1 /* 2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "oops/objArrayKlass.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/c2compiler.hpp" 36 #include "opto/callGenerator.hpp" 37 #include "opto/castnode.hpp" 38 #include "opto/cfgnode.hpp" 39 #include "opto/convertnode.hpp" 40 #include "opto/countbitsnode.hpp" 41 #include "opto/intrinsicnode.hpp" 42 #include "opto/idealKit.hpp" 43 #include "opto/mathexactnode.hpp" 44 #include "opto/movenode.hpp" 45 #include "opto/mulnode.hpp" 46 #include "opto/narrowptrnode.hpp" 47 #include "opto/opaquenode.hpp" 48 #include "opto/parse.hpp" 49 #include "opto/runtime.hpp" 50 #include "opto/rootnode.hpp" 51 #include "opto/subnode.hpp" 52 #include "prims/nativeLookup.hpp" 53 #include "prims/unsafe.hpp" 54 #include "runtime/sharedRuntime.hpp" 55 #ifdef TRACE_HAVE_INTRINSICS 56 #include "trace/traceMacros.hpp" 57 #endif 58 59 class LibraryIntrinsic : public InlineCallGenerator { 60 // Extend the set of intrinsics known to the runtime: 61 public: 62 private: 63 bool _is_virtual; 64 bool _does_virtual_dispatch; 65 int8_t _predicates_count; // Intrinsic is predicated by several conditions 66 int8_t _last_predicate; // Last generated predicate 67 vmIntrinsics::ID _intrinsic_id; 68 69 public: 70 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 71 : InlineCallGenerator(m), 72 _is_virtual(is_virtual), 73 _does_virtual_dispatch(does_virtual_dispatch), 74 _predicates_count((int8_t)predicates_count), 75 _last_predicate((int8_t)-1), 76 _intrinsic_id(id) 77 { 78 } 79 virtual bool is_intrinsic() const { return true; } 80 virtual bool is_virtual() const { return _is_virtual; } 81 virtual bool is_predicated() const { return _predicates_count > 0; } 82 virtual int predicates_count() const { return _predicates_count; } 83 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 84 virtual JVMState* generate(JVMState* jvms); 85 virtual Node* generate_predicate(JVMState* jvms, int predicate); 86 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 87 }; 88 89 90 // Local helper class for LibraryIntrinsic: 91 class LibraryCallKit : public GraphKit { 92 private: 93 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 94 Node* _result; // the result node, if any 95 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 96 97 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type); 98 99 public: 100 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 101 : GraphKit(jvms), 102 _intrinsic(intrinsic), 103 _result(NULL) 104 { 105 // Check if this is a root compile. In that case we don't have a caller. 106 if (!jvms->has_method()) { 107 _reexecute_sp = sp(); 108 } else { 109 // Find out how many arguments the interpreter needs when deoptimizing 110 // and save the stack pointer value so it can used by uncommon_trap. 111 // We find the argument count by looking at the declared signature. 112 bool ignored_will_link; 113 ciSignature* declared_signature = NULL; 114 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 115 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 116 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 117 } 118 } 119 120 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 121 122 ciMethod* caller() const { return jvms()->method(); } 123 int bci() const { return jvms()->bci(); } 124 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 125 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 126 ciMethod* callee() const { return _intrinsic->method(); } 127 128 bool try_to_inline(int predicate); 129 Node* try_to_predicate(int predicate); 130 131 void push_result() { 132 // Push the result onto the stack. 133 if (!stopped() && result() != NULL) { 134 BasicType bt = result()->bottom_type()->basic_type(); 135 push_node(bt, result()); 136 } 137 } 138 139 private: 140 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 141 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 142 } 143 144 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 145 void set_result(RegionNode* region, PhiNode* value); 146 Node* result() { return _result; } 147 148 virtual int reexecute_sp() { return _reexecute_sp; } 149 150 // Helper functions to inline natives 151 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 152 Node* generate_slow_guard(Node* test, RegionNode* region); 153 Node* generate_fair_guard(Node* test, RegionNode* region); 154 Node* generate_negative_guard(Node* index, RegionNode* region, 155 // resulting CastII of index: 156 Node* *pos_index = NULL); 157 Node* generate_limit_guard(Node* offset, Node* subseq_length, 158 Node* array_length, 159 RegionNode* region); 160 void generate_string_range_check(Node* array, Node* offset, 161 Node* length, bool char_count); 162 Node* generate_current_thread(Node* &tls_output); 163 Node* load_mirror_from_klass(Node* klass); 164 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 165 RegionNode* region, int null_path, 166 int offset); 167 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 168 RegionNode* region, int null_path) { 169 int offset = java_lang_Class::klass_offset_in_bytes(); 170 return load_klass_from_mirror_common(mirror, never_see_null, 171 region, null_path, 172 offset); 173 } 174 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 175 RegionNode* region, int null_path) { 176 int offset = java_lang_Class::array_klass_offset_in_bytes(); 177 return load_klass_from_mirror_common(mirror, never_see_null, 178 region, null_path, 179 offset); 180 } 181 Node* generate_access_flags_guard(Node* kls, 182 int modifier_mask, int modifier_bits, 183 RegionNode* region); 184 Node* generate_interface_guard(Node* kls, RegionNode* region); 185 Node* generate_array_guard(Node* kls, RegionNode* region) { 186 return generate_array_guard_common(kls, region, false, false); 187 } 188 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 189 return generate_array_guard_common(kls, region, false, true); 190 } 191 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 192 return generate_array_guard_common(kls, region, true, false); 193 } 194 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 195 return generate_array_guard_common(kls, region, true, true); 196 } 197 Node* generate_array_guard_common(Node* kls, RegionNode* region, 198 bool obj_array, bool not_array); 199 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 200 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 201 bool is_virtual = false, bool is_static = false); 202 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 203 return generate_method_call(method_id, false, true); 204 } 205 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 206 return generate_method_call(method_id, true, false); 207 } 208 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 209 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 210 211 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 212 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 213 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 214 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 215 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 216 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); 217 bool inline_string_indexOfChar(); 218 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 219 bool inline_string_toBytesU(); 220 bool inline_string_getCharsU(); 221 bool inline_string_copy(bool compress); 222 bool inline_string_char_access(bool is_store); 223 Node* round_double_node(Node* n); 224 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 225 bool inline_math_native(vmIntrinsics::ID id); 226 bool inline_math(vmIntrinsics::ID id); 227 template <typename OverflowOp> 228 bool inline_math_overflow(Node* arg1, Node* arg2); 229 void inline_math_mathExact(Node* math, Node* test); 230 bool inline_math_addExactI(bool is_increment); 231 bool inline_math_addExactL(bool is_increment); 232 bool inline_math_multiplyExactI(); 233 bool inline_math_multiplyExactL(); 234 bool inline_math_multiplyHigh(); 235 bool inline_math_negateExactI(); 236 bool inline_math_negateExactL(); 237 bool inline_math_subtractExactI(bool is_decrement); 238 bool inline_math_subtractExactL(bool is_decrement); 239 bool inline_min_max(vmIntrinsics::ID id); 240 bool inline_notify(vmIntrinsics::ID id); 241 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 242 // This returns Type::AnyPtr, RawPtr, or OopPtr. 243 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type); 244 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false); 245 // Helper for inline_unsafe_access. 246 // Generates the guards that check whether the result of 247 // Unsafe.getObject should be recorded in an SATB log buffer. 248 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 249 250 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind; 251 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned); 252 static bool klass_needs_init_guard(Node* kls); 253 bool inline_unsafe_allocate(); 254 bool inline_unsafe_newArray(bool uninitialized); 255 bool inline_unsafe_copyMemory(); 256 bool inline_native_currentThread(); 257 258 bool inline_native_time_funcs(address method, const char* funcName); 259 #ifdef TRACE_HAVE_INTRINSICS 260 bool inline_native_classID(); 261 bool inline_native_getBufferWriter(); 262 #endif 263 bool inline_native_isInterrupted(); 264 bool inline_native_Class_query(vmIntrinsics::ID id); 265 bool inline_native_subtype_check(); 266 bool inline_native_getLength(); 267 bool inline_array_copyOf(bool is_copyOfRange); 268 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 269 bool inline_preconditions_checkIndex(); 270 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 271 bool inline_native_clone(bool is_virtual); 272 bool inline_native_Reflection_getCallerClass(); 273 // Helper function for inlining native object hash method 274 bool inline_native_hashcode(bool is_virtual, bool is_static); 275 bool inline_native_getClass(); 276 277 // Helper functions for inlining arraycopy 278 bool inline_arraycopy(); 279 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 280 RegionNode* slow_region); 281 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 282 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp, 283 uint new_idx); 284 285 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind; 286 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store); 287 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind); 288 bool loadstore_requires_writeback_barrier(LoadStoreKind kind); 289 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind); 290 Node* make_cas_loadbarrier(CompareAndSwapNode* cas); 291 Node* make_cmpx_loadbarrier(CompareAndExchangePNode* cmpx); 292 bool inline_unsafe_fence(vmIntrinsics::ID id); 293 bool inline_onspinwait(); 294 bool inline_fp_conversions(vmIntrinsics::ID id); 295 bool inline_number_methods(vmIntrinsics::ID id); 296 bool inline_reference_get(); 297 bool inline_Class_cast(); 298 bool inline_aescrypt_Block(vmIntrinsics::ID id); 299 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 300 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 301 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 302 Node* inline_counterMode_AESCrypt_predicate(); 303 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 304 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 305 bool inline_ghash_processBlocks(); 306 bool inline_sha_implCompress(vmIntrinsics::ID id); 307 bool inline_digestBase_implCompressMB(int predicate); 308 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 309 bool long_state, address stubAddr, const char *stubName, 310 Node* src_start, Node* ofs, Node* limit); 311 Node* get_state_from_sha_object(Node *sha_object); 312 Node* get_state_from_sha5_object(Node *sha_object); 313 Node* inline_digestBase_implCompressMB_predicate(int predicate); 314 bool inline_encodeISOArray(); 315 bool inline_updateCRC32(); 316 bool inline_updateBytesCRC32(); 317 bool inline_updateByteBufferCRC32(); 318 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 319 bool inline_updateBytesCRC32C(); 320 bool inline_updateDirectByteBufferCRC32C(); 321 bool inline_updateBytesAdler32(); 322 bool inline_updateByteBufferAdler32(); 323 bool inline_multiplyToLen(); 324 bool inline_hasNegatives(); 325 bool inline_squareToLen(); 326 bool inline_mulAdd(); 327 bool inline_montgomeryMultiply(); 328 bool inline_montgomerySquare(); 329 bool inline_vectorizedMismatch(); 330 bool inline_fma(vmIntrinsics::ID id); 331 332 bool inline_profileBoolean(); 333 bool inline_isCompileConstant(); 334 void clear_upper_avx() { 335 #ifdef X86 336 if (UseAVX >= 2) { 337 C->set_clear_upper_avx(true); 338 } 339 #endif 340 } 341 }; 342 343 //---------------------------make_vm_intrinsic---------------------------- 344 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 345 vmIntrinsics::ID id = m->intrinsic_id(); 346 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 347 348 if (!m->is_loaded()) { 349 // Do not attempt to inline unloaded methods. 350 return NULL; 351 } 352 353 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 354 bool is_available = false; 355 356 { 357 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 358 // the compiler must transition to '_thread_in_vm' state because both 359 // methods access VM-internal data. 360 VM_ENTRY_MARK; 361 methodHandle mh(THREAD, m->get_Method()); 362 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) && 363 !C->directive()->is_intrinsic_disabled(mh) && 364 !vmIntrinsics::is_disabled_by_flags(mh); 365 366 } 367 368 if (is_available) { 369 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 370 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 371 return new LibraryIntrinsic(m, is_virtual, 372 vmIntrinsics::predicates_needed(id), 373 vmIntrinsics::does_virtual_dispatch(id), 374 (vmIntrinsics::ID) id); 375 } else { 376 return NULL; 377 } 378 } 379 380 //----------------------register_library_intrinsics----------------------- 381 // Initialize this file's data structures, for each Compile instance. 382 void Compile::register_library_intrinsics() { 383 // Nothing to do here. 384 } 385 386 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 387 LibraryCallKit kit(jvms, this); 388 Compile* C = kit.C; 389 int nodes = C->unique(); 390 #ifndef PRODUCT 391 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 392 char buf[1000]; 393 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 394 tty->print_cr("Intrinsic %s", str); 395 } 396 #endif 397 ciMethod* callee = kit.callee(); 398 const int bci = kit.bci(); 399 400 // Try to inline the intrinsic. 401 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 402 kit.try_to_inline(_last_predicate)) { 403 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)" 404 : "(intrinsic)"; 405 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 406 if (C->print_intrinsics() || C->print_inlining()) { 407 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 408 } 409 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 410 if (C->log()) { 411 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 412 vmIntrinsics::name_at(intrinsic_id()), 413 (is_virtual() ? " virtual='1'" : ""), 414 C->unique() - nodes); 415 } 416 // Push the result from the inlined method onto the stack. 417 kit.push_result(); 418 C->print_inlining_update(this); 419 return kit.transfer_exceptions_into_jvms(); 420 } 421 422 // The intrinsic bailed out 423 if (jvms->has_method()) { 424 // Not a root compile. 425 const char* msg; 426 if (callee->intrinsic_candidate()) { 427 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 428 } else { 429 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 430 : "failed to inline (intrinsic), method not annotated"; 431 } 432 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg); 433 if (C->print_intrinsics() || C->print_inlining()) { 434 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 435 } 436 } else { 437 // Root compile 438 ResourceMark rm; 439 stringStream msg_stream; 440 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 441 vmIntrinsics::name_at(intrinsic_id()), 442 is_virtual() ? " (virtual)" : "", bci); 443 const char *msg = msg_stream.as_string(); 444 log_debug(jit, inlining)("%s", msg); 445 if (C->print_intrinsics() || C->print_inlining()) { 446 tty->print("%s", msg); 447 } 448 } 449 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 450 C->print_inlining_update(this); 451 return NULL; 452 } 453 454 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 455 LibraryCallKit kit(jvms, this); 456 Compile* C = kit.C; 457 int nodes = C->unique(); 458 _last_predicate = predicate; 459 #ifndef PRODUCT 460 assert(is_predicated() && predicate < predicates_count(), "sanity"); 461 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 462 char buf[1000]; 463 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 464 tty->print_cr("Predicate for intrinsic %s", str); 465 } 466 #endif 467 ciMethod* callee = kit.callee(); 468 const int bci = kit.bci(); 469 470 Node* slow_ctl = kit.try_to_predicate(predicate); 471 if (!kit.failing()) { 472 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)" 473 : "(intrinsic, predicate)"; 474 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 475 if (C->print_intrinsics() || C->print_inlining()) { 476 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 477 } 478 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 479 if (C->log()) { 480 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 481 vmIntrinsics::name_at(intrinsic_id()), 482 (is_virtual() ? " virtual='1'" : ""), 483 C->unique() - nodes); 484 } 485 return slow_ctl; // Could be NULL if the check folds. 486 } 487 488 // The intrinsic bailed out 489 if (jvms->has_method()) { 490 // Not a root compile. 491 const char* msg = "failed to generate predicate for intrinsic"; 492 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg); 493 if (C->print_intrinsics() || C->print_inlining()) { 494 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 495 } 496 } else { 497 // Root compile 498 ResourceMark rm; 499 stringStream msg_stream; 500 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 501 vmIntrinsics::name_at(intrinsic_id()), 502 is_virtual() ? " (virtual)" : "", bci); 503 const char *msg = msg_stream.as_string(); 504 log_debug(jit, inlining)("%s", msg); 505 if (C->print_intrinsics() || C->print_inlining()) { 506 C->print_inlining_stream()->print("%s", msg); 507 } 508 } 509 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 510 return NULL; 511 } 512 513 bool LibraryCallKit::try_to_inline(int predicate) { 514 // Handle symbolic names for otherwise undistinguished boolean switches: 515 const bool is_store = true; 516 const bool is_compress = true; 517 const bool is_static = true; 518 const bool is_volatile = true; 519 520 if (!jvms()->has_method()) { 521 // Root JVMState has a null method. 522 assert(map()->memory()->Opcode() == Op_Parm, ""); 523 // Insert the memory aliasing node 524 set_all_memory(reset_memory()); 525 } 526 assert(merged_memory(), ""); 527 528 529 switch (intrinsic_id()) { 530 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 531 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 532 case vmIntrinsics::_getClass: return inline_native_getClass(); 533 534 case vmIntrinsics::_dsin: 535 case vmIntrinsics::_dcos: 536 case vmIntrinsics::_dtan: 537 case vmIntrinsics::_dabs: 538 case vmIntrinsics::_datan2: 539 case vmIntrinsics::_dsqrt: 540 case vmIntrinsics::_dexp: 541 case vmIntrinsics::_dlog: 542 case vmIntrinsics::_dlog10: 543 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 544 545 case vmIntrinsics::_min: 546 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 547 548 case vmIntrinsics::_notify: 549 case vmIntrinsics::_notifyAll: 550 if (InlineNotify) { 551 return inline_notify(intrinsic_id()); 552 } 553 return false; 554 555 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 556 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 557 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 558 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 559 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 560 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 561 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 562 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 563 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh(); 564 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 565 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 566 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 567 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 568 569 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 570 571 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 572 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 573 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 574 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 575 576 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 577 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 578 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 579 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 580 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 581 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 582 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 583 584 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 585 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 586 587 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 588 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 589 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 590 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 591 592 case vmIntrinsics::_compressStringC: 593 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 594 case vmIntrinsics::_inflateStringC: 595 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 596 597 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 598 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 599 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 600 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 601 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 602 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 603 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 604 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 605 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 606 607 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 608 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 609 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 610 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 611 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 612 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 613 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 614 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 615 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 616 617 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 618 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 619 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 620 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 621 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 622 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 623 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 624 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 625 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 626 627 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 628 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 629 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 630 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 631 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 632 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 633 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 634 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 635 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 636 637 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 638 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 639 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 640 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 641 642 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 643 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 644 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 645 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 646 647 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 648 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 649 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 650 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 651 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 652 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 653 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 654 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 655 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 656 657 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 658 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 659 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 660 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 661 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 662 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 663 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 664 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 665 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 666 667 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 668 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 669 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 670 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 671 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 672 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 673 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 674 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 675 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 676 677 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 678 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 679 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 680 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 681 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 682 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 683 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 684 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 685 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 686 687 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 688 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); 689 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); 690 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 691 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 692 693 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 694 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 695 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 696 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 697 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); 698 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); 699 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); 700 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); 701 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); 702 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); 703 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); 704 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); 705 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 706 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 707 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 708 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 709 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 710 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 711 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 712 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 713 714 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 715 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 716 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 717 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); 718 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); 719 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); 720 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); 721 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); 722 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); 723 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 724 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 725 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 726 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 727 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 728 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 729 730 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); 731 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); 732 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 733 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 734 735 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); 736 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); 737 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 738 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 739 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 740 741 case vmIntrinsics::_loadFence: 742 case vmIntrinsics::_storeFence: 743 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 744 745 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 746 747 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 748 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 749 750 #ifdef TRACE_HAVE_INTRINSICS 751 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 752 case vmIntrinsics::_getClassId: return inline_native_classID(); 753 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter(); 754 #endif 755 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 756 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 757 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 758 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 759 case vmIntrinsics::_getLength: return inline_native_getLength(); 760 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 761 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 762 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 763 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 764 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 765 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 766 767 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 768 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 769 770 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 771 772 case vmIntrinsics::_isInstance: 773 case vmIntrinsics::_getModifiers: 774 case vmIntrinsics::_isInterface: 775 case vmIntrinsics::_isArray: 776 case vmIntrinsics::_isPrimitive: 777 case vmIntrinsics::_getSuperclass: 778 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 779 780 case vmIntrinsics::_floatToRawIntBits: 781 case vmIntrinsics::_floatToIntBits: 782 case vmIntrinsics::_intBitsToFloat: 783 case vmIntrinsics::_doubleToRawLongBits: 784 case vmIntrinsics::_doubleToLongBits: 785 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 786 787 case vmIntrinsics::_numberOfLeadingZeros_i: 788 case vmIntrinsics::_numberOfLeadingZeros_l: 789 case vmIntrinsics::_numberOfTrailingZeros_i: 790 case vmIntrinsics::_numberOfTrailingZeros_l: 791 case vmIntrinsics::_bitCount_i: 792 case vmIntrinsics::_bitCount_l: 793 case vmIntrinsics::_reverseBytes_i: 794 case vmIntrinsics::_reverseBytes_l: 795 case vmIntrinsics::_reverseBytes_s: 796 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 797 798 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 799 800 case vmIntrinsics::_Reference_get: return inline_reference_get(); 801 802 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 803 804 case vmIntrinsics::_aescrypt_encryptBlock: 805 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 806 807 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 808 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 809 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 810 811 case vmIntrinsics::_counterMode_AESCrypt: 812 return inline_counterMode_AESCrypt(intrinsic_id()); 813 814 case vmIntrinsics::_sha_implCompress: 815 case vmIntrinsics::_sha2_implCompress: 816 case vmIntrinsics::_sha5_implCompress: 817 return inline_sha_implCompress(intrinsic_id()); 818 819 case vmIntrinsics::_digestBase_implCompressMB: 820 return inline_digestBase_implCompressMB(predicate); 821 822 case vmIntrinsics::_multiplyToLen: 823 return inline_multiplyToLen(); 824 825 case vmIntrinsics::_squareToLen: 826 return inline_squareToLen(); 827 828 case vmIntrinsics::_mulAdd: 829 return inline_mulAdd(); 830 831 case vmIntrinsics::_montgomeryMultiply: 832 return inline_montgomeryMultiply(); 833 case vmIntrinsics::_montgomerySquare: 834 return inline_montgomerySquare(); 835 836 case vmIntrinsics::_vectorizedMismatch: 837 return inline_vectorizedMismatch(); 838 839 case vmIntrinsics::_ghash_processBlocks: 840 return inline_ghash_processBlocks(); 841 842 case vmIntrinsics::_encodeISOArray: 843 case vmIntrinsics::_encodeByteISOArray: 844 return inline_encodeISOArray(); 845 846 case vmIntrinsics::_updateCRC32: 847 return inline_updateCRC32(); 848 case vmIntrinsics::_updateBytesCRC32: 849 return inline_updateBytesCRC32(); 850 case vmIntrinsics::_updateByteBufferCRC32: 851 return inline_updateByteBufferCRC32(); 852 853 case vmIntrinsics::_updateBytesCRC32C: 854 return inline_updateBytesCRC32C(); 855 case vmIntrinsics::_updateDirectByteBufferCRC32C: 856 return inline_updateDirectByteBufferCRC32C(); 857 858 case vmIntrinsics::_updateBytesAdler32: 859 return inline_updateBytesAdler32(); 860 case vmIntrinsics::_updateByteBufferAdler32: 861 return inline_updateByteBufferAdler32(); 862 863 case vmIntrinsics::_profileBoolean: 864 return inline_profileBoolean(); 865 case vmIntrinsics::_isCompileConstant: 866 return inline_isCompileConstant(); 867 868 case vmIntrinsics::_hasNegatives: 869 return inline_hasNegatives(); 870 871 case vmIntrinsics::_fmaD: 872 case vmIntrinsics::_fmaF: 873 return inline_fma(intrinsic_id()); 874 875 default: 876 // If you get here, it may be that someone has added a new intrinsic 877 // to the list in vmSymbols.hpp without implementing it here. 878 #ifndef PRODUCT 879 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 880 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 881 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 882 } 883 #endif 884 return false; 885 } 886 } 887 888 Node* LibraryCallKit::try_to_predicate(int predicate) { 889 if (!jvms()->has_method()) { 890 // Root JVMState has a null method. 891 assert(map()->memory()->Opcode() == Op_Parm, ""); 892 // Insert the memory aliasing node 893 set_all_memory(reset_memory()); 894 } 895 assert(merged_memory(), ""); 896 897 switch (intrinsic_id()) { 898 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 899 return inline_cipherBlockChaining_AESCrypt_predicate(false); 900 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 901 return inline_cipherBlockChaining_AESCrypt_predicate(true); 902 case vmIntrinsics::_counterMode_AESCrypt: 903 return inline_counterMode_AESCrypt_predicate(); 904 case vmIntrinsics::_digestBase_implCompressMB: 905 return inline_digestBase_implCompressMB_predicate(predicate); 906 907 default: 908 // If you get here, it may be that someone has added a new intrinsic 909 // to the list in vmSymbols.hpp without implementing it here. 910 #ifndef PRODUCT 911 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 912 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 913 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 914 } 915 #endif 916 Node* slow_ctl = control(); 917 set_control(top()); // No fast path instrinsic 918 return slow_ctl; 919 } 920 } 921 922 //------------------------------set_result------------------------------- 923 // Helper function for finishing intrinsics. 924 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 925 record_for_igvn(region); 926 set_control(_gvn.transform(region)); 927 set_result( _gvn.transform(value)); 928 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 929 } 930 931 //------------------------------generate_guard--------------------------- 932 // Helper function for generating guarded fast-slow graph structures. 933 // The given 'test', if true, guards a slow path. If the test fails 934 // then a fast path can be taken. (We generally hope it fails.) 935 // In all cases, GraphKit::control() is updated to the fast path. 936 // The returned value represents the control for the slow path. 937 // The return value is never 'top'; it is either a valid control 938 // or NULL if it is obvious that the slow path can never be taken. 939 // Also, if region and the slow control are not NULL, the slow edge 940 // is appended to the region. 941 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 942 if (stopped()) { 943 // Already short circuited. 944 return NULL; 945 } 946 947 // Build an if node and its projections. 948 // If test is true we take the slow path, which we assume is uncommon. 949 if (_gvn.type(test) == TypeInt::ZERO) { 950 // The slow branch is never taken. No need to build this guard. 951 return NULL; 952 } 953 954 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 955 956 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 957 if (if_slow == top()) { 958 // The slow branch is never taken. No need to build this guard. 959 return NULL; 960 } 961 962 if (region != NULL) 963 region->add_req(if_slow); 964 965 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 966 set_control(if_fast); 967 968 return if_slow; 969 } 970 971 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 972 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 973 } 974 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 975 return generate_guard(test, region, PROB_FAIR); 976 } 977 978 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 979 Node* *pos_index) { 980 if (stopped()) 981 return NULL; // already stopped 982 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 983 return NULL; // index is already adequately typed 984 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 985 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 986 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 987 if (is_neg != NULL && pos_index != NULL) { 988 // Emulate effect of Parse::adjust_map_after_if. 989 Node* ccast = new CastIINode(index, TypeInt::POS); 990 ccast->set_req(0, control()); 991 (*pos_index) = _gvn.transform(ccast); 992 } 993 return is_neg; 994 } 995 996 // Make sure that 'position' is a valid limit index, in [0..length]. 997 // There are two equivalent plans for checking this: 998 // A. (offset + copyLength) unsigned<= arrayLength 999 // B. offset <= (arrayLength - copyLength) 1000 // We require that all of the values above, except for the sum and 1001 // difference, are already known to be non-negative. 1002 // Plan A is robust in the face of overflow, if offset and copyLength 1003 // are both hugely positive. 1004 // 1005 // Plan B is less direct and intuitive, but it does not overflow at 1006 // all, since the difference of two non-negatives is always 1007 // representable. Whenever Java methods must perform the equivalent 1008 // check they generally use Plan B instead of Plan A. 1009 // For the moment we use Plan A. 1010 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1011 Node* subseq_length, 1012 Node* array_length, 1013 RegionNode* region) { 1014 if (stopped()) 1015 return NULL; // already stopped 1016 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1017 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1018 return NULL; // common case of whole-array copy 1019 Node* last = subseq_length; 1020 if (!zero_offset) // last += offset 1021 last = _gvn.transform(new AddINode(last, offset)); 1022 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1023 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1024 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1025 return is_over; 1026 } 1027 1028 // Emit range checks for the given String.value byte array 1029 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 1030 if (stopped()) { 1031 return; // already stopped 1032 } 1033 RegionNode* bailout = new RegionNode(1); 1034 record_for_igvn(bailout); 1035 if (char_count) { 1036 // Convert char count to byte count 1037 count = _gvn.transform(new LShiftINode(count, intcon(1))); 1038 } 1039 1040 // Offset and count must not be negative 1041 generate_negative_guard(offset, bailout); 1042 generate_negative_guard(count, bailout); 1043 // Offset + count must not exceed length of array 1044 generate_limit_guard(offset, count, load_array_length(array), bailout); 1045 1046 if (bailout->req() > 1) { 1047 PreserveJVMState pjvms(this); 1048 set_control(_gvn.transform(bailout)); 1049 uncommon_trap(Deoptimization::Reason_intrinsic, 1050 Deoptimization::Action_maybe_recompile); 1051 } 1052 } 1053 1054 //--------------------------generate_current_thread-------------------- 1055 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1056 ciKlass* thread_klass = env()->Thread_klass(); 1057 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1058 Node* thread = _gvn.transform(new ThreadLocalNode()); 1059 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1060 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1061 if (UseZGC) { 1062 threadObj = load_barrier(threadObj, p); 1063 } 1064 1065 tls_output = thread; 1066 return threadObj; 1067 } 1068 1069 1070 //------------------------------make_string_method_node------------------------ 1071 // Helper method for String intrinsic functions. This version is called with 1072 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 1073 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 1074 // containing the lengths of str1 and str2. 1075 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 1076 Node* result = NULL; 1077 switch (opcode) { 1078 case Op_StrIndexOf: 1079 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 1080 str1_start, cnt1, str2_start, cnt2, ae); 1081 break; 1082 case Op_StrComp: 1083 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 1084 str1_start, cnt1, str2_start, cnt2, ae); 1085 break; 1086 case Op_StrEquals: 1087 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 1088 // Use the constant length if there is one because optimized match rule may exist. 1089 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 1090 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 1091 break; 1092 default: 1093 ShouldNotReachHere(); 1094 return NULL; 1095 } 1096 1097 // All these intrinsics have checks. 1098 C->set_has_split_ifs(true); // Has chance for split-if optimization 1099 clear_upper_avx(); 1100 1101 return _gvn.transform(result); 1102 } 1103 1104 //------------------------------inline_string_compareTo------------------------ 1105 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 1106 Node* arg1 = argument(0); 1107 Node* arg2 = argument(1); 1108 1109 // Get start addr and length of first argument 1110 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1111 Node* arg1_cnt = load_array_length(arg1); 1112 1113 // Get start addr and length of second argument 1114 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1115 Node* arg2_cnt = load_array_length(arg2); 1116 1117 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1118 set_result(result); 1119 return true; 1120 } 1121 1122 //------------------------------inline_string_equals------------------------ 1123 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1124 Node* arg1 = argument(0); 1125 Node* arg2 = argument(1); 1126 1127 // paths (plus control) merge 1128 RegionNode* region = new RegionNode(3); 1129 Node* phi = new PhiNode(region, TypeInt::BOOL); 1130 1131 if (!stopped()) { 1132 // Get start addr and length of first argument 1133 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1134 Node* arg1_cnt = load_array_length(arg1); 1135 1136 // Get start addr and length of second argument 1137 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1138 Node* arg2_cnt = load_array_length(arg2); 1139 1140 // Check for arg1_cnt != arg2_cnt 1141 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1142 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1143 Node* if_ne = generate_slow_guard(bol, NULL); 1144 if (if_ne != NULL) { 1145 phi->init_req(2, intcon(0)); 1146 region->init_req(2, if_ne); 1147 } 1148 1149 // Check for count == 0 is done by assembler code for StrEquals. 1150 1151 if (!stopped()) { 1152 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1153 phi->init_req(1, equals); 1154 region->init_req(1, control()); 1155 } 1156 } 1157 1158 // post merge 1159 set_control(_gvn.transform(region)); 1160 record_for_igvn(region); 1161 1162 set_result(_gvn.transform(phi)); 1163 return true; 1164 } 1165 1166 //------------------------------inline_array_equals---------------------------- 1167 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1168 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1169 Node* arg1 = argument(0); 1170 Node* arg2 = argument(1); 1171 1172 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1173 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1174 clear_upper_avx(); 1175 1176 return true; 1177 } 1178 1179 //------------------------------inline_hasNegatives------------------------------ 1180 bool LibraryCallKit::inline_hasNegatives() { 1181 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1182 return false; 1183 } 1184 1185 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1186 // no receiver since it is static method 1187 Node* ba = argument(0); 1188 Node* offset = argument(1); 1189 Node* len = argument(2); 1190 1191 // Range checks 1192 generate_string_range_check(ba, offset, len, false); 1193 if (stopped()) { 1194 return true; 1195 } 1196 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1197 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1198 set_result(_gvn.transform(result)); 1199 return true; 1200 } 1201 1202 bool LibraryCallKit::inline_preconditions_checkIndex() { 1203 Node* index = argument(0); 1204 Node* length = argument(1); 1205 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1206 return false; 1207 } 1208 1209 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1210 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1211 1212 { 1213 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1214 uncommon_trap(Deoptimization::Reason_intrinsic, 1215 Deoptimization::Action_make_not_entrant); 1216 } 1217 1218 if (stopped()) { 1219 return false; 1220 } 1221 1222 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1223 BoolTest::mask btest = BoolTest::lt; 1224 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1225 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1226 _gvn.set_type(rc, rc->Value(&_gvn)); 1227 if (!rc_bool->is_Con()) { 1228 record_for_igvn(rc); 1229 } 1230 set_control(_gvn.transform(new IfTrueNode(rc))); 1231 { 1232 PreserveJVMState pjvms(this); 1233 set_control(_gvn.transform(new IfFalseNode(rc))); 1234 uncommon_trap(Deoptimization::Reason_range_check, 1235 Deoptimization::Action_make_not_entrant); 1236 } 1237 1238 if (stopped()) { 1239 return false; 1240 } 1241 1242 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1243 result->set_req(0, control()); 1244 result = _gvn.transform(result); 1245 set_result(result); 1246 replace_in_map(index, result); 1247 clear_upper_avx(); 1248 return true; 1249 } 1250 1251 //------------------------------inline_string_indexOf------------------------ 1252 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1253 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1254 return false; 1255 } 1256 Node* src = argument(0); 1257 Node* tgt = argument(1); 1258 1259 // Make the merge point 1260 RegionNode* result_rgn = new RegionNode(4); 1261 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1262 1263 // Get start addr and length of source string 1264 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1265 Node* src_count = load_array_length(src); 1266 1267 // Get start addr and length of substring 1268 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1269 Node* tgt_count = load_array_length(tgt); 1270 1271 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1272 // Divide src size by 2 if String is UTF16 encoded 1273 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1274 } 1275 if (ae == StrIntrinsicNode::UU) { 1276 // Divide substring size by 2 if String is UTF16 encoded 1277 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1278 } 1279 1280 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1281 if (result != NULL) { 1282 result_phi->init_req(3, result); 1283 result_rgn->init_req(3, control()); 1284 } 1285 set_control(_gvn.transform(result_rgn)); 1286 record_for_igvn(result_rgn); 1287 set_result(_gvn.transform(result_phi)); 1288 1289 return true; 1290 } 1291 1292 //-----------------------------inline_string_indexOf----------------------- 1293 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1294 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1295 return false; 1296 } 1297 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1298 return false; 1299 } 1300 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1301 Node* src = argument(0); // byte[] 1302 Node* src_count = argument(1); // char count 1303 Node* tgt = argument(2); // byte[] 1304 Node* tgt_count = argument(3); // char count 1305 Node* from_index = argument(4); // char index 1306 1307 // Multiply byte array index by 2 if String is UTF16 encoded 1308 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1309 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1310 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1311 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1312 1313 // Range checks 1314 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1315 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1316 if (stopped()) { 1317 return true; 1318 } 1319 1320 RegionNode* region = new RegionNode(5); 1321 Node* phi = new PhiNode(region, TypeInt::INT); 1322 1323 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1324 if (result != NULL) { 1325 // The result is index relative to from_index if substring was found, -1 otherwise. 1326 // Generate code which will fold into cmove. 1327 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1328 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1329 1330 Node* if_lt = generate_slow_guard(bol, NULL); 1331 if (if_lt != NULL) { 1332 // result == -1 1333 phi->init_req(3, result); 1334 region->init_req(3, if_lt); 1335 } 1336 if (!stopped()) { 1337 result = _gvn.transform(new AddINode(result, from_index)); 1338 phi->init_req(4, result); 1339 region->init_req(4, control()); 1340 } 1341 } 1342 1343 set_control(_gvn.transform(region)); 1344 record_for_igvn(region); 1345 set_result(_gvn.transform(phi)); 1346 clear_upper_avx(); 1347 1348 return true; 1349 } 1350 1351 // Create StrIndexOfNode with fast path checks 1352 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1353 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1354 // Check for substr count > string count 1355 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1356 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1357 Node* if_gt = generate_slow_guard(bol, NULL); 1358 if (if_gt != NULL) { 1359 phi->init_req(1, intcon(-1)); 1360 region->init_req(1, if_gt); 1361 } 1362 if (!stopped()) { 1363 // Check for substr count == 0 1364 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1365 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1366 Node* if_zero = generate_slow_guard(bol, NULL); 1367 if (if_zero != NULL) { 1368 phi->init_req(2, intcon(0)); 1369 region->init_req(2, if_zero); 1370 } 1371 } 1372 if (!stopped()) { 1373 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1374 } 1375 return NULL; 1376 } 1377 1378 //-----------------------------inline_string_indexOfChar----------------------- 1379 bool LibraryCallKit::inline_string_indexOfChar() { 1380 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1381 return false; 1382 } 1383 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1384 return false; 1385 } 1386 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1387 Node* src = argument(0); // byte[] 1388 Node* tgt = argument(1); // tgt is int ch 1389 Node* from_index = argument(2); 1390 Node* max = argument(3); 1391 1392 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1393 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1394 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1395 1396 // Range checks 1397 generate_string_range_check(src, src_offset, src_count, true); 1398 if (stopped()) { 1399 return true; 1400 } 1401 1402 RegionNode* region = new RegionNode(3); 1403 Node* phi = new PhiNode(region, TypeInt::INT); 1404 1405 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1406 C->set_has_split_ifs(true); // Has chance for split-if optimization 1407 _gvn.transform(result); 1408 1409 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1410 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1411 1412 Node* if_lt = generate_slow_guard(bol, NULL); 1413 if (if_lt != NULL) { 1414 // result == -1 1415 phi->init_req(2, result); 1416 region->init_req(2, if_lt); 1417 } 1418 if (!stopped()) { 1419 result = _gvn.transform(new AddINode(result, from_index)); 1420 phi->init_req(1, result); 1421 region->init_req(1, control()); 1422 } 1423 set_control(_gvn.transform(region)); 1424 record_for_igvn(region); 1425 set_result(_gvn.transform(phi)); 1426 1427 return true; 1428 } 1429 //---------------------------inline_string_copy--------------------- 1430 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1431 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1432 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1433 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1434 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1435 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1436 bool LibraryCallKit::inline_string_copy(bool compress) { 1437 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1438 return false; 1439 } 1440 int nargs = 5; // 2 oops, 3 ints 1441 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1442 1443 Node* src = argument(0); 1444 Node* src_offset = argument(1); 1445 Node* dst = argument(2); 1446 Node* dst_offset = argument(3); 1447 Node* length = argument(4); 1448 1449 // Check for allocation before we add nodes that would confuse 1450 // tightly_coupled_allocation() 1451 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1452 1453 // Figure out the size and type of the elements we will be copying. 1454 const Type* src_type = src->Value(&_gvn); 1455 const Type* dst_type = dst->Value(&_gvn); 1456 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1457 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1458 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1459 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1460 "Unsupported array types for inline_string_copy"); 1461 1462 // Convert char[] offsets to byte[] offsets 1463 bool convert_src = (compress && src_elem == T_BYTE); 1464 bool convert_dst = (!compress && dst_elem == T_BYTE); 1465 if (convert_src) { 1466 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1467 } else if (convert_dst) { 1468 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1469 } 1470 1471 // Range checks 1472 generate_string_range_check(src, src_offset, length, convert_src); 1473 generate_string_range_check(dst, dst_offset, length, convert_dst); 1474 if (stopped()) { 1475 return true; 1476 } 1477 1478 Node* src_start = array_element_address(src, src_offset, src_elem); 1479 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1480 // 'src_start' points to src array + scaled offset 1481 // 'dst_start' points to dst array + scaled offset 1482 Node* count = NULL; 1483 if (compress) { 1484 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1485 } else { 1486 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1487 } 1488 1489 if (alloc != NULL) { 1490 if (alloc->maybe_set_complete(&_gvn)) { 1491 // "You break it, you buy it." 1492 InitializeNode* init = alloc->initialization(); 1493 assert(init->is_complete(), "we just did this"); 1494 init->set_complete_with_arraycopy(); 1495 assert(dst->is_CheckCastPP(), "sanity"); 1496 assert(dst->in(0)->in(0) == init, "dest pinned"); 1497 } 1498 // Do not let stores that initialize this object be reordered with 1499 // a subsequent store that would make this object accessible by 1500 // other threads. 1501 // Record what AllocateNode this StoreStore protects so that 1502 // escape analysis can go from the MemBarStoreStoreNode to the 1503 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1504 // based on the escape status of the AllocateNode. 1505 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1506 } 1507 if (compress) { 1508 set_result(_gvn.transform(count)); 1509 } 1510 clear_upper_avx(); 1511 1512 return true; 1513 } 1514 1515 #ifdef _LP64 1516 #define XTOP ,top() /*additional argument*/ 1517 #else //_LP64 1518 #define XTOP /*no additional argument*/ 1519 #endif //_LP64 1520 1521 //------------------------inline_string_toBytesU-------------------------- 1522 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1523 bool LibraryCallKit::inline_string_toBytesU() { 1524 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1525 return false; 1526 } 1527 // Get the arguments. 1528 Node* value = argument(0); 1529 Node* offset = argument(1); 1530 Node* length = argument(2); 1531 1532 Node* newcopy = NULL; 1533 1534 // Set the original stack and the reexecute bit for the interpreter to reexecute 1535 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1536 { PreserveReexecuteState preexecs(this); 1537 jvms()->set_should_reexecute(true); 1538 1539 // Check if a null path was taken unconditionally. 1540 value = null_check(value); 1541 1542 RegionNode* bailout = new RegionNode(1); 1543 record_for_igvn(bailout); 1544 1545 // Range checks 1546 generate_negative_guard(offset, bailout); 1547 generate_negative_guard(length, bailout); 1548 generate_limit_guard(offset, length, load_array_length(value), bailout); 1549 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1550 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1551 1552 if (bailout->req() > 1) { 1553 PreserveJVMState pjvms(this); 1554 set_control(_gvn.transform(bailout)); 1555 uncommon_trap(Deoptimization::Reason_intrinsic, 1556 Deoptimization::Action_maybe_recompile); 1557 } 1558 if (stopped()) { 1559 return true; 1560 } 1561 1562 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1563 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1564 newcopy = new_array(klass_node, size, 0); // no arguments to push 1565 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1566 1567 // Calculate starting addresses. 1568 Node* src_start = array_element_address(value, offset, T_CHAR); 1569 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1570 1571 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1572 const TypeInt* toffset = gvn().type(offset)->is_int(); 1573 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1574 1575 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1576 const char* copyfunc_name = "arraycopy"; 1577 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1578 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1579 OptoRuntime::fast_arraycopy_Type(), 1580 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1581 src_start, dst_start, ConvI2X(length) XTOP); 1582 // Do not let reads from the cloned object float above the arraycopy. 1583 if (alloc != NULL) { 1584 if (alloc->maybe_set_complete(&_gvn)) { 1585 // "You break it, you buy it." 1586 InitializeNode* init = alloc->initialization(); 1587 assert(init->is_complete(), "we just did this"); 1588 init->set_complete_with_arraycopy(); 1589 assert(newcopy->is_CheckCastPP(), "sanity"); 1590 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1591 } 1592 // Do not let stores that initialize this object be reordered with 1593 // a subsequent store that would make this object accessible by 1594 // other threads. 1595 // Record what AllocateNode this StoreStore protects so that 1596 // escape analysis can go from the MemBarStoreStoreNode to the 1597 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1598 // based on the escape status of the AllocateNode. 1599 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1600 } else { 1601 insert_mem_bar(Op_MemBarCPUOrder); 1602 } 1603 } // original reexecute is set back here 1604 1605 C->set_has_split_ifs(true); // Has chance for split-if optimization 1606 if (!stopped()) { 1607 set_result(newcopy); 1608 } 1609 clear_upper_avx(); 1610 1611 return true; 1612 } 1613 1614 //------------------------inline_string_getCharsU-------------------------- 1615 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1616 bool LibraryCallKit::inline_string_getCharsU() { 1617 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1618 return false; 1619 } 1620 1621 // Get the arguments. 1622 Node* src = argument(0); 1623 Node* src_begin = argument(1); 1624 Node* src_end = argument(2); // exclusive offset (i < src_end) 1625 Node* dst = argument(3); 1626 Node* dst_begin = argument(4); 1627 1628 // Check for allocation before we add nodes that would confuse 1629 // tightly_coupled_allocation() 1630 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1631 1632 // Check if a null path was taken unconditionally. 1633 src = null_check(src); 1634 dst = null_check(dst); 1635 if (stopped()) { 1636 return true; 1637 } 1638 1639 // Get length and convert char[] offset to byte[] offset 1640 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1641 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1642 1643 // Range checks 1644 generate_string_range_check(src, src_begin, length, true); 1645 generate_string_range_check(dst, dst_begin, length, false); 1646 if (stopped()) { 1647 return true; 1648 } 1649 1650 if (!stopped()) { 1651 // Calculate starting addresses. 1652 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1653 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1654 1655 // Check if array addresses are aligned to HeapWordSize 1656 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1657 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1658 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1659 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1660 1661 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1662 const char* copyfunc_name = "arraycopy"; 1663 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1664 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1665 OptoRuntime::fast_arraycopy_Type(), 1666 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1667 src_start, dst_start, ConvI2X(length) XTOP); 1668 // Do not let reads from the cloned object float above the arraycopy. 1669 if (alloc != NULL) { 1670 if (alloc->maybe_set_complete(&_gvn)) { 1671 // "You break it, you buy it." 1672 InitializeNode* init = alloc->initialization(); 1673 assert(init->is_complete(), "we just did this"); 1674 init->set_complete_with_arraycopy(); 1675 assert(dst->is_CheckCastPP(), "sanity"); 1676 assert(dst->in(0)->in(0) == init, "dest pinned"); 1677 } 1678 // Do not let stores that initialize this object be reordered with 1679 // a subsequent store that would make this object accessible by 1680 // other threads. 1681 // Record what AllocateNode this StoreStore protects so that 1682 // escape analysis can go from the MemBarStoreStoreNode to the 1683 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1684 // based on the escape status of the AllocateNode. 1685 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1686 } else { 1687 insert_mem_bar(Op_MemBarCPUOrder); 1688 } 1689 } 1690 1691 C->set_has_split_ifs(true); // Has chance for split-if optimization 1692 return true; 1693 } 1694 1695 //----------------------inline_string_char_access---------------------------- 1696 // Store/Load char to/from byte[] array. 1697 // static void StringUTF16.putChar(byte[] val, int index, int c) 1698 // static char StringUTF16.getChar(byte[] val, int index) 1699 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1700 Node* value = argument(0); 1701 Node* index = argument(1); 1702 Node* ch = is_store ? argument(2) : NULL; 1703 1704 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1705 // correctly requires matched array shapes. 1706 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1707 "sanity: byte[] and char[] bases agree"); 1708 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1709 "sanity: byte[] and char[] scales agree"); 1710 1711 // Bail when getChar over constants is requested: constant folding would 1712 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1713 // Java method would constant fold nicely instead. 1714 if (!is_store && value->is_Con() && index->is_Con()) { 1715 return false; 1716 } 1717 1718 Node* adr = array_element_address(value, index, T_CHAR); 1719 if (adr->is_top()) { 1720 return false; 1721 } 1722 if (is_store) { 1723 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1724 false, false, true /* mismatched */); 1725 } else { 1726 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1727 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */); 1728 set_result(ch); 1729 } 1730 return true; 1731 } 1732 1733 //--------------------------round_double_node-------------------------------- 1734 // Round a double node if necessary. 1735 Node* LibraryCallKit::round_double_node(Node* n) { 1736 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1737 n = _gvn.transform(new RoundDoubleNode(0, n)); 1738 return n; 1739 } 1740 1741 //------------------------------inline_math----------------------------------- 1742 // public static double Math.abs(double) 1743 // public static double Math.sqrt(double) 1744 // public static double Math.log(double) 1745 // public static double Math.log10(double) 1746 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1747 Node* arg = round_double_node(argument(0)); 1748 Node* n = NULL; 1749 switch (id) { 1750 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1751 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1752 default: fatal_unexpected_iid(id); break; 1753 } 1754 set_result(_gvn.transform(n)); 1755 return true; 1756 } 1757 1758 //------------------------------runtime_math----------------------------- 1759 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1760 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1761 "must be (DD)D or (D)D type"); 1762 1763 // Inputs 1764 Node* a = round_double_node(argument(0)); 1765 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1766 1767 const TypePtr* no_memory_effects = NULL; 1768 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1769 no_memory_effects, 1770 a, top(), b, b ? top() : NULL); 1771 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1772 #ifdef ASSERT 1773 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1774 assert(value_top == top(), "second value must be top"); 1775 #endif 1776 1777 set_result(value); 1778 return true; 1779 } 1780 1781 //------------------------------inline_math_native----------------------------- 1782 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1783 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1784 switch (id) { 1785 // These intrinsics are not properly supported on all hardware 1786 case vmIntrinsics::_dsin: 1787 return StubRoutines::dsin() != NULL ? 1788 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1789 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1790 case vmIntrinsics::_dcos: 1791 return StubRoutines::dcos() != NULL ? 1792 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1793 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1794 case vmIntrinsics::_dtan: 1795 return StubRoutines::dtan() != NULL ? 1796 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1797 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1798 case vmIntrinsics::_dlog: 1799 return StubRoutines::dlog() != NULL ? 1800 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1801 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1802 case vmIntrinsics::_dlog10: 1803 return StubRoutines::dlog10() != NULL ? 1804 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1805 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1806 1807 // These intrinsics are supported on all hardware 1808 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1809 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1810 1811 case vmIntrinsics::_dexp: 1812 return StubRoutines::dexp() != NULL ? 1813 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1814 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1815 case vmIntrinsics::_dpow: { 1816 Node* exp = round_double_node(argument(2)); 1817 const TypeD* d = _gvn.type(exp)->isa_double_constant(); 1818 if (d != NULL && d->getd() == 2.0) { 1819 // Special case: pow(x, 2.0) => x * x 1820 Node* base = round_double_node(argument(0)); 1821 set_result(_gvn.transform(new MulDNode(base, base))); 1822 return true; 1823 } 1824 return StubRoutines::dexp() != NULL ? 1825 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1826 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1827 } 1828 #undef FN_PTR 1829 1830 // These intrinsics are not yet correctly implemented 1831 case vmIntrinsics::_datan2: 1832 return false; 1833 1834 default: 1835 fatal_unexpected_iid(id); 1836 return false; 1837 } 1838 } 1839 1840 static bool is_simple_name(Node* n) { 1841 return (n->req() == 1 // constant 1842 || (n->is_Type() && n->as_Type()->type()->singleton()) 1843 || n->is_Proj() // parameter or return value 1844 || n->is_Phi() // local of some sort 1845 ); 1846 } 1847 1848 //----------------------------inline_notify-----------------------------------* 1849 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1850 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1851 address func; 1852 if (id == vmIntrinsics::_notify) { 1853 func = OptoRuntime::monitor_notify_Java(); 1854 } else { 1855 func = OptoRuntime::monitor_notifyAll_Java(); 1856 } 1857 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1858 make_slow_call_ex(call, env()->Throwable_klass(), false); 1859 return true; 1860 } 1861 1862 1863 //----------------------------inline_min_max----------------------------------- 1864 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1865 set_result(generate_min_max(id, argument(0), argument(1))); 1866 return true; 1867 } 1868 1869 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1870 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1871 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1872 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1873 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1874 1875 { 1876 PreserveJVMState pjvms(this); 1877 PreserveReexecuteState preexecs(this); 1878 jvms()->set_should_reexecute(true); 1879 1880 set_control(slow_path); 1881 set_i_o(i_o()); 1882 1883 uncommon_trap(Deoptimization::Reason_intrinsic, 1884 Deoptimization::Action_none); 1885 } 1886 1887 set_control(fast_path); 1888 set_result(math); 1889 } 1890 1891 template <typename OverflowOp> 1892 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1893 typedef typename OverflowOp::MathOp MathOp; 1894 1895 MathOp* mathOp = new MathOp(arg1, arg2); 1896 Node* operation = _gvn.transform( mathOp ); 1897 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1898 inline_math_mathExact(operation, ofcheck); 1899 return true; 1900 } 1901 1902 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1903 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1904 } 1905 1906 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1907 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1908 } 1909 1910 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1911 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1912 } 1913 1914 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1915 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1916 } 1917 1918 bool LibraryCallKit::inline_math_negateExactI() { 1919 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1920 } 1921 1922 bool LibraryCallKit::inline_math_negateExactL() { 1923 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 1924 } 1925 1926 bool LibraryCallKit::inline_math_multiplyExactI() { 1927 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 1928 } 1929 1930 bool LibraryCallKit::inline_math_multiplyExactL() { 1931 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 1932 } 1933 1934 bool LibraryCallKit::inline_math_multiplyHigh() { 1935 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2)))); 1936 return true; 1937 } 1938 1939 Node* 1940 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1941 // These are the candidate return value: 1942 Node* xvalue = x0; 1943 Node* yvalue = y0; 1944 1945 if (xvalue == yvalue) { 1946 return xvalue; 1947 } 1948 1949 bool want_max = (id == vmIntrinsics::_max); 1950 1951 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1952 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1953 if (txvalue == NULL || tyvalue == NULL) return top(); 1954 // This is not really necessary, but it is consistent with a 1955 // hypothetical MaxINode::Value method: 1956 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1957 1958 // %%% This folding logic should (ideally) be in a different place. 1959 // Some should be inside IfNode, and there to be a more reliable 1960 // transformation of ?: style patterns into cmoves. We also want 1961 // more powerful optimizations around cmove and min/max. 1962 1963 // Try to find a dominating comparison of these guys. 1964 // It can simplify the index computation for Arrays.copyOf 1965 // and similar uses of System.arraycopy. 1966 // First, compute the normalized version of CmpI(x, y). 1967 int cmp_op = Op_CmpI; 1968 Node* xkey = xvalue; 1969 Node* ykey = yvalue; 1970 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 1971 if (ideal_cmpxy->is_Cmp()) { 1972 // E.g., if we have CmpI(length - offset, count), 1973 // it might idealize to CmpI(length, count + offset) 1974 cmp_op = ideal_cmpxy->Opcode(); 1975 xkey = ideal_cmpxy->in(1); 1976 ykey = ideal_cmpxy->in(2); 1977 } 1978 1979 // Start by locating any relevant comparisons. 1980 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1981 Node* cmpxy = NULL; 1982 Node* cmpyx = NULL; 1983 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1984 Node* cmp = start_from->fast_out(k); 1985 if (cmp->outcnt() > 0 && // must have prior uses 1986 cmp->in(0) == NULL && // must be context-independent 1987 cmp->Opcode() == cmp_op) { // right kind of compare 1988 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1989 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1990 } 1991 } 1992 1993 const int NCMPS = 2; 1994 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1995 int cmpn; 1996 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1997 if (cmps[cmpn] != NULL) break; // find a result 1998 } 1999 if (cmpn < NCMPS) { 2000 // Look for a dominating test that tells us the min and max. 2001 int depth = 0; // Limit search depth for speed 2002 Node* dom = control(); 2003 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2004 if (++depth >= 100) break; 2005 Node* ifproj = dom; 2006 if (!ifproj->is_Proj()) continue; 2007 Node* iff = ifproj->in(0); 2008 if (!iff->is_If()) continue; 2009 Node* bol = iff->in(1); 2010 if (!bol->is_Bool()) continue; 2011 Node* cmp = bol->in(1); 2012 if (cmp == NULL) continue; 2013 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2014 if (cmps[cmpn] == cmp) break; 2015 if (cmpn == NCMPS) continue; 2016 BoolTest::mask btest = bol->as_Bool()->_test._test; 2017 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2018 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2019 // At this point, we know that 'x btest y' is true. 2020 switch (btest) { 2021 case BoolTest::eq: 2022 // They are proven equal, so we can collapse the min/max. 2023 // Either value is the answer. Choose the simpler. 2024 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2025 return yvalue; 2026 return xvalue; 2027 case BoolTest::lt: // x < y 2028 case BoolTest::le: // x <= y 2029 return (want_max ? yvalue : xvalue); 2030 case BoolTest::gt: // x > y 2031 case BoolTest::ge: // x >= y 2032 return (want_max ? xvalue : yvalue); 2033 default: 2034 break; 2035 } 2036 } 2037 } 2038 2039 // We failed to find a dominating test. 2040 // Let's pick a test that might GVN with prior tests. 2041 Node* best_bol = NULL; 2042 BoolTest::mask best_btest = BoolTest::illegal; 2043 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2044 Node* cmp = cmps[cmpn]; 2045 if (cmp == NULL) continue; 2046 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2047 Node* bol = cmp->fast_out(j); 2048 if (!bol->is_Bool()) continue; 2049 BoolTest::mask btest = bol->as_Bool()->_test._test; 2050 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2051 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2052 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2053 best_bol = bol->as_Bool(); 2054 best_btest = btest; 2055 } 2056 } 2057 } 2058 2059 Node* answer_if_true = NULL; 2060 Node* answer_if_false = NULL; 2061 switch (best_btest) { 2062 default: 2063 if (cmpxy == NULL) 2064 cmpxy = ideal_cmpxy; 2065 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2066 // and fall through: 2067 case BoolTest::lt: // x < y 2068 case BoolTest::le: // x <= y 2069 answer_if_true = (want_max ? yvalue : xvalue); 2070 answer_if_false = (want_max ? xvalue : yvalue); 2071 break; 2072 case BoolTest::gt: // x > y 2073 case BoolTest::ge: // x >= y 2074 answer_if_true = (want_max ? xvalue : yvalue); 2075 answer_if_false = (want_max ? yvalue : xvalue); 2076 break; 2077 } 2078 2079 jint hi, lo; 2080 if (want_max) { 2081 // We can sharpen the minimum. 2082 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2083 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2084 } else { 2085 // We can sharpen the maximum. 2086 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2087 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2088 } 2089 2090 // Use a flow-free graph structure, to avoid creating excess control edges 2091 // which could hinder other optimizations. 2092 // Since Math.min/max is often used with arraycopy, we want 2093 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2094 Node* cmov = CMoveNode::make(NULL, best_bol, 2095 answer_if_false, answer_if_true, 2096 TypeInt::make(lo, hi, widen)); 2097 2098 return _gvn.transform(cmov); 2099 2100 /* 2101 // This is not as desirable as it may seem, since Min and Max 2102 // nodes do not have a full set of optimizations. 2103 // And they would interfere, anyway, with 'if' optimizations 2104 // and with CMoveI canonical forms. 2105 switch (id) { 2106 case vmIntrinsics::_min: 2107 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2108 case vmIntrinsics::_max: 2109 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2110 default: 2111 ShouldNotReachHere(); 2112 } 2113 */ 2114 } 2115 2116 inline int 2117 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) { 2118 const TypePtr* base_type = TypePtr::NULL_PTR; 2119 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2120 if (base_type == NULL) { 2121 // Unknown type. 2122 return Type::AnyPtr; 2123 } else if (base_type == TypePtr::NULL_PTR) { 2124 // Since this is a NULL+long form, we have to switch to a rawptr. 2125 base = _gvn.transform(new CastX2PNode(offset)); 2126 offset = MakeConX(0); 2127 return Type::RawPtr; 2128 } else if (base_type->base() == Type::RawPtr) { 2129 return Type::RawPtr; 2130 } else if (base_type->isa_oopptr()) { 2131 // Base is never null => always a heap address. 2132 if (!TypePtr::NULL_PTR->higher_equal(base_type)) { 2133 return Type::OopPtr; 2134 } 2135 // Offset is small => always a heap address. 2136 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2137 if (offset_type != NULL && 2138 base_type->offset() == 0 && // (should always be?) 2139 offset_type->_lo >= 0 && 2140 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2141 return Type::OopPtr; 2142 } else if (type == T_OBJECT) { 2143 // off heap access to an oop doesn't make any sense. Has to be on 2144 // heap. 2145 return Type::OopPtr; 2146 } 2147 // Otherwise, it might either be oop+off or NULL+addr. 2148 return Type::AnyPtr; 2149 } else { 2150 // No information: 2151 return Type::AnyPtr; 2152 } 2153 } 2154 2155 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) { 2156 Node* uncasted_base = base; 2157 int kind = classify_unsafe_addr(uncasted_base, offset, type); 2158 if (kind == Type::RawPtr) { 2159 return basic_plus_adr(top(), uncasted_base, offset); 2160 } else if (kind == Type::AnyPtr) { 2161 assert(base == uncasted_base, "unexpected base change"); 2162 if (can_cast) { 2163 if (!_gvn.type(base)->speculative_maybe_null() && 2164 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2165 // According to profiling, this access is always on 2166 // heap. Casting the base to not null and thus avoiding membars 2167 // around the access should allow better optimizations 2168 Node* null_ctl = top(); 2169 base = null_check_oop(base, &null_ctl, true, true, true); 2170 assert(null_ctl->is_top(), "no null control here"); 2171 return basic_plus_adr(base, offset); 2172 } else if (_gvn.type(base)->speculative_always_null() && 2173 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2174 // According to profiling, this access is always off 2175 // heap. 2176 base = null_assert(base); 2177 Node* raw_base = _gvn.transform(new CastX2PNode(offset)); 2178 offset = MakeConX(0); 2179 return basic_plus_adr(top(), raw_base, offset); 2180 } 2181 } 2182 // We don't know if it's an on heap or off heap access. Fall back 2183 // to raw memory access. 2184 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM)); 2185 return basic_plus_adr(top(), raw, offset); 2186 } else { 2187 assert(base == uncasted_base, "unexpected base change"); 2188 // We know it's an on heap access so base can't be null 2189 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) { 2190 base = must_be_not_null(base, true); 2191 } 2192 return basic_plus_adr(base, offset); 2193 } 2194 } 2195 2196 //--------------------------inline_number_methods----------------------------- 2197 // inline int Integer.numberOfLeadingZeros(int) 2198 // inline int Long.numberOfLeadingZeros(long) 2199 // 2200 // inline int Integer.numberOfTrailingZeros(int) 2201 // inline int Long.numberOfTrailingZeros(long) 2202 // 2203 // inline int Integer.bitCount(int) 2204 // inline int Long.bitCount(long) 2205 // 2206 // inline char Character.reverseBytes(char) 2207 // inline short Short.reverseBytes(short) 2208 // inline int Integer.reverseBytes(int) 2209 // inline long Long.reverseBytes(long) 2210 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2211 Node* arg = argument(0); 2212 Node* n = NULL; 2213 switch (id) { 2214 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2215 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2216 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2217 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2218 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2219 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2220 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2221 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2222 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2223 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2224 default: fatal_unexpected_iid(id); break; 2225 } 2226 set_result(_gvn.transform(n)); 2227 return true; 2228 } 2229 2230 //----------------------------inline_unsafe_access---------------------------- 2231 2232 // Helper that guards and inserts a pre-barrier. 2233 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2234 Node* pre_val, bool need_mem_bar) { 2235 // We could be accessing the referent field of a reference object. If so, when G1 2236 // is enabled, we need to log the value in the referent field in an SATB buffer. 2237 // This routine performs some compile time filters and generates suitable 2238 // runtime filters that guard the pre-barrier code. 2239 // Also add memory barrier for non volatile load from the referent field 2240 // to prevent commoning of loads across safepoint. 2241 if (!UseG1GC && !need_mem_bar) 2242 return; 2243 2244 // Some compile time checks. 2245 2246 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2247 const TypeX* otype = offset->find_intptr_t_type(); 2248 if (otype != NULL && otype->is_con() && 2249 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2250 // Constant offset but not the reference_offset so just return 2251 return; 2252 } 2253 2254 // We only need to generate the runtime guards for instances. 2255 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2256 if (btype != NULL) { 2257 if (btype->isa_aryptr()) { 2258 // Array type so nothing to do 2259 return; 2260 } 2261 2262 const TypeInstPtr* itype = btype->isa_instptr(); 2263 if (itype != NULL) { 2264 // Can the klass of base_oop be statically determined to be 2265 // _not_ a sub-class of Reference and _not_ Object? 2266 ciKlass* klass = itype->klass(); 2267 if ( klass->is_loaded() && 2268 !klass->is_subtype_of(env()->Reference_klass()) && 2269 !env()->Object_klass()->is_subtype_of(klass)) { 2270 return; 2271 } 2272 } 2273 } 2274 2275 // The compile time filters did not reject base_oop/offset so 2276 // we need to generate the following runtime filters 2277 // 2278 // if (offset == java_lang_ref_Reference::_reference_offset) { 2279 // if (instance_of(base, java.lang.ref.Reference)) { 2280 // pre_barrier(_, pre_val, ...); 2281 // } 2282 // } 2283 2284 float likely = PROB_LIKELY( 0.999); 2285 float unlikely = PROB_UNLIKELY(0.999); 2286 2287 IdealKit ideal(this); 2288 #define __ ideal. 2289 2290 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2291 2292 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2293 // Update graphKit memory and control from IdealKit. 2294 sync_kit(ideal); 2295 2296 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2297 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2298 2299 // Update IdealKit memory and control from graphKit. 2300 __ sync_kit(this); 2301 2302 Node* one = __ ConI(1); 2303 // is_instof == 0 if base_oop == NULL 2304 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2305 2306 // Update graphKit from IdeakKit. 2307 sync_kit(ideal); 2308 2309 // Use the pre-barrier to record the value in the referent field 2310 pre_barrier(false /* do_load */, 2311 __ ctrl(), 2312 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2313 pre_val /* pre_val */, 2314 T_OBJECT); 2315 if (need_mem_bar) { 2316 // Add memory barrier to prevent commoning reads from this field 2317 // across safepoint since GC can change its value. 2318 insert_mem_bar(Op_MemBarCPUOrder); 2319 } 2320 // Update IdealKit from graphKit. 2321 __ sync_kit(this); 2322 2323 } __ end_if(); // _ref_type != ref_none 2324 } __ end_if(); // offset == referent_offset 2325 2326 // Final sync IdealKit and GraphKit. 2327 final_sync(ideal); 2328 #undef __ 2329 } 2330 2331 2332 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2333 // Attempt to infer a sharper value type from the offset and base type. 2334 ciKlass* sharpened_klass = NULL; 2335 2336 // See if it is an instance field, with an object type. 2337 if (alias_type->field() != NULL) { 2338 if (alias_type->field()->type()->is_klass()) { 2339 sharpened_klass = alias_type->field()->type()->as_klass(); 2340 } 2341 } 2342 2343 // See if it is a narrow oop array. 2344 if (adr_type->isa_aryptr()) { 2345 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2346 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2347 if (elem_type != NULL) { 2348 sharpened_klass = elem_type->klass(); 2349 } 2350 } 2351 } 2352 2353 // The sharpened class might be unloaded if there is no class loader 2354 // contraint in place. 2355 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2356 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2357 2358 #ifndef PRODUCT 2359 if (C->print_intrinsics() || C->print_inlining()) { 2360 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2361 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2362 } 2363 #endif 2364 // Sharpen the value type. 2365 return tjp; 2366 } 2367 return NULL; 2368 } 2369 2370 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2371 if (callee()->is_static()) return false; // caller must have the capability! 2372 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2373 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2374 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2375 2376 #ifndef PRODUCT 2377 { 2378 ResourceMark rm; 2379 // Check the signatures. 2380 ciSignature* sig = callee()->signature(); 2381 #ifdef ASSERT 2382 if (!is_store) { 2383 // Object getObject(Object base, int/long offset), etc. 2384 BasicType rtype = sig->return_type()->basic_type(); 2385 assert(rtype == type, "getter must return the expected value"); 2386 assert(sig->count() == 2, "oop getter has 2 arguments"); 2387 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2388 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2389 } else { 2390 // void putObject(Object base, int/long offset, Object x), etc. 2391 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2392 assert(sig->count() == 3, "oop putter has 3 arguments"); 2393 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2394 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2395 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2396 assert(vtype == type, "putter must accept the expected value"); 2397 } 2398 #endif // ASSERT 2399 } 2400 #endif //PRODUCT 2401 2402 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2403 2404 Node* receiver = argument(0); // type: oop 2405 2406 // Build address expression. 2407 Node* adr; 2408 Node* heap_base_oop = top(); 2409 Node* offset = top(); 2410 Node* val; 2411 2412 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2413 Node* base = argument(1); // type: oop 2414 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2415 offset = argument(2); // type: long 2416 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2417 // to be plain byte offsets, which are also the same as those accepted 2418 // by oopDesc::field_addr. 2419 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2420 "fieldOffset must be byte-scaled"); 2421 // 32-bit machines ignore the high half! 2422 offset = ConvL2X(offset); 2423 adr = make_unsafe_address(base, offset, type, kind == Relaxed); 2424 2425 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) { 2426 heap_base_oop = base; 2427 } else if (type == T_OBJECT) { 2428 return false; // off-heap oop accesses are not supported 2429 } 2430 2431 // Can base be NULL? Otherwise, always on-heap access. 2432 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop)); 2433 2434 val = is_store ? argument(4) : NULL; 2435 2436 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2437 2438 // Try to categorize the address. 2439 Compile::AliasType* alias_type = C->alias_type(adr_type); 2440 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2441 2442 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2443 alias_type->adr_type() == TypeAryPtr::RANGE) { 2444 return false; // not supported 2445 } 2446 2447 bool mismatched = false; 2448 BasicType bt = alias_type->basic_type(); 2449 if (bt != T_ILLEGAL) { 2450 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2451 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2452 // Alias type doesn't differentiate between byte[] and boolean[]). 2453 // Use address type to get the element type. 2454 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2455 } 2456 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2457 // accessing an array field with getObject is not a mismatch 2458 bt = T_OBJECT; 2459 } 2460 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2461 // Don't intrinsify mismatched object accesses 2462 return false; 2463 } 2464 mismatched = (bt != type); 2465 } else if (alias_type->adr_type()->isa_oopptr()) { 2466 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched 2467 } 2468 2469 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2470 2471 // First guess at the value type. 2472 const Type *value_type = Type::get_const_basic_type(type); 2473 2474 // We will need memory barriers unless we can determine a unique 2475 // alias category for this reference. (Note: If for some reason 2476 // the barriers get omitted and the unsafe reference begins to "pollute" 2477 // the alias analysis of the rest of the graph, either Compile::can_alias 2478 // or Compile::must_alias will throw a diagnostic assert.) 2479 bool need_mem_bar = false; 2480 switch (kind) { 2481 case Relaxed: 2482 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap; 2483 break; 2484 case Opaque: 2485 // Opaque uses CPUOrder membars for protection against code movement. 2486 case Acquire: 2487 case Release: 2488 case Volatile: 2489 need_mem_bar = true; 2490 break; 2491 default: 2492 ShouldNotReachHere(); 2493 } 2494 2495 // Some accesses require access atomicity for all types, notably longs and doubles. 2496 // When AlwaysAtomicAccesses is enabled, all accesses are atomic. 2497 bool requires_atomic_access = false; 2498 switch (kind) { 2499 case Relaxed: 2500 requires_atomic_access = AlwaysAtomicAccesses; 2501 break; 2502 case Opaque: 2503 // Opaque accesses are atomic. 2504 case Acquire: 2505 case Release: 2506 case Volatile: 2507 requires_atomic_access = true; 2508 break; 2509 default: 2510 ShouldNotReachHere(); 2511 } 2512 2513 // Figure out the memory ordering. 2514 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd 2515 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store); 2516 2517 // If we are reading the value of the referent field of a Reference 2518 // object (either by using Unsafe directly or through reflection) 2519 // then, if G1 is enabled, we need to record the referent in an 2520 // SATB log buffer using the pre-barrier mechanism. 2521 // Also we need to add memory barrier to prevent commoning reads 2522 // from this field across safepoint since GC can change its value. 2523 bool need_read_barrier = !is_store && 2524 offset != top() && heap_base_oop != top(); 2525 2526 if (!is_store && type == T_OBJECT) { 2527 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2528 if (tjp != NULL) { 2529 value_type = tjp; 2530 } 2531 } 2532 2533 receiver = null_check(receiver); 2534 if (stopped()) { 2535 return true; 2536 } 2537 // Heap pointers get a null-check from the interpreter, 2538 // as a courtesy. However, this is not guaranteed by Unsafe, 2539 // and it is not possible to fully distinguish unintended nulls 2540 // from intended ones in this API. 2541 2542 // We need to emit leading and trailing CPU membars (see below) in 2543 // addition to memory membars for special access modes. This is a little 2544 // too strong, but avoids the need to insert per-alias-type 2545 // volatile membars (for stores; compare Parse::do_put_xxx), which 2546 // we cannot do effectively here because we probably only have a 2547 // rough approximation of type. 2548 2549 switch(kind) { 2550 case Relaxed: 2551 case Opaque: 2552 case Acquire: 2553 break; 2554 case Release: 2555 case Volatile: 2556 if (is_store) { 2557 insert_mem_bar(Op_MemBarRelease); 2558 } else { 2559 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2560 insert_mem_bar(Op_MemBarVolatile); 2561 } 2562 } 2563 break; 2564 default: 2565 ShouldNotReachHere(); 2566 } 2567 2568 // Memory barrier to prevent normal and 'unsafe' accesses from 2569 // bypassing each other. Happens after null checks, so the 2570 // exception paths do not take memory state from the memory barrier, 2571 // so there's no problems making a strong assert about mixing users 2572 // of safe & unsafe memory. 2573 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2574 2575 if (!is_store) { 2576 Node* p = NULL; 2577 // Try to constant fold a load from a constant field 2578 ciField* field = alias_type->field(); 2579 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { 2580 // final or stable field 2581 p = make_constant_from_field(field, heap_base_oop); 2582 } 2583 if (p == NULL) { 2584 // To be valid, unsafe loads may depend on other conditions than 2585 // the one that guards them: pin the Load node 2586 LoadNode::ControlDependency dep = LoadNode::Pinned; 2587 Node* ctrl = control(); 2588 // non volatile loads may be able to float 2589 if (!need_mem_bar && adr_type->isa_instptr()) { 2590 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null"); 2591 intptr_t offset = Type::OffsetBot; 2592 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset); 2593 if (offset >= 0) { 2594 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper()); 2595 if (offset < s) { 2596 // Guaranteed to be a valid access, no need to pin it 2597 dep = LoadNode::DependsOnlyOnTest; 2598 ctrl = NULL; 2599 } 2600 } 2601 } 2602 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched); 2603 // load value 2604 switch (type) { 2605 case T_BOOLEAN: 2606 { 2607 // Normalize the value returned by getBoolean in the following cases 2608 if (mismatched || 2609 heap_base_oop == top() || // - heap_base_oop is NULL or 2610 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL 2611 // and the unsafe access is made to large offset 2612 // (i.e., larger than the maximum offset necessary for any 2613 // field access) 2614 ) { 2615 IdealKit ideal = IdealKit(this); 2616 #define __ ideal. 2617 IdealVariable normalized_result(ideal); 2618 __ declarations_done(); 2619 __ set(normalized_result, p); 2620 __ if_then(p, BoolTest::ne, ideal.ConI(0)); 2621 __ set(normalized_result, ideal.ConI(1)); 2622 ideal.end_if(); 2623 final_sync(ideal); 2624 p = __ value(normalized_result); 2625 #undef __ 2626 } 2627 } 2628 case T_CHAR: 2629 case T_BYTE: 2630 case T_SHORT: 2631 case T_INT: 2632 case T_LONG: 2633 case T_FLOAT: 2634 case T_DOUBLE: 2635 break; 2636 case T_OBJECT: 2637 if (need_read_barrier) { 2638 // We do not require a mem bar inside pre_barrier if need_mem_bar 2639 // is set: the barriers would be emitted by us. 2640 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar); 2641 } 2642 if (UseZGC) { 2643 if (!VerifyLoadBarriers) { 2644 p = load_barrier(p, adr); 2645 } else { 2646 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2647 p = load_barrier(p, adr); 2648 } else { 2649 IdealKit ideal(this); 2650 IdealVariable res(ideal); 2651 #define __ ideal. 2652 __ declarations_done(); 2653 __ set(res, p); 2654 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2655 sync_kit(ideal); 2656 p = load_barrier(p, adr); 2657 __ set(res, p); 2658 __ sync_kit(this); 2659 } __ end_if(); 2660 final_sync(ideal); 2661 p = __ value(res); 2662 #undef __ 2663 } 2664 } 2665 } 2666 break; 2667 case T_ADDRESS: 2668 // Cast to an int type. 2669 p = _gvn.transform(new CastP2XNode(NULL, p)); 2670 p = ConvX2UL(p); 2671 break; 2672 default: 2673 fatal("unexpected type %d: %s", type, type2name(type)); 2674 break; 2675 } 2676 } 2677 // The load node has the control of the preceding MemBarCPUOrder. All 2678 // following nodes will have the control of the MemBarCPUOrder inserted at 2679 // the end of this method. So, pushing the load onto the stack at a later 2680 // point is fine. 2681 set_result(p); 2682 } else { 2683 // place effect of store into memory 2684 switch (type) { 2685 case T_DOUBLE: 2686 val = dstore_rounding(val); 2687 break; 2688 case T_ADDRESS: 2689 // Repackage the long as a pointer. 2690 val = ConvL2X(val); 2691 val = _gvn.transform(new CastX2PNode(val)); 2692 break; 2693 default: 2694 break; 2695 } 2696 2697 if (type == T_OBJECT) { 2698 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2699 } else { 2700 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched); 2701 } 2702 } 2703 2704 switch(kind) { 2705 case Relaxed: 2706 case Opaque: 2707 case Release: 2708 break; 2709 case Acquire: 2710 case Volatile: 2711 if (!is_store) { 2712 insert_mem_bar(Op_MemBarAcquire); 2713 } else { 2714 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2715 insert_mem_bar(Op_MemBarVolatile); 2716 } 2717 } 2718 break; 2719 default: 2720 ShouldNotReachHere(); 2721 } 2722 2723 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2724 2725 return true; 2726 } 2727 2728 bool LibraryCallKit::loadstore_requires_writeback_barrier(const LoadStoreKind kind) { 2729 return (kind != LS_get_set); 2730 } 2731 2732 //----------------------------inline_unsafe_load_store---------------------------- 2733 // This method serves a couple of different customers (depending on LoadStoreKind): 2734 // 2735 // LS_cmp_swap: 2736 // 2737 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x); 2738 // boolean compareAndSetInt( Object o, long offset, int expected, int x); 2739 // boolean compareAndSetLong( Object o, long offset, long expected, long x); 2740 // 2741 // LS_cmp_swap_weak: 2742 // 2743 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x); 2744 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x); 2745 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x); 2746 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x); 2747 // 2748 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x); 2749 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x); 2750 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x); 2751 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x); 2752 // 2753 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x); 2754 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x); 2755 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x); 2756 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x); 2757 // 2758 // LS_cmp_exchange: 2759 // 2760 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x); 2761 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x); 2762 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x); 2763 // 2764 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2765 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2766 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2767 // 2768 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2769 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2770 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2771 // 2772 // LS_get_add: 2773 // 2774 // int getAndAddInt( Object o, long offset, int delta) 2775 // long getAndAddLong(Object o, long offset, long delta) 2776 // 2777 // LS_get_set: 2778 // 2779 // int getAndSet(Object o, long offset, int newValue) 2780 // long getAndSet(Object o, long offset, long newValue) 2781 // Object getAndSet(Object o, long offset, Object newValue) 2782 // 2783 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2784 // This basic scheme here is the same as inline_unsafe_access, but 2785 // differs in enough details that combining them would make the code 2786 // overly confusing. (This is a true fact! I originally combined 2787 // them, but even I was confused by it!) As much code/comments as 2788 // possible are retained from inline_unsafe_access though to make 2789 // the correspondences clearer. - dl 2790 2791 if (callee()->is_static()) return false; // caller must have the capability! 2792 2793 #ifndef PRODUCT 2794 BasicType rtype; 2795 { 2796 ResourceMark rm; 2797 // Check the signatures. 2798 ciSignature* sig = callee()->signature(); 2799 rtype = sig->return_type()->basic_type(); 2800 switch(kind) { 2801 case LS_get_add: 2802 case LS_get_set: { 2803 // Check the signatures. 2804 #ifdef ASSERT 2805 assert(rtype == type, "get and set must return the expected type"); 2806 assert(sig->count() == 3, "get and set has 3 arguments"); 2807 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2808 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2809 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2810 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2811 #endif // ASSERT 2812 break; 2813 } 2814 case LS_cmp_swap: 2815 case LS_cmp_swap_weak: { 2816 // Check the signatures. 2817 #ifdef ASSERT 2818 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2819 assert(sig->count() == 4, "CAS has 4 arguments"); 2820 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2821 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2822 #endif // ASSERT 2823 break; 2824 } 2825 case LS_cmp_exchange: { 2826 // Check the signatures. 2827 #ifdef ASSERT 2828 assert(rtype == type, "CAS must return the expected type"); 2829 assert(sig->count() == 4, "CAS has 4 arguments"); 2830 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2831 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2832 #endif // ASSERT 2833 break; 2834 } 2835 default: 2836 ShouldNotReachHere(); 2837 } 2838 } 2839 #endif //PRODUCT 2840 2841 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2842 2843 // Get arguments: 2844 Node* receiver = NULL; 2845 Node* base = NULL; 2846 Node* offset = NULL; 2847 Node* oldval = NULL; 2848 Node* newval = NULL; 2849 switch(kind) { 2850 case LS_cmp_swap: 2851 case LS_cmp_swap_weak: 2852 case LS_cmp_exchange: { 2853 const bool two_slot_type = type2size[type] == 2; 2854 receiver = argument(0); // type: oop 2855 base = argument(1); // type: oop 2856 offset = argument(2); // type: long 2857 oldval = argument(4); // type: oop, int, or long 2858 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2859 break; 2860 } 2861 case LS_get_add: 2862 case LS_get_set: { 2863 receiver = argument(0); // type: oop 2864 base = argument(1); // type: oop 2865 offset = argument(2); // type: long 2866 oldval = NULL; 2867 newval = argument(4); // type: oop, int, or long 2868 break; 2869 } 2870 default: 2871 ShouldNotReachHere(); 2872 } 2873 2874 // Build field offset expression. 2875 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2876 // to be plain byte offsets, which are also the same as those accepted 2877 // by oopDesc::field_addr. 2878 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2879 // 32-bit machines ignore the high half of long offsets 2880 offset = ConvL2X(offset); 2881 Node* adr = make_unsafe_address(base, offset, type, false); 2882 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2883 2884 Compile::AliasType* alias_type = C->alias_type(adr_type); 2885 BasicType bt = alias_type->basic_type(); 2886 if (bt != T_ILLEGAL && 2887 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2888 // Don't intrinsify mismatched object accesses. 2889 return false; 2890 } 2891 2892 // For CAS, unlike inline_unsafe_access, there seems no point in 2893 // trying to refine types. Just use the coarse types here. 2894 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2895 const Type *value_type = Type::get_const_basic_type(type); 2896 2897 switch (kind) { 2898 case LS_get_set: 2899 case LS_cmp_exchange: { 2900 if (type == T_OBJECT) { 2901 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2902 if (tjp != NULL) { 2903 value_type = tjp; 2904 } 2905 } 2906 break; 2907 } 2908 case LS_cmp_swap: 2909 case LS_cmp_swap_weak: 2910 case LS_get_add: 2911 break; 2912 default: 2913 ShouldNotReachHere(); 2914 } 2915 2916 // Null check receiver. 2917 receiver = null_check(receiver); 2918 if (stopped()) { 2919 return true; 2920 } 2921 2922 int alias_idx = C->get_alias_index(adr_type); 2923 2924 // Memory-model-wise, a LoadStore acts like a little synchronized 2925 // block, so needs barriers on each side. These don't translate 2926 // into actual barriers on most machines, but we still need rest of 2927 // compiler to respect ordering. 2928 2929 switch (access_kind) { 2930 case Relaxed: 2931 case Acquire: 2932 break; 2933 case Release: 2934 insert_mem_bar(Op_MemBarRelease); 2935 break; 2936 case Volatile: 2937 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2938 insert_mem_bar(Op_MemBarVolatile); 2939 } else { 2940 insert_mem_bar(Op_MemBarRelease); 2941 } 2942 break; 2943 default: 2944 ShouldNotReachHere(); 2945 } 2946 insert_mem_bar(Op_MemBarCPUOrder); 2947 2948 // Figure out the memory ordering. 2949 MemNode::MemOrd mo = access_kind_to_memord(access_kind); 2950 2951 // 4984716: MemBars must be inserted before this 2952 // memory node in order to avoid a false 2953 // dependency which will confuse the scheduler. 2954 Node *mem = memory(alias_idx); 2955 2956 bool add_scmemproj = true; // Depending on barrier and kind we sometimes need a scmemproj to keep barrier alive 2957 2958 // For now, we handle only those cases that actually exist: ints, 2959 // longs, and Object. Adding others should be straightforward. 2960 Node* load_store = NULL; 2961 switch(type) { 2962 case T_BYTE: 2963 switch(kind) { 2964 case LS_get_add: 2965 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type)); 2966 break; 2967 case LS_get_set: 2968 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type)); 2969 break; 2970 case LS_cmp_swap_weak: 2971 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo)); 2972 break; 2973 case LS_cmp_swap: 2974 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo)); 2975 break; 2976 case LS_cmp_exchange: 2977 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo)); 2978 break; 2979 default: 2980 ShouldNotReachHere(); 2981 } 2982 break; 2983 case T_SHORT: 2984 switch(kind) { 2985 case LS_get_add: 2986 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type)); 2987 break; 2988 case LS_get_set: 2989 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type)); 2990 break; 2991 case LS_cmp_swap_weak: 2992 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo)); 2993 break; 2994 case LS_cmp_swap: 2995 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo)); 2996 break; 2997 case LS_cmp_exchange: 2998 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo)); 2999 break; 3000 default: 3001 ShouldNotReachHere(); 3002 } 3003 break; 3004 case T_INT: 3005 switch(kind) { 3006 case LS_get_add: 3007 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 3008 break; 3009 case LS_get_set: 3010 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 3011 break; 3012 case LS_cmp_swap_weak: 3013 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo)); 3014 break; 3015 case LS_cmp_swap: 3016 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo)); 3017 break; 3018 case LS_cmp_exchange: 3019 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo)); 3020 break; 3021 default: 3022 ShouldNotReachHere(); 3023 } 3024 break; 3025 case T_LONG: 3026 switch(kind) { 3027 case LS_get_add: 3028 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 3029 break; 3030 case LS_get_set: 3031 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 3032 break; 3033 case LS_cmp_swap_weak: 3034 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo)); 3035 break; 3036 case LS_cmp_swap: 3037 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo)); 3038 break; 3039 case LS_cmp_exchange: 3040 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo)); 3041 break; 3042 default: 3043 ShouldNotReachHere(); 3044 } 3045 break; 3046 case T_OBJECT: 3047 // Transformation of a value which could be NULL pointer (CastPP #NULL) 3048 // could be delayed during Parse (for example, in adjust_map_after_if()). 3049 // Execute transformation here to avoid barrier generation in such case. 3050 if (_gvn.type(newval) == TypePtr::NULL_PTR) 3051 newval = _gvn.makecon(TypePtr::NULL_PTR); 3052 3053 // Reference stores need a store barrier. 3054 switch(kind) { 3055 case LS_get_set: { 3056 // If pre-barrier must execute before the oop store, old value will require do_load here. 3057 if (!can_move_pre_barrier()) { 3058 pre_barrier(true /* do_load*/, 3059 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 3060 NULL /* pre_val*/, 3061 T_OBJECT); 3062 } // Else move pre_barrier to use load_store value, see below. 3063 break; 3064 } 3065 case LS_cmp_swap_weak: 3066 case LS_cmp_swap: 3067 case LS_cmp_exchange: { 3068 // Same as for newval above: 3069 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 3070 oldval = _gvn.makecon(TypePtr::NULL_PTR); 3071 } 3072 // The only known value which might get overwritten is oldval. 3073 pre_barrier(false /* do_load */, 3074 control(), NULL, NULL, max_juint, NULL, NULL, 3075 oldval /* pre_val */, 3076 T_OBJECT); 3077 break; 3078 } 3079 default: 3080 ShouldNotReachHere(); 3081 } 3082 3083 #ifdef _LP64 3084 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3085 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 3086 3087 switch(kind) { 3088 case LS_get_set: 3089 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop())); 3090 break; 3091 case LS_cmp_swap_weak: { 3092 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 3093 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo)); 3094 break; 3095 } 3096 case LS_cmp_swap: { 3097 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 3098 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo)); 3099 break; 3100 } 3101 case LS_cmp_exchange: { 3102 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 3103 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 3104 break; 3105 } 3106 default: 3107 ShouldNotReachHere(); 3108 } 3109 } else 3110 #endif 3111 switch (kind) { 3112 case LS_get_set: 3113 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 3114 break; 3115 case LS_cmp_swap_weak: 3116 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo)); 3117 break; 3118 case LS_cmp_swap: 3119 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo)); 3120 break; 3121 case LS_cmp_exchange: 3122 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo)); 3123 break; 3124 default: 3125 ShouldNotReachHere(); 3126 } 3127 3128 // Optimization: expected_is_null 3129 // For CAS and weakCAS - if the expected value (oldval) is null, then 3130 // we can avoid expanding a load barrier (null can't have bad bits) 3131 3132 // CMH - Remove flags and simplify code when final variants are stable 3133 3134 if (UseZGC) { 3135 switch (kind) { 3136 case LS_get_set: 3137 break; 3138 case LS_cmp_swap_weak: 3139 { 3140 bool expected_is_null = (oldval->get_ptr_type() == TypePtr::NULL_PTR); 3141 if (!expected_is_null && C->directive()->UseWeakCASLoadBarrierOption) { 3142 add_scmemproj = false; 3143 load_store = make_cas_loadbarrier((CompareAndSwapNode*) load_store); 3144 } 3145 } 3146 break; 3147 case LS_cmp_swap: 3148 { 3149 bool expected_is_null = (oldval->get_ptr_type() == TypePtr::NULL_PTR); 3150 if (!expected_is_null && C->directive()->UseCASLoadBarrierOption) { 3151 add_scmemproj = false; 3152 load_store = make_cas_loadbarrier((CompareAndSwapNode*) load_store); 3153 } 3154 } 3155 break; 3156 case LS_cmp_exchange: 3157 if (C->directive()->UseCMPXLoadBarrierOption) { 3158 add_scmemproj = false; 3159 load_store = make_cmpx_loadbarrier((CompareAndExchangePNode*) load_store); 3160 } 3161 break; 3162 default: 3163 ShouldNotReachHere(); 3164 } 3165 } 3166 3167 // Emit the post barrier only when the actual store happened. This makes sense 3168 // to check only for LS_cmp_* that can fail to set the value. 3169 // LS_cmp_exchange does not produce any branches by default, so there is no 3170 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with 3171 // CompareAndExchange and move branches here, it would make sense to conditionalize 3172 // post_barriers for LS_cmp_exchange as well. 3173 // 3174 // CAS success path is marked more likely since we anticipate this is a performance 3175 // critical path, while CAS failure path can use the penalty for going through unlikely 3176 // path as backoff. Which is still better than doing a store barrier there. 3177 if (has_post_barrier()) { 3178 switch (kind) { 3179 case LS_get_set: 3180 case LS_cmp_exchange: { 3181 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 3182 break; 3183 } 3184 case LS_cmp_swap_weak: 3185 case LS_cmp_swap: { 3186 IdealKit ideal(this); 3187 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); { 3188 sync_kit(ideal); 3189 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 3190 ideal.sync_kit(this); 3191 } ideal.end_if(); 3192 final_sync(ideal); 3193 break; 3194 } 3195 default: 3196 ShouldNotReachHere(); 3197 } 3198 } 3199 break; 3200 default: 3201 fatal("unexpected type %d: %s", type, type2name(type)); 3202 break; 3203 } 3204 3205 // SCMemProjNodes represent the memory state of a LoadStore. Their 3206 // main role is to prevent LoadStore nodes from being optimized away 3207 // when their results aren't used. 3208 if (add_scmemproj) { 3209 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 3210 set_memory(proj, alias_idx); 3211 } 3212 3213 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) { 3214 #ifdef _LP64 3215 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3216 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 3217 } 3218 #endif 3219 if (UseZGC && (kind == LS_get_set) && C->directive()->UseSwapLoadBarrierOption) { 3220 load_store = load_barrier(load_store, adr, false, loadstore_requires_writeback_barrier(kind), false); 3221 } 3222 3223 if (can_move_pre_barrier()) { 3224 // Don't need to load pre_val. The old value is returned by load_store. 3225 // The pre_barrier can execute after the xchg as long as no safepoint 3226 // gets inserted between them. 3227 pre_barrier(false /* do_load */, 3228 control(), NULL, NULL, max_juint, NULL, NULL, 3229 load_store /* pre_val */, 3230 T_OBJECT); 3231 } 3232 } 3233 3234 // Add the trailing membar surrounding the access 3235 insert_mem_bar(Op_MemBarCPUOrder); 3236 3237 switch (access_kind) { 3238 case Relaxed: 3239 case Release: 3240 break; // do nothing 3241 case Acquire: 3242 case Volatile: 3243 insert_mem_bar(Op_MemBarAcquire); 3244 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code 3245 break; 3246 default: 3247 ShouldNotReachHere(); 3248 } 3249 3250 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3251 set_result(load_store); 3252 return true; 3253 } 3254 3255 Node* LibraryCallKit::make_cas_loadbarrier(CompareAndSwapNode* cas) { 3256 3257 assert(UseZGC, "Must be turned on"); 3258 assert(!UseCompressedOops, "Not allowed"); 3259 3260 Node* in_ctrl = cas->in(MemNode::Control); 3261 Node* in_mem = cas->in(MemNode::Memory); 3262 Node* in_adr = cas->in(MemNode::Address); 3263 Node* in_val = cas->in(MemNode::ValueIn); 3264 Node* in_expected = cas->in(LoadStoreConditionalNode::ExpectedIn); 3265 3266 float likely = PROB_LIKELY(0.999); 3267 3268 const TypePtr *adr_type = _gvn.type(in_adr)->isa_ptr(); 3269 Compile::AliasType* alias_type = C->alias_type(adr_type); 3270 int alias_idx = C->get_alias_index(adr_type); 3271 3272 // Outer check - true: continue, false: load and check 3273 Node* region = new RegionNode(3); 3274 Node* phi = new PhiNode(region, TypeInt::BOOL); 3275 Node* phi_mem = new PhiNode(region, Type::MEMORY, adr_type); 3276 3277 // Inner check - is the healed ref equal to the expected 3278 Node* region2 = new RegionNode(3); 3279 Node* phi2 = new PhiNode(region2, TypeInt::BOOL); 3280 Node* phi_mem2 = new PhiNode(region2, Type::MEMORY, adr_type); 3281 3282 // CAS node returns 0 or 1 3283 Node* cmp = _gvn.transform(new CmpINode(cas, intcon(0))); 3284 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne))->as_Bool(); 3285 IfNode* iff = _gvn.transform(new IfNode(in_ctrl, bol, likely, COUNT_UNKNOWN))->as_If(); 3286 Node* then = _gvn.transform(new IfTrueNode(iff)); 3287 Node* elsen = _gvn.transform(new IfFalseNode(iff)); 3288 3289 Node* scmemproj1 = _gvn.transform(new SCMemProjNode(cas)); 3290 3291 set_memory(scmemproj1, alias_idx); 3292 phi_mem->init_req(1, scmemproj1); 3293 phi_mem2->init_req(2, scmemproj1); 3294 3295 // CAS fail - reload and heal oop 3296 Node* reload = make_load(elsen, in_adr, TypeOopPtr::BOTTOM, T_OBJECT, MemNode::unordered); 3297 Node* barrier = _gvn.transform(new LoadBarrierNode(C, elsen, scmemproj1, reload, in_adr, false, true, false)); 3298 Node* barrierctrl = _gvn.transform(new ProjNode(barrier, LoadBarrierNode::Control)); 3299 Node* barrierdata = _gvn.transform(new ProjNode(barrier, LoadBarrierNode::Oop)); 3300 3301 // Check load 3302 Node* tmpX = _gvn.transform(new CastP2XNode(NULL, barrierdata)); 3303 Node* in_expX = _gvn.transform(new CastP2XNode(NULL, in_expected)); 3304 Node* cmp2 = _gvn.transform(new CmpXNode(tmpX, in_expX)); 3305 Node *bol2 = _gvn.transform(new BoolNode(cmp2, BoolTest::ne))->as_Bool(); 3306 IfNode* iff2 = _gvn.transform(new IfNode(barrierctrl, bol2, likely, COUNT_UNKNOWN))->as_If(); 3307 Node* then2 = _gvn.transform(new IfTrueNode(iff2)); 3308 Node* elsen2 = _gvn.transform(new IfFalseNode(iff2)); 3309 3310 // redo CAS 3311 Node* cas2 = _gvn.transform(new CompareAndSwapPNode(elsen2, memory(alias_idx), in_adr, in_val, in_expected, cas->order())); 3312 Node* scmemproj2 = _gvn.transform(new SCMemProjNode(cas2)); 3313 set_control(elsen2); 3314 set_memory(scmemproj2, alias_idx); 3315 3316 // Merge inner flow - check if healed oop was equal too expected. 3317 region2->set_req(1, control()); 3318 region2->set_req(2, then2); 3319 phi2->set_req(1, cas2); 3320 phi2->set_req(2, intcon(0)); 3321 phi_mem2->init_req(1, scmemproj2); 3322 set_memory(phi_mem2, alias_idx); 3323 3324 // Merge outer flow - then check if first cas succeded 3325 region->set_req(1, then); 3326 region->set_req(2, region2); 3327 phi->set_req(1, intcon(1)); 3328 phi->set_req(2, phi2); 3329 phi_mem->init_req(2, phi_mem2); 3330 set_memory(phi_mem, alias_idx); 3331 3332 _gvn.transform(region2); 3333 _gvn.transform(phi2); 3334 _gvn.transform(phi_mem2); 3335 _gvn.transform(region); 3336 _gvn.transform(phi); 3337 _gvn.transform(phi_mem); 3338 3339 set_control(region); 3340 insert_mem_bar(Op_MemBarCPUOrder); 3341 3342 return phi; 3343 } 3344 3345 Node* LibraryCallKit::make_cmpx_loadbarrier(CompareAndExchangePNode* cmpx) { 3346 3347 assert(UseZGC, "Must be turned on"); 3348 assert(!UseCompressedOops, "Not allowed"); 3349 3350 Node* in_ctrl = cmpx->in(MemNode::Control); 3351 Node* in_mem = cmpx->in(MemNode::Memory); 3352 Node* in_adr = cmpx->in(MemNode::Address); 3353 Node* in_val = cmpx->in(MemNode::ValueIn); 3354 Node* in_expected = cmpx->in(LoadStoreConditionalNode::ExpectedIn); 3355 3356 float likely = PROB_LIKELY(0.999); 3357 3358 const TypePtr *adr_type = cmpx->get_ptr_type(); 3359 Compile::AliasType* alias_type = C->alias_type(adr_type); 3360 int alias_idx = C->get_alias_index(adr_type); 3361 3362 // Outer check - true: continue, false: load and check 3363 Node* region = new RegionNode(3); 3364 Node* phi = new PhiNode(region, adr_type); 3365 3366 // Inner check - is the healed ref equal to the expected 3367 Node* region2 = new RegionNode(3); 3368 Node* phi2 = new PhiNode(region2, adr_type); 3369 3370 // Check if cmpx succeded 3371 Node* cmp = _gvn.transform(new CmpPNode(cmpx, in_expected)); 3372 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq))->as_Bool(); 3373 IfNode* iff = _gvn.transform(new IfNode(in_ctrl, bol, likely, COUNT_UNKNOWN))->as_If(); 3374 Node* then = _gvn.transform(new IfTrueNode(iff)); 3375 Node* elsen = _gvn.transform(new IfFalseNode(iff)); 3376 3377 Node* scmemproj1 = _gvn.transform(new SCMemProjNode(cmpx)); 3378 set_memory(scmemproj1, alias_idx); 3379 3380 // CAS fail - reload and heal oop 3381 Node* reload = make_load(elsen, in_adr, TypeOopPtr::BOTTOM, T_OBJECT, MemNode::unordered); 3382 Node* barrier = _gvn.transform(new LoadBarrierNode(C, elsen, scmemproj1, reload, in_adr, false, true, false)); 3383 Node* barrierctrl = _gvn.transform(new ProjNode(barrier, LoadBarrierNode::Control)); 3384 Node* barrierdata = _gvn.transform(new ProjNode(barrier, LoadBarrierNode::Oop)); 3385 3386 // Check load 3387 Node* tmpX = _gvn.transform(new CastP2XNode(NULL, barrierdata)); 3388 Node* in_expX = _gvn.transform(new CastP2XNode(NULL, in_expected)); 3389 Node* cmp2 = _gvn.transform(new CmpXNode(tmpX, in_expX)); 3390 Node *bol2 = _gvn.transform(new BoolNode(cmp2, BoolTest::ne))->as_Bool(); 3391 IfNode* iff2 = _gvn.transform(new IfNode(barrierctrl, bol2, likely, COUNT_UNKNOWN))->as_If(); 3392 Node* then2 = _gvn.transform(new IfTrueNode(iff2)); 3393 Node* elsen2 = _gvn.transform(new IfFalseNode(iff2)); 3394 3395 // Redo CAS 3396 Node* cmpx2 = _gvn.transform(new CompareAndExchangePNode(elsen2, memory(alias_idx), in_adr, in_val, in_expected, adr_type, cmpx->get_ptr_type(), cmpx->order())); 3397 Node* scmemproj2 = _gvn.transform(new SCMemProjNode(cmpx2)); 3398 set_control(elsen2); 3399 set_memory(scmemproj2, alias_idx); 3400 3401 // Merge inner flow - check if healed oop was equal too expected. 3402 region2->set_req(1, control()); 3403 region2->set_req(2, then2); 3404 phi2->set_req(1, cmpx2); 3405 phi2->set_req(2, barrierdata); 3406 3407 // Merge outer flow - then check if first cas succeded 3408 region->set_req(1, then); 3409 region->set_req(2, region2); 3410 phi->set_req(1, cmpx); 3411 phi->set_req(2, phi2); 3412 3413 _gvn.transform(region2); 3414 _gvn.transform(phi2); 3415 _gvn.transform(region); 3416 _gvn.transform(phi); 3417 3418 set_control(region); 3419 set_memory(in_mem, alias_idx); 3420 insert_mem_bar(Op_MemBarCPUOrder); 3421 3422 return phi; 3423 } 3424 3425 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) { 3426 MemNode::MemOrd mo = MemNode::unset; 3427 switch(kind) { 3428 case Opaque: 3429 case Relaxed: mo = MemNode::unordered; break; 3430 case Acquire: mo = MemNode::acquire; break; 3431 case Release: mo = MemNode::release; break; 3432 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break; 3433 default: 3434 ShouldNotReachHere(); 3435 } 3436 guarantee(mo != MemNode::unset, "Should select memory ordering"); 3437 return mo; 3438 } 3439 3440 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) { 3441 MemNode::MemOrd mo = MemNode::unset; 3442 switch(kind) { 3443 case Opaque: 3444 case Relaxed: mo = MemNode::unordered; break; 3445 case Acquire: mo = MemNode::acquire; break; 3446 case Release: mo = MemNode::release; break; 3447 case Volatile: mo = MemNode::seqcst; break; 3448 default: 3449 ShouldNotReachHere(); 3450 } 3451 guarantee(mo != MemNode::unset, "Should select memory ordering"); 3452 return mo; 3453 } 3454 3455 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3456 // Regardless of form, don't allow previous ld/st to move down, 3457 // then issue acquire, release, or volatile mem_bar. 3458 insert_mem_bar(Op_MemBarCPUOrder); 3459 switch(id) { 3460 case vmIntrinsics::_loadFence: 3461 insert_mem_bar(Op_LoadFence); 3462 return true; 3463 case vmIntrinsics::_storeFence: 3464 insert_mem_bar(Op_StoreFence); 3465 return true; 3466 case vmIntrinsics::_fullFence: 3467 insert_mem_bar(Op_MemBarVolatile); 3468 return true; 3469 default: 3470 fatal_unexpected_iid(id); 3471 return false; 3472 } 3473 } 3474 3475 bool LibraryCallKit::inline_onspinwait() { 3476 insert_mem_bar(Op_OnSpinWait); 3477 return true; 3478 } 3479 3480 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3481 if (!kls->is_Con()) { 3482 return true; 3483 } 3484 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3485 if (klsptr == NULL) { 3486 return true; 3487 } 3488 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3489 // don't need a guard for a klass that is already initialized 3490 return !ik->is_initialized(); 3491 } 3492 3493 //----------------------------inline_unsafe_allocate--------------------------- 3494 // public native Object Unsafe.allocateInstance(Class<?> cls); 3495 bool LibraryCallKit::inline_unsafe_allocate() { 3496 if (callee()->is_static()) return false; // caller must have the capability! 3497 3498 null_check_receiver(); // null-check, then ignore 3499 Node* cls = null_check(argument(1)); 3500 if (stopped()) return true; 3501 3502 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3503 kls = null_check(kls); 3504 if (stopped()) return true; // argument was like int.class 3505 3506 Node* test = NULL; 3507 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3508 // Note: The argument might still be an illegal value like 3509 // Serializable.class or Object[].class. The runtime will handle it. 3510 // But we must make an explicit check for initialization. 3511 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3512 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3513 // can generate code to load it as unsigned byte. 3514 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3515 Node* bits = intcon(InstanceKlass::fully_initialized); 3516 test = _gvn.transform(new SubINode(inst, bits)); 3517 // The 'test' is non-zero if we need to take a slow path. 3518 } 3519 3520 Node* obj = new_instance(kls, test); 3521 set_result(obj); 3522 return true; 3523 } 3524 3525 //------------------------inline_native_time_funcs-------------- 3526 // inline code for System.currentTimeMillis() and System.nanoTime() 3527 // these have the same type and signature 3528 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3529 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3530 const TypePtr* no_memory_effects = NULL; 3531 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3532 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3533 #ifdef ASSERT 3534 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3535 assert(value_top == top(), "second value must be top"); 3536 #endif 3537 set_result(value); 3538 return true; 3539 } 3540 3541 #ifdef TRACE_HAVE_INTRINSICS 3542 3543 /* 3544 * oop -> myklass 3545 * myklass->trace_id |= USED 3546 * return myklass->trace_id & ~0x3 3547 */ 3548 bool LibraryCallKit::inline_native_classID() { 3549 Node* cls = null_check(argument(0), T_OBJECT); 3550 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3551 kls = null_check(kls, T_OBJECT); 3552 3553 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET; 3554 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3555 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3556 3557 Node* clsused = longcon(0x01l); // set the class bit 3558 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3559 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3560 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3561 3562 #ifdef TRACE_ID_META_BITS 3563 Node* mbits = longcon(~TRACE_ID_META_BITS); 3564 tvalue = _gvn.transform(new AndLNode(tvalue, mbits)); 3565 #endif 3566 #ifdef TRACE_ID_CLASS_SHIFT 3567 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT); 3568 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits)); 3569 #endif 3570 3571 set_result(tvalue); 3572 return true; 3573 3574 } 3575 3576 bool LibraryCallKit::inline_native_getBufferWriter() { 3577 Node* tls_ptr = _gvn.transform(new ThreadLocalNode()); 3578 3579 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr, 3580 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET) 3581 ); 3582 3583 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered); 3584 3585 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) ); 3586 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) ); 3587 3588 IfNode* iff_jobj_null = 3589 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN); 3590 3591 enum { _normal_path = 1, 3592 _null_path = 2, 3593 PATH_LIMIT }; 3594 3595 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3596 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM); 3597 3598 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null)); 3599 result_rgn->init_req(_null_path, jobj_is_null); 3600 result_val->init_req(_null_path, null()); 3601 3602 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null)); 3603 result_rgn->init_req(_normal_path, jobj_is_not_null); 3604 3605 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered); 3606 result_val->init_req(_normal_path, res); 3607 3608 set_result(result_rgn, result_val); 3609 3610 return true; 3611 } 3612 3613 #endif 3614 3615 //------------------------inline_native_currentThread------------------ 3616 bool LibraryCallKit::inline_native_currentThread() { 3617 Node* junk = NULL; 3618 set_result(generate_current_thread(junk)); 3619 return true; 3620 } 3621 3622 //------------------------inline_native_isInterrupted------------------ 3623 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3624 bool LibraryCallKit::inline_native_isInterrupted() { 3625 // Add a fast path to t.isInterrupted(clear_int): 3626 // (t == Thread.current() && 3627 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3628 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3629 // So, in the common case that the interrupt bit is false, 3630 // we avoid making a call into the VM. Even if the interrupt bit 3631 // is true, if the clear_int argument is false, we avoid the VM call. 3632 // However, if the receiver is not currentThread, we must call the VM, 3633 // because there must be some locking done around the operation. 3634 3635 // We only go to the fast case code if we pass two guards. 3636 // Paths which do not pass are accumulated in the slow_region. 3637 3638 enum { 3639 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3640 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3641 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3642 PATH_LIMIT 3643 }; 3644 3645 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3646 // out of the function. 3647 insert_mem_bar(Op_MemBarCPUOrder); 3648 3649 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3650 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3651 3652 RegionNode* slow_region = new RegionNode(1); 3653 record_for_igvn(slow_region); 3654 3655 // (a) Receiving thread must be the current thread. 3656 Node* rec_thr = argument(0); 3657 Node* tls_ptr = NULL; 3658 Node* cur_thr = generate_current_thread(tls_ptr); 3659 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3660 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3661 3662 generate_slow_guard(bol_thr, slow_region); 3663 3664 // (b) Interrupt bit on TLS must be false. 3665 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3666 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3667 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3668 3669 // Set the control input on the field _interrupted read to prevent it floating up. 3670 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3671 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3672 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3673 3674 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3675 3676 // First fast path: if (!TLS._interrupted) return false; 3677 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3678 result_rgn->init_req(no_int_result_path, false_bit); 3679 result_val->init_req(no_int_result_path, intcon(0)); 3680 3681 // drop through to next case 3682 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3683 3684 #ifndef _WINDOWS 3685 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3686 Node* clr_arg = argument(1); 3687 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3688 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3689 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3690 3691 // Second fast path: ... else if (!clear_int) return true; 3692 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3693 result_rgn->init_req(no_clear_result_path, false_arg); 3694 result_val->init_req(no_clear_result_path, intcon(1)); 3695 3696 // drop through to next case 3697 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3698 #else 3699 // To return true on Windows you must read the _interrupted field 3700 // and check the event state i.e. take the slow path. 3701 #endif // _WINDOWS 3702 3703 // (d) Otherwise, go to the slow path. 3704 slow_region->add_req(control()); 3705 set_control( _gvn.transform(slow_region)); 3706 3707 if (stopped()) { 3708 // There is no slow path. 3709 result_rgn->init_req(slow_result_path, top()); 3710 result_val->init_req(slow_result_path, top()); 3711 } else { 3712 // non-virtual because it is a private non-static 3713 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3714 3715 Node* slow_val = set_results_for_java_call(slow_call); 3716 // this->control() comes from set_results_for_java_call 3717 3718 Node* fast_io = slow_call->in(TypeFunc::I_O); 3719 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3720 3721 // These two phis are pre-filled with copies of of the fast IO and Memory 3722 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3723 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3724 3725 result_rgn->init_req(slow_result_path, control()); 3726 result_io ->init_req(slow_result_path, i_o()); 3727 result_mem->init_req(slow_result_path, reset_memory()); 3728 result_val->init_req(slow_result_path, slow_val); 3729 3730 set_all_memory(_gvn.transform(result_mem)); 3731 set_i_o( _gvn.transform(result_io)); 3732 } 3733 3734 C->set_has_split_ifs(true); // Has chance for split-if optimization 3735 set_result(result_rgn, result_val); 3736 return true; 3737 } 3738 3739 //---------------------------load_mirror_from_klass---------------------------- 3740 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3741 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3742 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3743 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3744 return make_load(NULL, load, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3745 } 3746 3747 //-----------------------load_klass_from_mirror_common------------------------- 3748 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3749 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3750 // and branch to the given path on the region. 3751 // If never_see_null, take an uncommon trap on null, so we can optimistically 3752 // compile for the non-null case. 3753 // If the region is NULL, force never_see_null = true. 3754 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3755 bool never_see_null, 3756 RegionNode* region, 3757 int null_path, 3758 int offset) { 3759 if (region == NULL) never_see_null = true; 3760 Node* p = basic_plus_adr(mirror, offset); 3761 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3762 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3763 Node* null_ctl = top(); 3764 kls = null_check_oop(kls, &null_ctl, never_see_null); 3765 if (region != NULL) { 3766 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3767 region->init_req(null_path, null_ctl); 3768 } else { 3769 assert(null_ctl == top(), "no loose ends"); 3770 } 3771 return kls; 3772 } 3773 3774 //--------------------(inline_native_Class_query helpers)--------------------- 3775 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3776 // Fall through if (mods & mask) == bits, take the guard otherwise. 3777 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3778 // Branch around if the given klass has the given modifier bit set. 3779 // Like generate_guard, adds a new path onto the region. 3780 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3781 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3782 Node* mask = intcon(modifier_mask); 3783 Node* bits = intcon(modifier_bits); 3784 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3785 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3786 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3787 return generate_fair_guard(bol, region); 3788 } 3789 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3790 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3791 } 3792 3793 //-------------------------inline_native_Class_query------------------- 3794 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3795 const Type* return_type = TypeInt::BOOL; 3796 Node* prim_return_value = top(); // what happens if it's a primitive class? 3797 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3798 bool expect_prim = false; // most of these guys expect to work on refs 3799 3800 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3801 3802 Node* mirror = argument(0); 3803 Node* obj = top(); 3804 3805 switch (id) { 3806 case vmIntrinsics::_isInstance: 3807 // nothing is an instance of a primitive type 3808 prim_return_value = intcon(0); 3809 obj = argument(1); 3810 break; 3811 case vmIntrinsics::_getModifiers: 3812 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3813 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3814 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3815 break; 3816 case vmIntrinsics::_isInterface: 3817 prim_return_value = intcon(0); 3818 break; 3819 case vmIntrinsics::_isArray: 3820 prim_return_value = intcon(0); 3821 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3822 break; 3823 case vmIntrinsics::_isPrimitive: 3824 prim_return_value = intcon(1); 3825 expect_prim = true; // obviously 3826 break; 3827 case vmIntrinsics::_getSuperclass: 3828 prim_return_value = null(); 3829 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3830 break; 3831 case vmIntrinsics::_getClassAccessFlags: 3832 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3833 return_type = TypeInt::INT; // not bool! 6297094 3834 break; 3835 default: 3836 fatal_unexpected_iid(id); 3837 break; 3838 } 3839 3840 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3841 if (mirror_con == NULL) return false; // cannot happen? 3842 3843 #ifndef PRODUCT 3844 if (C->print_intrinsics() || C->print_inlining()) { 3845 ciType* k = mirror_con->java_mirror_type(); 3846 if (k) { 3847 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3848 k->print_name(); 3849 tty->cr(); 3850 } 3851 } 3852 #endif 3853 3854 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3855 RegionNode* region = new RegionNode(PATH_LIMIT); 3856 record_for_igvn(region); 3857 PhiNode* phi = new PhiNode(region, return_type); 3858 3859 // The mirror will never be null of Reflection.getClassAccessFlags, however 3860 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3861 // if it is. See bug 4774291. 3862 3863 // For Reflection.getClassAccessFlags(), the null check occurs in 3864 // the wrong place; see inline_unsafe_access(), above, for a similar 3865 // situation. 3866 mirror = null_check(mirror); 3867 // If mirror or obj is dead, only null-path is taken. 3868 if (stopped()) return true; 3869 3870 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3871 3872 // Now load the mirror's klass metaobject, and null-check it. 3873 // Side-effects region with the control path if the klass is null. 3874 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3875 // If kls is null, we have a primitive mirror. 3876 phi->init_req(_prim_path, prim_return_value); 3877 if (stopped()) { set_result(region, phi); return true; } 3878 bool safe_for_replace = (region->in(_prim_path) == top()); 3879 3880 Node* p; // handy temp 3881 Node* null_ctl; 3882 3883 // Now that we have the non-null klass, we can perform the real query. 3884 // For constant classes, the query will constant-fold in LoadNode::Value. 3885 Node* query_value = top(); 3886 switch (id) { 3887 case vmIntrinsics::_isInstance: 3888 // nothing is an instance of a primitive type 3889 query_value = gen_instanceof(obj, kls, safe_for_replace); 3890 break; 3891 3892 case vmIntrinsics::_getModifiers: 3893 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3894 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3895 break; 3896 3897 case vmIntrinsics::_isInterface: 3898 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3899 if (generate_interface_guard(kls, region) != NULL) 3900 // A guard was added. If the guard is taken, it was an interface. 3901 phi->add_req(intcon(1)); 3902 // If we fall through, it's a plain class. 3903 query_value = intcon(0); 3904 break; 3905 3906 case vmIntrinsics::_isArray: 3907 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3908 if (generate_array_guard(kls, region) != NULL) 3909 // A guard was added. If the guard is taken, it was an array. 3910 phi->add_req(intcon(1)); 3911 // If we fall through, it's a plain class. 3912 query_value = intcon(0); 3913 break; 3914 3915 case vmIntrinsics::_isPrimitive: 3916 query_value = intcon(0); // "normal" path produces false 3917 break; 3918 3919 case vmIntrinsics::_getSuperclass: 3920 // The rules here are somewhat unfortunate, but we can still do better 3921 // with random logic than with a JNI call. 3922 // Interfaces store null or Object as _super, but must report null. 3923 // Arrays store an intermediate super as _super, but must report Object. 3924 // Other types can report the actual _super. 3925 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3926 if (generate_interface_guard(kls, region) != NULL) 3927 // A guard was added. If the guard is taken, it was an interface. 3928 phi->add_req(null()); 3929 if (generate_array_guard(kls, region) != NULL) 3930 // A guard was added. If the guard is taken, it was an array. 3931 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3932 // If we fall through, it's a plain class. Get its _super. 3933 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3934 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3935 null_ctl = top(); 3936 kls = null_check_oop(kls, &null_ctl); 3937 if (null_ctl != top()) { 3938 // If the guard is taken, Object.superClass is null (both klass and mirror). 3939 region->add_req(null_ctl); 3940 phi ->add_req(null()); 3941 } 3942 if (!stopped()) { 3943 query_value = load_mirror_from_klass(kls); 3944 } 3945 break; 3946 3947 case vmIntrinsics::_getClassAccessFlags: 3948 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3949 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3950 break; 3951 3952 default: 3953 fatal_unexpected_iid(id); 3954 break; 3955 } 3956 3957 // Fall-through is the normal case of a query to a real class. 3958 phi->init_req(1, query_value); 3959 region->init_req(1, control()); 3960 3961 C->set_has_split_ifs(true); // Has chance for split-if optimization 3962 set_result(region, phi); 3963 return true; 3964 } 3965 3966 //-------------------------inline_Class_cast------------------- 3967 bool LibraryCallKit::inline_Class_cast() { 3968 Node* mirror = argument(0); // Class 3969 Node* obj = argument(1); 3970 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3971 if (mirror_con == NULL) { 3972 return false; // dead path (mirror->is_top()). 3973 } 3974 if (obj == NULL || obj->is_top()) { 3975 return false; // dead path 3976 } 3977 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3978 3979 // First, see if Class.cast() can be folded statically. 3980 // java_mirror_type() returns non-null for compile-time Class constants. 3981 ciType* tm = mirror_con->java_mirror_type(); 3982 if (tm != NULL && tm->is_klass() && 3983 tp != NULL && tp->klass() != NULL) { 3984 if (!tp->klass()->is_loaded()) { 3985 // Don't use intrinsic when class is not loaded. 3986 return false; 3987 } else { 3988 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3989 if (static_res == Compile::SSC_always_true) { 3990 // isInstance() is true - fold the code. 3991 set_result(obj); 3992 return true; 3993 } else if (static_res == Compile::SSC_always_false) { 3994 // Don't use intrinsic, have to throw ClassCastException. 3995 // If the reference is null, the non-intrinsic bytecode will 3996 // be optimized appropriately. 3997 return false; 3998 } 3999 } 4000 } 4001 4002 // Bailout intrinsic and do normal inlining if exception path is frequent. 4003 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 4004 return false; 4005 } 4006 4007 // Generate dynamic checks. 4008 // Class.cast() is java implementation of _checkcast bytecode. 4009 // Do checkcast (Parse::do_checkcast()) optimizations here. 4010 4011 mirror = null_check(mirror); 4012 // If mirror is dead, only null-path is taken. 4013 if (stopped()) { 4014 return true; 4015 } 4016 4017 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 4018 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 4019 RegionNode* region = new RegionNode(PATH_LIMIT); 4020 record_for_igvn(region); 4021 4022 // Now load the mirror's klass metaobject, and null-check it. 4023 // If kls is null, we have a primitive mirror and 4024 // nothing is an instance of a primitive type. 4025 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 4026 4027 Node* res = top(); 4028 if (!stopped()) { 4029 Node* bad_type_ctrl = top(); 4030 // Do checkcast optimizations. 4031 res = gen_checkcast(obj, kls, &bad_type_ctrl); 4032 region->init_req(_bad_type_path, bad_type_ctrl); 4033 } 4034 if (region->in(_prim_path) != top() || 4035 region->in(_bad_type_path) != top()) { 4036 // Let Interpreter throw ClassCastException. 4037 PreserveJVMState pjvms(this); 4038 set_control(_gvn.transform(region)); 4039 uncommon_trap(Deoptimization::Reason_intrinsic, 4040 Deoptimization::Action_maybe_recompile); 4041 } 4042 if (!stopped()) { 4043 set_result(res); 4044 } 4045 return true; 4046 } 4047 4048 4049 //--------------------------inline_native_subtype_check------------------------ 4050 // This intrinsic takes the JNI calls out of the heart of 4051 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 4052 bool LibraryCallKit::inline_native_subtype_check() { 4053 // Pull both arguments off the stack. 4054 Node* args[2]; // two java.lang.Class mirrors: superc, subc 4055 args[0] = argument(0); 4056 args[1] = argument(1); 4057 Node* klasses[2]; // corresponding Klasses: superk, subk 4058 klasses[0] = klasses[1] = top(); 4059 4060 enum { 4061 // A full decision tree on {superc is prim, subc is prim}: 4062 _prim_0_path = 1, // {P,N} => false 4063 // {P,P} & superc!=subc => false 4064 _prim_same_path, // {P,P} & superc==subc => true 4065 _prim_1_path, // {N,P} => false 4066 _ref_subtype_path, // {N,N} & subtype check wins => true 4067 _both_ref_path, // {N,N} & subtype check loses => false 4068 PATH_LIMIT 4069 }; 4070 4071 RegionNode* region = new RegionNode(PATH_LIMIT); 4072 Node* phi = new PhiNode(region, TypeInt::BOOL); 4073 record_for_igvn(region); 4074 4075 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 4076 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 4077 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 4078 4079 // First null-check both mirrors and load each mirror's klass metaobject. 4080 int which_arg; 4081 for (which_arg = 0; which_arg <= 1; which_arg++) { 4082 Node* arg = args[which_arg]; 4083 arg = null_check(arg); 4084 if (stopped()) break; 4085 args[which_arg] = arg; 4086 4087 Node* p = basic_plus_adr(arg, class_klass_offset); 4088 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 4089 klasses[which_arg] = _gvn.transform(kls); 4090 } 4091 4092 // Having loaded both klasses, test each for null. 4093 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 4094 for (which_arg = 0; which_arg <= 1; which_arg++) { 4095 Node* kls = klasses[which_arg]; 4096 Node* null_ctl = top(); 4097 kls = null_check_oop(kls, &null_ctl, never_see_null); 4098 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 4099 region->init_req(prim_path, null_ctl); 4100 if (stopped()) break; 4101 klasses[which_arg] = kls; 4102 } 4103 4104 if (!stopped()) { 4105 // now we have two reference types, in klasses[0..1] 4106 Node* subk = klasses[1]; // the argument to isAssignableFrom 4107 Node* superk = klasses[0]; // the receiver 4108 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 4109 // now we have a successful reference subtype check 4110 region->set_req(_ref_subtype_path, control()); 4111 } 4112 4113 // If both operands are primitive (both klasses null), then 4114 // we must return true when they are identical primitives. 4115 // It is convenient to test this after the first null klass check. 4116 set_control(region->in(_prim_0_path)); // go back to first null check 4117 if (!stopped()) { 4118 // Since superc is primitive, make a guard for the superc==subc case. 4119 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 4120 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 4121 generate_guard(bol_eq, region, PROB_FAIR); 4122 if (region->req() == PATH_LIMIT+1) { 4123 // A guard was added. If the added guard is taken, superc==subc. 4124 region->swap_edges(PATH_LIMIT, _prim_same_path); 4125 region->del_req(PATH_LIMIT); 4126 } 4127 region->set_req(_prim_0_path, control()); // Not equal after all. 4128 } 4129 4130 // these are the only paths that produce 'true': 4131 phi->set_req(_prim_same_path, intcon(1)); 4132 phi->set_req(_ref_subtype_path, intcon(1)); 4133 4134 // pull together the cases: 4135 assert(region->req() == PATH_LIMIT, "sane region"); 4136 for (uint i = 1; i < region->req(); i++) { 4137 Node* ctl = region->in(i); 4138 if (ctl == NULL || ctl == top()) { 4139 region->set_req(i, top()); 4140 phi ->set_req(i, top()); 4141 } else if (phi->in(i) == NULL) { 4142 phi->set_req(i, intcon(0)); // all other paths produce 'false' 4143 } 4144 } 4145 4146 set_control(_gvn.transform(region)); 4147 set_result(_gvn.transform(phi)); 4148 return true; 4149 } 4150 4151 //---------------------generate_array_guard_common------------------------ 4152 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 4153 bool obj_array, bool not_array) { 4154 4155 if (stopped()) { 4156 return NULL; 4157 } 4158 4159 // If obj_array/non_array==false/false: 4160 // Branch around if the given klass is in fact an array (either obj or prim). 4161 // If obj_array/non_array==false/true: 4162 // Branch around if the given klass is not an array klass of any kind. 4163 // If obj_array/non_array==true/true: 4164 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 4165 // If obj_array/non_array==true/false: 4166 // Branch around if the kls is an oop array (Object[] or subtype) 4167 // 4168 // Like generate_guard, adds a new path onto the region. 4169 jint layout_con = 0; 4170 Node* layout_val = get_layout_helper(kls, layout_con); 4171 if (layout_val == NULL) { 4172 bool query = (obj_array 4173 ? Klass::layout_helper_is_objArray(layout_con) 4174 : Klass::layout_helper_is_array(layout_con)); 4175 if (query == not_array) { 4176 return NULL; // never a branch 4177 } else { // always a branch 4178 Node* always_branch = control(); 4179 if (region != NULL) 4180 region->add_req(always_branch); 4181 set_control(top()); 4182 return always_branch; 4183 } 4184 } 4185 // Now test the correct condition. 4186 jint nval = (obj_array 4187 ? (jint)(Klass::_lh_array_tag_type_value 4188 << Klass::_lh_array_tag_shift) 4189 : Klass::_lh_neutral_value); 4190 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 4191 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 4192 // invert the test if we are looking for a non-array 4193 if (not_array) btest = BoolTest(btest).negate(); 4194 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 4195 return generate_fair_guard(bol, region); 4196 } 4197 4198 4199 //-----------------------inline_native_newArray-------------------------- 4200 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 4201 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 4202 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 4203 Node* mirror; 4204 Node* count_val; 4205 if (uninitialized) { 4206 mirror = argument(1); 4207 count_val = argument(2); 4208 } else { 4209 mirror = argument(0); 4210 count_val = argument(1); 4211 } 4212 4213 mirror = null_check(mirror); 4214 // If mirror or obj is dead, only null-path is taken. 4215 if (stopped()) return true; 4216 4217 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 4218 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4219 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4220 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4221 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4222 4223 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 4224 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 4225 result_reg, _slow_path); 4226 Node* normal_ctl = control(); 4227 Node* no_array_ctl = result_reg->in(_slow_path); 4228 4229 // Generate code for the slow case. We make a call to newArray(). 4230 set_control(no_array_ctl); 4231 if (!stopped()) { 4232 // Either the input type is void.class, or else the 4233 // array klass has not yet been cached. Either the 4234 // ensuing call will throw an exception, or else it 4235 // will cache the array klass for next time. 4236 PreserveJVMState pjvms(this); 4237 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 4238 Node* slow_result = set_results_for_java_call(slow_call); 4239 // this->control() comes from set_results_for_java_call 4240 result_reg->set_req(_slow_path, control()); 4241 result_val->set_req(_slow_path, slow_result); 4242 result_io ->set_req(_slow_path, i_o()); 4243 result_mem->set_req(_slow_path, reset_memory()); 4244 } 4245 4246 set_control(normal_ctl); 4247 if (!stopped()) { 4248 // Normal case: The array type has been cached in the java.lang.Class. 4249 // The following call works fine even if the array type is polymorphic. 4250 // It could be a dynamic mix of int[], boolean[], Object[], etc. 4251 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 4252 result_reg->init_req(_normal_path, control()); 4253 result_val->init_req(_normal_path, obj); 4254 result_io ->init_req(_normal_path, i_o()); 4255 result_mem->init_req(_normal_path, reset_memory()); 4256 4257 if (uninitialized) { 4258 // Mark the allocation so that zeroing is skipped 4259 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 4260 alloc->maybe_set_complete(&_gvn); 4261 } 4262 } 4263 4264 // Return the combined state. 4265 set_i_o( _gvn.transform(result_io) ); 4266 set_all_memory( _gvn.transform(result_mem)); 4267 4268 C->set_has_split_ifs(true); // Has chance for split-if optimization 4269 set_result(result_reg, result_val); 4270 return true; 4271 } 4272 4273 //----------------------inline_native_getLength-------------------------- 4274 // public static native int java.lang.reflect.Array.getLength(Object array); 4275 bool LibraryCallKit::inline_native_getLength() { 4276 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 4277 4278 Node* array = null_check(argument(0)); 4279 // If array is dead, only null-path is taken. 4280 if (stopped()) return true; 4281 4282 // Deoptimize if it is a non-array. 4283 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 4284 4285 if (non_array != NULL) { 4286 PreserveJVMState pjvms(this); 4287 set_control(non_array); 4288 uncommon_trap(Deoptimization::Reason_intrinsic, 4289 Deoptimization::Action_maybe_recompile); 4290 } 4291 4292 // If control is dead, only non-array-path is taken. 4293 if (stopped()) return true; 4294 4295 // The works fine even if the array type is polymorphic. 4296 // It could be a dynamic mix of int[], boolean[], Object[], etc. 4297 Node* result = load_array_length(array); 4298 4299 C->set_has_split_ifs(true); // Has chance for split-if optimization 4300 set_result(result); 4301 return true; 4302 } 4303 4304 //------------------------inline_array_copyOf---------------------------- 4305 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 4306 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 4307 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 4308 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 4309 4310 // Get the arguments. 4311 Node* original = argument(0); 4312 Node* start = is_copyOfRange? argument(1): intcon(0); 4313 Node* end = is_copyOfRange? argument(2): argument(1); 4314 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 4315 4316 Node* newcopy = NULL; 4317 4318 // Set the original stack and the reexecute bit for the interpreter to reexecute 4319 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 4320 { PreserveReexecuteState preexecs(this); 4321 jvms()->set_should_reexecute(true); 4322 4323 array_type_mirror = null_check(array_type_mirror); 4324 original = null_check(original); 4325 4326 // Check if a null path was taken unconditionally. 4327 if (stopped()) return true; 4328 4329 Node* orig_length = load_array_length(original); 4330 4331 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 4332 klass_node = null_check(klass_node); 4333 4334 RegionNode* bailout = new RegionNode(1); 4335 record_for_igvn(bailout); 4336 4337 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 4338 // Bail out if that is so. 4339 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 4340 if (not_objArray != NULL) { 4341 // Improve the klass node's type from the new optimistic assumption: 4342 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 4343 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 4344 Node* cast = new CastPPNode(klass_node, akls); 4345 cast->init_req(0, control()); 4346 klass_node = _gvn.transform(cast); 4347 } 4348 4349 // Bail out if either start or end is negative. 4350 generate_negative_guard(start, bailout, &start); 4351 generate_negative_guard(end, bailout, &end); 4352 4353 Node* length = end; 4354 if (_gvn.type(start) != TypeInt::ZERO) { 4355 length = _gvn.transform(new SubINode(end, start)); 4356 } 4357 4358 // Bail out if length is negative. 4359 // Without this the new_array would throw 4360 // NegativeArraySizeException but IllegalArgumentException is what 4361 // should be thrown 4362 generate_negative_guard(length, bailout, &length); 4363 4364 if (bailout->req() > 1) { 4365 PreserveJVMState pjvms(this); 4366 set_control(_gvn.transform(bailout)); 4367 uncommon_trap(Deoptimization::Reason_intrinsic, 4368 Deoptimization::Action_maybe_recompile); 4369 } 4370 4371 if (!stopped()) { 4372 // How many elements will we copy from the original? 4373 // The answer is MinI(orig_length - start, length). 4374 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 4375 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 4376 4377 // Generate a direct call to the right arraycopy function(s). 4378 // We know the copy is disjoint but we might not know if the 4379 // oop stores need checking. 4380 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 4381 // This will fail a store-check if x contains any non-nulls. 4382 4383 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 4384 // loads/stores but it is legal only if we're sure the 4385 // Arrays.copyOf would succeed. So we need all input arguments 4386 // to the copyOf to be validated, including that the copy to the 4387 // new array won't trigger an ArrayStoreException. That subtype 4388 // check can be optimized if we know something on the type of 4389 // the input array from type speculation. 4390 if (_gvn.type(klass_node)->singleton()) { 4391 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 4392 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 4393 4394 int test = C->static_subtype_check(superk, subk); 4395 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 4396 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 4397 if (t_original->speculative_type() != NULL) { 4398 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 4399 } 4400 } 4401 } 4402 4403 bool validated = false; 4404 // Reason_class_check rather than Reason_intrinsic because we 4405 // want to intrinsify even if this traps. 4406 if (!too_many_traps(Deoptimization::Reason_class_check)) { 4407 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 4408 klass_node); 4409 4410 if (not_subtype_ctrl != top()) { 4411 PreserveJVMState pjvms(this); 4412 set_control(not_subtype_ctrl); 4413 uncommon_trap(Deoptimization::Reason_class_check, 4414 Deoptimization::Action_make_not_entrant); 4415 assert(stopped(), "Should be stopped"); 4416 } 4417 validated = true; 4418 } 4419 4420 if (!stopped()) { 4421 newcopy = new_array(klass_node, length, 0); // no arguments to push 4422 4423 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false, 4424 load_object_klass(original), klass_node); 4425 if (!is_copyOfRange) { 4426 ac->set_copyof(validated); 4427 } else { 4428 ac->set_copyofrange(validated); 4429 } 4430 Node* n = _gvn.transform(ac); 4431 if (n == ac) { 4432 ac->connect_outputs(this); 4433 } else { 4434 assert(validated, "shouldn't transform if all arguments not validated"); 4435 set_all_memory(n); 4436 } 4437 } 4438 } 4439 } // original reexecute is set back here 4440 4441 C->set_has_split_ifs(true); // Has chance for split-if optimization 4442 if (!stopped()) { 4443 set_result(newcopy); 4444 } 4445 return true; 4446 } 4447 4448 4449 //----------------------generate_virtual_guard--------------------------- 4450 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4451 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4452 RegionNode* slow_region) { 4453 ciMethod* method = callee(); 4454 int vtable_index = method->vtable_index(); 4455 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4456 "bad index %d", vtable_index); 4457 // Get the Method* out of the appropriate vtable entry. 4458 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 4459 vtable_index*vtableEntry::size_in_bytes() + 4460 vtableEntry::method_offset_in_bytes(); 4461 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4462 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4463 4464 // Compare the target method with the expected method (e.g., Object.hashCode). 4465 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4466 4467 Node* native_call = makecon(native_call_addr); 4468 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4469 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4470 4471 return generate_slow_guard(test_native, slow_region); 4472 } 4473 4474 //-----------------------generate_method_call---------------------------- 4475 // Use generate_method_call to make a slow-call to the real 4476 // method if the fast path fails. An alternative would be to 4477 // use a stub like OptoRuntime::slow_arraycopy_Java. 4478 // This only works for expanding the current library call, 4479 // not another intrinsic. (E.g., don't use this for making an 4480 // arraycopy call inside of the copyOf intrinsic.) 4481 CallJavaNode* 4482 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4483 // When compiling the intrinsic method itself, do not use this technique. 4484 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4485 4486 ciMethod* method = callee(); 4487 // ensure the JVMS we have will be correct for this call 4488 guarantee(method_id == method->intrinsic_id(), "must match"); 4489 4490 const TypeFunc* tf = TypeFunc::make(method); 4491 CallJavaNode* slow_call; 4492 if (is_static) { 4493 assert(!is_virtual, ""); 4494 slow_call = new CallStaticJavaNode(C, tf, 4495 SharedRuntime::get_resolve_static_call_stub(), 4496 method, bci()); 4497 } else if (is_virtual) { 4498 null_check_receiver(); 4499 int vtable_index = Method::invalid_vtable_index; 4500 if (UseInlineCaches) { 4501 // Suppress the vtable call 4502 } else { 4503 // hashCode and clone are not a miranda methods, 4504 // so the vtable index is fixed. 4505 // No need to use the linkResolver to get it. 4506 vtable_index = method->vtable_index(); 4507 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4508 "bad index %d", vtable_index); 4509 } 4510 slow_call = new CallDynamicJavaNode(tf, 4511 SharedRuntime::get_resolve_virtual_call_stub(), 4512 method, vtable_index, bci()); 4513 } else { // neither virtual nor static: opt_virtual 4514 null_check_receiver(); 4515 slow_call = new CallStaticJavaNode(C, tf, 4516 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4517 method, bci()); 4518 slow_call->set_optimized_virtual(true); 4519 } 4520 set_arguments_for_java_call(slow_call); 4521 set_edges_for_java_call(slow_call); 4522 return slow_call; 4523 } 4524 4525 4526 /** 4527 * Build special case code for calls to hashCode on an object. This call may 4528 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4529 * slightly different code. 4530 */ 4531 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4532 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4533 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4534 4535 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4536 4537 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4538 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4539 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4540 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4541 Node* obj = NULL; 4542 if (!is_static) { 4543 // Check for hashing null object 4544 obj = null_check_receiver(); 4545 if (stopped()) return true; // unconditionally null 4546 result_reg->init_req(_null_path, top()); 4547 result_val->init_req(_null_path, top()); 4548 } else { 4549 // Do a null check, and return zero if null. 4550 // System.identityHashCode(null) == 0 4551 obj = argument(0); 4552 Node* null_ctl = top(); 4553 obj = null_check_oop(obj, &null_ctl); 4554 result_reg->init_req(_null_path, null_ctl); 4555 result_val->init_req(_null_path, _gvn.intcon(0)); 4556 } 4557 4558 // Unconditionally null? Then return right away. 4559 if (stopped()) { 4560 set_control( result_reg->in(_null_path)); 4561 if (!stopped()) 4562 set_result(result_val->in(_null_path)); 4563 return true; 4564 } 4565 4566 // We only go to the fast case code if we pass a number of guards. The 4567 // paths which do not pass are accumulated in the slow_region. 4568 RegionNode* slow_region = new RegionNode(1); 4569 record_for_igvn(slow_region); 4570 4571 // If this is a virtual call, we generate a funny guard. We pull out 4572 // the vtable entry corresponding to hashCode() from the target object. 4573 // If the target method which we are calling happens to be the native 4574 // Object hashCode() method, we pass the guard. We do not need this 4575 // guard for non-virtual calls -- the caller is known to be the native 4576 // Object hashCode(). 4577 if (is_virtual) { 4578 // After null check, get the object's klass. 4579 Node* obj_klass = load_object_klass(obj); 4580 generate_virtual_guard(obj_klass, slow_region); 4581 } 4582 4583 // Get the header out of the object, use LoadMarkNode when available 4584 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4585 // The control of the load must be NULL. Otherwise, the load can move before 4586 // the null check after castPP removal. 4587 Node* no_ctrl = NULL; 4588 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4589 4590 // Test the header to see if it is unlocked. 4591 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4592 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4593 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4594 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4595 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4596 4597 generate_slow_guard(test_unlocked, slow_region); 4598 4599 // Get the hash value and check to see that it has been properly assigned. 4600 // We depend on hash_mask being at most 32 bits and avoid the use of 4601 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4602 // vm: see markOop.hpp. 4603 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4604 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4605 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4606 // This hack lets the hash bits live anywhere in the mark object now, as long 4607 // as the shift drops the relevant bits into the low 32 bits. Note that 4608 // Java spec says that HashCode is an int so there's no point in capturing 4609 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4610 hshifted_header = ConvX2I(hshifted_header); 4611 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4612 4613 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4614 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4615 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4616 4617 generate_slow_guard(test_assigned, slow_region); 4618 4619 Node* init_mem = reset_memory(); 4620 // fill in the rest of the null path: 4621 result_io ->init_req(_null_path, i_o()); 4622 result_mem->init_req(_null_path, init_mem); 4623 4624 result_val->init_req(_fast_path, hash_val); 4625 result_reg->init_req(_fast_path, control()); 4626 result_io ->init_req(_fast_path, i_o()); 4627 result_mem->init_req(_fast_path, init_mem); 4628 4629 // Generate code for the slow case. We make a call to hashCode(). 4630 set_control(_gvn.transform(slow_region)); 4631 if (!stopped()) { 4632 // No need for PreserveJVMState, because we're using up the present state. 4633 set_all_memory(init_mem); 4634 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4635 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4636 Node* slow_result = set_results_for_java_call(slow_call); 4637 // this->control() comes from set_results_for_java_call 4638 result_reg->init_req(_slow_path, control()); 4639 result_val->init_req(_slow_path, slow_result); 4640 result_io ->set_req(_slow_path, i_o()); 4641 result_mem ->set_req(_slow_path, reset_memory()); 4642 } 4643 4644 // Return the combined state. 4645 set_i_o( _gvn.transform(result_io) ); 4646 set_all_memory( _gvn.transform(result_mem)); 4647 4648 set_result(result_reg, result_val); 4649 return true; 4650 } 4651 4652 //---------------------------inline_native_getClass---------------------------- 4653 // public final native Class<?> java.lang.Object.getClass(); 4654 // 4655 // Build special case code for calls to getClass on an object. 4656 bool LibraryCallKit::inline_native_getClass() { 4657 Node* obj = null_check_receiver(); 4658 if (stopped()) return true; 4659 set_result(load_mirror_from_klass(load_object_klass(obj))); 4660 return true; 4661 } 4662 4663 //-----------------inline_native_Reflection_getCallerClass--------------------- 4664 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4665 // 4666 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4667 // 4668 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4669 // in that it must skip particular security frames and checks for 4670 // caller sensitive methods. 4671 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4672 #ifndef PRODUCT 4673 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4674 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4675 } 4676 #endif 4677 4678 if (!jvms()->has_method()) { 4679 #ifndef PRODUCT 4680 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4681 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4682 } 4683 #endif 4684 return false; 4685 } 4686 4687 // Walk back up the JVM state to find the caller at the required 4688 // depth. 4689 JVMState* caller_jvms = jvms(); 4690 4691 // Cf. JVM_GetCallerClass 4692 // NOTE: Start the loop at depth 1 because the current JVM state does 4693 // not include the Reflection.getCallerClass() frame. 4694 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4695 ciMethod* m = caller_jvms->method(); 4696 switch (n) { 4697 case 0: 4698 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4699 break; 4700 case 1: 4701 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4702 if (!m->caller_sensitive()) { 4703 #ifndef PRODUCT 4704 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4705 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4706 } 4707 #endif 4708 return false; // bail-out; let JVM_GetCallerClass do the work 4709 } 4710 break; 4711 default: 4712 if (!m->is_ignored_by_security_stack_walk()) { 4713 // We have reached the desired frame; return the holder class. 4714 // Acquire method holder as java.lang.Class and push as constant. 4715 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4716 ciInstance* caller_mirror = caller_klass->java_mirror(); 4717 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4718 4719 #ifndef PRODUCT 4720 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4721 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()); 4722 tty->print_cr(" JVM state at this point:"); 4723 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4724 ciMethod* m = jvms()->of_depth(i)->method(); 4725 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4726 } 4727 } 4728 #endif 4729 return true; 4730 } 4731 break; 4732 } 4733 } 4734 4735 #ifndef PRODUCT 4736 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4737 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4738 tty->print_cr(" JVM state at this point:"); 4739 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4740 ciMethod* m = jvms()->of_depth(i)->method(); 4741 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4742 } 4743 } 4744 #endif 4745 4746 return false; // bail-out; let JVM_GetCallerClass do the work 4747 } 4748 4749 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4750 Node* arg = argument(0); 4751 Node* result = NULL; 4752 4753 switch (id) { 4754 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4755 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4756 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4757 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4758 4759 case vmIntrinsics::_doubleToLongBits: { 4760 // two paths (plus control) merge in a wood 4761 RegionNode *r = new RegionNode(3); 4762 Node *phi = new PhiNode(r, TypeLong::LONG); 4763 4764 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4765 // Build the boolean node 4766 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4767 4768 // Branch either way. 4769 // NaN case is less traveled, which makes all the difference. 4770 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4771 Node *opt_isnan = _gvn.transform(ifisnan); 4772 assert( opt_isnan->is_If(), "Expect an IfNode"); 4773 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4774 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4775 4776 set_control(iftrue); 4777 4778 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4779 Node *slow_result = longcon(nan_bits); // return NaN 4780 phi->init_req(1, _gvn.transform( slow_result )); 4781 r->init_req(1, iftrue); 4782 4783 // Else fall through 4784 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4785 set_control(iffalse); 4786 4787 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4788 r->init_req(2, iffalse); 4789 4790 // Post merge 4791 set_control(_gvn.transform(r)); 4792 record_for_igvn(r); 4793 4794 C->set_has_split_ifs(true); // Has chance for split-if optimization 4795 result = phi; 4796 assert(result->bottom_type()->isa_long(), "must be"); 4797 break; 4798 } 4799 4800 case vmIntrinsics::_floatToIntBits: { 4801 // two paths (plus control) merge in a wood 4802 RegionNode *r = new RegionNode(3); 4803 Node *phi = new PhiNode(r, TypeInt::INT); 4804 4805 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4806 // Build the boolean node 4807 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4808 4809 // Branch either way. 4810 // NaN case is less traveled, which makes all the difference. 4811 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4812 Node *opt_isnan = _gvn.transform(ifisnan); 4813 assert( opt_isnan->is_If(), "Expect an IfNode"); 4814 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4815 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4816 4817 set_control(iftrue); 4818 4819 static const jint nan_bits = 0x7fc00000; 4820 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4821 phi->init_req(1, _gvn.transform( slow_result )); 4822 r->init_req(1, iftrue); 4823 4824 // Else fall through 4825 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4826 set_control(iffalse); 4827 4828 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4829 r->init_req(2, iffalse); 4830 4831 // Post merge 4832 set_control(_gvn.transform(r)); 4833 record_for_igvn(r); 4834 4835 C->set_has_split_ifs(true); // Has chance for split-if optimization 4836 result = phi; 4837 assert(result->bottom_type()->isa_int(), "must be"); 4838 break; 4839 } 4840 4841 default: 4842 fatal_unexpected_iid(id); 4843 break; 4844 } 4845 set_result(_gvn.transform(result)); 4846 return true; 4847 } 4848 4849 //----------------------inline_unsafe_copyMemory------------------------- 4850 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4851 bool LibraryCallKit::inline_unsafe_copyMemory() { 4852 if (callee()->is_static()) return false; // caller must have the capability! 4853 null_check_receiver(); // null-check receiver 4854 if (stopped()) return true; 4855 4856 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4857 4858 Node* src_ptr = argument(1); // type: oop 4859 Node* src_off = ConvL2X(argument(2)); // type: long 4860 Node* dst_ptr = argument(4); // type: oop 4861 Node* dst_off = ConvL2X(argument(5)); // type: long 4862 Node* size = ConvL2X(argument(7)); // type: long 4863 4864 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4865 "fieldOffset must be byte-scaled"); 4866 4867 Node* src = make_unsafe_address(src_ptr, src_off); 4868 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4869 4870 // Conservatively insert a memory barrier on all memory slices. 4871 // Do not let writes of the copy source or destination float below the copy. 4872 insert_mem_bar(Op_MemBarCPUOrder); 4873 4874 // Call it. Note that the length argument is not scaled. 4875 make_runtime_call(RC_LEAF|RC_NO_FP, 4876 OptoRuntime::fast_arraycopy_Type(), 4877 StubRoutines::unsafe_arraycopy(), 4878 "unsafe_arraycopy", 4879 TypeRawPtr::BOTTOM, 4880 src, dst, size XTOP); 4881 4882 // Do not let reads of the copy destination float above the copy. 4883 insert_mem_bar(Op_MemBarCPUOrder); 4884 4885 return true; 4886 } 4887 4888 //------------------------clone_coping----------------------------------- 4889 // Helper function for inline_native_clone. 4890 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4891 assert(obj_size != NULL, ""); 4892 Node* raw_obj = alloc_obj->in(1); 4893 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4894 4895 AllocateNode* alloc = NULL; 4896 if (ReduceBulkZeroing) { 4897 // We will be completely responsible for initializing this object - 4898 // mark Initialize node as complete. 4899 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4900 // The object was just allocated - there should be no any stores! 4901 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4902 // Mark as complete_with_arraycopy so that on AllocateNode 4903 // expansion, we know this AllocateNode is initialized by an array 4904 // copy and a StoreStore barrier exists after the array copy. 4905 alloc->initialization()->set_complete_with_arraycopy(); 4906 } 4907 4908 // Copy the fastest available way. 4909 // TODO: generate fields copies for small objects instead. 4910 Node* src = obj; 4911 Node* dest = alloc_obj; 4912 Node* size = _gvn.transform(obj_size); 4913 4914 // Exclude the header but include array length to copy by 8 bytes words. 4915 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4916 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4917 instanceOopDesc::base_offset_in_bytes(); 4918 // base_off: 4919 // 8 - 32-bit VM 4920 // 12 - 64-bit VM, compressed klass 4921 // 16 - 64-bit VM, normal klass 4922 if (base_off % BytesPerLong != 0) { 4923 assert(UseCompressedClassPointers, ""); 4924 if (is_array) { 4925 // Exclude length to copy by 8 bytes words. 4926 base_off += sizeof(int); 4927 } else { 4928 // Include klass to copy by 8 bytes words. 4929 base_off = instanceOopDesc::klass_offset_in_bytes(); 4930 } 4931 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4932 } 4933 src = basic_plus_adr(src, base_off); 4934 dest = basic_plus_adr(dest, base_off); 4935 4936 // Compute the length also, if needed: 4937 Node* countx = size; 4938 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4939 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4940 4941 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4942 4943 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false); 4944 ac->set_clonebasic(); 4945 Node* n = _gvn.transform(ac); 4946 if (n == ac) { 4947 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4948 } else { 4949 set_all_memory(n); 4950 } 4951 4952 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4953 if (card_mark) { 4954 assert(!is_array, ""); 4955 // Put in store barrier for any and all oops we are sticking 4956 // into this object. (We could avoid this if we could prove 4957 // that the object type contains no oop fields at all.) 4958 Node* no_particular_value = NULL; 4959 Node* no_particular_field = NULL; 4960 int raw_adr_idx = Compile::AliasIdxRaw; 4961 post_barrier(control(), 4962 memory(raw_adr_type), 4963 alloc_obj, 4964 no_particular_field, 4965 raw_adr_idx, 4966 no_particular_value, 4967 T_OBJECT, 4968 false); 4969 } 4970 4971 // Do not let reads from the cloned object float above the arraycopy. 4972 if (alloc != NULL) { 4973 // Do not let stores that initialize this object be reordered with 4974 // a subsequent store that would make this object accessible by 4975 // other threads. 4976 // Record what AllocateNode this StoreStore protects so that 4977 // escape analysis can go from the MemBarStoreStoreNode to the 4978 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4979 // based on the escape status of the AllocateNode. 4980 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 4981 } else { 4982 insert_mem_bar(Op_MemBarCPUOrder); 4983 } 4984 } 4985 4986 //------------------------inline_native_clone---------------------------- 4987 // protected native Object java.lang.Object.clone(); 4988 // 4989 // Here are the simple edge cases: 4990 // null receiver => normal trap 4991 // virtual and clone was overridden => slow path to out-of-line clone 4992 // not cloneable or finalizer => slow path to out-of-line Object.clone 4993 // 4994 // The general case has two steps, allocation and copying. 4995 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4996 // 4997 // Copying also has two cases, oop arrays and everything else. 4998 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4999 // Everything else uses the tight inline loop supplied by CopyArrayNode. 5000 // 5001 // These steps fold up nicely if and when the cloned object's klass 5002 // can be sharply typed as an object array, a type array, or an instance. 5003 // 5004 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 5005 PhiNode* result_val; 5006 5007 // Set the reexecute bit for the interpreter to reexecute 5008 // the bytecode that invokes Object.clone if deoptimization happens. 5009 { PreserveReexecuteState preexecs(this); 5010 jvms()->set_should_reexecute(true); 5011 5012 Node* obj = null_check_receiver(); 5013 if (stopped()) return true; 5014 5015 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 5016 5017 // If we are going to clone an instance, we need its exact type to 5018 // know the number and types of fields to convert the clone to 5019 // loads/stores. Maybe a speculative type can help us. 5020 if (!obj_type->klass_is_exact() && 5021 obj_type->speculative_type() != NULL && 5022 obj_type->speculative_type()->is_instance_klass()) { 5023 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 5024 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 5025 !spec_ik->has_injected_fields()) { 5026 ciKlass* k = obj_type->klass(); 5027 if (!k->is_instance_klass() || 5028 k->as_instance_klass()->is_interface() || 5029 k->as_instance_klass()->has_subklass()) { 5030 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 5031 } 5032 } 5033 } 5034 5035 Node* obj_klass = load_object_klass(obj); 5036 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 5037 const TypeOopPtr* toop = ((tklass != NULL) 5038 ? tklass->as_instance_type() 5039 : TypeInstPtr::NOTNULL); 5040 5041 // Conservatively insert a memory barrier on all memory slices. 5042 // Do not let writes into the original float below the clone. 5043 insert_mem_bar(Op_MemBarCPUOrder); 5044 5045 // paths into result_reg: 5046 enum { 5047 _slow_path = 1, // out-of-line call to clone method (virtual or not) 5048 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 5049 _array_path, // plain array allocation, plus arrayof_long_arraycopy 5050 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 5051 PATH_LIMIT 5052 }; 5053 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 5054 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 5055 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 5056 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 5057 record_for_igvn(result_reg); 5058 5059 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 5060 int raw_adr_idx = Compile::AliasIdxRaw; 5061 5062 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 5063 if (array_ctl != NULL) { 5064 // It's an array. 5065 PreserveJVMState pjvms(this); 5066 set_control(array_ctl); 5067 Node* obj_length = load_array_length(obj); 5068 Node* obj_size = NULL; 5069 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 5070 5071 if (!use_ReduceInitialCardMarks()) { 5072 // If it is an oop array, it requires very special treatment, 5073 // because card marking is required on each card of the array. 5074 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 5075 if (is_obja != NULL) { 5076 PreserveJVMState pjvms2(this); 5077 set_control(is_obja); 5078 // Generate a direct call to the right arraycopy function(s). 5079 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 5080 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false); 5081 ac->set_cloneoop(); 5082 Node* n = _gvn.transform(ac); 5083 assert(n == ac, "cannot disappear"); 5084 ac->connect_outputs(this); 5085 5086 result_reg->init_req(_objArray_path, control()); 5087 result_val->init_req(_objArray_path, alloc_obj); 5088 result_i_o ->set_req(_objArray_path, i_o()); 5089 result_mem ->set_req(_objArray_path, reset_memory()); 5090 } 5091 } 5092 // Otherwise, there are no card marks to worry about. 5093 // (We can dispense with card marks if we know the allocation 5094 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 5095 // causes the non-eden paths to take compensating steps to 5096 // simulate a fresh allocation, so that no further 5097 // card marks are required in compiled code to initialize 5098 // the object.) 5099 5100 if (!stopped()) { 5101 copy_to_clone(obj, alloc_obj, obj_size, true, false); 5102 5103 // Present the results of the copy. 5104 result_reg->init_req(_array_path, control()); 5105 result_val->init_req(_array_path, alloc_obj); 5106 result_i_o ->set_req(_array_path, i_o()); 5107 result_mem ->set_req(_array_path, reset_memory()); 5108 } 5109 } 5110 5111 // We only go to the instance fast case code if we pass a number of guards. 5112 // The paths which do not pass are accumulated in the slow_region. 5113 RegionNode* slow_region = new RegionNode(1); 5114 record_for_igvn(slow_region); 5115 if (!stopped()) { 5116 // It's an instance (we did array above). Make the slow-path tests. 5117 // If this is a virtual call, we generate a funny guard. We grab 5118 // the vtable entry corresponding to clone() from the target object. 5119 // If the target method which we are calling happens to be the 5120 // Object clone() method, we pass the guard. We do not need this 5121 // guard for non-virtual calls; the caller is known to be the native 5122 // Object clone(). 5123 if (is_virtual) { 5124 generate_virtual_guard(obj_klass, slow_region); 5125 } 5126 5127 // The object must be easily cloneable and must not have a finalizer. 5128 // Both of these conditions may be checked in a single test. 5129 // We could optimize the test further, but we don't care. 5130 generate_access_flags_guard(obj_klass, 5131 // Test both conditions: 5132 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 5133 // Must be cloneable but not finalizer: 5134 JVM_ACC_IS_CLONEABLE_FAST, 5135 slow_region); 5136 } 5137 5138 if (!stopped()) { 5139 // It's an instance, and it passed the slow-path tests. 5140 PreserveJVMState pjvms(this); 5141 Node* obj_size = NULL; 5142 // Need to deoptimize on exception from allocation since Object.clone intrinsic 5143 // is reexecuted if deoptimization occurs and there could be problems when merging 5144 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 5145 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 5146 5147 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 5148 5149 // Present the results of the slow call. 5150 result_reg->init_req(_instance_path, control()); 5151 result_val->init_req(_instance_path, alloc_obj); 5152 result_i_o ->set_req(_instance_path, i_o()); 5153 result_mem ->set_req(_instance_path, reset_memory()); 5154 } 5155 5156 // Generate code for the slow case. We make a call to clone(). 5157 set_control(_gvn.transform(slow_region)); 5158 if (!stopped()) { 5159 PreserveJVMState pjvms(this); 5160 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 5161 Node* slow_result = set_results_for_java_call(slow_call); 5162 // this->control() comes from set_results_for_java_call 5163 result_reg->init_req(_slow_path, control()); 5164 result_val->init_req(_slow_path, slow_result); 5165 result_i_o ->set_req(_slow_path, i_o()); 5166 result_mem ->set_req(_slow_path, reset_memory()); 5167 } 5168 5169 // Return the combined state. 5170 set_control( _gvn.transform(result_reg)); 5171 set_i_o( _gvn.transform(result_i_o)); 5172 set_all_memory( _gvn.transform(result_mem)); 5173 } // original reexecute is set back here 5174 5175 set_result(_gvn.transform(result_val)); 5176 return true; 5177 } 5178 5179 // If we have a tighly coupled allocation, the arraycopy may take care 5180 // of the array initialization. If one of the guards we insert between 5181 // the allocation and the arraycopy causes a deoptimization, an 5182 // unitialized array will escape the compiled method. To prevent that 5183 // we set the JVM state for uncommon traps between the allocation and 5184 // the arraycopy to the state before the allocation so, in case of 5185 // deoptimization, we'll reexecute the allocation and the 5186 // initialization. 5187 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 5188 if (alloc != NULL) { 5189 ciMethod* trap_method = alloc->jvms()->method(); 5190 int trap_bci = alloc->jvms()->bci(); 5191 5192 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 5193 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 5194 // Make sure there's no store between the allocation and the 5195 // arraycopy otherwise visible side effects could be rexecuted 5196 // in case of deoptimization and cause incorrect execution. 5197 bool no_interfering_store = true; 5198 Node* mem = alloc->in(TypeFunc::Memory); 5199 if (mem->is_MergeMem()) { 5200 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 5201 Node* n = mms.memory(); 5202 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 5203 assert(n->is_Store(), "what else?"); 5204 no_interfering_store = false; 5205 break; 5206 } 5207 } 5208 } else { 5209 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 5210 Node* n = mms.memory(); 5211 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 5212 assert(n->is_Store(), "what else?"); 5213 no_interfering_store = false; 5214 break; 5215 } 5216 } 5217 } 5218 5219 if (no_interfering_store) { 5220 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 5221 uint size = alloc->req(); 5222 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 5223 old_jvms->set_map(sfpt); 5224 for (uint i = 0; i < size; i++) { 5225 sfpt->init_req(i, alloc->in(i)); 5226 } 5227 // re-push array length for deoptimization 5228 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 5229 old_jvms->set_sp(old_jvms->sp()+1); 5230 old_jvms->set_monoff(old_jvms->monoff()+1); 5231 old_jvms->set_scloff(old_jvms->scloff()+1); 5232 old_jvms->set_endoff(old_jvms->endoff()+1); 5233 old_jvms->set_should_reexecute(true); 5234 5235 sfpt->set_i_o(map()->i_o()); 5236 sfpt->set_memory(map()->memory()); 5237 sfpt->set_control(map()->control()); 5238 5239 JVMState* saved_jvms = jvms(); 5240 saved_reexecute_sp = _reexecute_sp; 5241 5242 set_jvms(sfpt->jvms()); 5243 _reexecute_sp = jvms()->sp(); 5244 5245 return saved_jvms; 5246 } 5247 } 5248 } 5249 return NULL; 5250 } 5251 5252 // In case of a deoptimization, we restart execution at the 5253 // allocation, allocating a new array. We would leave an uninitialized 5254 // array in the heap that GCs wouldn't expect. Move the allocation 5255 // after the traps so we don't allocate the array if we 5256 // deoptimize. This is possible because tightly_coupled_allocation() 5257 // guarantees there's no observer of the allocated array at this point 5258 // and the control flow is simple enough. 5259 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, 5260 int saved_reexecute_sp, uint new_idx) { 5261 if (saved_jvms != NULL && !stopped()) { 5262 assert(alloc != NULL, "only with a tightly coupled allocation"); 5263 // restore JVM state to the state at the arraycopy 5264 saved_jvms->map()->set_control(map()->control()); 5265 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 5266 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 5267 // If we've improved the types of some nodes (null check) while 5268 // emitting the guards, propagate them to the current state 5269 map()->replaced_nodes().apply(saved_jvms->map(), new_idx); 5270 set_jvms(saved_jvms); 5271 _reexecute_sp = saved_reexecute_sp; 5272 5273 // Remove the allocation from above the guards 5274 CallProjections callprojs; 5275 alloc->extract_projections(&callprojs, true); 5276 InitializeNode* init = alloc->initialization(); 5277 Node* alloc_mem = alloc->in(TypeFunc::Memory); 5278 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 5279 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 5280 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 5281 5282 // move the allocation here (after the guards) 5283 _gvn.hash_delete(alloc); 5284 alloc->set_req(TypeFunc::Control, control()); 5285 alloc->set_req(TypeFunc::I_O, i_o()); 5286 Node *mem = reset_memory(); 5287 set_all_memory(mem); 5288 alloc->set_req(TypeFunc::Memory, mem); 5289 set_control(init->proj_out_or_null(TypeFunc::Control)); 5290 set_i_o(callprojs.fallthrough_ioproj); 5291 5292 // Update memory as done in GraphKit::set_output_for_allocation() 5293 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 5294 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 5295 if (ary_type->isa_aryptr() && length_type != NULL) { 5296 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 5297 } 5298 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 5299 int elemidx = C->get_alias_index(telemref); 5300 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw); 5301 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx); 5302 5303 Node* allocx = _gvn.transform(alloc); 5304 assert(allocx == alloc, "where has the allocation gone?"); 5305 assert(dest->is_CheckCastPP(), "not an allocation result?"); 5306 5307 _gvn.hash_delete(dest); 5308 dest->set_req(0, control()); 5309 Node* destx = _gvn.transform(dest); 5310 assert(destx == dest, "where has the allocation result gone?"); 5311 } 5312 } 5313 5314 5315 //------------------------------inline_arraycopy----------------------- 5316 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 5317 // Object dest, int destPos, 5318 // int length); 5319 bool LibraryCallKit::inline_arraycopy() { 5320 // Get the arguments. 5321 Node* src = argument(0); // type: oop 5322 Node* src_offset = argument(1); // type: int 5323 Node* dest = argument(2); // type: oop 5324 Node* dest_offset = argument(3); // type: int 5325 Node* length = argument(4); // type: int 5326 5327 uint new_idx = C->unique(); 5328 5329 // Check for allocation before we add nodes that would confuse 5330 // tightly_coupled_allocation() 5331 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 5332 5333 int saved_reexecute_sp = -1; 5334 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 5335 // See arraycopy_restore_alloc_state() comment 5336 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 5337 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 5338 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards 5339 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 5340 5341 // The following tests must be performed 5342 // (1) src and dest are arrays. 5343 // (2) src and dest arrays must have elements of the same BasicType 5344 // (3) src and dest must not be null. 5345 // (4) src_offset must not be negative. 5346 // (5) dest_offset must not be negative. 5347 // (6) length must not be negative. 5348 // (7) src_offset + length must not exceed length of src. 5349 // (8) dest_offset + length must not exceed length of dest. 5350 // (9) each element of an oop array must be assignable 5351 5352 // (3) src and dest must not be null. 5353 // always do this here because we need the JVM state for uncommon traps 5354 Node* null_ctl = top(); 5355 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 5356 assert(null_ctl->is_top(), "no null control here"); 5357 dest = null_check(dest, T_ARRAY); 5358 5359 if (!can_emit_guards) { 5360 // if saved_jvms == NULL and alloc != NULL, we don't emit any 5361 // guards but the arraycopy node could still take advantage of a 5362 // tightly allocated allocation. tightly_coupled_allocation() is 5363 // called again to make sure it takes the null check above into 5364 // account: the null check is mandatory and if it caused an 5365 // uncommon trap to be emitted then the allocation can't be 5366 // considered tightly coupled in this context. 5367 alloc = tightly_coupled_allocation(dest, NULL); 5368 } 5369 5370 bool validated = false; 5371 5372 const Type* src_type = _gvn.type(src); 5373 const Type* dest_type = _gvn.type(dest); 5374 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5375 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5376 5377 // Do we have the type of src? 5378 bool has_src = (top_src != NULL && top_src->klass() != NULL); 5379 // Do we have the type of dest? 5380 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 5381 // Is the type for src from speculation? 5382 bool src_spec = false; 5383 // Is the type for dest from speculation? 5384 bool dest_spec = false; 5385 5386 if ((!has_src || !has_dest) && can_emit_guards) { 5387 // We don't have sufficient type information, let's see if 5388 // speculative types can help. We need to have types for both src 5389 // and dest so that it pays off. 5390 5391 // Do we already have or could we have type information for src 5392 bool could_have_src = has_src; 5393 // Do we already have or could we have type information for dest 5394 bool could_have_dest = has_dest; 5395 5396 ciKlass* src_k = NULL; 5397 if (!has_src) { 5398 src_k = src_type->speculative_type_not_null(); 5399 if (src_k != NULL && src_k->is_array_klass()) { 5400 could_have_src = true; 5401 } 5402 } 5403 5404 ciKlass* dest_k = NULL; 5405 if (!has_dest) { 5406 dest_k = dest_type->speculative_type_not_null(); 5407 if (dest_k != NULL && dest_k->is_array_klass()) { 5408 could_have_dest = true; 5409 } 5410 } 5411 5412 if (could_have_src && could_have_dest) { 5413 // This is going to pay off so emit the required guards 5414 if (!has_src) { 5415 src = maybe_cast_profiled_obj(src, src_k, true); 5416 src_type = _gvn.type(src); 5417 top_src = src_type->isa_aryptr(); 5418 has_src = (top_src != NULL && top_src->klass() != NULL); 5419 src_spec = true; 5420 } 5421 if (!has_dest) { 5422 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5423 dest_type = _gvn.type(dest); 5424 top_dest = dest_type->isa_aryptr(); 5425 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 5426 dest_spec = true; 5427 } 5428 } 5429 } 5430 5431 if (has_src && has_dest && can_emit_guards) { 5432 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 5433 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 5434 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 5435 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 5436 5437 if (src_elem == dest_elem && src_elem == T_OBJECT) { 5438 // If both arrays are object arrays then having the exact types 5439 // for both will remove the need for a subtype check at runtime 5440 // before the call and may make it possible to pick a faster copy 5441 // routine (without a subtype check on every element) 5442 // Do we have the exact type of src? 5443 bool could_have_src = src_spec; 5444 // Do we have the exact type of dest? 5445 bool could_have_dest = dest_spec; 5446 ciKlass* src_k = top_src->klass(); 5447 ciKlass* dest_k = top_dest->klass(); 5448 if (!src_spec) { 5449 src_k = src_type->speculative_type_not_null(); 5450 if (src_k != NULL && src_k->is_array_klass()) { 5451 could_have_src = true; 5452 } 5453 } 5454 if (!dest_spec) { 5455 dest_k = dest_type->speculative_type_not_null(); 5456 if (dest_k != NULL && dest_k->is_array_klass()) { 5457 could_have_dest = true; 5458 } 5459 } 5460 if (could_have_src && could_have_dest) { 5461 // If we can have both exact types, emit the missing guards 5462 if (could_have_src && !src_spec) { 5463 src = maybe_cast_profiled_obj(src, src_k, true); 5464 } 5465 if (could_have_dest && !dest_spec) { 5466 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5467 } 5468 } 5469 } 5470 } 5471 5472 ciMethod* trap_method = method(); 5473 int trap_bci = bci(); 5474 if (saved_jvms != NULL) { 5475 trap_method = alloc->jvms()->method(); 5476 trap_bci = alloc->jvms()->bci(); 5477 } 5478 5479 bool negative_length_guard_generated = false; 5480 5481 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5482 can_emit_guards && 5483 !src->is_top() && !dest->is_top()) { 5484 // validate arguments: enables transformation the ArrayCopyNode 5485 validated = true; 5486 5487 RegionNode* slow_region = new RegionNode(1); 5488 record_for_igvn(slow_region); 5489 5490 // (1) src and dest are arrays. 5491 generate_non_array_guard(load_object_klass(src), slow_region); 5492 generate_non_array_guard(load_object_klass(dest), slow_region); 5493 5494 // (2) src and dest arrays must have elements of the same BasicType 5495 // done at macro expansion or at Ideal transformation time 5496 5497 // (4) src_offset must not be negative. 5498 generate_negative_guard(src_offset, slow_region); 5499 5500 // (5) dest_offset must not be negative. 5501 generate_negative_guard(dest_offset, slow_region); 5502 5503 // (7) src_offset + length must not exceed length of src. 5504 generate_limit_guard(src_offset, length, 5505 load_array_length(src), 5506 slow_region); 5507 5508 // (8) dest_offset + length must not exceed length of dest. 5509 generate_limit_guard(dest_offset, length, 5510 load_array_length(dest), 5511 slow_region); 5512 5513 // (6) length must not be negative. 5514 // This is also checked in generate_arraycopy() during macro expansion, but 5515 // we also have to check it here for the case where the ArrayCopyNode will 5516 // be eliminated by Escape Analysis. 5517 if (EliminateAllocations) { 5518 generate_negative_guard(length, slow_region); 5519 negative_length_guard_generated = true; 5520 } 5521 5522 // (9) each element of an oop array must be assignable 5523 Node* src_klass = load_object_klass(src); 5524 Node* dest_klass = load_object_klass(dest); 5525 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5526 5527 if (not_subtype_ctrl != top()) { 5528 PreserveJVMState pjvms(this); 5529 set_control(not_subtype_ctrl); 5530 uncommon_trap(Deoptimization::Reason_intrinsic, 5531 Deoptimization::Action_make_not_entrant); 5532 assert(stopped(), "Should be stopped"); 5533 } 5534 { 5535 PreserveJVMState pjvms(this); 5536 set_control(_gvn.transform(slow_region)); 5537 uncommon_trap(Deoptimization::Reason_intrinsic, 5538 Deoptimization::Action_make_not_entrant); 5539 assert(stopped(), "Should be stopped"); 5540 } 5541 5542 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr(); 5543 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass()); 5544 src = _gvn.transform(new CheckCastPPNode(control(), src, toop)); 5545 } 5546 5547 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx); 5548 5549 if (stopped()) { 5550 return true; 5551 } 5552 5553 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated, 5554 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5555 // so the compiler has a chance to eliminate them: during macro expansion, 5556 // we have to set their control (CastPP nodes are eliminated). 5557 load_object_klass(src), load_object_klass(dest), 5558 load_array_length(src), load_array_length(dest)); 5559 5560 ac->set_arraycopy(validated); 5561 5562 Node* n = _gvn.transform(ac); 5563 if (n == ac) { 5564 ac->connect_outputs(this); 5565 } else { 5566 assert(validated, "shouldn't transform if all arguments not validated"); 5567 set_all_memory(n); 5568 } 5569 clear_upper_avx(); 5570 5571 5572 return true; 5573 } 5574 5575 5576 // Helper function which determines if an arraycopy immediately follows 5577 // an allocation, with no intervening tests or other escapes for the object. 5578 AllocateArrayNode* 5579 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5580 RegionNode* slow_region) { 5581 if (stopped()) return NULL; // no fast path 5582 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5583 5584 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5585 if (alloc == NULL) return NULL; 5586 5587 Node* rawmem = memory(Compile::AliasIdxRaw); 5588 // Is the allocation's memory state untouched? 5589 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5590 // Bail out if there have been raw-memory effects since the allocation. 5591 // (Example: There might have been a call or safepoint.) 5592 return NULL; 5593 } 5594 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5595 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5596 return NULL; 5597 } 5598 5599 // There must be no unexpected observers of this allocation. 5600 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5601 Node* obs = ptr->fast_out(i); 5602 if (obs != this->map()) { 5603 return NULL; 5604 } 5605 } 5606 5607 // This arraycopy must unconditionally follow the allocation of the ptr. 5608 Node* alloc_ctl = ptr->in(0); 5609 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5610 5611 Node* ctl = control(); 5612 while (ctl != alloc_ctl) { 5613 // There may be guards which feed into the slow_region. 5614 // Any other control flow means that we might not get a chance 5615 // to finish initializing the allocated object. 5616 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5617 IfNode* iff = ctl->in(0)->as_If(); 5618 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con); 5619 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5620 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5621 ctl = iff->in(0); // This test feeds the known slow_region. 5622 continue; 5623 } 5624 // One more try: Various low-level checks bottom out in 5625 // uncommon traps. If the debug-info of the trap omits 5626 // any reference to the allocation, as we've already 5627 // observed, then there can be no objection to the trap. 5628 bool found_trap = false; 5629 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5630 Node* obs = not_ctl->fast_out(j); 5631 if (obs->in(0) == not_ctl && obs->is_Call() && 5632 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5633 found_trap = true; break; 5634 } 5635 } 5636 if (found_trap) { 5637 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5638 continue; 5639 } 5640 } 5641 return NULL; 5642 } 5643 5644 // If we get this far, we have an allocation which immediately 5645 // precedes the arraycopy, and we can take over zeroing the new object. 5646 // The arraycopy will finish the initialization, and provide 5647 // a new control state to which we will anchor the destination pointer. 5648 5649 return alloc; 5650 } 5651 5652 //-------------inline_encodeISOArray----------------------------------- 5653 // encode char[] to byte[] in ISO_8859_1 5654 bool LibraryCallKit::inline_encodeISOArray() { 5655 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5656 // no receiver since it is static method 5657 Node *src = argument(0); 5658 Node *src_offset = argument(1); 5659 Node *dst = argument(2); 5660 Node *dst_offset = argument(3); 5661 Node *length = argument(4); 5662 5663 const Type* src_type = src->Value(&_gvn); 5664 const Type* dst_type = dst->Value(&_gvn); 5665 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5666 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5667 if (top_src == NULL || top_src->klass() == NULL || 5668 top_dest == NULL || top_dest->klass() == NULL) { 5669 // failed array check 5670 return false; 5671 } 5672 5673 // Figure out the size and type of the elements we will be copying. 5674 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5675 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5676 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5677 return false; 5678 } 5679 5680 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5681 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5682 // 'src_start' points to src array + scaled offset 5683 // 'dst_start' points to dst array + scaled offset 5684 5685 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5686 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5687 enc = _gvn.transform(enc); 5688 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5689 set_memory(res_mem, mtype); 5690 set_result(enc); 5691 clear_upper_avx(); 5692 5693 return true; 5694 } 5695 5696 //-------------inline_multiplyToLen----------------------------------- 5697 bool LibraryCallKit::inline_multiplyToLen() { 5698 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5699 5700 address stubAddr = StubRoutines::multiplyToLen(); 5701 if (stubAddr == NULL) { 5702 return false; // Intrinsic's stub is not implemented on this platform 5703 } 5704 const char* stubName = "multiplyToLen"; 5705 5706 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5707 5708 // no receiver because it is a static method 5709 Node* x = argument(0); 5710 Node* xlen = argument(1); 5711 Node* y = argument(2); 5712 Node* ylen = argument(3); 5713 Node* z = argument(4); 5714 5715 const Type* x_type = x->Value(&_gvn); 5716 const Type* y_type = y->Value(&_gvn); 5717 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5718 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5719 if (top_x == NULL || top_x->klass() == NULL || 5720 top_y == NULL || top_y->klass() == NULL) { 5721 // failed array check 5722 return false; 5723 } 5724 5725 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5726 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5727 if (x_elem != T_INT || y_elem != T_INT) { 5728 return false; 5729 } 5730 5731 // Set the original stack and the reexecute bit for the interpreter to reexecute 5732 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5733 // on the return from z array allocation in runtime. 5734 { PreserveReexecuteState preexecs(this); 5735 jvms()->set_should_reexecute(true); 5736 5737 Node* x_start = array_element_address(x, intcon(0), x_elem); 5738 Node* y_start = array_element_address(y, intcon(0), y_elem); 5739 // 'x_start' points to x array + scaled xlen 5740 // 'y_start' points to y array + scaled ylen 5741 5742 // Allocate the result array 5743 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5744 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5745 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5746 5747 IdealKit ideal(this); 5748 5749 #define __ ideal. 5750 Node* one = __ ConI(1); 5751 Node* zero = __ ConI(0); 5752 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5753 __ set(need_alloc, zero); 5754 __ set(z_alloc, z); 5755 __ if_then(z, BoolTest::eq, null()); { 5756 __ increment (need_alloc, one); 5757 } __ else_(); { 5758 // Update graphKit memory and control from IdealKit. 5759 sync_kit(ideal); 5760 Node* zlen_arg = load_array_length(z); 5761 // Update IdealKit memory and control from graphKit. 5762 __ sync_kit(this); 5763 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5764 __ increment (need_alloc, one); 5765 } __ end_if(); 5766 } __ end_if(); 5767 5768 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5769 // Update graphKit memory and control from IdealKit. 5770 sync_kit(ideal); 5771 Node * narr = new_array(klass_node, zlen, 1); 5772 // Update IdealKit memory and control from graphKit. 5773 __ sync_kit(this); 5774 __ set(z_alloc, narr); 5775 } __ end_if(); 5776 5777 sync_kit(ideal); 5778 z = __ value(z_alloc); 5779 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5780 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5781 // Final sync IdealKit and GraphKit. 5782 final_sync(ideal); 5783 #undef __ 5784 5785 Node* z_start = array_element_address(z, intcon(0), T_INT); 5786 5787 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5788 OptoRuntime::multiplyToLen_Type(), 5789 stubAddr, stubName, TypePtr::BOTTOM, 5790 x_start, xlen, y_start, ylen, z_start, zlen); 5791 } // original reexecute is set back here 5792 5793 C->set_has_split_ifs(true); // Has chance for split-if optimization 5794 set_result(z); 5795 return true; 5796 } 5797 5798 //-------------inline_squareToLen------------------------------------ 5799 bool LibraryCallKit::inline_squareToLen() { 5800 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5801 5802 address stubAddr = StubRoutines::squareToLen(); 5803 if (stubAddr == NULL) { 5804 return false; // Intrinsic's stub is not implemented on this platform 5805 } 5806 const char* stubName = "squareToLen"; 5807 5808 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5809 5810 Node* x = argument(0); 5811 Node* len = argument(1); 5812 Node* z = argument(2); 5813 Node* zlen = argument(3); 5814 5815 const Type* x_type = x->Value(&_gvn); 5816 const Type* z_type = z->Value(&_gvn); 5817 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5818 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5819 if (top_x == NULL || top_x->klass() == NULL || 5820 top_z == NULL || top_z->klass() == NULL) { 5821 // failed array check 5822 return false; 5823 } 5824 5825 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5826 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5827 if (x_elem != T_INT || z_elem != T_INT) { 5828 return false; 5829 } 5830 5831 5832 Node* x_start = array_element_address(x, intcon(0), x_elem); 5833 Node* z_start = array_element_address(z, intcon(0), z_elem); 5834 5835 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5836 OptoRuntime::squareToLen_Type(), 5837 stubAddr, stubName, TypePtr::BOTTOM, 5838 x_start, len, z_start, zlen); 5839 5840 set_result(z); 5841 return true; 5842 } 5843 5844 //-------------inline_mulAdd------------------------------------------ 5845 bool LibraryCallKit::inline_mulAdd() { 5846 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5847 5848 address stubAddr = StubRoutines::mulAdd(); 5849 if (stubAddr == NULL) { 5850 return false; // Intrinsic's stub is not implemented on this platform 5851 } 5852 const char* stubName = "mulAdd"; 5853 5854 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5855 5856 Node* out = argument(0); 5857 Node* in = argument(1); 5858 Node* offset = argument(2); 5859 Node* len = argument(3); 5860 Node* k = argument(4); 5861 5862 const Type* out_type = out->Value(&_gvn); 5863 const Type* in_type = in->Value(&_gvn); 5864 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5865 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5866 if (top_out == NULL || top_out->klass() == NULL || 5867 top_in == NULL || top_in->klass() == NULL) { 5868 // failed array check 5869 return false; 5870 } 5871 5872 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5873 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5874 if (out_elem != T_INT || in_elem != T_INT) { 5875 return false; 5876 } 5877 5878 Node* outlen = load_array_length(out); 5879 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5880 Node* out_start = array_element_address(out, intcon(0), out_elem); 5881 Node* in_start = array_element_address(in, intcon(0), in_elem); 5882 5883 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5884 OptoRuntime::mulAdd_Type(), 5885 stubAddr, stubName, TypePtr::BOTTOM, 5886 out_start,in_start, new_offset, len, k); 5887 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5888 set_result(result); 5889 return true; 5890 } 5891 5892 //-------------inline_montgomeryMultiply----------------------------------- 5893 bool LibraryCallKit::inline_montgomeryMultiply() { 5894 address stubAddr = StubRoutines::montgomeryMultiply(); 5895 if (stubAddr == NULL) { 5896 return false; // Intrinsic's stub is not implemented on this platform 5897 } 5898 5899 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5900 const char* stubName = "montgomery_multiply"; 5901 5902 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5903 5904 Node* a = argument(0); 5905 Node* b = argument(1); 5906 Node* n = argument(2); 5907 Node* len = argument(3); 5908 Node* inv = argument(4); 5909 Node* m = argument(6); 5910 5911 const Type* a_type = a->Value(&_gvn); 5912 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5913 const Type* b_type = b->Value(&_gvn); 5914 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5915 const Type* n_type = a->Value(&_gvn); 5916 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5917 const Type* m_type = a->Value(&_gvn); 5918 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5919 if (top_a == NULL || top_a->klass() == NULL || 5920 top_b == NULL || top_b->klass() == NULL || 5921 top_n == NULL || top_n->klass() == NULL || 5922 top_m == NULL || top_m->klass() == NULL) { 5923 // failed array check 5924 return false; 5925 } 5926 5927 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5928 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5929 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5930 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5931 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5932 return false; 5933 } 5934 5935 // Make the call 5936 { 5937 Node* a_start = array_element_address(a, intcon(0), a_elem); 5938 Node* b_start = array_element_address(b, intcon(0), b_elem); 5939 Node* n_start = array_element_address(n, intcon(0), n_elem); 5940 Node* m_start = array_element_address(m, intcon(0), m_elem); 5941 5942 Node* call = make_runtime_call(RC_LEAF, 5943 OptoRuntime::montgomeryMultiply_Type(), 5944 stubAddr, stubName, TypePtr::BOTTOM, 5945 a_start, b_start, n_start, len, inv, top(), 5946 m_start); 5947 set_result(m); 5948 } 5949 5950 return true; 5951 } 5952 5953 bool LibraryCallKit::inline_montgomerySquare() { 5954 address stubAddr = StubRoutines::montgomerySquare(); 5955 if (stubAddr == NULL) { 5956 return false; // Intrinsic's stub is not implemented on this platform 5957 } 5958 5959 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5960 const char* stubName = "montgomery_square"; 5961 5962 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5963 5964 Node* a = argument(0); 5965 Node* n = argument(1); 5966 Node* len = argument(2); 5967 Node* inv = argument(3); 5968 Node* m = argument(5); 5969 5970 const Type* a_type = a->Value(&_gvn); 5971 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5972 const Type* n_type = a->Value(&_gvn); 5973 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5974 const Type* m_type = a->Value(&_gvn); 5975 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5976 if (top_a == NULL || top_a->klass() == NULL || 5977 top_n == NULL || top_n->klass() == NULL || 5978 top_m == NULL || top_m->klass() == NULL) { 5979 // failed array check 5980 return false; 5981 } 5982 5983 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5984 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5985 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5986 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5987 return false; 5988 } 5989 5990 // Make the call 5991 { 5992 Node* a_start = array_element_address(a, intcon(0), a_elem); 5993 Node* n_start = array_element_address(n, intcon(0), n_elem); 5994 Node* m_start = array_element_address(m, intcon(0), m_elem); 5995 5996 Node* call = make_runtime_call(RC_LEAF, 5997 OptoRuntime::montgomerySquare_Type(), 5998 stubAddr, stubName, TypePtr::BOTTOM, 5999 a_start, n_start, len, inv, top(), 6000 m_start); 6001 set_result(m); 6002 } 6003 6004 return true; 6005 } 6006 6007 //-------------inline_vectorizedMismatch------------------------------ 6008 bool LibraryCallKit::inline_vectorizedMismatch() { 6009 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 6010 6011 address stubAddr = StubRoutines::vectorizedMismatch(); 6012 if (stubAddr == NULL) { 6013 return false; // Intrinsic's stub is not implemented on this platform 6014 } 6015 const char* stubName = "vectorizedMismatch"; 6016 int size_l = callee()->signature()->size(); 6017 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 6018 6019 Node* obja = argument(0); 6020 Node* aoffset = argument(1); 6021 Node* objb = argument(3); 6022 Node* boffset = argument(4); 6023 Node* length = argument(6); 6024 Node* scale = argument(7); 6025 6026 const Type* a_type = obja->Value(&_gvn); 6027 const Type* b_type = objb->Value(&_gvn); 6028 const TypeAryPtr* top_a = a_type->isa_aryptr(); 6029 const TypeAryPtr* top_b = b_type->isa_aryptr(); 6030 if (top_a == NULL || top_a->klass() == NULL || 6031 top_b == NULL || top_b->klass() == NULL) { 6032 // failed array check 6033 return false; 6034 } 6035 6036 Node* call; 6037 jvms()->set_should_reexecute(true); 6038 6039 Node* obja_adr = make_unsafe_address(obja, aoffset); 6040 Node* objb_adr = make_unsafe_address(objb, boffset); 6041 6042 call = make_runtime_call(RC_LEAF, 6043 OptoRuntime::vectorizedMismatch_Type(), 6044 stubAddr, stubName, TypePtr::BOTTOM, 6045 obja_adr, objb_adr, length, scale); 6046 6047 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6048 set_result(result); 6049 return true; 6050 } 6051 6052 /** 6053 * Calculate CRC32 for byte. 6054 * int java.util.zip.CRC32.update(int crc, int b) 6055 */ 6056 bool LibraryCallKit::inline_updateCRC32() { 6057 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 6058 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 6059 // no receiver since it is static method 6060 Node* crc = argument(0); // type: int 6061 Node* b = argument(1); // type: int 6062 6063 /* 6064 * int c = ~ crc; 6065 * b = timesXtoThe32[(b ^ c) & 0xFF]; 6066 * b = b ^ (c >>> 8); 6067 * crc = ~b; 6068 */ 6069 6070 Node* M1 = intcon(-1); 6071 crc = _gvn.transform(new XorINode(crc, M1)); 6072 Node* result = _gvn.transform(new XorINode(crc, b)); 6073 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 6074 6075 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 6076 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 6077 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 6078 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 6079 6080 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 6081 result = _gvn.transform(new XorINode(crc, result)); 6082 result = _gvn.transform(new XorINode(result, M1)); 6083 set_result(result); 6084 return true; 6085 } 6086 6087 /** 6088 * Calculate CRC32 for byte[] array. 6089 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 6090 */ 6091 bool LibraryCallKit::inline_updateBytesCRC32() { 6092 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 6093 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 6094 // no receiver since it is static method 6095 Node* crc = argument(0); // type: int 6096 Node* src = argument(1); // type: oop 6097 Node* offset = argument(2); // type: int 6098 Node* length = argument(3); // type: int 6099 6100 const Type* src_type = src->Value(&_gvn); 6101 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6102 if (top_src == NULL || top_src->klass() == NULL) { 6103 // failed array check 6104 return false; 6105 } 6106 6107 // Figure out the size and type of the elements we will be copying. 6108 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6109 if (src_elem != T_BYTE) { 6110 return false; 6111 } 6112 6113 // 'src_start' points to src array + scaled offset 6114 Node* src_start = array_element_address(src, offset, src_elem); 6115 6116 // We assume that range check is done by caller. 6117 // TODO: generate range check (offset+length < src.length) in debug VM. 6118 6119 // Call the stub. 6120 address stubAddr = StubRoutines::updateBytesCRC32(); 6121 const char *stubName = "updateBytesCRC32"; 6122 6123 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6124 stubAddr, stubName, TypePtr::BOTTOM, 6125 crc, src_start, length); 6126 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6127 set_result(result); 6128 return true; 6129 } 6130 6131 /** 6132 * Calculate CRC32 for ByteBuffer. 6133 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 6134 */ 6135 bool LibraryCallKit::inline_updateByteBufferCRC32() { 6136 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 6137 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 6138 // no receiver since it is static method 6139 Node* crc = argument(0); // type: int 6140 Node* src = argument(1); // type: long 6141 Node* offset = argument(3); // type: int 6142 Node* length = argument(4); // type: int 6143 6144 src = ConvL2X(src); // adjust Java long to machine word 6145 Node* base = _gvn.transform(new CastX2PNode(src)); 6146 offset = ConvI2X(offset); 6147 6148 // 'src_start' points to src array + scaled offset 6149 Node* src_start = basic_plus_adr(top(), base, offset); 6150 6151 // Call the stub. 6152 address stubAddr = StubRoutines::updateBytesCRC32(); 6153 const char *stubName = "updateBytesCRC32"; 6154 6155 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6156 stubAddr, stubName, TypePtr::BOTTOM, 6157 crc, src_start, length); 6158 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6159 set_result(result); 6160 return true; 6161 } 6162 6163 //------------------------------get_table_from_crc32c_class----------------------- 6164 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 6165 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 6166 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 6167 6168 return table; 6169 } 6170 6171 //------------------------------inline_updateBytesCRC32C----------------------- 6172 // 6173 // Calculate CRC32C for byte[] array. 6174 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 6175 // 6176 bool LibraryCallKit::inline_updateBytesCRC32C() { 6177 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 6178 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 6179 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 6180 // no receiver since it is a static method 6181 Node* crc = argument(0); // type: int 6182 Node* src = argument(1); // type: oop 6183 Node* offset = argument(2); // type: int 6184 Node* end = argument(3); // type: int 6185 6186 Node* length = _gvn.transform(new SubINode(end, offset)); 6187 6188 const Type* src_type = src->Value(&_gvn); 6189 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6190 if (top_src == NULL || top_src->klass() == NULL) { 6191 // failed array check 6192 return false; 6193 } 6194 6195 // Figure out the size and type of the elements we will be copying. 6196 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6197 if (src_elem != T_BYTE) { 6198 return false; 6199 } 6200 6201 // 'src_start' points to src array + scaled offset 6202 Node* src_start = array_element_address(src, offset, src_elem); 6203 6204 // static final int[] byteTable in class CRC32C 6205 Node* table = get_table_from_crc32c_class(callee()->holder()); 6206 Node* table_start = array_element_address(table, intcon(0), T_INT); 6207 6208 // We assume that range check is done by caller. 6209 // TODO: generate range check (offset+length < src.length) in debug VM. 6210 6211 // Call the stub. 6212 address stubAddr = StubRoutines::updateBytesCRC32C(); 6213 const char *stubName = "updateBytesCRC32C"; 6214 6215 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 6216 stubAddr, stubName, TypePtr::BOTTOM, 6217 crc, src_start, length, table_start); 6218 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6219 set_result(result); 6220 return true; 6221 } 6222 6223 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 6224 // 6225 // Calculate CRC32C for DirectByteBuffer. 6226 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 6227 // 6228 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 6229 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 6230 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 6231 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 6232 // no receiver since it is a static method 6233 Node* crc = argument(0); // type: int 6234 Node* src = argument(1); // type: long 6235 Node* offset = argument(3); // type: int 6236 Node* end = argument(4); // type: int 6237 6238 Node* length = _gvn.transform(new SubINode(end, offset)); 6239 6240 src = ConvL2X(src); // adjust Java long to machine word 6241 Node* base = _gvn.transform(new CastX2PNode(src)); 6242 offset = ConvI2X(offset); 6243 6244 // 'src_start' points to src array + scaled offset 6245 Node* src_start = basic_plus_adr(top(), base, offset); 6246 6247 // static final int[] byteTable in class CRC32C 6248 Node* table = get_table_from_crc32c_class(callee()->holder()); 6249 Node* table_start = array_element_address(table, intcon(0), T_INT); 6250 6251 // Call the stub. 6252 address stubAddr = StubRoutines::updateBytesCRC32C(); 6253 const char *stubName = "updateBytesCRC32C"; 6254 6255 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 6256 stubAddr, stubName, TypePtr::BOTTOM, 6257 crc, src_start, length, table_start); 6258 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6259 set_result(result); 6260 return true; 6261 } 6262 6263 //------------------------------inline_updateBytesAdler32---------------------- 6264 // 6265 // Calculate Adler32 checksum for byte[] array. 6266 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 6267 // 6268 bool LibraryCallKit::inline_updateBytesAdler32() { 6269 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 6270 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 6271 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 6272 // no receiver since it is static method 6273 Node* crc = argument(0); // type: int 6274 Node* src = argument(1); // type: oop 6275 Node* offset = argument(2); // type: int 6276 Node* length = argument(3); // type: int 6277 6278 const Type* src_type = src->Value(&_gvn); 6279 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6280 if (top_src == NULL || top_src->klass() == NULL) { 6281 // failed array check 6282 return false; 6283 } 6284 6285 // Figure out the size and type of the elements we will be copying. 6286 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6287 if (src_elem != T_BYTE) { 6288 return false; 6289 } 6290 6291 // 'src_start' points to src array + scaled offset 6292 Node* src_start = array_element_address(src, offset, src_elem); 6293 6294 // We assume that range check is done by caller. 6295 // TODO: generate range check (offset+length < src.length) in debug VM. 6296 6297 // Call the stub. 6298 address stubAddr = StubRoutines::updateBytesAdler32(); 6299 const char *stubName = "updateBytesAdler32"; 6300 6301 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 6302 stubAddr, stubName, TypePtr::BOTTOM, 6303 crc, src_start, length); 6304 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6305 set_result(result); 6306 return true; 6307 } 6308 6309 //------------------------------inline_updateByteBufferAdler32--------------- 6310 // 6311 // Calculate Adler32 checksum for DirectByteBuffer. 6312 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 6313 // 6314 bool LibraryCallKit::inline_updateByteBufferAdler32() { 6315 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 6316 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 6317 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 6318 // no receiver since it is static method 6319 Node* crc = argument(0); // type: int 6320 Node* src = argument(1); // type: long 6321 Node* offset = argument(3); // type: int 6322 Node* length = argument(4); // type: int 6323 6324 src = ConvL2X(src); // adjust Java long to machine word 6325 Node* base = _gvn.transform(new CastX2PNode(src)); 6326 offset = ConvI2X(offset); 6327 6328 // 'src_start' points to src array + scaled offset 6329 Node* src_start = basic_plus_adr(top(), base, offset); 6330 6331 // Call the stub. 6332 address stubAddr = StubRoutines::updateBytesAdler32(); 6333 const char *stubName = "updateBytesAdler32"; 6334 6335 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 6336 stubAddr, stubName, TypePtr::BOTTOM, 6337 crc, src_start, length); 6338 6339 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6340 set_result(result); 6341 return true; 6342 } 6343 6344 //----------------------------inline_reference_get---------------------------- 6345 // public T java.lang.ref.Reference.get(); 6346 bool LibraryCallKit::inline_reference_get() { 6347 const int referent_offset = java_lang_ref_Reference::referent_offset; 6348 guarantee(referent_offset > 0, "should have already been set"); 6349 6350 // Get the argument: 6351 Node* reference_obj = null_check_receiver(); 6352 if (stopped()) return true; 6353 6354 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 6355 6356 ciInstanceKlass* klass = env()->Object_klass(); 6357 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 6358 6359 Node* no_ctrl = NULL; 6360 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 6361 if (UseZGC) { 6362 result = load_barrier(result, adr, true /* weak */); 6363 } 6364 6365 // Use the pre-barrier to record the value in the referent field 6366 pre_barrier(false /* do_load */, 6367 control(), 6368 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 6369 result /* pre_val */, 6370 T_OBJECT); 6371 6372 // Add memory barrier to prevent commoning reads from this field 6373 // across safepoint since GC can change its value. 6374 insert_mem_bar(Op_MemBarCPUOrder); 6375 6376 set_result(result); 6377 return true; 6378 } 6379 6380 6381 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 6382 bool is_exact=true, bool is_static=false, 6383 ciInstanceKlass * fromKls=NULL) { 6384 if (fromKls == NULL) { 6385 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 6386 assert(tinst != NULL, "obj is null"); 6387 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6388 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 6389 fromKls = tinst->klass()->as_instance_klass(); 6390 } else { 6391 assert(is_static, "only for static field access"); 6392 } 6393 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 6394 ciSymbol::make(fieldTypeString), 6395 is_static); 6396 6397 assert (field != NULL, "undefined field"); 6398 if (field == NULL) return (Node *) NULL; 6399 6400 if (is_static) { 6401 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 6402 fromObj = makecon(tip); 6403 } 6404 6405 // Next code copied from Parse::do_get_xxx(): 6406 6407 // Compute address and memory type. 6408 int offset = field->offset_in_bytes(); 6409 bool is_vol = field->is_volatile(); 6410 ciType* field_klass = field->type(); 6411 assert(field_klass->is_loaded(), "should be loaded"); 6412 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 6413 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 6414 BasicType bt = field->layout_type(); 6415 6416 // Build the resultant type of the load 6417 const Type *type; 6418 if (bt == T_OBJECT) { 6419 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 6420 } else { 6421 type = Type::get_const_basic_type(bt); 6422 } 6423 6424 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 6425 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 6426 } 6427 // Build the load. 6428 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 6429 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 6430 if (UseZGC) { 6431 loadedField = load_barrier(loadedField, adr); 6432 } 6433 6434 // If reference is volatile, prevent following memory ops from 6435 // floating up past the volatile read. Also prevents commoning 6436 // another volatile read. 6437 if (is_vol) { 6438 // Memory barrier includes bogus read of value to force load BEFORE membar 6439 insert_mem_bar(Op_MemBarAcquire, loadedField); 6440 } 6441 return loadedField; 6442 } 6443 6444 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 6445 bool is_exact = true, bool is_static = false, 6446 ciInstanceKlass * fromKls = NULL) { 6447 if (fromKls == NULL) { 6448 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 6449 assert(tinst != NULL, "obj is null"); 6450 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6451 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 6452 fromKls = tinst->klass()->as_instance_klass(); 6453 } 6454 else { 6455 assert(is_static, "only for static field access"); 6456 } 6457 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 6458 ciSymbol::make(fieldTypeString), 6459 is_static); 6460 6461 assert(field != NULL, "undefined field"); 6462 assert(!field->is_volatile(), "not defined for volatile fields"); 6463 6464 if (is_static) { 6465 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 6466 fromObj = makecon(tip); 6467 } 6468 6469 // Next code copied from Parse::do_get_xxx(): 6470 6471 // Compute address and memory type. 6472 int offset = field->offset_in_bytes(); 6473 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 6474 6475 return adr; 6476 } 6477 6478 //------------------------------inline_aescrypt_Block----------------------- 6479 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 6480 address stubAddr = NULL; 6481 const char *stubName; 6482 assert(UseAES, "need AES instruction support"); 6483 6484 switch(id) { 6485 case vmIntrinsics::_aescrypt_encryptBlock: 6486 stubAddr = StubRoutines::aescrypt_encryptBlock(); 6487 stubName = "aescrypt_encryptBlock"; 6488 break; 6489 case vmIntrinsics::_aescrypt_decryptBlock: 6490 stubAddr = StubRoutines::aescrypt_decryptBlock(); 6491 stubName = "aescrypt_decryptBlock"; 6492 break; 6493 default: 6494 break; 6495 } 6496 if (stubAddr == NULL) return false; 6497 6498 Node* aescrypt_object = argument(0); 6499 Node* src = argument(1); 6500 Node* src_offset = argument(2); 6501 Node* dest = argument(3); 6502 Node* dest_offset = argument(4); 6503 6504 // (1) src and dest are arrays. 6505 const Type* src_type = src->Value(&_gvn); 6506 const Type* dest_type = dest->Value(&_gvn); 6507 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6508 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6509 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6510 6511 // for the quick and dirty code we will skip all the checks. 6512 // we are just trying to get the call to be generated. 6513 Node* src_start = src; 6514 Node* dest_start = dest; 6515 if (src_offset != NULL || dest_offset != NULL) { 6516 assert(src_offset != NULL && dest_offset != NULL, ""); 6517 src_start = array_element_address(src, src_offset, T_BYTE); 6518 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6519 } 6520 6521 // now need to get the start of its expanded key array 6522 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6523 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6524 if (k_start == NULL) return false; 6525 6526 if (Matcher::pass_original_key_for_aes()) { 6527 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6528 // compatibility issues between Java key expansion and SPARC crypto instructions 6529 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6530 if (original_k_start == NULL) return false; 6531 6532 // Call the stub. 6533 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6534 stubAddr, stubName, TypePtr::BOTTOM, 6535 src_start, dest_start, k_start, original_k_start); 6536 } else { 6537 // Call the stub. 6538 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6539 stubAddr, stubName, TypePtr::BOTTOM, 6540 src_start, dest_start, k_start); 6541 } 6542 6543 return true; 6544 } 6545 6546 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 6547 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 6548 address stubAddr = NULL; 6549 const char *stubName = NULL; 6550 6551 assert(UseAES, "need AES instruction support"); 6552 6553 switch(id) { 6554 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6555 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6556 stubName = "cipherBlockChaining_encryptAESCrypt"; 6557 break; 6558 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6559 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6560 stubName = "cipherBlockChaining_decryptAESCrypt"; 6561 break; 6562 default: 6563 break; 6564 } 6565 if (stubAddr == NULL) return false; 6566 6567 Node* cipherBlockChaining_object = argument(0); 6568 Node* src = argument(1); 6569 Node* src_offset = argument(2); 6570 Node* len = argument(3); 6571 Node* dest = argument(4); 6572 Node* dest_offset = argument(5); 6573 6574 // (1) src and dest are arrays. 6575 const Type* src_type = src->Value(&_gvn); 6576 const Type* dest_type = dest->Value(&_gvn); 6577 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6578 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6579 assert (top_src != NULL && top_src->klass() != NULL 6580 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6581 6582 // checks are the responsibility of the caller 6583 Node* src_start = src; 6584 Node* dest_start = dest; 6585 if (src_offset != NULL || dest_offset != NULL) { 6586 assert(src_offset != NULL && dest_offset != NULL, ""); 6587 src_start = array_element_address(src, src_offset, T_BYTE); 6588 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6589 } 6590 6591 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6592 // (because of the predicated logic executed earlier). 6593 // so we cast it here safely. 6594 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6595 6596 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6597 if (embeddedCipherObj == NULL) return false; 6598 6599 // cast it to what we know it will be at runtime 6600 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6601 assert(tinst != NULL, "CBC obj is null"); 6602 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6603 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6604 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6605 6606 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6607 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6608 const TypeOopPtr* xtype = aklass->as_instance_type(); 6609 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6610 aescrypt_object = _gvn.transform(aescrypt_object); 6611 6612 // we need to get the start of the aescrypt_object's expanded key array 6613 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6614 if (k_start == NULL) return false; 6615 6616 // similarly, get the start address of the r vector 6617 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6618 if (objRvec == NULL) return false; 6619 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6620 6621 Node* cbcCrypt; 6622 if (Matcher::pass_original_key_for_aes()) { 6623 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6624 // compatibility issues between Java key expansion and SPARC crypto instructions 6625 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6626 if (original_k_start == NULL) return false; 6627 6628 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6629 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6630 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6631 stubAddr, stubName, TypePtr::BOTTOM, 6632 src_start, dest_start, k_start, r_start, len, original_k_start); 6633 } else { 6634 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6635 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6636 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6637 stubAddr, stubName, TypePtr::BOTTOM, 6638 src_start, dest_start, k_start, r_start, len); 6639 } 6640 6641 // return cipher length (int) 6642 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6643 set_result(retvalue); 6644 return true; 6645 } 6646 6647 //------------------------------inline_counterMode_AESCrypt----------------------- 6648 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6649 assert(UseAES, "need AES instruction support"); 6650 if (!UseAESCTRIntrinsics) return false; 6651 6652 address stubAddr = NULL; 6653 const char *stubName = NULL; 6654 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6655 stubAddr = StubRoutines::counterMode_AESCrypt(); 6656 stubName = "counterMode_AESCrypt"; 6657 } 6658 if (stubAddr == NULL) return false; 6659 6660 Node* counterMode_object = argument(0); 6661 Node* src = argument(1); 6662 Node* src_offset = argument(2); 6663 Node* len = argument(3); 6664 Node* dest = argument(4); 6665 Node* dest_offset = argument(5); 6666 6667 // (1) src and dest are arrays. 6668 const Type* src_type = src->Value(&_gvn); 6669 const Type* dest_type = dest->Value(&_gvn); 6670 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6671 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6672 assert(top_src != NULL && top_src->klass() != NULL && 6673 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6674 6675 // checks are the responsibility of the caller 6676 Node* src_start = src; 6677 Node* dest_start = dest; 6678 if (src_offset != NULL || dest_offset != NULL) { 6679 assert(src_offset != NULL && dest_offset != NULL, ""); 6680 src_start = array_element_address(src, src_offset, T_BYTE); 6681 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6682 } 6683 6684 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6685 // (because of the predicated logic executed earlier). 6686 // so we cast it here safely. 6687 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6688 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6689 if (embeddedCipherObj == NULL) return false; 6690 // cast it to what we know it will be at runtime 6691 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6692 assert(tinst != NULL, "CTR obj is null"); 6693 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6694 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6695 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6696 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6697 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6698 const TypeOopPtr* xtype = aklass->as_instance_type(); 6699 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6700 aescrypt_object = _gvn.transform(aescrypt_object); 6701 // we need to get the start of the aescrypt_object's expanded key array 6702 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6703 if (k_start == NULL) return false; 6704 // similarly, get the start address of the r vector 6705 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6706 if (obj_counter == NULL) return false; 6707 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6708 6709 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6710 if (saved_encCounter == NULL) return false; 6711 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6712 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6713 6714 Node* ctrCrypt; 6715 if (Matcher::pass_original_key_for_aes()) { 6716 // no SPARC version for AES/CTR intrinsics now. 6717 return false; 6718 } 6719 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6720 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6721 OptoRuntime::counterMode_aescrypt_Type(), 6722 stubAddr, stubName, TypePtr::BOTTOM, 6723 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6724 6725 // return cipher length (int) 6726 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6727 set_result(retvalue); 6728 return true; 6729 } 6730 6731 //------------------------------get_key_start_from_aescrypt_object----------------------- 6732 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6733 #if defined(PPC64) || defined(S390) 6734 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6735 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6736 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6737 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6738 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6739 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6740 if (objSessionK == NULL) { 6741 return (Node *) NULL; 6742 } 6743 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6744 #else 6745 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6746 #endif // PPC64 6747 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6748 if (objAESCryptKey == NULL) return (Node *) NULL; 6749 6750 // now have the array, need to get the start address of the K array 6751 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6752 return k_start; 6753 } 6754 6755 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6756 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6757 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6758 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6759 if (objAESCryptKey == NULL) return (Node *) NULL; 6760 6761 // now have the array, need to get the start address of the lastKey array 6762 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6763 return original_k_start; 6764 } 6765 6766 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6767 // Return node representing slow path of predicate check. 6768 // the pseudo code we want to emulate with this predicate is: 6769 // for encryption: 6770 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6771 // for decryption: 6772 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6773 // note cipher==plain is more conservative than the original java code but that's OK 6774 // 6775 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6776 // The receiver was checked for NULL already. 6777 Node* objCBC = argument(0); 6778 6779 // Load embeddedCipher field of CipherBlockChaining object. 6780 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6781 6782 // get AESCrypt klass for instanceOf check 6783 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6784 // will have same classloader as CipherBlockChaining object 6785 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6786 assert(tinst != NULL, "CBCobj is null"); 6787 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6788 6789 // we want to do an instanceof comparison against the AESCrypt class 6790 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6791 if (!klass_AESCrypt->is_loaded()) { 6792 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6793 Node* ctrl = control(); 6794 set_control(top()); // no regular fast path 6795 return ctrl; 6796 } 6797 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6798 6799 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6800 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6801 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6802 6803 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6804 6805 // for encryption, we are done 6806 if (!decrypting) 6807 return instof_false; // even if it is NULL 6808 6809 // for decryption, we need to add a further check to avoid 6810 // taking the intrinsic path when cipher and plain are the same 6811 // see the original java code for why. 6812 RegionNode* region = new RegionNode(3); 6813 region->init_req(1, instof_false); 6814 Node* src = argument(1); 6815 Node* dest = argument(4); 6816 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6817 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6818 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6819 region->init_req(2, src_dest_conjoint); 6820 6821 record_for_igvn(region); 6822 return _gvn.transform(region); 6823 } 6824 6825 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6826 // Return node representing slow path of predicate check. 6827 // the pseudo code we want to emulate with this predicate is: 6828 // for encryption: 6829 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6830 // for decryption: 6831 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6832 // note cipher==plain is more conservative than the original java code but that's OK 6833 // 6834 6835 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6836 // The receiver was checked for NULL already. 6837 Node* objCTR = argument(0); 6838 6839 // Load embeddedCipher field of CipherBlockChaining object. 6840 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6841 6842 // get AESCrypt klass for instanceOf check 6843 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6844 // will have same classloader as CipherBlockChaining object 6845 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6846 assert(tinst != NULL, "CTRobj is null"); 6847 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6848 6849 // we want to do an instanceof comparison against the AESCrypt class 6850 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6851 if (!klass_AESCrypt->is_loaded()) { 6852 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6853 Node* ctrl = control(); 6854 set_control(top()); // no regular fast path 6855 return ctrl; 6856 } 6857 6858 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6859 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6860 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6861 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6862 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6863 6864 return instof_false; // even if it is NULL 6865 } 6866 6867 //------------------------------inline_ghash_processBlocks 6868 bool LibraryCallKit::inline_ghash_processBlocks() { 6869 address stubAddr; 6870 const char *stubName; 6871 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6872 6873 stubAddr = StubRoutines::ghash_processBlocks(); 6874 stubName = "ghash_processBlocks"; 6875 6876 Node* data = argument(0); 6877 Node* offset = argument(1); 6878 Node* len = argument(2); 6879 Node* state = argument(3); 6880 Node* subkeyH = argument(4); 6881 6882 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6883 assert(state_start, "state is NULL"); 6884 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6885 assert(subkeyH_start, "subkeyH is NULL"); 6886 Node* data_start = array_element_address(data, offset, T_BYTE); 6887 assert(data_start, "data is NULL"); 6888 6889 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6890 OptoRuntime::ghash_processBlocks_Type(), 6891 stubAddr, stubName, TypePtr::BOTTOM, 6892 state_start, subkeyH_start, data_start, len); 6893 return true; 6894 } 6895 6896 //------------------------------inline_sha_implCompress----------------------- 6897 // 6898 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6899 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6900 // 6901 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6902 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6903 // 6904 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6905 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6906 // 6907 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6908 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6909 6910 Node* sha_obj = argument(0); 6911 Node* src = argument(1); // type oop 6912 Node* ofs = argument(2); // type int 6913 6914 const Type* src_type = src->Value(&_gvn); 6915 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6916 if (top_src == NULL || top_src->klass() == NULL) { 6917 // failed array check 6918 return false; 6919 } 6920 // Figure out the size and type of the elements we will be copying. 6921 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6922 if (src_elem != T_BYTE) { 6923 return false; 6924 } 6925 // 'src_start' points to src array + offset 6926 Node* src_start = array_element_address(src, ofs, src_elem); 6927 Node* state = NULL; 6928 address stubAddr; 6929 const char *stubName; 6930 6931 switch(id) { 6932 case vmIntrinsics::_sha_implCompress: 6933 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6934 state = get_state_from_sha_object(sha_obj); 6935 stubAddr = StubRoutines::sha1_implCompress(); 6936 stubName = "sha1_implCompress"; 6937 break; 6938 case vmIntrinsics::_sha2_implCompress: 6939 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6940 state = get_state_from_sha_object(sha_obj); 6941 stubAddr = StubRoutines::sha256_implCompress(); 6942 stubName = "sha256_implCompress"; 6943 break; 6944 case vmIntrinsics::_sha5_implCompress: 6945 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6946 state = get_state_from_sha5_object(sha_obj); 6947 stubAddr = StubRoutines::sha512_implCompress(); 6948 stubName = "sha512_implCompress"; 6949 break; 6950 default: 6951 fatal_unexpected_iid(id); 6952 return false; 6953 } 6954 if (state == NULL) return false; 6955 6956 // Call the stub. 6957 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6958 stubAddr, stubName, TypePtr::BOTTOM, 6959 src_start, state); 6960 6961 return true; 6962 } 6963 6964 //------------------------------inline_digestBase_implCompressMB----------------------- 6965 // 6966 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6967 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6968 // 6969 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6970 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6971 "need SHA1/SHA256/SHA512 instruction support"); 6972 assert((uint)predicate < 3, "sanity"); 6973 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6974 6975 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6976 Node* src = argument(1); // byte[] array 6977 Node* ofs = argument(2); // type int 6978 Node* limit = argument(3); // type int 6979 6980 const Type* src_type = src->Value(&_gvn); 6981 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6982 if (top_src == NULL || top_src->klass() == NULL) { 6983 // failed array check 6984 return false; 6985 } 6986 // Figure out the size and type of the elements we will be copying. 6987 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6988 if (src_elem != T_BYTE) { 6989 return false; 6990 } 6991 // 'src_start' points to src array + offset 6992 Node* src_start = array_element_address(src, ofs, src_elem); 6993 6994 const char* klass_SHA_name = NULL; 6995 const char* stub_name = NULL; 6996 address stub_addr = NULL; 6997 bool long_state = false; 6998 6999 switch (predicate) { 7000 case 0: 7001 if (UseSHA1Intrinsics) { 7002 klass_SHA_name = "sun/security/provider/SHA"; 7003 stub_name = "sha1_implCompressMB"; 7004 stub_addr = StubRoutines::sha1_implCompressMB(); 7005 } 7006 break; 7007 case 1: 7008 if (UseSHA256Intrinsics) { 7009 klass_SHA_name = "sun/security/provider/SHA2"; 7010 stub_name = "sha256_implCompressMB"; 7011 stub_addr = StubRoutines::sha256_implCompressMB(); 7012 } 7013 break; 7014 case 2: 7015 if (UseSHA512Intrinsics) { 7016 klass_SHA_name = "sun/security/provider/SHA5"; 7017 stub_name = "sha512_implCompressMB"; 7018 stub_addr = StubRoutines::sha512_implCompressMB(); 7019 long_state = true; 7020 } 7021 break; 7022 default: 7023 fatal("unknown SHA intrinsic predicate: %d", predicate); 7024 } 7025 if (klass_SHA_name != NULL) { 7026 // get DigestBase klass to lookup for SHA klass 7027 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 7028 assert(tinst != NULL, "digestBase_obj is not instance???"); 7029 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 7030 7031 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 7032 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 7033 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 7034 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 7035 } 7036 return false; 7037 } 7038 //------------------------------inline_sha_implCompressMB----------------------- 7039 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 7040 bool long_state, address stubAddr, const char *stubName, 7041 Node* src_start, Node* ofs, Node* limit) { 7042 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 7043 const TypeOopPtr* xtype = aklass->as_instance_type(); 7044 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 7045 sha_obj = _gvn.transform(sha_obj); 7046 7047 Node* state; 7048 if (long_state) { 7049 state = get_state_from_sha5_object(sha_obj); 7050 } else { 7051 state = get_state_from_sha_object(sha_obj); 7052 } 7053 if (state == NULL) return false; 7054 7055 // Call the stub. 7056 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 7057 OptoRuntime::digestBase_implCompressMB_Type(), 7058 stubAddr, stubName, TypePtr::BOTTOM, 7059 src_start, state, ofs, limit); 7060 // return ofs (int) 7061 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 7062 set_result(result); 7063 7064 return true; 7065 } 7066 7067 //------------------------------get_state_from_sha_object----------------------- 7068 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 7069 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 7070 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 7071 if (sha_state == NULL) return (Node *) NULL; 7072 7073 // now have the array, need to get the start address of the state array 7074 Node* state = array_element_address(sha_state, intcon(0), T_INT); 7075 return state; 7076 } 7077 7078 //------------------------------get_state_from_sha5_object----------------------- 7079 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 7080 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 7081 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 7082 if (sha_state == NULL) return (Node *) NULL; 7083 7084 // now have the array, need to get the start address of the state array 7085 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 7086 return state; 7087 } 7088 7089 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 7090 // Return node representing slow path of predicate check. 7091 // the pseudo code we want to emulate with this predicate is: 7092 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 7093 // 7094 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 7095 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 7096 "need SHA1/SHA256/SHA512 instruction support"); 7097 assert((uint)predicate < 3, "sanity"); 7098 7099 // The receiver was checked for NULL already. 7100 Node* digestBaseObj = argument(0); 7101 7102 // get DigestBase klass for instanceOf check 7103 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 7104 assert(tinst != NULL, "digestBaseObj is null"); 7105 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 7106 7107 const char* klass_SHA_name = NULL; 7108 switch (predicate) { 7109 case 0: 7110 if (UseSHA1Intrinsics) { 7111 // we want to do an instanceof comparison against the SHA class 7112 klass_SHA_name = "sun/security/provider/SHA"; 7113 } 7114 break; 7115 case 1: 7116 if (UseSHA256Intrinsics) { 7117 // we want to do an instanceof comparison against the SHA2 class 7118 klass_SHA_name = "sun/security/provider/SHA2"; 7119 } 7120 break; 7121 case 2: 7122 if (UseSHA512Intrinsics) { 7123 // we want to do an instanceof comparison against the SHA5 class 7124 klass_SHA_name = "sun/security/provider/SHA5"; 7125 } 7126 break; 7127 default: 7128 fatal("unknown SHA intrinsic predicate: %d", predicate); 7129 } 7130 7131 ciKlass* klass_SHA = NULL; 7132 if (klass_SHA_name != NULL) { 7133 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 7134 } 7135 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 7136 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 7137 Node* ctrl = control(); 7138 set_control(top()); // no intrinsic path 7139 return ctrl; 7140 } 7141 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 7142 7143 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 7144 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 7145 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 7146 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 7147 7148 return instof_false; // even if it is NULL 7149 } 7150 7151 //-------------inline_fma----------------------------------- 7152 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) { 7153 Node *a = NULL; 7154 Node *b = NULL; 7155 Node *c = NULL; 7156 Node* result = NULL; 7157 switch (id) { 7158 case vmIntrinsics::_fmaD: 7159 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each."); 7160 // no receiver since it is static method 7161 a = round_double_node(argument(0)); 7162 b = round_double_node(argument(2)); 7163 c = round_double_node(argument(4)); 7164 result = _gvn.transform(new FmaDNode(control(), a, b, c)); 7165 break; 7166 case vmIntrinsics::_fmaF: 7167 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each."); 7168 a = argument(0); 7169 b = argument(1); 7170 c = argument(2); 7171 result = _gvn.transform(new FmaFNode(control(), a, b, c)); 7172 break; 7173 default: 7174 fatal_unexpected_iid(id); break; 7175 } 7176 set_result(result); 7177 return true; 7178 } 7179 7180 bool LibraryCallKit::inline_profileBoolean() { 7181 Node* counts = argument(1); 7182 const TypeAryPtr* ary = NULL; 7183 ciArray* aobj = NULL; 7184 if (counts->is_Con() 7185 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 7186 && (aobj = ary->const_oop()->as_array()) != NULL 7187 && (aobj->length() == 2)) { 7188 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 7189 jint false_cnt = aobj->element_value(0).as_int(); 7190 jint true_cnt = aobj->element_value(1).as_int(); 7191 7192 if (C->log() != NULL) { 7193 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 7194 false_cnt, true_cnt); 7195 } 7196 7197 if (false_cnt + true_cnt == 0) { 7198 // According to profile, never executed. 7199 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 7200 Deoptimization::Action_reinterpret); 7201 return true; 7202 } 7203 7204 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 7205 // is a number of each value occurrences. 7206 Node* result = argument(0); 7207 if (false_cnt == 0 || true_cnt == 0) { 7208 // According to profile, one value has been never seen. 7209 int expected_val = (false_cnt == 0) ? 1 : 0; 7210 7211 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 7212 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 7213 7214 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 7215 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 7216 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 7217 7218 { // Slow path: uncommon trap for never seen value and then reexecute 7219 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 7220 // the value has been seen at least once. 7221 PreserveJVMState pjvms(this); 7222 PreserveReexecuteState preexecs(this); 7223 jvms()->set_should_reexecute(true); 7224 7225 set_control(slow_path); 7226 set_i_o(i_o()); 7227 7228 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 7229 Deoptimization::Action_reinterpret); 7230 } 7231 // The guard for never seen value enables sharpening of the result and 7232 // returning a constant. It allows to eliminate branches on the same value 7233 // later on. 7234 set_control(fast_path); 7235 result = intcon(expected_val); 7236 } 7237 // Stop profiling. 7238 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 7239 // By replacing method body with profile data (represented as ProfileBooleanNode 7240 // on IR level) we effectively disable profiling. 7241 // It enables full speed execution once optimized code is generated. 7242 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 7243 C->record_for_igvn(profile); 7244 set_result(profile); 7245 return true; 7246 } else { 7247 // Continue profiling. 7248 // Profile data isn't available at the moment. So, execute method's bytecode version. 7249 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 7250 // is compiled and counters aren't available since corresponding MethodHandle 7251 // isn't a compile-time constant. 7252 return false; 7253 } 7254 } 7255 7256 bool LibraryCallKit::inline_isCompileConstant() { 7257 Node* n = argument(0); 7258 set_result(n->is_Con() ? intcon(1) : intcon(0)); 7259 return true; 7260 }