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 assert(ValueArrayFlatten, "can never be flattened"); 216 return generate_array_guard_common(kls, region, ValueArray); 217 } 218 Node* generate_array_guard_common(Node* kls, RegionNode* region, ArrayKind kind); 219 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 220 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 221 bool is_virtual = false, bool is_static = false); 222 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 223 return generate_method_call(method_id, false, true); 224 } 225 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 226 return generate_method_call(method_id, true, false); 227 } 228 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 229 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 230 231 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 232 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 233 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 234 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 235 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 236 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); 237 bool inline_string_indexOfChar(); 238 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 239 bool inline_string_toBytesU(); 240 bool inline_string_getCharsU(); 241 bool inline_string_copy(bool compress); 242 bool inline_string_char_access(bool is_store); 243 Node* round_double_node(Node* n); 244 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 245 bool inline_math_native(vmIntrinsics::ID id); 246 bool inline_math(vmIntrinsics::ID id); 247 bool inline_double_math(vmIntrinsics::ID id); 248 template <typename OverflowOp> 249 bool inline_math_overflow(Node* arg1, Node* arg2); 250 void inline_math_mathExact(Node* math, Node* test); 251 bool inline_math_addExactI(bool is_increment); 252 bool inline_math_addExactL(bool is_increment); 253 bool inline_math_multiplyExactI(); 254 bool inline_math_multiplyExactL(); 255 bool inline_math_multiplyHigh(); 256 bool inline_math_negateExactI(); 257 bool inline_math_negateExactL(); 258 bool inline_math_subtractExactI(bool is_decrement); 259 bool inline_math_subtractExactL(bool is_decrement); 260 bool inline_min_max(vmIntrinsics::ID id); 261 bool inline_notify(vmIntrinsics::ID id); 262 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 263 // This returns Type::AnyPtr, RawPtr, or OopPtr. 264 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type); 265 Node* make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type = T_ILLEGAL, bool can_cast = false); 266 267 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind; 268 DecoratorSet mo_decorator_for_access_kind(AccessKind kind); 269 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned); 270 static bool klass_needs_init_guard(Node* kls); 271 bool inline_unsafe_allocate(); 272 bool inline_unsafe_newArray(bool uninitialized); 273 bool inline_unsafe_copyMemory(); 274 bool inline_unsafe_make_private_buffer(); 275 bool inline_unsafe_finish_private_buffer(); 276 bool inline_native_currentThread(); 277 278 bool inline_native_time_funcs(address method, const char* funcName); 279 #ifdef JFR_HAVE_INTRINSICS 280 bool inline_native_classID(); 281 bool inline_native_getEventWriter(); 282 #endif 283 bool inline_native_isInterrupted(); 284 bool inline_native_Class_query(vmIntrinsics::ID id); 285 bool inline_value_Class_conversion(vmIntrinsics::ID id); 286 bool inline_native_subtype_check(); 287 bool inline_native_getLength(); 288 bool inline_array_copyOf(bool is_copyOfRange); 289 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 290 bool inline_preconditions_checkIndex(); 291 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array); 292 bool inline_native_clone(bool is_virtual); 293 bool inline_native_Reflection_getCallerClass(); 294 // Helper function for inlining native object hash method 295 bool inline_native_hashcode(bool is_virtual, bool is_static); 296 bool inline_native_getClass(); 297 298 // Helper functions for inlining arraycopy 299 bool inline_arraycopy(); 300 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 301 RegionNode* slow_region); 302 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 303 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp, 304 uint new_idx); 305 306 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind; 307 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind); 308 bool inline_unsafe_fence(vmIntrinsics::ID id); 309 bool inline_onspinwait(); 310 bool inline_fp_conversions(vmIntrinsics::ID id); 311 bool inline_number_methods(vmIntrinsics::ID id); 312 bool inline_reference_get(); 313 bool inline_Class_cast(); 314 bool inline_aescrypt_Block(vmIntrinsics::ID id); 315 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 316 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 317 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 318 Node* inline_counterMode_AESCrypt_predicate(); 319 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 320 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 321 bool inline_ghash_processBlocks(); 322 bool inline_base64_encodeBlock(); 323 bool inline_sha_implCompress(vmIntrinsics::ID id); 324 bool inline_digestBase_implCompressMB(int predicate); 325 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 326 bool long_state, address stubAddr, const char *stubName, 327 Node* src_start, Node* ofs, Node* limit); 328 Node* get_state_from_sha_object(Node *sha_object); 329 Node* get_state_from_sha5_object(Node *sha_object); 330 Node* inline_digestBase_implCompressMB_predicate(int predicate); 331 bool inline_encodeISOArray(); 332 bool inline_updateCRC32(); 333 bool inline_updateBytesCRC32(); 334 bool inline_updateByteBufferCRC32(); 335 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 336 bool inline_updateBytesCRC32C(); 337 bool inline_updateDirectByteBufferCRC32C(); 338 bool inline_updateBytesAdler32(); 339 bool inline_updateByteBufferAdler32(); 340 bool inline_multiplyToLen(); 341 bool inline_hasNegatives(); 342 bool inline_squareToLen(); 343 bool inline_mulAdd(); 344 bool inline_montgomeryMultiply(); 345 bool inline_montgomerySquare(); 346 bool inline_vectorizedMismatch(); 347 bool inline_fma(vmIntrinsics::ID id); 348 bool inline_character_compare(vmIntrinsics::ID id); 349 bool inline_fp_min_max(vmIntrinsics::ID id); 350 351 bool inline_profileBoolean(); 352 bool inline_isCompileConstant(); 353 void clear_upper_avx() { 354 #ifdef X86 355 if (UseAVX >= 2) { 356 C->set_clear_upper_avx(true); 357 } 358 #endif 359 } 360 }; 361 362 //---------------------------make_vm_intrinsic---------------------------- 363 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 364 vmIntrinsics::ID id = m->intrinsic_id(); 365 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 366 367 if (!m->is_loaded()) { 368 // Do not attempt to inline unloaded methods. 369 return NULL; 370 } 371 372 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 373 bool is_available = false; 374 375 { 376 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 377 // the compiler must transition to '_thread_in_vm' state because both 378 // methods access VM-internal data. 379 VM_ENTRY_MARK; 380 methodHandle mh(THREAD, m->get_Method()); 381 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) && 382 !C->directive()->is_intrinsic_disabled(mh) && 383 !vmIntrinsics::is_disabled_by_flags(mh); 384 385 } 386 387 if (is_available) { 388 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 389 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 390 return new LibraryIntrinsic(m, is_virtual, 391 vmIntrinsics::predicates_needed(id), 392 vmIntrinsics::does_virtual_dispatch(id), 393 (vmIntrinsics::ID) id); 394 } else { 395 return NULL; 396 } 397 } 398 399 //----------------------register_library_intrinsics----------------------- 400 // Initialize this file's data structures, for each Compile instance. 401 void Compile::register_library_intrinsics() { 402 // Nothing to do here. 403 } 404 405 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 406 LibraryCallKit kit(jvms, this); 407 Compile* C = kit.C; 408 int nodes = C->unique(); 409 #ifndef PRODUCT 410 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 411 char buf[1000]; 412 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 413 tty->print_cr("Intrinsic %s", str); 414 } 415 #endif 416 ciMethod* callee = kit.callee(); 417 const int bci = kit.bci(); 418 419 // Try to inline the intrinsic. 420 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 421 kit.try_to_inline(_last_predicate)) { 422 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)" 423 : "(intrinsic)"; 424 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 425 if (C->print_intrinsics() || C->print_inlining()) { 426 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 427 } 428 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 429 if (C->log()) { 430 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 431 vmIntrinsics::name_at(intrinsic_id()), 432 (is_virtual() ? " virtual='1'" : ""), 433 C->unique() - nodes); 434 } 435 // Push the result from the inlined method onto the stack. 436 kit.push_result(); 437 C->print_inlining_update(this); 438 return kit.transfer_exceptions_into_jvms(); 439 } 440 441 // The intrinsic bailed out 442 if (jvms->has_method()) { 443 // Not a root compile. 444 const char* msg; 445 if (callee->intrinsic_candidate()) { 446 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 447 } else { 448 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 449 : "failed to inline (intrinsic), method not annotated"; 450 } 451 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg); 452 if (C->print_intrinsics() || C->print_inlining()) { 453 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 454 } 455 } else { 456 // Root compile 457 ResourceMark rm; 458 stringStream msg_stream; 459 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 460 vmIntrinsics::name_at(intrinsic_id()), 461 is_virtual() ? " (virtual)" : "", bci); 462 const char *msg = msg_stream.as_string(); 463 log_debug(jit, inlining)("%s", msg); 464 if (C->print_intrinsics() || C->print_inlining()) { 465 tty->print("%s", msg); 466 } 467 } 468 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 469 C->print_inlining_update(this); 470 return NULL; 471 } 472 473 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 474 LibraryCallKit kit(jvms, this); 475 Compile* C = kit.C; 476 int nodes = C->unique(); 477 _last_predicate = predicate; 478 #ifndef PRODUCT 479 assert(is_predicated() && predicate < predicates_count(), "sanity"); 480 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 481 char buf[1000]; 482 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 483 tty->print_cr("Predicate for intrinsic %s", str); 484 } 485 #endif 486 ciMethod* callee = kit.callee(); 487 const int bci = kit.bci(); 488 489 Node* slow_ctl = kit.try_to_predicate(predicate); 490 if (!kit.failing()) { 491 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)" 492 : "(intrinsic, predicate)"; 493 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 494 if (C->print_intrinsics() || C->print_inlining()) { 495 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 496 } 497 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 498 if (C->log()) { 499 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 500 vmIntrinsics::name_at(intrinsic_id()), 501 (is_virtual() ? " virtual='1'" : ""), 502 C->unique() - nodes); 503 } 504 return slow_ctl; // Could be NULL if the check folds. 505 } 506 507 // The intrinsic bailed out 508 if (jvms->has_method()) { 509 // Not a root compile. 510 const char* msg = "failed to generate predicate for intrinsic"; 511 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg); 512 if (C->print_intrinsics() || C->print_inlining()) { 513 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 514 } 515 } else { 516 // Root compile 517 ResourceMark rm; 518 stringStream msg_stream; 519 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 520 vmIntrinsics::name_at(intrinsic_id()), 521 is_virtual() ? " (virtual)" : "", bci); 522 const char *msg = msg_stream.as_string(); 523 log_debug(jit, inlining)("%s", msg); 524 if (C->print_intrinsics() || C->print_inlining()) { 525 C->print_inlining_stream()->print("%s", msg); 526 } 527 } 528 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 529 return NULL; 530 } 531 532 bool LibraryCallKit::try_to_inline(int predicate) { 533 // Handle symbolic names for otherwise undistinguished boolean switches: 534 const bool is_store = true; 535 const bool is_compress = true; 536 const bool is_static = true; 537 const bool is_volatile = true; 538 539 if (!jvms()->has_method()) { 540 // Root JVMState has a null method. 541 assert(map()->memory()->Opcode() == Op_Parm, ""); 542 // Insert the memory aliasing node 543 set_all_memory(reset_memory()); 544 } 545 assert(merged_memory(), ""); 546 547 548 switch (intrinsic_id()) { 549 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 550 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 551 case vmIntrinsics::_getClass: return inline_native_getClass(); 552 553 case vmIntrinsics::_dsin: 554 case vmIntrinsics::_dcos: 555 case vmIntrinsics::_dtan: 556 case vmIntrinsics::_dabs: 557 case vmIntrinsics::_fabs: 558 case vmIntrinsics::_iabs: 559 case vmIntrinsics::_labs: 560 case vmIntrinsics::_datan2: 561 case vmIntrinsics::_dsqrt: 562 case vmIntrinsics::_dexp: 563 case vmIntrinsics::_dlog: 564 case vmIntrinsics::_dlog10: 565 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 566 567 case vmIntrinsics::_min: 568 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 569 570 case vmIntrinsics::_notify: 571 case vmIntrinsics::_notifyAll: 572 return inline_notify(intrinsic_id()); 573 574 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 575 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 576 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 577 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 578 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 579 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 580 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 581 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 582 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh(); 583 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 584 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 585 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 586 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 587 588 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 589 590 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 591 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 592 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 593 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 594 595 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 596 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 597 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 598 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 599 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 600 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 601 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 602 603 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 604 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 605 606 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 607 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 608 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 609 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 610 611 case vmIntrinsics::_compressStringC: 612 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 613 case vmIntrinsics::_inflateStringC: 614 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 615 616 case vmIntrinsics::_makePrivateBuffer: return inline_unsafe_make_private_buffer(); 617 case vmIntrinsics::_finishPrivateBuffer: return inline_unsafe_finish_private_buffer(); 618 case vmIntrinsics::_getReference: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 619 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 620 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 621 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 622 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 623 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 624 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 625 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 626 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 627 case vmIntrinsics::_getValue: return inline_unsafe_access(!is_store, T_VALUETYPE,Relaxed, false); 628 629 case vmIntrinsics::_putReference: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 630 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 631 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 632 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 633 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 634 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 635 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 636 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 637 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 638 case vmIntrinsics::_putValue: return inline_unsafe_access( is_store, T_VALUETYPE,Relaxed, false); 639 640 case vmIntrinsics::_getReferenceVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 641 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 642 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 643 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 644 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 645 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 646 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 647 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 648 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 649 650 case vmIntrinsics::_putReferenceVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 651 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 652 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 653 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 654 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 655 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 656 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 657 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 658 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 659 660 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 661 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 662 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 663 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 664 665 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 666 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 667 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 668 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 669 670 case vmIntrinsics::_getReferenceAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 671 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 672 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 673 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 674 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 675 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 676 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 677 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 678 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 679 680 case vmIntrinsics::_putReferenceRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 681 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 682 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 683 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 684 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 685 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 686 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 687 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 688 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 689 690 case vmIntrinsics::_getReferenceOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 691 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 692 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 693 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 694 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 695 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 696 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 697 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 698 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 699 700 case vmIntrinsics::_putReferenceOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 701 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 702 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 703 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 704 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 705 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 706 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 707 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 708 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 709 710 case vmIntrinsics::_compareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 711 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); 712 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); 713 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 714 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 715 716 case vmIntrinsics::_weakCompareAndSetReferencePlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 717 case vmIntrinsics::_weakCompareAndSetReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 718 case vmIntrinsics::_weakCompareAndSetReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 719 case vmIntrinsics::_weakCompareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 720 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); 721 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); 722 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); 723 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); 724 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); 725 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); 726 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); 727 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); 728 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 729 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 730 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 731 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 732 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 733 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 734 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 735 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 736 737 case vmIntrinsics::_compareAndExchangeReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 738 case vmIntrinsics::_compareAndExchangeReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 739 case vmIntrinsics::_compareAndExchangeReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 740 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); 741 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); 742 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); 743 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); 744 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); 745 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); 746 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 747 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 748 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 749 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 750 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 751 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 752 753 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); 754 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); 755 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 756 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 757 758 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); 759 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); 760 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 761 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 762 case vmIntrinsics::_getAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 763 764 case vmIntrinsics::_loadFence: 765 case vmIntrinsics::_storeFence: 766 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 767 768 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 769 770 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 771 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 772 773 #ifdef JFR_HAVE_INTRINSICS 774 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime"); 775 case vmIntrinsics::_getClassId: return inline_native_classID(); 776 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter(); 777 #endif 778 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 779 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 780 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 781 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 782 case vmIntrinsics::_getLength: return inline_native_getLength(); 783 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 784 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 785 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 786 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 787 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 788 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 789 790 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 791 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 792 793 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 794 795 case vmIntrinsics::_isInstance: 796 case vmIntrinsics::_getModifiers: 797 case vmIntrinsics::_isInterface: 798 case vmIntrinsics::_isArray: 799 case vmIntrinsics::_isPrimitive: 800 case vmIntrinsics::_getSuperclass: 801 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 802 803 case vmIntrinsics::_asPrimaryType: 804 case vmIntrinsics::_asIndirectType: return inline_value_Class_conversion(intrinsic_id()); 805 806 case vmIntrinsics::_floatToRawIntBits: 807 case vmIntrinsics::_floatToIntBits: 808 case vmIntrinsics::_intBitsToFloat: 809 case vmIntrinsics::_doubleToRawLongBits: 810 case vmIntrinsics::_doubleToLongBits: 811 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 812 813 case vmIntrinsics::_numberOfLeadingZeros_i: 814 case vmIntrinsics::_numberOfLeadingZeros_l: 815 case vmIntrinsics::_numberOfTrailingZeros_i: 816 case vmIntrinsics::_numberOfTrailingZeros_l: 817 case vmIntrinsics::_bitCount_i: 818 case vmIntrinsics::_bitCount_l: 819 case vmIntrinsics::_reverseBytes_i: 820 case vmIntrinsics::_reverseBytes_l: 821 case vmIntrinsics::_reverseBytes_s: 822 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 823 824 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 825 826 case vmIntrinsics::_Reference_get: return inline_reference_get(); 827 828 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 829 830 case vmIntrinsics::_aescrypt_encryptBlock: 831 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 832 833 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 834 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 835 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 836 837 case vmIntrinsics::_counterMode_AESCrypt: 838 return inline_counterMode_AESCrypt(intrinsic_id()); 839 840 case vmIntrinsics::_sha_implCompress: 841 case vmIntrinsics::_sha2_implCompress: 842 case vmIntrinsics::_sha5_implCompress: 843 return inline_sha_implCompress(intrinsic_id()); 844 845 case vmIntrinsics::_digestBase_implCompressMB: 846 return inline_digestBase_implCompressMB(predicate); 847 848 case vmIntrinsics::_multiplyToLen: 849 return inline_multiplyToLen(); 850 851 case vmIntrinsics::_squareToLen: 852 return inline_squareToLen(); 853 854 case vmIntrinsics::_mulAdd: 855 return inline_mulAdd(); 856 857 case vmIntrinsics::_montgomeryMultiply: 858 return inline_montgomeryMultiply(); 859 case vmIntrinsics::_montgomerySquare: 860 return inline_montgomerySquare(); 861 862 case vmIntrinsics::_vectorizedMismatch: 863 return inline_vectorizedMismatch(); 864 865 case vmIntrinsics::_ghash_processBlocks: 866 return inline_ghash_processBlocks(); 867 case vmIntrinsics::_base64_encodeBlock: 868 return inline_base64_encodeBlock(); 869 870 case vmIntrinsics::_encodeISOArray: 871 case vmIntrinsics::_encodeByteISOArray: 872 return inline_encodeISOArray(); 873 874 case vmIntrinsics::_updateCRC32: 875 return inline_updateCRC32(); 876 case vmIntrinsics::_updateBytesCRC32: 877 return inline_updateBytesCRC32(); 878 case vmIntrinsics::_updateByteBufferCRC32: 879 return inline_updateByteBufferCRC32(); 880 881 case vmIntrinsics::_updateBytesCRC32C: 882 return inline_updateBytesCRC32C(); 883 case vmIntrinsics::_updateDirectByteBufferCRC32C: 884 return inline_updateDirectByteBufferCRC32C(); 885 886 case vmIntrinsics::_updateBytesAdler32: 887 return inline_updateBytesAdler32(); 888 case vmIntrinsics::_updateByteBufferAdler32: 889 return inline_updateByteBufferAdler32(); 890 891 case vmIntrinsics::_profileBoolean: 892 return inline_profileBoolean(); 893 case vmIntrinsics::_isCompileConstant: 894 return inline_isCompileConstant(); 895 896 case vmIntrinsics::_hasNegatives: 897 return inline_hasNegatives(); 898 899 case vmIntrinsics::_fmaD: 900 case vmIntrinsics::_fmaF: 901 return inline_fma(intrinsic_id()); 902 903 case vmIntrinsics::_isDigit: 904 case vmIntrinsics::_isLowerCase: 905 case vmIntrinsics::_isUpperCase: 906 case vmIntrinsics::_isWhitespace: 907 return inline_character_compare(intrinsic_id()); 908 909 case vmIntrinsics::_maxF: 910 case vmIntrinsics::_minF: 911 case vmIntrinsics::_maxD: 912 case vmIntrinsics::_minD: 913 return inline_fp_min_max(intrinsic_id()); 914 915 default: 916 // If you get here, it may be that someone has added a new intrinsic 917 // to the list in vmSymbols.hpp without implementing it here. 918 #ifndef PRODUCT 919 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 920 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 921 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 922 } 923 #endif 924 return false; 925 } 926 } 927 928 Node* LibraryCallKit::try_to_predicate(int predicate) { 929 if (!jvms()->has_method()) { 930 // Root JVMState has a null method. 931 assert(map()->memory()->Opcode() == Op_Parm, ""); 932 // Insert the memory aliasing node 933 set_all_memory(reset_memory()); 934 } 935 assert(merged_memory(), ""); 936 937 switch (intrinsic_id()) { 938 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 939 return inline_cipherBlockChaining_AESCrypt_predicate(false); 940 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 941 return inline_cipherBlockChaining_AESCrypt_predicate(true); 942 case vmIntrinsics::_counterMode_AESCrypt: 943 return inline_counterMode_AESCrypt_predicate(); 944 case vmIntrinsics::_digestBase_implCompressMB: 945 return inline_digestBase_implCompressMB_predicate(predicate); 946 947 default: 948 // If you get here, it may be that someone has added a new intrinsic 949 // to the list in vmSymbols.hpp without implementing it here. 950 #ifndef PRODUCT 951 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 952 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 953 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 954 } 955 #endif 956 Node* slow_ctl = control(); 957 set_control(top()); // No fast path instrinsic 958 return slow_ctl; 959 } 960 } 961 962 //------------------------------set_result------------------------------- 963 // Helper function for finishing intrinsics. 964 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 965 record_for_igvn(region); 966 set_control(_gvn.transform(region)); 967 set_result( _gvn.transform(value)); 968 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 969 } 970 971 //------------------------------generate_guard--------------------------- 972 // Helper function for generating guarded fast-slow graph structures. 973 // The given 'test', if true, guards a slow path. If the test fails 974 // then a fast path can be taken. (We generally hope it fails.) 975 // In all cases, GraphKit::control() is updated to the fast path. 976 // The returned value represents the control for the slow path. 977 // The return value is never 'top'; it is either a valid control 978 // or NULL if it is obvious that the slow path can never be taken. 979 // Also, if region and the slow control are not NULL, the slow edge 980 // is appended to the region. 981 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 982 if (stopped()) { 983 // Already short circuited. 984 return NULL; 985 } 986 987 // Build an if node and its projections. 988 // If test is true we take the slow path, which we assume is uncommon. 989 if (_gvn.type(test) == TypeInt::ZERO) { 990 // The slow branch is never taken. No need to build this guard. 991 return NULL; 992 } 993 994 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 995 996 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 997 if (if_slow == top()) { 998 // The slow branch is never taken. No need to build this guard. 999 return NULL; 1000 } 1001 1002 if (region != NULL) 1003 region->add_req(if_slow); 1004 1005 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1006 set_control(if_fast); 1007 1008 return if_slow; 1009 } 1010 1011 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1012 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1013 } 1014 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1015 return generate_guard(test, region, PROB_FAIR); 1016 } 1017 1018 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1019 Node* *pos_index) { 1020 if (stopped()) 1021 return NULL; // already stopped 1022 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1023 return NULL; // index is already adequately typed 1024 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1025 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1026 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1027 if (is_neg != NULL && pos_index != NULL) { 1028 // Emulate effect of Parse::adjust_map_after_if. 1029 Node* ccast = new CastIINode(index, TypeInt::POS); 1030 ccast->set_req(0, control()); 1031 (*pos_index) = _gvn.transform(ccast); 1032 } 1033 return is_neg; 1034 } 1035 1036 // Make sure that 'position' is a valid limit index, in [0..length]. 1037 // There are two equivalent plans for checking this: 1038 // A. (offset + copyLength) unsigned<= arrayLength 1039 // B. offset <= (arrayLength - copyLength) 1040 // We require that all of the values above, except for the sum and 1041 // difference, are already known to be non-negative. 1042 // Plan A is robust in the face of overflow, if offset and copyLength 1043 // are both hugely positive. 1044 // 1045 // Plan B is less direct and intuitive, but it does not overflow at 1046 // all, since the difference of two non-negatives is always 1047 // representable. Whenever Java methods must perform the equivalent 1048 // check they generally use Plan B instead of Plan A. 1049 // For the moment we use Plan A. 1050 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1051 Node* subseq_length, 1052 Node* array_length, 1053 RegionNode* region) { 1054 if (stopped()) 1055 return NULL; // already stopped 1056 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1057 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1058 return NULL; // common case of whole-array copy 1059 Node* last = subseq_length; 1060 if (!zero_offset) // last += offset 1061 last = _gvn.transform(new AddINode(last, offset)); 1062 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1063 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1064 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1065 return is_over; 1066 } 1067 1068 // Emit range checks for the given String.value byte array 1069 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 1070 if (stopped()) { 1071 return; // already stopped 1072 } 1073 RegionNode* bailout = new RegionNode(1); 1074 record_for_igvn(bailout); 1075 if (char_count) { 1076 // Convert char count to byte count 1077 count = _gvn.transform(new LShiftINode(count, intcon(1))); 1078 } 1079 1080 // Offset and count must not be negative 1081 generate_negative_guard(offset, bailout); 1082 generate_negative_guard(count, bailout); 1083 // Offset + count must not exceed length of array 1084 generate_limit_guard(offset, count, load_array_length(array), bailout); 1085 1086 if (bailout->req() > 1) { 1087 PreserveJVMState pjvms(this); 1088 set_control(_gvn.transform(bailout)); 1089 uncommon_trap(Deoptimization::Reason_intrinsic, 1090 Deoptimization::Action_maybe_recompile); 1091 } 1092 } 1093 1094 //--------------------------generate_current_thread-------------------- 1095 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1096 ciKlass* thread_klass = env()->Thread_klass(); 1097 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1098 Node* thread = _gvn.transform(new ThreadLocalNode()); 1099 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1100 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1101 tls_output = thread; 1102 return threadObj; 1103 } 1104 1105 1106 //------------------------------make_string_method_node------------------------ 1107 // Helper method for String intrinsic functions. This version is called with 1108 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 1109 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 1110 // containing the lengths of str1 and str2. 1111 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 1112 Node* result = NULL; 1113 switch (opcode) { 1114 case Op_StrIndexOf: 1115 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 1116 str1_start, cnt1, str2_start, cnt2, ae); 1117 break; 1118 case Op_StrComp: 1119 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 1120 str1_start, cnt1, str2_start, cnt2, ae); 1121 break; 1122 case Op_StrEquals: 1123 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 1124 // Use the constant length if there is one because optimized match rule may exist. 1125 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 1126 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 1127 break; 1128 default: 1129 ShouldNotReachHere(); 1130 return NULL; 1131 } 1132 1133 // All these intrinsics have checks. 1134 C->set_has_split_ifs(true); // Has chance for split-if optimization 1135 clear_upper_avx(); 1136 1137 return _gvn.transform(result); 1138 } 1139 1140 //------------------------------inline_string_compareTo------------------------ 1141 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 1142 Node* arg1 = argument(0); 1143 Node* arg2 = argument(1); 1144 1145 arg1 = must_be_not_null(arg1, true); 1146 arg2 = must_be_not_null(arg2, true); 1147 1148 arg1 = access_resolve(arg1, ACCESS_READ); 1149 arg2 = access_resolve(arg2, ACCESS_READ); 1150 1151 // Get start addr and length of first argument 1152 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1153 Node* arg1_cnt = load_array_length(arg1); 1154 1155 // Get start addr and length of second argument 1156 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1157 Node* arg2_cnt = load_array_length(arg2); 1158 1159 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1160 set_result(result); 1161 return true; 1162 } 1163 1164 //------------------------------inline_string_equals------------------------ 1165 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1166 Node* arg1 = argument(0); 1167 Node* arg2 = argument(1); 1168 1169 // paths (plus control) merge 1170 RegionNode* region = new RegionNode(3); 1171 Node* phi = new PhiNode(region, TypeInt::BOOL); 1172 1173 if (!stopped()) { 1174 1175 arg1 = must_be_not_null(arg1, true); 1176 arg2 = must_be_not_null(arg2, true); 1177 1178 arg1 = access_resolve(arg1, ACCESS_READ); 1179 arg2 = access_resolve(arg2, ACCESS_READ); 1180 1181 // Get start addr and length of first argument 1182 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1183 Node* arg1_cnt = load_array_length(arg1); 1184 1185 // Get start addr and length of second argument 1186 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1187 Node* arg2_cnt = load_array_length(arg2); 1188 1189 // Check for arg1_cnt != arg2_cnt 1190 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1191 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1192 Node* if_ne = generate_slow_guard(bol, NULL); 1193 if (if_ne != NULL) { 1194 phi->init_req(2, intcon(0)); 1195 region->init_req(2, if_ne); 1196 } 1197 1198 // Check for count == 0 is done by assembler code for StrEquals. 1199 1200 if (!stopped()) { 1201 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1202 phi->init_req(1, equals); 1203 region->init_req(1, control()); 1204 } 1205 } 1206 1207 // post merge 1208 set_control(_gvn.transform(region)); 1209 record_for_igvn(region); 1210 1211 set_result(_gvn.transform(phi)); 1212 return true; 1213 } 1214 1215 //------------------------------inline_array_equals---------------------------- 1216 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1217 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1218 Node* arg1 = argument(0); 1219 Node* arg2 = argument(1); 1220 1221 arg1 = access_resolve(arg1, ACCESS_READ); 1222 arg2 = access_resolve(arg2, ACCESS_READ); 1223 1224 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1225 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1226 clear_upper_avx(); 1227 1228 return true; 1229 } 1230 1231 //------------------------------inline_hasNegatives------------------------------ 1232 bool LibraryCallKit::inline_hasNegatives() { 1233 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1234 return false; 1235 } 1236 1237 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1238 // no receiver since it is static method 1239 Node* ba = argument(0); 1240 Node* offset = argument(1); 1241 Node* len = argument(2); 1242 1243 ba = must_be_not_null(ba, true); 1244 1245 // Range checks 1246 generate_string_range_check(ba, offset, len, false); 1247 if (stopped()) { 1248 return true; 1249 } 1250 ba = access_resolve(ba, ACCESS_READ); 1251 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1252 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1253 set_result(_gvn.transform(result)); 1254 return true; 1255 } 1256 1257 bool LibraryCallKit::inline_preconditions_checkIndex() { 1258 Node* index = argument(0); 1259 Node* length = argument(1); 1260 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1261 return false; 1262 } 1263 1264 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1265 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1266 1267 { 1268 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1269 uncommon_trap(Deoptimization::Reason_intrinsic, 1270 Deoptimization::Action_make_not_entrant); 1271 } 1272 1273 if (stopped()) { 1274 return false; 1275 } 1276 1277 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1278 BoolTest::mask btest = BoolTest::lt; 1279 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1280 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1281 _gvn.set_type(rc, rc->Value(&_gvn)); 1282 if (!rc_bool->is_Con()) { 1283 record_for_igvn(rc); 1284 } 1285 set_control(_gvn.transform(new IfTrueNode(rc))); 1286 { 1287 PreserveJVMState pjvms(this); 1288 set_control(_gvn.transform(new IfFalseNode(rc))); 1289 uncommon_trap(Deoptimization::Reason_range_check, 1290 Deoptimization::Action_make_not_entrant); 1291 } 1292 1293 if (stopped()) { 1294 return false; 1295 } 1296 1297 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1298 result->set_req(0, control()); 1299 result = _gvn.transform(result); 1300 set_result(result); 1301 replace_in_map(index, result); 1302 clear_upper_avx(); 1303 return true; 1304 } 1305 1306 //------------------------------inline_string_indexOf------------------------ 1307 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1308 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1309 return false; 1310 } 1311 Node* src = argument(0); 1312 Node* tgt = argument(1); 1313 1314 // Make the merge point 1315 RegionNode* result_rgn = new RegionNode(4); 1316 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1317 1318 src = must_be_not_null(src, true); 1319 tgt = must_be_not_null(tgt, true); 1320 1321 src = access_resolve(src, ACCESS_READ); 1322 tgt = access_resolve(tgt, ACCESS_READ); 1323 1324 // Get start addr and length of source string 1325 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1326 Node* src_count = load_array_length(src); 1327 1328 // Get start addr and length of substring 1329 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1330 Node* tgt_count = load_array_length(tgt); 1331 1332 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1333 // Divide src size by 2 if String is UTF16 encoded 1334 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1335 } 1336 if (ae == StrIntrinsicNode::UU) { 1337 // Divide substring size by 2 if String is UTF16 encoded 1338 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1339 } 1340 1341 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1342 if (result != NULL) { 1343 result_phi->init_req(3, result); 1344 result_rgn->init_req(3, control()); 1345 } 1346 set_control(_gvn.transform(result_rgn)); 1347 record_for_igvn(result_rgn); 1348 set_result(_gvn.transform(result_phi)); 1349 1350 return true; 1351 } 1352 1353 //-----------------------------inline_string_indexOf----------------------- 1354 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1355 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1356 return false; 1357 } 1358 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1359 return false; 1360 } 1361 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1362 Node* src = argument(0); // byte[] 1363 Node* src_count = argument(1); // char count 1364 Node* tgt = argument(2); // byte[] 1365 Node* tgt_count = argument(3); // char count 1366 Node* from_index = argument(4); // char index 1367 1368 src = must_be_not_null(src, true); 1369 tgt = must_be_not_null(tgt, true); 1370 1371 src = access_resolve(src, ACCESS_READ); 1372 tgt = access_resolve(tgt, ACCESS_READ); 1373 1374 // Multiply byte array index by 2 if String is UTF16 encoded 1375 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1376 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1377 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1378 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1379 1380 // Range checks 1381 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1382 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1383 if (stopped()) { 1384 return true; 1385 } 1386 1387 RegionNode* region = new RegionNode(5); 1388 Node* phi = new PhiNode(region, TypeInt::INT); 1389 1390 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1391 if (result != NULL) { 1392 // The result is index relative to from_index if substring was found, -1 otherwise. 1393 // Generate code which will fold into cmove. 1394 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1395 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1396 1397 Node* if_lt = generate_slow_guard(bol, NULL); 1398 if (if_lt != NULL) { 1399 // result == -1 1400 phi->init_req(3, result); 1401 region->init_req(3, if_lt); 1402 } 1403 if (!stopped()) { 1404 result = _gvn.transform(new AddINode(result, from_index)); 1405 phi->init_req(4, result); 1406 region->init_req(4, control()); 1407 } 1408 } 1409 1410 set_control(_gvn.transform(region)); 1411 record_for_igvn(region); 1412 set_result(_gvn.transform(phi)); 1413 clear_upper_avx(); 1414 1415 return true; 1416 } 1417 1418 // Create StrIndexOfNode with fast path checks 1419 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1420 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1421 // Check for substr count > string count 1422 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1423 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1424 Node* if_gt = generate_slow_guard(bol, NULL); 1425 if (if_gt != NULL) { 1426 phi->init_req(1, intcon(-1)); 1427 region->init_req(1, if_gt); 1428 } 1429 if (!stopped()) { 1430 // Check for substr count == 0 1431 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1432 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1433 Node* if_zero = generate_slow_guard(bol, NULL); 1434 if (if_zero != NULL) { 1435 phi->init_req(2, intcon(0)); 1436 region->init_req(2, if_zero); 1437 } 1438 } 1439 if (!stopped()) { 1440 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1441 } 1442 return NULL; 1443 } 1444 1445 //-----------------------------inline_string_indexOfChar----------------------- 1446 bool LibraryCallKit::inline_string_indexOfChar() { 1447 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1448 return false; 1449 } 1450 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1451 return false; 1452 } 1453 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1454 Node* src = argument(0); // byte[] 1455 Node* tgt = argument(1); // tgt is int ch 1456 Node* from_index = argument(2); 1457 Node* max = argument(3); 1458 1459 src = must_be_not_null(src, true); 1460 src = access_resolve(src, ACCESS_READ); 1461 1462 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1463 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1464 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1465 1466 // Range checks 1467 generate_string_range_check(src, src_offset, src_count, true); 1468 if (stopped()) { 1469 return true; 1470 } 1471 1472 RegionNode* region = new RegionNode(3); 1473 Node* phi = new PhiNode(region, TypeInt::INT); 1474 1475 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1476 C->set_has_split_ifs(true); // Has chance for split-if optimization 1477 _gvn.transform(result); 1478 1479 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1480 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1481 1482 Node* if_lt = generate_slow_guard(bol, NULL); 1483 if (if_lt != NULL) { 1484 // result == -1 1485 phi->init_req(2, result); 1486 region->init_req(2, if_lt); 1487 } 1488 if (!stopped()) { 1489 result = _gvn.transform(new AddINode(result, from_index)); 1490 phi->init_req(1, result); 1491 region->init_req(1, control()); 1492 } 1493 set_control(_gvn.transform(region)); 1494 record_for_igvn(region); 1495 set_result(_gvn.transform(phi)); 1496 1497 return true; 1498 } 1499 //---------------------------inline_string_copy--------------------- 1500 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1501 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1502 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1503 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1504 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1505 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1506 bool LibraryCallKit::inline_string_copy(bool compress) { 1507 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1508 return false; 1509 } 1510 int nargs = 5; // 2 oops, 3 ints 1511 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1512 1513 Node* src = argument(0); 1514 Node* src_offset = argument(1); 1515 Node* dst = argument(2); 1516 Node* dst_offset = argument(3); 1517 Node* length = argument(4); 1518 1519 // Check for allocation before we add nodes that would confuse 1520 // tightly_coupled_allocation() 1521 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1522 1523 // Figure out the size and type of the elements we will be copying. 1524 const Type* src_type = src->Value(&_gvn); 1525 const Type* dst_type = dst->Value(&_gvn); 1526 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1527 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1528 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1529 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1530 "Unsupported array types for inline_string_copy"); 1531 1532 src = must_be_not_null(src, true); 1533 dst = must_be_not_null(dst, true); 1534 1535 // Convert char[] offsets to byte[] offsets 1536 bool convert_src = (compress && src_elem == T_BYTE); 1537 bool convert_dst = (!compress && dst_elem == T_BYTE); 1538 if (convert_src) { 1539 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1540 } else if (convert_dst) { 1541 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1542 } 1543 1544 // Range checks 1545 generate_string_range_check(src, src_offset, length, convert_src); 1546 generate_string_range_check(dst, dst_offset, length, convert_dst); 1547 if (stopped()) { 1548 return true; 1549 } 1550 1551 src = access_resolve(src, ACCESS_READ); 1552 dst = access_resolve(dst, ACCESS_WRITE); 1553 1554 Node* src_start = array_element_address(src, src_offset, src_elem); 1555 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1556 // 'src_start' points to src array + scaled offset 1557 // 'dst_start' points to dst array + scaled offset 1558 Node* count = NULL; 1559 if (compress) { 1560 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1561 } else { 1562 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1563 } 1564 1565 if (alloc != NULL) { 1566 if (alloc->maybe_set_complete(&_gvn)) { 1567 // "You break it, you buy it." 1568 InitializeNode* init = alloc->initialization(); 1569 assert(init->is_complete(), "we just did this"); 1570 init->set_complete_with_arraycopy(); 1571 assert(dst->is_CheckCastPP(), "sanity"); 1572 assert(dst->in(0)->in(0) == init, "dest pinned"); 1573 } 1574 // Do not let stores that initialize this object be reordered with 1575 // a subsequent store that would make this object accessible by 1576 // other threads. 1577 // Record what AllocateNode this StoreStore protects so that 1578 // escape analysis can go from the MemBarStoreStoreNode to the 1579 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1580 // based on the escape status of the AllocateNode. 1581 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1582 } 1583 if (compress) { 1584 set_result(_gvn.transform(count)); 1585 } 1586 clear_upper_avx(); 1587 1588 return true; 1589 } 1590 1591 #ifdef _LP64 1592 #define XTOP ,top() /*additional argument*/ 1593 #else //_LP64 1594 #define XTOP /*no additional argument*/ 1595 #endif //_LP64 1596 1597 //------------------------inline_string_toBytesU-------------------------- 1598 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1599 bool LibraryCallKit::inline_string_toBytesU() { 1600 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1601 return false; 1602 } 1603 // Get the arguments. 1604 Node* value = argument(0); 1605 Node* offset = argument(1); 1606 Node* length = argument(2); 1607 1608 Node* newcopy = NULL; 1609 1610 // Set the original stack and the reexecute bit for the interpreter to reexecute 1611 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1612 { PreserveReexecuteState preexecs(this); 1613 jvms()->set_should_reexecute(true); 1614 1615 // Check if a null path was taken unconditionally. 1616 value = null_check(value); 1617 1618 RegionNode* bailout = new RegionNode(1); 1619 record_for_igvn(bailout); 1620 1621 // Range checks 1622 generate_negative_guard(offset, bailout); 1623 generate_negative_guard(length, bailout); 1624 generate_limit_guard(offset, length, load_array_length(value), bailout); 1625 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1626 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1627 1628 if (bailout->req() > 1) { 1629 PreserveJVMState pjvms(this); 1630 set_control(_gvn.transform(bailout)); 1631 uncommon_trap(Deoptimization::Reason_intrinsic, 1632 Deoptimization::Action_maybe_recompile); 1633 } 1634 if (stopped()) { 1635 return true; 1636 } 1637 1638 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1639 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1640 newcopy = new_array(klass_node, size, 0); // no arguments to push 1641 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1642 1643 // Calculate starting addresses. 1644 value = access_resolve(value, ACCESS_READ); 1645 Node* src_start = array_element_address(value, offset, T_CHAR); 1646 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1647 1648 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1649 const TypeInt* toffset = gvn().type(offset)->is_int(); 1650 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1651 1652 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1653 const char* copyfunc_name = "arraycopy"; 1654 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1655 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1656 OptoRuntime::fast_arraycopy_Type(), 1657 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1658 src_start, dst_start, ConvI2X(length) XTOP); 1659 // Do not let reads from the cloned object float above the arraycopy. 1660 if (alloc != NULL) { 1661 if (alloc->maybe_set_complete(&_gvn)) { 1662 // "You break it, you buy it." 1663 InitializeNode* init = alloc->initialization(); 1664 assert(init->is_complete(), "we just did this"); 1665 init->set_complete_with_arraycopy(); 1666 assert(newcopy->is_CheckCastPP(), "sanity"); 1667 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1668 } 1669 // Do not let stores that initialize this object be reordered with 1670 // a subsequent store that would make this object accessible by 1671 // other threads. 1672 // Record what AllocateNode this StoreStore protects so that 1673 // escape analysis can go from the MemBarStoreStoreNode to the 1674 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1675 // based on the escape status of the AllocateNode. 1676 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1677 } else { 1678 insert_mem_bar(Op_MemBarCPUOrder); 1679 } 1680 } // original reexecute is set back here 1681 1682 C->set_has_split_ifs(true); // Has chance for split-if optimization 1683 if (!stopped()) { 1684 set_result(newcopy); 1685 } 1686 clear_upper_avx(); 1687 1688 return true; 1689 } 1690 1691 //------------------------inline_string_getCharsU-------------------------- 1692 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1693 bool LibraryCallKit::inline_string_getCharsU() { 1694 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1695 return false; 1696 } 1697 1698 // Get the arguments. 1699 Node* src = argument(0); 1700 Node* src_begin = argument(1); 1701 Node* src_end = argument(2); // exclusive offset (i < src_end) 1702 Node* dst = argument(3); 1703 Node* dst_begin = argument(4); 1704 1705 // Check for allocation before we add nodes that would confuse 1706 // tightly_coupled_allocation() 1707 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1708 1709 // Check if a null path was taken unconditionally. 1710 src = null_check(src); 1711 dst = null_check(dst); 1712 if (stopped()) { 1713 return true; 1714 } 1715 1716 // Get length and convert char[] offset to byte[] offset 1717 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1718 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1719 1720 // Range checks 1721 generate_string_range_check(src, src_begin, length, true); 1722 generate_string_range_check(dst, dst_begin, length, false); 1723 if (stopped()) { 1724 return true; 1725 } 1726 1727 if (!stopped()) { 1728 src = access_resolve(src, ACCESS_READ); 1729 dst = access_resolve(dst, ACCESS_WRITE); 1730 1731 // Calculate starting addresses. 1732 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1733 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1734 1735 // Check if array addresses are aligned to HeapWordSize 1736 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1737 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1738 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1739 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1740 1741 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1742 const char* copyfunc_name = "arraycopy"; 1743 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1744 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1745 OptoRuntime::fast_arraycopy_Type(), 1746 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1747 src_start, dst_start, ConvI2X(length) XTOP); 1748 // Do not let reads from the cloned object float above the arraycopy. 1749 if (alloc != NULL) { 1750 if (alloc->maybe_set_complete(&_gvn)) { 1751 // "You break it, you buy it." 1752 InitializeNode* init = alloc->initialization(); 1753 assert(init->is_complete(), "we just did this"); 1754 init->set_complete_with_arraycopy(); 1755 assert(dst->is_CheckCastPP(), "sanity"); 1756 assert(dst->in(0)->in(0) == init, "dest pinned"); 1757 } 1758 // Do not let stores that initialize this object be reordered with 1759 // a subsequent store that would make this object accessible by 1760 // other threads. 1761 // Record what AllocateNode this StoreStore protects so that 1762 // escape analysis can go from the MemBarStoreStoreNode to the 1763 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1764 // based on the escape status of the AllocateNode. 1765 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1766 } else { 1767 insert_mem_bar(Op_MemBarCPUOrder); 1768 } 1769 } 1770 1771 C->set_has_split_ifs(true); // Has chance for split-if optimization 1772 return true; 1773 } 1774 1775 //----------------------inline_string_char_access---------------------------- 1776 // Store/Load char to/from byte[] array. 1777 // static void StringUTF16.putChar(byte[] val, int index, int c) 1778 // static char StringUTF16.getChar(byte[] val, int index) 1779 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1780 Node* value = argument(0); 1781 Node* index = argument(1); 1782 Node* ch = is_store ? argument(2) : NULL; 1783 1784 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1785 // correctly requires matched array shapes. 1786 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1787 "sanity: byte[] and char[] bases agree"); 1788 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1789 "sanity: byte[] and char[] scales agree"); 1790 1791 // Bail when getChar over constants is requested: constant folding would 1792 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1793 // Java method would constant fold nicely instead. 1794 if (!is_store && value->is_Con() && index->is_Con()) { 1795 return false; 1796 } 1797 1798 value = must_be_not_null(value, true); 1799 value = access_resolve(value, is_store ? ACCESS_WRITE : ACCESS_READ); 1800 1801 Node* adr = array_element_address(value, index, T_CHAR); 1802 if (adr->is_top()) { 1803 return false; 1804 } 1805 if (is_store) { 1806 access_store_at(value, adr, TypeAryPtr::BYTES, ch, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED); 1807 } else { 1808 ch = access_load_at(value, adr, TypeAryPtr::BYTES, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED | C2_CONTROL_DEPENDENT_LOAD); 1809 set_result(ch); 1810 } 1811 return true; 1812 } 1813 1814 //--------------------------round_double_node-------------------------------- 1815 // Round a double node if necessary. 1816 Node* LibraryCallKit::round_double_node(Node* n) { 1817 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1818 n = _gvn.transform(new RoundDoubleNode(0, n)); 1819 return n; 1820 } 1821 1822 //------------------------------inline_math----------------------------------- 1823 // public static double Math.abs(double) 1824 // public static double Math.sqrt(double) 1825 // public static double Math.log(double) 1826 // public static double Math.log10(double) 1827 bool LibraryCallKit::inline_double_math(vmIntrinsics::ID id) { 1828 Node* arg = round_double_node(argument(0)); 1829 Node* n = NULL; 1830 switch (id) { 1831 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1832 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1833 default: fatal_unexpected_iid(id); break; 1834 } 1835 set_result(_gvn.transform(n)); 1836 return true; 1837 } 1838 1839 //------------------------------inline_math----------------------------------- 1840 // public static float Math.abs(float) 1841 // public static int Math.abs(int) 1842 // public static long Math.abs(long) 1843 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1844 Node* arg = argument(0); 1845 Node* n = NULL; 1846 switch (id) { 1847 case vmIntrinsics::_fabs: n = new AbsFNode( arg); break; 1848 case vmIntrinsics::_iabs: n = new AbsINode( arg); break; 1849 case vmIntrinsics::_labs: n = new AbsLNode( arg); break; 1850 default: fatal_unexpected_iid(id); break; 1851 } 1852 set_result(_gvn.transform(n)); 1853 return true; 1854 } 1855 1856 //------------------------------runtime_math----------------------------- 1857 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1858 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1859 "must be (DD)D or (D)D type"); 1860 1861 // Inputs 1862 Node* a = round_double_node(argument(0)); 1863 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1864 1865 const TypePtr* no_memory_effects = NULL; 1866 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1867 no_memory_effects, 1868 a, top(), b, b ? top() : NULL); 1869 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1870 #ifdef ASSERT 1871 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1872 assert(value_top == top(), "second value must be top"); 1873 #endif 1874 1875 set_result(value); 1876 return true; 1877 } 1878 1879 //------------------------------inline_math_native----------------------------- 1880 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1881 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1882 switch (id) { 1883 // These intrinsics are not properly supported on all hardware 1884 case vmIntrinsics::_dsin: 1885 return StubRoutines::dsin() != NULL ? 1886 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1887 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1888 case vmIntrinsics::_dcos: 1889 return StubRoutines::dcos() != NULL ? 1890 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1891 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1892 case vmIntrinsics::_dtan: 1893 return StubRoutines::dtan() != NULL ? 1894 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1895 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1896 case vmIntrinsics::_dlog: 1897 return StubRoutines::dlog() != NULL ? 1898 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1899 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1900 case vmIntrinsics::_dlog10: 1901 return StubRoutines::dlog10() != NULL ? 1902 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1903 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1904 1905 // These intrinsics are supported on all hardware 1906 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_double_math(id) : false; 1907 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_double_math(id) : false; 1908 case vmIntrinsics::_fabs: return Matcher::match_rule_supported(Op_AbsF) ? inline_math(id) : false; 1909 case vmIntrinsics::_iabs: return Matcher::match_rule_supported(Op_AbsI) ? inline_math(id) : false; 1910 case vmIntrinsics::_labs: return Matcher::match_rule_supported(Op_AbsL) ? inline_math(id) : false; 1911 1912 case vmIntrinsics::_dexp: 1913 return StubRoutines::dexp() != NULL ? 1914 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1915 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1916 case vmIntrinsics::_dpow: { 1917 Node* exp = round_double_node(argument(2)); 1918 const TypeD* d = _gvn.type(exp)->isa_double_constant(); 1919 if (d != NULL && d->getd() == 2.0) { 1920 // Special case: pow(x, 2.0) => x * x 1921 Node* base = round_double_node(argument(0)); 1922 set_result(_gvn.transform(new MulDNode(base, base))); 1923 return true; 1924 } 1925 return StubRoutines::dpow() != NULL ? 1926 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1927 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1928 } 1929 #undef FN_PTR 1930 1931 // These intrinsics are not yet correctly implemented 1932 case vmIntrinsics::_datan2: 1933 return false; 1934 1935 default: 1936 fatal_unexpected_iid(id); 1937 return false; 1938 } 1939 } 1940 1941 static bool is_simple_name(Node* n) { 1942 return (n->req() == 1 // constant 1943 || (n->is_Type() && n->as_Type()->type()->singleton()) 1944 || n->is_Proj() // parameter or return value 1945 || n->is_Phi() // local of some sort 1946 ); 1947 } 1948 1949 //----------------------------inline_notify-----------------------------------* 1950 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1951 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1952 address func; 1953 if (id == vmIntrinsics::_notify) { 1954 func = OptoRuntime::monitor_notify_Java(); 1955 } else { 1956 func = OptoRuntime::monitor_notifyAll_Java(); 1957 } 1958 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1959 make_slow_call_ex(call, env()->Throwable_klass(), false); 1960 return true; 1961 } 1962 1963 1964 //----------------------------inline_min_max----------------------------------- 1965 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1966 set_result(generate_min_max(id, argument(0), argument(1))); 1967 return true; 1968 } 1969 1970 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1971 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1972 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1973 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1974 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1975 1976 { 1977 PreserveJVMState pjvms(this); 1978 PreserveReexecuteState preexecs(this); 1979 jvms()->set_should_reexecute(true); 1980 1981 set_control(slow_path); 1982 set_i_o(i_o()); 1983 1984 uncommon_trap(Deoptimization::Reason_intrinsic, 1985 Deoptimization::Action_none); 1986 } 1987 1988 set_control(fast_path); 1989 set_result(math); 1990 } 1991 1992 template <typename OverflowOp> 1993 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1994 typedef typename OverflowOp::MathOp MathOp; 1995 1996 MathOp* mathOp = new MathOp(arg1, arg2); 1997 Node* operation = _gvn.transform( mathOp ); 1998 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1999 inline_math_mathExact(operation, ofcheck); 2000 return true; 2001 } 2002 2003 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2004 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2005 } 2006 2007 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2008 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2009 } 2010 2011 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2012 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2013 } 2014 2015 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2016 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2017 } 2018 2019 bool LibraryCallKit::inline_math_negateExactI() { 2020 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2021 } 2022 2023 bool LibraryCallKit::inline_math_negateExactL() { 2024 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2025 } 2026 2027 bool LibraryCallKit::inline_math_multiplyExactI() { 2028 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2029 } 2030 2031 bool LibraryCallKit::inline_math_multiplyExactL() { 2032 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2033 } 2034 2035 bool LibraryCallKit::inline_math_multiplyHigh() { 2036 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2)))); 2037 return true; 2038 } 2039 2040 Node* 2041 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2042 // These are the candidate return value: 2043 Node* xvalue = x0; 2044 Node* yvalue = y0; 2045 2046 if (xvalue == yvalue) { 2047 return xvalue; 2048 } 2049 2050 bool want_max = (id == vmIntrinsics::_max); 2051 2052 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2053 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2054 if (txvalue == NULL || tyvalue == NULL) return top(); 2055 // This is not really necessary, but it is consistent with a 2056 // hypothetical MaxINode::Value method: 2057 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2058 2059 // %%% This folding logic should (ideally) be in a different place. 2060 // Some should be inside IfNode, and there to be a more reliable 2061 // transformation of ?: style patterns into cmoves. We also want 2062 // more powerful optimizations around cmove and min/max. 2063 2064 // Try to find a dominating comparison of these guys. 2065 // It can simplify the index computation for Arrays.copyOf 2066 // and similar uses of System.arraycopy. 2067 // First, compute the normalized version of CmpI(x, y). 2068 int cmp_op = Op_CmpI; 2069 Node* xkey = xvalue; 2070 Node* ykey = yvalue; 2071 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2072 if (ideal_cmpxy->is_Cmp()) { 2073 // E.g., if we have CmpI(length - offset, count), 2074 // it might idealize to CmpI(length, count + offset) 2075 cmp_op = ideal_cmpxy->Opcode(); 2076 xkey = ideal_cmpxy->in(1); 2077 ykey = ideal_cmpxy->in(2); 2078 } 2079 2080 // Start by locating any relevant comparisons. 2081 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2082 Node* cmpxy = NULL; 2083 Node* cmpyx = NULL; 2084 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2085 Node* cmp = start_from->fast_out(k); 2086 if (cmp->outcnt() > 0 && // must have prior uses 2087 cmp->in(0) == NULL && // must be context-independent 2088 cmp->Opcode() == cmp_op) { // right kind of compare 2089 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2090 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2091 } 2092 } 2093 2094 const int NCMPS = 2; 2095 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2096 int cmpn; 2097 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2098 if (cmps[cmpn] != NULL) break; // find a result 2099 } 2100 if (cmpn < NCMPS) { 2101 // Look for a dominating test that tells us the min and max. 2102 int depth = 0; // Limit search depth for speed 2103 Node* dom = control(); 2104 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2105 if (++depth >= 100) break; 2106 Node* ifproj = dom; 2107 if (!ifproj->is_Proj()) continue; 2108 Node* iff = ifproj->in(0); 2109 if (!iff->is_If()) continue; 2110 Node* bol = iff->in(1); 2111 if (!bol->is_Bool()) continue; 2112 Node* cmp = bol->in(1); 2113 if (cmp == NULL) continue; 2114 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2115 if (cmps[cmpn] == cmp) break; 2116 if (cmpn == NCMPS) continue; 2117 BoolTest::mask btest = bol->as_Bool()->_test._test; 2118 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2119 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2120 // At this point, we know that 'x btest y' is true. 2121 switch (btest) { 2122 case BoolTest::eq: 2123 // They are proven equal, so we can collapse the min/max. 2124 // Either value is the answer. Choose the simpler. 2125 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2126 return yvalue; 2127 return xvalue; 2128 case BoolTest::lt: // x < y 2129 case BoolTest::le: // x <= y 2130 return (want_max ? yvalue : xvalue); 2131 case BoolTest::gt: // x > y 2132 case BoolTest::ge: // x >= y 2133 return (want_max ? xvalue : yvalue); 2134 default: 2135 break; 2136 } 2137 } 2138 } 2139 2140 // We failed to find a dominating test. 2141 // Let's pick a test that might GVN with prior tests. 2142 Node* best_bol = NULL; 2143 BoolTest::mask best_btest = BoolTest::illegal; 2144 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2145 Node* cmp = cmps[cmpn]; 2146 if (cmp == NULL) continue; 2147 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2148 Node* bol = cmp->fast_out(j); 2149 if (!bol->is_Bool()) continue; 2150 BoolTest::mask btest = bol->as_Bool()->_test._test; 2151 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2152 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2153 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2154 best_bol = bol->as_Bool(); 2155 best_btest = btest; 2156 } 2157 } 2158 } 2159 2160 Node* answer_if_true = NULL; 2161 Node* answer_if_false = NULL; 2162 switch (best_btest) { 2163 default: 2164 if (cmpxy == NULL) 2165 cmpxy = ideal_cmpxy; 2166 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2167 // and fall through: 2168 case BoolTest::lt: // x < y 2169 case BoolTest::le: // x <= y 2170 answer_if_true = (want_max ? yvalue : xvalue); 2171 answer_if_false = (want_max ? xvalue : yvalue); 2172 break; 2173 case BoolTest::gt: // x > y 2174 case BoolTest::ge: // x >= y 2175 answer_if_true = (want_max ? xvalue : yvalue); 2176 answer_if_false = (want_max ? yvalue : xvalue); 2177 break; 2178 } 2179 2180 jint hi, lo; 2181 if (want_max) { 2182 // We can sharpen the minimum. 2183 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2184 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2185 } else { 2186 // We can sharpen the maximum. 2187 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2188 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2189 } 2190 2191 // Use a flow-free graph structure, to avoid creating excess control edges 2192 // which could hinder other optimizations. 2193 // Since Math.min/max is often used with arraycopy, we want 2194 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2195 Node* cmov = CMoveNode::make(NULL, best_bol, 2196 answer_if_false, answer_if_true, 2197 TypeInt::make(lo, hi, widen)); 2198 2199 return _gvn.transform(cmov); 2200 2201 /* 2202 // This is not as desirable as it may seem, since Min and Max 2203 // nodes do not have a full set of optimizations. 2204 // And they would interfere, anyway, with 'if' optimizations 2205 // and with CMoveI canonical forms. 2206 switch (id) { 2207 case vmIntrinsics::_min: 2208 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2209 case vmIntrinsics::_max: 2210 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2211 default: 2212 ShouldNotReachHere(); 2213 } 2214 */ 2215 } 2216 2217 inline int 2218 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) { 2219 const TypePtr* base_type = TypePtr::NULL_PTR; 2220 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2221 if (base_type == NULL) { 2222 // Unknown type. 2223 return Type::AnyPtr; 2224 } else if (base_type == TypePtr::NULL_PTR) { 2225 // Since this is a NULL+long form, we have to switch to a rawptr. 2226 base = _gvn.transform(new CastX2PNode(offset)); 2227 offset = MakeConX(0); 2228 return Type::RawPtr; 2229 } else if (base_type->base() == Type::RawPtr) { 2230 return Type::RawPtr; 2231 } else if (base_type->isa_oopptr()) { 2232 // Base is never null => always a heap address. 2233 if (!TypePtr::NULL_PTR->higher_equal(base_type)) { 2234 return Type::OopPtr; 2235 } 2236 // Offset is small => always a heap address. 2237 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2238 if (offset_type != NULL && 2239 base_type->offset() == 0 && // (should always be?) 2240 offset_type->_lo >= 0 && 2241 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2242 return Type::OopPtr; 2243 } else if (type == T_OBJECT) { 2244 // off heap access to an oop doesn't make any sense. Has to be on 2245 // heap. 2246 return Type::OopPtr; 2247 } 2248 // Otherwise, it might either be oop+off or NULL+addr. 2249 return Type::AnyPtr; 2250 } else { 2251 // No information: 2252 return Type::AnyPtr; 2253 } 2254 } 2255 2256 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type, bool can_cast) { 2257 Node* uncasted_base = base; 2258 int kind = classify_unsafe_addr(uncasted_base, offset, type); 2259 if (kind == Type::RawPtr) { 2260 return basic_plus_adr(top(), uncasted_base, offset); 2261 } else if (kind == Type::AnyPtr) { 2262 assert(base == uncasted_base, "unexpected base change"); 2263 if (can_cast) { 2264 if (!_gvn.type(base)->speculative_maybe_null() && 2265 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2266 // According to profiling, this access is always on 2267 // heap. Casting the base to not null and thus avoiding membars 2268 // around the access should allow better optimizations 2269 Node* null_ctl = top(); 2270 base = null_check_oop(base, &null_ctl, true, true, true); 2271 assert(null_ctl->is_top(), "no null control here"); 2272 return basic_plus_adr(base, offset); 2273 } else if (_gvn.type(base)->speculative_always_null() && 2274 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2275 // According to profiling, this access is always off 2276 // heap. 2277 base = null_assert(base); 2278 Node* raw_base = _gvn.transform(new CastX2PNode(offset)); 2279 offset = MakeConX(0); 2280 return basic_plus_adr(top(), raw_base, offset); 2281 } 2282 } 2283 // We don't know if it's an on heap or off heap access. Fall back 2284 // to raw memory access. 2285 base = access_resolve(base, decorators); 2286 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM)); 2287 return basic_plus_adr(top(), raw, offset); 2288 } else { 2289 assert(base == uncasted_base, "unexpected base change"); 2290 // We know it's an on heap access so base can't be null 2291 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) { 2292 base = must_be_not_null(base, true); 2293 } 2294 return basic_plus_adr(base, offset); 2295 } 2296 } 2297 2298 //--------------------------inline_number_methods----------------------------- 2299 // inline int Integer.numberOfLeadingZeros(int) 2300 // inline int Long.numberOfLeadingZeros(long) 2301 // 2302 // inline int Integer.numberOfTrailingZeros(int) 2303 // inline int Long.numberOfTrailingZeros(long) 2304 // 2305 // inline int Integer.bitCount(int) 2306 // inline int Long.bitCount(long) 2307 // 2308 // inline char Character.reverseBytes(char) 2309 // inline short Short.reverseBytes(short) 2310 // inline int Integer.reverseBytes(int) 2311 // inline long Long.reverseBytes(long) 2312 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2313 Node* arg = argument(0); 2314 Node* n = NULL; 2315 switch (id) { 2316 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2317 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2318 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2319 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2320 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2321 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2322 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2323 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2324 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2325 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2326 default: fatal_unexpected_iid(id); break; 2327 } 2328 set_result(_gvn.transform(n)); 2329 return true; 2330 } 2331 2332 //----------------------------inline_unsafe_access---------------------------- 2333 2334 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2335 // Attempt to infer a sharper value type from the offset and base type. 2336 ciKlass* sharpened_klass = NULL; 2337 2338 // See if it is an instance field, with an object type. 2339 if (alias_type->field() != NULL) { 2340 if (alias_type->field()->type()->is_klass()) { 2341 sharpened_klass = alias_type->field()->type()->as_klass(); 2342 } 2343 } 2344 2345 // See if it is a narrow oop array. 2346 if (adr_type->isa_aryptr()) { 2347 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2348 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2349 if (elem_type != NULL) { 2350 sharpened_klass = elem_type->klass(); 2351 } 2352 } 2353 } 2354 2355 // The sharpened class might be unloaded if there is no class loader 2356 // contraint in place. 2357 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2358 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2359 2360 #ifndef PRODUCT 2361 if (C->print_intrinsics() || C->print_inlining()) { 2362 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2363 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2364 } 2365 #endif 2366 // Sharpen the value type. 2367 return tjp; 2368 } 2369 return NULL; 2370 } 2371 2372 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) { 2373 switch (kind) { 2374 case Relaxed: 2375 return MO_UNORDERED; 2376 case Opaque: 2377 return MO_RELAXED; 2378 case Acquire: 2379 return MO_ACQUIRE; 2380 case Release: 2381 return MO_RELEASE; 2382 case Volatile: 2383 return MO_SEQ_CST; 2384 default: 2385 ShouldNotReachHere(); 2386 return 0; 2387 } 2388 } 2389 2390 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2391 if (callee()->is_static()) return false; // caller must have the capability! 2392 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2393 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2394 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2395 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2396 2397 if (type == T_OBJECT || type == T_ARRAY) { 2398 decorators |= ON_UNKNOWN_OOP_REF; 2399 } 2400 2401 if (unaligned) { 2402 decorators |= C2_UNALIGNED; 2403 } 2404 2405 #ifndef PRODUCT 2406 { 2407 ResourceMark rm; 2408 // Check the signatures. 2409 ciSignature* sig = callee()->signature(); 2410 #ifdef ASSERT 2411 if (!is_store) { 2412 // Object getReference(Object base, int/long offset), etc. 2413 BasicType rtype = sig->return_type()->basic_type(); 2414 assert(rtype == type || (rtype == T_OBJECT && type == T_VALUETYPE), "getter must return the expected value"); 2415 assert(sig->count() == 2 || (type == T_VALUETYPE && sig->count() == 3), "oop getter has 2 or 3 arguments"); 2416 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2417 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2418 } else { 2419 // void putReference(Object base, int/long offset, Object x), etc. 2420 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2421 assert(sig->count() == 3 || (type == T_VALUETYPE && sig->count() == 4), "oop putter has 3 arguments"); 2422 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2423 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2424 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2425 assert(vtype == type || (type == T_VALUETYPE && vtype == T_OBJECT), "putter must accept the expected value"); 2426 } 2427 #endif // ASSERT 2428 } 2429 #endif //PRODUCT 2430 2431 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2432 2433 Node* receiver = argument(0); // type: oop 2434 2435 // Build address expression. 2436 Node* adr; 2437 Node* heap_base_oop = top(); 2438 Node* offset = top(); 2439 Node* val; 2440 2441 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2442 Node* base = argument(1); // type: oop 2443 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2444 offset = argument(2); // type: long 2445 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2446 // to be plain byte offsets, which are also the same as those accepted 2447 // by oopDesc::field_addr. 2448 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2449 "fieldOffset must be byte-scaled"); 2450 2451 ciValueKlass* value_klass = NULL; 2452 if (type == T_VALUETYPE) { 2453 Node* cls = null_check(argument(4)); 2454 if (stopped()) { 2455 return true; 2456 } 2457 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2458 const TypeKlassPtr* kls_t = _gvn.type(kls)->isa_klassptr(); 2459 if (!kls_t->klass_is_exact()) { 2460 return false; 2461 } 2462 ciKlass* klass = kls_t->klass(); 2463 if (!klass->is_valuetype()) { 2464 return false; 2465 } 2466 value_klass = klass->as_value_klass(); 2467 } 2468 2469 receiver = null_check(receiver); 2470 if (stopped()) { 2471 return true; 2472 } 2473 2474 if (base->is_ValueType()) { 2475 ValueTypeNode* vt = base->as_ValueType(); 2476 2477 if (is_store) { 2478 if (!vt->is_allocated(&_gvn) || !_gvn.type(vt)->is_valuetype()->larval()) { 2479 return false; 2480 } 2481 base = vt->get_oop(); 2482 } else { 2483 if (offset->is_Con()) { 2484 long off = find_long_con(offset, 0); 2485 ciValueKlass* vk = vt->type()->value_klass(); 2486 if ((long)(int)off != off || !vk->contains_field_offset(off)) { 2487 return false; 2488 } 2489 2490 ciField* f = vk->get_non_flattened_field_by_offset((int)off); 2491 2492 if (f != NULL) { 2493 BasicType bt = f->layout_type(); 2494 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2495 bt = T_OBJECT; 2496 } 2497 if (bt == type) { 2498 if (bt != T_VALUETYPE || f->type() == value_klass) { 2499 set_result(vt->field_value_by_offset((int)off, false)); 2500 return true; 2501 } 2502 } 2503 } 2504 } 2505 vt = vt->allocate(this)->as_ValueType(); 2506 base = vt->get_oop(); 2507 } 2508 } 2509 2510 // 32-bit machines ignore the high half! 2511 offset = ConvL2X(offset); 2512 adr = make_unsafe_address(base, offset, is_store ? ACCESS_WRITE : ACCESS_READ, type, kind == Relaxed); 2513 2514 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) { 2515 heap_base_oop = base; 2516 } else if (type == T_OBJECT || (value_klass != NULL && value_klass->has_object_fields())) { 2517 return false; // off-heap oop accesses are not supported 2518 } 2519 2520 // Can base be NULL? Otherwise, always on-heap access. 2521 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base)); 2522 2523 if (!can_access_non_heap) { 2524 decorators |= IN_HEAP; 2525 } 2526 2527 val = is_store ? argument(4 + (type == T_VALUETYPE ? 1 : 0)) : NULL; 2528 2529 const TypePtr* adr_type = _gvn.type(adr)->isa_ptr(); 2530 if (adr_type == TypePtr::NULL_PTR) { 2531 return false; // off-heap access with zero address 2532 } 2533 2534 // Try to categorize the address. 2535 Compile::AliasType* alias_type = C->alias_type(adr_type); 2536 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2537 2538 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2539 alias_type->adr_type() == TypeAryPtr::RANGE) { 2540 return false; // not supported 2541 } 2542 2543 bool mismatched = false; 2544 BasicType bt = T_ILLEGAL; 2545 ciField* field = NULL; 2546 if (adr_type->isa_instptr()) { 2547 const TypeInstPtr* instptr = adr_type->is_instptr(); 2548 ciInstanceKlass* k = instptr->klass()->as_instance_klass(); 2549 int off = instptr->offset(); 2550 if (instptr->const_oop() != NULL && 2551 instptr->klass() == ciEnv::current()->Class_klass() && 2552 instptr->offset() >= (instptr->klass()->as_instance_klass()->size_helper() * wordSize)) { 2553 k = instptr->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 2554 field = k->get_field_by_offset(off, true); 2555 } else { 2556 field = k->get_non_flattened_field_by_offset(off); 2557 } 2558 if (field != NULL) { 2559 bt = field->layout_type(); 2560 } 2561 assert(bt == alias_type->basic_type() || bt == T_VALUETYPE, "should match"); 2562 if (field != NULL && bt == T_VALUETYPE && !field->is_flattened()) { 2563 bt = T_OBJECT; 2564 } 2565 } else { 2566 bt = alias_type->basic_type(); 2567 } 2568 2569 if (bt != T_ILLEGAL) { 2570 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2571 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2572 // Alias type doesn't differentiate between byte[] and boolean[]). 2573 // Use address type to get the element type. 2574 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2575 } 2576 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2577 // accessing an array field with getReference is not a mismatch 2578 bt = T_OBJECT; 2579 } 2580 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2581 // Don't intrinsify mismatched object accesses 2582 return false; 2583 } 2584 mismatched = (bt != type); 2585 } else if (alias_type->adr_type()->isa_oopptr()) { 2586 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched 2587 } 2588 2589 if (type == T_VALUETYPE) { 2590 if (adr_type->isa_instptr()) { 2591 if (field == NULL || field->type() != value_klass) { 2592 mismatched = true; 2593 } 2594 } else if (adr_type->isa_aryptr()) { 2595 const Type* elem = adr_type->is_aryptr()->elem(); 2596 if (!elem->isa_valuetype()) { 2597 mismatched = true; 2598 } else if (elem->value_klass() != value_klass) { 2599 mismatched = true; 2600 } 2601 } 2602 if (is_store) { 2603 const Type* val_t = _gvn.type(val); 2604 if (!val_t->isa_valuetype() || val_t->value_klass() != value_klass) { 2605 return false; 2606 } 2607 } 2608 } 2609 2610 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2611 2612 if (mismatched) { 2613 decorators |= C2_MISMATCHED; 2614 } 2615 2616 // First guess at the value type. 2617 const Type *value_type = Type::get_const_basic_type(type); 2618 2619 // Figure out the memory ordering. 2620 decorators |= mo_decorator_for_access_kind(kind); 2621 2622 if (!is_store) { 2623 if (type == T_OBJECT) { 2624 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2625 if (tjp != NULL) { 2626 value_type = tjp; 2627 } 2628 } else if (type == T_VALUETYPE) { 2629 value_type = NULL; 2630 } 2631 } 2632 2633 // Heap pointers get a null-check from the interpreter, 2634 // as a courtesy. However, this is not guaranteed by Unsafe, 2635 // and it is not possible to fully distinguish unintended nulls 2636 // from intended ones in this API. 2637 2638 if (!is_store) { 2639 Node* p = NULL; 2640 // Try to constant fold a load from a constant field 2641 2642 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { 2643 // final or stable field 2644 p = make_constant_from_field(field, heap_base_oop); 2645 } 2646 2647 if (p == NULL) { // Could not constant fold the load 2648 if (type == T_VALUETYPE) { 2649 if (adr_type->isa_instptr() && !mismatched) { 2650 ciInstanceKlass* holder = adr_type->is_instptr()->klass()->as_instance_klass(); 2651 int offset = adr_type->is_instptr()->offset(); 2652 p = ValueTypeNode::make_from_flattened(this, value_klass, base, base, holder, offset, decorators); 2653 } else { 2654 p = ValueTypeNode::make_from_flattened(this, value_klass, base, adr, NULL, 0, decorators); 2655 } 2656 } else { 2657 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators); 2658 } 2659 // Normalize the value returned by getBoolean in the following cases 2660 if (type == T_BOOLEAN && 2661 (mismatched || 2662 heap_base_oop == top() || // - heap_base_oop is NULL or 2663 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL 2664 // and the unsafe access is made to large offset 2665 // (i.e., larger than the maximum offset necessary for any 2666 // field access) 2667 ) { 2668 IdealKit ideal = IdealKit(this); 2669 #define __ ideal. 2670 IdealVariable normalized_result(ideal); 2671 __ declarations_done(); 2672 __ set(normalized_result, p); 2673 __ if_then(p, BoolTest::ne, ideal.ConI(0)); 2674 __ set(normalized_result, ideal.ConI(1)); 2675 ideal.end_if(); 2676 final_sync(ideal); 2677 p = __ value(normalized_result); 2678 #undef __ 2679 } 2680 } 2681 if (type == T_ADDRESS) { 2682 p = gvn().transform(new CastP2XNode(NULL, p)); 2683 p = ConvX2UL(p); 2684 } 2685 if (field != NULL && field->is_flattenable() && !field->is_flattened()) { 2686 // Load a non-flattened but flattenable value type from memory 2687 if (value_type->value_klass()->is_scalarizable()) { 2688 p = ValueTypeNode::make_from_oop(this, p, value_type->value_klass()); 2689 } else { 2690 p = null2default(p, value_type->value_klass()); 2691 } 2692 } 2693 // The load node has the control of the preceding MemBarCPUOrder. All 2694 // following nodes will have the control of the MemBarCPUOrder inserted at 2695 // the end of this method. So, pushing the load onto the stack at a later 2696 // point is fine. 2697 set_result(p); 2698 } else { 2699 if (bt == T_ADDRESS) { 2700 // Repackage the long as a pointer. 2701 val = ConvL2X(val); 2702 val = gvn().transform(new CastX2PNode(val)); 2703 } 2704 if (type == T_VALUETYPE) { 2705 if (adr_type->isa_instptr() && !mismatched) { 2706 ciInstanceKlass* holder = adr_type->is_instptr()->klass()->as_instance_klass(); 2707 int offset = adr_type->is_instptr()->offset(); 2708 val->as_ValueType()->store_flattened(this, base, base, holder, offset, decorators); 2709 } else { 2710 val->as_ValueType()->store_flattened(this, base, adr, NULL, 0, decorators); 2711 } 2712 } else { 2713 access_store_at(heap_base_oop, adr, adr_type, val, value_type, type, decorators); 2714 } 2715 } 2716 2717 if (argument(1)->is_ValueType() && is_store) { 2718 Node* value = ValueTypeNode::make_from_oop(this, base, _gvn.type(base)->value_klass()); 2719 value = value->as_ValueType()->make_larval(this, false); 2720 replace_in_map(argument(1), value); 2721 } 2722 2723 return true; 2724 } 2725 2726 bool LibraryCallKit::inline_unsafe_make_private_buffer() { 2727 Node* receiver = argument(0); 2728 Node* value = argument(1); 2729 2730 receiver = null_check(receiver); 2731 if (stopped()) { 2732 return true; 2733 } 2734 2735 if (!value->is_ValueType()) { 2736 return false; 2737 } 2738 2739 set_result(value->as_ValueType()->make_larval(this, true)); 2740 2741 return true; 2742 } 2743 2744 bool LibraryCallKit::inline_unsafe_finish_private_buffer() { 2745 Node* receiver = argument(0); 2746 Node* buffer = argument(1); 2747 2748 receiver = null_check(receiver); 2749 if (stopped()) { 2750 return true; 2751 } 2752 2753 if (!buffer->is_ValueType()) { 2754 return false; 2755 } 2756 2757 ValueTypeNode* vt = buffer->as_ValueType(); 2758 if (!vt->is_allocated(&_gvn) || !_gvn.type(vt)->is_valuetype()->larval()) { 2759 return false; 2760 } 2761 2762 set_result(vt->finish_larval(this)); 2763 2764 return true; 2765 } 2766 2767 //----------------------------inline_unsafe_load_store---------------------------- 2768 // This method serves a couple of different customers (depending on LoadStoreKind): 2769 // 2770 // LS_cmp_swap: 2771 // 2772 // boolean compareAndSetReference(Object o, long offset, Object expected, Object x); 2773 // boolean compareAndSetInt( Object o, long offset, int expected, int x); 2774 // boolean compareAndSetLong( Object o, long offset, long expected, long x); 2775 // 2776 // LS_cmp_swap_weak: 2777 // 2778 // boolean weakCompareAndSetReference( Object o, long offset, Object expected, Object x); 2779 // boolean weakCompareAndSetReferencePlain( Object o, long offset, Object expected, Object x); 2780 // boolean weakCompareAndSetReferenceAcquire(Object o, long offset, Object expected, Object x); 2781 // boolean weakCompareAndSetReferenceRelease(Object o, long offset, Object expected, Object x); 2782 // 2783 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x); 2784 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x); 2785 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x); 2786 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x); 2787 // 2788 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x); 2789 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x); 2790 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x); 2791 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x); 2792 // 2793 // LS_cmp_exchange: 2794 // 2795 // Object compareAndExchangeReferenceVolatile(Object o, long offset, Object expected, Object x); 2796 // Object compareAndExchangeReferenceAcquire( Object o, long offset, Object expected, Object x); 2797 // Object compareAndExchangeReferenceRelease( Object o, long offset, Object expected, Object x); 2798 // 2799 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2800 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2801 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2802 // 2803 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2804 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2805 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2806 // 2807 // LS_get_add: 2808 // 2809 // int getAndAddInt( Object o, long offset, int delta) 2810 // long getAndAddLong(Object o, long offset, long delta) 2811 // 2812 // LS_get_set: 2813 // 2814 // int getAndSet(Object o, long offset, int newValue) 2815 // long getAndSet(Object o, long offset, long newValue) 2816 // Object getAndSet(Object o, long offset, Object newValue) 2817 // 2818 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2819 // This basic scheme here is the same as inline_unsafe_access, but 2820 // differs in enough details that combining them would make the code 2821 // overly confusing. (This is a true fact! I originally combined 2822 // them, but even I was confused by it!) As much code/comments as 2823 // possible are retained from inline_unsafe_access though to make 2824 // the correspondences clearer. - dl 2825 2826 if (callee()->is_static()) return false; // caller must have the capability! 2827 2828 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2829 decorators |= mo_decorator_for_access_kind(access_kind); 2830 2831 #ifndef PRODUCT 2832 BasicType rtype; 2833 { 2834 ResourceMark rm; 2835 // Check the signatures. 2836 ciSignature* sig = callee()->signature(); 2837 rtype = sig->return_type()->basic_type(); 2838 switch(kind) { 2839 case LS_get_add: 2840 case LS_get_set: { 2841 // Check the signatures. 2842 #ifdef ASSERT 2843 assert(rtype == type, "get and set must return the expected type"); 2844 assert(sig->count() == 3, "get and set has 3 arguments"); 2845 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2846 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2847 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2848 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2849 #endif // ASSERT 2850 break; 2851 } 2852 case LS_cmp_swap: 2853 case LS_cmp_swap_weak: { 2854 // Check the signatures. 2855 #ifdef ASSERT 2856 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2857 assert(sig->count() == 4, "CAS has 4 arguments"); 2858 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2859 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2860 #endif // ASSERT 2861 break; 2862 } 2863 case LS_cmp_exchange: { 2864 // Check the signatures. 2865 #ifdef ASSERT 2866 assert(rtype == type, "CAS must return the expected type"); 2867 assert(sig->count() == 4, "CAS has 4 arguments"); 2868 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2869 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2870 #endif // ASSERT 2871 break; 2872 } 2873 default: 2874 ShouldNotReachHere(); 2875 } 2876 } 2877 #endif //PRODUCT 2878 2879 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2880 2881 // Get arguments: 2882 Node* receiver = NULL; 2883 Node* base = NULL; 2884 Node* offset = NULL; 2885 Node* oldval = NULL; 2886 Node* newval = NULL; 2887 switch(kind) { 2888 case LS_cmp_swap: 2889 case LS_cmp_swap_weak: 2890 case LS_cmp_exchange: { 2891 const bool two_slot_type = type2size[type] == 2; 2892 receiver = argument(0); // type: oop 2893 base = argument(1); // type: oop 2894 offset = argument(2); // type: long 2895 oldval = argument(4); // type: oop, int, or long 2896 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2897 break; 2898 } 2899 case LS_get_add: 2900 case LS_get_set: { 2901 receiver = argument(0); // type: oop 2902 base = argument(1); // type: oop 2903 offset = argument(2); // type: long 2904 oldval = NULL; 2905 newval = argument(4); // type: oop, int, or long 2906 break; 2907 } 2908 default: 2909 ShouldNotReachHere(); 2910 } 2911 2912 // Build field offset expression. 2913 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2914 // to be plain byte offsets, which are also the same as those accepted 2915 // by oopDesc::field_addr. 2916 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2917 // 32-bit machines ignore the high half of long offsets 2918 offset = ConvL2X(offset); 2919 Node* adr = make_unsafe_address(base, offset, ACCESS_WRITE | ACCESS_READ, type, false); 2920 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2921 2922 Compile::AliasType* alias_type = C->alias_type(adr_type); 2923 BasicType bt = alias_type->basic_type(); 2924 if (bt != T_ILLEGAL && 2925 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2926 // Don't intrinsify mismatched object accesses. 2927 return false; 2928 } 2929 2930 // For CAS, unlike inline_unsafe_access, there seems no point in 2931 // trying to refine types. Just use the coarse types here. 2932 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2933 const Type *value_type = Type::get_const_basic_type(type); 2934 2935 switch (kind) { 2936 case LS_get_set: 2937 case LS_cmp_exchange: { 2938 if (type == T_OBJECT) { 2939 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2940 if (tjp != NULL) { 2941 value_type = tjp; 2942 } 2943 } 2944 break; 2945 } 2946 case LS_cmp_swap: 2947 case LS_cmp_swap_weak: 2948 case LS_get_add: 2949 break; 2950 default: 2951 ShouldNotReachHere(); 2952 } 2953 2954 // Null check receiver. 2955 receiver = null_check(receiver); 2956 if (stopped()) { 2957 return true; 2958 } 2959 2960 int alias_idx = C->get_alias_index(adr_type); 2961 2962 if (type == T_OBJECT || type == T_ARRAY) { 2963 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF; 2964 2965 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2966 // could be delayed during Parse (for example, in adjust_map_after_if()). 2967 // Execute transformation here to avoid barrier generation in such case. 2968 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2969 newval = _gvn.makecon(TypePtr::NULL_PTR); 2970 2971 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) { 2972 // Refine the value to a null constant, when it is known to be null 2973 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2974 } 2975 } 2976 2977 Node* result = NULL; 2978 switch (kind) { 2979 case LS_cmp_exchange: { 2980 result = access_atomic_cmpxchg_val_at(base, adr, adr_type, alias_idx, 2981 oldval, newval, value_type, type, decorators); 2982 break; 2983 } 2984 case LS_cmp_swap_weak: 2985 decorators |= C2_WEAK_CMPXCHG; 2986 case LS_cmp_swap: { 2987 result = access_atomic_cmpxchg_bool_at(base, adr, adr_type, alias_idx, 2988 oldval, newval, value_type, type, decorators); 2989 break; 2990 } 2991 case LS_get_set: { 2992 result = access_atomic_xchg_at(base, adr, adr_type, alias_idx, 2993 newval, value_type, type, decorators); 2994 break; 2995 } 2996 case LS_get_add: { 2997 result = access_atomic_add_at(base, adr, adr_type, alias_idx, 2998 newval, value_type, type, decorators); 2999 break; 3000 } 3001 default: 3002 ShouldNotReachHere(); 3003 } 3004 3005 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3006 set_result(result); 3007 return true; 3008 } 3009 3010 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3011 // Regardless of form, don't allow previous ld/st to move down, 3012 // then issue acquire, release, or volatile mem_bar. 3013 insert_mem_bar(Op_MemBarCPUOrder); 3014 switch(id) { 3015 case vmIntrinsics::_loadFence: 3016 insert_mem_bar(Op_LoadFence); 3017 return true; 3018 case vmIntrinsics::_storeFence: 3019 insert_mem_bar(Op_StoreFence); 3020 return true; 3021 case vmIntrinsics::_fullFence: 3022 insert_mem_bar(Op_MemBarVolatile); 3023 return true; 3024 default: 3025 fatal_unexpected_iid(id); 3026 return false; 3027 } 3028 } 3029 3030 bool LibraryCallKit::inline_onspinwait() { 3031 insert_mem_bar(Op_OnSpinWait); 3032 return true; 3033 } 3034 3035 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3036 if (!kls->is_Con()) { 3037 return true; 3038 } 3039 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3040 if (klsptr == NULL) { 3041 return true; 3042 } 3043 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3044 // don't need a guard for a klass that is already initialized 3045 return !ik->is_initialized(); 3046 } 3047 3048 //----------------------------inline_unsafe_allocate--------------------------- 3049 // public native Object Unsafe.allocateInstance(Class<?> cls); 3050 bool LibraryCallKit::inline_unsafe_allocate() { 3051 if (callee()->is_static()) return false; // caller must have the capability! 3052 3053 null_check_receiver(); // null-check, then ignore 3054 Node* cls = null_check(argument(1)); 3055 if (stopped()) return true; 3056 3057 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3058 kls = null_check(kls); 3059 if (stopped()) return true; // argument was like int.class 3060 3061 Node* test = NULL; 3062 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3063 // Note: The argument might still be an illegal value like 3064 // Serializable.class or Object[].class. The runtime will handle it. 3065 // But we must make an explicit check for initialization. 3066 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3067 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3068 // can generate code to load it as unsigned byte. 3069 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3070 Node* bits = intcon(InstanceKlass::fully_initialized); 3071 test = _gvn.transform(new SubINode(inst, bits)); 3072 // The 'test' is non-zero if we need to take a slow path. 3073 } 3074 3075 Node* obj = new_instance(kls, test); 3076 set_result(obj); 3077 return true; 3078 } 3079 3080 //------------------------inline_native_time_funcs-------------- 3081 // inline code for System.currentTimeMillis() and System.nanoTime() 3082 // these have the same type and signature 3083 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3084 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3085 const TypePtr* no_memory_effects = NULL; 3086 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3087 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3088 #ifdef ASSERT 3089 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3090 assert(value_top == top(), "second value must be top"); 3091 #endif 3092 set_result(value); 3093 return true; 3094 } 3095 3096 #ifdef JFR_HAVE_INTRINSICS 3097 3098 /* 3099 * oop -> myklass 3100 * myklass->trace_id |= USED 3101 * return myklass->trace_id & ~0x3 3102 */ 3103 bool LibraryCallKit::inline_native_classID() { 3104 Node* cls = null_check(argument(0), T_OBJECT); 3105 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3106 kls = null_check(kls, T_OBJECT); 3107 3108 ByteSize offset = KLASS_TRACE_ID_OFFSET; 3109 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3110 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3111 3112 Node* clsused = longcon(0x01l); // set the class bit 3113 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3114 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3115 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3116 3117 #ifdef TRACE_ID_META_BITS 3118 Node* mbits = longcon(~TRACE_ID_META_BITS); 3119 tvalue = _gvn.transform(new AndLNode(tvalue, mbits)); 3120 #endif 3121 #ifdef TRACE_ID_SHIFT 3122 Node* cbits = intcon(TRACE_ID_SHIFT); 3123 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits)); 3124 #endif 3125 3126 set_result(tvalue); 3127 return true; 3128 3129 } 3130 3131 bool LibraryCallKit::inline_native_getEventWriter() { 3132 Node* tls_ptr = _gvn.transform(new ThreadLocalNode()); 3133 3134 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr, 3135 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)); 3136 3137 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered); 3138 3139 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) ); 3140 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) ); 3141 3142 IfNode* iff_jobj_null = 3143 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN); 3144 3145 enum { _normal_path = 1, 3146 _null_path = 2, 3147 PATH_LIMIT }; 3148 3149 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3150 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM); 3151 3152 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null)); 3153 result_rgn->init_req(_null_path, jobj_is_null); 3154 result_val->init_req(_null_path, null()); 3155 3156 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null)); 3157 set_control(jobj_is_not_null); 3158 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT, 3159 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD); 3160 result_rgn->init_req(_normal_path, control()); 3161 result_val->init_req(_normal_path, res); 3162 3163 set_result(result_rgn, result_val); 3164 3165 return true; 3166 } 3167 3168 #endif // JFR_HAVE_INTRINSICS 3169 3170 //------------------------inline_native_currentThread------------------ 3171 bool LibraryCallKit::inline_native_currentThread() { 3172 Node* junk = NULL; 3173 set_result(generate_current_thread(junk)); 3174 return true; 3175 } 3176 3177 //------------------------inline_native_isInterrupted------------------ 3178 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3179 bool LibraryCallKit::inline_native_isInterrupted() { 3180 // Add a fast path to t.isInterrupted(clear_int): 3181 // (t == Thread.current() && 3182 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3183 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3184 // So, in the common case that the interrupt bit is false, 3185 // we avoid making a call into the VM. Even if the interrupt bit 3186 // is true, if the clear_int argument is false, we avoid the VM call. 3187 // However, if the receiver is not currentThread, we must call the VM, 3188 // because there must be some locking done around the operation. 3189 3190 // We only go to the fast case code if we pass two guards. 3191 // Paths which do not pass are accumulated in the slow_region. 3192 3193 enum { 3194 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3195 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3196 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3197 PATH_LIMIT 3198 }; 3199 3200 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3201 // out of the function. 3202 insert_mem_bar(Op_MemBarCPUOrder); 3203 3204 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3205 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3206 3207 RegionNode* slow_region = new RegionNode(1); 3208 record_for_igvn(slow_region); 3209 3210 // (a) Receiving thread must be the current thread. 3211 Node* rec_thr = argument(0); 3212 Node* tls_ptr = NULL; 3213 Node* cur_thr = generate_current_thread(tls_ptr); 3214 3215 // Resolve oops to stable for CmpP below. 3216 cur_thr = access_resolve(cur_thr, 0); 3217 rec_thr = access_resolve(rec_thr, 0); 3218 3219 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3220 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3221 3222 generate_slow_guard(bol_thr, slow_region); 3223 3224 // (b) Interrupt bit on TLS must be false. 3225 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3226 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3227 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3228 3229 // Set the control input on the field _interrupted read to prevent it floating up. 3230 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3231 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3232 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3233 3234 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3235 3236 // First fast path: if (!TLS._interrupted) return false; 3237 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3238 result_rgn->init_req(no_int_result_path, false_bit); 3239 result_val->init_req(no_int_result_path, intcon(0)); 3240 3241 // drop through to next case 3242 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3243 3244 #ifndef _WINDOWS 3245 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3246 Node* clr_arg = argument(1); 3247 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3248 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3249 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3250 3251 // Second fast path: ... else if (!clear_int) return true; 3252 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3253 result_rgn->init_req(no_clear_result_path, false_arg); 3254 result_val->init_req(no_clear_result_path, intcon(1)); 3255 3256 // drop through to next case 3257 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3258 #else 3259 // To return true on Windows you must read the _interrupted field 3260 // and check the event state i.e. take the slow path. 3261 #endif // _WINDOWS 3262 3263 // (d) Otherwise, go to the slow path. 3264 slow_region->add_req(control()); 3265 set_control( _gvn.transform(slow_region)); 3266 3267 if (stopped()) { 3268 // There is no slow path. 3269 result_rgn->init_req(slow_result_path, top()); 3270 result_val->init_req(slow_result_path, top()); 3271 } else { 3272 // non-virtual because it is a private non-static 3273 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3274 3275 Node* slow_val = set_results_for_java_call(slow_call); 3276 // this->control() comes from set_results_for_java_call 3277 3278 Node* fast_io = slow_call->in(TypeFunc::I_O); 3279 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3280 3281 // These two phis are pre-filled with copies of of the fast IO and Memory 3282 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3283 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3284 3285 result_rgn->init_req(slow_result_path, control()); 3286 result_io ->init_req(slow_result_path, i_o()); 3287 result_mem->init_req(slow_result_path, reset_memory()); 3288 result_val->init_req(slow_result_path, slow_val); 3289 3290 set_all_memory(_gvn.transform(result_mem)); 3291 set_i_o( _gvn.transform(result_io)); 3292 } 3293 3294 C->set_has_split_ifs(true); // Has chance for split-if optimization 3295 set_result(result_rgn, result_val); 3296 return true; 3297 } 3298 3299 //-----------------------load_klass_from_mirror_common------------------------- 3300 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3301 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3302 // and branch to the given path on the region. 3303 // If never_see_null, take an uncommon trap on null, so we can optimistically 3304 // compile for the non-null case. 3305 // If the region is NULL, force never_see_null = true. 3306 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3307 bool never_see_null, 3308 RegionNode* region, 3309 int null_path, 3310 int offset) { 3311 if (region == NULL) never_see_null = true; 3312 Node* p = basic_plus_adr(mirror, offset); 3313 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3314 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3315 Node* null_ctl = top(); 3316 kls = null_check_oop(kls, &null_ctl, never_see_null); 3317 if (region != NULL) { 3318 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3319 region->init_req(null_path, null_ctl); 3320 } else { 3321 assert(null_ctl == top(), "no loose ends"); 3322 } 3323 return kls; 3324 } 3325 3326 //--------------------(inline_native_Class_query helpers)--------------------- 3327 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3328 // Fall through if (mods & mask) == bits, take the guard otherwise. 3329 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3330 // Branch around if the given klass has the given modifier bit set. 3331 // Like generate_guard, adds a new path onto the region. 3332 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3333 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3334 Node* mask = intcon(modifier_mask); 3335 Node* bits = intcon(modifier_bits); 3336 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3337 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3338 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3339 return generate_fair_guard(bol, region); 3340 } 3341 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3342 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3343 } 3344 3345 Node* LibraryCallKit::generate_value_guard(Node* kls, RegionNode* region) { 3346 return generate_access_flags_guard(kls, JVM_ACC_VALUE, 0, region); 3347 } 3348 3349 //-------------------------inline_native_Class_query------------------- 3350 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3351 const Type* return_type = TypeInt::BOOL; 3352 Node* prim_return_value = top(); // what happens if it's a primitive class? 3353 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3354 bool expect_prim = false; // most of these guys expect to work on refs 3355 3356 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3357 3358 Node* mirror = argument(0); 3359 Node* obj = top(); 3360 3361 switch (id) { 3362 case vmIntrinsics::_isInstance: 3363 // nothing is an instance of a primitive type 3364 prim_return_value = intcon(0); 3365 obj = argument(1); 3366 break; 3367 case vmIntrinsics::_getModifiers: 3368 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3369 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3370 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3371 break; 3372 case vmIntrinsics::_isInterface: 3373 prim_return_value = intcon(0); 3374 break; 3375 case vmIntrinsics::_isArray: 3376 prim_return_value = intcon(0); 3377 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3378 break; 3379 case vmIntrinsics::_isPrimitive: 3380 prim_return_value = intcon(1); 3381 expect_prim = true; // obviously 3382 break; 3383 case vmIntrinsics::_getSuperclass: 3384 prim_return_value = null(); 3385 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3386 break; 3387 case vmIntrinsics::_getClassAccessFlags: 3388 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3389 return_type = TypeInt::INT; // not bool! 6297094 3390 break; 3391 default: 3392 fatal_unexpected_iid(id); 3393 break; 3394 } 3395 3396 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3397 if (mirror_con == NULL) return false; // cannot happen? 3398 3399 #ifndef PRODUCT 3400 if (C->print_intrinsics() || C->print_inlining()) { 3401 ciType* k = mirror_con->java_mirror_type(); 3402 if (k) { 3403 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3404 k->print_name(); 3405 tty->cr(); 3406 } 3407 } 3408 #endif 3409 3410 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3411 RegionNode* region = new RegionNode(PATH_LIMIT); 3412 record_for_igvn(region); 3413 PhiNode* phi = new PhiNode(region, return_type); 3414 3415 // The mirror will never be null of Reflection.getClassAccessFlags, however 3416 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3417 // if it is. See bug 4774291. 3418 3419 // For Reflection.getClassAccessFlags(), the null check occurs in 3420 // the wrong place; see inline_unsafe_access(), above, for a similar 3421 // situation. 3422 mirror = null_check(mirror); 3423 // If mirror or obj is dead, only null-path is taken. 3424 if (stopped()) return true; 3425 3426 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3427 3428 // Now load the mirror's klass metaobject, and null-check it. 3429 // Side-effects region with the control path if the klass is null. 3430 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3431 // If kls is null, we have a primitive mirror. 3432 phi->init_req(_prim_path, prim_return_value); 3433 if (stopped()) { set_result(region, phi); return true; } 3434 bool safe_for_replace = (region->in(_prim_path) == top()); 3435 3436 Node* p; // handy temp 3437 Node* null_ctl; 3438 3439 // Now that we have the non-null klass, we can perform the real query. 3440 // For constant classes, the query will constant-fold in LoadNode::Value. 3441 Node* query_value = top(); 3442 switch (id) { 3443 case vmIntrinsics::_isInstance: 3444 // nothing is an instance of a primitive type 3445 query_value = gen_instanceof(obj, kls, safe_for_replace); 3446 break; 3447 3448 case vmIntrinsics::_getModifiers: 3449 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3450 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3451 break; 3452 3453 case vmIntrinsics::_isInterface: 3454 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3455 if (generate_interface_guard(kls, region) != NULL) 3456 // A guard was added. If the guard is taken, it was an interface. 3457 phi->add_req(intcon(1)); 3458 // If we fall through, it's a plain class. 3459 query_value = intcon(0); 3460 break; 3461 3462 case vmIntrinsics::_isArray: 3463 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3464 if (generate_array_guard(kls, region) != NULL) 3465 // A guard was added. If the guard is taken, it was an array. 3466 phi->add_req(intcon(1)); 3467 // If we fall through, it's a plain class. 3468 query_value = intcon(0); 3469 break; 3470 3471 case vmIntrinsics::_isPrimitive: 3472 query_value = intcon(0); // "normal" path produces false 3473 break; 3474 3475 case vmIntrinsics::_getSuperclass: 3476 // The rules here are somewhat unfortunate, but we can still do better 3477 // with random logic than with a JNI call. 3478 // Interfaces store null or Object as _super, but must report null. 3479 // Arrays store an intermediate super as _super, but must report Object. 3480 // Other types can report the actual _super. 3481 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3482 if (generate_interface_guard(kls, region) != NULL) 3483 // A guard was added. If the guard is taken, it was an interface. 3484 phi->add_req(null()); 3485 if (generate_array_guard(kls, region) != NULL) 3486 // A guard was added. If the guard is taken, it was an array. 3487 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3488 // If we fall through, it's a plain class. Get its _super. 3489 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3490 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3491 null_ctl = top(); 3492 kls = null_check_oop(kls, &null_ctl); 3493 if (null_ctl != top()) { 3494 // If the guard is taken, Object.superClass is null (both klass and mirror). 3495 region->add_req(null_ctl); 3496 phi ->add_req(null()); 3497 } 3498 if (!stopped()) { 3499 query_value = load_mirror_from_klass(kls); 3500 } 3501 break; 3502 3503 case vmIntrinsics::_getClassAccessFlags: 3504 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3505 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3506 break; 3507 3508 default: 3509 fatal_unexpected_iid(id); 3510 break; 3511 } 3512 3513 // Fall-through is the normal case of a query to a real class. 3514 phi->init_req(1, query_value); 3515 region->init_req(1, control()); 3516 3517 C->set_has_split_ifs(true); // Has chance for split-if optimization 3518 set_result(region, phi); 3519 return true; 3520 } 3521 3522 //-------------------------inline_value_Class_conversion------------------- 3523 // public Class<T> java.lang.Class.asPrimaryType(); 3524 // public Class<T> java.lang.Class.asIndirectType() 3525 bool LibraryCallKit::inline_value_Class_conversion(vmIntrinsics::ID id) { 3526 Node* mirror = argument(0); // Receiver Class 3527 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3528 if (mirror_con == NULL) { 3529 return false; 3530 } 3531 3532 bool is_indirect_type = true; 3533 ciType* tm = mirror_con->java_mirror_type(&is_indirect_type); 3534 if (tm != NULL) { 3535 Node* result = mirror; 3536 if (tm->is_valuetype()) { 3537 if (id == vmIntrinsics::_asPrimaryType && is_indirect_type) { 3538 result = _gvn.makecon(TypeInstPtr::make(tm->as_value_klass()->inline_mirror_instance())); 3539 } else if (id == vmIntrinsics::_asIndirectType && !is_indirect_type) { 3540 result = _gvn.makecon(TypeInstPtr::make(tm->as_value_klass()->indirect_mirror_instance())); 3541 } 3542 } 3543 set_result(result); 3544 return true; 3545 } 3546 return false; 3547 } 3548 3549 //-------------------------inline_Class_cast------------------- 3550 bool LibraryCallKit::inline_Class_cast() { 3551 Node* mirror = argument(0); // Class 3552 Node* obj = argument(1); 3553 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3554 if (mirror_con == NULL) { 3555 return false; // dead path (mirror->is_top()). 3556 } 3557 if (obj == NULL || obj->is_top()) { 3558 return false; // dead path 3559 } 3560 3561 ciKlass* obj_klass = NULL; 3562 if (obj->is_ValueType()) { 3563 obj_klass = _gvn.type(obj)->value_klass(); 3564 } else { 3565 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3566 if (tp != NULL) { 3567 obj_klass = tp->klass(); 3568 } 3569 } 3570 3571 // First, see if Class.cast() can be folded statically. 3572 // java_mirror_type() returns non-null for compile-time Class constants. 3573 bool is_indirect_type = true; 3574 ciType* tm = mirror_con->java_mirror_type(&is_indirect_type); 3575 if (!obj->is_ValueType() && !is_indirect_type) { 3576 obj = null_check(obj); 3577 if (stopped()) { 3578 return true; 3579 } 3580 } 3581 if (tm != NULL && tm->is_klass() && obj_klass != NULL) { 3582 if (!obj_klass->is_loaded()) { 3583 // Don't use intrinsic when class is not loaded. 3584 return false; 3585 } else { 3586 int static_res = C->static_subtype_check(tm->as_klass(), obj_klass); 3587 if (static_res == Compile::SSC_always_true) { 3588 // isInstance() is true - fold the code. 3589 set_result(obj); 3590 return true; 3591 } else if (static_res == Compile::SSC_always_false) { 3592 // Don't use intrinsic, have to throw ClassCastException. 3593 // If the reference is null, the non-intrinsic bytecode will 3594 // be optimized appropriately. 3595 return false; 3596 } 3597 } 3598 } 3599 3600 // Bailout intrinsic and do normal inlining if exception path is frequent. 3601 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3602 return false; 3603 } 3604 3605 // Generate dynamic checks. 3606 // Class.cast() is java implementation of _checkcast bytecode. 3607 // Do checkcast (Parse::do_checkcast()) optimizations here. 3608 3609 mirror = null_check(mirror); 3610 // If mirror is dead, only null-path is taken. 3611 if (stopped()) { 3612 return true; 3613 } 3614 3615 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3616 enum { _bad_type_path = 1, _prim_path = 2, _npe_path = 3, PATH_LIMIT }; 3617 RegionNode* region = new RegionNode(PATH_LIMIT); 3618 record_for_igvn(region); 3619 3620 // Now load the mirror's klass metaobject, and null-check it. 3621 // If kls is null, we have a primitive mirror and 3622 // nothing is an instance of a primitive type. 3623 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3624 3625 Node* res = top(); 3626 if (!stopped()) { 3627 if (EnableValhalla && !obj->is_ValueType() && is_indirect_type) { 3628 // Check if (mirror == inline_mirror && obj == null) 3629 Node* is_val_mirror = generate_fair_guard(is_value_mirror(mirror), NULL); 3630 if (is_val_mirror != NULL) { 3631 RegionNode* r = new RegionNode(3); 3632 record_for_igvn(r); 3633 r->init_req(1, control()); 3634 3635 // Casting to .val, check for null 3636 set_control(is_val_mirror); 3637 Node *null_ctr = top(); 3638 null_check_oop(obj, &null_ctr); 3639 region->init_req(_npe_path, null_ctr); 3640 r->init_req(2, control()); 3641 3642 set_control(_gvn.transform(r)); 3643 } 3644 } 3645 3646 Node* bad_type_ctrl = top(); 3647 // Do checkcast optimizations. 3648 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3649 region->init_req(_bad_type_path, bad_type_ctrl); 3650 } 3651 if (region->in(_prim_path) != top() || 3652 region->in(_bad_type_path) != top() || 3653 region->in(_npe_path) != top()) { 3654 // Let Interpreter throw ClassCastException. 3655 PreserveJVMState pjvms(this); 3656 set_control(_gvn.transform(region)); 3657 uncommon_trap(Deoptimization::Reason_intrinsic, 3658 Deoptimization::Action_maybe_recompile); 3659 } 3660 if (!stopped()) { 3661 set_result(res); 3662 } 3663 return true; 3664 } 3665 3666 3667 //--------------------------inline_native_subtype_check------------------------ 3668 // This intrinsic takes the JNI calls out of the heart of 3669 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3670 bool LibraryCallKit::inline_native_subtype_check() { 3671 // Pull both arguments off the stack. 3672 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3673 args[0] = argument(0); 3674 args[1] = argument(1); 3675 Node* klasses[2]; // corresponding Klasses: superk, subk 3676 klasses[0] = klasses[1] = top(); 3677 3678 enum { 3679 // A full decision tree on {superc is prim, subc is prim}: 3680 _prim_0_path = 1, // {P,N} => false 3681 // {P,P} & superc!=subc => false 3682 _prim_same_path, // {P,P} & superc==subc => true 3683 _prim_1_path, // {N,P} => false 3684 _ref_subtype_path, // {N,N} & subtype check wins => true 3685 _both_ref_path, // {N,N} & subtype check loses => false 3686 PATH_LIMIT 3687 }; 3688 3689 RegionNode* region = new RegionNode(PATH_LIMIT); 3690 RegionNode* prim_region = new RegionNode(2); 3691 Node* phi = new PhiNode(region, TypeInt::BOOL); 3692 record_for_igvn(region); 3693 record_for_igvn(prim_region); 3694 3695 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3696 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3697 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3698 3699 // First null-check both mirrors and load each mirror's klass metaobject. 3700 int which_arg; 3701 for (which_arg = 0; which_arg <= 1; which_arg++) { 3702 Node* arg = args[which_arg]; 3703 arg = null_check(arg); 3704 if (stopped()) break; 3705 args[which_arg] = arg; 3706 3707 Node* p = basic_plus_adr(arg, class_klass_offset); 3708 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3709 klasses[which_arg] = _gvn.transform(kls); 3710 } 3711 3712 // Resolve oops to stable for CmpP below. 3713 args[0] = access_resolve(args[0], 0); 3714 args[1] = access_resolve(args[1], 0); 3715 3716 // Having loaded both klasses, test each for null. 3717 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3718 for (which_arg = 0; which_arg <= 1; which_arg++) { 3719 Node* kls = klasses[which_arg]; 3720 Node* null_ctl = top(); 3721 kls = null_check_oop(kls, &null_ctl, never_see_null); 3722 if (which_arg == 0) { 3723 prim_region->init_req(1, null_ctl); 3724 } else { 3725 region->init_req(_prim_1_path, null_ctl); 3726 } 3727 if (stopped()) break; 3728 klasses[which_arg] = kls; 3729 } 3730 3731 if (!stopped()) { 3732 // now we have two reference types, in klasses[0..1] 3733 Node* subk = klasses[1]; // the argument to isAssignableFrom 3734 Node* superk = klasses[0]; // the receiver 3735 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3736 // If superc is a value mirror, we also need to check if superc == subc because 3737 // V? is not a subtype of V but due to subk == superk the subtype check will pass. 3738 generate_fair_guard(is_value_mirror(args[0]), prim_region); 3739 // now we have a successful reference subtype check 3740 region->set_req(_ref_subtype_path, control()); 3741 } 3742 3743 // If both operands are primitive (both klasses null), then 3744 // we must return true when they are identical primitives. 3745 // It is convenient to test this after the first null klass check. 3746 // This path is also used if superc is a value mirror. 3747 set_control(_gvn.transform(prim_region)); 3748 if (!stopped()) { 3749 // Since superc is primitive, make a guard for the superc==subc case. 3750 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3751 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3752 generate_fair_guard(bol_eq, region); 3753 if (region->req() == PATH_LIMIT+1) { 3754 // A guard was added. If the added guard is taken, superc==subc. 3755 region->swap_edges(PATH_LIMIT, _prim_same_path); 3756 region->del_req(PATH_LIMIT); 3757 } 3758 region->set_req(_prim_0_path, control()); // Not equal after all. 3759 } 3760 3761 // these are the only paths that produce 'true': 3762 phi->set_req(_prim_same_path, intcon(1)); 3763 phi->set_req(_ref_subtype_path, intcon(1)); 3764 3765 // pull together the cases: 3766 assert(region->req() == PATH_LIMIT, "sane region"); 3767 for (uint i = 1; i < region->req(); i++) { 3768 Node* ctl = region->in(i); 3769 if (ctl == NULL || ctl == top()) { 3770 region->set_req(i, top()); 3771 phi ->set_req(i, top()); 3772 } else if (phi->in(i) == NULL) { 3773 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3774 } 3775 } 3776 3777 set_control(_gvn.transform(region)); 3778 set_result(_gvn.transform(phi)); 3779 return true; 3780 } 3781 3782 //---------------------generate_array_guard_common------------------------ 3783 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, ArrayKind kind) { 3784 3785 if (stopped()) { 3786 return NULL; 3787 } 3788 3789 // Like generate_guard, adds a new path onto the region. 3790 jint layout_con = 0; 3791 Node* layout_val = get_layout_helper(kls, layout_con); 3792 if (layout_val == NULL) { 3793 bool query = 0; 3794 switch(kind) { 3795 case ObjectArray: query = Klass::layout_helper_is_objArray(layout_con); break; 3796 case NonObjectArray: query = !Klass::layout_helper_is_objArray(layout_con); break; 3797 case TypeArray: query = Klass::layout_helper_is_typeArray(layout_con); break; 3798 case ValueArray: query = Klass::layout_helper_is_valueArray(layout_con); break; 3799 case AnyArray: query = Klass::layout_helper_is_array(layout_con); break; 3800 case NonArray: query = !Klass::layout_helper_is_array(layout_con); break; 3801 default: 3802 ShouldNotReachHere(); 3803 } 3804 if (!query) { 3805 return NULL; // never a branch 3806 } else { // always a branch 3807 Node* always_branch = control(); 3808 if (region != NULL) 3809 region->add_req(always_branch); 3810 set_control(top()); 3811 return always_branch; 3812 } 3813 } 3814 unsigned int value = 0; 3815 BoolTest::mask btest = BoolTest::illegal; 3816 switch(kind) { 3817 case ObjectArray: 3818 case NonObjectArray: { 3819 value = Klass::_lh_array_tag_obj_value; 3820 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3821 btest = kind == ObjectArray ? BoolTest::eq : BoolTest::ne; 3822 break; 3823 } 3824 case TypeArray: { 3825 value = Klass::_lh_array_tag_type_value; 3826 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3827 btest = BoolTest::eq; 3828 break; 3829 } 3830 case ValueArray: { 3831 value = Klass::_lh_array_tag_vt_value; 3832 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3833 btest = BoolTest::eq; 3834 break; 3835 } 3836 case AnyArray: value = Klass::_lh_neutral_value; btest = BoolTest::lt; break; 3837 case NonArray: value = Klass::_lh_neutral_value; btest = BoolTest::gt; break; 3838 default: 3839 ShouldNotReachHere(); 3840 } 3841 // Now test the correct condition. 3842 jint nval = (jint)value; 3843 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3844 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3845 return generate_fair_guard(bol, region); 3846 } 3847 3848 3849 //-----------------------inline_native_newArray-------------------------- 3850 // private static native Object java.lang.reflect.Array.newArray(Class<?> componentType, int length); 3851 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 3852 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 3853 Node* mirror; 3854 Node* count_val; 3855 if (uninitialized) { 3856 mirror = argument(1); 3857 count_val = argument(2); 3858 } else { 3859 mirror = argument(0); 3860 count_val = argument(1); 3861 } 3862 3863 mirror = null_check(mirror); 3864 // If mirror or obj is dead, only null-path is taken. 3865 if (stopped()) return true; 3866 3867 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3868 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3869 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3870 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3871 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3872 3873 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3874 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3875 result_reg, _slow_path); 3876 Node* normal_ctl = control(); 3877 Node* no_array_ctl = result_reg->in(_slow_path); 3878 3879 // Generate code for the slow case. We make a call to newArray(). 3880 set_control(no_array_ctl); 3881 if (!stopped()) { 3882 // Either the input type is void.class, or else the 3883 // array klass has not yet been cached. Either the 3884 // ensuing call will throw an exception, or else it 3885 // will cache the array klass for next time. 3886 PreserveJVMState pjvms(this); 3887 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3888 Node* slow_result = set_results_for_java_call(slow_call); 3889 // this->control() comes from set_results_for_java_call 3890 result_reg->set_req(_slow_path, control()); 3891 result_val->set_req(_slow_path, slow_result); 3892 result_io ->set_req(_slow_path, i_o()); 3893 result_mem->set_req(_slow_path, reset_memory()); 3894 } 3895 3896 set_control(normal_ctl); 3897 if (!stopped()) { 3898 // Normal case: The array type has been cached in the java.lang.Class. 3899 // The following call works fine even if the array type is polymorphic. 3900 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3901 Node* obj = new_array(klass_node, count_val, 0, NULL, false, mirror); // no arguments to push 3902 result_reg->init_req(_normal_path, control()); 3903 result_val->init_req(_normal_path, obj); 3904 result_io ->init_req(_normal_path, i_o()); 3905 result_mem->init_req(_normal_path, reset_memory()); 3906 3907 if (uninitialized) { 3908 // Mark the allocation so that zeroing is skipped 3909 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 3910 alloc->maybe_set_complete(&_gvn); 3911 } 3912 } 3913 3914 // Return the combined state. 3915 set_i_o( _gvn.transform(result_io) ); 3916 set_all_memory( _gvn.transform(result_mem)); 3917 3918 C->set_has_split_ifs(true); // Has chance for split-if optimization 3919 set_result(result_reg, result_val); 3920 return true; 3921 } 3922 3923 //----------------------inline_native_getLength-------------------------- 3924 // public static native int java.lang.reflect.Array.getLength(Object array); 3925 bool LibraryCallKit::inline_native_getLength() { 3926 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3927 3928 Node* array = null_check(argument(0)); 3929 // If array is dead, only null-path is taken. 3930 if (stopped()) return true; 3931 3932 // Deoptimize if it is a non-array. 3933 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3934 3935 if (non_array != NULL) { 3936 PreserveJVMState pjvms(this); 3937 set_control(non_array); 3938 uncommon_trap(Deoptimization::Reason_intrinsic, 3939 Deoptimization::Action_maybe_recompile); 3940 } 3941 3942 // If control is dead, only non-array-path is taken. 3943 if (stopped()) return true; 3944 3945 // The works fine even if the array type is polymorphic. 3946 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3947 Node* result = load_array_length(array); 3948 3949 C->set_has_split_ifs(true); // Has chance for split-if optimization 3950 set_result(result); 3951 return true; 3952 } 3953 3954 //------------------------inline_array_copyOf---------------------------- 3955 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3956 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3957 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3958 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3959 3960 // Get the arguments. 3961 Node* original = argument(0); 3962 Node* start = is_copyOfRange? argument(1): intcon(0); 3963 Node* end = is_copyOfRange? argument(2): argument(1); 3964 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3965 3966 const TypeAryPtr* original_t = _gvn.type(original)->isa_aryptr(); 3967 const TypeInstPtr* mirror_t = _gvn.type(array_type_mirror)->isa_instptr(); 3968 if (EnableValhalla && ValueArrayFlatten && 3969 (original_t == NULL || mirror_t == NULL || 3970 (mirror_t->java_mirror_type() == NULL && 3971 (original_t->elem()->isa_valuetype() || 3972 (original_t->elem()->make_oopptr() != NULL && 3973 original_t->elem()->make_oopptr()->can_be_value_type()))))) { 3974 // We need to know statically if the copy is to a flattened array 3975 // or not but can't tell. 3976 return false; 3977 } 3978 3979 Node* newcopy = NULL; 3980 3981 // Set the original stack and the reexecute bit for the interpreter to reexecute 3982 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3983 { PreserveReexecuteState preexecs(this); 3984 jvms()->set_should_reexecute(true); 3985 3986 array_type_mirror = null_check(array_type_mirror); 3987 original = null_check(original); 3988 3989 // Check if a null path was taken unconditionally. 3990 if (stopped()) return true; 3991 3992 Node* orig_length = load_array_length(original); 3993 3994 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3995 klass_node = null_check(klass_node); 3996 3997 RegionNode* bailout = new RegionNode(1); 3998 record_for_igvn(bailout); 3999 4000 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 4001 // Bail out if that is so. 4002 // Value type array may have object field that would require a 4003 // write barrier. Conservatively, go to slow path. 4004 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4005 Node* not_objArray = !bs->array_copy_requires_gc_barriers(false, T_OBJECT, false, BarrierSetC2::Parsing) ? 4006 generate_typeArray_guard(klass_node, bailout) : generate_non_objArray_guard(klass_node, bailout); 4007 if (not_objArray != NULL) { 4008 // Improve the klass node's type from the new optimistic assumption: 4009 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 4010 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0)); 4011 Node* cast = new CastPPNode(klass_node, akls); 4012 cast->init_req(0, control()); 4013 klass_node = _gvn.transform(cast); 4014 } 4015 4016 Node* original_kls = load_object_klass(original); 4017 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 4018 // loads/stores but it is legal only if we're sure the 4019 // Arrays.copyOf would succeed. So we need all input arguments 4020 // to the copyOf to be validated, including that the copy to the 4021 // new array won't trigger an ArrayStoreException. That subtype 4022 // check can be optimized if we know something on the type of 4023 // the input array from type speculation. 4024 if (_gvn.type(klass_node)->singleton() && !stopped()) { 4025 ciKlass* subk = _gvn.type(original_kls)->is_klassptr()->klass(); 4026 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 4027 4028 int test = C->static_subtype_check(superk, subk); 4029 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 4030 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 4031 if (t_original->speculative_type() != NULL) { 4032 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 4033 original_kls = load_object_klass(original); 4034 } 4035 } 4036 } 4037 4038 if (ValueArrayFlatten) { 4039 // Either both or neither new array klass and original array 4040 // klass must be flattened 4041 Node* is_flat = generate_valueArray_guard(klass_node, NULL); 4042 if (!original_t->is_not_flat()) { 4043 generate_valueArray_guard(original_kls, bailout); 4044 } 4045 if (is_flat != NULL) { 4046 RegionNode* r = new RegionNode(2); 4047 record_for_igvn(r); 4048 r->init_req(1, control()); 4049 set_control(is_flat); 4050 if (!original_t->is_not_flat()) { 4051 generate_valueArray_guard(original_kls, r); 4052 } 4053 bailout->add_req(control()); 4054 set_control(_gvn.transform(r)); 4055 } 4056 } 4057 4058 // Bail out if either start or end is negative. 4059 generate_negative_guard(start, bailout, &start); 4060 generate_negative_guard(end, bailout, &end); 4061 4062 Node* length = end; 4063 if (_gvn.type(start) != TypeInt::ZERO) { 4064 length = _gvn.transform(new SubINode(end, start)); 4065 } 4066 4067 // Bail out if length is negative. 4068 // Without this the new_array would throw 4069 // NegativeArraySizeException but IllegalArgumentException is what 4070 // should be thrown 4071 generate_negative_guard(length, bailout, &length); 4072 4073 if (bailout->req() > 1) { 4074 PreserveJVMState pjvms(this); 4075 set_control(_gvn.transform(bailout)); 4076 uncommon_trap(Deoptimization::Reason_intrinsic, 4077 Deoptimization::Action_maybe_recompile); 4078 } 4079 4080 if (!stopped()) { 4081 // How many elements will we copy from the original? 4082 // The answer is MinI(orig_length - start, length). 4083 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 4084 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 4085 4086 original = access_resolve(original, ACCESS_READ); 4087 4088 // Generate a direct call to the right arraycopy function(s). 4089 // We know the copy is disjoint but we might not know if the 4090 // oop stores need checking. 4091 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 4092 // This will fail a store-check if x contains any non-nulls. 4093 4094 bool validated = false; 4095 // Reason_class_check rather than Reason_intrinsic because we 4096 // want to intrinsify even if this traps. 4097 if (!too_many_traps(Deoptimization::Reason_class_check)) { 4098 Node* not_subtype_ctrl = gen_subtype_check(original_kls, 4099 klass_node); 4100 4101 if (not_subtype_ctrl != top()) { 4102 PreserveJVMState pjvms(this); 4103 set_control(not_subtype_ctrl); 4104 uncommon_trap(Deoptimization::Reason_class_check, 4105 Deoptimization::Action_make_not_entrant); 4106 assert(stopped(), "Should be stopped"); 4107 } 4108 validated = true; 4109 } 4110 4111 if (!stopped()) { 4112 // Load element mirror 4113 Node* p = basic_plus_adr(array_type_mirror, java_lang_Class::component_mirror_offset_in_bytes()); 4114 Node* elem_mirror = access_load_at(array_type_mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR, T_OBJECT, IN_HEAP); 4115 4116 newcopy = new_array(klass_node, length, 0, NULL, false, elem_mirror); 4117 4118 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false, 4119 original_kls, klass_node); 4120 if (!is_copyOfRange) { 4121 ac->set_copyof(validated); 4122 } else { 4123 ac->set_copyofrange(validated); 4124 } 4125 Node* n = _gvn.transform(ac); 4126 if (n == ac) { 4127 ac->connect_outputs(this); 4128 } else { 4129 assert(validated, "shouldn't transform if all arguments not validated"); 4130 set_all_memory(n); 4131 } 4132 } 4133 } 4134 } // original reexecute is set back here 4135 4136 C->set_has_split_ifs(true); // Has chance for split-if optimization 4137 if (!stopped()) { 4138 set_result(newcopy); 4139 } 4140 return true; 4141 } 4142 4143 4144 //----------------------generate_virtual_guard--------------------------- 4145 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4146 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4147 RegionNode* slow_region) { 4148 ciMethod* method = callee(); 4149 int vtable_index = method->vtable_index(); 4150 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4151 "bad index %d", vtable_index); 4152 // Get the Method* out of the appropriate vtable entry. 4153 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 4154 vtable_index*vtableEntry::size_in_bytes() + 4155 vtableEntry::method_offset_in_bytes(); 4156 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4157 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4158 4159 // Compare the target method with the expected method (e.g., Object.hashCode). 4160 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4161 4162 Node* native_call = makecon(native_call_addr); 4163 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4164 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4165 4166 return generate_slow_guard(test_native, slow_region); 4167 } 4168 4169 //-----------------------generate_method_call---------------------------- 4170 // Use generate_method_call to make a slow-call to the real 4171 // method if the fast path fails. An alternative would be to 4172 // use a stub like OptoRuntime::slow_arraycopy_Java. 4173 // This only works for expanding the current library call, 4174 // not another intrinsic. (E.g., don't use this for making an 4175 // arraycopy call inside of the copyOf intrinsic.) 4176 CallJavaNode* 4177 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4178 // When compiling the intrinsic method itself, do not use this technique. 4179 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4180 4181 ciMethod* method = callee(); 4182 // ensure the JVMS we have will be correct for this call 4183 guarantee(method_id == method->intrinsic_id(), "must match"); 4184 4185 const TypeFunc* tf = TypeFunc::make(method); 4186 CallJavaNode* slow_call; 4187 if (is_static) { 4188 assert(!is_virtual, ""); 4189 slow_call = new CallStaticJavaNode(C, tf, 4190 SharedRuntime::get_resolve_static_call_stub(), 4191 method, bci()); 4192 } else if (is_virtual) { 4193 null_check_receiver(); 4194 int vtable_index = Method::invalid_vtable_index; 4195 if (UseInlineCaches) { 4196 // Suppress the vtable call 4197 } else { 4198 // hashCode and clone are not a miranda methods, 4199 // so the vtable index is fixed. 4200 // No need to use the linkResolver to get it. 4201 vtable_index = method->vtable_index(); 4202 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4203 "bad index %d", vtable_index); 4204 } 4205 slow_call = new CallDynamicJavaNode(tf, 4206 SharedRuntime::get_resolve_virtual_call_stub(), 4207 method, vtable_index, bci()); 4208 } else { // neither virtual nor static: opt_virtual 4209 null_check_receiver(); 4210 slow_call = new CallStaticJavaNode(C, tf, 4211 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4212 method, bci()); 4213 slow_call->set_optimized_virtual(true); 4214 } 4215 if (CallGenerator::is_inlined_method_handle_intrinsic(this->method(), bci(), callee())) { 4216 // To be able to issue a direct call (optimized virtual or virtual) 4217 // and skip a call to MH.linkTo*/invokeBasic adapter, additional information 4218 // about the method being invoked should be attached to the call site to 4219 // make resolution logic work (see SharedRuntime::resolve_{virtual,opt_virtual}_call_C). 4220 slow_call->set_override_symbolic_info(true); 4221 } 4222 set_arguments_for_java_call(slow_call); 4223 set_edges_for_java_call(slow_call); 4224 return slow_call; 4225 } 4226 4227 4228 /** 4229 * Build special case code for calls to hashCode on an object. This call may 4230 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4231 * slightly different code. 4232 */ 4233 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4234 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4235 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4236 4237 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4238 4239 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4240 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4241 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4242 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4243 Node* obj = argument(0); 4244 4245 if (obj->is_ValueType() || gvn().type(obj)->is_valuetypeptr()) { 4246 return false; 4247 } 4248 4249 if (!is_static) { 4250 // Check for hashing null object 4251 obj = null_check_receiver(); 4252 if (stopped()) return true; // unconditionally null 4253 result_reg->init_req(_null_path, top()); 4254 result_val->init_req(_null_path, top()); 4255 } else { 4256 // Do a null check, and return zero if null. 4257 // System.identityHashCode(null) == 0 4258 Node* null_ctl = top(); 4259 obj = null_check_oop(obj, &null_ctl); 4260 result_reg->init_req(_null_path, null_ctl); 4261 result_val->init_req(_null_path, _gvn.intcon(0)); 4262 } 4263 4264 // Unconditionally null? Then return right away. 4265 if (stopped()) { 4266 set_control( result_reg->in(_null_path)); 4267 if (!stopped()) 4268 set_result(result_val->in(_null_path)); 4269 return true; 4270 } 4271 4272 // We only go to the fast case code if we pass a number of guards. The 4273 // paths which do not pass are accumulated in the slow_region. 4274 RegionNode* slow_region = new RegionNode(1); 4275 record_for_igvn(slow_region); 4276 4277 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4278 assert(!obj_type->isa_valuetype() || !obj_type->is_valuetypeptr(), "no value type here"); 4279 if (is_static && obj_type->can_be_value_type()) { 4280 Node* obj_klass = load_object_klass(obj); 4281 generate_value_guard(obj_klass, slow_region); 4282 } 4283 4284 // If this is a virtual call, we generate a funny guard. We pull out 4285 // the vtable entry corresponding to hashCode() from the target object. 4286 // If the target method which we are calling happens to be the native 4287 // Object hashCode() method, we pass the guard. We do not need this 4288 // guard for non-virtual calls -- the caller is known to be the native 4289 // Object hashCode(). 4290 if (is_virtual) { 4291 // After null check, get the object's klass. 4292 Node* obj_klass = load_object_klass(obj); 4293 generate_virtual_guard(obj_klass, slow_region); 4294 } 4295 4296 // Get the header out of the object, use LoadMarkNode when available 4297 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4298 // The control of the load must be NULL. Otherwise, the load can move before 4299 // the null check after castPP removal. 4300 Node* no_ctrl = NULL; 4301 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4302 4303 // Test the header to see if it is unlocked. 4304 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4305 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4306 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4307 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4308 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4309 4310 generate_slow_guard(test_unlocked, slow_region); 4311 4312 // Get the hash value and check to see that it has been properly assigned. 4313 // We depend on hash_mask being at most 32 bits and avoid the use of 4314 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4315 // vm: see markOop.hpp. 4316 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4317 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4318 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4319 // This hack lets the hash bits live anywhere in the mark object now, as long 4320 // as the shift drops the relevant bits into the low 32 bits. Note that 4321 // Java spec says that HashCode is an int so there's no point in capturing 4322 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4323 hshifted_header = ConvX2I(hshifted_header); 4324 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4325 4326 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4327 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4328 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4329 4330 generate_slow_guard(test_assigned, slow_region); 4331 4332 Node* init_mem = reset_memory(); 4333 // fill in the rest of the null path: 4334 result_io ->init_req(_null_path, i_o()); 4335 result_mem->init_req(_null_path, init_mem); 4336 4337 result_val->init_req(_fast_path, hash_val); 4338 result_reg->init_req(_fast_path, control()); 4339 result_io ->init_req(_fast_path, i_o()); 4340 result_mem->init_req(_fast_path, init_mem); 4341 4342 // Generate code for the slow case. We make a call to hashCode(). 4343 set_control(_gvn.transform(slow_region)); 4344 if (!stopped()) { 4345 // No need for PreserveJVMState, because we're using up the present state. 4346 set_all_memory(init_mem); 4347 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4348 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4349 Node* slow_result = set_results_for_java_call(slow_call); 4350 // this->control() comes from set_results_for_java_call 4351 result_reg->init_req(_slow_path, control()); 4352 result_val->init_req(_slow_path, slow_result); 4353 result_io ->set_req(_slow_path, i_o()); 4354 result_mem ->set_req(_slow_path, reset_memory()); 4355 } 4356 4357 // Return the combined state. 4358 set_i_o( _gvn.transform(result_io) ); 4359 set_all_memory( _gvn.transform(result_mem)); 4360 4361 set_result(result_reg, result_val); 4362 return true; 4363 } 4364 4365 //---------------------------inline_native_getClass---------------------------- 4366 // public final native Class<?> java.lang.Object.getClass(); 4367 // 4368 // Build special case code for calls to getClass on an object. 4369 bool LibraryCallKit::inline_native_getClass() { 4370 Node* obj = argument(0); 4371 if (obj->is_ValueType()) { 4372 ciKlass* vk = _gvn.type(obj)->value_klass(); 4373 set_result(makecon(TypeInstPtr::make(vk->java_mirror()))); 4374 return true; 4375 } 4376 obj = null_check_receiver(); 4377 if (stopped()) return true; 4378 set_result(load_mirror_from_klass(load_object_klass(obj))); 4379 return true; 4380 } 4381 4382 //-----------------inline_native_Reflection_getCallerClass--------------------- 4383 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4384 // 4385 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4386 // 4387 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4388 // in that it must skip particular security frames and checks for 4389 // caller sensitive methods. 4390 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4391 #ifndef PRODUCT 4392 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4393 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4394 } 4395 #endif 4396 4397 if (!jvms()->has_method()) { 4398 #ifndef PRODUCT 4399 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4400 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4401 } 4402 #endif 4403 return false; 4404 } 4405 4406 // Walk back up the JVM state to find the caller at the required 4407 // depth. 4408 JVMState* caller_jvms = jvms(); 4409 4410 // Cf. JVM_GetCallerClass 4411 // NOTE: Start the loop at depth 1 because the current JVM state does 4412 // not include the Reflection.getCallerClass() frame. 4413 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4414 ciMethod* m = caller_jvms->method(); 4415 switch (n) { 4416 case 0: 4417 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4418 break; 4419 case 1: 4420 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4421 if (!m->caller_sensitive()) { 4422 #ifndef PRODUCT 4423 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4424 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4425 } 4426 #endif 4427 return false; // bail-out; let JVM_GetCallerClass do the work 4428 } 4429 break; 4430 default: 4431 if (!m->is_ignored_by_security_stack_walk()) { 4432 // We have reached the desired frame; return the holder class. 4433 // Acquire method holder as java.lang.Class and push as constant. 4434 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4435 ciInstance* caller_mirror = caller_klass->java_mirror(); 4436 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4437 4438 #ifndef PRODUCT 4439 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4440 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()); 4441 tty->print_cr(" JVM state at this point:"); 4442 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4443 ciMethod* m = jvms()->of_depth(i)->method(); 4444 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4445 } 4446 } 4447 #endif 4448 return true; 4449 } 4450 break; 4451 } 4452 } 4453 4454 #ifndef PRODUCT 4455 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4456 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4457 tty->print_cr(" JVM state at this point:"); 4458 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4459 ciMethod* m = jvms()->of_depth(i)->method(); 4460 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4461 } 4462 } 4463 #endif 4464 4465 return false; // bail-out; let JVM_GetCallerClass do the work 4466 } 4467 4468 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4469 Node* arg = argument(0); 4470 Node* result = NULL; 4471 4472 switch (id) { 4473 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4474 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4475 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4476 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4477 4478 case vmIntrinsics::_doubleToLongBits: { 4479 // two paths (plus control) merge in a wood 4480 RegionNode *r = new RegionNode(3); 4481 Node *phi = new PhiNode(r, TypeLong::LONG); 4482 4483 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4484 // Build the boolean node 4485 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4486 4487 // Branch either way. 4488 // NaN case is less traveled, which makes all the difference. 4489 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4490 Node *opt_isnan = _gvn.transform(ifisnan); 4491 assert( opt_isnan->is_If(), "Expect an IfNode"); 4492 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4493 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4494 4495 set_control(iftrue); 4496 4497 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4498 Node *slow_result = longcon(nan_bits); // return NaN 4499 phi->init_req(1, _gvn.transform( slow_result )); 4500 r->init_req(1, iftrue); 4501 4502 // Else fall through 4503 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4504 set_control(iffalse); 4505 4506 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4507 r->init_req(2, iffalse); 4508 4509 // Post merge 4510 set_control(_gvn.transform(r)); 4511 record_for_igvn(r); 4512 4513 C->set_has_split_ifs(true); // Has chance for split-if optimization 4514 result = phi; 4515 assert(result->bottom_type()->isa_long(), "must be"); 4516 break; 4517 } 4518 4519 case vmIntrinsics::_floatToIntBits: { 4520 // two paths (plus control) merge in a wood 4521 RegionNode *r = new RegionNode(3); 4522 Node *phi = new PhiNode(r, TypeInt::INT); 4523 4524 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4525 // Build the boolean node 4526 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4527 4528 // Branch either way. 4529 // NaN case is less traveled, which makes all the difference. 4530 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4531 Node *opt_isnan = _gvn.transform(ifisnan); 4532 assert( opt_isnan->is_If(), "Expect an IfNode"); 4533 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4534 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4535 4536 set_control(iftrue); 4537 4538 static const jint nan_bits = 0x7fc00000; 4539 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4540 phi->init_req(1, _gvn.transform( slow_result )); 4541 r->init_req(1, iftrue); 4542 4543 // Else fall through 4544 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4545 set_control(iffalse); 4546 4547 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4548 r->init_req(2, iffalse); 4549 4550 // Post merge 4551 set_control(_gvn.transform(r)); 4552 record_for_igvn(r); 4553 4554 C->set_has_split_ifs(true); // Has chance for split-if optimization 4555 result = phi; 4556 assert(result->bottom_type()->isa_int(), "must be"); 4557 break; 4558 } 4559 4560 default: 4561 fatal_unexpected_iid(id); 4562 break; 4563 } 4564 set_result(_gvn.transform(result)); 4565 return true; 4566 } 4567 4568 //----------------------inline_unsafe_copyMemory------------------------- 4569 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4570 bool LibraryCallKit::inline_unsafe_copyMemory() { 4571 if (callee()->is_static()) return false; // caller must have the capability! 4572 null_check_receiver(); // null-check receiver 4573 if (stopped()) return true; 4574 4575 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4576 4577 Node* src_ptr = argument(1); // type: oop 4578 Node* src_off = ConvL2X(argument(2)); // type: long 4579 Node* dst_ptr = argument(4); // type: oop 4580 Node* dst_off = ConvL2X(argument(5)); // type: long 4581 Node* size = ConvL2X(argument(7)); // type: long 4582 4583 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4584 "fieldOffset must be byte-scaled"); 4585 4586 src_ptr = access_resolve(src_ptr, ACCESS_READ); 4587 dst_ptr = access_resolve(dst_ptr, ACCESS_WRITE); 4588 Node* src = make_unsafe_address(src_ptr, src_off, ACCESS_READ); 4589 Node* dst = make_unsafe_address(dst_ptr, dst_off, ACCESS_WRITE); 4590 4591 // Conservatively insert a memory barrier on all memory slices. 4592 // Do not let writes of the copy source or destination float below the copy. 4593 insert_mem_bar(Op_MemBarCPUOrder); 4594 4595 // Call it. Note that the length argument is not scaled. 4596 make_runtime_call(RC_LEAF|RC_NO_FP, 4597 OptoRuntime::fast_arraycopy_Type(), 4598 StubRoutines::unsafe_arraycopy(), 4599 "unsafe_arraycopy", 4600 TypeRawPtr::BOTTOM, 4601 src, dst, size XTOP); 4602 4603 // Do not let reads of the copy destination float above the copy. 4604 insert_mem_bar(Op_MemBarCPUOrder); 4605 4606 return true; 4607 } 4608 4609 //------------------------clone_coping----------------------------------- 4610 // Helper function for inline_native_clone. 4611 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) { 4612 assert(obj_size != NULL, ""); 4613 Node* raw_obj = alloc_obj->in(1); 4614 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4615 4616 AllocateNode* alloc = NULL; 4617 if (ReduceBulkZeroing) { 4618 // We will be completely responsible for initializing this object - 4619 // mark Initialize node as complete. 4620 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4621 // The object was just allocated - there should be no any stores! 4622 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4623 // Mark as complete_with_arraycopy so that on AllocateNode 4624 // expansion, we know this AllocateNode is initialized by an array 4625 // copy and a StoreStore barrier exists after the array copy. 4626 alloc->initialization()->set_complete_with_arraycopy(); 4627 } 4628 4629 // Copy the fastest available way. 4630 // TODO: generate fields copies for small objects instead. 4631 Node* size = _gvn.transform(obj_size); 4632 4633 // Exclude the header but include array length to copy by 8 bytes words. 4634 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4635 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4636 instanceOopDesc::base_offset_in_bytes(); 4637 // base_off: 4638 // 8 - 32-bit VM 4639 // 12 - 64-bit VM, compressed klass 4640 // 16 - 64-bit VM, normal klass 4641 if (base_off % BytesPerLong != 0) { 4642 assert(UseCompressedClassPointers, ""); 4643 if (is_array) { 4644 // Exclude length to copy by 8 bytes words. 4645 base_off += sizeof(int); 4646 } else { 4647 // Include klass to copy by 8 bytes words. 4648 base_off = instanceOopDesc::klass_offset_in_bytes(); 4649 } 4650 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4651 } 4652 Node* src_base = basic_plus_adr(obj, base_off); 4653 Node* dst_base = basic_plus_adr(alloc_obj, base_off); 4654 4655 // Compute the length also, if needed: 4656 Node* countx = size; 4657 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4658 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong))); 4659 4660 access_clone(src_base, dst_base, countx, is_array); 4661 4662 // Do not let reads from the cloned object float above the arraycopy. 4663 if (alloc != NULL) { 4664 // Do not let stores that initialize this object be reordered with 4665 // a subsequent store that would make this object accessible by 4666 // other threads. 4667 // Record what AllocateNode this StoreStore protects so that 4668 // escape analysis can go from the MemBarStoreStoreNode to the 4669 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4670 // based on the escape status of the AllocateNode. 4671 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 4672 } else { 4673 insert_mem_bar(Op_MemBarCPUOrder); 4674 } 4675 } 4676 4677 //------------------------inline_native_clone---------------------------- 4678 // protected native Object java.lang.Object.clone(); 4679 // 4680 // Here are the simple edge cases: 4681 // null receiver => normal trap 4682 // virtual and clone was overridden => slow path to out-of-line clone 4683 // not cloneable or finalizer => slow path to out-of-line Object.clone 4684 // 4685 // The general case has two steps, allocation and copying. 4686 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4687 // 4688 // Copying also has two cases, oop arrays and everything else. 4689 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4690 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4691 // 4692 // These steps fold up nicely if and when the cloned object's klass 4693 // can be sharply typed as an object array, a type array, or an instance. 4694 // 4695 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4696 PhiNode* result_val; 4697 4698 // Set the reexecute bit for the interpreter to reexecute 4699 // the bytecode that invokes Object.clone if deoptimization happens. 4700 { PreserveReexecuteState preexecs(this); 4701 jvms()->set_should_reexecute(true); 4702 4703 Node* obj = argument(0); 4704 if (obj->is_ValueType()) { 4705 return false; 4706 } 4707 4708 obj = null_check_receiver(); 4709 if (stopped()) return true; 4710 4711 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4712 4713 // If we are going to clone an instance, we need its exact type to 4714 // know the number and types of fields to convert the clone to 4715 // loads/stores. Maybe a speculative type can help us. 4716 if (!obj_type->klass_is_exact() && 4717 obj_type->speculative_type() != NULL && 4718 obj_type->speculative_type()->is_instance_klass() && 4719 !obj_type->speculative_type()->is_valuetype()) { 4720 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4721 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4722 !spec_ik->has_injected_fields()) { 4723 ciKlass* k = obj_type->klass(); 4724 if (!k->is_instance_klass() || 4725 k->as_instance_klass()->is_interface() || 4726 k->as_instance_klass()->has_subklass()) { 4727 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4728 } 4729 } 4730 } 4731 4732 Node* obj_klass = load_object_klass(obj); 4733 4734 // Conservatively insert a memory barrier on all memory slices. 4735 // Do not let writes into the original float below the clone. 4736 insert_mem_bar(Op_MemBarCPUOrder); 4737 4738 // paths into result_reg: 4739 enum { 4740 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4741 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4742 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4743 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4744 PATH_LIMIT 4745 }; 4746 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4747 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4748 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4749 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4750 record_for_igvn(result_reg); 4751 4752 // We only go to the fast case code if we pass a number of guards. 4753 // The paths which do not pass are accumulated in the slow_region. 4754 RegionNode* slow_region = new RegionNode(1); 4755 record_for_igvn(slow_region); 4756 4757 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4758 if (array_ctl != NULL) { 4759 // It's an array. 4760 PreserveJVMState pjvms(this); 4761 set_control(array_ctl); 4762 4763 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4764 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing) && 4765 (!obj_type->isa_aryptr() || !obj_type->is_aryptr()->is_not_flat())) { 4766 // Flattened value type array may have object field that would require a 4767 // write barrier. Conservatively, go to slow path. 4768 generate_valueArray_guard(obj_klass, slow_region); 4769 } 4770 4771 if (!stopped()) { 4772 Node* obj_length = load_array_length(obj); 4773 Node* obj_size = NULL; 4774 // Load element mirror 4775 Node* array_type_mirror = load_mirror_from_klass(obj_klass); 4776 Node* p = basic_plus_adr(array_type_mirror, java_lang_Class::component_mirror_offset_in_bytes()); 4777 Node* elem_mirror = access_load_at(array_type_mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR, T_OBJECT, IN_HEAP); 4778 4779 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size, false, elem_mirror); 4780 4781 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4782 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) { 4783 // If it is an oop array, it requires very special treatment, 4784 // because gc barriers are required when accessing the array. 4785 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4786 if (is_obja != NULL) { 4787 PreserveJVMState pjvms2(this); 4788 set_control(is_obja); 4789 // Generate a direct call to the right arraycopy function(s). 4790 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4791 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false); 4792 ac->set_cloneoop(); 4793 Node* n = _gvn.transform(ac); 4794 assert(n == ac, "cannot disappear"); 4795 ac->connect_outputs(this); 4796 4797 result_reg->init_req(_objArray_path, control()); 4798 result_val->init_req(_objArray_path, alloc_obj); 4799 result_i_o ->set_req(_objArray_path, i_o()); 4800 result_mem ->set_req(_objArray_path, reset_memory()); 4801 } 4802 } 4803 4804 // Otherwise, there are no barriers to worry about. 4805 // (We can dispense with card marks if we know the allocation 4806 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4807 // causes the non-eden paths to take compensating steps to 4808 // simulate a fresh allocation, so that no further 4809 // card marks are required in compiled code to initialize 4810 // the object.) 4811 4812 if (!stopped()) { 4813 copy_to_clone(obj, alloc_obj, obj_size, true); 4814 4815 // Present the results of the copy. 4816 result_reg->init_req(_array_path, control()); 4817 result_val->init_req(_array_path, alloc_obj); 4818 result_i_o ->set_req(_array_path, i_o()); 4819 result_mem ->set_req(_array_path, reset_memory()); 4820 } 4821 } 4822 } 4823 4824 if (!stopped()) { 4825 // It's an instance (we did array above). Make the slow-path tests. 4826 // If this is a virtual call, we generate a funny guard. We grab 4827 // the vtable entry corresponding to clone() from the target object. 4828 // If the target method which we are calling happens to be the 4829 // Object clone() method, we pass the guard. We do not need this 4830 // guard for non-virtual calls; the caller is known to be the native 4831 // Object clone(). 4832 if (is_virtual) { 4833 generate_virtual_guard(obj_klass, slow_region); 4834 } 4835 4836 // The object must be easily cloneable and must not have a finalizer. 4837 // Both of these conditions may be checked in a single test. 4838 // We could optimize the test further, but we don't care. 4839 generate_access_flags_guard(obj_klass, 4840 // Test both conditions: 4841 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4842 // Must be cloneable but not finalizer: 4843 JVM_ACC_IS_CLONEABLE_FAST, 4844 slow_region); 4845 } 4846 4847 if (!stopped()) { 4848 // It's an instance, and it passed the slow-path tests. 4849 PreserveJVMState pjvms(this); 4850 Node* obj_size = NULL; 4851 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4852 // is reexecuted if deoptimization occurs and there could be problems when merging 4853 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4854 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4855 4856 copy_to_clone(obj, alloc_obj, obj_size, false); 4857 4858 // Present the results of the slow call. 4859 result_reg->init_req(_instance_path, control()); 4860 result_val->init_req(_instance_path, alloc_obj); 4861 result_i_o ->set_req(_instance_path, i_o()); 4862 result_mem ->set_req(_instance_path, reset_memory()); 4863 } 4864 4865 // Generate code for the slow case. We make a call to clone(). 4866 set_control(_gvn.transform(slow_region)); 4867 if (!stopped()) { 4868 PreserveJVMState pjvms(this); 4869 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4870 // We need to deoptimize on exception (see comment above) 4871 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true); 4872 // this->control() comes from set_results_for_java_call 4873 result_reg->init_req(_slow_path, control()); 4874 result_val->init_req(_slow_path, slow_result); 4875 result_i_o ->set_req(_slow_path, i_o()); 4876 result_mem ->set_req(_slow_path, reset_memory()); 4877 } 4878 4879 // Return the combined state. 4880 set_control( _gvn.transform(result_reg)); 4881 set_i_o( _gvn.transform(result_i_o)); 4882 set_all_memory( _gvn.transform(result_mem)); 4883 } // original reexecute is set back here 4884 4885 set_result(_gvn.transform(result_val)); 4886 return true; 4887 } 4888 4889 // If we have a tightly coupled allocation, the arraycopy may take care 4890 // of the array initialization. If one of the guards we insert between 4891 // the allocation and the arraycopy causes a deoptimization, an 4892 // unitialized array will escape the compiled method. To prevent that 4893 // we set the JVM state for uncommon traps between the allocation and 4894 // the arraycopy to the state before the allocation so, in case of 4895 // deoptimization, we'll reexecute the allocation and the 4896 // initialization. 4897 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4898 if (alloc != NULL) { 4899 ciMethod* trap_method = alloc->jvms()->method(); 4900 int trap_bci = alloc->jvms()->bci(); 4901 4902 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 4903 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4904 // Make sure there's no store between the allocation and the 4905 // arraycopy otherwise visible side effects could be rexecuted 4906 // in case of deoptimization and cause incorrect execution. 4907 bool no_interfering_store = true; 4908 Node* mem = alloc->in(TypeFunc::Memory); 4909 if (mem->is_MergeMem()) { 4910 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4911 Node* n = mms.memory(); 4912 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4913 assert(n->is_Store(), "what else?"); 4914 no_interfering_store = false; 4915 break; 4916 } 4917 } 4918 } else { 4919 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4920 Node* n = mms.memory(); 4921 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4922 assert(n->is_Store(), "what else?"); 4923 no_interfering_store = false; 4924 break; 4925 } 4926 } 4927 } 4928 4929 if (no_interfering_store) { 4930 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4931 uint size = alloc->req(); 4932 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4933 old_jvms->set_map(sfpt); 4934 for (uint i = 0; i < size; i++) { 4935 sfpt->init_req(i, alloc->in(i)); 4936 } 4937 // re-push array length for deoptimization 4938 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4939 old_jvms->set_sp(old_jvms->sp()+1); 4940 old_jvms->set_monoff(old_jvms->monoff()+1); 4941 old_jvms->set_scloff(old_jvms->scloff()+1); 4942 old_jvms->set_endoff(old_jvms->endoff()+1); 4943 old_jvms->set_should_reexecute(true); 4944 4945 sfpt->set_i_o(map()->i_o()); 4946 sfpt->set_memory(map()->memory()); 4947 sfpt->set_control(map()->control()); 4948 4949 JVMState* saved_jvms = jvms(); 4950 saved_reexecute_sp = _reexecute_sp; 4951 4952 set_jvms(sfpt->jvms()); 4953 _reexecute_sp = jvms()->sp(); 4954 4955 return saved_jvms; 4956 } 4957 } 4958 } 4959 return NULL; 4960 } 4961 4962 // In case of a deoptimization, we restart execution at the 4963 // allocation, allocating a new array. We would leave an uninitialized 4964 // array in the heap that GCs wouldn't expect. Move the allocation 4965 // after the traps so we don't allocate the array if we 4966 // deoptimize. This is possible because tightly_coupled_allocation() 4967 // guarantees there's no observer of the allocated array at this point 4968 // and the control flow is simple enough. 4969 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, 4970 int saved_reexecute_sp, uint new_idx) { 4971 if (saved_jvms != NULL && !stopped()) { 4972 assert(alloc != NULL, "only with a tightly coupled allocation"); 4973 // restore JVM state to the state at the arraycopy 4974 saved_jvms->map()->set_control(map()->control()); 4975 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4976 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4977 // If we've improved the types of some nodes (null check) while 4978 // emitting the guards, propagate them to the current state 4979 map()->replaced_nodes().apply(saved_jvms->map(), new_idx); 4980 set_jvms(saved_jvms); 4981 _reexecute_sp = saved_reexecute_sp; 4982 4983 // Remove the allocation from above the guards 4984 CallProjections* callprojs = alloc->extract_projections(true); 4985 InitializeNode* init = alloc->initialization(); 4986 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4987 C->gvn_replace_by(callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4988 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4989 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4990 4991 // move the allocation here (after the guards) 4992 _gvn.hash_delete(alloc); 4993 alloc->set_req(TypeFunc::Control, control()); 4994 alloc->set_req(TypeFunc::I_O, i_o()); 4995 Node *mem = reset_memory(); 4996 set_all_memory(mem); 4997 alloc->set_req(TypeFunc::Memory, mem); 4998 set_control(init->proj_out_or_null(TypeFunc::Control)); 4999 set_i_o(callprojs->fallthrough_ioproj); 5000 5001 // Update memory as done in GraphKit::set_output_for_allocation() 5002 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 5003 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 5004 if (ary_type->isa_aryptr() && length_type != NULL) { 5005 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 5006 } 5007 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 5008 int elemidx = C->get_alias_index(telemref); 5009 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw); 5010 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx); 5011 5012 Node* allocx = _gvn.transform(alloc); 5013 assert(allocx == alloc, "where has the allocation gone?"); 5014 assert(dest->is_CheckCastPP(), "not an allocation result?"); 5015 5016 _gvn.hash_delete(dest); 5017 dest->set_req(0, control()); 5018 Node* destx = _gvn.transform(dest); 5019 assert(destx == dest, "where has the allocation result gone?"); 5020 } 5021 } 5022 5023 5024 //------------------------------inline_arraycopy----------------------- 5025 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 5026 // Object dest, int destPos, 5027 // int length); 5028 bool LibraryCallKit::inline_arraycopy() { 5029 // Get the arguments. 5030 Node* src = argument(0); // type: oop 5031 Node* src_offset = argument(1); // type: int 5032 Node* dest = argument(2); // type: oop 5033 Node* dest_offset = argument(3); // type: int 5034 Node* length = argument(4); // type: int 5035 5036 uint new_idx = C->unique(); 5037 5038 // Check for allocation before we add nodes that would confuse 5039 // tightly_coupled_allocation() 5040 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 5041 5042 int saved_reexecute_sp = -1; 5043 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 5044 // See arraycopy_restore_alloc_state() comment 5045 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 5046 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 5047 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards 5048 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 5049 5050 // The following tests must be performed 5051 // (1) src and dest are arrays. 5052 // (2) src and dest arrays must have elements of the same BasicType 5053 // (3) src and dest must not be null. 5054 // (4) src_offset must not be negative. 5055 // (5) dest_offset must not be negative. 5056 // (6) length must not be negative. 5057 // (7) src_offset + length must not exceed length of src. 5058 // (8) dest_offset + length must not exceed length of dest. 5059 // (9) each element of an oop array must be assignable 5060 5061 // (3) src and dest must not be null. 5062 // always do this here because we need the JVM state for uncommon traps 5063 Node* null_ctl = top(); 5064 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 5065 assert(null_ctl->is_top(), "no null control here"); 5066 dest = null_check(dest, T_ARRAY); 5067 5068 if (!can_emit_guards) { 5069 // if saved_jvms == NULL and alloc != NULL, we don't emit any 5070 // guards but the arraycopy node could still take advantage of a 5071 // tightly allocated allocation. tightly_coupled_allocation() is 5072 // called again to make sure it takes the null check above into 5073 // account: the null check is mandatory and if it caused an 5074 // uncommon trap to be emitted then the allocation can't be 5075 // considered tightly coupled in this context. 5076 alloc = tightly_coupled_allocation(dest, NULL); 5077 } 5078 5079 bool validated = false; 5080 5081 const Type* src_type = _gvn.type(src); 5082 const Type* dest_type = _gvn.type(dest); 5083 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5084 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5085 5086 // Do we have the type of src? 5087 bool has_src = (top_src != NULL && top_src->klass() != NULL); 5088 // Do we have the type of dest? 5089 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 5090 // Is the type for src from speculation? 5091 bool src_spec = false; 5092 // Is the type for dest from speculation? 5093 bool dest_spec = false; 5094 5095 if ((!has_src || !has_dest) && can_emit_guards) { 5096 // We don't have sufficient type information, let's see if 5097 // speculative types can help. We need to have types for both src 5098 // and dest so that it pays off. 5099 5100 // Do we already have or could we have type information for src 5101 bool could_have_src = has_src; 5102 // Do we already have or could we have type information for dest 5103 bool could_have_dest = has_dest; 5104 5105 ciKlass* src_k = NULL; 5106 if (!has_src) { 5107 src_k = src_type->speculative_type_not_null(); 5108 if (src_k != NULL && src_k->is_array_klass()) { 5109 could_have_src = true; 5110 } 5111 } 5112 5113 ciKlass* dest_k = NULL; 5114 if (!has_dest) { 5115 dest_k = dest_type->speculative_type_not_null(); 5116 if (dest_k != NULL && dest_k->is_array_klass()) { 5117 could_have_dest = true; 5118 } 5119 } 5120 5121 if (could_have_src && could_have_dest) { 5122 // This is going to pay off so emit the required guards 5123 if (!has_src) { 5124 src = maybe_cast_profiled_obj(src, src_k, true); 5125 src_type = _gvn.type(src); 5126 top_src = src_type->isa_aryptr(); 5127 has_src = (top_src != NULL && top_src->klass() != NULL); 5128 src_spec = true; 5129 } 5130 if (!has_dest) { 5131 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5132 dest_type = _gvn.type(dest); 5133 top_dest = dest_type->isa_aryptr(); 5134 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 5135 dest_spec = true; 5136 } 5137 } 5138 } 5139 5140 if (has_src && has_dest && can_emit_guards) { 5141 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 5142 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 5143 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 5144 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 5145 5146 if (src_elem == dest_elem && src_elem == T_OBJECT) { 5147 // If both arrays are object arrays then having the exact types 5148 // for both will remove the need for a subtype check at runtime 5149 // before the call and may make it possible to pick a faster copy 5150 // routine (without a subtype check on every element) 5151 // Do we have the exact type of src? 5152 bool could_have_src = src_spec; 5153 // Do we have the exact type of dest? 5154 bool could_have_dest = dest_spec; 5155 ciKlass* src_k = top_src->klass(); 5156 ciKlass* dest_k = top_dest->klass(); 5157 if (!src_spec) { 5158 src_k = src_type->speculative_type_not_null(); 5159 if (src_k != NULL && src_k->is_array_klass()) { 5160 could_have_src = true; 5161 } 5162 } 5163 if (!dest_spec) { 5164 dest_k = dest_type->speculative_type_not_null(); 5165 if (dest_k != NULL && dest_k->is_array_klass()) { 5166 could_have_dest = true; 5167 } 5168 } 5169 if (could_have_src && could_have_dest) { 5170 // If we can have both exact types, emit the missing guards 5171 if (could_have_src && !src_spec) { 5172 src = maybe_cast_profiled_obj(src, src_k, true); 5173 } 5174 if (could_have_dest && !dest_spec) { 5175 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5176 } 5177 } 5178 } 5179 } 5180 5181 ciMethod* trap_method = method(); 5182 int trap_bci = bci(); 5183 if (saved_jvms != NULL) { 5184 trap_method = alloc->jvms()->method(); 5185 trap_bci = alloc->jvms()->bci(); 5186 } 5187 5188 bool negative_length_guard_generated = false; 5189 5190 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5191 can_emit_guards && 5192 !src->is_top() && !dest->is_top()) { 5193 // validate arguments: enables transformation the ArrayCopyNode 5194 validated = true; 5195 5196 RegionNode* slow_region = new RegionNode(1); 5197 record_for_igvn(slow_region); 5198 5199 // (1) src and dest are arrays. 5200 generate_non_array_guard(load_object_klass(src), slow_region); 5201 generate_non_array_guard(load_object_klass(dest), slow_region); 5202 5203 // (2) src and dest arrays must have elements of the same BasicType 5204 // done at macro expansion or at Ideal transformation time 5205 5206 // (4) src_offset must not be negative. 5207 generate_negative_guard(src_offset, slow_region); 5208 5209 // (5) dest_offset must not be negative. 5210 generate_negative_guard(dest_offset, slow_region); 5211 5212 // (7) src_offset + length must not exceed length of src. 5213 generate_limit_guard(src_offset, length, 5214 load_array_length(src), 5215 slow_region); 5216 5217 // (8) dest_offset + length must not exceed length of dest. 5218 generate_limit_guard(dest_offset, length, 5219 load_array_length(dest), 5220 slow_region); 5221 5222 // (6) length must not be negative. 5223 // This is also checked in generate_arraycopy() during macro expansion, but 5224 // we also have to check it here for the case where the ArrayCopyNode will 5225 // be eliminated by Escape Analysis. 5226 if (EliminateAllocations) { 5227 generate_negative_guard(length, slow_region); 5228 negative_length_guard_generated = true; 5229 } 5230 5231 // (9) each element of an oop array must be assignable 5232 Node* src_klass = load_object_klass(src); 5233 Node* dest_klass = load_object_klass(dest); 5234 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5235 5236 if (not_subtype_ctrl != top()) { 5237 PreserveJVMState pjvms(this); 5238 set_control(not_subtype_ctrl); 5239 uncommon_trap(Deoptimization::Reason_intrinsic, 5240 Deoptimization::Action_make_not_entrant); 5241 assert(stopped(), "Should be stopped"); 5242 } 5243 5244 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr(); 5245 const Type* toop = TypeOopPtr::make_from_klass(dest_klass_t->klass()); 5246 src = _gvn.transform(new CheckCastPPNode(control(), src, toop)); 5247 src_type = _gvn.type(src); 5248 top_src = src_type->isa_aryptr(); 5249 5250 if (top_dest != NULL && !top_dest->elem()->isa_valuetype() && !top_dest->is_not_flat()) { 5251 generate_valueArray_guard(dest_klass, slow_region); 5252 } 5253 5254 if (top_src != NULL && !top_src->elem()->isa_valuetype() && !top_src->is_not_flat()) { 5255 generate_valueArray_guard(src_klass, slow_region); 5256 } 5257 5258 { 5259 PreserveJVMState pjvms(this); 5260 set_control(_gvn.transform(slow_region)); 5261 uncommon_trap(Deoptimization::Reason_intrinsic, 5262 Deoptimization::Action_make_not_entrant); 5263 assert(stopped(), "Should be stopped"); 5264 } 5265 } 5266 5267 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx); 5268 5269 if (stopped()) { 5270 return true; 5271 } 5272 5273 Node* new_src = access_resolve(src, ACCESS_READ); 5274 Node* new_dest = access_resolve(dest, ACCESS_WRITE); 5275 5276 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, new_src, src_offset, new_dest, dest_offset, length, alloc != NULL, negative_length_guard_generated, 5277 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5278 // so the compiler has a chance to eliminate them: during macro expansion, 5279 // we have to set their control (CastPP nodes are eliminated). 5280 load_object_klass(src), load_object_klass(dest), 5281 load_array_length(src), load_array_length(dest)); 5282 5283 ac->set_arraycopy(validated); 5284 5285 Node* n = _gvn.transform(ac); 5286 if (n == ac) { 5287 ac->connect_outputs(this); 5288 } else { 5289 assert(validated, "shouldn't transform if all arguments not validated"); 5290 set_all_memory(n); 5291 } 5292 clear_upper_avx(); 5293 5294 5295 return true; 5296 } 5297 5298 5299 // Helper function which determines if an arraycopy immediately follows 5300 // an allocation, with no intervening tests or other escapes for the object. 5301 AllocateArrayNode* 5302 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5303 RegionNode* slow_region) { 5304 if (stopped()) return NULL; // no fast path 5305 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5306 5307 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5308 if (alloc == NULL) return NULL; 5309 5310 Node* rawmem = memory(Compile::AliasIdxRaw); 5311 // Is the allocation's memory state untouched? 5312 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5313 // Bail out if there have been raw-memory effects since the allocation. 5314 // (Example: There might have been a call or safepoint.) 5315 return NULL; 5316 } 5317 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5318 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5319 return NULL; 5320 } 5321 5322 // There must be no unexpected observers of this allocation. 5323 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5324 Node* obs = ptr->fast_out(i); 5325 if (obs != this->map()) { 5326 return NULL; 5327 } 5328 } 5329 5330 // This arraycopy must unconditionally follow the allocation of the ptr. 5331 Node* alloc_ctl = ptr->in(0); 5332 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5333 5334 Node* ctl = control(); 5335 while (ctl != alloc_ctl) { 5336 // There may be guards which feed into the slow_region. 5337 // Any other control flow means that we might not get a chance 5338 // to finish initializing the allocated object. 5339 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5340 IfNode* iff = ctl->in(0)->as_If(); 5341 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con); 5342 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5343 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5344 ctl = iff->in(0); // This test feeds the known slow_region. 5345 continue; 5346 } 5347 // One more try: Various low-level checks bottom out in 5348 // uncommon traps. If the debug-info of the trap omits 5349 // any reference to the allocation, as we've already 5350 // observed, then there can be no objection to the trap. 5351 bool found_trap = false; 5352 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5353 Node* obs = not_ctl->fast_out(j); 5354 if (obs->in(0) == not_ctl && obs->is_Call() && 5355 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5356 found_trap = true; break; 5357 } 5358 } 5359 if (found_trap) { 5360 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5361 continue; 5362 } 5363 } 5364 return NULL; 5365 } 5366 5367 // If we get this far, we have an allocation which immediately 5368 // precedes the arraycopy, and we can take over zeroing the new object. 5369 // The arraycopy will finish the initialization, and provide 5370 // a new control state to which we will anchor the destination pointer. 5371 5372 return alloc; 5373 } 5374 5375 //-------------inline_encodeISOArray----------------------------------- 5376 // encode char[] to byte[] in ISO_8859_1 5377 bool LibraryCallKit::inline_encodeISOArray() { 5378 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5379 // no receiver since it is static method 5380 Node *src = argument(0); 5381 Node *src_offset = argument(1); 5382 Node *dst = argument(2); 5383 Node *dst_offset = argument(3); 5384 Node *length = argument(4); 5385 5386 src = must_be_not_null(src, true); 5387 dst = must_be_not_null(dst, true); 5388 5389 src = access_resolve(src, ACCESS_READ); 5390 dst = access_resolve(dst, ACCESS_WRITE); 5391 5392 const Type* src_type = src->Value(&_gvn); 5393 const Type* dst_type = dst->Value(&_gvn); 5394 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5395 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5396 if (top_src == NULL || top_src->klass() == NULL || 5397 top_dest == NULL || top_dest->klass() == NULL) { 5398 // failed array check 5399 return false; 5400 } 5401 5402 // Figure out the size and type of the elements we will be copying. 5403 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5404 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5405 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5406 return false; 5407 } 5408 5409 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5410 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5411 // 'src_start' points to src array + scaled offset 5412 // 'dst_start' points to dst array + scaled offset 5413 5414 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5415 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5416 enc = _gvn.transform(enc); 5417 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5418 set_memory(res_mem, mtype); 5419 set_result(enc); 5420 clear_upper_avx(); 5421 5422 return true; 5423 } 5424 5425 //-------------inline_multiplyToLen----------------------------------- 5426 bool LibraryCallKit::inline_multiplyToLen() { 5427 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5428 5429 address stubAddr = StubRoutines::multiplyToLen(); 5430 if (stubAddr == NULL) { 5431 return false; // Intrinsic's stub is not implemented on this platform 5432 } 5433 const char* stubName = "multiplyToLen"; 5434 5435 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5436 5437 // no receiver because it is a static method 5438 Node* x = argument(0); 5439 Node* xlen = argument(1); 5440 Node* y = argument(2); 5441 Node* ylen = argument(3); 5442 Node* z = argument(4); 5443 5444 x = must_be_not_null(x, true); 5445 y = must_be_not_null(y, true); 5446 5447 x = access_resolve(x, ACCESS_READ); 5448 y = access_resolve(y, ACCESS_READ); 5449 z = access_resolve(z, ACCESS_WRITE); 5450 5451 const Type* x_type = x->Value(&_gvn); 5452 const Type* y_type = y->Value(&_gvn); 5453 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5454 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5455 if (top_x == NULL || top_x->klass() == NULL || 5456 top_y == NULL || top_y->klass() == NULL) { 5457 // failed array check 5458 return false; 5459 } 5460 5461 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5462 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5463 if (x_elem != T_INT || y_elem != T_INT) { 5464 return false; 5465 } 5466 5467 // Set the original stack and the reexecute bit for the interpreter to reexecute 5468 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5469 // on the return from z array allocation in runtime. 5470 { PreserveReexecuteState preexecs(this); 5471 jvms()->set_should_reexecute(true); 5472 5473 Node* x_start = array_element_address(x, intcon(0), x_elem); 5474 Node* y_start = array_element_address(y, intcon(0), y_elem); 5475 // 'x_start' points to x array + scaled xlen 5476 // 'y_start' points to y array + scaled ylen 5477 5478 // Allocate the result array 5479 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5480 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5481 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5482 5483 IdealKit ideal(this); 5484 5485 #define __ ideal. 5486 Node* one = __ ConI(1); 5487 Node* zero = __ ConI(0); 5488 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5489 __ set(need_alloc, zero); 5490 __ set(z_alloc, z); 5491 __ if_then(z, BoolTest::eq, null()); { 5492 __ increment (need_alloc, one); 5493 } __ else_(); { 5494 // Update graphKit memory and control from IdealKit. 5495 sync_kit(ideal); 5496 Node *cast = new CastPPNode(z, TypePtr::NOTNULL); 5497 cast->init_req(0, control()); 5498 _gvn.set_type(cast, cast->bottom_type()); 5499 C->record_for_igvn(cast); 5500 5501 Node* zlen_arg = load_array_length(cast); 5502 // Update IdealKit memory and control from graphKit. 5503 __ sync_kit(this); 5504 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5505 __ increment (need_alloc, one); 5506 } __ end_if(); 5507 } __ end_if(); 5508 5509 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5510 // Update graphKit memory and control from IdealKit. 5511 sync_kit(ideal); 5512 Node * narr = new_array(klass_node, zlen, 1); 5513 // Update IdealKit memory and control from graphKit. 5514 __ sync_kit(this); 5515 __ set(z_alloc, narr); 5516 } __ end_if(); 5517 5518 sync_kit(ideal); 5519 z = __ value(z_alloc); 5520 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5521 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5522 // Final sync IdealKit and GraphKit. 5523 final_sync(ideal); 5524 #undef __ 5525 5526 Node* z_start = array_element_address(z, intcon(0), T_INT); 5527 5528 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5529 OptoRuntime::multiplyToLen_Type(), 5530 stubAddr, stubName, TypePtr::BOTTOM, 5531 x_start, xlen, y_start, ylen, z_start, zlen); 5532 } // original reexecute is set back here 5533 5534 C->set_has_split_ifs(true); // Has chance for split-if optimization 5535 set_result(z); 5536 return true; 5537 } 5538 5539 //-------------inline_squareToLen------------------------------------ 5540 bool LibraryCallKit::inline_squareToLen() { 5541 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5542 5543 address stubAddr = StubRoutines::squareToLen(); 5544 if (stubAddr == NULL) { 5545 return false; // Intrinsic's stub is not implemented on this platform 5546 } 5547 const char* stubName = "squareToLen"; 5548 5549 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5550 5551 Node* x = argument(0); 5552 Node* len = argument(1); 5553 Node* z = argument(2); 5554 Node* zlen = argument(3); 5555 5556 x = must_be_not_null(x, true); 5557 z = must_be_not_null(z, true); 5558 5559 x = access_resolve(x, ACCESS_READ); 5560 z = access_resolve(z, ACCESS_WRITE); 5561 5562 const Type* x_type = x->Value(&_gvn); 5563 const Type* z_type = z->Value(&_gvn); 5564 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5565 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5566 if (top_x == NULL || top_x->klass() == NULL || 5567 top_z == NULL || top_z->klass() == NULL) { 5568 // failed array check 5569 return false; 5570 } 5571 5572 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5573 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5574 if (x_elem != T_INT || z_elem != T_INT) { 5575 return false; 5576 } 5577 5578 5579 Node* x_start = array_element_address(x, intcon(0), x_elem); 5580 Node* z_start = array_element_address(z, intcon(0), z_elem); 5581 5582 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5583 OptoRuntime::squareToLen_Type(), 5584 stubAddr, stubName, TypePtr::BOTTOM, 5585 x_start, len, z_start, zlen); 5586 5587 set_result(z); 5588 return true; 5589 } 5590 5591 //-------------inline_mulAdd------------------------------------------ 5592 bool LibraryCallKit::inline_mulAdd() { 5593 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5594 5595 address stubAddr = StubRoutines::mulAdd(); 5596 if (stubAddr == NULL) { 5597 return false; // Intrinsic's stub is not implemented on this platform 5598 } 5599 const char* stubName = "mulAdd"; 5600 5601 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5602 5603 Node* out = argument(0); 5604 Node* in = argument(1); 5605 Node* offset = argument(2); 5606 Node* len = argument(3); 5607 Node* k = argument(4); 5608 5609 out = must_be_not_null(out, true); 5610 5611 in = access_resolve(in, ACCESS_READ); 5612 out = access_resolve(out, ACCESS_WRITE); 5613 5614 const Type* out_type = out->Value(&_gvn); 5615 const Type* in_type = in->Value(&_gvn); 5616 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5617 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5618 if (top_out == NULL || top_out->klass() == NULL || 5619 top_in == NULL || top_in->klass() == NULL) { 5620 // failed array check 5621 return false; 5622 } 5623 5624 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5625 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5626 if (out_elem != T_INT || in_elem != T_INT) { 5627 return false; 5628 } 5629 5630 Node* outlen = load_array_length(out); 5631 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5632 Node* out_start = array_element_address(out, intcon(0), out_elem); 5633 Node* in_start = array_element_address(in, intcon(0), in_elem); 5634 5635 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5636 OptoRuntime::mulAdd_Type(), 5637 stubAddr, stubName, TypePtr::BOTTOM, 5638 out_start,in_start, new_offset, len, k); 5639 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5640 set_result(result); 5641 return true; 5642 } 5643 5644 //-------------inline_montgomeryMultiply----------------------------------- 5645 bool LibraryCallKit::inline_montgomeryMultiply() { 5646 address stubAddr = StubRoutines::montgomeryMultiply(); 5647 if (stubAddr == NULL) { 5648 return false; // Intrinsic's stub is not implemented on this platform 5649 } 5650 5651 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5652 const char* stubName = "montgomery_multiply"; 5653 5654 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5655 5656 Node* a = argument(0); 5657 Node* b = argument(1); 5658 Node* n = argument(2); 5659 Node* len = argument(3); 5660 Node* inv = argument(4); 5661 Node* m = argument(6); 5662 5663 a = access_resolve(a, ACCESS_READ); 5664 b = access_resolve(b, ACCESS_READ); 5665 n = access_resolve(n, ACCESS_READ); 5666 m = access_resolve(m, ACCESS_WRITE); 5667 5668 const Type* a_type = a->Value(&_gvn); 5669 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5670 const Type* b_type = b->Value(&_gvn); 5671 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5672 const Type* n_type = a->Value(&_gvn); 5673 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5674 const Type* m_type = a->Value(&_gvn); 5675 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5676 if (top_a == NULL || top_a->klass() == NULL || 5677 top_b == NULL || top_b->klass() == NULL || 5678 top_n == NULL || top_n->klass() == NULL || 5679 top_m == NULL || top_m->klass() == NULL) { 5680 // failed array check 5681 return false; 5682 } 5683 5684 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5685 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5686 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5687 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5688 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5689 return false; 5690 } 5691 5692 // Make the call 5693 { 5694 Node* a_start = array_element_address(a, intcon(0), a_elem); 5695 Node* b_start = array_element_address(b, intcon(0), b_elem); 5696 Node* n_start = array_element_address(n, intcon(0), n_elem); 5697 Node* m_start = array_element_address(m, intcon(0), m_elem); 5698 5699 Node* call = make_runtime_call(RC_LEAF, 5700 OptoRuntime::montgomeryMultiply_Type(), 5701 stubAddr, stubName, TypePtr::BOTTOM, 5702 a_start, b_start, n_start, len, inv, top(), 5703 m_start); 5704 set_result(m); 5705 } 5706 5707 return true; 5708 } 5709 5710 bool LibraryCallKit::inline_montgomerySquare() { 5711 address stubAddr = StubRoutines::montgomerySquare(); 5712 if (stubAddr == NULL) { 5713 return false; // Intrinsic's stub is not implemented on this platform 5714 } 5715 5716 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5717 const char* stubName = "montgomery_square"; 5718 5719 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5720 5721 Node* a = argument(0); 5722 Node* n = argument(1); 5723 Node* len = argument(2); 5724 Node* inv = argument(3); 5725 Node* m = argument(5); 5726 5727 a = access_resolve(a, ACCESS_READ); 5728 n = access_resolve(n, ACCESS_READ); 5729 m = access_resolve(m, ACCESS_WRITE); 5730 5731 const Type* a_type = a->Value(&_gvn); 5732 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5733 const Type* n_type = a->Value(&_gvn); 5734 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5735 const Type* m_type = a->Value(&_gvn); 5736 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5737 if (top_a == NULL || top_a->klass() == NULL || 5738 top_n == NULL || top_n->klass() == NULL || 5739 top_m == NULL || top_m->klass() == NULL) { 5740 // failed array check 5741 return false; 5742 } 5743 5744 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5745 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5746 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5747 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5748 return false; 5749 } 5750 5751 // Make the call 5752 { 5753 Node* a_start = array_element_address(a, intcon(0), a_elem); 5754 Node* n_start = array_element_address(n, intcon(0), n_elem); 5755 Node* m_start = array_element_address(m, intcon(0), m_elem); 5756 5757 Node* call = make_runtime_call(RC_LEAF, 5758 OptoRuntime::montgomerySquare_Type(), 5759 stubAddr, stubName, TypePtr::BOTTOM, 5760 a_start, n_start, len, inv, top(), 5761 m_start); 5762 set_result(m); 5763 } 5764 5765 return true; 5766 } 5767 5768 //-------------inline_vectorizedMismatch------------------------------ 5769 bool LibraryCallKit::inline_vectorizedMismatch() { 5770 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5771 5772 address stubAddr = StubRoutines::vectorizedMismatch(); 5773 if (stubAddr == NULL) { 5774 return false; // Intrinsic's stub is not implemented on this platform 5775 } 5776 const char* stubName = "vectorizedMismatch"; 5777 int size_l = callee()->signature()->size(); 5778 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5779 5780 Node* obja = argument(0); 5781 Node* aoffset = argument(1); 5782 Node* objb = argument(3); 5783 Node* boffset = argument(4); 5784 Node* length = argument(6); 5785 Node* scale = argument(7); 5786 5787 const Type* a_type = obja->Value(&_gvn); 5788 const Type* b_type = objb->Value(&_gvn); 5789 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5790 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5791 if (top_a == NULL || top_a->klass() == NULL || 5792 top_b == NULL || top_b->klass() == NULL) { 5793 // failed array check 5794 return false; 5795 } 5796 5797 Node* call; 5798 jvms()->set_should_reexecute(true); 5799 5800 obja = access_resolve(obja, ACCESS_READ); 5801 objb = access_resolve(objb, ACCESS_READ); 5802 Node* obja_adr = make_unsafe_address(obja, aoffset, ACCESS_READ); 5803 Node* objb_adr = make_unsafe_address(objb, boffset, ACCESS_READ); 5804 5805 call = make_runtime_call(RC_LEAF, 5806 OptoRuntime::vectorizedMismatch_Type(), 5807 stubAddr, stubName, TypePtr::BOTTOM, 5808 obja_adr, objb_adr, length, scale); 5809 5810 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5811 set_result(result); 5812 return true; 5813 } 5814 5815 /** 5816 * Calculate CRC32 for byte. 5817 * int java.util.zip.CRC32.update(int crc, int b) 5818 */ 5819 bool LibraryCallKit::inline_updateCRC32() { 5820 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5821 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5822 // no receiver since it is static method 5823 Node* crc = argument(0); // type: int 5824 Node* b = argument(1); // type: int 5825 5826 /* 5827 * int c = ~ crc; 5828 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5829 * b = b ^ (c >>> 8); 5830 * crc = ~b; 5831 */ 5832 5833 Node* M1 = intcon(-1); 5834 crc = _gvn.transform(new XorINode(crc, M1)); 5835 Node* result = _gvn.transform(new XorINode(crc, b)); 5836 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5837 5838 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5839 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5840 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5841 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5842 5843 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5844 result = _gvn.transform(new XorINode(crc, result)); 5845 result = _gvn.transform(new XorINode(result, M1)); 5846 set_result(result); 5847 return true; 5848 } 5849 5850 /** 5851 * Calculate CRC32 for byte[] array. 5852 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5853 */ 5854 bool LibraryCallKit::inline_updateBytesCRC32() { 5855 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5856 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5857 // no receiver since it is static method 5858 Node* crc = argument(0); // type: int 5859 Node* src = argument(1); // type: oop 5860 Node* offset = argument(2); // type: int 5861 Node* length = argument(3); // type: int 5862 5863 const Type* src_type = src->Value(&_gvn); 5864 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5865 if (top_src == NULL || top_src->klass() == NULL) { 5866 // failed array check 5867 return false; 5868 } 5869 5870 // Figure out the size and type of the elements we will be copying. 5871 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5872 if (src_elem != T_BYTE) { 5873 return false; 5874 } 5875 5876 // 'src_start' points to src array + scaled offset 5877 src = must_be_not_null(src, true); 5878 src = access_resolve(src, ACCESS_READ); 5879 Node* src_start = array_element_address(src, offset, src_elem); 5880 5881 // We assume that range check is done by caller. 5882 // TODO: generate range check (offset+length < src.length) in debug VM. 5883 5884 // Call the stub. 5885 address stubAddr = StubRoutines::updateBytesCRC32(); 5886 const char *stubName = "updateBytesCRC32"; 5887 5888 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5889 stubAddr, stubName, TypePtr::BOTTOM, 5890 crc, src_start, length); 5891 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5892 set_result(result); 5893 return true; 5894 } 5895 5896 /** 5897 * Calculate CRC32 for ByteBuffer. 5898 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5899 */ 5900 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5901 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5902 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5903 // no receiver since it is static method 5904 Node* crc = argument(0); // type: int 5905 Node* src = argument(1); // type: long 5906 Node* offset = argument(3); // type: int 5907 Node* length = argument(4); // type: int 5908 5909 src = ConvL2X(src); // adjust Java long to machine word 5910 Node* base = _gvn.transform(new CastX2PNode(src)); 5911 offset = ConvI2X(offset); 5912 5913 // 'src_start' points to src array + scaled offset 5914 Node* src_start = basic_plus_adr(top(), base, offset); 5915 5916 // Call the stub. 5917 address stubAddr = StubRoutines::updateBytesCRC32(); 5918 const char *stubName = "updateBytesCRC32"; 5919 5920 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5921 stubAddr, stubName, TypePtr::BOTTOM, 5922 crc, src_start, length); 5923 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5924 set_result(result); 5925 return true; 5926 } 5927 5928 //------------------------------get_table_from_crc32c_class----------------------- 5929 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5930 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5931 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5932 5933 return table; 5934 } 5935 5936 //------------------------------inline_updateBytesCRC32C----------------------- 5937 // 5938 // Calculate CRC32C for byte[] array. 5939 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5940 // 5941 bool LibraryCallKit::inline_updateBytesCRC32C() { 5942 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5943 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5944 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5945 // no receiver since it is a static method 5946 Node* crc = argument(0); // type: int 5947 Node* src = argument(1); // type: oop 5948 Node* offset = argument(2); // type: int 5949 Node* end = argument(3); // type: int 5950 5951 Node* length = _gvn.transform(new SubINode(end, offset)); 5952 5953 const Type* src_type = src->Value(&_gvn); 5954 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5955 if (top_src == NULL || top_src->klass() == NULL) { 5956 // failed array check 5957 return false; 5958 } 5959 5960 // Figure out the size and type of the elements we will be copying. 5961 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5962 if (src_elem != T_BYTE) { 5963 return false; 5964 } 5965 5966 // 'src_start' points to src array + scaled offset 5967 src = must_be_not_null(src, true); 5968 src = access_resolve(src, ACCESS_READ); 5969 Node* src_start = array_element_address(src, offset, src_elem); 5970 5971 // static final int[] byteTable in class CRC32C 5972 Node* table = get_table_from_crc32c_class(callee()->holder()); 5973 table = must_be_not_null(table, true); 5974 table = access_resolve(table, ACCESS_READ); 5975 Node* table_start = array_element_address(table, intcon(0), T_INT); 5976 5977 // We assume that range check is done by caller. 5978 // TODO: generate range check (offset+length < src.length) in debug VM. 5979 5980 // Call the stub. 5981 address stubAddr = StubRoutines::updateBytesCRC32C(); 5982 const char *stubName = "updateBytesCRC32C"; 5983 5984 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5985 stubAddr, stubName, TypePtr::BOTTOM, 5986 crc, src_start, length, table_start); 5987 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5988 set_result(result); 5989 return true; 5990 } 5991 5992 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5993 // 5994 // Calculate CRC32C for DirectByteBuffer. 5995 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5996 // 5997 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5998 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5999 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 6000 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 6001 // no receiver since it is a static method 6002 Node* crc = argument(0); // type: int 6003 Node* src = argument(1); // type: long 6004 Node* offset = argument(3); // type: int 6005 Node* end = argument(4); // type: int 6006 6007 Node* length = _gvn.transform(new SubINode(end, offset)); 6008 6009 src = ConvL2X(src); // adjust Java long to machine word 6010 Node* base = _gvn.transform(new CastX2PNode(src)); 6011 offset = ConvI2X(offset); 6012 6013 // 'src_start' points to src array + scaled offset 6014 Node* src_start = basic_plus_adr(top(), base, offset); 6015 6016 // static final int[] byteTable in class CRC32C 6017 Node* table = get_table_from_crc32c_class(callee()->holder()); 6018 table = must_be_not_null(table, true); 6019 table = access_resolve(table, ACCESS_READ); 6020 Node* table_start = array_element_address(table, intcon(0), T_INT); 6021 6022 // Call the stub. 6023 address stubAddr = StubRoutines::updateBytesCRC32C(); 6024 const char *stubName = "updateBytesCRC32C"; 6025 6026 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 6027 stubAddr, stubName, TypePtr::BOTTOM, 6028 crc, src_start, length, table_start); 6029 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6030 set_result(result); 6031 return true; 6032 } 6033 6034 //------------------------------inline_updateBytesAdler32---------------------- 6035 // 6036 // Calculate Adler32 checksum for byte[] array. 6037 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 6038 // 6039 bool LibraryCallKit::inline_updateBytesAdler32() { 6040 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 6041 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 6042 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 6043 // no receiver since it is static method 6044 Node* crc = argument(0); // type: int 6045 Node* src = argument(1); // type: oop 6046 Node* offset = argument(2); // type: int 6047 Node* length = argument(3); // type: int 6048 6049 const Type* src_type = src->Value(&_gvn); 6050 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6051 if (top_src == NULL || top_src->klass() == NULL) { 6052 // failed array check 6053 return false; 6054 } 6055 6056 // Figure out the size and type of the elements we will be copying. 6057 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6058 if (src_elem != T_BYTE) { 6059 return false; 6060 } 6061 6062 // 'src_start' points to src array + scaled offset 6063 src = access_resolve(src, ACCESS_READ); 6064 Node* src_start = array_element_address(src, offset, src_elem); 6065 6066 // We assume that range check is done by caller. 6067 // TODO: generate range check (offset+length < src.length) in debug VM. 6068 6069 // Call the stub. 6070 address stubAddr = StubRoutines::updateBytesAdler32(); 6071 const char *stubName = "updateBytesAdler32"; 6072 6073 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 6074 stubAddr, stubName, TypePtr::BOTTOM, 6075 crc, src_start, length); 6076 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6077 set_result(result); 6078 return true; 6079 } 6080 6081 //------------------------------inline_updateByteBufferAdler32--------------- 6082 // 6083 // Calculate Adler32 checksum for DirectByteBuffer. 6084 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 6085 // 6086 bool LibraryCallKit::inline_updateByteBufferAdler32() { 6087 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 6088 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 6089 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 6090 // no receiver since it is static method 6091 Node* crc = argument(0); // type: int 6092 Node* src = argument(1); // type: long 6093 Node* offset = argument(3); // type: int 6094 Node* length = argument(4); // type: int 6095 6096 src = ConvL2X(src); // adjust Java long to machine word 6097 Node* base = _gvn.transform(new CastX2PNode(src)); 6098 offset = ConvI2X(offset); 6099 6100 // 'src_start' points to src array + scaled offset 6101 Node* src_start = basic_plus_adr(top(), base, offset); 6102 6103 // Call the stub. 6104 address stubAddr = StubRoutines::updateBytesAdler32(); 6105 const char *stubName = "updateBytesAdler32"; 6106 6107 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 6108 stubAddr, stubName, TypePtr::BOTTOM, 6109 crc, src_start, length); 6110 6111 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6112 set_result(result); 6113 return true; 6114 } 6115 6116 //----------------------------inline_reference_get---------------------------- 6117 // public T java.lang.ref.Reference.get(); 6118 bool LibraryCallKit::inline_reference_get() { 6119 const int referent_offset = java_lang_ref_Reference::referent_offset; 6120 guarantee(referent_offset > 0, "should have already been set"); 6121 6122 // Get the argument: 6123 Node* reference_obj = null_check_receiver(); 6124 if (stopped()) return true; 6125 6126 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr(); 6127 assert(tinst != NULL, "obj is null"); 6128 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6129 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass(); 6130 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"), 6131 ciSymbol::make("Ljava/lang/Object;"), 6132 false); 6133 assert (field != NULL, "undefined field"); 6134 6135 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 6136 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 6137 6138 ciInstanceKlass* klass = env()->Object_klass(); 6139 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 6140 6141 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF; 6142 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators); 6143 // Add memory barrier to prevent commoning reads from this field 6144 // across safepoint since GC can change its value. 6145 insert_mem_bar(Op_MemBarCPUOrder); 6146 6147 set_result(result); 6148 return true; 6149 } 6150 6151 6152 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 6153 bool is_exact=true, bool is_static=false, 6154 ciInstanceKlass * fromKls=NULL) { 6155 if (fromKls == NULL) { 6156 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 6157 assert(tinst != NULL, "obj is null"); 6158 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6159 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 6160 fromKls = tinst->klass()->as_instance_klass(); 6161 } else { 6162 assert(is_static, "only for static field access"); 6163 } 6164 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 6165 ciSymbol::make(fieldTypeString), 6166 is_static); 6167 6168 assert (field != NULL, "undefined field"); 6169 if (field == NULL) return (Node *) NULL; 6170 6171 if (is_static) { 6172 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 6173 fromObj = makecon(tip); 6174 } 6175 6176 // Next code copied from Parse::do_get_xxx(): 6177 6178 // Compute address and memory type. 6179 int offset = field->offset_in_bytes(); 6180 bool is_vol = field->is_volatile(); 6181 ciType* field_klass = field->type(); 6182 assert(field_klass->is_loaded(), "should be loaded"); 6183 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 6184 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 6185 BasicType bt = field->layout_type(); 6186 6187 // Build the resultant type of the load 6188 const Type *type; 6189 if (bt == T_OBJECT) { 6190 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 6191 } else { 6192 type = Type::get_const_basic_type(bt); 6193 } 6194 6195 DecoratorSet decorators = IN_HEAP; 6196 6197 if (is_vol) { 6198 decorators |= MO_SEQ_CST; 6199 } 6200 6201 return access_load_at(fromObj, adr, adr_type, type, bt, decorators); 6202 } 6203 6204 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 6205 bool is_exact = true, bool is_static = false, 6206 ciInstanceKlass * fromKls = NULL) { 6207 if (fromKls == NULL) { 6208 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 6209 assert(tinst != NULL, "obj is null"); 6210 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6211 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 6212 fromKls = tinst->klass()->as_instance_klass(); 6213 } 6214 else { 6215 assert(is_static, "only for static field access"); 6216 } 6217 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 6218 ciSymbol::make(fieldTypeString), 6219 is_static); 6220 6221 assert(field != NULL, "undefined field"); 6222 assert(!field->is_volatile(), "not defined for volatile fields"); 6223 6224 if (is_static) { 6225 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 6226 fromObj = makecon(tip); 6227 } 6228 6229 // Next code copied from Parse::do_get_xxx(): 6230 6231 // Compute address and memory type. 6232 int offset = field->offset_in_bytes(); 6233 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 6234 6235 return adr; 6236 } 6237 6238 //------------------------------inline_aescrypt_Block----------------------- 6239 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 6240 address stubAddr = NULL; 6241 const char *stubName; 6242 assert(UseAES, "need AES instruction support"); 6243 6244 switch(id) { 6245 case vmIntrinsics::_aescrypt_encryptBlock: 6246 stubAddr = StubRoutines::aescrypt_encryptBlock(); 6247 stubName = "aescrypt_encryptBlock"; 6248 break; 6249 case vmIntrinsics::_aescrypt_decryptBlock: 6250 stubAddr = StubRoutines::aescrypt_decryptBlock(); 6251 stubName = "aescrypt_decryptBlock"; 6252 break; 6253 default: 6254 break; 6255 } 6256 if (stubAddr == NULL) return false; 6257 6258 Node* aescrypt_object = argument(0); 6259 Node* src = argument(1); 6260 Node* src_offset = argument(2); 6261 Node* dest = argument(3); 6262 Node* dest_offset = argument(4); 6263 6264 src = must_be_not_null(src, true); 6265 dest = must_be_not_null(dest, true); 6266 6267 src = access_resolve(src, ACCESS_READ); 6268 dest = access_resolve(dest, ACCESS_WRITE); 6269 6270 // (1) src and dest are arrays. 6271 const Type* src_type = src->Value(&_gvn); 6272 const Type* dest_type = dest->Value(&_gvn); 6273 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6274 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6275 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6276 6277 // for the quick and dirty code we will skip all the checks. 6278 // we are just trying to get the call to be generated. 6279 Node* src_start = src; 6280 Node* dest_start = dest; 6281 if (src_offset != NULL || dest_offset != NULL) { 6282 assert(src_offset != NULL && dest_offset != NULL, ""); 6283 src_start = array_element_address(src, src_offset, T_BYTE); 6284 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6285 } 6286 6287 // now need to get the start of its expanded key array 6288 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6289 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6290 if (k_start == NULL) return false; 6291 6292 if (Matcher::pass_original_key_for_aes()) { 6293 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6294 // compatibility issues between Java key expansion and SPARC crypto instructions 6295 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6296 if (original_k_start == NULL) return false; 6297 6298 // Call the stub. 6299 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6300 stubAddr, stubName, TypePtr::BOTTOM, 6301 src_start, dest_start, k_start, original_k_start); 6302 } else { 6303 // Call the stub. 6304 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6305 stubAddr, stubName, TypePtr::BOTTOM, 6306 src_start, dest_start, k_start); 6307 } 6308 6309 return true; 6310 } 6311 6312 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 6313 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 6314 address stubAddr = NULL; 6315 const char *stubName = NULL; 6316 6317 assert(UseAES, "need AES instruction support"); 6318 6319 switch(id) { 6320 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6321 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6322 stubName = "cipherBlockChaining_encryptAESCrypt"; 6323 break; 6324 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6325 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6326 stubName = "cipherBlockChaining_decryptAESCrypt"; 6327 break; 6328 default: 6329 break; 6330 } 6331 if (stubAddr == NULL) return false; 6332 6333 Node* cipherBlockChaining_object = argument(0); 6334 Node* src = argument(1); 6335 Node* src_offset = argument(2); 6336 Node* len = argument(3); 6337 Node* dest = argument(4); 6338 Node* dest_offset = argument(5); 6339 6340 src = must_be_not_null(src, false); 6341 dest = must_be_not_null(dest, false); 6342 6343 src = access_resolve(src, ACCESS_READ); 6344 dest = access_resolve(dest, ACCESS_WRITE); 6345 6346 // (1) src and dest are arrays. 6347 const Type* src_type = src->Value(&_gvn); 6348 const Type* dest_type = dest->Value(&_gvn); 6349 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6350 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6351 assert (top_src != NULL && top_src->klass() != NULL 6352 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6353 6354 // checks are the responsibility of the caller 6355 Node* src_start = src; 6356 Node* dest_start = dest; 6357 if (src_offset != NULL || dest_offset != NULL) { 6358 assert(src_offset != NULL && dest_offset != NULL, ""); 6359 src_start = array_element_address(src, src_offset, T_BYTE); 6360 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6361 } 6362 6363 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6364 // (because of the predicated logic executed earlier). 6365 // so we cast it here safely. 6366 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6367 6368 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6369 if (embeddedCipherObj == NULL) return false; 6370 6371 // cast it to what we know it will be at runtime 6372 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6373 assert(tinst != NULL, "CBC obj is null"); 6374 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6375 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6376 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6377 6378 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6379 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6380 const TypeOopPtr* xtype = aklass->as_instance_type(); 6381 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6382 aescrypt_object = _gvn.transform(aescrypt_object); 6383 6384 // we need to get the start of the aescrypt_object's expanded key array 6385 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6386 if (k_start == NULL) return false; 6387 6388 // similarly, get the start address of the r vector 6389 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6390 if (objRvec == NULL) return false; 6391 objRvec = access_resolve(objRvec, ACCESS_WRITE); 6392 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6393 6394 Node* cbcCrypt; 6395 if (Matcher::pass_original_key_for_aes()) { 6396 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6397 // compatibility issues between Java key expansion and SPARC crypto instructions 6398 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6399 if (original_k_start == NULL) return false; 6400 6401 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6402 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6403 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6404 stubAddr, stubName, TypePtr::BOTTOM, 6405 src_start, dest_start, k_start, r_start, len, original_k_start); 6406 } else { 6407 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6408 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6409 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6410 stubAddr, stubName, TypePtr::BOTTOM, 6411 src_start, dest_start, k_start, r_start, len); 6412 } 6413 6414 // return cipher length (int) 6415 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6416 set_result(retvalue); 6417 return true; 6418 } 6419 6420 //------------------------------inline_counterMode_AESCrypt----------------------- 6421 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6422 assert(UseAES, "need AES instruction support"); 6423 if (!UseAESCTRIntrinsics) return false; 6424 6425 address stubAddr = NULL; 6426 const char *stubName = NULL; 6427 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6428 stubAddr = StubRoutines::counterMode_AESCrypt(); 6429 stubName = "counterMode_AESCrypt"; 6430 } 6431 if (stubAddr == NULL) return false; 6432 6433 Node* counterMode_object = argument(0); 6434 Node* src = argument(1); 6435 Node* src_offset = argument(2); 6436 Node* len = argument(3); 6437 Node* dest = argument(4); 6438 Node* dest_offset = argument(5); 6439 6440 src = access_resolve(src, ACCESS_READ); 6441 dest = access_resolve(dest, ACCESS_WRITE); 6442 counterMode_object = access_resolve(counterMode_object, ACCESS_WRITE); 6443 6444 // (1) src and dest are arrays. 6445 const Type* src_type = src->Value(&_gvn); 6446 const Type* dest_type = dest->Value(&_gvn); 6447 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6448 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6449 assert(top_src != NULL && top_src->klass() != NULL && 6450 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6451 6452 // checks are the responsibility of the caller 6453 Node* src_start = src; 6454 Node* dest_start = dest; 6455 if (src_offset != NULL || dest_offset != NULL) { 6456 assert(src_offset != NULL && dest_offset != NULL, ""); 6457 src_start = array_element_address(src, src_offset, T_BYTE); 6458 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6459 } 6460 6461 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6462 // (because of the predicated logic executed earlier). 6463 // so we cast it here safely. 6464 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6465 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6466 if (embeddedCipherObj == NULL) return false; 6467 // cast it to what we know it will be at runtime 6468 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6469 assert(tinst != NULL, "CTR obj is null"); 6470 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6471 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6472 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6473 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6474 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6475 const TypeOopPtr* xtype = aklass->as_instance_type(); 6476 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6477 aescrypt_object = _gvn.transform(aescrypt_object); 6478 // we need to get the start of the aescrypt_object's expanded key array 6479 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6480 if (k_start == NULL) return false; 6481 // similarly, get the start address of the r vector 6482 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6483 if (obj_counter == NULL) return false; 6484 obj_counter = access_resolve(obj_counter, ACCESS_WRITE); 6485 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6486 6487 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6488 if (saved_encCounter == NULL) return false; 6489 saved_encCounter = access_resolve(saved_encCounter, ACCESS_WRITE); 6490 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6491 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6492 6493 Node* ctrCrypt; 6494 if (Matcher::pass_original_key_for_aes()) { 6495 // no SPARC version for AES/CTR intrinsics now. 6496 return false; 6497 } 6498 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6499 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6500 OptoRuntime::counterMode_aescrypt_Type(), 6501 stubAddr, stubName, TypePtr::BOTTOM, 6502 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6503 6504 // return cipher length (int) 6505 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6506 set_result(retvalue); 6507 return true; 6508 } 6509 6510 //------------------------------get_key_start_from_aescrypt_object----------------------- 6511 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6512 #if defined(PPC64) || defined(S390) 6513 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6514 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6515 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6516 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6517 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6518 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6519 if (objSessionK == NULL) { 6520 return (Node *) NULL; 6521 } 6522 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6523 #else 6524 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6525 #endif // PPC64 6526 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6527 if (objAESCryptKey == NULL) return (Node *) NULL; 6528 6529 // now have the array, need to get the start address of the K array 6530 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ); 6531 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6532 return k_start; 6533 } 6534 6535 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6536 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6537 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6538 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6539 if (objAESCryptKey == NULL) return (Node *) NULL; 6540 6541 // now have the array, need to get the start address of the lastKey array 6542 objAESCryptKey = access_resolve(objAESCryptKey, ACCESS_READ); 6543 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6544 return original_k_start; 6545 } 6546 6547 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6548 // Return node representing slow path of predicate check. 6549 // the pseudo code we want to emulate with this predicate is: 6550 // for encryption: 6551 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6552 // for decryption: 6553 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6554 // note cipher==plain is more conservative than the original java code but that's OK 6555 // 6556 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6557 // The receiver was checked for NULL already. 6558 Node* objCBC = argument(0); 6559 6560 Node* src = argument(1); 6561 Node* dest = argument(4); 6562 6563 // Load embeddedCipher field of CipherBlockChaining object. 6564 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6565 6566 // get AESCrypt klass for instanceOf check 6567 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6568 // will have same classloader as CipherBlockChaining object 6569 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6570 assert(tinst != NULL, "CBCobj is null"); 6571 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6572 6573 // we want to do an instanceof comparison against the AESCrypt class 6574 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6575 if (!klass_AESCrypt->is_loaded()) { 6576 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6577 Node* ctrl = control(); 6578 set_control(top()); // no regular fast path 6579 return ctrl; 6580 } 6581 6582 src = must_be_not_null(src, true); 6583 dest = must_be_not_null(dest, true); 6584 6585 // Resolve oops to stable for CmpP below. 6586 src = access_resolve(src, 0); 6587 dest = access_resolve(dest, 0); 6588 6589 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6590 6591 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6592 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6593 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6594 6595 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6596 6597 // for encryption, we are done 6598 if (!decrypting) 6599 return instof_false; // even if it is NULL 6600 6601 // for decryption, we need to add a further check to avoid 6602 // taking the intrinsic path when cipher and plain are the same 6603 // see the original java code for why. 6604 RegionNode* region = new RegionNode(3); 6605 region->init_req(1, instof_false); 6606 6607 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6608 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6609 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6610 region->init_req(2, src_dest_conjoint); 6611 6612 record_for_igvn(region); 6613 return _gvn.transform(region); 6614 } 6615 6616 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6617 // Return node representing slow path of predicate check. 6618 // the pseudo code we want to emulate with this predicate is: 6619 // for encryption: 6620 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6621 // for decryption: 6622 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6623 // note cipher==plain is more conservative than the original java code but that's OK 6624 // 6625 6626 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6627 // The receiver was checked for NULL already. 6628 Node* objCTR = argument(0); 6629 6630 // Load embeddedCipher field of CipherBlockChaining object. 6631 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6632 6633 // get AESCrypt klass for instanceOf check 6634 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6635 // will have same classloader as CipherBlockChaining object 6636 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6637 assert(tinst != NULL, "CTRobj is null"); 6638 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6639 6640 // we want to do an instanceof comparison against the AESCrypt class 6641 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6642 if (!klass_AESCrypt->is_loaded()) { 6643 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6644 Node* ctrl = control(); 6645 set_control(top()); // no regular fast path 6646 return ctrl; 6647 } 6648 6649 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6650 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6651 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6652 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6653 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6654 6655 return instof_false; // even if it is NULL 6656 } 6657 6658 //------------------------------inline_ghash_processBlocks 6659 bool LibraryCallKit::inline_ghash_processBlocks() { 6660 address stubAddr; 6661 const char *stubName; 6662 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6663 6664 stubAddr = StubRoutines::ghash_processBlocks(); 6665 stubName = "ghash_processBlocks"; 6666 6667 Node* data = argument(0); 6668 Node* offset = argument(1); 6669 Node* len = argument(2); 6670 Node* state = argument(3); 6671 Node* subkeyH = argument(4); 6672 6673 state = must_be_not_null(state, true); 6674 subkeyH = must_be_not_null(subkeyH, true); 6675 data = must_be_not_null(data, true); 6676 6677 state = access_resolve(state, ACCESS_WRITE); 6678 subkeyH = access_resolve(subkeyH, ACCESS_READ); 6679 data = access_resolve(data, ACCESS_READ); 6680 6681 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6682 assert(state_start, "state is NULL"); 6683 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6684 assert(subkeyH_start, "subkeyH is NULL"); 6685 Node* data_start = array_element_address(data, offset, T_BYTE); 6686 assert(data_start, "data is NULL"); 6687 6688 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6689 OptoRuntime::ghash_processBlocks_Type(), 6690 stubAddr, stubName, TypePtr::BOTTOM, 6691 state_start, subkeyH_start, data_start, len); 6692 return true; 6693 } 6694 6695 bool LibraryCallKit::inline_base64_encodeBlock() { 6696 address stubAddr; 6697 const char *stubName; 6698 assert(UseBASE64Intrinsics, "need Base64 intrinsics support"); 6699 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters"); 6700 stubAddr = StubRoutines::base64_encodeBlock(); 6701 stubName = "encodeBlock"; 6702 6703 if (!stubAddr) return false; 6704 Node* base64obj = argument(0); 6705 Node* src = argument(1); 6706 Node* offset = argument(2); 6707 Node* len = argument(3); 6708 Node* dest = argument(4); 6709 Node* dp = argument(5); 6710 Node* isURL = argument(6); 6711 6712 src = must_be_not_null(src, true); 6713 src = access_resolve(src, ACCESS_READ); 6714 dest = must_be_not_null(dest, true); 6715 dest = access_resolve(dest, ACCESS_WRITE); 6716 6717 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 6718 assert(src_start, "source array is NULL"); 6719 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE); 6720 assert(dest_start, "destination array is NULL"); 6721 6722 Node* base64 = make_runtime_call(RC_LEAF, 6723 OptoRuntime::base64_encodeBlock_Type(), 6724 stubAddr, stubName, TypePtr::BOTTOM, 6725 src_start, offset, len, dest_start, dp, isURL); 6726 return true; 6727 } 6728 6729 //------------------------------inline_sha_implCompress----------------------- 6730 // 6731 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6732 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6733 // 6734 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6735 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6736 // 6737 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6738 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6739 // 6740 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6741 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6742 6743 Node* sha_obj = argument(0); 6744 Node* src = argument(1); // type oop 6745 Node* ofs = argument(2); // type int 6746 6747 const Type* src_type = src->Value(&_gvn); 6748 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6749 if (top_src == NULL || top_src->klass() == NULL) { 6750 // failed array check 6751 return false; 6752 } 6753 // Figure out the size and type of the elements we will be copying. 6754 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6755 if (src_elem != T_BYTE) { 6756 return false; 6757 } 6758 // 'src_start' points to src array + offset 6759 src = must_be_not_null(src, true); 6760 src = access_resolve(src, ACCESS_READ); 6761 Node* src_start = array_element_address(src, ofs, src_elem); 6762 Node* state = NULL; 6763 address stubAddr; 6764 const char *stubName; 6765 6766 switch(id) { 6767 case vmIntrinsics::_sha_implCompress: 6768 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6769 state = get_state_from_sha_object(sha_obj); 6770 stubAddr = StubRoutines::sha1_implCompress(); 6771 stubName = "sha1_implCompress"; 6772 break; 6773 case vmIntrinsics::_sha2_implCompress: 6774 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6775 state = get_state_from_sha_object(sha_obj); 6776 stubAddr = StubRoutines::sha256_implCompress(); 6777 stubName = "sha256_implCompress"; 6778 break; 6779 case vmIntrinsics::_sha5_implCompress: 6780 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6781 state = get_state_from_sha5_object(sha_obj); 6782 stubAddr = StubRoutines::sha512_implCompress(); 6783 stubName = "sha512_implCompress"; 6784 break; 6785 default: 6786 fatal_unexpected_iid(id); 6787 return false; 6788 } 6789 if (state == NULL) return false; 6790 6791 assert(stubAddr != NULL, "Stub is generated"); 6792 if (stubAddr == NULL) return false; 6793 6794 // Call the stub. 6795 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6796 stubAddr, stubName, TypePtr::BOTTOM, 6797 src_start, state); 6798 6799 return true; 6800 } 6801 6802 //------------------------------inline_digestBase_implCompressMB----------------------- 6803 // 6804 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6805 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6806 // 6807 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6808 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6809 "need SHA1/SHA256/SHA512 instruction support"); 6810 assert((uint)predicate < 3, "sanity"); 6811 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6812 6813 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6814 Node* src = argument(1); // byte[] array 6815 Node* ofs = argument(2); // type int 6816 Node* limit = argument(3); // type int 6817 6818 const Type* src_type = src->Value(&_gvn); 6819 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6820 if (top_src == NULL || top_src->klass() == NULL) { 6821 // failed array check 6822 return false; 6823 } 6824 // Figure out the size and type of the elements we will be copying. 6825 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6826 if (src_elem != T_BYTE) { 6827 return false; 6828 } 6829 // 'src_start' points to src array + offset 6830 src = must_be_not_null(src, false); 6831 src = access_resolve(src, ACCESS_READ); 6832 Node* src_start = array_element_address(src, ofs, src_elem); 6833 6834 const char* klass_SHA_name = NULL; 6835 const char* stub_name = NULL; 6836 address stub_addr = NULL; 6837 bool long_state = false; 6838 6839 switch (predicate) { 6840 case 0: 6841 if (UseSHA1Intrinsics) { 6842 klass_SHA_name = "sun/security/provider/SHA"; 6843 stub_name = "sha1_implCompressMB"; 6844 stub_addr = StubRoutines::sha1_implCompressMB(); 6845 } 6846 break; 6847 case 1: 6848 if (UseSHA256Intrinsics) { 6849 klass_SHA_name = "sun/security/provider/SHA2"; 6850 stub_name = "sha256_implCompressMB"; 6851 stub_addr = StubRoutines::sha256_implCompressMB(); 6852 } 6853 break; 6854 case 2: 6855 if (UseSHA512Intrinsics) { 6856 klass_SHA_name = "sun/security/provider/SHA5"; 6857 stub_name = "sha512_implCompressMB"; 6858 stub_addr = StubRoutines::sha512_implCompressMB(); 6859 long_state = true; 6860 } 6861 break; 6862 default: 6863 fatal("unknown SHA intrinsic predicate: %d", predicate); 6864 } 6865 if (klass_SHA_name != NULL) { 6866 assert(stub_addr != NULL, "Stub is generated"); 6867 if (stub_addr == NULL) return false; 6868 6869 // get DigestBase klass to lookup for SHA klass 6870 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6871 assert(tinst != NULL, "digestBase_obj is not instance???"); 6872 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6873 6874 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6875 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6876 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6877 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6878 } 6879 return false; 6880 } 6881 //------------------------------inline_sha_implCompressMB----------------------- 6882 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6883 bool long_state, address stubAddr, const char *stubName, 6884 Node* src_start, Node* ofs, Node* limit) { 6885 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6886 const TypeOopPtr* xtype = aklass->as_instance_type(); 6887 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6888 sha_obj = _gvn.transform(sha_obj); 6889 6890 Node* state; 6891 if (long_state) { 6892 state = get_state_from_sha5_object(sha_obj); 6893 } else { 6894 state = get_state_from_sha_object(sha_obj); 6895 } 6896 if (state == NULL) return false; 6897 6898 // Call the stub. 6899 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6900 OptoRuntime::digestBase_implCompressMB_Type(), 6901 stubAddr, stubName, TypePtr::BOTTOM, 6902 src_start, state, ofs, limit); 6903 // return ofs (int) 6904 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6905 set_result(result); 6906 6907 return true; 6908 } 6909 6910 //------------------------------get_state_from_sha_object----------------------- 6911 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6912 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6913 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6914 if (sha_state == NULL) return (Node *) NULL; 6915 6916 // now have the array, need to get the start address of the state array 6917 sha_state = access_resolve(sha_state, ACCESS_WRITE); 6918 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6919 return state; 6920 } 6921 6922 //------------------------------get_state_from_sha5_object----------------------- 6923 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6924 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6925 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6926 if (sha_state == NULL) return (Node *) NULL; 6927 6928 // now have the array, need to get the start address of the state array 6929 sha_state = access_resolve(sha_state, ACCESS_WRITE); 6930 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6931 return state; 6932 } 6933 6934 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6935 // Return node representing slow path of predicate check. 6936 // the pseudo code we want to emulate with this predicate is: 6937 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6938 // 6939 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6940 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6941 "need SHA1/SHA256/SHA512 instruction support"); 6942 assert((uint)predicate < 3, "sanity"); 6943 6944 // The receiver was checked for NULL already. 6945 Node* digestBaseObj = argument(0); 6946 6947 // get DigestBase klass for instanceOf check 6948 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6949 assert(tinst != NULL, "digestBaseObj is null"); 6950 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6951 6952 const char* klass_SHA_name = NULL; 6953 switch (predicate) { 6954 case 0: 6955 if (UseSHA1Intrinsics) { 6956 // we want to do an instanceof comparison against the SHA class 6957 klass_SHA_name = "sun/security/provider/SHA"; 6958 } 6959 break; 6960 case 1: 6961 if (UseSHA256Intrinsics) { 6962 // we want to do an instanceof comparison against the SHA2 class 6963 klass_SHA_name = "sun/security/provider/SHA2"; 6964 } 6965 break; 6966 case 2: 6967 if (UseSHA512Intrinsics) { 6968 // we want to do an instanceof comparison against the SHA5 class 6969 klass_SHA_name = "sun/security/provider/SHA5"; 6970 } 6971 break; 6972 default: 6973 fatal("unknown SHA intrinsic predicate: %d", predicate); 6974 } 6975 6976 ciKlass* klass_SHA = NULL; 6977 if (klass_SHA_name != NULL) { 6978 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6979 } 6980 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6981 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6982 Node* ctrl = control(); 6983 set_control(top()); // no intrinsic path 6984 return ctrl; 6985 } 6986 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6987 6988 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6989 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6990 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6991 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6992 6993 return instof_false; // even if it is NULL 6994 } 6995 6996 //-------------inline_fma----------------------------------- 6997 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) { 6998 Node *a = NULL; 6999 Node *b = NULL; 7000 Node *c = NULL; 7001 Node* result = NULL; 7002 switch (id) { 7003 case vmIntrinsics::_fmaD: 7004 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each."); 7005 // no receiver since it is static method 7006 a = round_double_node(argument(0)); 7007 b = round_double_node(argument(2)); 7008 c = round_double_node(argument(4)); 7009 result = _gvn.transform(new FmaDNode(control(), a, b, c)); 7010 break; 7011 case vmIntrinsics::_fmaF: 7012 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each."); 7013 a = argument(0); 7014 b = argument(1); 7015 c = argument(2); 7016 result = _gvn.transform(new FmaFNode(control(), a, b, c)); 7017 break; 7018 default: 7019 fatal_unexpected_iid(id); break; 7020 } 7021 set_result(result); 7022 return true; 7023 } 7024 7025 bool LibraryCallKit::inline_character_compare(vmIntrinsics::ID id) { 7026 // argument(0) is receiver 7027 Node* codePoint = argument(1); 7028 Node* n = NULL; 7029 7030 switch (id) { 7031 case vmIntrinsics::_isDigit : 7032 n = new DigitNode(control(), codePoint); 7033 break; 7034 case vmIntrinsics::_isLowerCase : 7035 n = new LowerCaseNode(control(), codePoint); 7036 break; 7037 case vmIntrinsics::_isUpperCase : 7038 n = new UpperCaseNode(control(), codePoint); 7039 break; 7040 case vmIntrinsics::_isWhitespace : 7041 n = new WhitespaceNode(control(), codePoint); 7042 break; 7043 default: 7044 fatal_unexpected_iid(id); 7045 } 7046 7047 set_result(_gvn.transform(n)); 7048 return true; 7049 } 7050 7051 //------------------------------inline_fp_min_max------------------------------ 7052 bool LibraryCallKit::inline_fp_min_max(vmIntrinsics::ID id) { 7053 /* DISABLED BECAUSE METHOD DATA ISN'T COLLECTED PER CALL-SITE, SEE JDK-8015416. 7054 7055 // The intrinsic should be used only when the API branches aren't predictable, 7056 // the last one performing the most important comparison. The following heuristic 7057 // uses the branch statistics to eventually bail out if necessary. 7058 7059 ciMethodData *md = callee()->method_data(); 7060 7061 if ( md != NULL && md->is_mature() && md->invocation_count() > 0 ) { 7062 ciCallProfile cp = caller()->call_profile_at_bci(bci()); 7063 7064 if ( ((double)cp.count()) / ((double)md->invocation_count()) < 0.8 ) { 7065 // Bail out if the call-site didn't contribute enough to the statistics. 7066 return false; 7067 } 7068 7069 uint taken = 0, not_taken = 0; 7070 7071 for (ciProfileData *p = md->first_data(); md->is_valid(p); p = md->next_data(p)) { 7072 if (p->is_BranchData()) { 7073 taken = ((ciBranchData*)p)->taken(); 7074 not_taken = ((ciBranchData*)p)->not_taken(); 7075 } 7076 } 7077 7078 double balance = (((double)taken) - ((double)not_taken)) / ((double)md->invocation_count()); 7079 balance = balance < 0 ? -balance : balance; 7080 if ( balance > 0.2 ) { 7081 // Bail out if the most important branch is predictable enough. 7082 return false; 7083 } 7084 } 7085 */ 7086 7087 Node *a = NULL; 7088 Node *b = NULL; 7089 Node *n = NULL; 7090 switch (id) { 7091 case vmIntrinsics::_maxF: 7092 case vmIntrinsics::_minF: 7093 assert(callee()->signature()->size() == 2, "minF/maxF has 2 parameters of size 1 each."); 7094 a = argument(0); 7095 b = argument(1); 7096 break; 7097 case vmIntrinsics::_maxD: 7098 case vmIntrinsics::_minD: 7099 assert(callee()->signature()->size() == 4, "minD/maxD has 2 parameters of size 2 each."); 7100 a = round_double_node(argument(0)); 7101 b = round_double_node(argument(2)); 7102 break; 7103 default: 7104 fatal_unexpected_iid(id); 7105 break; 7106 } 7107 if (a->is_Con() || b->is_Con()) { 7108 return false; 7109 } 7110 switch (id) { 7111 case vmIntrinsics::_maxF: n = new MaxFNode(a, b); break; 7112 case vmIntrinsics::_minF: n = new MinFNode(a, b); break; 7113 case vmIntrinsics::_maxD: n = new MaxDNode(a, b); break; 7114 case vmIntrinsics::_minD: n = new MinDNode(a, b); break; 7115 default: fatal_unexpected_iid(id); break; 7116 } 7117 set_result(_gvn.transform(n)); 7118 return true; 7119 } 7120 7121 bool LibraryCallKit::inline_profileBoolean() { 7122 Node* counts = argument(1); 7123 const TypeAryPtr* ary = NULL; 7124 ciArray* aobj = NULL; 7125 if (counts->is_Con() 7126 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 7127 && (aobj = ary->const_oop()->as_array()) != NULL 7128 && (aobj->length() == 2)) { 7129 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 7130 jint false_cnt = aobj->element_value(0).as_int(); 7131 jint true_cnt = aobj->element_value(1).as_int(); 7132 7133 if (C->log() != NULL) { 7134 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 7135 false_cnt, true_cnt); 7136 } 7137 7138 if (false_cnt + true_cnt == 0) { 7139 // According to profile, never executed. 7140 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 7141 Deoptimization::Action_reinterpret); 7142 return true; 7143 } 7144 7145 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 7146 // is a number of each value occurrences. 7147 Node* result = argument(0); 7148 if (false_cnt == 0 || true_cnt == 0) { 7149 // According to profile, one value has been never seen. 7150 int expected_val = (false_cnt == 0) ? 1 : 0; 7151 7152 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 7153 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 7154 7155 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 7156 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 7157 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 7158 7159 { // Slow path: uncommon trap for never seen value and then reexecute 7160 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 7161 // the value has been seen at least once. 7162 PreserveJVMState pjvms(this); 7163 PreserveReexecuteState preexecs(this); 7164 jvms()->set_should_reexecute(true); 7165 7166 set_control(slow_path); 7167 set_i_o(i_o()); 7168 7169 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 7170 Deoptimization::Action_reinterpret); 7171 } 7172 // The guard for never seen value enables sharpening of the result and 7173 // returning a constant. It allows to eliminate branches on the same value 7174 // later on. 7175 set_control(fast_path); 7176 result = intcon(expected_val); 7177 } 7178 // Stop profiling. 7179 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 7180 // By replacing method body with profile data (represented as ProfileBooleanNode 7181 // on IR level) we effectively disable profiling. 7182 // It enables full speed execution once optimized code is generated. 7183 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 7184 C->record_for_igvn(profile); 7185 set_result(profile); 7186 return true; 7187 } else { 7188 // Continue profiling. 7189 // Profile data isn't available at the moment. So, execute method's bytecode version. 7190 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 7191 // is compiled and counters aren't available since corresponding MethodHandle 7192 // isn't a compile-time constant. 7193 return false; 7194 } 7195 } 7196 7197 bool LibraryCallKit::inline_isCompileConstant() { 7198 Node* n = argument(0); 7199 set_result(n->is_Con() ? intcon(1) : intcon(0)); 7200 return true; 7201 }