 } 2654 if (type == T_ADDRESS) { 2655 p = gvn().transform(new CastP2XNode(NULL, p)); 2656 p = ConvX2UL(p); 2657 } 2658 if (field != NULL && field->is_flattenable() && !field->is_flattened()) { 2659 // Load a non-flattened but flattenable value type from memory 2660 if (value_type->value_klass()->is_scalarizable()) { 2661 p = ValueTypeNode::make_from_oop(this, p, value_type->value_klass()); 2662 } else { 2663 p = null2default(p, value_type->value_klass()); 2664 } 2665 } 2666 // The load node has the control of the preceding MemBarCPUOrder. All 2667 // following nodes will have the control of the MemBarCPUOrder inserted at 2668 // the end of this method. So, pushing the load onto the stack at a later 2669 // point is fine. 2670 set_result(p); 2671 } else { 2672 if (bt == T_ADDRESS) { 2673 // Repackage the long as a pointer. 2674 val = ConvL2X(val); 2675 val = gvn().transform(new CastX2PNode(val)); 2676 } 2677 if (type == T_VALUETYPE) { 2678 if (adr_type->isa_instptr() && !mismatched) { 2679 ciInstanceKlass* holder = adr_type->is_instptr()->klass()->as_instance_klass(); 2680 int offset = adr_type->is_instptr()->offset(); 2681 val->as_ValueType()->store_flattened(this, base, base, holder, offset, decorators); 2682 } else { 2683 val->as_ValueType()->store_flattened(this, base, adr, NULL, 0, decorators); 2684 } 2685 } else { 2686 access_store_at(heap_base_oop, adr, adr_type, val, value_type, type, decorators); 2687 } 2688 } 2689 2690 if (argument(1)->is_ValueType() && is_store) { 2691 Node* value = ValueTypeNode::make_from_oop(this, base, _gvn.type(base)->value_klass()); 2692 value = value->as_ValueType()->make_larval(this, false); 2693 replace_in_map(argument(1), value); 2694 } 2695 2696 return true; 2697 } 2698 2699 bool LibraryCallKit::inline_unsafe_make_private_buffer() { 2700 Node* receiver = argument(0); 2701 Node* value = argument(1); 2702 2703 receiver = null_check(receiver); 2704 if (stopped()) { 2705 return true; 2706 } 2707 2708 if (!value->is_ValueType()) { 2709 return false; 2710 } 2711 2712 set_result(value->as_ValueType()->make_larval(this, true)); 2713 2714 return true; 2715 } 2716 2717 bool LibraryCallKit::inline_unsafe_finish_private_buffer() { 2718 Node* receiver = argument(0); 2719 Node* buffer = argument(1); 2720 2721 receiver = null_check(receiver); 2722 if (stopped()) { 2723 return true; 2724 } 2725 2726 if (!buffer->is_ValueType()) { 2727 return false; 2728 } 2729 2730 ValueTypeNode* vt = buffer->as_ValueType(); 2731 if (!vt->is_allocated(&_gvn) || !_gvn.type(vt)->is_valuetype()->larval()) { 2732 return false; 2733 } 2734 2735 set_result(vt->finish_larval(this)); 2736 2737 return true; 2738 } 2739 2740 //----------------------------inline_unsafe_load_store---------------------------- 2741 // This method serves a couple of different customers (depending on LoadStoreKind): 2742 // 2743 // LS_cmp_swap: 2744 // 2745 // boolean compareAndSetReference(Object o, long offset, Object expected, Object x); 2746 // boolean compareAndSetInt( Object o, long offset, int expected, int x); 2747 // boolean compareAndSetLong( Object o, long offset, long expected, long x); 2748 // 2749 // LS_cmp_swap_weak: 2750 // 2751 // boolean weakCompareAndSetReference( Object o, long offset, Object expected, Object x); 2752 // boolean weakCompareAndSetReferencePlain( Object o, long offset, Object expected, Object x); 2753 // boolean weakCompareAndSetReferenceAcquire(Object o, long offset, Object expected, Object x); 2754 // boolean weakCompareAndSetReferenceRelease(Object o, long offset, Object expected, Object x); 2755 // 2756 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x); 2757 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x); 2758 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x); 2759 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x); 2760 // 2761 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x); 2762 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x); 2763 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x); 2764 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x); 2765 // 2766 // LS_cmp_exchange: 2767 // 2768 // Object compareAndExchangeReferenceVolatile(Object o, long offset, Object expected, Object x); 2769 // Object compareAndExchangeReferenceAcquire( Object o, long offset, Object expected, Object x); 2770 // Object compareAndExchangeReferenceRelease( Object o, long offset, Object expected, Object x); 2771 // 2772 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2773 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2774 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2775 // 2776 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2777 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2778 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2779 // 2780 // LS_get_add: 2781 // 2782 // int getAndAddInt( Object o, long offset, int delta) 2783 // long getAndAddLong(Object o, long offset, long delta) 2784 // 2785 // LS_get_set: 2786 // 2787 // int getAndSet(Object o, long offset, int newValue) 2788 // long getAndSet(Object o, long offset, long newValue) 2789 // Object getAndSet(Object o, long offset, Object newValue) 2790 // 2791 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2792 // This basic scheme here is the same as inline_unsafe_access, but 2793 // differs in enough details that combining them would make the code 2794 // overly confusing. (This is a true fact! I originally combined 2795 // them, but even I was confused by it!) As much code/comments as 2796 // possible are retained from inline_unsafe_access though to make 2797 // the correspondences clearer. - dl 2798 2799 if (callee()->is_static()) return false; // caller must have the capability! 2800 2801 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2802 decorators |= mo_decorator_for_access_kind(access_kind); 2803 2804 #ifndef PRODUCT 2805 BasicType rtype; 2806 { 2807 ResourceMark rm; 2808 // Check the signatures. 2809 ciSignature* sig = callee()->signature(); 2810 rtype = sig->return_type()->basic_type(); 2811 switch(kind) { 2812 case LS_get_add: 2813 case LS_get_set: { 2814 // Check the signatures. 2815 #ifdef ASSERT 2816 assert(rtype == type, "get and set must return the expected type"); 2817 assert(sig->count() == 3, "get and set has 3 arguments"); 2818 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2819 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2820 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2821 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2822 #endif // ASSERT 2823 break; 2824 } 2825 case LS_cmp_swap: 2826 case LS_cmp_swap_weak: { 2827 // Check the signatures. 2828 #ifdef ASSERT 2829 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2830 assert(sig->count() == 4, "CAS has 4 arguments"); 2831 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2832 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2833 #endif // ASSERT 2834 break; 2835 } 2836 case LS_cmp_exchange: { 2837 // Check the signatures. 2838 #ifdef ASSERT 2839 assert(rtype == type, "CAS must return the expected type"); 2840 assert(sig->count() == 4, "CAS has 4 arguments"); 2841 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2842 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2843 #endif // ASSERT 2844 break; 2845 } 2846 default: 2847 ShouldNotReachHere(); 2848 } 2849 } 2850 #endif //PRODUCT 2851 2852 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2853 2854 // Get arguments: 2855 Node* receiver = NULL; 2856 Node* base = NULL; 2857 Node* offset = NULL; 2858 Node* oldval = NULL; 2859 Node* newval = NULL; 2860 switch(kind) { 2861 case LS_cmp_swap: 2862 case LS_cmp_swap_weak: 2863 case LS_cmp_exchange: { 2864 const bool two_slot_type = type2size[type] == 2; 2865 receiver = argument(0); // type: oop 2866 base = argument(1); // type: oop 2867 offset = argument(2); // type: long 2868 oldval = argument(4); // type: oop, int, or long 2869 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2870 break; 2871 } 2872 case LS_get_add: 2873 case LS_get_set: { 2874 receiver = argument(0); // type: oop 2875 base = argument(1); // type: oop 2876 offset = argument(2); // type: long 2877 oldval = NULL; 2878 newval = argument(4); // type: oop, int, or long 2879 break; 2880 } 2881 default: 2882 ShouldNotReachHere(); 2883 } 2884 2885 // Build field offset expression. 2886 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2887 // to be plain byte offsets, which are also the same as those accepted 2888 // by oopDesc::field_addr. 2889 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2890 // 32-bit machines ignore the high half of long offsets 2891 offset = ConvL2X(offset); 2892 Node* adr = make_unsafe_address(base, offset, ACCESS_WRITE | ACCESS_READ, type, false); 2893 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2894 2895 Compile::AliasType* alias_type = C->alias_type(adr_type); 2896 BasicType bt = alias_type->basic_type(); 2897 if (bt != T_ILLEGAL && 2898 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2899 // Don't intrinsify mismatched object accesses. 2900 return false; 2901 } 2902 2903 // For CAS, unlike inline_unsafe_access, there seems no point in 2904 // trying to refine types. Just use the coarse types here. 2905 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2906 const Type *value_type = Type::get_const_basic_type(type); 2907 2908 switch (kind) { 2909 case LS_get_set: 2910 case LS_cmp_exchange: { 2911 if (type == T_OBJECT) { 2912 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2913 if (tjp != NULL) { 2914 value_type = tjp; 2915 } 2916 } 2917 break; 2918 } 2919 case LS_cmp_swap: 2920 case LS_cmp_swap_weak: 2921 case LS_get_add: 2922 break; 2923 default: 2924 ShouldNotReachHere(); 2925 } 2926 2927 // Null check receiver. 2928 receiver = null_check(receiver); 2929 if (stopped()) { 2930 return true; 2931 } 2932 2933 int alias_idx = C->get_alias_index(adr_type); 2934 2935 if (type == T_OBJECT || type == T_ARRAY) { 2936 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF; 2937 2938 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2939 // could be delayed during Parse (for example, in adjust_map_after_if()). 2940 // Execute transformation here to avoid barrier generation in such case. 2941 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2942 newval = _gvn.makecon(TypePtr::NULL_PTR); 2943 2944 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) { 2945 // Refine the value to a null constant, when it is known to be null 2946 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2947 } 2948 } 2949 2950 Node* result = NULL; 2951 switch (kind) { 2952 case LS_cmp_exchange: { 2953 result = access_atomic_cmpxchg_val_at(base, adr, adr_type, alias_idx, 2954 oldval, newval, value_type, type, decorators); 2955 break; 2956 } 2957 case LS_cmp_swap_weak: 2958 decorators |= C2_WEAK_CMPXCHG; 2959 case LS_cmp_swap: { 2960 result = access_atomic_cmpxchg_bool_at(base, adr, adr_type, alias_idx, 2961 oldval, newval, value_type, type, decorators); 2962 break; 2963 } 2964 case LS_get_set: { 2965 result = access_atomic_xchg_at(base, adr, adr_type, alias_idx, 2966 newval, value_type, type, decorators); 2967 break; 2968 } 2969 case LS_get_add: { 2970 result = access_atomic_add_at(base, adr, adr_type, alias_idx, 2971 newval, value_type, type, decorators); 2972 break; 2973 } 2974 default: 2975 ShouldNotReachHere(); 2976 } 2977 2978 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2979 set_result(result); 2980 return true; 2981 } 2982 2983 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2984 // Regardless of form, don't allow previous ld/st to move down, 2985 // then issue acquire, release, or volatile mem_bar. 2986 insert_mem_bar(Op_MemBarCPUOrder); 2987 switch(id) { 2988 case vmIntrinsics::_loadFence: 2989 insert_mem_bar(Op_LoadFence); 2990 return true; 2991 case vmIntrinsics::_storeFence: 2992 insert_mem_bar(Op_StoreFence); 2993 return true; 2994 case vmIntrinsics::_fullFence: 2995 insert_mem_bar(Op_MemBarVolatile); 2996 return true; 2997 default: 2998 fatal_unexpected_iid(id); 2999 return false; 3000 } 3001 } 3002 3003 bool LibraryCallKit::inline_onspinwait() { 3004 insert_mem_bar(Op_OnSpinWait); 3005 return true; 3006 } 3007 3008 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3009 if (!kls->is_Con()) { 3010 return true; 3011 } 3012 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3013 if (klsptr == NULL) { 3014 return true; 3015 } 3016 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3017 // don't need a guard for a klass that is already initialized 3018 return !ik->is_initialized(); 3019 } 3020 3021 //----------------------------inline_unsafe_allocate--------------------------- 3022 // public native Object Unsafe.allocateInstance(Class<?> cls); 3023 bool LibraryCallKit::inline_unsafe_allocate() { 3024 if (callee()->is_static()) return false; // caller must have the capability! 3025 3026 null_check_receiver(); // null-check, then ignore 3027 Node* cls = null_check(argument(1)); 3028 if (stopped()) return true; 3029 3030 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3031 kls = null_check(kls); 3032 if (stopped()) return true; // argument was like int.class 3033 3034 Node* test = NULL; 3035 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3036 // Note: The argument might still be an illegal value like 3037 // Serializable.class or Object[].class. The runtime will handle it. 3038 // But we must make an explicit check for initialization. 3039 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3040 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3041 // can generate code to load it as unsigned byte. 3042 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3043 Node* bits = intcon(InstanceKlass::fully_initialized); 3044 test = _gvn.transform(new SubINode(inst, bits)); 3045 // The 'test' is non-zero if we need to take a slow path. 3046 } 3047 3048 Node* obj = new_instance(kls, test); 3049 set_result(obj); 3050 return true; 3051 } 3052 3053 //------------------------inline_native_time_funcs-------------- 3054 // inline code for System.currentTimeMillis() and System.nanoTime() 3055 // these have the same type and signature 3056 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3057 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3058 const TypePtr* no_memory_effects = NULL; 3059 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3060 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3061 #ifdef ASSERT 3062 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3063 assert(value_top == top(), "second value must be top"); 3064 #endif 3065 set_result(value); 3066 return true; 3067 } 3068 3069 #ifdef JFR_HAVE_INTRINSICS 3070 3071 /* 3072 * oop -> myklass 3073 * myklass->trace_id |= USED 3074 * return myklass->trace_id & ~0x3 3075 */ 3076 bool LibraryCallKit::inline_native_classID() { 3077 Node* cls = null_check(argument(0), T_OBJECT); 3078 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3079 kls = null_check(kls, T_OBJECT); 3080 3081 ByteSize offset = KLASS_TRACE_ID_OFFSET; 3082 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3083 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3084 3085 Node* clsused = longcon(0x01l); // set the class bit 3086 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3087 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3088 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3089 3090 #ifdef TRACE_ID_META_BITS 3091 Node* mbits = longcon(~TRACE_ID_META_BITS); 3092 tvalue = _gvn.transform(new AndLNode(tvalue, mbits)); 3093 #endif 3094 #ifdef TRACE_ID_SHIFT 3095 Node* cbits = intcon(TRACE_ID_SHIFT); 3096 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits)); 3097 #endif 3098 3099 set_result(tvalue); 3100 return true; 3101 3102 } 3103 3104 bool LibraryCallKit::inline_native_getEventWriter() { 3105 Node* tls_ptr = _gvn.transform(new ThreadLocalNode()); 3106 3107 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr, 3108 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)); 3109 3110 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered); 3111 3112 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) ); 3113 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) ); 3114 3115 IfNode* iff_jobj_null = 3116 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN); 3117 3118 enum { _normal_path = 1, 3119 _null_path = 2, 3120 PATH_LIMIT }; 3121 3122 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3123 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM); 3124 3125 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null)); 3126 result_rgn->init_req(_null_path, jobj_is_null); 3127 result_val->init_req(_null_path, null()); 3128 3129 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null)); 3130 set_control(jobj_is_not_null); 3131 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT, 3132 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD); 3133 result_rgn->init_req(_normal_path, control()); 3134 result_val->init_req(_normal_path, res); 3135 3136 set_result(result_rgn, result_val); 3137 3138 return true; 3139 } 3140 3141 #endif // JFR_HAVE_INTRINSICS 3142 3143 //------------------------inline_native_currentThread------------------ 3144 bool LibraryCallKit::inline_native_currentThread() { 3145 Node* junk = NULL; 3146 set_result(generate_current_thread(junk)); 3147 return true; 3148 } 3149 3150 //------------------------inline_native_isInterrupted------------------ 3151 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3152 bool LibraryCallKit::inline_native_isInterrupted() { 3153 // Add a fast path to t.isInterrupted(clear_int): 3154 // (t == Thread.current() && 3155 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3156 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3157 // So, in the common case that the interrupt bit is false, 3158 // we avoid making a call into the VM. Even if the interrupt bit 3159 // is true, if the clear_int argument is false, we avoid the VM call. 3160 // However, if the receiver is not currentThread, we must call the VM, 3161 // because there must be some locking done around the operation. 3162 3163 // We only go to the fast case code if we pass two guards. 3164 // Paths which do not pass are accumulated in the slow_region. 3165 3166 enum { 3167 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3168 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3169 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3170 PATH_LIMIT 3171 }; 3172 3173 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3174 // out of the function. 3175 insert_mem_bar(Op_MemBarCPUOrder); 3176 3177 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3178 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3179 3180 RegionNode* slow_region = new RegionNode(1); 3181 record_for_igvn(slow_region); 3182 3183 // (a) Receiving thread must be the current thread. 3184 Node* rec_thr = argument(0); 3185 Node* tls_ptr = NULL; 3186 Node* cur_thr = generate_current_thread(tls_ptr); 3187 3188 // Resolve oops to stable for CmpP below. 3189 cur_thr = access_resolve(cur_thr, 0); 3190 rec_thr = access_resolve(rec_thr, 0); 3191 3192 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3193 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3194 3195 generate_slow_guard(bol_thr, slow_region); 3196 3197 // (b) Interrupt bit on TLS must be false. 3198 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3199 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3200 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3201 3202 // Set the control input on the field _interrupted read to prevent it floating up. 3203 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3204 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3205 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3206 3207 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3208 3209 // First fast path: if (!TLS._interrupted) return false; 3210 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3211 result_rgn->init_req(no_int_result_path, false_bit); 3212 result_val->init_req(no_int_result_path, intcon(0)); 3213 3214 // drop through to next case 3215 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3216 3217 #ifndef _WINDOWS 3218 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3219 Node* clr_arg = argument(1); 3220 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3221 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3222 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3223 3224 // Second fast path: ... else if (!clear_int) return true; 3225 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3226 result_rgn->init_req(no_clear_result_path, false_arg); 3227 result_val->init_req(no_clear_result_path, intcon(1)); 3228 3229 // drop through to next case 3230 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3231 #else 3232 // To return true on Windows you must read the _interrupted field 3233 // and check the event state i.e. take the slow path. 3234 #endif // _WINDOWS 3235 3236 // (d) Otherwise, go to the slow path. 3237 slow_region->add_req(control()); 3238 set_control( _gvn.transform(slow_region)); 3239 3240 if (stopped()) { 3241 // There is no slow path. 3242 result_rgn->init_req(slow_result_path, top()); 3243 result_val->init_req(slow_result_path, top()); 3244 } else { 3245 // non-virtual because it is a private non-static 3246 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3247 3248 Node* slow_val = set_results_for_java_call(slow_call); 3249 // this->control() comes from set_results_for_java_call 3250 3251 Node* fast_io = slow_call->in(TypeFunc::I_O); 3252 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3253 3254 // These two phis are pre-filled with copies of of the fast IO and Memory 3255 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3256 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3257 3258 result_rgn->init_req(slow_result_path, control()); 3259 result_io ->init_req(slow_result_path, i_o()); 3260 result_mem->init_req(slow_result_path, reset_memory()); 3261 result_val->init_req(slow_result_path, slow_val); 3262 3263 set_all_memory(_gvn.transform(result_mem)); 3264 set_i_o( _gvn.transform(result_io)); 3265 } 3266 3267 C->set_has_split_ifs(true); // Has chance for split-if optimization 3268 set_result(result_rgn, result_val); 3269 return true; 3270 } 3271 3272 //-----------------------load_klass_from_mirror_common------------------------- 3273 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3274 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3275 // and branch to the given path on the region. 3276 // If never_see_null, take an uncommon trap on null, so we can optimistically 3277 // compile for the non-null case. 3278 // If the region is NULL, force never_see_null = true. 3279 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3280 bool never_see_null, 3281 RegionNode* region, 3282 int null_path, 3283 int offset) { 3284 if (region == NULL) never_see_null = true; 3285 Node* p = basic_plus_adr(mirror, offset); 3286 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3287 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3288 Node* null_ctl = top(); 3289 kls = null_check_oop(kls, &null_ctl, never_see_null); 3290 if (region != NULL) { 3291 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3292 region->init_req(null_path, null_ctl); 3293 } else { 3294 assert(null_ctl == top(), "no loose ends"); 3295 } 3296 return kls; 3297 } 3298 3299 //--------------------(inline_native_Class_query helpers)--------------------- 3300 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3301 // Fall through if (mods & mask) == bits, take the guard otherwise. 3302 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3303 // Branch around if the given klass has the given modifier bit set. 3304 // Like generate_guard, adds a new path onto the region. 3305 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3306 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3307 Node* mask = intcon(modifier_mask); 3308 Node* bits = intcon(modifier_bits); 3309 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3310 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3311 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3312 return generate_fair_guard(bol, region); 3313 } 3314 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3315 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3316 } 3317 3318 Node* LibraryCallKit::generate_value_guard(Node* kls, RegionNode* region) { 3319 return generate_access_flags_guard(kls, JVM_ACC_VALUE, 0, region); 3320 } 3321 3322 //-------------------------inline_native_Class_query------------------- 3323 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3324 const Type* return_type = TypeInt::BOOL; 3325 Node* prim_return_value = top(); // what happens if it's a primitive class? 3326 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3327 bool expect_prim = false; // most of these guys expect to work on refs 3328 3329 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3330 3331 Node* mirror = argument(0); 3332 Node* obj = top(); 3333 3334 switch (id) { 3335 case vmIntrinsics::_isInstance: 3336 // nothing is an instance of a primitive type 3337 prim_return_value = intcon(0); 3338 obj = argument(1); 3339 break; 3340 case vmIntrinsics::_getModifiers: 3341 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3342 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3343 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3344 break; 3345 case vmIntrinsics::_isInterface: 3346 prim_return_value = intcon(0); 3347 break; 3348 case vmIntrinsics::_isArray: 3349 prim_return_value = intcon(0); 3350 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3351 break; 3352 case vmIntrinsics::_isPrimitive: 3353 prim_return_value = intcon(1); 3354 expect_prim = true; // obviously 3355 break; 3356 case vmIntrinsics::_getSuperclass: 3357 prim_return_value = null(); 3358 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3359 break; 3360 case vmIntrinsics::_getClassAccessFlags: 3361 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3362 return_type = TypeInt::INT; // not bool! 6297094 3363 break; 3364 default: 3365 fatal_unexpected_iid(id); 3366 break; 3367 } 3368 3369 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3370 if (mirror_con == NULL) return false; // cannot happen? 3371 3372 #ifndef PRODUCT 3373 if (C->print_intrinsics() || C->print_inlining()) { 3374 ciType* k = mirror_con->java_mirror_type(); 3375 if (k) { 3376 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3377 k->print_name(); 3378 tty->cr(); 3379 } 3380 } 3381 #endif 3382 3383 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3384 RegionNode* region = new RegionNode(PATH_LIMIT); 3385 record_for_igvn(region); 3386 PhiNode* phi = new PhiNode(region, return_type); 3387 3388 // The mirror will never be null of Reflection.getClassAccessFlags, however 3389 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3390 // if it is. See bug 4774291. 3391 3392 // For Reflection.getClassAccessFlags(), the null check occurs in 3393 // the wrong place; see inline_unsafe_access(), above, for a similar 3394 // situation. 3395 mirror = null_check(mirror); 3396 // If mirror or obj is dead, only null-path is taken. 3397 if (stopped()) return true; 3398 3399 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3400 3401 // Now load the mirror's klass metaobject, and null-check it. 3402 // Side-effects region with the control path if the klass is null. 3403 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3404 // If kls is null, we have a primitive mirror. 3405 phi->init_req(_prim_path, prim_return_value); 3406 if (stopped()) { set_result(region, phi); return true; } 3407 bool safe_for_replace = (region->in(_prim_path) == top()); 3408 3409 Node* p; // handy temp 3410 Node* null_ctl; 3411 3412 // Now that we have the non-null klass, we can perform the real query. 3413 // For constant classes, the query will constant-fold in LoadNode::Value. 3414 Node* query_value = top(); 3415 switch (id) { 3416 case vmIntrinsics::_isInstance: 3417 // nothing is an instance of a primitive type 3418 query_value = gen_instanceof(obj, kls, safe_for_replace); 3419 break; 3420 3421 case vmIntrinsics::_getModifiers: 3422 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3423 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3424 break; 3425 3426 case vmIntrinsics::_isInterface: 3427 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3428 if (generate_interface_guard(kls, region) != NULL) 3429 // A guard was added. If the guard is taken, it was an interface. 3430 phi->add_req(intcon(1)); 3431 // If we fall through, it's a plain class. 3432 query_value = intcon(0); 3433 break; 3434 3435 case vmIntrinsics::_isArray: 3436 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3437 if (generate_array_guard(kls, region) != NULL) 3438 // A guard was added. If the guard is taken, it was an array. 3439 phi->add_req(intcon(1)); 3440 // If we fall through, it's a plain class. 3441 query_value = intcon(0); 3442 break; 3443 3444 case vmIntrinsics::_isPrimitive: 3445 query_value = intcon(0); // "normal" path produces false 3446 break; 3447 3448 case vmIntrinsics::_getSuperclass: 3449 // The rules here are somewhat unfortunate, but we can still do better 3450 // with random logic than with a JNI call. 3451 // Interfaces store null or Object as _super, but must report null. 3452 // Arrays store an intermediate super as _super, but must report Object. 3453 // Other types can report the actual _super. 3454 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3455 if (generate_interface_guard(kls, region) != NULL) 3456 // A guard was added. If the guard is taken, it was an interface. 3457 phi->add_req(null()); 3458 if (generate_array_guard(kls, region) != NULL) 3459 // A guard was added. If the guard is taken, it was an array. 3460 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3461 // If we fall through, it's a plain class. Get its _super. 3462 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3463 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3464 null_ctl = top(); 3465 kls = null_check_oop(kls, &null_ctl); 3466 if (null_ctl != top()) { 3467 // If the guard is taken, Object.superClass is null (both klass and mirror). 3468 region->add_req(null_ctl); 3469 phi ->add_req(null()); 3470 } 3471 if (!stopped()) { 3472 query_value = load_mirror_from_klass(kls); 3473 } 3474 break; 3475 3476 case vmIntrinsics::_getClassAccessFlags: 3477 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3478 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3479 break; 3480 3481 default: 3482 fatal_unexpected_iid(id); 3483 break; 3484 } 3485 3486 // Fall-through is the normal case of a query to a real class. 3487 phi->init_req(1, query_value); 3488 region->init_req(1, control()); 3489 3490 C->set_has_split_ifs(true); // Has chance for split-if optimization 3491 set_result(region, phi); 3492 return true; 3493 } 3494 3495 //-------------------------inline_value_Class_conversion------------------- 3496 // public Class<T> java.lang.Class.asBoxType(); 3497 // public Class<T> java.lang.Class.asValueType() 3498 bool LibraryCallKit::inline_value_Class_conversion(vmIntrinsics::ID id) { 3499 Node* mirror = argument(0); // Receiver Class 3500 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3501 if (mirror_con == NULL) { 3502 return false; 3503 } 3504 3505 bool is_val_type = false; 3506 ciType* tm = mirror_con->java_mirror_type(&is_val_type); 3507 if (tm != NULL && tm->is_valuetype()) { 3508 Node* result = mirror; 3509 if (id == vmIntrinsics::_asValueType && !is_val_type) { 3510 result = _gvn.makecon(TypeInstPtr::make(tm->as_value_klass()->value_mirror_instance())); 3511 } else if (id == vmIntrinsics::_asBoxType && is_val_type) { 3512 result = _gvn.makecon(TypeInstPtr::make(tm->as_value_klass()->box_mirror_instance())); 3513 } 3514 set_result(result); 3515 return true; 3516 } 3517 return false; 3518 } 3519 3520 //-------------------------inline_Class_cast------------------- 3521 bool LibraryCallKit::inline_Class_cast() { 3522 Node* mirror = argument(0); // Class 3523 Node* obj = argument(1); 3524 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3525 if (mirror_con == NULL) { 3526 return false; // dead path (mirror->is_top()). 3527 } 3528 if (obj == NULL || obj->is_top()) { 3529 return false; // dead path 3530 } 3531 3532 ciKlass* obj_klass = NULL; 3533 if (obj->is_ValueType()) { 3534 const TypeValueType* tvt = _gvn.type(obj)->is_valuetype(); 3535 obj_klass = tvt->value_klass(); 3536 } else { 3537 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3538 if (tp != NULL) { 3539 obj_klass = tp->klass(); 3540 } 3541 } 3542 3543 // First, see if Class.cast() can be folded statically. 3544 // java_mirror_type() returns non-null for compile-time Class constants. 3545 bool is_val_type = false; 3546 ciType* tm = mirror_con->java_mirror_type(&is_val_type); 3547 if (!obj->is_ValueType() && is_val_type) { 3548 obj = null_check(obj); 3549 } 3550 if (tm != NULL && tm->is_klass() && obj_klass != NULL) { 3551 if (!obj_klass->is_loaded()) { 3552 // Don't use intrinsic when class is not loaded. 3553 return false; 3554 } else { 3555 int static_res = C->static_subtype_check(tm->as_klass(), obj_klass); 3556 if (static_res == Compile::SSC_always_true) { 3557 // isInstance() is true - fold the code. 3558 set_result(obj); 3559 return true; 3560 } else if (static_res == Compile::SSC_always_false) { 3561 // Don't use intrinsic, have to throw ClassCastException. 3562 // If the reference is null, the non-intrinsic bytecode will 3563 // be optimized appropriately. 3564 return false; 3565 } 3566 } 3567 } 3568 3569 // Bailout intrinsic and do normal inlining if exception path is frequent. 3570 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3571 return false; 3572 } 3573 3574 // Generate dynamic checks. 3575 // Class.cast() is java implementation of _checkcast bytecode. 3576 // Do checkcast (Parse::do_checkcast()) optimizations here. 3577 3578 mirror = null_check(mirror); 3579 // If mirror is dead, only null-path is taken. 3580 if (stopped()) { 3581 return true; 3582 } 3583 3584 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3585 enum { _bad_type_path = 1, _prim_path = 2, _npe_path = 3, PATH_LIMIT }; 3586 RegionNode* region = new RegionNode(PATH_LIMIT); 3587 record_for_igvn(region); 3588 3589 // Now load the mirror's klass metaobject, and null-check it. 3590 // If kls is null, we have a primitive mirror and 3591 // nothing is an instance of a primitive type. 3592 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3593 3594 Node* res = top(); 3595 if (!stopped()) { 3596 // TODO move this into do_checkcast? 3597 if (EnableValhalla && !obj->is_ValueType() && !is_val_type) { 3598 // Check if (mirror == value_mirror && obj == null) 3599 RegionNode* r = new RegionNode(3); 3600 Node* p = basic_plus_adr(mirror, java_lang_Class::value_mirror_offset_in_bytes()); 3601 Node* value_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR, T_OBJECT, IN_HEAP); 3602 Node* cmp = _gvn.transform(new CmpPNode(mirror, value_mirror)); 3603 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3604 Node* if_ne = generate_fair_guard(bol, NULL); 3605 r->init_req(1, if_ne); 3606 3607 // Casting to .val, check for null 3608 Node* null_ctr = top(); 3609 null_check_oop(obj, &null_ctr); 3610 region->init_req(_npe_path, null_ctr); 3611 r->init_req(2, control()); 3612 3613 set_control(_gvn.transform(r)); 3614 } 3615 3616 Node* bad_type_ctrl = top(); 3617 // Do checkcast optimizations. 3618 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3619 region->init_req(_bad_type_path, bad_type_ctrl); 3620 } 3621 if (region->in(_prim_path) != top() || 3622 region->in(_bad_type_path) != top() || 3623 region->in(_npe_path) != top()) { 3624 // Let Interpreter throw ClassCastException. 3625 PreserveJVMState pjvms(this); 3626 set_control(_gvn.transform(region)); 3627 uncommon_trap(Deoptimization::Reason_intrinsic, 3628 Deoptimization::Action_maybe_recompile); 3629 } 3630 if (!stopped()) { 3631 set_result(res); 3632 } 3633 return true; 3634 } 3635 3636 3637 //--------------------------inline_native_subtype_check------------------------ 3638 // This intrinsic takes the JNI calls out of the heart of 3639 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3640 bool LibraryCallKit::inline_native_subtype_check() { 3641 // Pull both arguments off the stack. 3642 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3643 args[0] = argument(0); 3644 args[1] = argument(1); 3645 Node* klasses[2]; // corresponding Klasses: superk, subk 3646 klasses[0] = klasses[1] = top(); 3647 3648 enum { 3649 // A full decision tree on {superc is prim, subc is prim}: 3650 _prim_0_path = 1, // {P,N} => false 3651 // {P,P} & superc!=subc => false 3652 _prim_same_path, // {P,P} & superc==subc => true 3653 _prim_1_path, // {N,P} => false 3654 _ref_subtype_path, // {N,N} & subtype check wins => true 3655 _both_ref_path, // {N,N} & subtype check loses => false 3656 PATH_LIMIT 3657 }; 3658 3659 RegionNode* region = new RegionNode(PATH_LIMIT); 3660 Node* phi = new PhiNode(region, TypeInt::BOOL); 3661 record_for_igvn(region); 3662 3663 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3664 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3665 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3666 3667 // First null-check both mirrors and load each mirror's klass metaobject. 3668 int which_arg; 3669 for (which_arg = 0; which_arg <= 1; which_arg++) { 3670 Node* arg = args[which_arg]; 3671 arg = null_check(arg); 3672 if (stopped()) break; 3673 args[which_arg] = arg; 3674 3675 Node* p = basic_plus_adr(arg, class_klass_offset); 3676 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3677 klasses[which_arg] = _gvn.transform(kls); 3678 } 3679 3680 // Resolve oops to stable for CmpP below. 3681 args[0] = access_resolve(args[0], 0); 3682 args[1] = access_resolve(args[1], 0); 3683 3684 // Having loaded both klasses, test each for null. 3685 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3686 for (which_arg = 0; which_arg <= 1; which_arg++) { 3687 Node* kls = klasses[which_arg]; 3688 Node* null_ctl = top(); 3689 kls = null_check_oop(kls, &null_ctl, never_see_null); 3690 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3691 region->init_req(prim_path, null_ctl); 3692 if (stopped()) break; 3693 klasses[which_arg] = kls; 3694 } 3695 3696 if (!stopped()) { 3697 // now we have two reference types, in klasses[0..1] 3698 Node* subk = klasses[1]; // the argument to isAssignableFrom 3699 Node* superk = klasses[0]; // the receiver 3700 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3701 // now we have a successful reference subtype check 3702 region->set_req(_ref_subtype_path, control()); 3703 } 3704 3705 // If both operands are primitive (both klasses null), then 3706 // we must return true when they are identical primitives. 3707 // It is convenient to test this after the first null klass check. 3708 set_control(region->in(_prim_0_path)); // go back to first null check 3709 if (!stopped()) { 3710 // Since superc is primitive, make a guard for the superc==subc case. 3711 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3712 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3713 generate_guard(bol_eq, region, PROB_FAIR); 3714 if (region->req() == PATH_LIMIT+1) { 3715 // A guard was added. If the added guard is taken, superc==subc. 3716 region->swap_edges(PATH_LIMIT, _prim_same_path); 3717 region->del_req(PATH_LIMIT); 3718 } 3719 region->set_req(_prim_0_path, control()); // Not equal after all. 3720 } 3721 3722 // these are the only paths that produce 'true': 3723 phi->set_req(_prim_same_path, intcon(1)); 3724 phi->set_req(_ref_subtype_path, intcon(1)); 3725 3726 // pull together the cases: 3727 assert(region->req() == PATH_LIMIT, "sane region"); 3728 for (uint i = 1; i < region->req(); i++) { 3729 Node* ctl = region->in(i); 3730 if (ctl == NULL || ctl == top()) { 3731 region->set_req(i, top()); 3732 phi ->set_req(i, top()); 3733 } else if (phi->in(i) == NULL) { 3734 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3735 } 3736 } 3737 3738 set_control(_gvn.transform(region)); 3739 set_result(_gvn.transform(phi)); 3740 return true; 3741 } 3742 3743 //---------------------generate_array_guard_common------------------------ 3744 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, ArrayKind kind) { 3745 3746 if (stopped()) { 3747 return NULL; 3748 } 3749 3750 // Like generate_guard, adds a new path onto the region. 3751 jint layout_con = 0; 3752 Node* layout_val = get_layout_helper(kls, layout_con); 3753 if (layout_val == NULL) { 3754 bool query = 0; 3755 switch(kind) { 3756 case ObjectArray: query = Klass::layout_helper_is_objArray(layout_con); break; 3757 case NonObjectArray: query = !Klass::layout_helper_is_objArray(layout_con); break; 3758 case TypeArray: query = Klass::layout_helper_is_typeArray(layout_con); break; 3759 case ValueArray: query = Klass::layout_helper_is_valueArray(layout_con); break; 3760 case AnyArray: query = Klass::layout_helper_is_array(layout_con); break; 3761 case NonArray: query = !Klass::layout_helper_is_array(layout_con); break; 3762 default: 3763 ShouldNotReachHere(); 3764 } 3765 if (!query) { 3766 return NULL; // never a branch 3767 } else { // always a branch 3768 Node* always_branch = control(); 3769 if (region != NULL) 3770 region->add_req(always_branch); 3771 set_control(top()); 3772 return always_branch; 3773 } 3774 } 3775 unsigned int value = 0; 3776 BoolTest::mask btest = BoolTest::illegal; 3777 switch(kind) { 3778 case ObjectArray: 3779 case NonObjectArray: { 3780 value = Klass::_lh_array_tag_obj_value; 3781 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3782 btest = kind == ObjectArray ? BoolTest::eq : BoolTest::ne; 3783 break; 3784 } 3785 case TypeArray: { 3786 value = Klass::_lh_array_tag_type_value; 3787 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3788 btest = BoolTest::eq; 3789 break; 3790 } 3791 case ValueArray: { 3792 value = Klass::_lh_array_tag_vt_value; 3793 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3794 btest = BoolTest::eq; 3795 break; 3796 } 3797 case AnyArray: value = Klass::_lh_neutral_value; btest = BoolTest::lt; break; 3798 case NonArray: value = Klass::_lh_neutral_value; btest = BoolTest::gt; break; 3799 default: 3800 ShouldNotReachHere(); 3801 } 3802 // Now test the correct condition. 3803 jint nval = (jint)value; 3804 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3805 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3806 return generate_fair_guard(bol, region); 3807 } 3808 3809 3810 //-----------------------inline_native_newArray-------------------------- 3811 // private static native Object java.lang.reflect.Array.newArray(Class<?> componentType, int length); 3812 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 3813 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 3814 Node* mirror; 3815 Node* count_val; 3816 if (uninitialized) { 3817 mirror = argument(1); 3818 count_val = argument(2); 3819 } else { 3820 mirror = argument(0); 3821 count_val = argument(1); 3822 } 3823 3824 mirror = null_check(mirror); 3825 // If mirror or obj is dead, only null-path is taken. 3826 if (stopped()) return true; 3827 3828 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3829 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3830 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3831 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3832 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3833 3834 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3835 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3836 result_reg, _slow_path); 3837 Node* normal_ctl = control(); 3838 Node* no_array_ctl = result_reg->in(_slow_path); 3839 3840 // Generate code for the slow case. We make a call to newArray(). 3841 set_control(no_array_ctl); 3842 if (!stopped()) { 3843 // Either the input type is void.class, or else the 3844 // array klass has not yet been cached. Either the 3845 // ensuing call will throw an exception, or else it 3846 // will cache the array klass for next time. 3847 PreserveJVMState pjvms(this); 3848 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3849 Node* slow_result = set_results_for_java_call(slow_call); 3850 // this->control() comes from set_results_for_java_call 3851 result_reg->set_req(_slow_path, control()); 3852 result_val->set_req(_slow_path, slow_result); 3853 result_io ->set_req(_slow_path, i_o()); 3854 result_mem->set_req(_slow_path, reset_memory()); 3855 } 3856 3857 set_control(normal_ctl); 3858 if (!stopped()) { 3859 // Normal case: The array type has been cached in the java.lang.Class. 3860 // The following call works fine even if the array type is polymorphic. 3861 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3862 Node* obj = new_array(klass_node, count_val, 0, NULL, false, mirror); // no arguments to push 3863 result_reg->init_req(_normal_path, control()); 3864 result_val->init_req(_normal_path, obj); 3865 result_io ->init_req(_normal_path, i_o()); 3866 result_mem->init_req(_normal_path, reset_memory()); 3867 3868 if (uninitialized) { 3869 // Mark the allocation so that zeroing is skipped 3870 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 3871 alloc->maybe_set_complete(&_gvn); 3872 } 3873 } 3874 3875 // Return the combined state. 3876 set_i_o( _gvn.transform(result_io) ); 3877 set_all_memory( _gvn.transform(result_mem)); 3878 3879 C->set_has_split_ifs(true); // Has chance for split-if optimization 3880 set_result(result_reg, result_val); 3881 return true; 3882 } 3883 3884 //----------------------inline_native_getLength-------------------------- 3885 // public static native int java.lang.reflect.Array.getLength(Object array); 3886 bool LibraryCallKit::inline_native_getLength() { 3887 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3888 3889 Node* array = null_check(argument(0)); 3890 // If array is dead, only null-path is taken. 3891 if (stopped()) return true; 3892 3893 // Deoptimize if it is a non-array. 3894 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3895 3896 if (non_array != NULL) { 3897 PreserveJVMState pjvms(this); 3898 set_control(non_array); 3899 uncommon_trap(Deoptimization::Reason_intrinsic, 3900 Deoptimization::Action_maybe_recompile); 3901 } 3902 3903 // If control is dead, only non-array-path is taken. 3904 if (stopped()) return true; 3905 3906 // The works fine even if the array type is polymorphic. 3907 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3908 Node* result = load_array_length(array); 3909 3910 C->set_has_split_ifs(true); // Has chance for split-if optimization 3911 set_result(result); 3912 return true; 3913 } 3914 3915 //------------------------inline_array_copyOf---------------------------- 3916 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3917 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3918 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3919 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3920 3921 // Get the arguments. 3922 Node* original = argument(0); 3923 Node* start = is_copyOfRange? argument(1): intcon(0); 3924 Node* end = is_copyOfRange? argument(2): argument(1); 3925 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3926 3927 const TypeAryPtr* original_t = _gvn.type(original)->isa_aryptr(); 3928 const TypeInstPtr* mirror_t = _gvn.type(array_type_mirror)->isa_instptr(); 3929 if (EnableValhalla && ValueArrayFlatten && 3930 (original_t == NULL || mirror_t == NULL || 3931 (mirror_t->java_mirror_type() == NULL && 3932 (original_t->elem()->isa_valuetype() || 3933 (original_t->elem()->make_oopptr() != NULL && 3934 original_t->elem()->make_oopptr()->can_be_value_type()))))) { 3935 // We need to know statically if the copy is to a flattened array 3936 // or not but can't tell. 3937 return false; 3938 } 3939 3940 Node* newcopy = NULL; 3941 3942 // Set the original stack and the reexecute bit for the interpreter to reexecute 3943 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3944 { PreserveReexecuteState preexecs(this); 3945 jvms()->set_should_reexecute(true); 3946 3947 array_type_mirror = null_check(array_type_mirror); 3948 original = null_check(original); 3949 3950 // Check if a null path was taken unconditionally. 3951 if (stopped()) return true; 3952 3953 Node* orig_length = load_array_length(original); 3954 3955 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3956 klass_node = null_check(klass_node); 3957 3958 RegionNode* bailout = new RegionNode(1); 3959 record_for_igvn(bailout); 3960 3961 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3962 // Bail out if that is so. 3963 // Value type array may have object field that would require a 3964 // write barrier. Conservatively, go to slow path. 3965 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 3966 Node* not_objArray = !bs->array_copy_requires_gc_barriers(false, T_OBJECT, false, BarrierSetC2::Parsing) ? 3967 generate_typeArray_guard(klass_node, bailout) : generate_non_objArray_guard(klass_node, bailout); 3968 if (not_objArray != NULL) { 3969 // Improve the klass node's type from the new optimistic assumption: 3970 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3971 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0)); 3972 Node* cast = new CastPPNode(klass_node, akls); 3973 cast->init_req(0, control()); 3974 klass_node = _gvn.transform(cast); 3975 } 3976 3977 Node* original_kls = load_object_klass(original); 3978 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3979 // loads/stores but it is legal only if we're sure the 3980 // Arrays.copyOf would succeed. So we need all input arguments 3981 // to the copyOf to be validated, including that the copy to the 3982 // new array won't trigger an ArrayStoreException. That subtype 3983 // check can be optimized if we know something on the type of 3984 // the input array from type speculation. 3985 if (_gvn.type(klass_node)->singleton() && !stopped()) { 3986 ciKlass* subk = _gvn.type(original_kls)->is_klassptr()->klass(); 3987 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3988 3989 int test = C->static_subtype_check(superk, subk); 3990 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3991 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3992 if (t_original->speculative_type() != NULL) { 3993 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3994 original_kls = load_object_klass(original); 3995 } 3996 } 3997 } 3998 3999 if (EnableValhalla) { 4000 // Either both or neither new array klass and original array 4001 // klass must be flattened 4002 Node* flattened_klass = generate_valueArray_guard(klass_node, NULL); 4003 generate_valueArray_guard(original_kls, bailout); 4004 if (flattened_klass != NULL) { 4005 RegionNode* r = new RegionNode(2); 4006 record_for_igvn(r); 4007 r->init_req(1, control()); 4008 set_control(flattened_klass); 4009 generate_valueArray_guard(original_kls, r); 4010 bailout->add_req(control()); 4011 set_control(_gvn.transform(r)); 4012 } 4013 } 4014 4015 // Bail out if either start or end is negative. 4016 generate_negative_guard(start, bailout, &start); 4017 generate_negative_guard(end, bailout, &end); 4018 4019 Node* length = end; 4020 if (_gvn.type(start) != TypeInt::ZERO) { 4021 length = _gvn.transform(new SubINode(end, start)); 4022 } 4023 4024 // Bail out if length is negative. 4025 // Without this the new_array would throw 4026 // NegativeArraySizeException but IllegalArgumentException is what 4027 // should be thrown 4028 generate_negative_guard(length, bailout, &length); 4029 4030 if (bailout->req() > 1) { 4031 PreserveJVMState pjvms(this); 4032 set_control(_gvn.transform(bailout)); 4033 uncommon_trap(Deoptimization::Reason_intrinsic, 4034 Deoptimization::Action_maybe_recompile); 4035 } 4036 4037 if (!stopped()) { 4038 // How many elements will we copy from the original? 4039 // The answer is MinI(orig_length - start, length). 4040 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 4041 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 4042 4043 original = access_resolve(original, ACCESS_READ); 4044 4045 // Generate a direct call to the right arraycopy function(s). 4046 // We know the copy is disjoint but we might not know if the 4047 // oop stores need checking. 4048 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 4049 // This will fail a store-check if x contains any non-nulls. 4050 4051 bool validated = false; 4052 // Reason_class_check rather than Reason_intrinsic because we 4053 // want to intrinsify even if this traps. 4054 if (!too_many_traps(Deoptimization::Reason_class_check)) { 4055 Node* not_subtype_ctrl = gen_subtype_check(original_kls, 4056 klass_node); 4057 4058 if (not_subtype_ctrl != top()) { 4059 PreserveJVMState pjvms(this); 4060 set_control(not_subtype_ctrl); 4061 uncommon_trap(Deoptimization::Reason_class_check, 4062 Deoptimization::Action_make_not_entrant); 4063 assert(stopped(), "Should be stopped"); 4064 } 4065 validated = true; 4066 } 4067 4068 if (!stopped()) { 4069 // Load element mirror 4070 Node* p = basic_plus_adr(array_type_mirror, java_lang_Class::component_mirror_offset_in_bytes()); 4071 Node* elem_mirror = access_load_at(array_type_mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR, T_OBJECT, IN_HEAP); 4072 4073 newcopy = new_array(klass_node, length, 0, NULL, false, elem_mirror); 4074 4075 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false, 4076 original_kls, klass_node); 4077 if (!is_copyOfRange) { 4078 ac->set_copyof(validated); 4079 } else { 4080 ac->set_copyofrange(validated); 4081 } 4082 Node* n = _gvn.transform(ac); 4083 if (n == ac) { 4084 ac->connect_outputs(this); 4085 } else { 4086 assert(validated, "shouldn't transform if all arguments not validated"); 4087 set_all_memory(n); 4088 } 4089 } 4090 } 4091 } // original reexecute is set back here 4092 4093 C->set_has_split_ifs(true); // Has chance for split-if optimization 4094 if (!stopped()) { 4095 set_result(newcopy); 4096 } 4097 return true; 4098 } 4099 4100 4101 //----------------------generate_virtual_guard--------------------------- 4102 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4103 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4104 RegionNode* slow_region) { 4105 ciMethod* method = callee(); 4106 int vtable_index = method->vtable_index(); 4107 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4108 "bad index %d", vtable_index); 4109 // Get the Method* out of the appropriate vtable entry. 4110 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 4111 vtable_index*vtableEntry::size_in_bytes() + 4112 vtableEntry::method_offset_in_bytes(); 4113 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4114 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4115 4116 // Compare the target method with the expected method (e.g., Object.hashCode). 4117 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4118 4119 Node* native_call = makecon(native_call_addr); 4120 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4121 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4122 4123 return generate_slow_guard(test_native, slow_region); 4124 } 4125 4126 //-----------------------generate_method_call---------------------------- 4127 // Use generate_method_call to make a slow-call to the real 4128 // method if the fast path fails. An alternative would be to 4129 // use a stub like OptoRuntime::slow_arraycopy_Java. 4130 // This only works for expanding the current library call, 4131 // not another intrinsic. (E.g., don't use this for making an 4132 // arraycopy call inside of the copyOf intrinsic.) 4133 CallJavaNode* 4134 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4135 // When compiling the intrinsic method itself, do not use this technique. 4136 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4137 4138 ciMethod* method = callee(); 4139 // ensure the JVMS we have will be correct for this call 4140 guarantee(method_id == method->intrinsic_id(), "must match"); 4141 4142 const TypeFunc* tf = TypeFunc::make(method); 4143 CallJavaNode* slow_call; 4144 if (is_static) { 4145 assert(!is_virtual, ""); 4146 slow_call = new CallStaticJavaNode(C, tf, 4147 SharedRuntime::get_resolve_static_call_stub(), 4148 method, bci()); 4149 } else if (is_virtual) { 4150 null_check_receiver(); 4151 int vtable_index = Method::invalid_vtable_index; 4152 if (UseInlineCaches) { 4153 // Suppress the vtable call 4154 } else { 4155 // hashCode and clone are not a miranda methods, 4156 // so the vtable index is fixed. 4157 // No need to use the linkResolver to get it. 4158 vtable_index = method->vtable_index(); 4159 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4160 "bad index %d", vtable_index); 4161 } 4162 slow_call = new CallDynamicJavaNode(tf, 4163 SharedRuntime::get_resolve_virtual_call_stub(), 4164 method, vtable_index, bci()); 4165 } else { // neither virtual nor static: opt_virtual 4166 null_check_receiver(); 4167 slow_call = new CallStaticJavaNode(C, tf, 4168 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4169 method, bci()); 4170 slow_call->set_optimized_virtual(true); 4171 } 4172 if (CallGenerator::is_inlined_method_handle_intrinsic(this->method(), bci(), callee())) { 4173 // To be able to issue a direct call (optimized virtual or virtual) 4174 // and skip a call to MH.linkTo*/invokeBasic adapter, additional information 4175 // about the method being invoked should be attached to the call site to 4176 // make resolution logic work (see SharedRuntime::resolve_{virtual,opt_virtual}_call_C). 4177 slow_call->set_override_symbolic_info(true); 4178 } 4179 set_arguments_for_java_call(slow_call); 4180 set_edges_for_java_call(slow_call); 4181 return slow_call; 4182 } 4183 4184 4185 /** 4186 * Build special case code for calls to hashCode on an object. This call may 4187 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4188 * slightly different code. 4189 */ 4190 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4191 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4192 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4193 4194 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4195 4196 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4197 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4198 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4199 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4200 Node* obj = argument(0); 4201 4202 if (obj->is_ValueType() || gvn().type(obj)->is_valuetypeptr()) { 4203 return false; 4204 } 4205 4206 if (!is_static) { 4207 // Check for hashing null object 4208 obj = null_check_receiver(); 4209 if (stopped()) return true; // unconditionally null 4210 result_reg->init_req(_null_path, top()); 4211 result_val->init_req(_null_path, top()); 4212 } else { 4213 // Do a null check, and return zero if null. 4214 // System.identityHashCode(null) == 0 4215 Node* null_ctl = top(); 4216 obj = null_check_oop(obj, &null_ctl); 4217 result_reg->init_req(_null_path, null_ctl); 4218 result_val->init_req(_null_path, _gvn.intcon(0)); 4219 } 4220 4221 // Unconditionally null? Then return right away. 4222 if (stopped()) { 4223 set_control( result_reg->in(_null_path)); 4224 if (!stopped()) 4225 set_result(result_val->in(_null_path)); 4226 return true; 4227 } 4228 4229 // We only go to the fast case code if we pass a number of guards. The 4230 // paths which do not pass are accumulated in the slow_region. 4231 RegionNode* slow_region = new RegionNode(1); 4232 record_for_igvn(slow_region); 4233 4234 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4235 assert(!obj_type->isa_valuetype() || !obj_type->is_valuetypeptr(), "no value type here"); 4236 if (is_static && obj_type->can_be_value_type()) { 4237 Node* obj_klass = load_object_klass(obj); 4238 generate_value_guard(obj_klass, slow_region); 4239 } 4240 4241 // If this is a virtual call, we generate a funny guard. We pull out 4242 // the vtable entry corresponding to hashCode() from the target object. 4243 // If the target method which we are calling happens to be the native 4244 // Object hashCode() method, we pass the guard. We do not need this 4245 // guard for non-virtual calls -- the caller is known to be the native 4246 // Object hashCode(). 4247 if (is_virtual) { 4248 // After null check, get the object's klass. 4249 Node* obj_klass = load_object_klass(obj); 4250 generate_virtual_guard(obj_klass, slow_region); 4251 } 4252 4253 // Get the header out of the object, use LoadMarkNode when available 4254 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4255 // The control of the load must be NULL. Otherwise, the load can move before 4256 // the null check after castPP removal. 4257 Node* no_ctrl = NULL; 4258 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4259 4260 // Test the header to see if it is unlocked. 4261 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4262 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4263 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4264 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4265 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4266 4267 generate_slow_guard(test_unlocked, slow_region); 4268 4269 // Get the hash value and check to see that it has been properly assigned. 4270 // We depend on hash_mask being at most 32 bits and avoid the use of 4271 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4272 // vm: see markOop.hpp. 4273 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4274 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4275 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4276 // This hack lets the hash bits live anywhere in the mark object now, as long 4277 // as the shift drops the relevant bits into the low 32 bits. Note that 4278 // Java spec says that HashCode is an int so there's no point in capturing 4279 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4280 hshifted_header = ConvX2I(hshifted_header); 4281 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4282 4283 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4284 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4285 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4286 4287 generate_slow_guard(test_assigned, slow_region); 4288 4289 Node* init_mem = reset_memory(); 4290 // fill in the rest of the null path: 4291 result_io ->init_req(_null_path, i_o()); 4292 result_mem->init_req(_null_path, init_mem); 4293 4294 result_val->init_req(_fast_path, hash_val); 4295 result_reg->init_req(_fast_path, control()); 4296 result_io ->init_req(_fast_path, i_o()); 4297 result_mem->init_req(_fast_path, init_mem); 4298 4299 // Generate code for the slow case. We make a call to hashCode(). 4300 set_control(_gvn.transform(slow_region)); 4301 if (!stopped()) { 4302 // No need for PreserveJVMState, because we're using up the present state. 4303 set_all_memory(init_mem); 4304 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4305 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4306 Node* slow_result = set_results_for_java_call(slow_call); 4307 // this->control() comes from set_results_for_java_call 4308 result_reg->init_req(_slow_path, control()); 4309 result_val->init_req(_slow_path, slow_result); 4310 result_io ->set_req(_slow_path, i_o()); 4311 result_mem ->set_req(_slow_path, reset_memory()); 4312 } 4313 4314 // Return the combined state. 4315 set_i_o( _gvn.transform(result_io) ); 4316 set_all_memory( _gvn.transform(result_mem)); 4317 4318 set_result(result_reg, result_val); 4319 return true; 4320 } 4321 4322 //---------------------------inline_native_getClass---------------------------- 4323 // public final native Class<?> java.lang.Object.getClass(); 4324 // 4325 // Build special case code for calls to getClass on an object. 4326 bool LibraryCallKit::inline_native_getClass() { 4327 Node* obj = argument(0); 4328 if (obj->is_ValueType()) { 4329 ciKlass* vk = _gvn.type(obj)->is_valuetype()->value_klass(); 4330 set_result(makecon(TypeInstPtr::make(vk->java_mirror()))); 4331 return true; 4332 } 4333 obj = null_check_receiver(); 4334 if (stopped()) return true; 4335 set_result(load_mirror_from_klass(load_object_klass(obj))); 4336 return true; 4337 } 4338 4339 //-----------------inline_native_Reflection_getCallerClass--------------------- 4340 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4341 // 4342 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4343 // 4344 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4345 // in that it must skip particular security frames and checks for 4346 // caller sensitive methods. 4347 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4348 #ifndef PRODUCT 4349 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4350 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4351 } 4352 #endif 4353 4354 if (!jvms()->has_method()) { 4355 #ifndef PRODUCT 4356 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4357 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4358 } 4359 #endif 4360 return false; 4361 } 4362 4363 // Walk back up the JVM state to find the caller at the required 4364 // depth. 4365 JVMState* caller_jvms = jvms(); 4366 4367 // Cf. JVM_GetCallerClass 4368 // NOTE: Start the loop at depth 1 because the current JVM state does 4369 // not include the Reflection.getCallerClass() frame. 4370 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4371 ciMethod* m = caller_jvms->method(); 4372 switch (n) { 4373 case 0: 4374 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4375 break; 4376 case 1: 4377 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4378 if (!m->caller_sensitive()) { 4379 #ifndef PRODUCT 4380 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4381 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4382 } 4383 #endif 4384 return false; // bail-out; let JVM_GetCallerClass do the work 4385 } 4386 break; 4387 default: 4388 if (!m->is_ignored_by_security_stack_walk()) { 4389 // We have reached the desired frame; return the holder class. 4390 // Acquire method holder as java.lang.Class and push as constant. 4391 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4392 ciInstance* caller_mirror = caller_klass->java_mirror(); 4393 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4394 4395 #ifndef PRODUCT 4396 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4397 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()); 4398 tty->print_cr(" JVM state at this point:"); 4399 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4400 ciMethod* m = jvms()->of_depth(i)->method(); 4401 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4402 } 4403 } 4404 #endif 4405 return true; 4406 } 4407 break; 4408 } 4409 } 4410 4411 #ifndef PRODUCT 4412 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4413 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4414 tty->print_cr(" JVM state at this point:"); 4415 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4416 ciMethod* m = jvms()->of_depth(i)->method(); 4417 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4418 } 4419 } 4420 #endif 4421 4422 return false; // bail-out; let JVM_GetCallerClass do the work 4423 } 4424 4425 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4426 Node* arg = argument(0); 4427 Node* result = NULL; 4428 4429 switch (id) { 4430 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4431 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4432 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4433 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4434 4435 case vmIntrinsics::_doubleToLongBits: { 4436 // two paths (plus control) merge in a wood 4437 RegionNode *r = new RegionNode(3); 4438 Node *phi = new PhiNode(r, TypeLong::LONG); 4439 4440 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4441 // Build the boolean node 4442 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4443 4444 // Branch either way. 4445 // NaN case is less traveled, which makes all the difference. 4446 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4447 Node *opt_isnan = _gvn.transform(ifisnan); 4448 assert( opt_isnan->is_If(), "Expect an IfNode"); 4449 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4450 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4451 4452 set_control(iftrue); 4453 4454 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4455 Node *slow_result = longcon(nan_bits); // return NaN 4456 phi->init_req(1, _gvn.transform( slow_result )); 4457 r->init_req(1, iftrue); 4458 4459 // Else fall through 4460 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4461 set_control(iffalse); 4462 4463 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4464 r->init_req(2, iffalse); 4465 4466 // Post merge 4467 set_control(_gvn.transform(r)); 4468 record_for_igvn(r); 4469 4470 C->set_has_split_ifs(true); // Has chance for split-if optimization 4471 result = phi; 4472 assert(result->bottom_type()->isa_long(), "must be"); 4473 break; 4474 } 4475 4476 case vmIntrinsics::_floatToIntBits: { 4477 // two paths (plus control) merge in a wood 4478 RegionNode *r = new RegionNode(3); 4479 Node *phi = new PhiNode(r, TypeInt::INT); 4480 4481 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4482 // Build the boolean node 4483 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4484 4485 // Branch either way. 4486 // NaN case is less traveled, which makes all the difference. 4487 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4488 Node *opt_isnan = _gvn.transform(ifisnan); 4489 assert( opt_isnan->is_If(), "Expect an IfNode"); 4490 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4491 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4492 4493 set_control(iftrue); 4494 4495 static const jint nan_bits = 0x7fc00000; 4496 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4497 phi->init_req(1, _gvn.transform( slow_result )); 4498 r->init_req(1, iftrue); 4499 4500 // Else fall through 4501 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4502 set_control(iffalse); 4503 4504 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4505 r->init_req(2, iffalse); 4506 4507 // Post merge 4508 set_control(_gvn.transform(r)); 4509 record_for_igvn(r); 4510 4511 C->set_has_split_ifs(true); // Has chance for split-if optimization 4512 result = phi; 4513 assert(result->bottom_type()->isa_int(), "must be"); 4514 break; 4515 } 4516 4517 default: 4518 fatal_unexpected_iid(id); 4519 break; 4520 } 4521 set_result(_gvn.transform(result)); 4522 return true; 4523 } 4524 4525 //----------------------inline_unsafe_copyMemory------------------------- 4526 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4527 bool LibraryCallKit::inline_unsafe_copyMemory() { 4528 if (callee()->is_static()) return false; // caller must have the capability! 4529 null_check_receiver(); // null-check receiver 4530 if (stopped()) return true; 4531 4532 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4533 4534 Node* src_ptr = argument(1); // type: oop 4535 Node* src_off = ConvL2X(argument(2)); // type: long 4536 Node* dst_ptr = argument(4); // type: oop 4537 Node* dst_off = ConvL2X(argument(5)); // type: long 4538 Node* size = ConvL2X(argument(7)); // type: long 4539 4540 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4541 "fieldOffset must be byte-scaled"); 4542 4543 src_ptr = access_resolve(src_ptr, ACCESS_READ); 4544 dst_ptr = access_resolve(dst_ptr, ACCESS_WRITE); 4545 Node* src = make_unsafe_address(src_ptr, src_off, ACCESS_READ); 4546 Node* dst = make_unsafe_address(dst_ptr, dst_off, ACCESS_WRITE); 4547 4548 // Conservatively insert a memory barrier on all memory slices. 4549 // Do not let writes of the copy source or destination float below the copy. 4550 insert_mem_bar(Op_MemBarCPUOrder); 4551 4552 // Call it. Note that the length argument is not scaled. 4553 make_runtime_call(RC_LEAF|RC_NO_FP, 4554 OptoRuntime::fast_arraycopy_Type(), 4555 StubRoutines::unsafe_arraycopy(), 4556 "unsafe_arraycopy", 4557 TypeRawPtr::BOTTOM, 4558 src, dst, size XTOP); 4559 4560 // Do not let reads of the copy destination float above the copy. 4561 insert_mem_bar(Op_MemBarCPUOrder); 4562 4563 return true; 4564 } 4565 4566 //------------------------clone_coping----------------------------------- 4567 // Helper function for inline_native_clone. 4568 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) { 4569 assert(obj_size != NULL, ""); 4570 Node* raw_obj = alloc_obj->in(1); 4571 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4572 4573 AllocateNode* alloc = NULL; 4574 if (ReduceBulkZeroing) { 4575 // We will be completely responsible for initializing this object - 4576 // mark Initialize node as complete. 4577 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4578 // The object was just allocated - there should be no any stores! 4579 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4580 // Mark as complete_with_arraycopy so that on AllocateNode 4581 // expansion, we know this AllocateNode is initialized by an array 4582 // copy and a StoreStore barrier exists after the array copy. 4583 alloc->initialization()->set_complete_with_arraycopy(); 4584 } 4585 4586 // Copy the fastest available way. 4587 // TODO: generate fields copies for small objects instead. 4588 Node* size = _gvn.transform(obj_size); 4589 4590 // Exclude the header but include array length to copy by 8 bytes words. 4591 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4592 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4593 instanceOopDesc::base_offset_in_bytes(); 4594 // base_off: 4595 // 8 - 32-bit VM 4596 // 12 - 64-bit VM, compressed klass 4597 // 16 - 64-bit VM, normal klass 4598 if (base_off % BytesPerLong != 0) { 4599 assert(UseCompressedClassPointers, ""); 4600 if (is_array) { 4601 // Exclude length to copy by 8 bytes words. 4602 base_off += sizeof(int); 4603 } else { 4604 // Include klass to copy by 8 bytes words. 4605 base_off = instanceOopDesc::klass_offset_in_bytes(); 4606 } 4607 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4608 } 4609 Node* src_base = basic_plus_adr(obj, base_off); 4610 Node* dst_base = basic_plus_adr(alloc_obj, base_off); 4611 4612 // Compute the length also, if needed: 4613 Node* countx = size; 4614 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4615 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong))); 4616 4617 access_clone(src_base, dst_base, countx, is_array); 4618 4619 // Do not let reads from the cloned object float above the arraycopy. 4620 if (alloc != NULL) { 4621 // Do not let stores that initialize this object be reordered with 4622 // a subsequent store that would make this object accessible by 4623 // other threads. 4624 // Record what AllocateNode this StoreStore protects so that 4625 // escape analysis can go from the MemBarStoreStoreNode to the 4626 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4627 // based on the escape status of the AllocateNode. 4628 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 4629 } else { 4630 insert_mem_bar(Op_MemBarCPUOrder); 4631 } 4632 } 4633 4634 //------------------------inline_native_clone---------------------------- 4635 // protected native Object java.lang.Object.clone(); 4636 // 4637 // Here are the simple edge cases: 4638 // null receiver => normal trap 4639 // virtual and clone was overridden => slow path to out-of-line clone 4640 // not cloneable or finalizer => slow path to out-of-line Object.clone 4641 // 4642 // The general case has two steps, allocation and copying. 4643 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4644 // 4645 // Copying also has two cases, oop arrays and everything else. 4646 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4647 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4648 // 4649 // These steps fold up nicely if and when the cloned object's klass 4650 // can be sharply typed as an object array, a type array, or an instance. 4651 // 4652 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4653 PhiNode* result_val; 4654 4655 // Set the reexecute bit for the interpreter to reexecute 4656 // the bytecode that invokes Object.clone if deoptimization happens. 4657 { PreserveReexecuteState preexecs(this); 4658 jvms()->set_should_reexecute(true); 4659 4660 Node* obj = argument(0); 4661 if (obj->is_ValueType()) { 4662 return false; 4663 } 4664 4665 obj = null_check_receiver(); 4666 if (stopped()) return true; 4667 4668 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4669 4670 // If we are going to clone an instance, we need its exact type to 4671 // know the number and types of fields to convert the clone to 4672 // loads/stores. Maybe a speculative type can help us. 4673 if (!obj_type->klass_is_exact() && 4674 obj_type->speculative_type() != NULL && 4675 obj_type->speculative_type()->is_instance_klass() && 4676 !obj_type->speculative_type()->is_valuetype()) { 4677 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4678 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4679 !spec_ik->has_injected_fields()) { 4680 ciKlass* k = obj_type->klass(); 4681 if (!k->is_instance_klass() || 4682 k->as_instance_klass()->is_interface() || 4683 k->as_instance_klass()->has_subklass()) { 4684 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4685 } 4686 } 4687 } 4688 4689 Node* obj_klass = load_object_klass(obj); 4690 4691 // Conservatively insert a memory barrier on all memory slices. 4692 // Do not let writes into the original float below the clone. 4693 insert_mem_bar(Op_MemBarCPUOrder); 4694 4695 // paths into result_reg: 4696 enum { 4697 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4698 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4699 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4700 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4701 PATH_LIMIT 4702 }; 4703 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4704 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4705 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4706 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4707 record_for_igvn(result_reg); 4708 4709 // We only go to the fast case code if we pass a number of guards. 4710 // The paths which do not pass are accumulated in the slow_region. 4711 RegionNode* slow_region = new RegionNode(1); 4712 record_for_igvn(slow_region); 4713 4714 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4715 if (array_ctl != NULL) { 4716 // It's an array. 4717 PreserveJVMState pjvms(this); 4718 set_control(array_ctl); 4719 4720 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4721 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) { 4722 // Flattened value type array may have object field that would require a 4723 // write barrier. Conservatively, go to slow path. 4724 generate_valueArray_guard(obj_klass, slow_region); 4725 } 4726 4727 if (!stopped()) { 4728 Node* obj_length = load_array_length(obj); 4729 Node* obj_size = NULL; 4730 // Load element mirror 4731 Node* array_type_mirror = load_mirror_from_klass(obj_klass); 4732 Node* p = basic_plus_adr(array_type_mirror, java_lang_Class::component_mirror_offset_in_bytes()); 4733 Node* elem_mirror = access_load_at(array_type_mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR, T_OBJECT, IN_HEAP); 4734 4735 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size, false, elem_mirror); 4736 4737 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4738 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) { 4739 // If it is an oop array, it requires very special treatment, 4740 // because gc barriers are required when accessing the array. 4741 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4742 if (is_obja != NULL) { 4743 PreserveJVMState pjvms2(this); 4744 set_control(is_obja); 4745 // Generate a direct call to the right arraycopy function(s). 4746 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4747 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false); 4748 ac->set_cloneoop(); 4749 Node* n = _gvn.transform(ac); 4750 assert(n == ac, "cannot disappear"); 4751 ac->connect_outputs(this); 4752 4753 result_reg->init_req(_objArray_path, control()); 4754 result_val->init_req(_objArray_path, alloc_obj); 4755 result_i_o ->set_req(_objArray_path, i_o()); 4756 result_mem ->set_req(_objArray_path, reset_memory()); 4757 } 4758 } 4759 4760 // Otherwise, there are no barriers to worry about. 4761 // (We can dispense with card marks if we know the allocation 4762 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4763 // causes the non-eden paths to take compensating steps to 4764 // simulate a fresh allocation, so that no further 4765 // card marks are required in compiled code to initialize 4766 // the object.) 4767 4768 if (!stopped()) { 4769 copy_to_clone(obj, alloc_obj, obj_size, true); 4770 4771 // Present the results of the copy. 4772 result_reg->init_req(_array_path, control()); 4773 result_val->init_req(_array_path, alloc_obj); 4774 result_i_o ->set_req(_array_path, i_o()); 4775 result_mem ->set_req(_array_path, reset_memory()); 4776 } 4777 } 4778 } 4779 4780 if (!stopped()) { 4781 // It's an instance (we did array above). Make the slow-path tests. 4782 // If this is a virtual call, we generate a funny guard. We grab 4783 // the vtable entry corresponding to clone() from the target object. 4784 // If the target method which we are calling happens to be the 4785 // Object clone() method, we pass the guard. We do not need this 4786 // guard for non-virtual calls; the caller is known to be the native 4787 // Object clone(). 4788 if (is_virtual) { 4789 generate_virtual_guard(obj_klass, slow_region); 4790 } 4791 4792 // The object must be easily cloneable and must not have a finalizer. 4793 // Both of these conditions may be checked in a single test. 4794 // We could optimize the test further, but we don't care. 4795 generate_access_flags_guard(obj_klass, 4796 // Test both conditions: 4797 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4798 // Must be cloneable but not finalizer: 4799 JVM_ACC_IS_CLONEABLE_FAST, 4800 slow_region); 4801 } 4802 4803 if (!stopped()) { 4804 // It's an instance, and it passed the slow-path tests. 4805 PreserveJVMState pjvms(this); 4806 Node* obj_size = NULL; 4807 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4808 // is reexecuted if deoptimization occurs and there could be problems when merging 4809 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4810 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4811 4812 copy_to_clone(obj, alloc_obj, obj_size, false); 4813 4814 // Present the results of the slow call. 4815 result_reg->init_req(_instance_path, control()); 4816 result_val->init_req(_instance_path, alloc_obj); 4817 result_i_o ->set_req(_instance_path, i_o()); 4818 result_mem ->set_req(_instance_path, reset_memory()); 4819 } 4820 4821 // Generate code for the slow case. We make a call to clone(). 4822 set_control(_gvn.transform(slow_region)); 4823 if (!stopped()) { 4824 PreserveJVMState pjvms(this); 4825 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4826 // We need to deoptimize on exception (see comment above) 4827 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true); 4828 // this->control() comes from set_results_for_java_call 4829 result_reg->init_req(_slow_path, control()); 4830 result_val->init_req(_slow_path, slow_result); 4831 result_i_o ->set_req(_slow_path, i_o()); 4832 result_mem ->set_req(_slow_path, reset_memory()); 4833 } 4834 4835 // Return the combined state. 4836 set_control( _gvn.transform(result_reg)); 4837 set_i_o( _gvn.transform(result_i_o)); 4838 set_all_memory( _gvn.transform(result_mem)); 4839 } // original reexecute is set back here 4840 4841 set_result(_gvn.transform(result_val)); 4842 return true; 4843 } 4844 4845 // If we have a tightly coupled allocation, the arraycopy may take care 4846 // of the array initialization. If one of the guards we insert between 4847 // the allocation and the arraycopy causes a deoptimization, an 4848 // unitialized array will escape the compiled method. To prevent that 4849 // we set the JVM state for uncommon traps between the allocation and 4850 // the arraycopy to the state before the allocation so, in case of 4851 // deoptimization, we'll reexecute the allocation and the 4852 // initialization. 4853 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4854 if (alloc != NULL) { 4855 ciMethod* trap_method = alloc->jvms()->method(); 4856 int trap_bci = alloc->jvms()->bci(); 4857 4858 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4859 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4860 // Make sure there's no store between the allocation and the 4861 // arraycopy otherwise visible side effects could be rexecuted 4862 // in case of deoptimization and cause incorrect execution. 4863 bool no_interfering_store = true; 4864 Node* mem = alloc->in(TypeFunc::Memory); 4865 if (mem->is_MergeMem()) { 4866 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4867 Node* n = mms.memory(); 4868 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4869 assert(n->is_Store(), "what else?"); 4870 no_interfering_store = false; 4871 break; 4872 } 4873 } 4874 } else { 4875 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4876 Node* n = mms.memory(); 4877 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4878 assert(n->is_Store(), "what else?"); 4879 no_interfering_store = false; 4880 break; 4881 } 4882 } 4883 } 4884 4885 if (no_interfering_store) { 4886 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4887 uint size = alloc->req(); 4888 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4889 old_jvms->set_map(sfpt); 4890 for (uint i = 0; i < size; i++) { 4891 sfpt->init_req(i, alloc->in(i)); 4892 } 4893 // re-push array length for deoptimization 4894 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4895 old_jvms->set_sp(old_jvms->sp()+1); 4896 old_jvms->set_monoff(old_jvms->monoff()+1); 4897 old_jvms->set_scloff(old_jvms->scloff()+1); 4898 old_jvms->set_endoff(old_jvms->endoff()+1); 4899 old_jvms->set_should_reexecute(true); 4900 4901 sfpt->set_i_o(map()->i_o()); 4902 sfpt->set_memory(map()->memory()); 4903 sfpt->set_control(map()->control()); 4904 4905 JVMState* saved_jvms = jvms(); 4906 saved_reexecute_sp = _reexecute_sp; 4907 4908 set_jvms(sfpt->jvms()); 4909 _reexecute_sp = jvms()->sp(); 4910 4911 return saved_jvms; 4912 } 4913 } 4914 } 4915 return NULL; 4916 } 4917 4918 // In case of a deoptimization, we restart execution at the 4919 // allocation, allocating a new array. We would leave an uninitialized 4920 // array in the heap that GCs wouldn't expect. Move the allocation 4921 // after the traps so we don't allocate the array if we 4922 // deoptimize. This is possible because tightly_coupled_allocation() 4923 // guarantees there's no observer of the allocated array at this point 4924 // and the control flow is simple enough. 4925 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, 4926 int saved_reexecute_sp, uint new_idx) { 4927 if (saved_jvms != NULL && !stopped()) { 4928 assert(alloc != NULL, "only with a tightly coupled allocation"); 4929 // restore JVM state to the state at the arraycopy 4930 saved_jvms->map()->set_control(map()->control()); 4931 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4932 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4933 // If we've improved the types of some nodes (null check) while 4934 // emitting the guards, propagate them to the current state 4935 map()->replaced_nodes().apply(saved_jvms->map(), new_idx); 4936 set_jvms(saved_jvms); 4937 _reexecute_sp = saved_reexecute_sp; 4938 4939 // Remove the allocation from above the guards 4940 CallProjections* callprojs = alloc->extract_projections(true); 4941 InitializeNode* init = alloc->initialization(); 4942 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4943 C->gvn_replace_by(callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4944 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4945 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4946 4947 // move the allocation here (after the guards) 4948 _gvn.hash_delete(alloc); 4949 alloc->set_req(TypeFunc::Control, control()); 4950 alloc->set_req(TypeFunc::I_O, i_o()); 4951 Node *mem = reset_memory(); 4952 set_all_memory(mem); 4953 alloc->set_req(TypeFunc::Memory, mem); 4954 set_control(init->proj_out_or_null(TypeFunc::Control)); 4955 set_i_o(callprojs->fallthrough_ioproj); 4956 4957 // Update memory as done in GraphKit::set_output_for_allocation() 4958 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4959 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4960 if (ary_type->isa_aryptr() && length_type != NULL) { 4961 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4962 } 4963 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4964 int elemidx = C->get_alias_index(telemref); 4965 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw); 4966 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx); 4967 4968 Node* allocx = _gvn.transform(alloc); 4969 assert(allocx == alloc, "where has the allocation gone?"); 4970 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4971 4972 _gvn.hash_delete(dest); 4973 dest->set_req(0, control()); 4974 Node* destx = _gvn.transform(dest); 4975 assert(destx == dest, "where has the allocation result gone?"); 4976 } 4977 } 4978 4979 4980 //------------------------------inline_arraycopy----------------------- 4981 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4982 // Object dest, int destPos, 4983 // int length); 4984 bool LibraryCallKit::inline_arraycopy() { 4985 // Get the arguments. 4986 Node* src = argument(0); // type: oop 4987 Node* src_offset = argument(1); // type: int 4988 Node* dest = argument(2); // type: oop 4989 Node* dest_offset = argument(3); // type: int 4990 Node* length = argument(4); // type: int 4991 4992 uint new_idx = C->unique(); 4993 4994 // Check for allocation before we add nodes that would confuse 4995 // tightly_coupled_allocation() 4996 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4997 4998 int saved_reexecute_sp = -1; 4999 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 5000 // See arraycopy_restore_alloc_state() comment 5001 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 5002 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 5003 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards 5004 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 5005 5006 // The following tests must be performed 5007 // (1) src and dest are arrays. 5008 // (2) src and dest arrays must have elements of the same BasicType 5009 // (3) src and dest must not be null. 5010 // (4) src_offset must not be negative. 5011 // (5) dest_offset must not be negative. 5012 // (6) length must not be negative. 5013 // (7) src_offset + length must not exceed length of src. 5014 // (8) dest_offset + length must not exceed length of dest. 5015 // (9) each element of an oop array must be assignable 5016 5017 // (3) src and dest must not be null. 5018 // always do this here because we need the JVM state for uncommon traps 5019 Node* null_ctl = top(); 5020 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 5021 assert(null_ctl->is_top(), "no null control here"); 5022 dest = null_check(dest, T_ARRAY); 5023 5024 if (!can_emit_guards) { 5025 // if saved_jvms == NULL and alloc != NULL, we don't emit any 5026 // guards but the arraycopy node could still take advantage of a 5027 // tightly allocated allocation. tightly_coupled_allocation() is 5028 // called again to make sure it takes the null check above into 5029 // account: the null check is mandatory and if it caused an 5030 // uncommon trap to be emitted then the allocation can't be 5031 // considered tightly coupled in this context. 5032 alloc = tightly_coupled_allocation(dest, NULL); 5033 } 5034 5035 bool validated = false; 5036 5037 const Type* src_type = _gvn.type(src); 5038 const Type* dest_type = _gvn.type(dest); 5039 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5040 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5041 5042 // Do we have the type of src? 5043 bool has_src = (top_src != NULL && top_src->klass() != NULL); 5044 // Do we have the type of dest? 5045 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 5046 // Is the type for src from speculation? 5047 bool src_spec = false; 5048 // Is the type for dest from speculation? 5049 bool dest_spec = false; 5050 5051 if ((!has_src || !has_dest) && can_emit_guards) { 5052 // We don't have sufficient type information, let's see if 5053 // speculative types can help. We need to have types for both src 5054 // and dest so that it pays off. 5055 5056 // Do we already have or could we have type information for src 5057 bool could_have_src = has_src; 5058 // Do we already have or could we have type information for dest 5059 bool could_have_dest = has_dest; 5060 5061 ciKlass* src_k = NULL; 5062 if (!has_src) { 5063 src_k = src_type->speculative_type_not_null(); 5064 if (src_k != NULL && src_k->is_array_klass()) { 5065 could_have_src = true; 5066 } 5067 } 5068 5069 ciKlass* dest_k = NULL; 5070 if (!has_dest) { 5071 dest_k = dest_type->speculative_type_not_null(); 5072 if (dest_k != NULL && dest_k->is_array_klass()) { 5073 could_have_dest = true; 5074 } 5075 } 5076 5077 if (could_have_src && could_have_dest) { 5078 // This is going to pay off so emit the required guards 5079 if (!has_src) { 5080 src = maybe_cast_profiled_obj(src, src_k, true); 5081 src_type = _gvn.type(src); 5082 top_src = src_type->isa_aryptr(); 5083 has_src = (top_src != NULL && top_src->klass() != NULL); 5084 src_spec = true; 5085 } 5086 if (!has_dest) { 5087 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5088 dest_type = _gvn.type(dest); 5089 top_dest = dest_type->isa_aryptr(); 5090 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 5091 dest_spec = true; 5092 } 5093 } 5094 } 5095 5096 if (has_src && has_dest && can_emit_guards) { 5097 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 5098 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 5099 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 5100 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 5101 5102 if (src_elem == dest_elem && src_elem == T_OBJECT) { 5103 // If both arrays are object arrays then having the exact types 5104 // for both will remove the need for a subtype check at runtime 5105 // before the call and may make it possible to pick a faster copy 5106 // routine (without a subtype check on every element) 5107 // Do we have the exact type of src? 5108 bool could_have_src = src_spec; 5109 // Do we have the exact type of dest? 5110 bool could_have_dest = dest_spec; 5111 ciKlass* src_k = top_src->klass(); 5112 ciKlass* dest_k = top_dest->klass(); 5113 if (!src_spec) { 5114 src_k = src_type->speculative_type_not_null(); 5115 if (src_k != NULL && src_k->is_array_klass()) { 5116 could_have_src = true; 5117 } 5118 } 5119 if (!dest_spec) { 5120 dest_k = dest_type->speculative_type_not_null(); 5121 if (dest_k != NULL && dest_k->is_array_klass()) { 5122 could_have_dest = true; 5123 } 5124 } 5125 if (could_have_src && could_have_dest) { 5126 // If we can have both exact types, emit the missing guards 5127 if (could_have_src && !src_spec) { 5128 src = maybe_cast_profiled_obj(src, src_k, true); 5129 } 5130 if (could_have_dest && !dest_spec) { 5131 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5132 } 5133 } 5134 } 5135 } 5136 5137 ciMethod* trap_method = method(); 5138 int trap_bci = bci(); 5139 if (saved_jvms != NULL) { 5140 trap_method = alloc->jvms()->method(); 5141 trap_bci = alloc->jvms()->bci(); 5142 } 5143 5144 bool negative_length_guard_generated = false; 5145 5146 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5147 can_emit_guards && 5148 !src->is_top() && !dest->is_top()) { 5149 // validate arguments: enables transformation the ArrayCopyNode 5150 validated = true; 5151 5152 RegionNode* slow_region = new RegionNode(1); 5153 record_for_igvn(slow_region); 5154 5155 // (1) src and dest are arrays. 5156 generate_non_array_guard(load_object_klass(src), slow_region); 5157 generate_non_array_guard(load_object_klass(dest), slow_region); 5158 5159 // (2) src and dest arrays must have elements of the same BasicType 5160 // done at macro expansion or at Ideal transformation time 5161 5162 // (4) src_offset must not be negative. 5163 generate_negative_guard(src_offset, slow_region); 5164 5165 // (5) dest_offset must not be negative. 5166 generate_negative_guard(dest_offset, slow_region); 5167 5168 // (7) src_offset + length must not exceed length of src. 5169 generate_limit_guard(src_offset, length, 5170 load_array_length(src), 5171 slow_region); 5172 5173 // (8) dest_offset + length must not exceed length of dest. 5174 generate_limit_guard(dest_offset, length, 5175 load_array_length(dest), 5176 slow_region); 5177 5178 // (6) length must not be negative. 5179 // This is also checked in generate_arraycopy() during macro expansion, but 5180 // we also have to check it here for the case where the ArrayCopyNode will 5181 // be eliminated by Escape Analysis. 5182 if (EliminateAllocations) { 5183 generate_negative_guard(length, slow_region); 5184 negative_length_guard_generated = true; 5185 } 5186 5187 // (9) each element of an oop array must be assignable 5188 Node* src_klass = load_object_klass(src); 5189 Node* dest_klass = load_object_klass(dest); 5190 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5191 5192 if (not_subtype_ctrl != top()) { 5193 PreserveJVMState pjvms(this); 5194 set_control(not_subtype_ctrl); 5195 uncommon_trap(Deoptimization::Reason_intrinsic, 5196 Deoptimization::Action_make_not_entrant); 5197 assert(stopped(), "Should be stopped"); 5198 } 5199 5200 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr(); 5201 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass()); 5202 src = _gvn.transform(new CheckCastPPNode(control(), src, toop)); 5203 5204 src_type = _gvn.type(src); 5205 top_src = src_type->isa_aryptr(); 5206 5207 if (top_dest != NULL && 5208 top_dest->elem()->make_oopptr() != NULL && 5209 top_dest->elem()->make_oopptr()->can_be_value_type()) { 5210 generate_valueArray_guard(load_object_klass(dest), slow_region); 5211 } 5212 5213 if (top_src != NULL && 5214 top_src->elem()->make_oopptr() != NULL && 5215 top_src->elem()->make_oopptr()->can_be_value_type()) { 5216 generate_valueArray_guard(load_object_klass(src), slow_region); 5217 } 5218 5219 { 5220 PreserveJVMState pjvms(this); 5221 set_control(_gvn.transform(slow_region)); 5222 uncommon_trap(Deoptimization::Reason_intrinsic, 5223 Deoptimization::Action_make_not_entrant); 5224 assert(stopped(), "Should be stopped"); 5225 } 5226 } 5227 5228 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx); 5229 5230 if (stopped()) { 5231 return true; 5232 } 5233 5234 Node* new_src = access_resolve(src, ACCESS_READ); 5235 Node* new_dest = access_resolve(dest, ACCESS_WRITE); 5236 5237 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, new_src, src_offset, new_dest, dest_offset, length, alloc != NULL, negative_length_guard_generated, 5238 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5239 // so the compiler has a chance to eliminate them: during macro expansion, 5240 // we have to set their control (CastPP nodes are eliminated). 5241 load_object_klass(src), load_object_klass(dest), 5242 load_array_length(src), load_array_length(dest)); 5243 5244 ac->set_arraycopy(validated); 5245 5246 Node* n = _gvn.transform(ac); 5247 if (n == ac) { 5248 ac->connect_outputs(this); 5249 } else { 5250 assert(validated, "shouldn't transform if all arguments not validated"); 5251 set_all_memory(n); 5252 } 5253 clear_upper_avx(); 5254 5255 5256 return true; 5257 } 5258 5259 5260 // Helper function which determines if an arraycopy immediately follows 5261 // an allocation, with no intervening tests or other escapes for the object. 5262 AllocateArrayNode* 5263 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5264 RegionNode* slow_region) { 5265 if (stopped()) return NULL; // no fast path 5266 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5267 5268 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5269 if (alloc == NULL) return NULL; 5270 5271 Node* rawmem = memory(Compile::AliasIdxRaw); 5272 // Is the allocation's memory state untouched? 5273 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5274 // Bail out if there have been raw-memory effects since the allocation. 5275 // (Example: There might have been a call or safepoint.) 5276 return NULL; 5277 } 5278 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5279 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5280 return NULL; 5281 } 5282 5283 // There must be no unexpected observers of this allocation. 5284 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5285 Node* obs = ptr->fast_out(i); 5286 if (obs != this->map()) { 5287 return NULL; 5288 } 5289 } 5290 5291 // This arraycopy must unconditionally follow the allocation of the ptr. 5292 Node* alloc_ctl = ptr->in(0); 5293 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5294 5295 Node* ctl = control(); 5296 while (ctl != alloc_ctl) { 5297 // There may be guards which feed into the slow_region. 5298 // Any other control flow means that we might not get a chance 5299 // to finish initializing the allocated object. 5300 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5301 IfNode* iff = ctl->in(0)->as_If(); 5302 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con); 5303 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5304 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5305 ctl = iff->in(0); // This test feeds the known slow_region. 5306 continue; 5307 } 5308 // One more try: Various low-level checks bottom out in 5309 // uncommon traps. If the debug-info of the trap omits 5310 // any reference to the allocation, as we've already 5311 // observed, then there can be no objection to the trap. 5312 bool found_trap = false; 5313 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5314 Node* obs = not_ctl->fast_out(j); 5315 if (obs->in(0) == not_ctl && obs->is_Call() && 5316 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5317 found_trap = true; break; 5318 } 5319 } 5320 if (found_trap) { 5321 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5322 continue; 5323 } 5324 } 5325 return NULL; 5326 } 5327 5328 // If we get this far, we have an allocation which immediately 5329 // precedes the arraycopy, and we can take over zeroing the new object. 5330 // The arraycopy will finish the initialization, and provide 5331 // a new control state to which we will anchor the destination pointer. 5332 5333 return alloc; 5334 } 5335 5336 //-------------inline_encodeISOArray----------------------------------- 5337 // encode char[] to byte[] in ISO_8859_1 5338 bool LibraryCallKit::inline_encodeISOArray() { 5339 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5340 // no receiver since it is static method 5341 Node *src = argument(0); 5342 Node *src_offset = argument(1); 5343 Node *dst = argument(2); 5344 Node *dst_offset = argument(3); 5345 Node *length = argument(4); 5346 5347 src = must_be_not_null(src, true); 5348 dst = must_be_not_null(dst, true); 5349 5350 src = access_resolve(src, ACCESS_READ); 5351 dst = access_resolve(dst, ACCESS_WRITE); 5352 5353 const Type* src_type = src->Value(&_gvn); 5354 const Type* dst_type = dst->Value(&_gvn); 5355 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5356 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5357 if (top_src == NULL || top_src->klass() == NULL || 5358 top_dest == NULL || top_dest->klass() == NULL) { 5359 // failed array check 5360 return false; 5361 } 5362 5363 // Figure out the size and type of the elements we will be copying. 5364 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5365 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5366 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5367 return false; 5368 } 5369 5370 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5371 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5372 // 'src_start' points to src array + scaled offset 5373 // 'dst_start' points to dst array + scaled offset 5374 5375 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5376 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5377 enc = _gvn.transform(enc); 5378 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5379 set_memory(res_mem, mtype); 5380 set_result(enc); 5381 clear_upper_avx(); 5382 5383 return true; 5384 } 5385 5386 //-------------inline_multiplyToLen----------------------------------- 5387 bool LibraryCallKit::inline_multiplyToLen() { 5388 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5389 5390 address stubAddr = StubRoutines::multiplyToLen(); 5391 if (stubAddr == NULL) { 5392 return false; // Intrinsic's stub is not implemented on this platform 5393 } 5394 const char* stubName = "multiplyToLen"; 5395 5396 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5397 5398 // no receiver because it is a static method 5399 Node* x = argument(0); 5400 Node* xlen = argument(1); 5401 Node* y = argument(2); 5402 Node* ylen = argument(3); 5403 Node* z = argument(4); 5404 5405 x = must_be_not_null(x, true); 5406 y = must_be_not_null(y, true); 5407 5408 x = access_resolve(x, ACCESS_READ); 5409 y = access_resolve(y, ACCESS_READ); 5410 z = access_resolve(z, ACCESS_WRITE); 5411 5412 const Type* x_type = x->Value(&_gvn); 5413 const Type* y_type = y->Value(&_gvn); 5414 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5415 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5416 if (top_x == NULL || top_x->klass() == NULL || 5417 top_y == NULL || top_y->klass() == NULL) { 5418 // failed array check 5419 return false; 5420 } 5421 5422 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5423 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5424 if (x_elem != T_INT || y_elem != T_INT) { 5425 return false; 5426 } 5427 5428 // Set the original stack and the reexecute bit for the interpreter to reexecute 5429 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5430 // on the return from z array allocation in runtime. 5431 { PreserveReexecuteState preexecs(this); 5432 jvms()->set_should_reexecute(true); 5433 5434 Node* x_start = array_element_address(x, intcon(0), x_elem); 5435 Node* y_start = array_element_address(y, intcon(0), y_elem); 5436 // 'x_start' points to x array + scaled xlen 5437 // 'y_start' points to y array + scaled ylen 5438 5439 // Allocate the result array 5440 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5441 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5442 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5443 5444 IdealKit ideal(this); 5445 5446 #define __ ideal. 5447 Node* one = __ ConI(1); 5448 Node* zero = __ ConI(0); 5449 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5450 __ set(need_alloc, zero); 5451 __ set(z_alloc, z); 5452 __ if_then(z, BoolTest::eq, null()); { 5453 __ increment (need_alloc, one); 5454 } __ else_(); { 5455 // Update graphKit memory and control from IdealKit. 5456 sync_kit(ideal); 5457 Node *cast = new CastPPNode(z, TypePtr::NOTNULL); 5458 cast->init_req(0, control()); 5459 _gvn.set_type(cast, cast->bottom_type()); 5460 C->record_for_igvn(cast); 5461 5462 Node* zlen_arg = load_array_length(cast); 5463 // Update IdealKit memory and control from graphKit. 5464 __ sync_kit(this); 5465 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5466 __ increment (need_alloc, one); 5467 } __ end_if(); 5468 } __ end_if(); 5469 5470 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5471 // Update graphKit memory and control from IdealKit. 5472 sync_kit(ideal); 5473 Node * narr = new_array(klass_node, zlen, 1); 5474 // Update IdealKit memory and control from graphKit. 5475 __ sync_kit(this); 5476 __ set(z_alloc, narr); 5477 } __ end_if(); 5478 5479 sync_kit(ideal); 5480 z = __ value(z_alloc); 5481 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5482 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5483 // Final sync IdealKit and GraphKit. 5484 final_sync(ideal); 5485 #undef __ 5486 5487 Node* z_start = array_element_address(z, intcon(0), T_INT); 5488 5489 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5490 OptoRuntime::multiplyToLen_Type(), 5491 stubAddr, stubName, TypePtr::BOTTOM, 5492 x_start, xlen, y_start, ylen, z_start, zlen); 5493 } // original reexecute is set back here 5494 5495 C->set_has_split_ifs(true); // Has chance for split-if optimization 5496 set_result(z); 5497 return true; 5498 } 5499 5500 //-------------inline_squareToLen------------------------------------ 5501 bool LibraryCallKit::inline_squareToLen() { 5502 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5503 5504 address stubAddr = StubRoutines::squareToLen(); 5505 if (stubAddr == NULL) { 5506 return false; // Intrinsic's stub is not implemented on this platform 5507 } 5508 const char* stubName = "squareToLen"; 5509 5510 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5511 5512 Node* x = argument(0); 5513 Node* len = argument(1); 5514 Node* z = argument(2); 5515 Node* zlen = argument(3); 5516 5517 x = must_be_not_null(x, true); 5518 z = must_be_not_null(z, true); 5519 5520 x = access_resolve(x, ACCESS_READ); 5521 z = access_resolve(z, ACCESS_WRITE); 5522 5523 const Type* x_type = x->Value(&_gvn); 5524 const Type* z_type = z->Value(&_gvn); 5525 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5526 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5527 if (top_x == NULL || top_x->klass() == NULL || 5528 top_z == NULL || top_z->klass() == NULL) { 5529 // failed array check 5530 return false; 5531 } 5532 5533 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5534 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5535 if (x_elem != T_INT || z_elem != T_INT) { 5536 return false; 5537 } 5538 5539 5540 Node* x_start = array_element_address(x, intcon(0), x_elem); 5541 Node* z_start = array_element_address(z, intcon(0), z_elem); 5542 5543 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5544 OptoRuntime::squareToLen_Type(), 5545 stubAddr, stubName, TypePtr::BOTTOM, 5546 x_start, len, z_start, zlen); 5547 5548 set_result(z); 5549 return true; 5550 } 5551 5552 //-------------inline_mulAdd------------------------------------------ 5553 bool LibraryCallKit::inline_mulAdd() { 5554 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5555 5556 address stubAddr = StubRoutines::mulAdd(); 5557 if (stubAddr == NULL) { 5558 return false; // Intrinsic's stub is not implemented on this platform 5559 } 5560 const char* stubName = "mulAdd"; 5561 5562 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5563 5564 Node* out = argument(0); 5565 Node* in = argument(1); 5566 Node* offset = argument(2); 5567 Node* len = argument(3); 5568 Node* k = argument(4); 5569 5570 out = must_be_not_null(out, true); 5571 5572 in = access_resolve(in, ACCESS_READ); 5573 out = access_resolve(out, ACCESS_WRITE); 5574 5575 const Type* out_type = out->Value(&_gvn); 5576 const Type* in_type = in->Value(&_gvn); 5577 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5578 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5579 if (top_out == NULL || top_out->klass() == NULL || 5580 top_in == NULL || top_in->klass() == NULL) { 5581 // failed array check 5582 return false; 5583 } 5584 5585 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5586 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5587 if (out_elem != T_INT || in_elem != T_INT) { 5588 return false; 5589 } 5590 5591 Node* outlen = load_array_length(out); 5592 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5593 Node* out_start = array_element_address(out, intcon(0), out_elem); 5594 Node* in_start = array_element_address(in, intcon(0), in_elem); 5595 5596 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5597 OptoRuntime::mulAdd_Type(), 5598 stubAddr, stubName, TypePtr::BOTTOM, 5599 out_start,in_start, new_offset, len, k); 5600 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5601 set_result(result); 5602 return true; 5603 } 5604 5605 //-------------inline_montgomeryMultiply----------------------------------- 5606 bool LibraryCallKit::inline_montgomeryMultiply() { 5607 address stubAddr = StubRoutines::montgomeryMultiply(); 5608 if (stubAddr == NULL) { 5609 return false; // Intrinsic's stub is not implemented on this platform 5610 } 5611 5612 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5613 const char* stubName = "montgomery_multiply"; 5614 5615 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5616 5617 Node* a = argument(0); 5618 Node* b = argument(1); 5619 Node* n = argument(2); 5620 Node* len = argument(3); 5621 Node* inv = argument(4); 5622 Node* m = argument(6); 5623 5624 a = access_resolve(a, ACCESS_READ); 5625 b = access_resolve(b, ACCESS_READ); 5626 n = access_resolve(n, ACCESS_READ); 5627 m = access_resolve(m, ACCESS_WRITE); 5628 5629 const Type* a_type = a->Value(&_gvn); 5630 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5631 const Type* b_type = b->Value(&_gvn); 5632 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5633 const Type* n_type = a->Value(&_gvn); 5634 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5635 const Type* m_type = a->Value(&_gvn); 5636 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5637 if (top_a == NULL || top_a->klass() == NULL || 5638 top_b == NULL || top_b->klass() == NULL || 5639 top_n == NULL || top_n->klass() == NULL || 5640 top_m == NULL || top_m->klass() == NULL) { 5641 // failed array check 5642 return false; 5643 } 5644 5645 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5646 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5647 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5648 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5649 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5650 return false; 5651 } 5652 5653 // Make the call 5654 { 5655 Node* a_start = array_element_address(a, intcon(0), a_elem); 5656 Node* b_start = array_element_address(b, intcon(0), b_elem); 5657 Node* n_start = array_element_address(n, intcon(0), n_elem); 5658 Node* m_start = array_element_address(m, intcon(0), m_elem); 5659 5660 Node* call = make_runtime_call(RC_LEAF, 5661 OptoRuntime::montgomeryMultiply_Type(), 5662 stubAddr, stubName, TypePtr::BOTTOM, 5663 a_start, b_start, n_start, len, inv, top(), 5664 m_start); 5665 set_result(m); 5666 } 5667 5668 return true; 5669 } 5670 5671 bool LibraryCallKit::inline_montgomerySquare() { 5672 address stubAddr = StubRoutines::montgomerySquare(); 5673 if (stubAddr == NULL) { 5674 return false; // Intrinsic's stub is not implemented on this platform 5675 } 5676 5677 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5678 const char* stubName = "montgomery_square"; 5679 5680 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5681 5682 Node* a = argument(0); 5683 Node* n = argument(1); 5684 Node* len = argument(2); 5685 Node* inv = argument(3); 5686 Node* m = argument(5); 5687 5688 a = access_resolve(a, ACCESS_READ); 5689 n = access_resolve(n, ACCESS_READ); 5690 m = access_resolve(m, ACCESS_WRITE); 5691 5692 const Type* a_type = a->Value(&_gvn); 5693 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5694 const Type* n_type = a->Value(&_gvn); 5695 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5696 const Type* m_type = a->Value(&_gvn); 5697 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5698 if (top_a == NULL || top_a->klass() == NULL || 5699 top_n == NULL || top_n->klass() == NULL || 5700 top_m == NULL || top_m->klass() == NULL) { 5701 // failed array check 5702 return false; 5703 } 5704 5705 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5706 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5707 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5708 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5709 return false; 5710 } 5711 5712 // Make the call 5713 { 5714 Node* a_start = array_element_address(a, intcon(0), a_elem); 5715 Node* n_start = array_element_address(n, intcon(0), n_elem); 5716 Node* m_start = array_element_address(m, intcon(0), m_elem); 5717 5718 Node* call = make_runtime_call(RC_LEAF, 5719 OptoRuntime::montgomerySquare_Type(), 5720 stubAddr, stubName, TypePtr::BOTTOM, 5721 a_start, n_start, len, inv, top(), 5722 m_start); 5723 set_result(m); 5724 } 5725 5726 return true; 5727 } 5728 5729 //-------------inline_vectorizedMismatch------------------------------ 5730 bool LibraryCallKit::inline_vectorizedMismatch() { 5731 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5732 5733 address stubAddr = StubRoutines::vectorizedMismatch(); 5734 if (stubAddr == NULL) { 5735 return false; // Intrinsic's stub is not implemented on this platform 5736 } 5737 const char* stubName = "vectorizedMismatch"; 5738 int size_l = callee()->signature()->size(); 5739 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5740 5741 Node* obja = argument(0); 5742 Node* aoffset = argument(1); 5743 Node* objb = argument(3); 5744 Node* boffset = argument(4); 5745 Node* length = argument(6); 5746 Node* scale = argument(7); 5747 5748 const Type* a_type = obja->Value(&_gvn); 5749 const Type* b_type = objb->Value(&_gvn); 5750 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5751 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5752 if (top_a == NULL || top_a->klass() == NULL || 5753 top_b == NULL || top_b->klass() == NULL) { 5754 // failed array check 5755 return false; 5756 } 5757 5758 Node* call; 5759 jvms()->set_should_reexecute(true); 5760 5761 obja = access_resolve(obja, ACCESS_READ); 5762 objb = access_resolve(objb, ACCESS_READ); 5763 Node* obja_adr = make_unsafe_address(obja, aoffset, ACCESS_READ); 5764 Node* objb_adr = make_unsafe_address(objb, boffset, ACCESS_READ); 5765 5766 call = make_runtime_call(RC_LEAF, 5767 OptoRuntime::vectorizedMismatch_Type(), 5768 stubAddr, stubName, TypePtr::BOTTOM, 5769 obja_adr, objb_adr, length, scale); 5770 5771 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5772 set_result(result); 5773 return true; 5774 } 5775 5776 /** 5777 * Calculate CRC32 for byte. 5778 * int java.util.zip.CRC32.update(int crc, int b) 5779 */ 5780 bool LibraryCallKit::inline_updateCRC32() { 5781 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5782 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5783 // no receiver since it is static method 5784 Node* crc = argument(0); // type: int 5785 Node* b = argument(1); // type: int 5786 5787 /* 5788 * int c = ~ crc; 5789 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5790 * b = b ^ (c >>> 8); 5791 * crc = ~b; 5792 */ 5793 5794 Node* M1 = intcon(-1); 5795 crc = _gvn.transform(new XorINode(crc, M1)); 5796 Node* result = _gvn.transform(new XorINode(crc, b)); 5797 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5798 5799 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5800 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5801 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5802 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5803 5804 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5805 result = _gvn.transform(new XorINode(crc, result)); 5806 result = _gvn.transform(new XorINode(result, M1)); 5807 set_result(result); 5808 return true; 5809 } 5810 5811 /** 5812 * Calculate CRC32 for byte[] array. 5813 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5814 */ 5815 bool LibraryCallKit::inline_updateBytesCRC32() { 5816 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5817 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5818 // no receiver since it is static method 5819 Node* crc = argument(0); // type: int 5820 Node* src = argument(1); // type: oop 5821 Node* offset = argument(2); // type: int 5822 Node* length = argument(3); // type: int 5823 5824 const Type* src_type = src->Value(&_gvn); 5825 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5826 if (top_src == NULL || top_src->klass() == NULL) { 5827 // failed array check 5828 return false; 5829 } 5830 5831 // Figure out the size and type of the elements we will be copying. 5832 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5833 if (src_elem != T_BYTE) { 5834 return false; 5835 } 5836 5837 // 'src_start' points to src array + scaled offset 5838 src = must_be_not_null(src, true); 5839 src = access_resolve(src, ACCESS_READ); 5840 Node* src_start = array_element_address(src, offset, src_elem); 5841 5842 // We assume that range check is done by caller. 5843 // TODO: generate range check (offset+length < src.length) in debug VM. 5844 5845 // Call the stub. 5846 address stubAddr = StubRoutines::updateBytesCRC32(); 5847 const char *stubName = "updateBytesCRC32"; 5848 5849 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5850 stubAddr, stubName, TypePtr::BOTTOM, 5851 crc, src_start, length); 5852 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5853 set_result(result); 5854 return true; 5855 } 5856 5857 /** 5858 * Calculate CRC32 for ByteBuffer. 5859 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5860 */ 5861 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5862 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5863 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5864 // no receiver since it is static method 5865 Node* crc = argument(0); // type: int 5866 Node* src = argument(1); // type: long 5867 Node* offset = argument(3); // type: int 5868 Node* length = argument(4); // type: int 5869 5870 src = ConvL2X(src); // adjust Java long to machine word 5871 Node* base = _gvn.transform(new CastX2PNode(src)); 5872 offset = ConvI2X(offset); 5873 5874 // 'src_start' points to src array + scaled offset 5875 Node* src_start = basic_plus_adr(top(), base, offset); 5876 5877 // Call the stub. 5878 address stubAddr = StubRoutines::updateBytesCRC32(); 5879 const char *stubName = "updateBytesCRC32"; 5880 5881 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5882 stubAddr, stubName, TypePtr::BOTTOM, 5883 crc, src_start, length); 5884 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5885 set_result(result); 5886 return true; 5887 } 5888 5889 //------------------------------get_table_from_crc32c_class----------------------- 5890 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5891 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5892 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5893 5894 return table; 5895 } 5896 5897 //------------------------------inline_updateBytesCRC32C----------------------- 5898 // 5899 // Calculate CRC32C for byte[] array. 5900 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5901 // 5902 bool LibraryCallKit::inline_updateBytesCRC32C() { 5903 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5904 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5905 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5906 // no receiver since it is a static method 5907 Node* crc = argument(0); // type: int 5908 Node* src = argument(1); // type: oop 5909 Node* offset = argument(2); // type: int 5910 Node* end = argument(3); // type: int 5911 5912 Node* length = _gvn.transform(new SubINode(end, offset)); 5913 5914 const Type* src_type = src->Value(&_gvn); 5915 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5916 if (top_src == NULL || top_src->klass() == NULL) { 5917 // failed array check 5918 return false; 5919 } 5920 5921 // Figure out the size and type of the elements we will be copying. 5922 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5923 if (src_elem != T_BYTE) { 5924 return false; 5925 } 5926 5927 // 'src_start' points to src array + scaled offset 5928 src = must_be_not_null(src, true); 5929 src = access_resolve(src, ACCESS_READ); 5930 Node* src_start = array_element_address(src, offset, src_elem); 5931 5932 // static final int[] byteTable in class CRC32C 5933 Node* table = get_table_from_crc32c_class(callee()->holder()); 5934 table = must_be_not_null(table, true); 5935 table = access_resolve(table, ACCESS_READ); 5936 Node* table_start = array_element_address(table, intcon(0), T_INT); 5937 5938 // We assume that range check is done by caller. 5939 // TODO: generate range check (offset+length < src.length) in debug VM. 5940 5941 // Call the stub. 5942 address stubAddr = StubRoutines::updateBytesCRC32C(); 5943 const char *stubName = "updateBytesCRC32C"; 5944 5945 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5946 stubAddr, stubName, TypePtr::BOTTOM, 5947 crc, src_start, length, table_start); 5948 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5949 set_result(result); 5950 return true; 5951 } 5952 5953 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5954 // 5955 // Calculate CRC32C for DirectByteBuffer. 5956 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5957 // 5958 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5959 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5960 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5961 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5962 // no receiver since it is a static method 5963 Node* crc = argument(0); // type: int 5964 Node* src = argument(1); // type: long 5965 Node* offset = argument(3); // type: int 5966 Node* end = argument(4); // type: int 5967 5968 Node* length = _gvn.transform(new SubINode(end, offset)); 5969 5970 src = ConvL2X(src); // adjust Java long to machine word 5971 Node* base = _gvn.transform(new CastX2PNode(src)); 5972 offset = ConvI2X(offset); 5973 5974 // 'src_start' points to src array + scaled offset 5975 Node* src_start = basic_plus_adr(top(), base, offset); 5976 5977 // static final int[] byteTable in class CRC32C 5978 Node* table = get_table_from_crc32c_class(callee()->holder()); 5979 table = must_be_not_null(table, true); 5980 table = access_resolve(table, ACCESS_READ); 5981 Node* table_start = array_element_address(table, intcon(0), T_INT); 5982 5983 // Call the stub. 5984 address stubAddr = StubRoutines::updateBytesCRC32C(); 5985 const char *stubName = "updateBytesCRC32C"; 5986 5987 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5988 stubAddr, stubName, TypePtr::BOTTOM, 5989 crc, src_start, length, table_start); 5990 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5991 set_result(result); 5992 return true; 5993 } 5994 5995 //------------------------------inline_updateBytesAdler32---------------------- 5996 // 5997 // Calculate Adler32 checksum for byte[] array. 5998 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5999 // 6000 bool LibraryCallKit::inline_updateBytesAdler32() { 6001 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 6002 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 6003 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 6004 // no receiver since it is static method 6005 Node* crc = argument(0); // type: int 6006 Node* src = argument(1); // type: oop 6007 Node* offset = argument(2); // type: int 6008 Node* length = argument(3); // type: int 6009 6010 const Type* src_type = src->Value(&_gvn); 6011 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6012 if (top_src == NULL || top_src->klass() == NULL) { 6013 // failed array check 6014 return false; 6015 } 6016 6017 // Figure out the size and type of the elements we will be copying. 6018 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6019 if (src_elem != T_BYTE) { 6020 return false; 6021 } 6022 6023 // 'src_start' points to src array + scaled offset 6024 src = access_resolve(src, ACCESS_READ); 6025 Node* src_start = array_element_address(src, offset, src_elem); 6026 6027 // We assume that range check is done by caller. 6028 // TODO: generate range check (offset+length < src.length) in debug VM. 6029 6030 // Call the stub. 6031 address stubAddr = StubRoutines::updateBytesAdler32(); 6032 const char *stubName = "updateBytesAdler32"; 6033 6034 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 6035 stubAddr, stubName, TypePtr::BOTTOM, 6036 crc, src_start, length); 6037 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6038 set_result(result); 6039 return true; 6040 } 6041 6042 //------------------------------inline_updateByteBufferAdler32--------------- 6043 // 6044 // Calculate Adler32 checksum for DirectByteBuffer. 6045 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 6046 // 6047 bool LibraryCallKit::inline_updateByteBufferAdler32() { 6048 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 6049 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 6050 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 6051 // no receiver since it is static method 6052 Node* crc = argument(0); // type: int 6053 Node* src = argument(1); // type: long 6054 Node* offset = argument(3); // type: int 6055 Node* length = argument(4); // type: int 6056 6057 src = ConvL2X(src); // adjust Java long to machine word 6058 Node* base = _gvn.transform(new CastX2PNode(src)); 6059 offset = ConvI2X(offset); 6060 6061 // 'src_start' points to src array + scaled offset 6062 Node* src_start = basic_plus_adr(top(), base, offset); 6063 6064 // Call the stub. 6065 address stubAddr = StubRoutines::updateBytesAdler32(); 6066 const char *stubName = "updateBytesAdler32"; 6067 6068 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 6069 stubAddr, stubName, TypePtr::BOTTOM, 6070 crc, src_start, length); 6071 6072 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6073 set_result(result); 6074 return true; 6075 } 6076 6077 //----------------------------inline_reference_get---------------------------- 6078 // public T java.lang.ref.Reference.get(); 6079 bool LibraryCallKit::inline_reference_get() { 6080 const int referent_offset = java_lang_ref_Reference::referent_offset; 6081 guarantee(referent_offset > 0, "should have already been set"); 6082 6083 // Get the argument: 6084 Node* reference_obj = null_check_receiver(); 6085 if (stopped()) return true; 6086 6087 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr(); 6088 assert(tinst != NULL, "obj is null"); 6089 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6090 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass(); 6091 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"), 6092 ciSymbol::make("Ljava/lang/Object;"), 6093 false); 6094 assert (field != NULL, "undefined field"); 6095 6096 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 6097 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 6098 6099 ciInstanceKlass* klass = env()->Object_klass(); 6100 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 6101 6102 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF; 6103 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators); 6104 // Add memory barrier to prevent commoning reads from this field 6105 // across safepoint since GC can change its value. 6106 insert_mem_bar(Op_MemBarCPUOrder); 6107 6108 set_result(result); 6109 return true; 6110 } 6111 6112 6113 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 6114 bool is_exact=true, bool is_static=false, 6115 ciInstanceKlass * fromKls=NULL) { 6116 if (fromKls == NULL) { 6117 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 6118 assert(tinst != NULL, "obj is null"); 6119 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6120 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 6121 fromKls = tinst->klass()->as_instance_klass(); 6122 } else { 6123 assert(is_static, "only for static field access"); 6124 } 6125 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 6126 ciSymbol::make(fieldTypeString), 6127 is_static); 6128 6129 assert (field != NULL, "undefined field"); 6130 if (field == NULL) return (Node *) NULL; 6131 6132 if (is_static) { 6133 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 6134 fromObj = makecon(tip); 6135 } 6136 6137 // Next code copied from Parse::do_get_xxx(): 6138 6139 // Compute address and memory type. 6140 int offset = field->offset_in_bytes(); 6141 bool is_vol = field->is_volatile(); 6142 ciType* field_klass = field->type(); 6143 assert(field_klass->is_loaded(), "should be loaded"); 6144 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 6145 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 6146 BasicType bt = field->layout_type(); 6147 6148 // Build the resultant type of the load 6149 const Type *type; 6150 if (bt == T_OBJECT) { 6151 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 6152 } else { 6153 type = Type::get_const_basic_type(bt); 6154 } 6155 6156 DecoratorSet decorators = IN_HEAP; 6157 6158 if (is_vol) { 6159 decorators |= MO_SEQ_CST; 6160 } 6161 6162 return access_load_at(fromObj, adr, adr_type, type, bt, decorators); 6163 } 6164 6165 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 6166 bool is_exact = true, bool is_static = false, 6167 ciInstanceKlass * fromKls = NULL) { 6168 if (fromKls == NULL) { 6169 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 6170 assert(tinst != NULL, "obj is null"); 6171 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6172 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 6173 fromKls = tinst->klass()->as_instance_klass(); 6174 } 6175 else { 6176 assert(is_static, "only for static field access"); 6177 } 6178 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 6179 ciSymbol::make(fieldTypeString), 6180 is_static); 6181 6182 assert(field != NULL, "undefined field"); 6183 assert(!field->is_volatile(), "not defined for volatile fields"); 6184 6185 if (is_static) { 6186 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 6187 fromObj = makecon(tip); 6188 } 6189 6190 // Next code copied from Parse::do_get_xxx(): 6191 6192 // Compute address and memory type. 6193 int offset = field->offset_in_bytes(); 6194 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 6195 6196 return adr; 6197 } 6198 6199 //------------------------------inline_aescrypt_Block----------------------- 6200 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 6201 address stubAddr = NULL; 6202 const char *stubName; 6203 assert(UseAES, "need AES instruction support"); 6204 6205 switch(id) { 6206 case vmIntrinsics::_aescrypt_encryptBlock: 6207 stubAddr = StubRoutines::aescrypt_encryptBlock(); 6208 stubName = "aescrypt_encryptBlock"; 6209 break; 6210 case vmIntrinsics::_aescrypt_decryptBlock: 6211 stubAddr = StubRoutines::aescrypt_decryptBlock(); 6212 stubName = "aescrypt_decryptBlock"; 6213 break; 6214 default: 6215 break; 6216 } 6217 if (stubAddr == NULL) return false; 6218 6219 Node* aescrypt_object = argument(0); 6220 Node* src = argument(1); 6221 Node* src_offset = argument(2); 6222 Node* dest = argument(3); 6223 Node* dest_offset = argument(4); 6224 6225 src = must_be_not_null(src, true); 6226 dest = must_be_not_null(dest, true); 6227 6228 src = access_resolve(src, ACCESS_READ); 6229 dest = access_resolve(dest, ACCESS_WRITE); 6230 6231 // (1) src and dest are arrays. 6232 const Type* src_type = src->Value(&_gvn); 6233 const Type* dest_type = dest->Value(&_gvn); 6234 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6235 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6236 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6237 6238 // for the quick and dirty code we will skip all the checks. 6239 // we are just trying to get the call to be generated. 6240 Node* src_start = src; 6241 Node* dest_start = dest; 6242 if (src_offset != NULL || dest_offset != NULL) { 6243 assert(src_offset != NULL && dest_offset != NULL, ""); 6244 src_start = array_element_address(src, src_offset, T_BYTE); 6245 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6246 } 6247 6248 // now need to get the start of its expanded key array 6249 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6250 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6251 if (k_start == NULL) return false; 6252 6253 if (Matcher::pass_original_key_for_aes()) { 6254 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6255 // compatibility issues between Java key expansion and SPARC crypto instructions 6256 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6257 if (original_k_start == NULL) return false; 6258 6259 // Call the stub. 6260 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6261 stubAddr, stubName, TypePtr::BOTTOM, 6262 src_start, dest_start, k_start, original_k_start); 6263 } else { 6264 // Call the stub. 6265 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6266 stubAddr, stubName, TypePtr::BOTTOM, 6267 src_start, dest_start, k_start); 6268 } 6269 6270 return true; 6271 } 6272 6273 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 6274 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 6275 address stubAddr = NULL; 6276 const char *stubName = NULL; 6277 6278 assert(UseAES, "need AES instruction support"); 6279 6280 switch(id) { 6281 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6282 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6283 stubName = "cipherBlockChaining_encryptAESCrypt"; 6284 break; 6285 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6286 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6287 stubName = "cipherBlockChaining_decryptAESCrypt"; 6288 break; 6289 default: 6290 break; 6291 } 6292 if (stubAddr == NULL) return false; 6293 6294 Node* cipherBlockChaining_object = argument(0); 6295 Node* src = argument(1); 6296 Node* src_offset = argument(2); 6297 Node* len = argument(3); 6298 Node* dest = argument(4); 6299 Node* dest_offset = argument(5); 6300 6301 src = must_be_not_null(src, false); 6302 dest = must_be_not_null(dest, false); 6303 6304 src = access_resolve(src, ACCESS_READ); 6305 dest = access_resolve(dest, ACCESS_WRITE); 6306 6307 // (1) src and dest are arrays. 6308 const Type* src_type = src->Value(&_gvn); 6309 const Type* dest_type = dest->Value(&_gvn); 6310 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6311 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6312 assert (top_src != NULL && top_src->klass() != NULL 6313 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6314 6315 // checks are the responsibility of the caller 6316 Node* src_start = src; 6317 Node* dest_start = dest; 6318 if (src_offset != NULL || dest_offset != NULL) { 6319 assert(src_offset != NULL && dest_offset != NULL, ""); 6320 src_start = array_element_address(src, src_offset, T_BYTE); 6321 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6322 } 6323 6324 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6325 // (because of the predicated logic executed earlier). 6326 // so we cast it here safely. 6327 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6328 6329 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6330 if (embeddedCipherObj == NULL) return false; 6331 6332 // cast it to what we know it will be at runtime 6333 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6334 assert(tinst != NULL, "CBC obj is null"); 6335 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6336 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6337 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6338 6339 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6340 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6341 const TypeOopPtr* xtype = aklass->as_instance_type(); 6342 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6343 aescrypt_object = _gvn.transform(aescrypt_object); 6344 6345 // we need to get the start of the aescrypt_object's expanded key array 6346 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6347 if (k_start == NULL) return false; 6348 6349 // similarly, get the start address of the r vector 6350 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6351 if (objRvec == NULL) return false; 6352 objRvec = access_resolve(objRvec, ACCESS_WRITE); 6353 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6354 6355 Node* cbcCrypt; 6356 if (Matcher::pass_original_key_for_aes()) { 6357 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6358 // compatibility issues between Java key expansion and SPARC crypto instructions 6359 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6360 if (original_k_start == NULL) return false; 6361 6362 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6363 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6364 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6365 stubAddr, stubName, TypePtr::BOTTOM, 6366 src_start, dest_start, k_start, r_start, len, original_k_start); 6367 } else { 6368 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6369 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6370 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6371 stubAddr, stubName, TypePtr::BOTTOM, 6372 src_start, dest_start, k_start, r_start, len); 6373 } 6374 6375 // return cipher length (int) 6376 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6377 set_result(retvalue); 6378 return true; 6379 } 6380 6381 //------------------------------inline_counterMode_AESCrypt----------------------- 6382 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6383 assert(UseAES, "need AES instruction support"); 6384 if (!UseAESCTRIntrinsics) return false; 6385 6386 address stubAddr = NULL; 6387 const char *stubName = NULL; 6388 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6389 stubAddr = StubRoutines::counterMode_AESCrypt(); 6390 stubName = "counterMode_AESCrypt"; 6391 } 6392 if (stubAddr == NULL) return false; 6393 6394 Node* counterMode_object = argument(0); 6395 Node* src = argument(1); 6396 Node* src_offset = argument(2); 6397 Node* len = argument(3); 6398 Node* dest = argument(4); 6399 Node* dest_offset = argument(5); 6400 6401 src = access_resolve(src, ACCESS_READ); 6402 dest = access_resolve(dest, ACCESS_WRITE); 6403 counterMode_object = access_resolve(counterMode_object, ACCESS_WRITE); 6404 6405 // (1) src and dest are arrays. 6406 const Type* src_type = src->Value(&_gvn); 6407 const Type* dest_type = dest->Value(&_gvn); 6408 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6409 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6410 assert(top_src != NULL && top_src->klass() != NULL && 6411 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6412 6413 // checks are the responsibility of the caller 6414 Node* src_start = src; 6415 Node* dest_start = dest; 6416 if (src_offset != NULL || dest_offset != NULL) { 6417 assert(src_offset != NULL && dest_offset != NULL, ""); 6418 src_start = array_element_address(src, src_offset, T_BYTE); 6419 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6420 } 6421 6422 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6423 // (because of the predicated logic executed earlier). 6424 // so we cast it here safely. 6425 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6426 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6427 if (embeddedCipherObj == NULL) return false; 6428 // cast it to what we know it will be at runtime 6429 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6430 assert(tinst != NULL, "CTR obj is null"); 6431 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6432 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6433 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6434 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6435 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6436 const TypeOopPtr* xtype = aklass->as_instance_type(); 6437 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6438 aescrypt_object = _gvn.transform(aescrypt_object); 6439 // we need to get the start of the aescrypt_object's expanded key array 6440 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6441 if (k_start == NULL) return false; 6442 // similarly, get the start address of the r vector 6443 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6444 if (obj_counter == NULL) return false; 6445 obj_counter = access_resolve(obj_counter, ACCESS_WRITE); 6446 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6447 6448 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6449 if (saved_encCounter == NULL) return false; 6450 saved_encCounter = access_resolve(saved_encCounter, ACCESS_WRITE); 6451 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6452 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6453 6454 Node* ctrCrypt; 6455 if (Matcher::pass_original_key_for_aes()) { 6456 // no SPARC version for AES/CTR intrinsics now. 6457 return false; 6458 } 6459 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6460 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6461 OptoRuntime::counterMode_aescrypt_Type(), 6462 stubAddr, stubName, TypePtr::BOTTOM, 6463 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6464 6465 // return cipher length (int) 6466 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6467 set_result(retvalue); 6468 return true; 6469 } 6470 6471 //------------------------------get_key_start_from_aescrypt_object----------------------- 6472 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6473 #if defined(PPC64) || defined(S390) 6474 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6475 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6476 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6477 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6478 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6479 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6480 if (objSessionK == NULL) { 6481 return (Node *) NULL; 6482 } 6483 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6484 #else 6485 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6486 #endif // PPC64 6487 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6488 if (objAESCryptKey == NULL) return (Node *) NULL; 6489 6490 // now have the array, need to get the start address of the K array 6491 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ); 6492 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6493 return k_start; 6494 } 6495 6496 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6497 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6498 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6499 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6500 if (objAESCryptKey == NULL) return (Node *) NULL; 6501 6502 // now have the array, need to get the start address of the lastKey array 6503 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ); 6504 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6505 return original_k_start; 6506 } 6507 6508 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6509 // Return node representing slow path of predicate check. 6510 // the pseudo code we want to emulate with this predicate is: 6511 // for encryption: 6512 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6513 // for decryption: 6514 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6515 // note cipher==plain is more conservative than the original java code but that's OK 6516 // 6517 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6518 // The receiver was checked for NULL already. 6519 Node* objCBC = argument(0); 6520 6521 Node* src = argument(1); 6522 Node* dest = argument(4); 6523 6524 // Load embeddedCipher field of CipherBlockChaining object. 6525 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6526 6527 // get AESCrypt klass for instanceOf check 6528 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6529 // will have same classloader as CipherBlockChaining object 6530 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6531 assert(tinst != NULL, "CBCobj is null"); 6532 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6533 6534 // we want to do an instanceof comparison against the AESCrypt class 6535 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6536 if (!klass_AESCrypt->is_loaded()) { 6537 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6538 Node* ctrl = control(); 6539 set_control(top()); // no regular fast path 6540 return ctrl; 6541 } 6542 6543 src = must_be_not_null(src, true); 6544 dest = must_be_not_null(dest, true); 6545 6546 // Resolve oops to stable for CmpP below. 6547 src = access_resolve(src, 0); 6548 dest = access_resolve(dest, 0); 6549 6550 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6551 6552 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6553 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6554 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6555 6556 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6557 6558 // for encryption, we are done 6559 if (!decrypting) 6560 return instof_false; // even if it is NULL 6561 6562 // for decryption, we need to add a further check to avoid 6563 // taking the intrinsic path when cipher and plain are the same 6564 // see the original java code for why. 6565 RegionNode* region = new RegionNode(3); 6566 region->init_req(1, instof_false); 6567 6568 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6569 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6570 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6571 region->init_req(2, src_dest_conjoint); 6572 6573 record_for_igvn(region); 6574 return _gvn.transform(region); 6575 } 6576 6577 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6578 // Return node representing slow path of predicate check. 6579 // the pseudo code we want to emulate with this predicate is: 6580 // for encryption: 6581 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6582 // for decryption: 6583 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6584 // note cipher==plain is more conservative than the original java code but that's OK 6585 // 6586 6587 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6588 // The receiver was checked for NULL already. 6589 Node* objCTR = argument(0); 6590 6591 // Load embeddedCipher field of CipherBlockChaining object. 6592 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6593 6594 // get AESCrypt klass for instanceOf check 6595 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6596 // will have same classloader as CipherBlockChaining object 6597 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6598 assert(tinst != NULL, "CTRobj is null"); 6599 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6600 6601 // we want to do an instanceof comparison against the AESCrypt class 6602 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6603 if (!klass_AESCrypt->is_loaded()) { 6604 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6605 Node* ctrl = control(); 6606 set_control(top()); // no regular fast path 6607 return ctrl; 6608 } 6609 6610 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6611 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6612 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6613 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6614 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6615 6616 return instof_false; // even if it is NULL 6617 } 6618 6619 //------------------------------inline_ghash_processBlocks 6620 bool LibraryCallKit::inline_ghash_processBlocks() { 6621 address stubAddr; 6622 const char *stubName; 6623 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6624 6625 stubAddr = StubRoutines::ghash_processBlocks(); 6626 stubName = "ghash_processBlocks"; 6627 6628 Node* data = argument(0); 6629 Node* offset = argument(1); 6630 Node* len = argument(2); 6631 Node* state = argument(3); 6632 Node* subkeyH = argument(4); 6633 6634 state = must_be_not_null(state, true); 6635 subkeyH = must_be_not_null(subkeyH, true); 6636 data = must_be_not_null(data, true); 6637 6638 state = access_resolve(state, ACCESS_WRITE); 6639 subkeyH = access_resolve(subkeyH, ACCESS_READ); 6640 data = access_resolve(data, ACCESS_READ); 6641 6642 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6643 assert(state_start, "state is NULL"); 6644 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6645 assert(subkeyH_start, "subkeyH is NULL"); 6646 Node* data_start = array_element_address(data, offset, T_BYTE); 6647 assert(data_start, "data is NULL"); 6648 6649 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6650 OptoRuntime::ghash_processBlocks_Type(), 6651 stubAddr, stubName, TypePtr::BOTTOM, 6652 state_start, subkeyH_start, data_start, len); 6653 return true; 6654 } 6655 6656 bool LibraryCallKit::inline_base64_encodeBlock() { 6657 address stubAddr; 6658 const char *stubName; 6659 assert(UseBASE64Intrinsics, "need Base64 intrinsics support"); 6660 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters"); 6661 stubAddr = StubRoutines::base64_encodeBlock(); 6662 stubName = "encodeBlock"; 6663 6664 if (!stubAddr) return false; 6665 Node* base64obj = argument(0); 6666 Node* src = argument(1); 6667 Node* offset = argument(2); 6668 Node* len = argument(3); 6669 Node* dest = argument(4); 6670 Node* dp = argument(5); 6671 Node* isURL = argument(6); 6672 6673 src = must_be_not_null(src, true); 6674 src = access_resolve(src, ACCESS_READ); 6675 dest = must_be_not_null(dest, true); 6676 dest = access_resolve(dest, ACCESS_WRITE); 6677 6678 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 6679 assert(src_start, "source array is NULL"); 6680 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE); 6681 assert(dest_start, "destination array is NULL"); 6682 6683 Node* base64 = make_runtime_call(RC_LEAF, 6684 OptoRuntime::base64_encodeBlock_Type(), 6685 stubAddr, stubName, TypePtr::BOTTOM, 6686 src_start, offset, len, dest_start, dp, isURL); 6687 return true; 6688 } 6689 6690 //------------------------------inline_sha_implCompress----------------------- 6691 // 6692 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6693 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6694 // 6695 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6696 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6697 // 6698 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6699 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6700 // 6701 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6702 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6703 6704 Node* sha_obj = argument(0); 6705 Node* src = argument(1); // type oop 6706 Node* ofs = argument(2); // type int 6707 6708 const Type* src_type = src->Value(&_gvn); 6709 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6710 if (top_src == NULL || top_src->klass() == NULL) { 6711 // failed array check 6712 return false; 6713 } 6714 // Figure out the size and type of the elements we will be copying. 6715 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6716 if (src_elem != T_BYTE) { 6717 return false; 6718 } 6719 // 'src_start' points to src array + offset 6720 src = must_be_not_null(src, true); 6721 src = access_resolve(src, ACCESS_READ); 6722 Node* src_start = array_element_address(src, ofs, src_elem); 6723 Node* state = NULL; 6724 address stubAddr; 6725 const char *stubName; 6726 6727 switch(id) { 6728 case vmIntrinsics::_sha_implCompress: 6729 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6730 state = get_state_from_sha_object(sha_obj); 6731 stubAddr = StubRoutines::sha1_implCompress(); 6732 stubName = "sha1_implCompress"; 6733 break; 6734 case vmIntrinsics::_sha2_implCompress: 6735 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6736 state = get_state_from_sha_object(sha_obj); 6737 stubAddr = StubRoutines::sha256_implCompress(); 6738 stubName = "sha256_implCompress"; 6739 break; 6740 case vmIntrinsics::_sha5_implCompress: 6741 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6742 state = get_state_from_sha5_object(sha_obj); 6743 stubAddr = StubRoutines::sha512_implCompress(); 6744 stubName = "sha512_implCompress"; 6745 break; 6746 default: 6747 fatal_unexpected_iid(id); 6748 return false; 6749 } 6750 if (state == NULL) return false; 6751 6752 assert(stubAddr != NULL, "Stub is generated"); 6753 if (stubAddr == NULL) return false; 6754 6755 // Call the stub. 6756 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6757 stubAddr, stubName, TypePtr::BOTTOM, 6758 src_start, state); 6759 6760 return true; 6761 } 6762 6763 //------------------------------inline_digestBase_implCompressMB----------------------- 6764 // 6765 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6766 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6767 // 6768 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6769 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6770 "need SHA1/SHA256/SHA512 instruction support"); 6771 assert((uint)predicate < 3, "sanity"); 6772 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6773 6774 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6775 Node* src = argument(1); // byte[] array 6776 Node* ofs = argument(2); // type int 6777 Node* limit = argument(3); // type int 6778 6779 const Type* src_type = src->Value(&_gvn); 6780 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6781 if (top_src == NULL || top_src->klass() == NULL) { 6782 // failed array check 6783 return false; 6784 } 6785 // Figure out the size and type of the elements we will be copying. 6786 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6787 if (src_elem != T_BYTE) { 6788 return false; 6789 } 6790 // 'src_start' points to src array + offset 6791 src = must_be_not_null(src, false); 6792 src = access_resolve(src, ACCESS_READ); 6793 Node* src_start = array_element_address(src, ofs, src_elem); 6794 6795 const char* klass_SHA_name = NULL; 6796 const char* stub_name = NULL; 6797 address stub_addr = NULL; 6798 bool long_state = false; 6799 6800 switch (predicate) { 6801 case 0: 6802 if (UseSHA1Intrinsics) { 6803 klass_SHA_name = "sun/security/provider/SHA"; 6804 stub_name = "sha1_implCompressMB"; 6805 stub_addr = StubRoutines::sha1_implCompressMB(); 6806 } 6807 break; 6808 case 1: 6809 if (UseSHA256Intrinsics) { 6810 klass_SHA_name = "sun/security/provider/SHA2"; 6811 stub_name = "sha256_implCompressMB"; 6812 stub_addr = StubRoutines::sha256_implCompressMB(); 6813 } 6814 break; 6815 case 2: 6816 if (UseSHA512Intrinsics) { 6817 klass_SHA_name = "sun/security/provider/SHA5"; 6818 stub_name = "sha512_implCompressMB"; 6819 stub_addr = StubRoutines::sha512_implCompressMB(); 6820 long_state = true; 6821 } 6822 break; 6823 default: 6824 fatal("unknown SHA intrinsic predicate: %d", predicate); 6825 } 6826 if (klass_SHA_name != NULL) { 6827 assert(stub_addr != NULL, "Stub is generated"); 6828 if (stub_addr == NULL) return false; 6829 6830 // get DigestBase klass to lookup for SHA klass 6831 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6832 assert(tinst != NULL, "digestBase_obj is not instance???"); 6833 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6834 6835 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6836 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6837 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6838 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6839 } 6840 return false; 6841 } 6842 //------------------------------inline_sha_implCompressMB----------------------- 6843 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6844 bool long_state, address stubAddr, const char *stubName, 6845 Node* src_start, Node* ofs, Node* limit) { 6846 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6847 const TypeOopPtr* xtype = aklass->as_instance_type(); 6848 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6849 sha_obj = _gvn.transform(sha_obj); 6850 6851 Node* state; 6852 if (long_state) { 6853 state = get_state_from_sha5_object(sha_obj); 6854 } else { 6855 state = get_state_from_sha_object(sha_obj); 6856 } 6857 if (state == NULL) return false; 6858 6859 // Call the stub. 6860 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6861 OptoRuntime::digestBase_implCompressMB_Type(), 6862 stubAddr, stubName, TypePtr::BOTTOM, 6863 src_start, state, ofs, limit); 6864 // return ofs (int) 6865 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6866 set_result(result); 6867 6868 return true; 6869 } 6870 6871 //------------------------------get_state_from_sha_object----------------------- 6872 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6873 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6874 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6875 if (sha_state == NULL) return (Node *) NULL; 6876 6877 // now have the array, need to get the start address of the state array 6878 sha_state = access_resolve(sha_state, ACCESS_WRITE); 6879 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6880 return state; 6881 } 6882 6883 //------------------------------get_state_from_sha5_object----------------------- 6884 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6885 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6886 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6887 if (sha_state == NULL) return (Node *) NULL; 6888 6889 // now have the array, need to get the start address of the state array 6890 sha_state = access_resolve(sha_state, ACCESS_WRITE); 6891 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6892 return state; 6893 } 6894 6895 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6896 // Return node representing slow path of predicate check. 6897 // the pseudo code we want to emulate with this predicate is: 6898 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6899 // 6900 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6901 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6902 "need SHA1/SHA256/SHA512 instruction support"); 6903 assert((uint)predicate < 3, "sanity"); 6904 6905 // The receiver was checked for NULL already. 6906 Node* digestBaseObj = argument(0); 6907 6908 // get DigestBase klass for instanceOf check 6909 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6910 assert(tinst != NULL, "digestBaseObj is null"); 6911 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6912 6913 const char* klass_SHA_name = NULL; 6914 switch (predicate) { 6915 case 0: 6916 if (UseSHA1Intrinsics) { 6917 // we want to do an instanceof comparison against the SHA class 6918 klass_SHA_name = "sun/security/provider/SHA"; 6919 } 6920 break; 6921 case 1: 6922 if (UseSHA256Intrinsics) { 6923 // we want to do an instanceof comparison against the SHA2 class 6924 klass_SHA_name = "sun/security/provider/SHA2"; 6925 } 6926 break; 6927 case 2: 6928 if (UseSHA512Intrinsics) { 6929 // we want to do an instanceof comparison against the SHA5 class 6930 klass_SHA_name = "sun/security/provider/SHA5"; 6931 } 6932 break; 6933 default: 6934 fatal("unknown SHA intrinsic predicate: %d", predicate); 6935 } 6936 6937 ciKlass* klass_SHA = NULL; 6938 if (klass_SHA_name != NULL) { 6939 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6940 } 6941 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6942 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6943 Node* ctrl = control(); 6944 set_control(top()); // no intrinsic path 6945 return ctrl; 6946 } 6947 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6948 6949 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6950 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6951 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6952 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6953 6954 return instof_false; // even if it is NULL 6955 } 6956 6957 //-------------inline_fma----------------------------------- 6958 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) { 6959 Node *a = NULL; 6960 Node *b = NULL; 6961 Node *c = NULL; 6962 Node* result = NULL; 6963 switch (id) { 6964 case vmIntrinsics::_fmaD: 6965 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each."); 6966 // no receiver since it is static method 6967 a = round_double_node(argument(0)); 6968 b = round_double_node(argument(2)); 6969 c = round_double_node(argument(4)); 6970 result = _gvn.transform(new FmaDNode(control(), a, b, c)); 6971 break; 6972 case vmIntrinsics::_fmaF: 6973 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each."); 6974 a = argument(0); 6975 b = argument(1); 6976 c = argument(2); 6977 result = _gvn.transform(new FmaFNode(control(), a, b, c)); 6978 break; 6979 default: 6980 fatal_unexpected_iid(id); break; 6981 } 6982 set_result(result); 6983 return true; 6984 } 6985 6986 bool LibraryCallKit::inline_character_compare(vmIntrinsics::ID id) { 6987 // argument(0) is receiver 6988 Node* codePoint = argument(1); 6989 Node* n = NULL; 6990 6991 switch (id) { 6992 case vmIntrinsics::_isDigit : 6993 n = new DigitNode(control(), codePoint); 6994 break; 6995 case vmIntrinsics::_isLowerCase : 6996 n = new LowerCaseNode(control(), codePoint); 6997 break; 6998 case vmIntrinsics::_isUpperCase : 6999 n = new UpperCaseNode(control(), codePoint); 7000 break; 7001 case vmIntrinsics::_isWhitespace : 7002 n = new WhitespaceNode(control(), codePoint); 7003 break; 7004 default: 7005 fatal_unexpected_iid(id); 7006 } 7007 7008 set_result(_gvn.transform(n)); 7009 return true; 7010 } 7011 7012 //------------------------------inline_fp_min_max------------------------------ 7013 bool LibraryCallKit::inline_fp_min_max(vmIntrinsics::ID id) { 7014 /* DISABLED BECAUSE METHOD DATA ISN'T COLLECTED PER CALL-SITE, SEE JDK-8015416. 7015 7016 // The intrinsic should be used only when the API branches aren't predictable, 7017 // the last one performing the most important comparison. The following heuristic 7018 // uses the branch statistics to eventually bail out if necessary. 7019 7020 ciMethodData *md = callee()->method_data(); 7021 7022 if ( md != NULL && md->is_mature() && md->invocation_count() > 0 ) { 7023 ciCallProfile cp = caller()->call_profile_at_bci(bci()); 7024 7025 if ( ((double)cp.count()) / ((double)md->invocation_count()) < 0.8 ) { 7026 // Bail out if the call-site didn't contribute enough to the statistics. 7027 return false; 7028 } 7029 7030 uint taken = 0, not_taken = 0; 7031 7032 for (ciProfileData *p = md->first_data(); md->is_valid(p); p = md->next_data(p)) { 7033 if (p->is_BranchData()) { 7034 taken = ((ciBranchData*)p)->taken(); 7035 not_taken = ((ciBranchData*)p)->not_taken(); 7036 } 7037 } 7038 7039 double balance = (((double)taken) - ((double)not_taken)) / ((double)md->invocation_count()); 7040 balance = balance < 0 ? -balance : balance; 7041 if ( balance > 0.2 ) { 7042 // Bail out if the most important branch is predictable enough. 7043 return false; 7044 } 7045 } 7046 */ 7047 7048 Node *a = NULL; 7049 Node *b = NULL; 7050 Node *n = NULL; 7051 switch (id) { 7052 case vmIntrinsics::_maxF: 7053 case vmIntrinsics::_minF: 7054 assert(callee()->signature()->size() == 2, "minF/maxF has 2 parameters of size 1 each."); 7055 a = argument(0); 7056 b = argument(1); 7057 break; 7058 case vmIntrinsics::_maxD: 7059 case vmIntrinsics::_minD: 7060 assert(callee()->signature()->size() == 4, "minD/maxD has 2 parameters of size 2 each."); 7061 a = round_double_node(argument(0)); 7062 b = round_double_node(argument(2)); 7063 break; 7064 default: 7065 fatal_unexpected_iid(id); 7066 break; 7067 } 7068 if (a->is_Con() || b->is_Con()) { 7069 return false; 7070 } 7071 switch (id) { 7072 case vmIntrinsics::_maxF: n = new MaxFNode(a, b); break; 7073 case vmIntrinsics::_minF: n = new MinFNode(a, b); break; 7074 case vmIntrinsics::_maxD: n = new MaxDNode(a, b); break; 7075 case vmIntrinsics::_minD: n = new MinDNode(a, b); break; 7076 default: fatal_unexpected_iid(id); break; 7077 } 7078 set_result(_gvn.transform(n)); 7079 return true; 7080 } 7081 7082 bool LibraryCallKit::inline_profileBoolean() { 7083 Node* counts = argument(1); 7084 const TypeAryPtr* ary = NULL; 7085 ciArray* aobj = NULL; 7086 if (counts->is_Con() 7087 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 7088 && (aobj = ary->const_oop()->as_array()) != NULL 7089 && (aobj->length() == 2)) { 7090 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 7091 jint false_cnt = aobj->element_value(0).as_int(); 7092 jint true_cnt = aobj->element_value(1).as_int(); 7093 7094 if (C->log() != NULL) { 7095 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 7096 false_cnt, true_cnt); 7097 } 7098 7099 if (false_cnt + true_cnt == 0) { 7100 // According to profile, never executed. 7101 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 7102 Deoptimization::Action_reinterpret); 7103 return true; 7104 } 7105 7106 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 7107 // is a number of each value occurrences. 7108 Node* result = argument(0); 7109 if (false_cnt == 0 || true_cnt == 0) { 7110 // According to profile, one value has been never seen. 7111 int expected_val = (false_cnt == 0) ? 1 : 0; 7112 7113 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 7114 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 7115 7116 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 7117 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 7118 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 7119 7120 { // Slow path: uncommon trap for never seen value and then reexecute 7121 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 7122 // the value has been seen at least once. 7123 PreserveJVMState pjvms(this); 7124 PreserveReexecuteState preexecs(this); 7125 jvms()->set_should_reexecute(true); 7126 7127 set_control(slow_path); 7128 set_i_o(i_o()); 7129 7130 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 7131 Deoptimization::Action_reinterpret); 7132 } 7133 // The guard for never seen value enables sharpening of the result and 7134 // returning a constant. It allows to eliminate branches on the same value 7135 // later on. 7136 set_control(fast_path); 7137 result = intcon(expected_val); 7138 } 7139 // Stop profiling. 7140 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 7141 // By replacing method body with profile data (represented as ProfileBooleanNode 7142 // on IR level) we effectively disable profiling. 7143 // It enables full speed execution once optimized code is generated. 7144 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 7145 C->record_for_igvn(profile); 7146 set_result(profile); 7147 return true; 7148 } else { 7149 // Continue profiling. 7150 // Profile data isn't available at the moment. So, execute method's bytecode version. 7151 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 7152 // is compiled and counters aren't available since corresponding MethodHandle 7153 // isn't a compile-time constant. 7154 return false; 7155 } 7156 } 7157 7158 bool LibraryCallKit::inline_isCompileConstant() { 7159 Node* n = argument(0); 7160 set_result(n->is_Con() ? intcon(1) : intcon(0)); 7161 return true; 7162 }