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