1 /* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "gc/shenandoah/shenandoahRuntime.hpp" 32 #include "oops/objArrayKlass.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/arraycopynode.hpp" 35 #include "opto/c2compiler.hpp" 36 #include "opto/callGenerator.hpp" 37 #include "opto/castnode.hpp" 38 #include "opto/cfgnode.hpp" 39 #include "opto/convertnode.hpp" 40 #include "opto/countbitsnode.hpp" 41 #include "opto/intrinsicnode.hpp" 42 #include "opto/idealKit.hpp" 43 #include "opto/mathexactnode.hpp" 44 #include "opto/movenode.hpp" 45 #include "opto/mulnode.hpp" 46 #include "opto/narrowptrnode.hpp" 47 #include "opto/opaquenode.hpp" 48 #include "opto/parse.hpp" 49 #include "opto/runtime.hpp" 50 #include "opto/shenandoahSupport.hpp" 51 #include "opto/subnode.hpp" 52 #include "prims/nativeLookup.hpp" 53 #include "runtime/sharedRuntime.hpp" 54 #include "trace/traceMacros.hpp" 55 56 class LibraryIntrinsic : public InlineCallGenerator { 57 // Extend the set of intrinsics known to the runtime: 58 public: 59 private: 60 bool _is_virtual; 61 bool _does_virtual_dispatch; 62 int8_t _predicates_count; // Intrinsic is predicated by several conditions 63 int8_t _last_predicate; // Last generated predicate 64 vmIntrinsics::ID _intrinsic_id; 65 66 public: 67 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 68 : InlineCallGenerator(m), 69 _is_virtual(is_virtual), 70 _does_virtual_dispatch(does_virtual_dispatch), 71 _predicates_count((int8_t)predicates_count), 72 _last_predicate((int8_t)-1), 73 _intrinsic_id(id) 74 { 75 } 76 virtual bool is_intrinsic() const { return true; } 77 virtual bool is_virtual() const { return _is_virtual; } 78 virtual bool is_predicated() const { return _predicates_count > 0; } 79 virtual int predicates_count() const { return _predicates_count; } 80 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 81 virtual JVMState* generate(JVMState* jvms); 82 virtual Node* generate_predicate(JVMState* jvms, int predicate); 83 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 84 }; 85 86 87 // Local helper class for LibraryIntrinsic: 88 class LibraryCallKit : public GraphKit { 89 private: 90 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 91 Node* _result; // the result node, if any 92 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 93 94 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 95 96 public: 97 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 98 : GraphKit(jvms), 99 _intrinsic(intrinsic), 100 _result(NULL) 101 { 102 // Check if this is a root compile. In that case we don't have a caller. 103 if (!jvms->has_method()) { 104 _reexecute_sp = sp(); 105 } else { 106 // Find out how many arguments the interpreter needs when deoptimizing 107 // and save the stack pointer value so it can used by uncommon_trap. 108 // We find the argument count by looking at the declared signature. 109 bool ignored_will_link; 110 ciSignature* declared_signature = NULL; 111 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 112 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 113 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 114 } 115 } 116 117 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 118 119 ciMethod* caller() const { return jvms()->method(); } 120 int bci() const { return jvms()->bci(); } 121 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 122 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 123 ciMethod* callee() const { return _intrinsic->method(); } 124 125 bool try_to_inline(int predicate); 126 Node* try_to_predicate(int predicate); 127 128 void push_result() { 129 // Push the result onto the stack. 130 if (!stopped() && result() != NULL) { 131 BasicType bt = result()->bottom_type()->basic_type(); 132 push_node(bt, result()); 133 } 134 } 135 136 private: 137 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 138 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 139 } 140 141 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 142 void set_result(RegionNode* region, PhiNode* value); 143 Node* result() { return _result; } 144 145 virtual int reexecute_sp() { return _reexecute_sp; } 146 147 // Helper functions to inline natives 148 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 149 Node* generate_slow_guard(Node* test, RegionNode* region); 150 Node* generate_fair_guard(Node* test, RegionNode* region); 151 Node* generate_negative_guard(Node* index, RegionNode* region, 152 // resulting CastII of index: 153 Node* *pos_index = NULL); 154 Node* generate_limit_guard(Node* offset, Node* subseq_length, 155 Node* array_length, 156 RegionNode* region); 157 Node* generate_current_thread(Node* &tls_output); 158 Node* load_mirror_from_klass(Node* klass); 159 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path, 161 int offset); 162 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 163 RegionNode* region, int null_path) { 164 int offset = java_lang_Class::klass_offset_in_bytes(); 165 return load_klass_from_mirror_common(mirror, never_see_null, 166 region, null_path, 167 offset); 168 } 169 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 170 RegionNode* region, int null_path) { 171 int offset = java_lang_Class::array_klass_offset_in_bytes(); 172 return load_klass_from_mirror_common(mirror, never_see_null, 173 region, null_path, 174 offset); 175 } 176 Node* generate_access_flags_guard(Node* kls, 177 int modifier_mask, int modifier_bits, 178 RegionNode* region); 179 Node* generate_interface_guard(Node* kls, RegionNode* region); 180 Node* generate_array_guard(Node* kls, RegionNode* region) { 181 return generate_array_guard_common(kls, region, false, false); 182 } 183 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 184 return generate_array_guard_common(kls, region, false, true); 185 } 186 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 187 return generate_array_guard_common(kls, region, true, false); 188 } 189 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 190 return generate_array_guard_common(kls, region, true, true); 191 } 192 Node* generate_array_guard_common(Node* kls, RegionNode* region, 193 bool obj_array, bool not_array); 194 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 195 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 196 bool is_virtual = false, bool is_static = false); 197 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 198 return generate_method_call(method_id, false, true); 199 } 200 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 201 return generate_method_call(method_id, true, false); 202 } 203 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 204 205 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 206 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 207 bool inline_string_compareTo(); 208 bool inline_string_indexOf(); 209 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 210 bool inline_string_equals(); 211 Node* round_double_node(Node* n); 212 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 213 bool inline_math_native(vmIntrinsics::ID id); 214 bool inline_trig(vmIntrinsics::ID id); 215 bool inline_math(vmIntrinsics::ID id); 216 template <typename OverflowOp> 217 bool inline_math_overflow(Node* arg1, Node* arg2); 218 void inline_math_mathExact(Node* math, Node* test); 219 bool inline_math_addExactI(bool is_increment); 220 bool inline_math_addExactL(bool is_increment); 221 bool inline_math_multiplyExactI(); 222 bool inline_math_multiplyExactL(); 223 bool inline_math_negateExactI(); 224 bool inline_math_negateExactL(); 225 bool inline_math_subtractExactI(bool is_decrement); 226 bool inline_math_subtractExactL(bool is_decrement); 227 bool inline_exp(); 228 bool inline_pow(); 229 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 230 bool inline_min_max(vmIntrinsics::ID id); 231 bool inline_notify(vmIntrinsics::ID id); 232 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 233 // This returns Type::AnyPtr, RawPtr, or OopPtr. 234 int classify_unsafe_addr(Node* &base, Node* &offset); 235 Node* make_unsafe_address(Node* base, Node* offset); 236 // Helper for inline_unsafe_access. 237 // Generates the guards that check whether the result of 238 // Unsafe.getObject should be recorded in an SATB log buffer. 239 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 240 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 241 static bool klass_needs_init_guard(Node* kls); 242 bool inline_unsafe_allocate(); 243 bool inline_unsafe_copyMemory(); 244 bool inline_native_currentThread(); 245 #ifdef TRACE_HAVE_INTRINSICS 246 bool inline_native_classID(); 247 bool inline_native_threadID(); 248 #endif 249 bool inline_native_time_funcs(address method, const char* funcName); 250 bool inline_native_isInterrupted(); 251 bool inline_native_Class_query(vmIntrinsics::ID id); 252 bool inline_native_subtype_check(); 253 254 bool inline_native_newArray(); 255 bool inline_native_getLength(); 256 bool inline_array_copyOf(bool is_copyOfRange); 257 bool inline_array_equals(); 258 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 259 bool inline_native_clone(bool is_virtual); 260 bool inline_native_Reflection_getCallerClass(); 261 // Helper function for inlining native object hash method 262 bool inline_native_hashcode(bool is_virtual, bool is_static); 263 bool inline_native_getClass(); 264 265 // Helper functions for inlining arraycopy 266 bool inline_arraycopy(); 267 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 268 RegionNode* slow_region); 269 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 270 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 271 272 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 273 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 274 bool inline_unsafe_ordered_store(BasicType type); 275 bool inline_unsafe_fence(vmIntrinsics::ID id); 276 bool inline_fp_conversions(vmIntrinsics::ID id); 277 bool inline_number_methods(vmIntrinsics::ID id); 278 bool inline_reference_get(); 279 bool inline_Class_cast(); 280 bool inline_aescrypt_Block(vmIntrinsics::ID id); 281 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 282 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 283 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 284 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 285 bool inline_ghash_processBlocks(); 286 bool inline_sha_implCompress(vmIntrinsics::ID id); 287 bool inline_digestBase_implCompressMB(int predicate); 288 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 289 bool long_state, address stubAddr, const char *stubName, 290 Node* src_start, Node* ofs, Node* limit); 291 Node* get_state_from_sha_object(Node *sha_object); 292 Node* get_state_from_sha5_object(Node *sha_object); 293 Node* inline_digestBase_implCompressMB_predicate(int predicate); 294 bool inline_encodeISOArray(); 295 bool inline_updateCRC32(); 296 bool inline_updateBytesCRC32(); 297 bool inline_updateByteBufferCRC32(); 298 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 299 bool inline_updateBytesCRC32C(); 300 bool inline_updateDirectByteBufferCRC32C(); 301 bool inline_updateBytesAdler32(); 302 bool inline_updateByteBufferAdler32(); 303 bool inline_multiplyToLen(); 304 bool inline_squareToLen(); 305 bool inline_mulAdd(); 306 bool inline_montgomeryMultiply(); 307 bool inline_montgomerySquare(); 308 309 bool inline_profileBoolean(); 310 bool inline_isCompileConstant(); 311 }; 312 313 //---------------------------make_vm_intrinsic---------------------------- 314 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 315 vmIntrinsics::ID id = m->intrinsic_id(); 316 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 317 318 if (!m->is_loaded()) { 319 // Do not attempt to inline unloaded methods. 320 return NULL; 321 } 322 323 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 324 bool is_available = false; 325 326 { 327 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 328 // the compiler must transition to '_thread_in_vm' state because both 329 // methods access VM-internal data. 330 VM_ENTRY_MARK; 331 methodHandle mh(THREAD, m->get_Method()); 332 methodHandle ct(THREAD, method()->get_Method()); 333 is_available = compiler->is_intrinsic_supported(mh, is_virtual) && 334 !vmIntrinsics::is_disabled_by_flags(mh, ct); 335 } 336 337 if (is_available) { 338 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 339 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 340 return new LibraryIntrinsic(m, is_virtual, 341 vmIntrinsics::predicates_needed(id), 342 vmIntrinsics::does_virtual_dispatch(id), 343 (vmIntrinsics::ID) id); 344 } else { 345 return NULL; 346 } 347 } 348 349 //----------------------register_library_intrinsics----------------------- 350 // Initialize this file's data structures, for each Compile instance. 351 void Compile::register_library_intrinsics() { 352 // Nothing to do here. 353 } 354 355 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 356 LibraryCallKit kit(jvms, this); 357 Compile* C = kit.C; 358 int nodes = C->unique(); 359 #ifndef PRODUCT 360 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 361 char buf[1000]; 362 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 363 tty->print_cr("Intrinsic %s", str); 364 } 365 #endif 366 ciMethod* callee = kit.callee(); 367 const int bci = kit.bci(); 368 369 // Try to inline the intrinsic. 370 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 371 kit.try_to_inline(_last_predicate)) { 372 if (C->print_intrinsics() || C->print_inlining()) { 373 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 374 } 375 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 376 if (C->log()) { 377 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 378 vmIntrinsics::name_at(intrinsic_id()), 379 (is_virtual() ? " virtual='1'" : ""), 380 C->unique() - nodes); 381 } 382 // Push the result from the inlined method onto the stack. 383 kit.push_result(); 384 C->print_inlining_update(this); 385 return kit.transfer_exceptions_into_jvms(); 386 } 387 388 // The intrinsic bailed out 389 if (C->print_intrinsics() || C->print_inlining()) { 390 if (jvms->has_method()) { 391 // Not a root compile. 392 const char* msg; 393 if (callee->intrinsic_candidate()) { 394 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 395 } else { 396 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 397 : "failed to inline (intrinsic), method not annotated"; 398 } 399 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 400 } else { 401 // Root compile 402 tty->print("Did not generate intrinsic %s%s at bci:%d in", 403 vmIntrinsics::name_at(intrinsic_id()), 404 (is_virtual() ? " (virtual)" : ""), bci); 405 } 406 } 407 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 408 C->print_inlining_update(this); 409 return NULL; 410 } 411 412 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 413 LibraryCallKit kit(jvms, this); 414 Compile* C = kit.C; 415 int nodes = C->unique(); 416 _last_predicate = predicate; 417 #ifndef PRODUCT 418 assert(is_predicated() && predicate < predicates_count(), "sanity"); 419 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 420 char buf[1000]; 421 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 422 tty->print_cr("Predicate for intrinsic %s", str); 423 } 424 #endif 425 ciMethod* callee = kit.callee(); 426 const int bci = kit.bci(); 427 428 Node* slow_ctl = kit.try_to_predicate(predicate); 429 if (!kit.failing()) { 430 if (C->print_intrinsics() || C->print_inlining()) { 431 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 432 } 433 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 434 if (C->log()) { 435 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 436 vmIntrinsics::name_at(intrinsic_id()), 437 (is_virtual() ? " virtual='1'" : ""), 438 C->unique() - nodes); 439 } 440 return slow_ctl; // Could be NULL if the check folds. 441 } 442 443 // The intrinsic bailed out 444 if (C->print_intrinsics() || C->print_inlining()) { 445 if (jvms->has_method()) { 446 // Not a root compile. 447 const char* msg = "failed to generate predicate for intrinsic"; 448 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 449 } else { 450 // Root compile 451 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 452 vmIntrinsics::name_at(intrinsic_id()), 453 (is_virtual() ? " (virtual)" : ""), bci); 454 } 455 } 456 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 457 return NULL; 458 } 459 460 bool LibraryCallKit::try_to_inline(int predicate) { 461 // Handle symbolic names for otherwise undistinguished boolean switches: 462 const bool is_store = true; 463 const bool is_native_ptr = true; 464 const bool is_static = true; 465 const bool is_volatile = true; 466 467 if (!jvms()->has_method()) { 468 // Root JVMState has a null method. 469 assert(map()->memory()->Opcode() == Op_Parm, ""); 470 // Insert the memory aliasing node 471 set_all_memory(reset_memory()); 472 } 473 assert(merged_memory(), ""); 474 475 476 switch (intrinsic_id()) { 477 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 478 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 479 case vmIntrinsics::_getClass: return inline_native_getClass(); 480 481 case vmIntrinsics::_dsin: 482 case vmIntrinsics::_dcos: 483 case vmIntrinsics::_dtan: 484 case vmIntrinsics::_dabs: 485 case vmIntrinsics::_datan2: 486 case vmIntrinsics::_dsqrt: 487 case vmIntrinsics::_dexp: 488 case vmIntrinsics::_dlog: 489 case vmIntrinsics::_dlog10: 490 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 491 492 case vmIntrinsics::_min: 493 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 494 495 case vmIntrinsics::_notify: 496 case vmIntrinsics::_notifyAll: 497 if (InlineNotify) { 498 return inline_notify(intrinsic_id()); 499 } 500 return false; 501 502 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 503 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 504 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 505 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 506 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 507 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 508 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 509 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 510 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 511 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 512 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 513 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 514 515 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 516 517 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 518 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 519 case vmIntrinsics::_equals: return inline_string_equals(); 520 521 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 522 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 523 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 524 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 525 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 526 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 527 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 528 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 529 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 530 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 531 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 532 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 533 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 534 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 535 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 536 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 537 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 538 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 539 540 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 541 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 542 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 543 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 544 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 545 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 546 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 547 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 548 549 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 550 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 551 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 552 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 553 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 554 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 555 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 556 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 557 558 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 559 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 560 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 561 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 562 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 563 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 564 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 565 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 566 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 567 568 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 569 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 570 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 571 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 572 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 573 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 574 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 575 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 576 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 577 578 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 579 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 580 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 581 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 582 583 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 584 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 585 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 586 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 587 588 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 589 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 590 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 591 592 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 593 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 594 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 595 596 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 597 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 598 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 599 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 600 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 601 602 case vmIntrinsics::_loadFence: 603 case vmIntrinsics::_storeFence: 604 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 605 606 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 607 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 608 609 #ifdef TRACE_HAVE_INTRINSICS 610 case vmIntrinsics::_classID: return inline_native_classID(); 611 case vmIntrinsics::_threadID: return inline_native_threadID(); 612 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 613 #endif 614 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 615 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 616 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 617 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 618 case vmIntrinsics::_newArray: return inline_native_newArray(); 619 case vmIntrinsics::_getLength: return inline_native_getLength(); 620 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 621 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 622 case vmIntrinsics::_equalsC: return inline_array_equals(); 623 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 624 625 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 626 627 case vmIntrinsics::_isInstance: 628 case vmIntrinsics::_getModifiers: 629 case vmIntrinsics::_isInterface: 630 case vmIntrinsics::_isArray: 631 case vmIntrinsics::_isPrimitive: 632 case vmIntrinsics::_getSuperclass: 633 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 634 635 case vmIntrinsics::_floatToRawIntBits: 636 case vmIntrinsics::_floatToIntBits: 637 case vmIntrinsics::_intBitsToFloat: 638 case vmIntrinsics::_doubleToRawLongBits: 639 case vmIntrinsics::_doubleToLongBits: 640 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 641 642 case vmIntrinsics::_numberOfLeadingZeros_i: 643 case vmIntrinsics::_numberOfLeadingZeros_l: 644 case vmIntrinsics::_numberOfTrailingZeros_i: 645 case vmIntrinsics::_numberOfTrailingZeros_l: 646 case vmIntrinsics::_bitCount_i: 647 case vmIntrinsics::_bitCount_l: 648 case vmIntrinsics::_reverseBytes_i: 649 case vmIntrinsics::_reverseBytes_l: 650 case vmIntrinsics::_reverseBytes_s: 651 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 652 653 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 654 655 case vmIntrinsics::_Reference_get: return inline_reference_get(); 656 657 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 658 659 case vmIntrinsics::_aescrypt_encryptBlock: 660 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 661 662 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 663 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 664 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 665 666 case vmIntrinsics::_sha_implCompress: 667 case vmIntrinsics::_sha2_implCompress: 668 case vmIntrinsics::_sha5_implCompress: 669 return inline_sha_implCompress(intrinsic_id()); 670 671 case vmIntrinsics::_digestBase_implCompressMB: 672 return inline_digestBase_implCompressMB(predicate); 673 674 case vmIntrinsics::_multiplyToLen: 675 return inline_multiplyToLen(); 676 677 case vmIntrinsics::_squareToLen: 678 return inline_squareToLen(); 679 680 case vmIntrinsics::_mulAdd: 681 return inline_mulAdd(); 682 683 case vmIntrinsics::_montgomeryMultiply: 684 return inline_montgomeryMultiply(); 685 case vmIntrinsics::_montgomerySquare: 686 return inline_montgomerySquare(); 687 688 case vmIntrinsics::_ghash_processBlocks: 689 return inline_ghash_processBlocks(); 690 691 case vmIntrinsics::_encodeISOArray: 692 return inline_encodeISOArray(); 693 694 case vmIntrinsics::_updateCRC32: 695 return inline_updateCRC32(); 696 case vmIntrinsics::_updateBytesCRC32: 697 return inline_updateBytesCRC32(); 698 case vmIntrinsics::_updateByteBufferCRC32: 699 return inline_updateByteBufferCRC32(); 700 701 case vmIntrinsics::_updateBytesCRC32C: 702 return inline_updateBytesCRC32C(); 703 case vmIntrinsics::_updateDirectByteBufferCRC32C: 704 return inline_updateDirectByteBufferCRC32C(); 705 706 case vmIntrinsics::_updateBytesAdler32: 707 return inline_updateBytesAdler32(); 708 case vmIntrinsics::_updateByteBufferAdler32: 709 return inline_updateByteBufferAdler32(); 710 711 case vmIntrinsics::_profileBoolean: 712 return inline_profileBoolean(); 713 case vmIntrinsics::_isCompileConstant: 714 return inline_isCompileConstant(); 715 716 default: 717 // If you get here, it may be that someone has added a new intrinsic 718 // to the list in vmSymbols.hpp without implementing it here. 719 #ifndef PRODUCT 720 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 721 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 722 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 723 } 724 #endif 725 return false; 726 } 727 } 728 729 Node* LibraryCallKit::try_to_predicate(int predicate) { 730 if (!jvms()->has_method()) { 731 // Root JVMState has a null method. 732 assert(map()->memory()->Opcode() == Op_Parm, ""); 733 // Insert the memory aliasing node 734 set_all_memory(reset_memory()); 735 } 736 assert(merged_memory(), ""); 737 738 switch (intrinsic_id()) { 739 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 740 return inline_cipherBlockChaining_AESCrypt_predicate(false); 741 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 742 return inline_cipherBlockChaining_AESCrypt_predicate(true); 743 case vmIntrinsics::_digestBase_implCompressMB: 744 return inline_digestBase_implCompressMB_predicate(predicate); 745 746 default: 747 // If you get here, it may be that someone has added a new intrinsic 748 // to the list in vmSymbols.hpp without implementing it here. 749 #ifndef PRODUCT 750 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 751 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 752 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 753 } 754 #endif 755 Node* slow_ctl = control(); 756 set_control(top()); // No fast path instrinsic 757 return slow_ctl; 758 } 759 } 760 761 //------------------------------set_result------------------------------- 762 // Helper function for finishing intrinsics. 763 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 764 record_for_igvn(region); 765 set_control(_gvn.transform(region)); 766 set_result( _gvn.transform(value)); 767 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 768 } 769 770 //------------------------------generate_guard--------------------------- 771 // Helper function for generating guarded fast-slow graph structures. 772 // The given 'test', if true, guards a slow path. If the test fails 773 // then a fast path can be taken. (We generally hope it fails.) 774 // In all cases, GraphKit::control() is updated to the fast path. 775 // The returned value represents the control for the slow path. 776 // The return value is never 'top'; it is either a valid control 777 // or NULL if it is obvious that the slow path can never be taken. 778 // Also, if region and the slow control are not NULL, the slow edge 779 // is appended to the region. 780 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 781 if (stopped()) { 782 // Already short circuited. 783 return NULL; 784 } 785 786 // Build an if node and its projections. 787 // If test is true we take the slow path, which we assume is uncommon. 788 if (_gvn.type(test) == TypeInt::ZERO) { 789 // The slow branch is never taken. No need to build this guard. 790 return NULL; 791 } 792 793 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 794 795 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 796 if (if_slow == top()) { 797 // The slow branch is never taken. No need to build this guard. 798 return NULL; 799 } 800 801 if (region != NULL) 802 region->add_req(if_slow); 803 804 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 805 set_control(if_fast); 806 807 return if_slow; 808 } 809 810 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 811 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 812 } 813 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 814 return generate_guard(test, region, PROB_FAIR); 815 } 816 817 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 818 Node* *pos_index) { 819 if (stopped()) 820 return NULL; // already stopped 821 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 822 return NULL; // index is already adequately typed 823 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 824 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 825 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 826 if (is_neg != NULL && pos_index != NULL) { 827 // Emulate effect of Parse::adjust_map_after_if. 828 Node* ccast = new CastIINode(index, TypeInt::POS); 829 ccast->set_req(0, control()); 830 (*pos_index) = _gvn.transform(ccast); 831 } 832 return is_neg; 833 } 834 835 // Make sure that 'position' is a valid limit index, in [0..length]. 836 // There are two equivalent plans for checking this: 837 // A. (offset + copyLength) unsigned<= arrayLength 838 // B. offset <= (arrayLength - copyLength) 839 // We require that all of the values above, except for the sum and 840 // difference, are already known to be non-negative. 841 // Plan A is robust in the face of overflow, if offset and copyLength 842 // are both hugely positive. 843 // 844 // Plan B is less direct and intuitive, but it does not overflow at 845 // all, since the difference of two non-negatives is always 846 // representable. Whenever Java methods must perform the equivalent 847 // check they generally use Plan B instead of Plan A. 848 // For the moment we use Plan A. 849 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 850 Node* subseq_length, 851 Node* array_length, 852 RegionNode* region) { 853 if (stopped()) 854 return NULL; // already stopped 855 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 856 if (zero_offset && subseq_length->eqv_uncast(array_length)) 857 return NULL; // common case of whole-array copy 858 Node* last = subseq_length; 859 if (!zero_offset) // last += offset 860 last = _gvn.transform(new AddINode(last, offset)); 861 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 862 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 863 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 864 return is_over; 865 } 866 867 868 //--------------------------generate_current_thread-------------------- 869 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 870 ciKlass* thread_klass = env()->Thread_klass(); 871 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 872 Node* thread = _gvn.transform(new ThreadLocalNode()); 873 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 874 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 875 tls_output = thread; 876 return threadObj; 877 } 878 879 880 //------------------------------make_string_method_node------------------------ 881 // Helper method for String intrinsic functions. This version is called 882 // with str1 and str2 pointing to String object nodes. 883 // 884 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 885 Node* no_ctrl = NULL; 886 887 // Get start addr of string 888 Node* str1_value = load_String_value(no_ctrl, str1); 889 Node* str1_offset = load_String_offset(no_ctrl, str1); 890 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 891 892 // Get length of string 1 893 Node* str1_len = load_String_length(no_ctrl, str1); 894 895 Node* str2_value = load_String_value(no_ctrl, str2); 896 Node* str2_offset = load_String_offset(no_ctrl, str2); 897 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 898 899 Node* str2_len = NULL; 900 Node* result = NULL; 901 902 switch (opcode) { 903 case Op_StrIndexOf: 904 // Get length of string 2 905 str2_len = load_String_length(no_ctrl, str2); 906 907 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 908 str1_start, str1_len, str2_start, str2_len); 909 break; 910 case Op_StrComp: 911 // Get length of string 2 912 str2_len = load_String_length(no_ctrl, str2); 913 914 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 915 str1_start, str1_len, str2_start, str2_len); 916 break; 917 case Op_StrEquals: 918 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 919 str1_start, str2_start, str1_len); 920 break; 921 default: 922 ShouldNotReachHere(); 923 return NULL; 924 } 925 926 // All these intrinsics have checks. 927 C->set_has_split_ifs(true); // Has chance for split-if optimization 928 929 return _gvn.transform(result); 930 } 931 932 // Helper method for String intrinsic functions. This version is called 933 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 934 // to Int nodes containing the lenghts of str1 and str2. 935 // 936 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 937 Node* result = NULL; 938 switch (opcode) { 939 case Op_StrIndexOf: 940 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 941 str1_start, cnt1, str2_start, cnt2); 942 break; 943 case Op_StrComp: 944 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 945 str1_start, cnt1, str2_start, cnt2); 946 break; 947 case Op_StrEquals: 948 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 949 str1_start, str2_start, cnt1); 950 break; 951 default: 952 ShouldNotReachHere(); 953 return NULL; 954 } 955 956 // All these intrinsics have checks. 957 C->set_has_split_ifs(true); // Has chance for split-if optimization 958 959 return _gvn.transform(result); 960 } 961 962 //------------------------------inline_string_compareTo------------------------ 963 // public int java.lang.String.compareTo(String anotherString); 964 bool LibraryCallKit::inline_string_compareTo() { 965 Node* receiver = null_check(argument(0)); 966 Node* arg = null_check(argument(1)); 967 if (stopped()) { 968 return true; 969 } 970 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 971 return true; 972 } 973 974 //------------------------------inline_string_equals------------------------ 975 bool LibraryCallKit::inline_string_equals() { 976 Node* receiver = null_check_receiver(); 977 978 if (ShenandoahVerifyReadsToFromSpace) { 979 receiver = shenandoah_read_barrier(receiver); 980 } 981 982 // NOTE: Do not null check argument for String.equals() because spec 983 // allows to specify NULL as argument. 984 Node* argument = this->argument(1); 985 986 if (ShenandoahVerifyReadsToFromSpace) { 987 argument = shenandoah_read_barrier(argument); 988 } 989 990 if (stopped()) { 991 return true; 992 } 993 994 // paths (plus control) merge 995 RegionNode* region = new RegionNode(5); 996 Node* phi = new PhiNode(region, TypeInt::BOOL); 997 998 // does source == target string? 999 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1000 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1001 1002 Node* if_eq = generate_slow_guard(bol, NULL); 1003 if (if_eq != NULL) { 1004 // receiver == argument 1005 phi->init_req(2, intcon(1)); 1006 region->init_req(2, if_eq); 1007 } 1008 1009 // get String klass for instanceOf 1010 ciInstanceKlass* klass = env()->String_klass(); 1011 1012 if (!stopped()) { 1013 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1014 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1015 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1016 1017 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1018 //instanceOf == true, fallthrough 1019 1020 if (inst_false != NULL) { 1021 phi->init_req(3, intcon(0)); 1022 region->init_req(3, inst_false); 1023 } 1024 } 1025 1026 if (!stopped()) { 1027 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1028 1029 // Properly cast the argument to String 1030 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1031 // This path is taken only when argument's type is String:NotNull. 1032 argument = cast_not_null(argument, false); 1033 1034 Node* no_ctrl = NULL; 1035 1036 // Get start addr of receiver 1037 Node* receiver_val = load_String_value(no_ctrl, receiver); 1038 1039 if (ShenandoahVerifyReadsToFromSpace) { 1040 receiver_val = shenandoah_read_barrier(receiver_val); 1041 } 1042 1043 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1044 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1045 1046 // Get length of receiver 1047 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1048 1049 // Get start addr of argument 1050 Node* argument_val = load_String_value(no_ctrl, argument); 1051 1052 if (ShenandoahVerifyReadsToFromSpace) { 1053 argument_val = shenandoah_read_barrier(argument_val); 1054 } 1055 1056 Node* argument_offset = load_String_offset(no_ctrl, argument); 1057 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1058 1059 // Get length of argument 1060 Node* argument_cnt = load_String_length(no_ctrl, argument); 1061 1062 // Check for receiver count != argument count 1063 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1064 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1065 Node* if_ne = generate_slow_guard(bol, NULL); 1066 if (if_ne != NULL) { 1067 phi->init_req(4, intcon(0)); 1068 region->init_req(4, if_ne); 1069 } 1070 1071 // Check for count == 0 is done by assembler code for StrEquals. 1072 1073 if (!stopped()) { 1074 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1075 phi->init_req(1, equals); 1076 region->init_req(1, control()); 1077 } 1078 } 1079 1080 // post merge 1081 set_control(_gvn.transform(region)); 1082 record_for_igvn(region); 1083 1084 set_result(_gvn.transform(phi)); 1085 return true; 1086 } 1087 1088 //------------------------------inline_array_equals---------------------------- 1089 bool LibraryCallKit::inline_array_equals() { 1090 Node* arg1 = argument(0); 1091 Node* arg2 = argument(1); 1092 1093 arg1 = shenandoah_read_barrier(arg1); 1094 arg2 = shenandoah_read_barrier(arg2); 1095 1096 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1097 return true; 1098 } 1099 1100 // Java version of String.indexOf(constant string) 1101 // class StringDecl { 1102 // StringDecl(char[] ca) { 1103 // offset = 0; 1104 // count = ca.length; 1105 // value = ca; 1106 // } 1107 // int offset; 1108 // int count; 1109 // char[] value; 1110 // } 1111 // 1112 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1113 // int targetOffset, int cache_i, int md2) { 1114 // int cache = cache_i; 1115 // int sourceOffset = string_object.offset; 1116 // int sourceCount = string_object.count; 1117 // int targetCount = target_object.length; 1118 // 1119 // int targetCountLess1 = targetCount - 1; 1120 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1121 // 1122 // char[] source = string_object.value; 1123 // char[] target = target_object; 1124 // int lastChar = target[targetCountLess1]; 1125 // 1126 // outer_loop: 1127 // for (int i = sourceOffset; i < sourceEnd; ) { 1128 // int src = source[i + targetCountLess1]; 1129 // if (src == lastChar) { 1130 // // With random strings and a 4-character alphabet, 1131 // // reverse matching at this point sets up 0.8% fewer 1132 // // frames, but (paradoxically) makes 0.3% more probes. 1133 // // Since those probes are nearer the lastChar probe, 1134 // // there is may be a net D$ win with reverse matching. 1135 // // But, reversing loop inhibits unroll of inner loop 1136 // // for unknown reason. So, does running outer loop from 1137 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1138 // for (int j = 0; j < targetCountLess1; j++) { 1139 // if (target[targetOffset + j] != source[i+j]) { 1140 // if ((cache & (1 << source[i+j])) == 0) { 1141 // if (md2 < j+1) { 1142 // i += j+1; 1143 // continue outer_loop; 1144 // } 1145 // } 1146 // i += md2; 1147 // continue outer_loop; 1148 // } 1149 // } 1150 // return i - sourceOffset; 1151 // } 1152 // if ((cache & (1 << src)) == 0) { 1153 // i += targetCountLess1; 1154 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1155 // i++; 1156 // } 1157 // return -1; 1158 // } 1159 1160 //------------------------------string_indexOf------------------------ 1161 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1162 jint cache_i, jint md2_i) { 1163 1164 Node* no_ctrl = NULL; 1165 float likely = PROB_LIKELY(0.9); 1166 float unlikely = PROB_UNLIKELY(0.9); 1167 1168 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1169 1170 Node* source = load_String_value(no_ctrl, string_object); 1171 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1172 Node* sourceCount = load_String_length(no_ctrl, string_object); 1173 1174 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1175 jint target_length = target_array->length(); 1176 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1177 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1178 1179 // String.value field is known to be @Stable. 1180 if (UseImplicitStableValues) { 1181 target = cast_array_to_stable(target, target_type); 1182 } 1183 1184 IdealKit kit(this, false, true); 1185 #define __ kit. 1186 Node* zero = __ ConI(0); 1187 Node* one = __ ConI(1); 1188 Node* cache = __ ConI(cache_i); 1189 Node* md2 = __ ConI(md2_i); 1190 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1191 Node* targetCountLess1 = __ ConI(target_length - 1); 1192 Node* targetOffset = __ ConI(targetOffset_i); 1193 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1194 1195 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1196 Node* outer_loop = __ make_label(2 /* goto */); 1197 Node* return_ = __ make_label(1); 1198 1199 __ set(rtn,__ ConI(-1)); 1200 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1201 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1202 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1203 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1204 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1205 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1206 Node* tpj = __ AddI(targetOffset, __ value(j)); 1207 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1208 Node* ipj = __ AddI(__ value(i), __ value(j)); 1209 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1210 __ if_then(targ, BoolTest::ne, src2); { 1211 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1212 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1213 __ increment(i, __ AddI(__ value(j), one)); 1214 __ goto_(outer_loop); 1215 } __ end_if(); __ dead(j); 1216 }__ end_if(); __ dead(j); 1217 __ increment(i, md2); 1218 __ goto_(outer_loop); 1219 }__ end_if(); 1220 __ increment(j, one); 1221 }__ end_loop(); __ dead(j); 1222 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1223 __ goto_(return_); 1224 }__ end_if(); 1225 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1226 __ increment(i, targetCountLess1); 1227 }__ end_if(); 1228 __ increment(i, one); 1229 __ bind(outer_loop); 1230 }__ end_loop(); __ dead(i); 1231 __ bind(return_); 1232 1233 // Final sync IdealKit and GraphKit. 1234 final_sync(kit); 1235 Node* result = __ value(rtn); 1236 #undef __ 1237 C->set_has_loops(true); 1238 return result; 1239 } 1240 1241 //------------------------------inline_string_indexOf------------------------ 1242 bool LibraryCallKit::inline_string_indexOf() { 1243 Node* receiver = argument(0); 1244 Node* arg = argument(1); 1245 1246 Node* result; 1247 if (Matcher::has_match_rule(Op_StrIndexOf) && 1248 UseSSE42Intrinsics) { 1249 // Generate SSE4.2 version of indexOf 1250 // We currently only have match rules that use SSE4.2 1251 1252 receiver = null_check(receiver); 1253 arg = null_check(arg); 1254 if (stopped()) { 1255 return true; 1256 } 1257 1258 // Make the merge point 1259 RegionNode* result_rgn = new RegionNode(4); 1260 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1261 Node* no_ctrl = NULL; 1262 1263 // Get start addr of source string 1264 Node* source = load_String_value(no_ctrl, receiver); 1265 Node* source_offset = load_String_offset(no_ctrl, receiver); 1266 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1267 1268 // Get length of source string 1269 Node* source_cnt = load_String_length(no_ctrl, receiver); 1270 1271 // Get start addr of substring 1272 Node* substr = load_String_value(no_ctrl, arg); 1273 Node* substr_offset = load_String_offset(no_ctrl, arg); 1274 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1275 1276 // Get length of source string 1277 Node* substr_cnt = load_String_length(no_ctrl, arg); 1278 1279 // Check for substr count > string count 1280 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1281 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1282 Node* if_gt = generate_slow_guard(bol, NULL); 1283 if (if_gt != NULL) { 1284 result_phi->init_req(2, intcon(-1)); 1285 result_rgn->init_req(2, if_gt); 1286 } 1287 1288 if (!stopped()) { 1289 // Check for substr count == 0 1290 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1291 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1292 Node* if_zero = generate_slow_guard(bol, NULL); 1293 if (if_zero != NULL) { 1294 result_phi->init_req(3, intcon(0)); 1295 result_rgn->init_req(3, if_zero); 1296 } 1297 } 1298 1299 if (!stopped()) { 1300 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1301 result_phi->init_req(1, result); 1302 result_rgn->init_req(1, control()); 1303 } 1304 set_control(_gvn.transform(result_rgn)); 1305 record_for_igvn(result_rgn); 1306 result = _gvn.transform(result_phi); 1307 1308 } else { // Use LibraryCallKit::string_indexOf 1309 // don't intrinsify if argument isn't a constant string. 1310 if (!arg->is_Con()) { 1311 return false; 1312 } 1313 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1314 if (str_type == NULL) { 1315 return false; 1316 } 1317 ciInstanceKlass* klass = env()->String_klass(); 1318 ciObject* str_const = str_type->const_oop(); 1319 if (str_const == NULL || str_const->klass() != klass) { 1320 return false; 1321 } 1322 ciInstance* str = str_const->as_instance(); 1323 assert(str != NULL, "must be instance"); 1324 1325 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1326 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1327 1328 int o; 1329 int c; 1330 if (java_lang_String::has_offset_field()) { 1331 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1332 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1333 } else { 1334 o = 0; 1335 c = pat->length(); 1336 } 1337 1338 // constant strings have no offset and count == length which 1339 // simplifies the resulting code somewhat so lets optimize for that. 1340 if (o != 0 || c != pat->length()) { 1341 return false; 1342 } 1343 1344 receiver = null_check(receiver, T_OBJECT); 1345 // NOTE: No null check on the argument is needed since it's a constant String oop. 1346 if (stopped()) { 1347 return true; 1348 } 1349 1350 // The null string as a pattern always returns 0 (match at beginning of string) 1351 if (c == 0) { 1352 set_result(intcon(0)); 1353 return true; 1354 } 1355 1356 // Generate default indexOf 1357 jchar lastChar = pat->char_at(o + (c - 1)); 1358 int cache = 0; 1359 int i; 1360 for (i = 0; i < c - 1; i++) { 1361 assert(i < pat->length(), "out of range"); 1362 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1363 } 1364 1365 int md2 = c; 1366 for (i = 0; i < c - 1; i++) { 1367 assert(i < pat->length(), "out of range"); 1368 if (pat->char_at(o + i) == lastChar) { 1369 md2 = (c - 1) - i; 1370 } 1371 } 1372 1373 result = string_indexOf(receiver, pat, o, cache, md2); 1374 } 1375 set_result(result); 1376 return true; 1377 } 1378 1379 //--------------------------round_double_node-------------------------------- 1380 // Round a double node if necessary. 1381 Node* LibraryCallKit::round_double_node(Node* n) { 1382 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1383 n = _gvn.transform(new RoundDoubleNode(0, n)); 1384 return n; 1385 } 1386 1387 //------------------------------inline_math----------------------------------- 1388 // public static double Math.abs(double) 1389 // public static double Math.sqrt(double) 1390 // public static double Math.log(double) 1391 // public static double Math.log10(double) 1392 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1393 Node* arg = round_double_node(argument(0)); 1394 Node* n; 1395 switch (id) { 1396 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1397 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1398 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1399 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1400 default: fatal_unexpected_iid(id); break; 1401 } 1402 set_result(_gvn.transform(n)); 1403 return true; 1404 } 1405 1406 //------------------------------inline_trig---------------------------------- 1407 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1408 // argument reduction which will turn into a fast/slow diamond. 1409 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1410 Node* arg = round_double_node(argument(0)); 1411 Node* n = NULL; 1412 1413 switch (id) { 1414 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1415 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1416 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1417 default: fatal_unexpected_iid(id); break; 1418 } 1419 n = _gvn.transform(n); 1420 1421 // Rounding required? Check for argument reduction! 1422 if (Matcher::strict_fp_requires_explicit_rounding) { 1423 static const double pi_4 = 0.7853981633974483; 1424 static const double neg_pi_4 = -0.7853981633974483; 1425 // pi/2 in 80-bit extended precision 1426 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1427 // -pi/2 in 80-bit extended precision 1428 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00}; 1429 // Cutoff value for using this argument reduction technique 1430 //static const double pi_2_minus_epsilon = 1.564660403643354; 1431 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1432 1433 // Pseudocode for sin: 1434 // if (x <= Math.PI / 4.0) { 1435 // if (x >= -Math.PI / 4.0) return fsin(x); 1436 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1437 // } else { 1438 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1439 // } 1440 // return StrictMath.sin(x); 1441 1442 // Pseudocode for cos: 1443 // if (x <= Math.PI / 4.0) { 1444 // if (x >= -Math.PI / 4.0) return fcos(x); 1445 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1446 // } else { 1447 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1448 // } 1449 // return StrictMath.cos(x); 1450 1451 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1452 // requires a special machine instruction to load it. Instead we'll try 1453 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1454 // probably do the math inside the SIN encoding. 1455 1456 // Make the merge point 1457 RegionNode* r = new RegionNode(3); 1458 Node* phi = new PhiNode(r, Type::DOUBLE); 1459 1460 // Flatten arg so we need only 1 test 1461 Node *abs = _gvn.transform(new AbsDNode(arg)); 1462 // Node for PI/4 constant 1463 Node *pi4 = makecon(TypeD::make(pi_4)); 1464 // Check PI/4 : abs(arg) 1465 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1466 // Check: If PI/4 < abs(arg) then go slow 1467 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1468 // Branch either way 1469 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1470 set_control(opt_iff(r,iff)); 1471 1472 // Set fast path result 1473 phi->init_req(2, n); 1474 1475 // Slow path - non-blocking leaf call 1476 Node* call = NULL; 1477 switch (id) { 1478 case vmIntrinsics::_dsin: 1479 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1480 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1481 "Sin", NULL, arg, top()); 1482 break; 1483 case vmIntrinsics::_dcos: 1484 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1485 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1486 "Cos", NULL, arg, top()); 1487 break; 1488 case vmIntrinsics::_dtan: 1489 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1490 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1491 "Tan", NULL, arg, top()); 1492 break; 1493 } 1494 assert(control()->in(0) == call, ""); 1495 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1496 r->init_req(1, control()); 1497 phi->init_req(1, slow_result); 1498 1499 // Post-merge 1500 set_control(_gvn.transform(r)); 1501 record_for_igvn(r); 1502 n = _gvn.transform(phi); 1503 1504 C->set_has_split_ifs(true); // Has chance for split-if optimization 1505 } 1506 set_result(n); 1507 return true; 1508 } 1509 1510 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1511 //------------------- 1512 //result=(result.isNaN())? funcAddr():result; 1513 // Check: If isNaN() by checking result!=result? then either trap 1514 // or go to runtime 1515 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1516 // Build the boolean node 1517 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1518 1519 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1520 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1521 // The pow or exp intrinsic returned a NaN, which requires a call 1522 // to the runtime. Recompile with the runtime call. 1523 uncommon_trap(Deoptimization::Reason_intrinsic, 1524 Deoptimization::Action_make_not_entrant); 1525 } 1526 return result; 1527 } else { 1528 // If this inlining ever returned NaN in the past, we compile a call 1529 // to the runtime to properly handle corner cases 1530 1531 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1532 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1533 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1534 1535 if (!if_slow->is_top()) { 1536 RegionNode* result_region = new RegionNode(3); 1537 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1538 1539 result_region->init_req(1, if_fast); 1540 result_val->init_req(1, result); 1541 1542 set_control(if_slow); 1543 1544 const TypePtr* no_memory_effects = NULL; 1545 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1546 no_memory_effects, 1547 x, top(), y, y ? top() : NULL); 1548 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1549 #ifdef ASSERT 1550 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1551 assert(value_top == top(), "second value must be top"); 1552 #endif 1553 1554 result_region->init_req(2, control()); 1555 result_val->init_req(2, value); 1556 set_control(_gvn.transform(result_region)); 1557 return _gvn.transform(result_val); 1558 } else { 1559 return result; 1560 } 1561 } 1562 } 1563 1564 //------------------------------inline_exp------------------------------------- 1565 // Inline exp instructions, if possible. The Intel hardware only misses 1566 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1567 bool LibraryCallKit::inline_exp() { 1568 Node* arg = round_double_node(argument(0)); 1569 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1570 1571 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1572 set_result(n); 1573 1574 C->set_has_split_ifs(true); // Has chance for split-if optimization 1575 return true; 1576 } 1577 1578 //------------------------------inline_pow------------------------------------- 1579 // Inline power instructions, if possible. 1580 bool LibraryCallKit::inline_pow() { 1581 // Pseudocode for pow 1582 // if (y == 2) { 1583 // return x * x; 1584 // } else { 1585 // if (x <= 0.0) { 1586 // long longy = (long)y; 1587 // if ((double)longy == y) { // if y is long 1588 // if (y + 1 == y) longy = 0; // huge number: even 1589 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1590 // } else { 1591 // result = NaN; 1592 // } 1593 // } else { 1594 // result = DPow(x,y); 1595 // } 1596 // if (result != result)? { 1597 // result = uncommon_trap() or runtime_call(); 1598 // } 1599 // return result; 1600 // } 1601 1602 Node* x = round_double_node(argument(0)); 1603 Node* y = round_double_node(argument(2)); 1604 1605 Node* result = NULL; 1606 1607 Node* const_two_node = makecon(TypeD::make(2.0)); 1608 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1609 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1610 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1611 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1612 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1613 1614 RegionNode* region_node = new RegionNode(3); 1615 region_node->init_req(1, if_true); 1616 1617 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1618 // special case for x^y where y == 2, we can convert it to x * x 1619 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1620 1621 // set control to if_false since we will now process the false branch 1622 set_control(if_false); 1623 1624 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1625 // Short form: skip the fancy tests and just check for NaN result. 1626 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1627 } else { 1628 // If this inlining ever returned NaN in the past, include all 1629 // checks + call to the runtime. 1630 1631 // Set the merge point for If node with condition of (x <= 0.0) 1632 // There are four possible paths to region node and phi node 1633 RegionNode *r = new RegionNode(4); 1634 Node *phi = new PhiNode(r, Type::DOUBLE); 1635 1636 // Build the first if node: if (x <= 0.0) 1637 // Node for 0 constant 1638 Node *zeronode = makecon(TypeD::ZERO); 1639 // Check x:0 1640 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1641 // Check: If (x<=0) then go complex path 1642 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1643 // Branch either way 1644 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1645 // Fast path taken; set region slot 3 1646 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1647 r->init_req(3,fast_taken); // Capture fast-control 1648 1649 // Fast path not-taken, i.e. slow path 1650 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1651 1652 // Set fast path result 1653 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1654 phi->init_req(3, fast_result); 1655 1656 // Complex path 1657 // Build the second if node (if y is long) 1658 // Node for (long)y 1659 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1660 // Node for (double)((long) y) 1661 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1662 // Check (double)((long) y) : y 1663 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1664 // Check if (y isn't long) then go to slow path 1665 1666 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1667 // Branch either way 1668 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1669 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1670 1671 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1672 1673 // Calculate DPow(abs(x), y)*(1 & (long)y) 1674 // Node for constant 1 1675 Node *conone = longcon(1); 1676 // 1& (long)y 1677 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1678 1679 // A huge number is always even. Detect a huge number by checking 1680 // if y + 1 == y and set integer to be tested for parity to 0. 1681 // Required for corner case: 1682 // (long)9.223372036854776E18 = max_jlong 1683 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1684 // max_jlong is odd but 9.223372036854776E18 is even 1685 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1686 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1687 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1688 Node* correctedsign = NULL; 1689 if (ConditionalMoveLimit != 0) { 1690 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1691 } else { 1692 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1693 RegionNode *r = new RegionNode(3); 1694 Node *phi = new PhiNode(r, TypeLong::LONG); 1695 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1696 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1697 phi->init_req(1, signnode); 1698 phi->init_req(2, longcon(0)); 1699 correctedsign = _gvn.transform(phi); 1700 ylong_path = _gvn.transform(r); 1701 record_for_igvn(r); 1702 } 1703 1704 // zero node 1705 Node *conzero = longcon(0); 1706 // Check (1&(long)y)==0? 1707 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1708 // Check if (1&(long)y)!=0?, if so the result is negative 1709 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1710 // abs(x) 1711 Node *absx=_gvn.transform(new AbsDNode(x)); 1712 // abs(x)^y 1713 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1714 // -abs(x)^y 1715 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1716 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1717 Node *signresult = NULL; 1718 if (ConditionalMoveLimit != 0) { 1719 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1720 } else { 1721 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1722 RegionNode *r = new RegionNode(3); 1723 Node *phi = new PhiNode(r, Type::DOUBLE); 1724 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1725 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1726 phi->init_req(1, absxpowy); 1727 phi->init_req(2, negabsxpowy); 1728 signresult = _gvn.transform(phi); 1729 ylong_path = _gvn.transform(r); 1730 record_for_igvn(r); 1731 } 1732 // Set complex path fast result 1733 r->init_req(2, ylong_path); 1734 phi->init_req(2, signresult); 1735 1736 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1737 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1738 r->init_req(1,slow_path); 1739 phi->init_req(1,slow_result); 1740 1741 // Post merge 1742 set_control(_gvn.transform(r)); 1743 record_for_igvn(r); 1744 result = _gvn.transform(phi); 1745 } 1746 1747 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1748 1749 // control from finish_pow_exp is now input to the region node 1750 region_node->set_req(2, control()); 1751 // the result from finish_pow_exp is now input to the phi node 1752 phi_node->init_req(2, result); 1753 set_control(_gvn.transform(region_node)); 1754 record_for_igvn(region_node); 1755 set_result(_gvn.transform(phi_node)); 1756 1757 C->set_has_split_ifs(true); // Has chance for split-if optimization 1758 return true; 1759 } 1760 1761 //------------------------------runtime_math----------------------------- 1762 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1763 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1764 "must be (DD)D or (D)D type"); 1765 1766 // Inputs 1767 Node* a = round_double_node(argument(0)); 1768 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1769 1770 const TypePtr* no_memory_effects = NULL; 1771 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1772 no_memory_effects, 1773 a, top(), b, b ? top() : NULL); 1774 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1775 #ifdef ASSERT 1776 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1777 assert(value_top == top(), "second value must be top"); 1778 #endif 1779 1780 set_result(value); 1781 return true; 1782 } 1783 1784 //------------------------------inline_math_native----------------------------- 1785 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1786 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1787 switch (id) { 1788 // These intrinsics are not properly supported on all hardware 1789 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1790 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1791 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1792 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1793 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1794 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1795 1796 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1797 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1798 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1799 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1800 1801 // These intrinsics are supported on all hardware 1802 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1803 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1804 1805 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1806 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1807 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1808 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1809 #undef FN_PTR 1810 1811 // These intrinsics are not yet correctly implemented 1812 case vmIntrinsics::_datan2: 1813 return false; 1814 1815 default: 1816 fatal_unexpected_iid(id); 1817 return false; 1818 } 1819 } 1820 1821 static bool is_simple_name(Node* n) { 1822 return (n->req() == 1 // constant 1823 || (n->is_Type() && n->as_Type()->type()->singleton()) 1824 || n->is_Proj() // parameter or return value 1825 || n->is_Phi() // local of some sort 1826 ); 1827 } 1828 1829 //----------------------------inline_notify-----------------------------------* 1830 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1831 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1832 address func; 1833 if (id == vmIntrinsics::_notify) { 1834 func = OptoRuntime::monitor_notify_Java(); 1835 } else { 1836 func = OptoRuntime::monitor_notifyAll_Java(); 1837 } 1838 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1839 make_slow_call_ex(call, env()->Throwable_klass(), false); 1840 return true; 1841 } 1842 1843 1844 //----------------------------inline_min_max----------------------------------- 1845 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1846 set_result(generate_min_max(id, argument(0), argument(1))); 1847 return true; 1848 } 1849 1850 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1851 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1852 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1853 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1854 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1855 1856 { 1857 PreserveJVMState pjvms(this); 1858 PreserveReexecuteState preexecs(this); 1859 jvms()->set_should_reexecute(true); 1860 1861 set_control(slow_path); 1862 set_i_o(i_o()); 1863 1864 uncommon_trap(Deoptimization::Reason_intrinsic, 1865 Deoptimization::Action_none); 1866 } 1867 1868 set_control(fast_path); 1869 set_result(math); 1870 } 1871 1872 template <typename OverflowOp> 1873 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1874 typedef typename OverflowOp::MathOp MathOp; 1875 1876 MathOp* mathOp = new MathOp(arg1, arg2); 1877 Node* operation = _gvn.transform( mathOp ); 1878 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1879 inline_math_mathExact(operation, ofcheck); 1880 return true; 1881 } 1882 1883 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1884 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1885 } 1886 1887 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1888 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1889 } 1890 1891 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1892 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1893 } 1894 1895 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1896 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1897 } 1898 1899 bool LibraryCallKit::inline_math_negateExactI() { 1900 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1901 } 1902 1903 bool LibraryCallKit::inline_math_negateExactL() { 1904 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 1905 } 1906 1907 bool LibraryCallKit::inline_math_multiplyExactI() { 1908 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 1909 } 1910 1911 bool LibraryCallKit::inline_math_multiplyExactL() { 1912 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 1913 } 1914 1915 Node* 1916 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1917 // These are the candidate return value: 1918 Node* xvalue = x0; 1919 Node* yvalue = y0; 1920 1921 if (xvalue == yvalue) { 1922 return xvalue; 1923 } 1924 1925 bool want_max = (id == vmIntrinsics::_max); 1926 1927 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1928 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1929 if (txvalue == NULL || tyvalue == NULL) return top(); 1930 // This is not really necessary, but it is consistent with a 1931 // hypothetical MaxINode::Value method: 1932 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1933 1934 // %%% This folding logic should (ideally) be in a different place. 1935 // Some should be inside IfNode, and there to be a more reliable 1936 // transformation of ?: style patterns into cmoves. We also want 1937 // more powerful optimizations around cmove and min/max. 1938 1939 // Try to find a dominating comparison of these guys. 1940 // It can simplify the index computation for Arrays.copyOf 1941 // and similar uses of System.arraycopy. 1942 // First, compute the normalized version of CmpI(x, y). 1943 int cmp_op = Op_CmpI; 1944 Node* xkey = xvalue; 1945 Node* ykey = yvalue; 1946 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 1947 if (ideal_cmpxy->is_Cmp()) { 1948 // E.g., if we have CmpI(length - offset, count), 1949 // it might idealize to CmpI(length, count + offset) 1950 cmp_op = ideal_cmpxy->Opcode(); 1951 xkey = ideal_cmpxy->in(1); 1952 ykey = ideal_cmpxy->in(2); 1953 } 1954 1955 // Start by locating any relevant comparisons. 1956 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1957 Node* cmpxy = NULL; 1958 Node* cmpyx = NULL; 1959 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1960 Node* cmp = start_from->fast_out(k); 1961 if (cmp->outcnt() > 0 && // must have prior uses 1962 cmp->in(0) == NULL && // must be context-independent 1963 cmp->Opcode() == cmp_op) { // right kind of compare 1964 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1965 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1966 } 1967 } 1968 1969 const int NCMPS = 2; 1970 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1971 int cmpn; 1972 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1973 if (cmps[cmpn] != NULL) break; // find a result 1974 } 1975 if (cmpn < NCMPS) { 1976 // Look for a dominating test that tells us the min and max. 1977 int depth = 0; // Limit search depth for speed 1978 Node* dom = control(); 1979 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1980 if (++depth >= 100) break; 1981 Node* ifproj = dom; 1982 if (!ifproj->is_Proj()) continue; 1983 Node* iff = ifproj->in(0); 1984 if (!iff->is_If()) continue; 1985 Node* bol = iff->in(1); 1986 if (!bol->is_Bool()) continue; 1987 Node* cmp = bol->in(1); 1988 if (cmp == NULL) continue; 1989 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1990 if (cmps[cmpn] == cmp) break; 1991 if (cmpn == NCMPS) continue; 1992 BoolTest::mask btest = bol->as_Bool()->_test._test; 1993 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1994 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1995 // At this point, we know that 'x btest y' is true. 1996 switch (btest) { 1997 case BoolTest::eq: 1998 // They are proven equal, so we can collapse the min/max. 1999 // Either value is the answer. Choose the simpler. 2000 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2001 return yvalue; 2002 return xvalue; 2003 case BoolTest::lt: // x < y 2004 case BoolTest::le: // x <= y 2005 return (want_max ? yvalue : xvalue); 2006 case BoolTest::gt: // x > y 2007 case BoolTest::ge: // x >= y 2008 return (want_max ? xvalue : yvalue); 2009 } 2010 } 2011 } 2012 2013 // We failed to find a dominating test. 2014 // Let's pick a test that might GVN with prior tests. 2015 Node* best_bol = NULL; 2016 BoolTest::mask best_btest = BoolTest::illegal; 2017 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2018 Node* cmp = cmps[cmpn]; 2019 if (cmp == NULL) continue; 2020 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2021 Node* bol = cmp->fast_out(j); 2022 if (!bol->is_Bool()) continue; 2023 BoolTest::mask btest = bol->as_Bool()->_test._test; 2024 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2025 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2026 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2027 best_bol = bol->as_Bool(); 2028 best_btest = btest; 2029 } 2030 } 2031 } 2032 2033 Node* answer_if_true = NULL; 2034 Node* answer_if_false = NULL; 2035 switch (best_btest) { 2036 default: 2037 if (cmpxy == NULL) 2038 cmpxy = ideal_cmpxy; 2039 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2040 // and fall through: 2041 case BoolTest::lt: // x < y 2042 case BoolTest::le: // x <= y 2043 answer_if_true = (want_max ? yvalue : xvalue); 2044 answer_if_false = (want_max ? xvalue : yvalue); 2045 break; 2046 case BoolTest::gt: // x > y 2047 case BoolTest::ge: // x >= y 2048 answer_if_true = (want_max ? xvalue : yvalue); 2049 answer_if_false = (want_max ? yvalue : xvalue); 2050 break; 2051 } 2052 2053 jint hi, lo; 2054 if (want_max) { 2055 // We can sharpen the minimum. 2056 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2057 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2058 } else { 2059 // We can sharpen the maximum. 2060 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2061 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2062 } 2063 2064 // Use a flow-free graph structure, to avoid creating excess control edges 2065 // which could hinder other optimizations. 2066 // Since Math.min/max is often used with arraycopy, we want 2067 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2068 Node* cmov = CMoveNode::make(NULL, best_bol, 2069 answer_if_false, answer_if_true, 2070 TypeInt::make(lo, hi, widen)); 2071 2072 return _gvn.transform(cmov); 2073 2074 /* 2075 // This is not as desirable as it may seem, since Min and Max 2076 // nodes do not have a full set of optimizations. 2077 // And they would interfere, anyway, with 'if' optimizations 2078 // and with CMoveI canonical forms. 2079 switch (id) { 2080 case vmIntrinsics::_min: 2081 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2082 case vmIntrinsics::_max: 2083 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2084 default: 2085 ShouldNotReachHere(); 2086 } 2087 */ 2088 } 2089 2090 inline int 2091 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2092 const TypePtr* base_type = TypePtr::NULL_PTR; 2093 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2094 if (base_type == NULL) { 2095 // Unknown type. 2096 return Type::AnyPtr; 2097 } else if (base_type == TypePtr::NULL_PTR) { 2098 // Since this is a NULL+long form, we have to switch to a rawptr. 2099 base = _gvn.transform(new CastX2PNode(offset)); 2100 offset = MakeConX(0); 2101 return Type::RawPtr; 2102 } else if (base_type->base() == Type::RawPtr) { 2103 return Type::RawPtr; 2104 } else if (base_type->isa_oopptr()) { 2105 // Base is never null => always a heap address. 2106 if (base_type->ptr() == TypePtr::NotNull) { 2107 return Type::OopPtr; 2108 } 2109 // Offset is small => always a heap address. 2110 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2111 if (offset_type != NULL && 2112 base_type->offset() == 0 && // (should always be?) 2113 offset_type->_lo >= 0 && 2114 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2115 return Type::OopPtr; 2116 } 2117 // Otherwise, it might either be oop+off or NULL+addr. 2118 return Type::AnyPtr; 2119 } else { 2120 // No information: 2121 return Type::AnyPtr; 2122 } 2123 } 2124 2125 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2126 int kind = classify_unsafe_addr(base, offset); 2127 if (kind == Type::RawPtr) { 2128 return basic_plus_adr(top(), base, offset); 2129 } else { 2130 return basic_plus_adr(base, offset); 2131 } 2132 } 2133 2134 //--------------------------inline_number_methods----------------------------- 2135 // inline int Integer.numberOfLeadingZeros(int) 2136 // inline int Long.numberOfLeadingZeros(long) 2137 // 2138 // inline int Integer.numberOfTrailingZeros(int) 2139 // inline int Long.numberOfTrailingZeros(long) 2140 // 2141 // inline int Integer.bitCount(int) 2142 // inline int Long.bitCount(long) 2143 // 2144 // inline char Character.reverseBytes(char) 2145 // inline short Short.reverseBytes(short) 2146 // inline int Integer.reverseBytes(int) 2147 // inline long Long.reverseBytes(long) 2148 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2149 Node* arg = argument(0); 2150 Node* n; 2151 switch (id) { 2152 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2153 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2154 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2155 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2156 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2157 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2158 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2159 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2160 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2161 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2162 default: fatal_unexpected_iid(id); break; 2163 } 2164 set_result(_gvn.transform(n)); 2165 return true; 2166 } 2167 2168 //----------------------------inline_unsafe_access---------------------------- 2169 2170 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2171 2172 // Helper that guards and inserts a pre-barrier. 2173 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2174 Node* pre_val, bool need_mem_bar) { 2175 // We could be accessing the referent field of a reference object. If so, when G1 2176 // is enabled, we need to log the value in the referent field in an SATB buffer. 2177 // This routine performs some compile time filters and generates suitable 2178 // runtime filters that guard the pre-barrier code. 2179 // Also add memory barrier for non volatile load from the referent field 2180 // to prevent commoning of loads across safepoint. 2181 if (!(UseG1GC || UseShenandoahGC) && !need_mem_bar) 2182 return; 2183 2184 // Some compile time checks. 2185 2186 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2187 const TypeX* otype = offset->find_intptr_t_type(); 2188 if (otype != NULL && otype->is_con() && 2189 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2190 // Constant offset but not the reference_offset so just return 2191 return; 2192 } 2193 2194 // We only need to generate the runtime guards for instances. 2195 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2196 if (btype != NULL) { 2197 if (btype->isa_aryptr()) { 2198 // Array type so nothing to do 2199 return; 2200 } 2201 2202 const TypeInstPtr* itype = btype->isa_instptr(); 2203 if (itype != NULL) { 2204 // Can the klass of base_oop be statically determined to be 2205 // _not_ a sub-class of Reference and _not_ Object? 2206 ciKlass* klass = itype->klass(); 2207 if ( klass->is_loaded() && 2208 !klass->is_subtype_of(env()->Reference_klass()) && 2209 !env()->Object_klass()->is_subtype_of(klass)) { 2210 return; 2211 } 2212 } 2213 } 2214 2215 // The compile time filters did not reject base_oop/offset so 2216 // we need to generate the following runtime filters 2217 // 2218 // if (offset == java_lang_ref_Reference::_reference_offset) { 2219 // if (instance_of(base, java.lang.ref.Reference)) { 2220 // pre_barrier(_, pre_val, ...); 2221 // } 2222 // } 2223 2224 float likely = PROB_LIKELY( 0.999); 2225 float unlikely = PROB_UNLIKELY(0.999); 2226 2227 IdealKit ideal(this); 2228 #define __ ideal. 2229 2230 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2231 2232 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2233 // Update graphKit memory and control from IdealKit. 2234 sync_kit(ideal); 2235 2236 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2237 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2238 2239 // Update IdealKit memory and control from graphKit. 2240 __ sync_kit(this); 2241 2242 Node* one = __ ConI(1); 2243 // is_instof == 0 if base_oop == NULL 2244 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2245 2246 // Update graphKit from IdeakKit. 2247 sync_kit(ideal); 2248 2249 // Use the pre-barrier to record the value in the referent field 2250 pre_barrier(false /* do_load */, 2251 __ ctrl(), 2252 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2253 pre_val /* pre_val */, 2254 T_OBJECT); 2255 if (need_mem_bar) { 2256 // Add memory barrier to prevent commoning reads from this field 2257 // across safepoint since GC can change its value. 2258 insert_mem_bar(Op_MemBarCPUOrder); 2259 } 2260 // Update IdealKit from graphKit. 2261 __ sync_kit(this); 2262 2263 } __ end_if(); // _ref_type != ref_none 2264 } __ end_if(); // offset == referent_offset 2265 2266 // Final sync IdealKit and GraphKit. 2267 final_sync(ideal); 2268 #undef __ 2269 } 2270 2271 2272 // Interpret Unsafe.fieldOffset cookies correctly: 2273 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2274 2275 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2276 // Attempt to infer a sharper value type from the offset and base type. 2277 ciKlass* sharpened_klass = NULL; 2278 2279 // See if it is an instance field, with an object type. 2280 if (alias_type->field() != NULL) { 2281 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2282 if (alias_type->field()->type()->is_klass()) { 2283 sharpened_klass = alias_type->field()->type()->as_klass(); 2284 } 2285 } 2286 2287 // See if it is a narrow oop array. 2288 if (adr_type->isa_aryptr()) { 2289 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2290 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2291 if (elem_type != NULL) { 2292 sharpened_klass = elem_type->klass(); 2293 } 2294 } 2295 } 2296 2297 // The sharpened class might be unloaded if there is no class loader 2298 // contraint in place. 2299 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2300 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2301 2302 #ifndef PRODUCT 2303 if (C->print_intrinsics() || C->print_inlining()) { 2304 tty->print(" from base type: "); adr_type->dump(); 2305 tty->print(" sharpened value: "); tjp->dump(); 2306 } 2307 #endif 2308 // Sharpen the value type. 2309 return tjp; 2310 } 2311 return NULL; 2312 } 2313 2314 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2315 if (callee()->is_static()) return false; // caller must have the capability! 2316 2317 #ifndef PRODUCT 2318 { 2319 ResourceMark rm; 2320 // Check the signatures. 2321 ciSignature* sig = callee()->signature(); 2322 #ifdef ASSERT 2323 if (!is_store) { 2324 // Object getObject(Object base, int/long offset), etc. 2325 BasicType rtype = sig->return_type()->basic_type(); 2326 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2327 rtype = T_ADDRESS; // it is really a C void* 2328 assert(rtype == type, "getter must return the expected value"); 2329 if (!is_native_ptr) { 2330 assert(sig->count() == 2, "oop getter has 2 arguments"); 2331 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2332 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2333 } else { 2334 assert(sig->count() == 1, "native getter has 1 argument"); 2335 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2336 } 2337 } else { 2338 // void putObject(Object base, int/long offset, Object x), etc. 2339 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2340 if (!is_native_ptr) { 2341 assert(sig->count() == 3, "oop putter has 3 arguments"); 2342 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2343 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2344 } else { 2345 assert(sig->count() == 2, "native putter has 2 arguments"); 2346 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2347 } 2348 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2349 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2350 vtype = T_ADDRESS; // it is really a C void* 2351 assert(vtype == type, "putter must accept the expected value"); 2352 } 2353 #endif // ASSERT 2354 } 2355 #endif //PRODUCT 2356 2357 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2358 2359 Node* receiver = argument(0); // type: oop 2360 2361 // Build address expression. 2362 Node* adr; 2363 Node* heap_base_oop = top(); 2364 Node* offset = top(); 2365 Node* val; 2366 2367 if (!is_native_ptr) { 2368 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2369 Node* base = argument(1); // type: oop 2370 if (UseShenandoahGC) { 2371 // Note: if we don't null-check here, we generate a read barrier with a built-in 2372 // null-check. This will later be attempted to be split on the phi, which 2373 // results in a load on a NULL-based address on the null-path, which blows up. 2374 // It will go away when we do late-insertion of read barriers. 2375 base = null_check(base); 2376 } 2377 if (is_store) { 2378 base = shenandoah_write_barrier(base); 2379 } else { 2380 base = shenandoah_read_barrier(base); 2381 } 2382 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2383 offset = argument(2); // type: long 2384 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2385 // to be plain byte offsets, which are also the same as those accepted 2386 // by oopDesc::field_base. 2387 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2388 "fieldOffset must be byte-scaled"); 2389 // 32-bit machines ignore the high half! 2390 offset = ConvL2X(offset); 2391 adr = make_unsafe_address(base, offset); 2392 heap_base_oop = base; 2393 val = is_store ? argument(4) : NULL; 2394 } else { 2395 Node* ptr = argument(1); // type: long 2396 ptr = ConvL2X(ptr); // adjust Java long to machine word 2397 adr = make_unsafe_address(NULL, ptr); 2398 val = is_store ? argument(3) : NULL; 2399 } 2400 2401 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2402 2403 // First guess at the value type. 2404 const Type *value_type = Type::get_const_basic_type(type); 2405 2406 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2407 // there was not enough information to nail it down. 2408 Compile::AliasType* alias_type = C->alias_type(adr_type); 2409 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2410 2411 // We will need memory barriers unless we can determine a unique 2412 // alias category for this reference. (Note: If for some reason 2413 // the barriers get omitted and the unsafe reference begins to "pollute" 2414 // the alias analysis of the rest of the graph, either Compile::can_alias 2415 // or Compile::must_alias will throw a diagnostic assert.) 2416 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2417 2418 // If we are reading the value of the referent field of a Reference 2419 // object (either by using Unsafe directly or through reflection) 2420 // then, if G1 is enabled, we need to record the referent in an 2421 // SATB log buffer using the pre-barrier mechanism. 2422 // Also we need to add memory barrier to prevent commoning reads 2423 // from this field across safepoint since GC can change its value. 2424 bool need_read_barrier = !is_native_ptr && !is_store && 2425 offset != top() && heap_base_oop != top(); 2426 2427 if (!is_store && type == T_OBJECT) { 2428 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2429 if (tjp != NULL) { 2430 value_type = tjp; 2431 } 2432 } 2433 2434 receiver = null_check(receiver); 2435 if (stopped()) { 2436 return true; 2437 } 2438 // Heap pointers get a null-check from the interpreter, 2439 // as a courtesy. However, this is not guaranteed by Unsafe, 2440 // and it is not possible to fully distinguish unintended nulls 2441 // from intended ones in this API. 2442 2443 if (is_volatile) { 2444 // We need to emit leading and trailing CPU membars (see below) in 2445 // addition to memory membars when is_volatile. This is a little 2446 // too strong, but avoids the need to insert per-alias-type 2447 // volatile membars (for stores; compare Parse::do_put_xxx), which 2448 // we cannot do effectively here because we probably only have a 2449 // rough approximation of type. 2450 need_mem_bar = true; 2451 // For Stores, place a memory ordering barrier now. 2452 if (is_store) { 2453 insert_mem_bar(Op_MemBarRelease); 2454 } else { 2455 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2456 insert_mem_bar(Op_MemBarVolatile); 2457 } 2458 } 2459 } 2460 2461 // Memory barrier to prevent normal and 'unsafe' accesses from 2462 // bypassing each other. Happens after null checks, so the 2463 // exception paths do not take memory state from the memory barrier, 2464 // so there's no problems making a strong assert about mixing users 2465 // of safe & unsafe memory. 2466 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2467 2468 if (!is_store) { 2469 Node* p = NULL; 2470 // Try to constant fold a load from a constant field 2471 ciField* field = alias_type->field(); 2472 if (heap_base_oop != top() && 2473 field != NULL && field->is_constant() && field->layout_type() == type) { 2474 // final or stable field 2475 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop); 2476 if (con_type != NULL) { 2477 p = makecon(con_type); 2478 } 2479 } 2480 if (p == NULL) { 2481 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2482 // To be valid, unsafe loads may depend on other conditions than 2483 // the one that guards them: pin the Load node 2484 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile); 2485 // load value 2486 switch (type) { 2487 case T_BOOLEAN: 2488 case T_CHAR: 2489 case T_BYTE: 2490 case T_SHORT: 2491 case T_INT: 2492 case T_LONG: 2493 case T_FLOAT: 2494 case T_DOUBLE: 2495 break; 2496 case T_OBJECT: 2497 if (need_read_barrier) { 2498 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2499 } 2500 break; 2501 case T_ADDRESS: 2502 // Cast to an int type. 2503 p = _gvn.transform(new CastP2XNode(NULL, p)); 2504 p = ConvX2UL(p); 2505 break; 2506 default: 2507 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2508 break; 2509 } 2510 } 2511 // The load node has the control of the preceding MemBarCPUOrder. All 2512 // following nodes will have the control of the MemBarCPUOrder inserted at 2513 // the end of this method. So, pushing the load onto the stack at a later 2514 // point is fine. 2515 set_result(p); 2516 } else { 2517 // place effect of store into memory 2518 switch (type) { 2519 case T_DOUBLE: 2520 val = dstore_rounding(val); 2521 break; 2522 case T_ADDRESS: 2523 // Repackage the long as a pointer. 2524 val = ConvL2X(val); 2525 val = _gvn.transform(new CastX2PNode(val)); 2526 break; 2527 } 2528 2529 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2530 if (type != T_OBJECT ) { 2531 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2532 } else { 2533 val = shenandoah_read_barrier_nomem(val); 2534 // Possibly an oop being stored to Java heap or native memory 2535 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2536 // oop to Java heap. 2537 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2538 } else { 2539 // We can't tell at compile time if we are storing in the Java heap or outside 2540 // of it. So we need to emit code to conditionally do the proper type of 2541 // store. 2542 2543 IdealKit ideal(this); 2544 #define __ ideal. 2545 // QQQ who knows what probability is here?? 2546 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2547 // Sync IdealKit and graphKit. 2548 sync_kit(ideal); 2549 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2550 // Update IdealKit memory. 2551 __ sync_kit(this); 2552 } __ else_(); { 2553 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2554 } __ end_if(); 2555 // Final sync IdealKit and GraphKit. 2556 final_sync(ideal); 2557 #undef __ 2558 } 2559 } 2560 } 2561 2562 if (is_volatile) { 2563 if (!is_store) { 2564 insert_mem_bar(Op_MemBarAcquire); 2565 } else { 2566 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2567 insert_mem_bar(Op_MemBarVolatile); 2568 } 2569 } 2570 } 2571 2572 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2573 2574 return true; 2575 } 2576 2577 //----------------------------inline_unsafe_load_store---------------------------- 2578 // This method serves a couple of different customers (depending on LoadStoreKind): 2579 // 2580 // LS_cmpxchg: 2581 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2582 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2583 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2584 // 2585 // LS_xadd: 2586 // public int getAndAddInt( Object o, long offset, int delta) 2587 // public long getAndAddLong(Object o, long offset, long delta) 2588 // 2589 // LS_xchg: 2590 // int getAndSet(Object o, long offset, int newValue) 2591 // long getAndSet(Object o, long offset, long newValue) 2592 // Object getAndSet(Object o, long offset, Object newValue) 2593 // 2594 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2595 // This basic scheme here is the same as inline_unsafe_access, but 2596 // differs in enough details that combining them would make the code 2597 // overly confusing. (This is a true fact! I originally combined 2598 // them, but even I was confused by it!) As much code/comments as 2599 // possible are retained from inline_unsafe_access though to make 2600 // the correspondences clearer. - dl 2601 2602 if (callee()->is_static()) return false; // caller must have the capability! 2603 2604 #ifndef PRODUCT 2605 BasicType rtype; 2606 { 2607 ResourceMark rm; 2608 // Check the signatures. 2609 ciSignature* sig = callee()->signature(); 2610 rtype = sig->return_type()->basic_type(); 2611 if (kind == LS_xadd || kind == LS_xchg) { 2612 // Check the signatures. 2613 #ifdef ASSERT 2614 assert(rtype == type, "get and set must return the expected type"); 2615 assert(sig->count() == 3, "get and set has 3 arguments"); 2616 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2617 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2618 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2619 #endif // ASSERT 2620 } else if (kind == LS_cmpxchg) { 2621 // Check the signatures. 2622 #ifdef ASSERT 2623 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2624 assert(sig->count() == 4, "CAS has 4 arguments"); 2625 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2626 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2627 #endif // ASSERT 2628 } else { 2629 ShouldNotReachHere(); 2630 } 2631 } 2632 #endif //PRODUCT 2633 2634 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2635 2636 // Get arguments: 2637 Node* receiver = NULL; 2638 Node* base = NULL; 2639 Node* offset = NULL; 2640 Node* oldval = NULL; 2641 Node* newval = NULL; 2642 if (kind == LS_cmpxchg) { 2643 const bool two_slot_type = type2size[type] == 2; 2644 receiver = argument(0); // type: oop 2645 base = argument(1); // type: oop 2646 offset = argument(2); // type: long 2647 oldval = argument(4); // type: oop, int, or long 2648 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2649 } else if (kind == LS_xadd || kind == LS_xchg){ 2650 receiver = argument(0); // type: oop 2651 base = argument(1); // type: oop 2652 offset = argument(2); // type: long 2653 oldval = NULL; 2654 newval = argument(4); // type: oop, int, or long 2655 } 2656 2657 // Null check receiver. 2658 receiver = null_check(receiver); 2659 if (stopped()) { 2660 return true; 2661 } 2662 2663 base = shenandoah_write_barrier(base); 2664 2665 // Build field offset expression. 2666 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2667 // to be plain byte offsets, which are also the same as those accepted 2668 // by oopDesc::field_base. 2669 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2670 // 32-bit machines ignore the high half of long offsets 2671 offset = ConvL2X(offset); 2672 Node* adr = make_unsafe_address(base, offset); 2673 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2674 2675 // For CAS, unlike inline_unsafe_access, there seems no point in 2676 // trying to refine types. Just use the coarse types here. 2677 const Type *value_type = Type::get_const_basic_type(type); 2678 Compile::AliasType* alias_type = C->alias_type(adr_type); 2679 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2680 2681 if (kind == LS_xchg && type == T_OBJECT) { 2682 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2683 if (tjp != NULL) { 2684 value_type = tjp; 2685 } 2686 } 2687 2688 int alias_idx = C->get_alias_index(adr_type); 2689 2690 // Memory-model-wise, a LoadStore acts like a little synchronized 2691 // block, so needs barriers on each side. These don't translate 2692 // into actual barriers on most machines, but we still need rest of 2693 // compiler to respect ordering. 2694 2695 insert_mem_bar(Op_MemBarRelease); 2696 insert_mem_bar(Op_MemBarCPUOrder); 2697 2698 // 4984716: MemBars must be inserted before this 2699 // memory node in order to avoid a false 2700 // dependency which will confuse the scheduler. 2701 Node *mem = memory(alias_idx); 2702 2703 // For now, we handle only those cases that actually exist: ints, 2704 // longs, and Object. Adding others should be straightforward. 2705 Node* load_store; 2706 Node* result; 2707 switch(type) { 2708 case T_INT: 2709 if (kind == LS_xadd) { 2710 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2711 } else if (kind == LS_xchg) { 2712 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2713 } else if (kind == LS_cmpxchg) { 2714 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2715 } else { 2716 ShouldNotReachHere(); 2717 } 2718 result = load_store; 2719 break; 2720 case T_LONG: 2721 if (kind == LS_xadd) { 2722 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2723 } else if (kind == LS_xchg) { 2724 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2725 } else if (kind == LS_cmpxchg) { 2726 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2727 } else { 2728 ShouldNotReachHere(); 2729 } 2730 result = load_store; 2731 break; 2732 case T_OBJECT: 2733 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2734 // could be delayed during Parse (for example, in adjust_map_after_if()). 2735 // Execute transformation here to avoid barrier generation in such case. 2736 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2737 newval = _gvn.makecon(TypePtr::NULL_PTR); 2738 2739 newval = shenandoah_read_barrier_nomem(newval); 2740 2741 // Reference stores need a store barrier. 2742 if (kind == LS_xchg) { 2743 // If pre-barrier must execute before the oop store, old value will require do_load here. 2744 if (!can_move_pre_barrier()) { 2745 pre_barrier(true /* do_load*/, 2746 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2747 NULL /* pre_val*/, 2748 T_OBJECT); 2749 } // Else move pre_barrier to use load_store value, see below. 2750 } else if (kind == LS_cmpxchg) { 2751 // Same as for newval above: 2752 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2753 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2754 } 2755 // The only known value which might get overwritten is oldval. 2756 pre_barrier(false /* do_load */, 2757 control(), NULL, NULL, max_juint, NULL, NULL, 2758 oldval /* pre_val */, 2759 T_OBJECT); 2760 } else { 2761 ShouldNotReachHere(); 2762 } 2763 2764 #ifdef _LP64 2765 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2766 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2767 if (kind == LS_xchg) { 2768 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2769 newval_enc, adr_type, value_type->make_narrowoop())); 2770 } else { 2771 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2772 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2773 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2774 newval_enc, oldval_enc)); 2775 } 2776 result = load_store; 2777 } else 2778 #endif 2779 { 2780 if (kind == LS_xchg) { 2781 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2782 result = load_store; 2783 } else { 2784 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2785 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2786 result = load_store; 2787 2788 if (UseShenandoahGC) { 2789 // if (! success) 2790 Node* cmp_true = _gvn.transform(new CmpINode(load_store, intcon(1))); 2791 Node* tst_true = _gvn.transform(new BoolNode(cmp_true, BoolTest::eq)); 2792 IfNode* iff = create_and_map_if(control(), tst_true, PROB_LIKELY_MAG(2), COUNT_UNKNOWN); 2793 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 2794 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 2795 2796 enum { _success_path = 1, _fail_path, _shenandoah_path, PATH_LIMIT }; 2797 RegionNode* region = new RegionNode(PATH_LIMIT); 2798 Node* phi = new PhiNode(region, TypeInt::BOOL); 2799 // success -> return result of CAS1. 2800 region->init_req(_success_path, iftrue); 2801 phi ->init_req(_success_path, load_store); 2802 2803 // failure 2804 set_control(iffalse); 2805 2806 // if (read_barrier(expected) == read_barrier(old) 2807 oldval = shenandoah_read_barrier(oldval); 2808 2809 // Load old value from memory. We shuold really use what we get back from the CAS, 2810 // if we can. 2811 Node* current = make_load(control(), adr, TypeInstPtr::BOTTOM, type, MemNode::unordered); 2812 // read_barrier(old) 2813 Node* new_current = shenandoah_read_barrier(current); 2814 2815 Node* chk = _gvn.transform(new CmpPNode(new_current, oldval)); 2816 Node* test = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 2817 2818 IfNode* iff2 = create_and_map_if(control(), test, PROB_UNLIKELY_MAG(2), COUNT_UNKNOWN); 2819 Node* iftrue2 = _gvn.transform(new IfTrueNode(iff2)); 2820 Node* iffalse2 = _gvn.transform(new IfFalseNode(iff2)); 2821 2822 // If they are not equal, it's a legitimate failure and we return the result of CAS1. 2823 region->init_req(_fail_path, iffalse2); 2824 phi ->init_req(_fail_path, load_store); 2825 2826 // Otherwise we retry with old. 2827 set_control(iftrue2); 2828 2829 Node *call = make_runtime_call(RC_LEAF | RC_NO_IO, 2830 OptoRuntime::shenandoah_cas_obj_Type(), 2831 CAST_FROM_FN_PTR(address, ShenandoahRuntime::compare_and_swap_object), 2832 "shenandoah_cas_obj", 2833 NULL, 2834 adr, newval, current); 2835 2836 Node* retval = _gvn.transform(new ProjNode(call, TypeFunc::Parms + 0)); 2837 2838 region->init_req(_shenandoah_path, control()); 2839 phi ->init_req(_shenandoah_path, retval); 2840 2841 set_control(_gvn.transform(region)); 2842 record_for_igvn(region); 2843 phi = _gvn.transform(phi); 2844 result = phi; 2845 } 2846 2847 } 2848 } 2849 if (kind == LS_cmpxchg) { 2850 // Emit the post barrier only when the actual store happened. 2851 // This makes sense to check only for compareAndSet that can fail to set the value. 2852 // CAS success path is marked more likely since we anticipate this is a performance 2853 // critical path, while CAS failure path can use the penalty for going through unlikely 2854 // path as backoff. Which is still better than doing a store barrier there. 2855 IdealKit ideal(this); 2856 ideal.if_then(result, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); { 2857 sync_kit(ideal); 2858 post_barrier(ideal.ctrl(), result, base, adr, alias_idx, newval, T_OBJECT, true); 2859 ideal.sync_kit(this); 2860 } ideal.end_if(); 2861 final_sync(ideal); 2862 } else { 2863 post_barrier(control(), result, base, adr, alias_idx, newval, T_OBJECT, true); 2864 } 2865 break; 2866 default: 2867 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2868 break; 2869 } 2870 2871 // SCMemProjNodes represent the memory state of a LoadStore. Their 2872 // main role is to prevent LoadStore nodes from being optimized away 2873 // when their results aren't used. 2874 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2875 set_memory(proj, alias_idx); 2876 2877 if (type == T_OBJECT && kind == LS_xchg) { 2878 #ifdef _LP64 2879 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2880 result = _gvn.transform(new DecodeNNode(result, result->get_ptr_type())); 2881 } 2882 #endif 2883 if (can_move_pre_barrier()) { 2884 // Don't need to load pre_val. The old value is returned by load_store. 2885 // The pre_barrier can execute after the xchg as long as no safepoint 2886 // gets inserted between them. 2887 pre_barrier(false /* do_load */, 2888 control(), NULL, NULL, max_juint, NULL, NULL, 2889 result /* pre_val */, 2890 T_OBJECT); 2891 } 2892 } 2893 2894 // Add the trailing membar surrounding the access 2895 insert_mem_bar(Op_MemBarCPUOrder); 2896 insert_mem_bar(Op_MemBarAcquire); 2897 2898 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2899 set_result(result); 2900 return true; 2901 } 2902 2903 //----------------------------inline_unsafe_ordered_store---------------------- 2904 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2905 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 2906 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 2907 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2908 // This is another variant of inline_unsafe_access, differing in 2909 // that it always issues store-store ("release") barrier and ensures 2910 // store-atomicity (which only matters for "long"). 2911 2912 if (callee()->is_static()) return false; // caller must have the capability! 2913 2914 #ifndef PRODUCT 2915 { 2916 ResourceMark rm; 2917 // Check the signatures. 2918 ciSignature* sig = callee()->signature(); 2919 #ifdef ASSERT 2920 BasicType rtype = sig->return_type()->basic_type(); 2921 assert(rtype == T_VOID, "must return void"); 2922 assert(sig->count() == 3, "has 3 arguments"); 2923 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2924 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2925 #endif // ASSERT 2926 } 2927 #endif //PRODUCT 2928 2929 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2930 2931 // Get arguments: 2932 Node* receiver = argument(0); // type: oop 2933 Node* base = argument(1); // type: oop 2934 Node* offset = argument(2); // type: long 2935 Node* val = argument(4); // type: oop, int, or long 2936 2937 // Null check receiver. 2938 receiver = null_check(receiver); 2939 if (stopped()) { 2940 return true; 2941 } 2942 2943 base = shenandoah_write_barrier(base); 2944 2945 // Build field offset expression. 2946 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2947 // 32-bit machines ignore the high half of long offsets 2948 offset = ConvL2X(offset); 2949 Node* adr = make_unsafe_address(base, offset); 2950 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2951 const Type *value_type = Type::get_const_basic_type(type); 2952 Compile::AliasType* alias_type = C->alias_type(adr_type); 2953 2954 insert_mem_bar(Op_MemBarRelease); 2955 insert_mem_bar(Op_MemBarCPUOrder); 2956 // Ensure that the store is atomic for longs: 2957 const bool require_atomic_access = true; 2958 Node* store; 2959 if (type == T_OBJECT) { // reference stores need a store barrier. 2960 val = shenandoah_read_barrier_nomem(val); 2961 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 2962 } 2963 else { 2964 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 2965 } 2966 insert_mem_bar(Op_MemBarCPUOrder); 2967 return true; 2968 } 2969 2970 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2971 // Regardless of form, don't allow previous ld/st to move down, 2972 // then issue acquire, release, or volatile mem_bar. 2973 insert_mem_bar(Op_MemBarCPUOrder); 2974 switch(id) { 2975 case vmIntrinsics::_loadFence: 2976 insert_mem_bar(Op_LoadFence); 2977 return true; 2978 case vmIntrinsics::_storeFence: 2979 insert_mem_bar(Op_StoreFence); 2980 return true; 2981 case vmIntrinsics::_fullFence: 2982 insert_mem_bar(Op_MemBarVolatile); 2983 return true; 2984 default: 2985 fatal_unexpected_iid(id); 2986 return false; 2987 } 2988 } 2989 2990 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 2991 if (!kls->is_Con()) { 2992 return true; 2993 } 2994 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 2995 if (klsptr == NULL) { 2996 return true; 2997 } 2998 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 2999 // don't need a guard for a klass that is already initialized 3000 return !ik->is_initialized(); 3001 } 3002 3003 //----------------------------inline_unsafe_allocate--------------------------- 3004 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3005 bool LibraryCallKit::inline_unsafe_allocate() { 3006 if (callee()->is_static()) return false; // caller must have the capability! 3007 3008 null_check_receiver(); // null-check, then ignore 3009 Node* cls = null_check(argument(1)); 3010 if (stopped()) return true; 3011 3012 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3013 kls = null_check(kls); 3014 if (stopped()) return true; // argument was like int.class 3015 3016 Node* test = NULL; 3017 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3018 // Note: The argument might still be an illegal value like 3019 // Serializable.class or Object[].class. The runtime will handle it. 3020 // But we must make an explicit check for initialization. 3021 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3022 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3023 // can generate code to load it as unsigned byte. 3024 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3025 Node* bits = intcon(InstanceKlass::fully_initialized); 3026 test = _gvn.transform(new SubINode(inst, bits)); 3027 // The 'test' is non-zero if we need to take a slow path. 3028 } 3029 3030 Node* obj = new_instance(kls, test); 3031 set_result(obj); 3032 return true; 3033 } 3034 3035 #ifdef TRACE_HAVE_INTRINSICS 3036 /* 3037 * oop -> myklass 3038 * myklass->trace_id |= USED 3039 * return myklass->trace_id & ~0x3 3040 */ 3041 bool LibraryCallKit::inline_native_classID() { 3042 null_check_receiver(); // null-check, then ignore 3043 Node* cls = null_check(argument(1), T_OBJECT); 3044 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3045 kls = null_check(kls, T_OBJECT); 3046 ByteSize offset = TRACE_ID_OFFSET; 3047 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3048 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3049 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3050 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3051 Node* clsused = longcon(0x01l); // set the class bit 3052 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3053 3054 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3055 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3056 set_result(andl); 3057 return true; 3058 } 3059 3060 bool LibraryCallKit::inline_native_threadID() { 3061 Node* tls_ptr = NULL; 3062 Node* cur_thr = generate_current_thread(tls_ptr); 3063 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3064 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3065 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3066 3067 Node* threadid = NULL; 3068 size_t thread_id_size = OSThread::thread_id_size(); 3069 if (thread_id_size == (size_t) BytesPerLong) { 3070 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3071 } else if (thread_id_size == (size_t) BytesPerInt) { 3072 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3073 } else { 3074 ShouldNotReachHere(); 3075 } 3076 set_result(threadid); 3077 return true; 3078 } 3079 #endif 3080 3081 //------------------------inline_native_time_funcs-------------- 3082 // inline code for System.currentTimeMillis() and System.nanoTime() 3083 // these have the same type and signature 3084 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3085 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3086 const TypePtr* no_memory_effects = NULL; 3087 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3088 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3089 #ifdef ASSERT 3090 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3091 assert(value_top == top(), "second value must be top"); 3092 #endif 3093 set_result(value); 3094 return true; 3095 } 3096 3097 //------------------------inline_native_currentThread------------------ 3098 bool LibraryCallKit::inline_native_currentThread() { 3099 Node* junk = NULL; 3100 set_result(generate_current_thread(junk)); 3101 return true; 3102 } 3103 3104 //------------------------inline_native_isInterrupted------------------ 3105 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3106 bool LibraryCallKit::inline_native_isInterrupted() { 3107 // Add a fast path to t.isInterrupted(clear_int): 3108 // (t == Thread.current() && 3109 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3110 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3111 // So, in the common case that the interrupt bit is false, 3112 // we avoid making a call into the VM. Even if the interrupt bit 3113 // is true, if the clear_int argument is false, we avoid the VM call. 3114 // However, if the receiver is not currentThread, we must call the VM, 3115 // because there must be some locking done around the operation. 3116 3117 // We only go to the fast case code if we pass two guards. 3118 // Paths which do not pass are accumulated in the slow_region. 3119 3120 enum { 3121 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3122 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3123 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3124 PATH_LIMIT 3125 }; 3126 3127 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3128 // out of the function. 3129 insert_mem_bar(Op_MemBarCPUOrder); 3130 3131 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3132 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3133 3134 RegionNode* slow_region = new RegionNode(1); 3135 record_for_igvn(slow_region); 3136 3137 // (a) Receiving thread must be the current thread. 3138 Node* rec_thr = argument(0); 3139 Node* tls_ptr = NULL; 3140 Node* cur_thr = generate_current_thread(tls_ptr); 3141 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3142 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3143 3144 generate_slow_guard(bol_thr, slow_region); 3145 3146 // (b) Interrupt bit on TLS must be false. 3147 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3148 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3149 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3150 3151 // Set the control input on the field _interrupted read to prevent it floating up. 3152 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3153 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3154 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3155 3156 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3157 3158 // First fast path: if (!TLS._interrupted) return false; 3159 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3160 result_rgn->init_req(no_int_result_path, false_bit); 3161 result_val->init_req(no_int_result_path, intcon(0)); 3162 3163 // drop through to next case 3164 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3165 3166 #ifndef TARGET_OS_FAMILY_windows 3167 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3168 Node* clr_arg = argument(1); 3169 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3170 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3171 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3172 3173 // Second fast path: ... else if (!clear_int) return true; 3174 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3175 result_rgn->init_req(no_clear_result_path, false_arg); 3176 result_val->init_req(no_clear_result_path, intcon(1)); 3177 3178 // drop through to next case 3179 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3180 #else 3181 // To return true on Windows you must read the _interrupted field 3182 // and check the the event state i.e. take the slow path. 3183 #endif // TARGET_OS_FAMILY_windows 3184 3185 // (d) Otherwise, go to the slow path. 3186 slow_region->add_req(control()); 3187 set_control( _gvn.transform(slow_region)); 3188 3189 if (stopped()) { 3190 // There is no slow path. 3191 result_rgn->init_req(slow_result_path, top()); 3192 result_val->init_req(slow_result_path, top()); 3193 } else { 3194 // non-virtual because it is a private non-static 3195 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3196 3197 Node* slow_val = set_results_for_java_call(slow_call); 3198 // this->control() comes from set_results_for_java_call 3199 3200 Node* fast_io = slow_call->in(TypeFunc::I_O); 3201 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3202 3203 // These two phis are pre-filled with copies of of the fast IO and Memory 3204 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3205 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3206 3207 result_rgn->init_req(slow_result_path, control()); 3208 result_io ->init_req(slow_result_path, i_o()); 3209 result_mem->init_req(slow_result_path, reset_memory()); 3210 result_val->init_req(slow_result_path, slow_val); 3211 3212 set_all_memory(_gvn.transform(result_mem)); 3213 set_i_o( _gvn.transform(result_io)); 3214 } 3215 3216 C->set_has_split_ifs(true); // Has chance for split-if optimization 3217 set_result(result_rgn, result_val); 3218 return true; 3219 } 3220 3221 //---------------------------load_mirror_from_klass---------------------------- 3222 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3223 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3224 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3225 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3226 } 3227 3228 //-----------------------load_klass_from_mirror_common------------------------- 3229 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3230 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3231 // and branch to the given path on the region. 3232 // If never_see_null, take an uncommon trap on null, so we can optimistically 3233 // compile for the non-null case. 3234 // If the region is NULL, force never_see_null = true. 3235 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3236 bool never_see_null, 3237 RegionNode* region, 3238 int null_path, 3239 int offset) { 3240 if (region == NULL) never_see_null = true; 3241 Node* p = basic_plus_adr(mirror, offset); 3242 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3243 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3244 Node* null_ctl = top(); 3245 kls = null_check_oop(kls, &null_ctl, never_see_null); 3246 if (region != NULL) { 3247 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3248 region->init_req(null_path, null_ctl); 3249 } else { 3250 assert(null_ctl == top(), "no loose ends"); 3251 } 3252 return kls; 3253 } 3254 3255 //--------------------(inline_native_Class_query helpers)--------------------- 3256 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3257 // Fall through if (mods & mask) == bits, take the guard otherwise. 3258 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3259 // Branch around if the given klass has the given modifier bit set. 3260 // Like generate_guard, adds a new path onto the region. 3261 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3262 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3263 Node* mask = intcon(modifier_mask); 3264 Node* bits = intcon(modifier_bits); 3265 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3266 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3267 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3268 return generate_fair_guard(bol, region); 3269 } 3270 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3271 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3272 } 3273 3274 //-------------------------inline_native_Class_query------------------- 3275 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3276 const Type* return_type = TypeInt::BOOL; 3277 Node* prim_return_value = top(); // what happens if it's a primitive class? 3278 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3279 bool expect_prim = false; // most of these guys expect to work on refs 3280 3281 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3282 3283 Node* mirror = argument(0); 3284 3285 if (ShenandoahVerifyReadsToFromSpace) { 3286 mirror = shenandoah_read_barrier(mirror); 3287 } 3288 3289 Node* obj = top(); 3290 3291 switch (id) { 3292 case vmIntrinsics::_isInstance: 3293 // nothing is an instance of a primitive type 3294 prim_return_value = intcon(0); 3295 obj = argument(1); 3296 if (ShenandoahVerifyReadsToFromSpace) { 3297 obj = shenandoah_read_barrier(obj); 3298 } 3299 break; 3300 case vmIntrinsics::_getModifiers: 3301 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3302 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3303 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3304 break; 3305 case vmIntrinsics::_isInterface: 3306 prim_return_value = intcon(0); 3307 break; 3308 case vmIntrinsics::_isArray: 3309 prim_return_value = intcon(0); 3310 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3311 break; 3312 case vmIntrinsics::_isPrimitive: 3313 prim_return_value = intcon(1); 3314 expect_prim = true; // obviously 3315 break; 3316 case vmIntrinsics::_getSuperclass: 3317 prim_return_value = null(); 3318 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3319 break; 3320 case vmIntrinsics::_getClassAccessFlags: 3321 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3322 return_type = TypeInt::INT; // not bool! 6297094 3323 break; 3324 default: 3325 fatal_unexpected_iid(id); 3326 break; 3327 } 3328 3329 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3330 if (mirror_con == NULL) return false; // cannot happen? 3331 3332 #ifndef PRODUCT 3333 if (C->print_intrinsics() || C->print_inlining()) { 3334 ciType* k = mirror_con->java_mirror_type(); 3335 if (k) { 3336 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3337 k->print_name(); 3338 tty->cr(); 3339 } 3340 } 3341 #endif 3342 3343 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3344 RegionNode* region = new RegionNode(PATH_LIMIT); 3345 record_for_igvn(region); 3346 PhiNode* phi = new PhiNode(region, return_type); 3347 3348 // The mirror will never be null of Reflection.getClassAccessFlags, however 3349 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3350 // if it is. See bug 4774291. 3351 3352 // For Reflection.getClassAccessFlags(), the null check occurs in 3353 // the wrong place; see inline_unsafe_access(), above, for a similar 3354 // situation. 3355 mirror = null_check(mirror); 3356 // If mirror or obj is dead, only null-path is taken. 3357 if (stopped()) return true; 3358 3359 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3360 3361 // Now load the mirror's klass metaobject, and null-check it. 3362 // Side-effects region with the control path if the klass is null. 3363 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3364 // If kls is null, we have a primitive mirror. 3365 phi->init_req(_prim_path, prim_return_value); 3366 if (stopped()) { set_result(region, phi); return true; } 3367 bool safe_for_replace = (region->in(_prim_path) == top()); 3368 3369 Node* p; // handy temp 3370 Node* null_ctl; 3371 3372 // Now that we have the non-null klass, we can perform the real query. 3373 // For constant classes, the query will constant-fold in LoadNode::Value. 3374 Node* query_value = top(); 3375 switch (id) { 3376 case vmIntrinsics::_isInstance: 3377 // nothing is an instance of a primitive type 3378 query_value = gen_instanceof(obj, kls, safe_for_replace); 3379 break; 3380 3381 case vmIntrinsics::_getModifiers: 3382 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3383 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3384 break; 3385 3386 case vmIntrinsics::_isInterface: 3387 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3388 if (generate_interface_guard(kls, region) != NULL) 3389 // A guard was added. If the guard is taken, it was an interface. 3390 phi->add_req(intcon(1)); 3391 // If we fall through, it's a plain class. 3392 query_value = intcon(0); 3393 break; 3394 3395 case vmIntrinsics::_isArray: 3396 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3397 if (generate_array_guard(kls, region) != NULL) 3398 // A guard was added. If the guard is taken, it was an array. 3399 phi->add_req(intcon(1)); 3400 // If we fall through, it's a plain class. 3401 query_value = intcon(0); 3402 break; 3403 3404 case vmIntrinsics::_isPrimitive: 3405 query_value = intcon(0); // "normal" path produces false 3406 break; 3407 3408 case vmIntrinsics::_getSuperclass: 3409 // The rules here are somewhat unfortunate, but we can still do better 3410 // with random logic than with a JNI call. 3411 // Interfaces store null or Object as _super, but must report null. 3412 // Arrays store an intermediate super as _super, but must report Object. 3413 // Other types can report the actual _super. 3414 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3415 if (generate_interface_guard(kls, region) != NULL) 3416 // A guard was added. If the guard is taken, it was an interface. 3417 phi->add_req(null()); 3418 if (generate_array_guard(kls, region) != NULL) 3419 // A guard was added. If the guard is taken, it was an array. 3420 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3421 // If we fall through, it's a plain class. Get its _super. 3422 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3423 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3424 null_ctl = top(); 3425 kls = null_check_oop(kls, &null_ctl); 3426 if (null_ctl != top()) { 3427 // If the guard is taken, Object.superClass is null (both klass and mirror). 3428 region->add_req(null_ctl); 3429 phi ->add_req(null()); 3430 } 3431 if (!stopped()) { 3432 query_value = load_mirror_from_klass(kls); 3433 } 3434 break; 3435 3436 case vmIntrinsics::_getClassAccessFlags: 3437 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3438 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3439 break; 3440 3441 default: 3442 fatal_unexpected_iid(id); 3443 break; 3444 } 3445 3446 // Fall-through is the normal case of a query to a real class. 3447 phi->init_req(1, query_value); 3448 region->init_req(1, control()); 3449 3450 C->set_has_split_ifs(true); // Has chance for split-if optimization 3451 set_result(region, phi); 3452 return true; 3453 } 3454 3455 //-------------------------inline_Class_cast------------------- 3456 bool LibraryCallKit::inline_Class_cast() { 3457 Node* mirror = argument(0); // Class 3458 Node* obj = argument(1); 3459 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3460 if (mirror_con == NULL) { 3461 return false; // dead path (mirror->is_top()). 3462 } 3463 if (obj == NULL || obj->is_top()) { 3464 return false; // dead path 3465 } 3466 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3467 3468 // First, see if Class.cast() can be folded statically. 3469 // java_mirror_type() returns non-null for compile-time Class constants. 3470 ciType* tm = mirror_con->java_mirror_type(); 3471 if (tm != NULL && tm->is_klass() && 3472 tp != NULL && tp->klass() != NULL) { 3473 if (!tp->klass()->is_loaded()) { 3474 // Don't use intrinsic when class is not loaded. 3475 return false; 3476 } else { 3477 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3478 if (static_res == Compile::SSC_always_true) { 3479 // isInstance() is true - fold the code. 3480 set_result(obj); 3481 return true; 3482 } else if (static_res == Compile::SSC_always_false) { 3483 // Don't use intrinsic, have to throw ClassCastException. 3484 // If the reference is null, the non-intrinsic bytecode will 3485 // be optimized appropriately. 3486 return false; 3487 } 3488 } 3489 } 3490 3491 // Bailout intrinsic and do normal inlining if exception path is frequent. 3492 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3493 return false; 3494 } 3495 3496 // Generate dynamic checks. 3497 // Class.cast() is java implementation of _checkcast bytecode. 3498 // Do checkcast (Parse::do_checkcast()) optimizations here. 3499 3500 mirror = null_check(mirror); 3501 // If mirror is dead, only null-path is taken. 3502 if (stopped()) { 3503 return true; 3504 } 3505 3506 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3507 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3508 RegionNode* region = new RegionNode(PATH_LIMIT); 3509 record_for_igvn(region); 3510 3511 // Now load the mirror's klass metaobject, and null-check it. 3512 // If kls is null, we have a primitive mirror and 3513 // nothing is an instance of a primitive type. 3514 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3515 3516 Node* res = top(); 3517 if (!stopped()) { 3518 Node* bad_type_ctrl = top(); 3519 // Do checkcast optimizations. 3520 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3521 region->init_req(_bad_type_path, bad_type_ctrl); 3522 } 3523 if (region->in(_prim_path) != top() || 3524 region->in(_bad_type_path) != top()) { 3525 // Let Interpreter throw ClassCastException. 3526 PreserveJVMState pjvms(this); 3527 set_control(_gvn.transform(region)); 3528 uncommon_trap(Deoptimization::Reason_intrinsic, 3529 Deoptimization::Action_maybe_recompile); 3530 } 3531 if (!stopped()) { 3532 set_result(res); 3533 } 3534 return true; 3535 } 3536 3537 3538 //--------------------------inline_native_subtype_check------------------------ 3539 // This intrinsic takes the JNI calls out of the heart of 3540 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3541 bool LibraryCallKit::inline_native_subtype_check() { 3542 // Pull both arguments off the stack. 3543 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3544 args[0] = argument(0); 3545 args[1] = argument(1); 3546 3547 // We need write barriers here, because for primitive types we later compare 3548 // the two Class objects using ==, and those would give false negatives 3549 // if one obj is in from-space, and one in to-space. 3550 // TODO: Consider doing improved == comparison that only needs read barriers 3551 // on the false-path. 3552 args[0] = shenandoah_write_barrier(args[0]); 3553 args[1] = shenandoah_write_barrier(args[1]); 3554 3555 Node* klasses[2]; // corresponding Klasses: superk, subk 3556 klasses[0] = klasses[1] = top(); 3557 3558 enum { 3559 // A full decision tree on {superc is prim, subc is prim}: 3560 _prim_0_path = 1, // {P,N} => false 3561 // {P,P} & superc!=subc => false 3562 _prim_same_path, // {P,P} & superc==subc => true 3563 _prim_1_path, // {N,P} => false 3564 _ref_subtype_path, // {N,N} & subtype check wins => true 3565 _both_ref_path, // {N,N} & subtype check loses => false 3566 PATH_LIMIT 3567 }; 3568 3569 RegionNode* region = new RegionNode(PATH_LIMIT); 3570 Node* phi = new PhiNode(region, TypeInt::BOOL); 3571 record_for_igvn(region); 3572 3573 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3574 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3575 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3576 3577 // First null-check both mirrors and load each mirror's klass metaobject. 3578 int which_arg; 3579 for (which_arg = 0; which_arg <= 1; which_arg++) { 3580 Node* arg = args[which_arg]; 3581 arg = null_check(arg); 3582 if (stopped()) break; 3583 args[which_arg] = arg; 3584 3585 Node* p = basic_plus_adr(arg, class_klass_offset); 3586 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3587 klasses[which_arg] = _gvn.transform(kls); 3588 } 3589 3590 // Having loaded both klasses, test each for null. 3591 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3592 for (which_arg = 0; which_arg <= 1; which_arg++) { 3593 Node* kls = klasses[which_arg]; 3594 Node* null_ctl = top(); 3595 kls = null_check_oop(kls, &null_ctl, never_see_null); 3596 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3597 region->init_req(prim_path, null_ctl); 3598 if (stopped()) break; 3599 klasses[which_arg] = kls; 3600 } 3601 3602 if (!stopped()) { 3603 // now we have two reference types, in klasses[0..1] 3604 Node* subk = klasses[1]; // the argument to isAssignableFrom 3605 Node* superk = klasses[0]; // the receiver 3606 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3607 // now we have a successful reference subtype check 3608 region->set_req(_ref_subtype_path, control()); 3609 } 3610 3611 // If both operands are primitive (both klasses null), then 3612 // we must return true when they are identical primitives. 3613 // It is convenient to test this after the first null klass check. 3614 set_control(region->in(_prim_0_path)); // go back to first null check 3615 if (!stopped()) { 3616 // Since superc is primitive, make a guard for the superc==subc case. 3617 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3618 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3619 generate_guard(bol_eq, region, PROB_FAIR); 3620 if (region->req() == PATH_LIMIT+1) { 3621 // A guard was added. If the added guard is taken, superc==subc. 3622 region->swap_edges(PATH_LIMIT, _prim_same_path); 3623 region->del_req(PATH_LIMIT); 3624 } 3625 region->set_req(_prim_0_path, control()); // Not equal after all. 3626 } 3627 3628 // these are the only paths that produce 'true': 3629 phi->set_req(_prim_same_path, intcon(1)); 3630 phi->set_req(_ref_subtype_path, intcon(1)); 3631 3632 // pull together the cases: 3633 assert(region->req() == PATH_LIMIT, "sane region"); 3634 for (uint i = 1; i < region->req(); i++) { 3635 Node* ctl = region->in(i); 3636 if (ctl == NULL || ctl == top()) { 3637 region->set_req(i, top()); 3638 phi ->set_req(i, top()); 3639 } else if (phi->in(i) == NULL) { 3640 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3641 } 3642 } 3643 3644 set_control(_gvn.transform(region)); 3645 set_result(_gvn.transform(phi)); 3646 return true; 3647 } 3648 3649 //---------------------generate_array_guard_common------------------------ 3650 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3651 bool obj_array, bool not_array) { 3652 3653 if (stopped()) { 3654 return NULL; 3655 } 3656 3657 // If obj_array/non_array==false/false: 3658 // Branch around if the given klass is in fact an array (either obj or prim). 3659 // If obj_array/non_array==false/true: 3660 // Branch around if the given klass is not an array klass of any kind. 3661 // If obj_array/non_array==true/true: 3662 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3663 // If obj_array/non_array==true/false: 3664 // Branch around if the kls is an oop array (Object[] or subtype) 3665 // 3666 // Like generate_guard, adds a new path onto the region. 3667 jint layout_con = 0; 3668 Node* layout_val = get_layout_helper(kls, layout_con); 3669 if (layout_val == NULL) { 3670 bool query = (obj_array 3671 ? Klass::layout_helper_is_objArray(layout_con) 3672 : Klass::layout_helper_is_array(layout_con)); 3673 if (query == not_array) { 3674 return NULL; // never a branch 3675 } else { // always a branch 3676 Node* always_branch = control(); 3677 if (region != NULL) 3678 region->add_req(always_branch); 3679 set_control(top()); 3680 return always_branch; 3681 } 3682 } 3683 // Now test the correct condition. 3684 jint nval = (obj_array 3685 ? ((jint)Klass::_lh_array_tag_type_value 3686 << Klass::_lh_array_tag_shift) 3687 : Klass::_lh_neutral_value); 3688 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3689 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3690 // invert the test if we are looking for a non-array 3691 if (not_array) btest = BoolTest(btest).negate(); 3692 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3693 return generate_fair_guard(bol, region); 3694 } 3695 3696 3697 //-----------------------inline_native_newArray-------------------------- 3698 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3699 bool LibraryCallKit::inline_native_newArray() { 3700 Node* mirror = argument(0); 3701 Node* count_val = argument(1); 3702 3703 mirror = null_check(mirror); 3704 // If mirror or obj is dead, only null-path is taken. 3705 if (stopped()) return true; 3706 3707 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3708 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3709 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3710 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3711 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3712 3713 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3714 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3715 result_reg, _slow_path); 3716 Node* normal_ctl = control(); 3717 Node* no_array_ctl = result_reg->in(_slow_path); 3718 3719 // Generate code for the slow case. We make a call to newArray(). 3720 set_control(no_array_ctl); 3721 if (!stopped()) { 3722 // Either the input type is void.class, or else the 3723 // array klass has not yet been cached. Either the 3724 // ensuing call will throw an exception, or else it 3725 // will cache the array klass for next time. 3726 PreserveJVMState pjvms(this); 3727 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3728 Node* slow_result = set_results_for_java_call(slow_call); 3729 // this->control() comes from set_results_for_java_call 3730 result_reg->set_req(_slow_path, control()); 3731 result_val->set_req(_slow_path, slow_result); 3732 result_io ->set_req(_slow_path, i_o()); 3733 result_mem->set_req(_slow_path, reset_memory()); 3734 } 3735 3736 set_control(normal_ctl); 3737 if (!stopped()) { 3738 // Normal case: The array type has been cached in the java.lang.Class. 3739 // The following call works fine even if the array type is polymorphic. 3740 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3741 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3742 result_reg->init_req(_normal_path, control()); 3743 result_val->init_req(_normal_path, obj); 3744 result_io ->init_req(_normal_path, i_o()); 3745 result_mem->init_req(_normal_path, reset_memory()); 3746 } 3747 3748 // Return the combined state. 3749 set_i_o( _gvn.transform(result_io) ); 3750 set_all_memory( _gvn.transform(result_mem)); 3751 3752 C->set_has_split_ifs(true); // Has chance for split-if optimization 3753 set_result(result_reg, result_val); 3754 return true; 3755 } 3756 3757 //----------------------inline_native_getLength-------------------------- 3758 // public static native int java.lang.reflect.Array.getLength(Object array); 3759 bool LibraryCallKit::inline_native_getLength() { 3760 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3761 3762 Node* array = null_check(argument(0)); 3763 // If array is dead, only null-path is taken. 3764 if (stopped()) return true; 3765 3766 // Deoptimize if it is a non-array. 3767 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3768 3769 if (non_array != NULL) { 3770 PreserveJVMState pjvms(this); 3771 set_control(non_array); 3772 uncommon_trap(Deoptimization::Reason_intrinsic, 3773 Deoptimization::Action_maybe_recompile); 3774 } 3775 3776 // If control is dead, only non-array-path is taken. 3777 if (stopped()) return true; 3778 3779 // The works fine even if the array type is polymorphic. 3780 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3781 Node* result = load_array_length(array); 3782 3783 C->set_has_split_ifs(true); // Has chance for split-if optimization 3784 set_result(result); 3785 return true; 3786 } 3787 3788 //------------------------inline_array_copyOf---------------------------- 3789 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3790 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3791 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3792 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3793 3794 // Get the arguments. 3795 Node* original = argument(0); 3796 Node* start = is_copyOfRange? argument(1): intcon(0); 3797 Node* end = is_copyOfRange? argument(2): argument(1); 3798 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3799 3800 Node* newcopy; 3801 3802 // Set the original stack and the reexecute bit for the interpreter to reexecute 3803 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3804 { PreserveReexecuteState preexecs(this); 3805 jvms()->set_should_reexecute(true); 3806 3807 array_type_mirror = null_check(array_type_mirror); 3808 original = null_check(original); 3809 3810 // Check if a null path was taken unconditionally. 3811 if (stopped()) return true; 3812 3813 Node* orig_length = load_array_length(original); 3814 3815 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3816 klass_node = null_check(klass_node); 3817 3818 RegionNode* bailout = new RegionNode(1); 3819 record_for_igvn(bailout); 3820 3821 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3822 // Bail out if that is so. 3823 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3824 if (not_objArray != NULL) { 3825 // Improve the klass node's type from the new optimistic assumption: 3826 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3827 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3828 Node* cast = new CastPPNode(klass_node, akls); 3829 cast->init_req(0, control()); 3830 klass_node = _gvn.transform(cast); 3831 } 3832 3833 // Bail out if either start or end is negative. 3834 generate_negative_guard(start, bailout, &start); 3835 generate_negative_guard(end, bailout, &end); 3836 3837 Node* length = end; 3838 if (_gvn.type(start) != TypeInt::ZERO) { 3839 length = _gvn.transform(new SubINode(end, start)); 3840 } 3841 3842 // Bail out if length is negative. 3843 // Without this the new_array would throw 3844 // NegativeArraySizeException but IllegalArgumentException is what 3845 // should be thrown 3846 generate_negative_guard(length, bailout, &length); 3847 3848 if (bailout->req() > 1) { 3849 PreserveJVMState pjvms(this); 3850 set_control(_gvn.transform(bailout)); 3851 uncommon_trap(Deoptimization::Reason_intrinsic, 3852 Deoptimization::Action_maybe_recompile); 3853 } 3854 3855 if (!stopped()) { 3856 // How many elements will we copy from the original? 3857 // The answer is MinI(orig_length - start, length). 3858 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3859 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3860 3861 original = shenandoah_read_barrier(original); 3862 3863 // Generate a direct call to the right arraycopy function(s). 3864 // We know the copy is disjoint but we might not know if the 3865 // oop stores need checking. 3866 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3867 // This will fail a store-check if x contains any non-nulls. 3868 3869 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3870 // loads/stores but it is legal only if we're sure the 3871 // Arrays.copyOf would succeed. So we need all input arguments 3872 // to the copyOf to be validated, including that the copy to the 3873 // new array won't trigger an ArrayStoreException. That subtype 3874 // check can be optimized if we know something on the type of 3875 // the input array from type speculation. 3876 if (_gvn.type(klass_node)->singleton()) { 3877 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3878 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3879 3880 int test = C->static_subtype_check(superk, subk); 3881 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3882 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3883 if (t_original->speculative_type() != NULL) { 3884 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3885 } 3886 } 3887 } 3888 3889 bool validated = false; 3890 // Reason_class_check rather than Reason_intrinsic because we 3891 // want to intrinsify even if this traps. 3892 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3893 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3894 klass_node); 3895 3896 if (not_subtype_ctrl != top()) { 3897 PreserveJVMState pjvms(this); 3898 set_control(not_subtype_ctrl); 3899 uncommon_trap(Deoptimization::Reason_class_check, 3900 Deoptimization::Action_make_not_entrant); 3901 assert(stopped(), "Should be stopped"); 3902 } 3903 validated = true; 3904 } 3905 3906 if (!stopped()) { 3907 newcopy = new_array(klass_node, length, 0); // no arguments to push 3908 3909 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3910 load_object_klass(original), klass_node); 3911 if (!is_copyOfRange) { 3912 ac->set_copyof(validated); 3913 } else { 3914 ac->set_copyofrange(validated); 3915 } 3916 Node* n = _gvn.transform(ac); 3917 if (n == ac) { 3918 ac->connect_outputs(this); 3919 } else { 3920 assert(validated, "shouldn't transform if all arguments not validated"); 3921 set_all_memory(n); 3922 } 3923 } 3924 } 3925 } // original reexecute is set back here 3926 3927 C->set_has_split_ifs(true); // Has chance for split-if optimization 3928 if (!stopped()) { 3929 set_result(newcopy); 3930 } 3931 return true; 3932 } 3933 3934 3935 //----------------------generate_virtual_guard--------------------------- 3936 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3937 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3938 RegionNode* slow_region) { 3939 ciMethod* method = callee(); 3940 int vtable_index = method->vtable_index(); 3941 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3942 err_msg_res("bad index %d", vtable_index)); 3943 // Get the Method* out of the appropriate vtable entry. 3944 int entry_offset = (InstanceKlass::vtable_start_offset() + 3945 vtable_index*vtableEntry::size()) * wordSize + 3946 vtableEntry::method_offset_in_bytes(); 3947 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3948 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3949 3950 // Compare the target method with the expected method (e.g., Object.hashCode). 3951 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3952 3953 Node* native_call = makecon(native_call_addr); 3954 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3955 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3956 3957 return generate_slow_guard(test_native, slow_region); 3958 } 3959 3960 //-----------------------generate_method_call---------------------------- 3961 // Use generate_method_call to make a slow-call to the real 3962 // method if the fast path fails. An alternative would be to 3963 // use a stub like OptoRuntime::slow_arraycopy_Java. 3964 // This only works for expanding the current library call, 3965 // not another intrinsic. (E.g., don't use this for making an 3966 // arraycopy call inside of the copyOf intrinsic.) 3967 CallJavaNode* 3968 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3969 // When compiling the intrinsic method itself, do not use this technique. 3970 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3971 3972 ciMethod* method = callee(); 3973 // ensure the JVMS we have will be correct for this call 3974 guarantee(method_id == method->intrinsic_id(), "must match"); 3975 3976 const TypeFunc* tf = TypeFunc::make(method); 3977 CallJavaNode* slow_call; 3978 if (is_static) { 3979 assert(!is_virtual, ""); 3980 slow_call = new CallStaticJavaNode(C, tf, 3981 SharedRuntime::get_resolve_static_call_stub(), 3982 method, bci()); 3983 } else if (is_virtual) { 3984 null_check_receiver(); 3985 int vtable_index = Method::invalid_vtable_index; 3986 if (UseInlineCaches) { 3987 // Suppress the vtable call 3988 } else { 3989 // hashCode and clone are not a miranda methods, 3990 // so the vtable index is fixed. 3991 // No need to use the linkResolver to get it. 3992 vtable_index = method->vtable_index(); 3993 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3994 err_msg_res("bad index %d", vtable_index)); 3995 } 3996 slow_call = new CallDynamicJavaNode(tf, 3997 SharedRuntime::get_resolve_virtual_call_stub(), 3998 method, vtable_index, bci()); 3999 } else { // neither virtual nor static: opt_virtual 4000 null_check_receiver(); 4001 slow_call = new CallStaticJavaNode(C, tf, 4002 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4003 method, bci()); 4004 slow_call->set_optimized_virtual(true); 4005 } 4006 set_arguments_for_java_call(slow_call); 4007 set_edges_for_java_call(slow_call); 4008 return slow_call; 4009 } 4010 4011 4012 /** 4013 * Build special case code for calls to hashCode on an object. This call may 4014 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4015 * slightly different code. 4016 */ 4017 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4018 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4019 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4020 4021 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4022 4023 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4024 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4025 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4026 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4027 Node* obj = NULL; 4028 if (!is_static) { 4029 // Check for hashing null object 4030 obj = null_check_receiver(); 4031 if (stopped()) return true; // unconditionally null 4032 result_reg->init_req(_null_path, top()); 4033 result_val->init_req(_null_path, top()); 4034 } else { 4035 // Do a null check, and return zero if null. 4036 // System.identityHashCode(null) == 0 4037 obj = argument(0); 4038 Node* null_ctl = top(); 4039 obj = null_check_oop(obj, &null_ctl); 4040 result_reg->init_req(_null_path, null_ctl); 4041 result_val->init_req(_null_path, _gvn.intcon(0)); 4042 } 4043 4044 if (ShenandoahVerifyReadsToFromSpace) { 4045 obj = shenandoah_read_barrier(obj); 4046 } 4047 4048 // Unconditionally null? Then return right away. 4049 if (stopped()) { 4050 set_control( result_reg->in(_null_path)); 4051 if (!stopped()) 4052 set_result(result_val->in(_null_path)); 4053 return true; 4054 } 4055 4056 // We only go to the fast case code if we pass a number of guards. The 4057 // paths which do not pass are accumulated in the slow_region. 4058 RegionNode* slow_region = new RegionNode(1); 4059 record_for_igvn(slow_region); 4060 4061 // If this is a virtual call, we generate a funny guard. We pull out 4062 // the vtable entry corresponding to hashCode() from the target object. 4063 // If the target method which we are calling happens to be the native 4064 // Object hashCode() method, we pass the guard. We do not need this 4065 // guard for non-virtual calls -- the caller is known to be the native 4066 // Object hashCode(). 4067 if (is_virtual) { 4068 // After null check, get the object's klass. 4069 Node* obj_klass = load_object_klass(obj); 4070 generate_virtual_guard(obj_klass, slow_region); 4071 } 4072 4073 // Get the header out of the object, use LoadMarkNode when available 4074 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4075 // The control of the load must be NULL. Otherwise, the load can move before 4076 // the null check after castPP removal. 4077 Node* no_ctrl = NULL; 4078 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4079 4080 // Test the header to see if it is unlocked. 4081 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4082 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4083 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4084 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4085 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4086 4087 generate_slow_guard(test_unlocked, slow_region); 4088 4089 // Get the hash value and check to see that it has been properly assigned. 4090 // We depend on hash_mask being at most 32 bits and avoid the use of 4091 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4092 // vm: see markOop.hpp. 4093 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4094 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4095 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4096 // This hack lets the hash bits live anywhere in the mark object now, as long 4097 // as the shift drops the relevant bits into the low 32 bits. Note that 4098 // Java spec says that HashCode is an int so there's no point in capturing 4099 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4100 hshifted_header = ConvX2I(hshifted_header); 4101 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4102 4103 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4104 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4105 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4106 4107 generate_slow_guard(test_assigned, slow_region); 4108 4109 Node* init_mem = reset_memory(); 4110 // fill in the rest of the null path: 4111 result_io ->init_req(_null_path, i_o()); 4112 result_mem->init_req(_null_path, init_mem); 4113 4114 result_val->init_req(_fast_path, hash_val); 4115 result_reg->init_req(_fast_path, control()); 4116 result_io ->init_req(_fast_path, i_o()); 4117 result_mem->init_req(_fast_path, init_mem); 4118 4119 // Generate code for the slow case. We make a call to hashCode(). 4120 set_control(_gvn.transform(slow_region)); 4121 if (!stopped()) { 4122 // No need for PreserveJVMState, because we're using up the present state. 4123 set_all_memory(init_mem); 4124 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4125 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4126 Node* slow_result = set_results_for_java_call(slow_call); 4127 // this->control() comes from set_results_for_java_call 4128 result_reg->init_req(_slow_path, control()); 4129 result_val->init_req(_slow_path, slow_result); 4130 result_io ->set_req(_slow_path, i_o()); 4131 result_mem ->set_req(_slow_path, reset_memory()); 4132 } 4133 4134 // Return the combined state. 4135 set_i_o( _gvn.transform(result_io) ); 4136 set_all_memory( _gvn.transform(result_mem)); 4137 4138 set_result(result_reg, result_val); 4139 return true; 4140 } 4141 4142 //---------------------------inline_native_getClass---------------------------- 4143 // public final native Class<?> java.lang.Object.getClass(); 4144 // 4145 // Build special case code for calls to getClass on an object. 4146 bool LibraryCallKit::inline_native_getClass() { 4147 Node* obj = null_check_receiver(); 4148 if (stopped()) return true; 4149 set_result(load_mirror_from_klass(load_object_klass(obj))); 4150 return true; 4151 } 4152 4153 //-----------------inline_native_Reflection_getCallerClass--------------------- 4154 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4155 // 4156 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4157 // 4158 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4159 // in that it must skip particular security frames and checks for 4160 // caller sensitive methods. 4161 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4162 #ifndef PRODUCT 4163 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4164 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4165 } 4166 #endif 4167 4168 if (!jvms()->has_method()) { 4169 #ifndef PRODUCT 4170 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4171 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4172 } 4173 #endif 4174 return false; 4175 } 4176 4177 // Walk back up the JVM state to find the caller at the required 4178 // depth. 4179 JVMState* caller_jvms = jvms(); 4180 4181 // Cf. JVM_GetCallerClass 4182 // NOTE: Start the loop at depth 1 because the current JVM state does 4183 // not include the Reflection.getCallerClass() frame. 4184 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4185 ciMethod* m = caller_jvms->method(); 4186 switch (n) { 4187 case 0: 4188 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4189 break; 4190 case 1: 4191 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4192 if (!m->caller_sensitive()) { 4193 #ifndef PRODUCT 4194 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4195 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4196 } 4197 #endif 4198 return false; // bail-out; let JVM_GetCallerClass do the work 4199 } 4200 break; 4201 default: 4202 if (!m->is_ignored_by_security_stack_walk()) { 4203 // We have reached the desired frame; return the holder class. 4204 // Acquire method holder as java.lang.Class and push as constant. 4205 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4206 ciInstance* caller_mirror = caller_klass->java_mirror(); 4207 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4208 4209 #ifndef PRODUCT 4210 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4211 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()); 4212 tty->print_cr(" JVM state at this point:"); 4213 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4214 ciMethod* m = jvms()->of_depth(i)->method(); 4215 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4216 } 4217 } 4218 #endif 4219 return true; 4220 } 4221 break; 4222 } 4223 } 4224 4225 #ifndef PRODUCT 4226 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4227 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4228 tty->print_cr(" JVM state at this point:"); 4229 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4230 ciMethod* m = jvms()->of_depth(i)->method(); 4231 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4232 } 4233 } 4234 #endif 4235 4236 return false; // bail-out; let JVM_GetCallerClass do the work 4237 } 4238 4239 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4240 Node* arg = argument(0); 4241 Node* result; 4242 4243 switch (id) { 4244 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4245 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4246 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4247 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4248 4249 case vmIntrinsics::_doubleToLongBits: { 4250 // two paths (plus control) merge in a wood 4251 RegionNode *r = new RegionNode(3); 4252 Node *phi = new PhiNode(r, TypeLong::LONG); 4253 4254 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4255 // Build the boolean node 4256 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4257 4258 // Branch either way. 4259 // NaN case is less traveled, which makes all the difference. 4260 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4261 Node *opt_isnan = _gvn.transform(ifisnan); 4262 assert( opt_isnan->is_If(), "Expect an IfNode"); 4263 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4264 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4265 4266 set_control(iftrue); 4267 4268 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4269 Node *slow_result = longcon(nan_bits); // return NaN 4270 phi->init_req(1, _gvn.transform( slow_result )); 4271 r->init_req(1, iftrue); 4272 4273 // Else fall through 4274 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4275 set_control(iffalse); 4276 4277 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4278 r->init_req(2, iffalse); 4279 4280 // Post merge 4281 set_control(_gvn.transform(r)); 4282 record_for_igvn(r); 4283 4284 C->set_has_split_ifs(true); // Has chance for split-if optimization 4285 result = phi; 4286 assert(result->bottom_type()->isa_long(), "must be"); 4287 break; 4288 } 4289 4290 case vmIntrinsics::_floatToIntBits: { 4291 // two paths (plus control) merge in a wood 4292 RegionNode *r = new RegionNode(3); 4293 Node *phi = new PhiNode(r, TypeInt::INT); 4294 4295 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4296 // Build the boolean node 4297 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4298 4299 // Branch either way. 4300 // NaN case is less traveled, which makes all the difference. 4301 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4302 Node *opt_isnan = _gvn.transform(ifisnan); 4303 assert( opt_isnan->is_If(), "Expect an IfNode"); 4304 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4305 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4306 4307 set_control(iftrue); 4308 4309 static const jint nan_bits = 0x7fc00000; 4310 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4311 phi->init_req(1, _gvn.transform( slow_result )); 4312 r->init_req(1, iftrue); 4313 4314 // Else fall through 4315 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4316 set_control(iffalse); 4317 4318 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4319 r->init_req(2, iffalse); 4320 4321 // Post merge 4322 set_control(_gvn.transform(r)); 4323 record_for_igvn(r); 4324 4325 C->set_has_split_ifs(true); // Has chance for split-if optimization 4326 result = phi; 4327 assert(result->bottom_type()->isa_int(), "must be"); 4328 break; 4329 } 4330 4331 default: 4332 fatal_unexpected_iid(id); 4333 break; 4334 } 4335 set_result(_gvn.transform(result)); 4336 return true; 4337 } 4338 4339 #ifdef _LP64 4340 #define XTOP ,top() /*additional argument*/ 4341 #else //_LP64 4342 #define XTOP /*no additional argument*/ 4343 #endif //_LP64 4344 4345 //----------------------inline_unsafe_copyMemory------------------------- 4346 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4347 bool LibraryCallKit::inline_unsafe_copyMemory() { 4348 if (callee()->is_static()) return false; // caller must have the capability! 4349 null_check_receiver(); // null-check receiver 4350 if (stopped()) return true; 4351 4352 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4353 4354 Node* src_ptr = argument(1); // type: oop 4355 Node* src_off = ConvL2X(argument(2)); // type: long 4356 Node* dst_ptr = argument(4); // type: oop 4357 Node* dst_off = ConvL2X(argument(5)); // type: long 4358 Node* size = ConvL2X(argument(7)); // type: long 4359 4360 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4361 "fieldOffset must be byte-scaled"); 4362 4363 src_ptr = shenandoah_read_barrier(src_ptr); 4364 dst_ptr = shenandoah_write_barrier(dst_ptr); 4365 4366 Node* src = make_unsafe_address(src_ptr, src_off); 4367 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4368 4369 // Conservatively insert a memory barrier on all memory slices. 4370 // Do not let writes of the copy source or destination float below the copy. 4371 insert_mem_bar(Op_MemBarCPUOrder); 4372 4373 // Call it. Note that the length argument is not scaled. 4374 make_runtime_call(RC_LEAF|RC_NO_FP, 4375 OptoRuntime::fast_arraycopy_Type(), 4376 StubRoutines::unsafe_arraycopy(), 4377 "unsafe_arraycopy", 4378 TypeRawPtr::BOTTOM, 4379 src, dst, size XTOP); 4380 4381 // Do not let reads of the copy destination float above the copy. 4382 insert_mem_bar(Op_MemBarCPUOrder); 4383 4384 return true; 4385 } 4386 4387 //------------------------clone_coping----------------------------------- 4388 // Helper function for inline_native_clone. 4389 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4390 assert(obj_size != NULL, ""); 4391 Node* raw_obj = alloc_obj->in(1); 4392 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4393 4394 obj = shenandoah_read_barrier(obj); 4395 4396 AllocateNode* alloc = NULL; 4397 if (ReduceBulkZeroing) { 4398 // We will be completely responsible for initializing this object - 4399 // mark Initialize node as complete. 4400 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4401 // The object was just allocated - there should be no any stores! 4402 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4403 // Mark as complete_with_arraycopy so that on AllocateNode 4404 // expansion, we know this AllocateNode is initialized by an array 4405 // copy and a StoreStore barrier exists after the array copy. 4406 alloc->initialization()->set_complete_with_arraycopy(); 4407 } 4408 4409 // Copy the fastest available way. 4410 // TODO: generate fields copies for small objects instead. 4411 Node* src = obj; 4412 Node* dest = alloc_obj; 4413 Node* size = _gvn.transform(obj_size); 4414 4415 // Exclude the header but include array length to copy by 8 bytes words. 4416 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4417 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4418 instanceOopDesc::base_offset_in_bytes(); 4419 // base_off: 4420 // 8 - 32-bit VM 4421 // 12 - 64-bit VM, compressed klass 4422 // 16 - 64-bit VM, normal klass 4423 if (base_off % BytesPerLong != 0) { 4424 assert(UseCompressedClassPointers, ""); 4425 if (is_array) { 4426 // Exclude length to copy by 8 bytes words. 4427 base_off += sizeof(int); 4428 } else { 4429 // Include klass to copy by 8 bytes words. 4430 base_off = instanceOopDesc::klass_offset_in_bytes(); 4431 } 4432 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4433 } 4434 src = basic_plus_adr(src, base_off); 4435 dest = basic_plus_adr(dest, base_off); 4436 4437 // Compute the length also, if needed: 4438 Node* countx = size; 4439 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4440 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4441 4442 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4443 4444 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4445 ac->set_clonebasic(); 4446 Node* n = _gvn.transform(ac); 4447 if (n == ac) { 4448 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4449 } else { 4450 set_all_memory(n); 4451 } 4452 4453 if (UseShenandoahGC) { 4454 // Make sure that references in the cloned object are updated for Shenandoah. 4455 make_runtime_call(RC_LEAF|RC_NO_FP, 4456 OptoRuntime::shenandoah_clone_barrier_Type(), 4457 CAST_FROM_FN_PTR(address, SharedRuntime::shenandoah_clone_barrier), 4458 "shenandoah_clone_barrier", TypePtr::BOTTOM, 4459 alloc_obj); 4460 } 4461 4462 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4463 if (card_mark) { 4464 assert(!is_array, ""); 4465 // Put in store barrier for any and all oops we are sticking 4466 // into this object. (We could avoid this if we could prove 4467 // that the object type contains no oop fields at all.) 4468 Node* no_particular_value = NULL; 4469 Node* no_particular_field = NULL; 4470 int raw_adr_idx = Compile::AliasIdxRaw; 4471 post_barrier(control(), 4472 memory(raw_adr_type), 4473 alloc_obj, 4474 no_particular_field, 4475 raw_adr_idx, 4476 no_particular_value, 4477 T_OBJECT, 4478 false); 4479 } 4480 4481 // Do not let reads from the cloned object float above the arraycopy. 4482 if (alloc != NULL) { 4483 // Do not let stores that initialize this object be reordered with 4484 // a subsequent store that would make this object accessible by 4485 // other threads. 4486 // Record what AllocateNode this StoreStore protects so that 4487 // escape analysis can go from the MemBarStoreStoreNode to the 4488 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4489 // based on the escape status of the AllocateNode. 4490 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4491 } else { 4492 insert_mem_bar(Op_MemBarCPUOrder); 4493 } 4494 } 4495 4496 //------------------------inline_native_clone---------------------------- 4497 // protected native Object java.lang.Object.clone(); 4498 // 4499 // Here are the simple edge cases: 4500 // null receiver => normal trap 4501 // virtual and clone was overridden => slow path to out-of-line clone 4502 // not cloneable or finalizer => slow path to out-of-line Object.clone 4503 // 4504 // The general case has two steps, allocation and copying. 4505 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4506 // 4507 // Copying also has two cases, oop arrays and everything else. 4508 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4509 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4510 // 4511 // These steps fold up nicely if and when the cloned object's klass 4512 // can be sharply typed as an object array, a type array, or an instance. 4513 // 4514 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4515 PhiNode* result_val; 4516 4517 // Set the reexecute bit for the interpreter to reexecute 4518 // the bytecode that invokes Object.clone if deoptimization happens. 4519 { PreserveReexecuteState preexecs(this); 4520 jvms()->set_should_reexecute(true); 4521 4522 Node* obj = null_check_receiver(); 4523 if (stopped()) return true; 4524 4525 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4526 4527 // If we are going to clone an instance, we need its exact type to 4528 // know the number and types of fields to convert the clone to 4529 // loads/stores. Maybe a speculative type can help us. 4530 if (!obj_type->klass_is_exact() && 4531 obj_type->speculative_type() != NULL && 4532 obj_type->speculative_type()->is_instance_klass()) { 4533 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4534 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4535 !spec_ik->has_injected_fields()) { 4536 ciKlass* k = obj_type->klass(); 4537 if (!k->is_instance_klass() || 4538 k->as_instance_klass()->is_interface() || 4539 k->as_instance_klass()->has_subklass()) { 4540 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4541 } 4542 } 4543 } 4544 4545 Node* obj_klass = load_object_klass(obj); 4546 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4547 const TypeOopPtr* toop = ((tklass != NULL) 4548 ? tklass->as_instance_type() 4549 : TypeInstPtr::NOTNULL); 4550 4551 // Conservatively insert a memory barrier on all memory slices. 4552 // Do not let writes into the original float below the clone. 4553 insert_mem_bar(Op_MemBarCPUOrder); 4554 4555 // paths into result_reg: 4556 enum { 4557 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4558 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4559 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4560 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4561 PATH_LIMIT 4562 }; 4563 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4564 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4565 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4566 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4567 record_for_igvn(result_reg); 4568 4569 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4570 int raw_adr_idx = Compile::AliasIdxRaw; 4571 4572 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4573 if (array_ctl != NULL) { 4574 // It's an array. 4575 PreserveJVMState pjvms(this); 4576 set_control(array_ctl); 4577 Node* obj_length = load_array_length(obj); 4578 Node* obj_size = NULL; 4579 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4580 4581 if (!use_ReduceInitialCardMarks()) { 4582 // If it is an oop array, it requires very special treatment, 4583 // because card marking is required on each card of the array. 4584 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4585 if (is_obja != NULL) { 4586 PreserveJVMState pjvms2(this); 4587 set_control(is_obja); 4588 4589 obj = shenandoah_read_barrier(obj); 4590 4591 // Generate a direct call to the right arraycopy function(s). 4592 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4593 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4594 ac->set_cloneoop(); 4595 Node* n = _gvn.transform(ac); 4596 assert(n == ac, "cannot disappear"); 4597 ac->connect_outputs(this); 4598 4599 result_reg->init_req(_objArray_path, control()); 4600 result_val->init_req(_objArray_path, alloc_obj); 4601 result_i_o ->set_req(_objArray_path, i_o()); 4602 result_mem ->set_req(_objArray_path, reset_memory()); 4603 } 4604 } 4605 // Otherwise, there are no card marks to worry about. 4606 // (We can dispense with card marks if we know the allocation 4607 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4608 // causes the non-eden paths to take compensating steps to 4609 // simulate a fresh allocation, so that no further 4610 // card marks are required in compiled code to initialize 4611 // the object.) 4612 4613 if (!stopped()) { 4614 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4615 4616 // Present the results of the copy. 4617 result_reg->init_req(_array_path, control()); 4618 result_val->init_req(_array_path, alloc_obj); 4619 result_i_o ->set_req(_array_path, i_o()); 4620 result_mem ->set_req(_array_path, reset_memory()); 4621 } 4622 } 4623 4624 // We only go to the instance fast case code if we pass a number of guards. 4625 // The paths which do not pass are accumulated in the slow_region. 4626 RegionNode* slow_region = new RegionNode(1); 4627 record_for_igvn(slow_region); 4628 if (!stopped()) { 4629 // It's an instance (we did array above). Make the slow-path tests. 4630 // If this is a virtual call, we generate a funny guard. We grab 4631 // the vtable entry corresponding to clone() from the target object. 4632 // If the target method which we are calling happens to be the 4633 // Object clone() method, we pass the guard. We do not need this 4634 // guard for non-virtual calls; the caller is known to be the native 4635 // Object clone(). 4636 if (is_virtual) { 4637 generate_virtual_guard(obj_klass, slow_region); 4638 } 4639 4640 // The object must be cloneable and must not have a finalizer. 4641 // Both of these conditions may be checked in a single test. 4642 // We could optimize the cloneable test further, but we don't care. 4643 generate_access_flags_guard(obj_klass, 4644 // Test both conditions: 4645 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4646 // Must be cloneable but not finalizer: 4647 JVM_ACC_IS_CLONEABLE, 4648 slow_region); 4649 } 4650 4651 if (!stopped()) { 4652 // It's an instance, and it passed the slow-path tests. 4653 PreserveJVMState pjvms(this); 4654 Node* obj_size = NULL; 4655 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4656 // is reexecuted if deoptimization occurs and there could be problems when merging 4657 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4658 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4659 4660 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4661 4662 // Present the results of the slow call. 4663 result_reg->init_req(_instance_path, control()); 4664 result_val->init_req(_instance_path, alloc_obj); 4665 result_i_o ->set_req(_instance_path, i_o()); 4666 result_mem ->set_req(_instance_path, reset_memory()); 4667 } 4668 4669 // Generate code for the slow case. We make a call to clone(). 4670 set_control(_gvn.transform(slow_region)); 4671 if (!stopped()) { 4672 PreserveJVMState pjvms(this); 4673 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4674 Node* slow_result = set_results_for_java_call(slow_call); 4675 // this->control() comes from set_results_for_java_call 4676 result_reg->init_req(_slow_path, control()); 4677 result_val->init_req(_slow_path, slow_result); 4678 result_i_o ->set_req(_slow_path, i_o()); 4679 result_mem ->set_req(_slow_path, reset_memory()); 4680 } 4681 4682 // Return the combined state. 4683 set_control( _gvn.transform(result_reg)); 4684 set_i_o( _gvn.transform(result_i_o)); 4685 set_all_memory( _gvn.transform(result_mem)); 4686 } // original reexecute is set back here 4687 4688 set_result(_gvn.transform(result_val)); 4689 return true; 4690 } 4691 4692 // If we have a tighly coupled allocation, the arraycopy may take care 4693 // of the array initialization. If one of the guards we insert between 4694 // the allocation and the arraycopy causes a deoptimization, an 4695 // unitialized array will escape the compiled method. To prevent that 4696 // we set the JVM state for uncommon traps between the allocation and 4697 // the arraycopy to the state before the allocation so, in case of 4698 // deoptimization, we'll reexecute the allocation and the 4699 // initialization. 4700 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4701 if (alloc != NULL) { 4702 ciMethod* trap_method = alloc->jvms()->method(); 4703 int trap_bci = alloc->jvms()->bci(); 4704 4705 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4706 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4707 // Make sure there's no store between the allocation and the 4708 // arraycopy otherwise visible side effects could be rexecuted 4709 // in case of deoptimization and cause incorrect execution. 4710 bool no_interfering_store = true; 4711 Node* mem = alloc->in(TypeFunc::Memory); 4712 if (mem->is_MergeMem()) { 4713 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4714 Node* n = mms.memory(); 4715 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4716 assert(n->is_Store(), "what else?"); 4717 no_interfering_store = false; 4718 break; 4719 } 4720 } 4721 } else { 4722 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4723 Node* n = mms.memory(); 4724 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4725 assert(n->is_Store(), "what else?"); 4726 no_interfering_store = false; 4727 break; 4728 } 4729 } 4730 } 4731 4732 if (no_interfering_store) { 4733 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4734 uint size = alloc->req(); 4735 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4736 old_jvms->set_map(sfpt); 4737 for (uint i = 0; i < size; i++) { 4738 sfpt->init_req(i, alloc->in(i)); 4739 } 4740 // re-push array length for deoptimization 4741 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4742 old_jvms->set_sp(old_jvms->sp()+1); 4743 old_jvms->set_monoff(old_jvms->monoff()+1); 4744 old_jvms->set_scloff(old_jvms->scloff()+1); 4745 old_jvms->set_endoff(old_jvms->endoff()+1); 4746 old_jvms->set_should_reexecute(true); 4747 4748 sfpt->set_i_o(map()->i_o()); 4749 sfpt->set_memory(map()->memory()); 4750 sfpt->set_control(map()->control()); 4751 4752 JVMState* saved_jvms = jvms(); 4753 saved_reexecute_sp = _reexecute_sp; 4754 4755 set_jvms(sfpt->jvms()); 4756 _reexecute_sp = jvms()->sp(); 4757 4758 return saved_jvms; 4759 } 4760 } 4761 } 4762 return NULL; 4763 } 4764 4765 // In case of a deoptimization, we restart execution at the 4766 // allocation, allocating a new array. We would leave an uninitialized 4767 // array in the heap that GCs wouldn't expect. Move the allocation 4768 // after the traps so we don't allocate the array if we 4769 // deoptimize. This is possible because tightly_coupled_allocation() 4770 // guarantees there's no observer of the allocated array at this point 4771 // and the control flow is simple enough. 4772 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4773 if (saved_jvms != NULL && !stopped()) { 4774 assert(alloc != NULL, "only with a tightly coupled allocation"); 4775 // restore JVM state to the state at the arraycopy 4776 saved_jvms->map()->set_control(map()->control()); 4777 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4778 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4779 // If we've improved the types of some nodes (null check) while 4780 // emitting the guards, propagate them to the current state 4781 map()->replaced_nodes().apply(saved_jvms->map()); 4782 set_jvms(saved_jvms); 4783 _reexecute_sp = saved_reexecute_sp; 4784 4785 // Remove the allocation from above the guards 4786 CallProjections callprojs; 4787 alloc->extract_projections(&callprojs, true); 4788 InitializeNode* init = alloc->initialization(); 4789 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4790 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4791 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4792 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4793 4794 // move the allocation here (after the guards) 4795 _gvn.hash_delete(alloc); 4796 alloc->set_req(TypeFunc::Control, control()); 4797 alloc->set_req(TypeFunc::I_O, i_o()); 4798 Node *mem = reset_memory(); 4799 set_all_memory(mem); 4800 alloc->set_req(TypeFunc::Memory, mem); 4801 set_control(init->proj_out(TypeFunc::Control)); 4802 set_i_o(callprojs.fallthrough_ioproj); 4803 4804 // Update memory as done in GraphKit::set_output_for_allocation() 4805 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4806 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4807 if (ary_type->isa_aryptr() && length_type != NULL) { 4808 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4809 } 4810 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4811 int elemidx = C->get_alias_index(telemref); 4812 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4813 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4814 4815 Node* allocx = _gvn.transform(alloc); 4816 assert(allocx == alloc, "where has the allocation gone?"); 4817 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4818 4819 _gvn.hash_delete(dest); 4820 dest->set_req(0, control()); 4821 Node* destx = _gvn.transform(dest); 4822 assert(destx == dest, "where has the allocation result gone?"); 4823 } 4824 } 4825 4826 4827 //------------------------------inline_arraycopy----------------------- 4828 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4829 // Object dest, int destPos, 4830 // int length); 4831 bool LibraryCallKit::inline_arraycopy() { 4832 // Get the arguments. 4833 Node* src = argument(0); // type: oop 4834 Node* src_offset = argument(1); // type: int 4835 Node* dest = argument(2); // type: oop 4836 Node* dest_offset = argument(3); // type: int 4837 Node* length = argument(4); // type: int 4838 4839 src = shenandoah_read_barrier(src); 4840 dest = shenandoah_write_barrier(dest); 4841 4842 // Check for allocation before we add nodes that would confuse 4843 // tightly_coupled_allocation() 4844 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4845 4846 int saved_reexecute_sp = -1; 4847 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4848 // See arraycopy_restore_alloc_state() comment 4849 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4850 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4851 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4852 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4853 4854 // The following tests must be performed 4855 // (1) src and dest are arrays. 4856 // (2) src and dest arrays must have elements of the same BasicType 4857 // (3) src and dest must not be null. 4858 // (4) src_offset must not be negative. 4859 // (5) dest_offset must not be negative. 4860 // (6) length must not be negative. 4861 // (7) src_offset + length must not exceed length of src. 4862 // (8) dest_offset + length must not exceed length of dest. 4863 // (9) each element of an oop array must be assignable 4864 4865 // (3) src and dest must not be null. 4866 // always do this here because we need the JVM state for uncommon traps 4867 Node* null_ctl = top(); 4868 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4869 assert(null_ctl->is_top(), "no null control here"); 4870 dest = null_check(dest, T_ARRAY); 4871 4872 if (!can_emit_guards) { 4873 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4874 // guards but the arraycopy node could still take advantage of a 4875 // tightly allocated allocation. tightly_coupled_allocation() is 4876 // called again to make sure it takes the null check above into 4877 // account: the null check is mandatory and if it caused an 4878 // uncommon trap to be emitted then the allocation can't be 4879 // considered tightly coupled in this context. 4880 alloc = tightly_coupled_allocation(dest, NULL); 4881 } 4882 4883 bool validated = false; 4884 4885 const Type* src_type = _gvn.type(src); 4886 const Type* dest_type = _gvn.type(dest); 4887 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4888 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4889 4890 // Do we have the type of src? 4891 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4892 // Do we have the type of dest? 4893 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4894 // Is the type for src from speculation? 4895 bool src_spec = false; 4896 // Is the type for dest from speculation? 4897 bool dest_spec = false; 4898 4899 if ((!has_src || !has_dest) && can_emit_guards) { 4900 // We don't have sufficient type information, let's see if 4901 // speculative types can help. We need to have types for both src 4902 // and dest so that it pays off. 4903 4904 // Do we already have or could we have type information for src 4905 bool could_have_src = has_src; 4906 // Do we already have or could we have type information for dest 4907 bool could_have_dest = has_dest; 4908 4909 ciKlass* src_k = NULL; 4910 if (!has_src) { 4911 src_k = src_type->speculative_type_not_null(); 4912 if (src_k != NULL && src_k->is_array_klass()) { 4913 could_have_src = true; 4914 } 4915 } 4916 4917 ciKlass* dest_k = NULL; 4918 if (!has_dest) { 4919 dest_k = dest_type->speculative_type_not_null(); 4920 if (dest_k != NULL && dest_k->is_array_klass()) { 4921 could_have_dest = true; 4922 } 4923 } 4924 4925 if (could_have_src && could_have_dest) { 4926 // This is going to pay off so emit the required guards 4927 if (!has_src) { 4928 src = maybe_cast_profiled_obj(src, src_k, true); 4929 src_type = _gvn.type(src); 4930 top_src = src_type->isa_aryptr(); 4931 has_src = (top_src != NULL && top_src->klass() != NULL); 4932 src_spec = true; 4933 } 4934 if (!has_dest) { 4935 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4936 dest_type = _gvn.type(dest); 4937 top_dest = dest_type->isa_aryptr(); 4938 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4939 dest_spec = true; 4940 } 4941 } 4942 } 4943 4944 if (has_src && has_dest && can_emit_guards) { 4945 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4946 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4947 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4948 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4949 4950 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4951 // If both arrays are object arrays then having the exact types 4952 // for both will remove the need for a subtype check at runtime 4953 // before the call and may make it possible to pick a faster copy 4954 // routine (without a subtype check on every element) 4955 // Do we have the exact type of src? 4956 bool could_have_src = src_spec; 4957 // Do we have the exact type of dest? 4958 bool could_have_dest = dest_spec; 4959 ciKlass* src_k = top_src->klass(); 4960 ciKlass* dest_k = top_dest->klass(); 4961 if (!src_spec) { 4962 src_k = src_type->speculative_type_not_null(); 4963 if (src_k != NULL && src_k->is_array_klass()) { 4964 could_have_src = true; 4965 } 4966 } 4967 if (!dest_spec) { 4968 dest_k = dest_type->speculative_type_not_null(); 4969 if (dest_k != NULL && dest_k->is_array_klass()) { 4970 could_have_dest = true; 4971 } 4972 } 4973 if (could_have_src && could_have_dest) { 4974 // If we can have both exact types, emit the missing guards 4975 if (could_have_src && !src_spec) { 4976 src = maybe_cast_profiled_obj(src, src_k, true); 4977 } 4978 if (could_have_dest && !dest_spec) { 4979 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4980 } 4981 } 4982 } 4983 } 4984 4985 ciMethod* trap_method = method(); 4986 int trap_bci = bci(); 4987 if (saved_jvms != NULL) { 4988 trap_method = alloc->jvms()->method(); 4989 trap_bci = alloc->jvms()->bci(); 4990 } 4991 4992 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 4993 can_emit_guards && 4994 !src->is_top() && !dest->is_top()) { 4995 // validate arguments: enables transformation the ArrayCopyNode 4996 validated = true; 4997 4998 RegionNode* slow_region = new RegionNode(1); 4999 record_for_igvn(slow_region); 5000 5001 // (1) src and dest are arrays. 5002 generate_non_array_guard(load_object_klass(src), slow_region); 5003 generate_non_array_guard(load_object_klass(dest), slow_region); 5004 5005 // (2) src and dest arrays must have elements of the same BasicType 5006 // done at macro expansion or at Ideal transformation time 5007 5008 // (4) src_offset must not be negative. 5009 generate_negative_guard(src_offset, slow_region); 5010 5011 // (5) dest_offset must not be negative. 5012 generate_negative_guard(dest_offset, slow_region); 5013 5014 // (7) src_offset + length must not exceed length of src. 5015 generate_limit_guard(src_offset, length, 5016 load_array_length(src), 5017 slow_region); 5018 5019 // (8) dest_offset + length must not exceed length of dest. 5020 generate_limit_guard(dest_offset, length, 5021 load_array_length(dest), 5022 slow_region); 5023 5024 // (9) each element of an oop array must be assignable 5025 Node* src_klass = load_object_klass(src); 5026 Node* dest_klass = load_object_klass(dest); 5027 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5028 5029 if (not_subtype_ctrl != top()) { 5030 PreserveJVMState pjvms(this); 5031 set_control(not_subtype_ctrl); 5032 uncommon_trap(Deoptimization::Reason_intrinsic, 5033 Deoptimization::Action_make_not_entrant); 5034 assert(stopped(), "Should be stopped"); 5035 } 5036 { 5037 PreserveJVMState pjvms(this); 5038 set_control(_gvn.transform(slow_region)); 5039 uncommon_trap(Deoptimization::Reason_intrinsic, 5040 Deoptimization::Action_make_not_entrant); 5041 assert(stopped(), "Should be stopped"); 5042 } 5043 } 5044 5045 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5046 5047 if (stopped()) { 5048 return true; 5049 } 5050 5051 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5052 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5053 // so the compiler has a chance to eliminate them: during macro expansion, 5054 // we have to set their control (CastPP nodes are eliminated). 5055 load_object_klass(src), load_object_klass(dest), 5056 load_array_length(src), load_array_length(dest)); 5057 5058 ac->set_arraycopy(validated); 5059 5060 Node* n = _gvn.transform(ac); 5061 if (n == ac) { 5062 ac->connect_outputs(this); 5063 } else { 5064 assert(validated, "shouldn't transform if all arguments not validated"); 5065 set_all_memory(n); 5066 } 5067 5068 return true; 5069 } 5070 5071 5072 // Helper function which determines if an arraycopy immediately follows 5073 // an allocation, with no intervening tests or other escapes for the object. 5074 AllocateArrayNode* 5075 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5076 RegionNode* slow_region) { 5077 if (stopped()) return NULL; // no fast path 5078 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5079 5080 ptr = ShenandoahBarrierNode::skip_through_barrier(ptr); 5081 5082 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5083 if (alloc == NULL) return NULL; 5084 5085 Node* rawmem = memory(Compile::AliasIdxRaw); 5086 // Is the allocation's memory state untouched? 5087 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5088 // Bail out if there have been raw-memory effects since the allocation. 5089 // (Example: There might have been a call or safepoint.) 5090 return NULL; 5091 } 5092 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5093 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5094 return NULL; 5095 } 5096 5097 // There must be no unexpected observers of this allocation. 5098 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5099 Node* obs = ptr->fast_out(i); 5100 if (obs != this->map()) { 5101 return NULL; 5102 } 5103 } 5104 5105 // This arraycopy must unconditionally follow the allocation of the ptr. 5106 Node* alloc_ctl = ptr->in(0); 5107 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5108 5109 Node* ctl = control(); 5110 while (ctl != alloc_ctl) { 5111 // There may be guards which feed into the slow_region. 5112 // Any other control flow means that we might not get a chance 5113 // to finish initializing the allocated object. 5114 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5115 IfNode* iff = ctl->in(0)->as_If(); 5116 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5117 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5118 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5119 ctl = iff->in(0); // This test feeds the known slow_region. 5120 continue; 5121 } 5122 // One more try: Various low-level checks bottom out in 5123 // uncommon traps. If the debug-info of the trap omits 5124 // any reference to the allocation, as we've already 5125 // observed, then there can be no objection to the trap. 5126 bool found_trap = false; 5127 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5128 Node* obs = not_ctl->fast_out(j); 5129 if (obs->in(0) == not_ctl && obs->is_Call() && 5130 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5131 found_trap = true; break; 5132 } 5133 } 5134 if (found_trap) { 5135 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5136 continue; 5137 } 5138 } 5139 return NULL; 5140 } 5141 5142 // If we get this far, we have an allocation which immediately 5143 // precedes the arraycopy, and we can take over zeroing the new object. 5144 // The arraycopy will finish the initialization, and provide 5145 // a new control state to which we will anchor the destination pointer. 5146 5147 return alloc; 5148 } 5149 5150 //-------------inline_encodeISOArray----------------------------------- 5151 // encode char[] to byte[] in ISO_8859_1 5152 bool LibraryCallKit::inline_encodeISOArray() { 5153 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5154 // no receiver since it is static method 5155 Node *src = argument(0); 5156 Node *src_offset = argument(1); 5157 Node *dst = argument(2); 5158 Node *dst_offset = argument(3); 5159 Node *length = argument(4); 5160 5161 src = shenandoah_read_barrier(src); 5162 dst = shenandoah_write_barrier(dst); 5163 5164 const Type* src_type = src->Value(&_gvn); 5165 const Type* dst_type = dst->Value(&_gvn); 5166 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5167 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5168 if (top_src == NULL || top_src->klass() == NULL || 5169 top_dest == NULL || top_dest->klass() == NULL) { 5170 // failed array check 5171 return false; 5172 } 5173 5174 // Figure out the size and type of the elements we will be copying. 5175 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5176 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5177 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5178 return false; 5179 } 5180 Node* src_start = array_element_address(src, src_offset, src_elem); 5181 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5182 // 'src_start' points to src array + scaled offset 5183 // 'dst_start' points to dst array + scaled offset 5184 5185 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5186 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5187 enc = _gvn.transform(enc); 5188 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5189 set_memory(res_mem, mtype); 5190 set_result(enc); 5191 return true; 5192 } 5193 5194 //-------------inline_multiplyToLen----------------------------------- 5195 bool LibraryCallKit::inline_multiplyToLen() { 5196 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5197 5198 address stubAddr = StubRoutines::multiplyToLen(); 5199 if (stubAddr == NULL) { 5200 return false; // Intrinsic's stub is not implemented on this platform 5201 } 5202 const char* stubName = "multiplyToLen"; 5203 5204 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5205 5206 // no receiver because it is a static method 5207 Node* x = argument(0); 5208 Node* xlen = argument(1); 5209 Node* y = argument(2); 5210 Node* ylen = argument(3); 5211 Node* z = argument(4); 5212 5213 x = shenandoah_read_barrier(x); 5214 y = shenandoah_read_barrier(y); 5215 z = shenandoah_write_barrier(z); 5216 5217 const Type* x_type = x->Value(&_gvn); 5218 const Type* y_type = y->Value(&_gvn); 5219 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5220 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5221 if (top_x == NULL || top_x->klass() == NULL || 5222 top_y == NULL || top_y->klass() == NULL) { 5223 // failed array check 5224 return false; 5225 } 5226 5227 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5228 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5229 if (x_elem != T_INT || y_elem != T_INT) { 5230 return false; 5231 } 5232 5233 // Set the original stack and the reexecute bit for the interpreter to reexecute 5234 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5235 // on the return from z array allocation in runtime. 5236 { PreserveReexecuteState preexecs(this); 5237 jvms()->set_should_reexecute(true); 5238 5239 Node* x_start = array_element_address(x, intcon(0), x_elem); 5240 Node* y_start = array_element_address(y, intcon(0), y_elem); 5241 // 'x_start' points to x array + scaled xlen 5242 // 'y_start' points to y array + scaled ylen 5243 5244 // Allocate the result array 5245 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5246 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5247 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5248 5249 IdealKit ideal(this); 5250 5251 #define __ ideal. 5252 Node* one = __ ConI(1); 5253 Node* zero = __ ConI(0); 5254 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5255 __ set(need_alloc, zero); 5256 __ set(z_alloc, z); 5257 __ if_then(z, BoolTest::eq, null()); { 5258 __ increment (need_alloc, one); 5259 } __ else_(); { 5260 // Update graphKit memory and control from IdealKit. 5261 sync_kit(ideal); 5262 Node* zlen_arg = load_array_length(z); 5263 // Update IdealKit memory and control from graphKit. 5264 __ sync_kit(this); 5265 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5266 __ increment (need_alloc, one); 5267 } __ end_if(); 5268 } __ end_if(); 5269 5270 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5271 // Update graphKit memory and control from IdealKit. 5272 sync_kit(ideal); 5273 Node * narr = new_array(klass_node, zlen, 1); 5274 // Update IdealKit memory and control from graphKit. 5275 __ sync_kit(this); 5276 __ set(z_alloc, narr); 5277 } __ end_if(); 5278 5279 sync_kit(ideal); 5280 z = __ value(z_alloc); 5281 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5282 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5283 // Final sync IdealKit and GraphKit. 5284 final_sync(ideal); 5285 #undef __ 5286 5287 Node* z_start = array_element_address(z, intcon(0), T_INT); 5288 5289 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5290 OptoRuntime::multiplyToLen_Type(), 5291 stubAddr, stubName, TypePtr::BOTTOM, 5292 x_start, xlen, y_start, ylen, z_start, zlen); 5293 } // original reexecute is set back here 5294 5295 C->set_has_split_ifs(true); // Has chance for split-if optimization 5296 set_result(z); 5297 return true; 5298 } 5299 5300 //-------------inline_squareToLen------------------------------------ 5301 bool LibraryCallKit::inline_squareToLen() { 5302 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5303 5304 address stubAddr = StubRoutines::squareToLen(); 5305 if (stubAddr == NULL) { 5306 return false; // Intrinsic's stub is not implemented on this platform 5307 } 5308 const char* stubName = "squareToLen"; 5309 5310 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5311 5312 Node* x = argument(0); 5313 Node* len = argument(1); 5314 Node* z = argument(2); 5315 Node* zlen = argument(3); 5316 5317 x = shenandoah_read_barrier(x); 5318 z = shenandoah_write_barrier(z); 5319 5320 const Type* x_type = x->Value(&_gvn); 5321 const Type* z_type = z->Value(&_gvn); 5322 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5323 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5324 if (top_x == NULL || top_x->klass() == NULL || 5325 top_z == NULL || top_z->klass() == NULL) { 5326 // failed array check 5327 return false; 5328 } 5329 5330 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5331 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5332 if (x_elem != T_INT || z_elem != T_INT) { 5333 return false; 5334 } 5335 5336 5337 Node* x_start = array_element_address(x, intcon(0), x_elem); 5338 Node* z_start = array_element_address(z, intcon(0), z_elem); 5339 5340 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5341 OptoRuntime::squareToLen_Type(), 5342 stubAddr, stubName, TypePtr::BOTTOM, 5343 x_start, len, z_start, zlen); 5344 5345 set_result(z); 5346 return true; 5347 } 5348 5349 //-------------inline_mulAdd------------------------------------------ 5350 bool LibraryCallKit::inline_mulAdd() { 5351 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5352 5353 address stubAddr = StubRoutines::mulAdd(); 5354 if (stubAddr == NULL) { 5355 return false; // Intrinsic's stub is not implemented on this platform 5356 } 5357 const char* stubName = "mulAdd"; 5358 5359 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5360 5361 Node* out = argument(0); 5362 Node* in = argument(1); 5363 Node* offset = argument(2); 5364 Node* len = argument(3); 5365 Node* k = argument(4); 5366 5367 in = shenandoah_read_barrier(in); 5368 out = shenandoah_write_barrier(out); 5369 5370 const Type* out_type = out->Value(&_gvn); 5371 const Type* in_type = in->Value(&_gvn); 5372 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5373 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5374 if (top_out == NULL || top_out->klass() == NULL || 5375 top_in == NULL || top_in->klass() == NULL) { 5376 // failed array check 5377 return false; 5378 } 5379 5380 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5381 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5382 if (out_elem != T_INT || in_elem != T_INT) { 5383 return false; 5384 } 5385 5386 Node* outlen = load_array_length(out); 5387 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5388 Node* out_start = array_element_address(out, intcon(0), out_elem); 5389 Node* in_start = array_element_address(in, intcon(0), in_elem); 5390 5391 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5392 OptoRuntime::mulAdd_Type(), 5393 stubAddr, stubName, TypePtr::BOTTOM, 5394 out_start,in_start, new_offset, len, k); 5395 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5396 set_result(result); 5397 return true; 5398 } 5399 5400 //-------------inline_montgomeryMultiply----------------------------------- 5401 bool LibraryCallKit::inline_montgomeryMultiply() { 5402 address stubAddr = StubRoutines::montgomeryMultiply(); 5403 if (stubAddr == NULL) { 5404 return false; // Intrinsic's stub is not implemented on this platform 5405 } 5406 5407 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5408 const char* stubName = "montgomery_square"; 5409 5410 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5411 5412 Node* a = argument(0); 5413 Node* b = argument(1); 5414 Node* n = argument(2); 5415 Node* len = argument(3); 5416 Node* inv = argument(4); 5417 Node* m = argument(6); 5418 5419 a = shenandoah_read_barrier(a); 5420 b = shenandoah_read_barrier(b); 5421 n = shenandoah_read_barrier(n); 5422 m = shenandoah_write_barrier(m); 5423 5424 const Type* a_type = a->Value(&_gvn); 5425 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5426 const Type* b_type = b->Value(&_gvn); 5427 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5428 const Type* n_type = a->Value(&_gvn); 5429 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5430 const Type* m_type = a->Value(&_gvn); 5431 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5432 if (top_a == NULL || top_a->klass() == NULL || 5433 top_b == NULL || top_b->klass() == NULL || 5434 top_n == NULL || top_n->klass() == NULL || 5435 top_m == NULL || top_m->klass() == NULL) { 5436 // failed array check 5437 return false; 5438 } 5439 5440 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5441 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5442 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5443 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5444 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5445 return false; 5446 } 5447 5448 // Make the call 5449 { 5450 Node* a_start = array_element_address(a, intcon(0), a_elem); 5451 Node* b_start = array_element_address(b, intcon(0), b_elem); 5452 Node* n_start = array_element_address(n, intcon(0), n_elem); 5453 Node* m_start = array_element_address(m, intcon(0), m_elem); 5454 5455 Node* call = make_runtime_call(RC_LEAF, 5456 OptoRuntime::montgomeryMultiply_Type(), 5457 stubAddr, stubName, TypePtr::BOTTOM, 5458 a_start, b_start, n_start, len, inv, top(), 5459 m_start); 5460 set_result(m); 5461 } 5462 5463 return true; 5464 } 5465 5466 bool LibraryCallKit::inline_montgomerySquare() { 5467 address stubAddr = StubRoutines::montgomerySquare(); 5468 if (stubAddr == NULL) { 5469 return false; // Intrinsic's stub is not implemented on this platform 5470 } 5471 5472 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5473 const char* stubName = "montgomery_square"; 5474 5475 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5476 5477 Node* a = argument(0); 5478 Node* n = argument(1); 5479 Node* len = argument(2); 5480 Node* inv = argument(3); 5481 Node* m = argument(5); 5482 5483 a = shenandoah_read_barrier(a); 5484 n = shenandoah_read_barrier(n); 5485 m = shenandoah_write_barrier(m); 5486 5487 const Type* a_type = a->Value(&_gvn); 5488 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5489 const Type* n_type = a->Value(&_gvn); 5490 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5491 const Type* m_type = a->Value(&_gvn); 5492 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5493 if (top_a == NULL || top_a->klass() == NULL || 5494 top_n == NULL || top_n->klass() == NULL || 5495 top_m == NULL || top_m->klass() == NULL) { 5496 // failed array check 5497 return false; 5498 } 5499 5500 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5501 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5502 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5503 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5504 return false; 5505 } 5506 5507 // Make the call 5508 { 5509 Node* a_start = array_element_address(a, intcon(0), a_elem); 5510 Node* n_start = array_element_address(n, intcon(0), n_elem); 5511 Node* m_start = array_element_address(m, intcon(0), m_elem); 5512 5513 Node* call = make_runtime_call(RC_LEAF, 5514 OptoRuntime::montgomerySquare_Type(), 5515 stubAddr, stubName, TypePtr::BOTTOM, 5516 a_start, n_start, len, inv, top(), 5517 m_start); 5518 set_result(m); 5519 } 5520 5521 return true; 5522 } 5523 5524 5525 /** 5526 * Calculate CRC32 for byte. 5527 * int java.util.zip.CRC32.update(int crc, int b) 5528 */ 5529 bool LibraryCallKit::inline_updateCRC32() { 5530 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5531 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5532 // no receiver since it is static method 5533 Node* crc = argument(0); // type: int 5534 Node* b = argument(1); // type: int 5535 5536 /* 5537 * int c = ~ crc; 5538 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5539 * b = b ^ (c >>> 8); 5540 * crc = ~b; 5541 */ 5542 5543 Node* M1 = intcon(-1); 5544 crc = _gvn.transform(new XorINode(crc, M1)); 5545 Node* result = _gvn.transform(new XorINode(crc, b)); 5546 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5547 5548 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5549 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5550 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5551 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5552 5553 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5554 result = _gvn.transform(new XorINode(crc, result)); 5555 result = _gvn.transform(new XorINode(result, M1)); 5556 set_result(result); 5557 return true; 5558 } 5559 5560 /** 5561 * Calculate CRC32 for byte[] array. 5562 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5563 */ 5564 bool LibraryCallKit::inline_updateBytesCRC32() { 5565 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5566 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5567 // no receiver since it is static method 5568 Node* crc = argument(0); // type: int 5569 Node* src = argument(1); // type: oop 5570 Node* offset = argument(2); // type: int 5571 Node* length = argument(3); // type: int 5572 5573 src = shenandoah_read_barrier(src); 5574 5575 const Type* src_type = src->Value(&_gvn); 5576 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5577 if (top_src == NULL || top_src->klass() == NULL) { 5578 // failed array check 5579 return false; 5580 } 5581 5582 // Figure out the size and type of the elements we will be copying. 5583 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5584 if (src_elem != T_BYTE) { 5585 return false; 5586 } 5587 5588 // 'src_start' points to src array + scaled offset 5589 Node* src_start = array_element_address(src, offset, src_elem); 5590 5591 // We assume that range check is done by caller. 5592 // TODO: generate range check (offset+length < src.length) in debug VM. 5593 5594 // Call the stub. 5595 address stubAddr = StubRoutines::updateBytesCRC32(); 5596 const char *stubName = "updateBytesCRC32"; 5597 5598 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5599 stubAddr, stubName, TypePtr::BOTTOM, 5600 crc, src_start, length); 5601 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5602 set_result(result); 5603 return true; 5604 } 5605 5606 /** 5607 * Calculate CRC32 for ByteBuffer. 5608 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5609 */ 5610 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5611 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5612 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5613 // no receiver since it is static method 5614 Node* crc = argument(0); // type: int 5615 Node* src = argument(1); // type: long 5616 Node* offset = argument(3); // type: int 5617 Node* length = argument(4); // type: int 5618 5619 src = ConvL2X(src); // adjust Java long to machine word 5620 Node* base = _gvn.transform(new CastX2PNode(src)); 5621 offset = ConvI2X(offset); 5622 5623 // 'src_start' points to src array + scaled offset 5624 Node* src_start = basic_plus_adr(top(), base, offset); 5625 5626 // Call the stub. 5627 address stubAddr = StubRoutines::updateBytesCRC32(); 5628 const char *stubName = "updateBytesCRC32"; 5629 5630 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5631 stubAddr, stubName, TypePtr::BOTTOM, 5632 crc, src_start, length); 5633 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5634 set_result(result); 5635 return true; 5636 } 5637 5638 //------------------------------get_table_from_crc32c_class----------------------- 5639 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5640 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5641 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5642 5643 return table; 5644 } 5645 5646 //------------------------------inline_updateBytesCRC32C----------------------- 5647 // 5648 // Calculate CRC32C for byte[] array. 5649 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5650 // 5651 bool LibraryCallKit::inline_updateBytesCRC32C() { 5652 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5653 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5654 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5655 // no receiver since it is a static method 5656 Node* crc = argument(0); // type: int 5657 Node* src = argument(1); // type: oop 5658 Node* offset = argument(2); // type: int 5659 Node* end = argument(3); // type: int 5660 5661 Node* length = _gvn.transform(new SubINode(end, offset)); 5662 5663 const Type* src_type = src->Value(&_gvn); 5664 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5665 if (top_src == NULL || top_src->klass() == NULL) { 5666 // failed array check 5667 return false; 5668 } 5669 5670 // Figure out the size and type of the elements we will be copying. 5671 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5672 if (src_elem != T_BYTE) { 5673 return false; 5674 } 5675 5676 // 'src_start' points to src array + scaled offset 5677 src = shenandoah_read_barrier(src); 5678 Node* src_start = array_element_address(src, offset, src_elem); 5679 5680 // static final int[] byteTable in class CRC32C 5681 Node* table = get_table_from_crc32c_class(callee()->holder()); 5682 table = shenandoah_read_barrier(table); 5683 Node* table_start = array_element_address(table, intcon(0), T_INT); 5684 5685 // We assume that range check is done by caller. 5686 // TODO: generate range check (offset+length < src.length) in debug VM. 5687 5688 // Call the stub. 5689 address stubAddr = StubRoutines::updateBytesCRC32C(); 5690 const char *stubName = "updateBytesCRC32C"; 5691 5692 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5693 stubAddr, stubName, TypePtr::BOTTOM, 5694 crc, src_start, length, table_start); 5695 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5696 set_result(result); 5697 return true; 5698 } 5699 5700 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5701 // 5702 // Calculate CRC32C for DirectByteBuffer. 5703 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5704 // 5705 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5706 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5707 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5708 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5709 // no receiver since it is a static method 5710 Node* crc = argument(0); // type: int 5711 Node* src = argument(1); // type: long 5712 Node* offset = argument(3); // type: int 5713 Node* end = argument(4); // type: int 5714 5715 Node* length = _gvn.transform(new SubINode(end, offset)); 5716 5717 src = ConvL2X(src); // adjust Java long to machine word 5718 Node* base = _gvn.transform(new CastX2PNode(src)); 5719 offset = ConvI2X(offset); 5720 5721 // 'src_start' points to src array + scaled offset 5722 Node* src_start = basic_plus_adr(top(), base, offset); 5723 5724 // static final int[] byteTable in class CRC32C 5725 Node* table = get_table_from_crc32c_class(callee()->holder()); 5726 table = shenandoah_read_barrier(table); 5727 Node* table_start = array_element_address(table, intcon(0), T_INT); 5728 5729 // Call the stub. 5730 address stubAddr = StubRoutines::updateBytesCRC32C(); 5731 const char *stubName = "updateBytesCRC32C"; 5732 5733 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5734 stubAddr, stubName, TypePtr::BOTTOM, 5735 crc, src_start, length, table_start); 5736 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5737 set_result(result); 5738 return true; 5739 } 5740 5741 //------------------------------inline_updateBytesAdler32---------------------- 5742 // 5743 // Calculate Adler32 checksum for byte[] array. 5744 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5745 // 5746 bool LibraryCallKit::inline_updateBytesAdler32() { 5747 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5748 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5749 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5750 // no receiver since it is static method 5751 Node* crc = argument(0); // type: int 5752 Node* src = argument(1); // type: oop 5753 Node* offset = argument(2); // type: int 5754 Node* length = argument(3); // type: int 5755 5756 const Type* src_type = src->Value(&_gvn); 5757 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5758 if (top_src == NULL || top_src->klass() == NULL) { 5759 // failed array check 5760 return false; 5761 } 5762 5763 // Figure out the size and type of the elements we will be copying. 5764 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5765 if (src_elem != T_BYTE) { 5766 return false; 5767 } 5768 5769 // 'src_start' points to src array + scaled offset 5770 src = shenandoah_read_barrier(src); 5771 Node* src_start = array_element_address(src, offset, src_elem); 5772 5773 // We assume that range check is done by caller. 5774 // TODO: generate range check (offset+length < src.length) in debug VM. 5775 5776 // Call the stub. 5777 address stubAddr = StubRoutines::updateBytesAdler32(); 5778 const char *stubName = "updateBytesAdler32"; 5779 5780 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5781 stubAddr, stubName, TypePtr::BOTTOM, 5782 crc, src_start, length); 5783 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5784 set_result(result); 5785 return true; 5786 } 5787 5788 //------------------------------inline_updateByteBufferAdler32--------------- 5789 // 5790 // Calculate Adler32 checksum for DirectByteBuffer. 5791 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5792 // 5793 bool LibraryCallKit::inline_updateByteBufferAdler32() { 5794 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5795 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5796 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5797 // no receiver since it is static method 5798 Node* crc = argument(0); // type: int 5799 Node* src = argument(1); // type: long 5800 Node* offset = argument(3); // type: int 5801 Node* length = argument(4); // type: int 5802 5803 src = ConvL2X(src); // adjust Java long to machine word 5804 Node* base = _gvn.transform(new CastX2PNode(src)); 5805 offset = ConvI2X(offset); 5806 5807 // 'src_start' points to src array + scaled offset 5808 Node* src_start = basic_plus_adr(top(), base, offset); 5809 5810 // Call the stub. 5811 address stubAddr = StubRoutines::updateBytesAdler32(); 5812 const char *stubName = "updateBytesAdler32"; 5813 5814 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5815 stubAddr, stubName, TypePtr::BOTTOM, 5816 crc, src_start, length); 5817 5818 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5819 set_result(result); 5820 return true; 5821 } 5822 5823 //----------------------------inline_reference_get---------------------------- 5824 // public T java.lang.ref.Reference.get(); 5825 bool LibraryCallKit::inline_reference_get() { 5826 const int referent_offset = java_lang_ref_Reference::referent_offset; 5827 guarantee(referent_offset > 0, "should have already been set"); 5828 5829 // Get the argument: 5830 Node* reference_obj = null_check_receiver(); 5831 if (stopped()) return true; 5832 5833 if (ShenandoahVerifyReadsToFromSpace) { 5834 reference_obj = shenandoah_read_barrier(reference_obj); 5835 } 5836 5837 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5838 5839 ciInstanceKlass* klass = env()->Object_klass(); 5840 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5841 5842 Node* no_ctrl = NULL; 5843 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5844 5845 // Use the pre-barrier to record the value in the referent field 5846 pre_barrier(false /* do_load */, 5847 control(), 5848 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5849 result /* pre_val */, 5850 T_OBJECT); 5851 5852 // Add memory barrier to prevent commoning reads from this field 5853 // across safepoint since GC can change its value. 5854 insert_mem_bar(Op_MemBarCPUOrder); 5855 5856 set_result(result); 5857 return true; 5858 } 5859 5860 5861 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5862 bool is_exact=true, bool is_static=false, 5863 ciInstanceKlass * fromKls=NULL) { 5864 if (fromKls == NULL) { 5865 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5866 assert(tinst != NULL, "obj is null"); 5867 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5868 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5869 fromKls = tinst->klass()->as_instance_klass(); 5870 } else { 5871 assert(is_static, "only for static field access"); 5872 } 5873 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5874 ciSymbol::make(fieldTypeString), 5875 is_static); 5876 5877 assert (field != NULL, "undefined field"); 5878 if (field == NULL) return (Node *) NULL; 5879 5880 if (is_static) { 5881 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5882 fromObj = makecon(tip); 5883 } 5884 5885 fromObj = shenandoah_read_barrier(fromObj); 5886 5887 // Next code copied from Parse::do_get_xxx(): 5888 5889 // Compute address and memory type. 5890 int offset = field->offset_in_bytes(); 5891 bool is_vol = field->is_volatile(); 5892 ciType* field_klass = field->type(); 5893 assert(field_klass->is_loaded(), "should be loaded"); 5894 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5895 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5896 BasicType bt = field->layout_type(); 5897 5898 // Build the resultant type of the load 5899 const Type *type; 5900 if (bt == T_OBJECT) { 5901 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5902 } else { 5903 type = Type::get_const_basic_type(bt); 5904 } 5905 5906 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5907 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5908 } 5909 // Build the load. 5910 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5911 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5912 // If reference is volatile, prevent following memory ops from 5913 // floating up past the volatile read. Also prevents commoning 5914 // another volatile read. 5915 if (is_vol) { 5916 // Memory barrier includes bogus read of value to force load BEFORE membar 5917 insert_mem_bar(Op_MemBarAcquire, loadedField); 5918 } 5919 return loadedField; 5920 } 5921 5922 5923 //------------------------------inline_aescrypt_Block----------------------- 5924 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5925 address stubAddr; 5926 const char *stubName; 5927 assert(UseAES, "need AES instruction support"); 5928 5929 switch(id) { 5930 case vmIntrinsics::_aescrypt_encryptBlock: 5931 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5932 stubName = "aescrypt_encryptBlock"; 5933 break; 5934 case vmIntrinsics::_aescrypt_decryptBlock: 5935 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5936 stubName = "aescrypt_decryptBlock"; 5937 break; 5938 } 5939 if (stubAddr == NULL) return false; 5940 5941 Node* aescrypt_object = argument(0); 5942 Node* src = argument(1); 5943 Node* src_offset = argument(2); 5944 Node* dest = argument(3); 5945 Node* dest_offset = argument(4); 5946 5947 // Resolve src and dest arrays for ShenandoahGC. 5948 src = shenandoah_read_barrier(src); 5949 dest = shenandoah_write_barrier(dest); 5950 5951 // (1) src and dest are arrays. 5952 const Type* src_type = src->Value(&_gvn); 5953 const Type* dest_type = dest->Value(&_gvn); 5954 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5955 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5956 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5957 5958 // for the quick and dirty code we will skip all the checks. 5959 // we are just trying to get the call to be generated. 5960 Node* src_start = src; 5961 Node* dest_start = dest; 5962 if (src_offset != NULL || dest_offset != NULL) { 5963 assert(src_offset != NULL && dest_offset != NULL, ""); 5964 src_start = array_element_address(src, src_offset, T_BYTE); 5965 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5966 } 5967 5968 // now need to get the start of its expanded key array 5969 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5970 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5971 if (k_start == NULL) return false; 5972 5973 if (Matcher::pass_original_key_for_aes()) { 5974 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5975 // compatibility issues between Java key expansion and SPARC crypto instructions 5976 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5977 if (original_k_start == NULL) return false; 5978 5979 // Call the stub. 5980 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5981 stubAddr, stubName, TypePtr::BOTTOM, 5982 src_start, dest_start, k_start, original_k_start); 5983 } else { 5984 // Call the stub. 5985 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5986 stubAddr, stubName, TypePtr::BOTTOM, 5987 src_start, dest_start, k_start); 5988 } 5989 5990 return true; 5991 } 5992 5993 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5994 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5995 address stubAddr; 5996 const char *stubName; 5997 5998 assert(UseAES, "need AES instruction support"); 5999 6000 switch(id) { 6001 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6002 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6003 stubName = "cipherBlockChaining_encryptAESCrypt"; 6004 break; 6005 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6006 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6007 stubName = "cipherBlockChaining_decryptAESCrypt"; 6008 break; 6009 } 6010 if (stubAddr == NULL) return false; 6011 6012 Node* cipherBlockChaining_object = argument(0); 6013 Node* src = argument(1); 6014 Node* src_offset = argument(2); 6015 Node* len = argument(3); 6016 Node* dest = argument(4); 6017 Node* dest_offset = argument(5); 6018 6019 // Resolve src and dest arrays for ShenandoahGC. 6020 src = shenandoah_read_barrier(src); 6021 dest = shenandoah_write_barrier(dest); 6022 6023 // (1) src and dest are arrays. 6024 const Type* src_type = src->Value(&_gvn); 6025 const Type* dest_type = dest->Value(&_gvn); 6026 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6027 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6028 assert (top_src != NULL && top_src->klass() != NULL 6029 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6030 6031 // checks are the responsibility of the caller 6032 Node* src_start = src; 6033 Node* dest_start = dest; 6034 if (src_offset != NULL || dest_offset != NULL) { 6035 assert(src_offset != NULL && dest_offset != NULL, ""); 6036 src_start = array_element_address(src, src_offset, T_BYTE); 6037 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6038 } 6039 6040 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6041 // (because of the predicated logic executed earlier). 6042 // so we cast it here safely. 6043 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6044 6045 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6046 if (embeddedCipherObj == NULL) return false; 6047 6048 // cast it to what we know it will be at runtime 6049 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6050 assert(tinst != NULL, "CBC obj is null"); 6051 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6052 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6053 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6054 6055 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6056 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6057 const TypeOopPtr* xtype = aklass->as_instance_type(); 6058 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6059 aescrypt_object = _gvn.transform(aescrypt_object); 6060 6061 // we need to get the start of the aescrypt_object's expanded key array 6062 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6063 if (k_start == NULL) return false; 6064 6065 // similarly, get the start address of the r vector 6066 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6067 6068 objRvec = shenandoah_write_barrier(objRvec); 6069 6070 if (objRvec == NULL) return false; 6071 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6072 6073 Node* cbcCrypt; 6074 if (Matcher::pass_original_key_for_aes()) { 6075 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6076 // compatibility issues between Java key expansion and SPARC crypto instructions 6077 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6078 if (original_k_start == NULL) return false; 6079 6080 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6081 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6082 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6083 stubAddr, stubName, TypePtr::BOTTOM, 6084 src_start, dest_start, k_start, r_start, len, original_k_start); 6085 } else { 6086 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6087 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6088 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6089 stubAddr, stubName, TypePtr::BOTTOM, 6090 src_start, dest_start, k_start, r_start, len); 6091 } 6092 6093 // return cipher length (int) 6094 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6095 set_result(retvalue); 6096 return true; 6097 } 6098 6099 //------------------------------get_key_start_from_aescrypt_object----------------------- 6100 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6101 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6102 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6103 if (objAESCryptKey == NULL) return (Node *) NULL; 6104 6105 objAESCryptKey = shenandoah_read_barrier(objAESCryptKey); 6106 6107 // now have the array, need to get the start address of the K array 6108 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6109 return k_start; 6110 } 6111 6112 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6113 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6114 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6115 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6116 if (objAESCryptKey == NULL) return (Node *) NULL; 6117 6118 // now have the array, need to get the start address of the lastKey array 6119 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6120 return original_k_start; 6121 } 6122 6123 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6124 // Return node representing slow path of predicate check. 6125 // the pseudo code we want to emulate with this predicate is: 6126 // for encryption: 6127 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6128 // for decryption: 6129 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6130 // note cipher==plain is more conservative than the original java code but that's OK 6131 // 6132 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6133 // The receiver was checked for NULL already. 6134 Node* objCBC = argument(0); 6135 6136 // Load embeddedCipher field of CipherBlockChaining object. 6137 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6138 6139 // get AESCrypt klass for instanceOf check 6140 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6141 // will have same classloader as CipherBlockChaining object 6142 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6143 assert(tinst != NULL, "CBCobj is null"); 6144 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6145 6146 // we want to do an instanceof comparison against the AESCrypt class 6147 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6148 if (!klass_AESCrypt->is_loaded()) { 6149 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6150 Node* ctrl = control(); 6151 set_control(top()); // no regular fast path 6152 return ctrl; 6153 } 6154 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6155 6156 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6157 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6158 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6159 6160 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6161 6162 // for encryption, we are done 6163 if (!decrypting) 6164 return instof_false; // even if it is NULL 6165 6166 // for decryption, we need to add a further check to avoid 6167 // taking the intrinsic path when cipher and plain are the same 6168 // see the original java code for why. 6169 RegionNode* region = new RegionNode(3); 6170 region->init_req(1, instof_false); 6171 Node* src = argument(1); 6172 Node* dest = argument(4); 6173 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6174 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6175 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6176 region->init_req(2, src_dest_conjoint); 6177 6178 record_for_igvn(region); 6179 return _gvn.transform(region); 6180 } 6181 6182 //------------------------------inline_ghash_processBlocks 6183 bool LibraryCallKit::inline_ghash_processBlocks() { 6184 address stubAddr; 6185 const char *stubName; 6186 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6187 6188 stubAddr = StubRoutines::ghash_processBlocks(); 6189 stubName = "ghash_processBlocks"; 6190 6191 Node* data = argument(0); 6192 Node* offset = argument(1); 6193 Node* len = argument(2); 6194 Node* state = argument(3); 6195 Node* subkeyH = argument(4); 6196 6197 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6198 assert(state_start, "state is NULL"); 6199 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6200 assert(subkeyH_start, "subkeyH is NULL"); 6201 Node* data_start = array_element_address(data, offset, T_BYTE); 6202 assert(data_start, "data is NULL"); 6203 6204 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6205 OptoRuntime::ghash_processBlocks_Type(), 6206 stubAddr, stubName, TypePtr::BOTTOM, 6207 state_start, subkeyH_start, data_start, len); 6208 return true; 6209 } 6210 6211 //------------------------------inline_sha_implCompress----------------------- 6212 // 6213 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6214 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6215 // 6216 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6217 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6218 // 6219 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6220 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6221 // 6222 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6223 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6224 6225 Node* sha_obj = argument(0); 6226 Node* src = argument(1); // type oop 6227 Node* ofs = argument(2); // type int 6228 6229 const Type* src_type = src->Value(&_gvn); 6230 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6231 if (top_src == NULL || top_src->klass() == NULL) { 6232 // failed array check 6233 return false; 6234 } 6235 // Figure out the size and type of the elements we will be copying. 6236 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6237 if (src_elem != T_BYTE) { 6238 return false; 6239 } 6240 // 'src_start' points to src array + offset 6241 Node* src_start = array_element_address(src, ofs, src_elem); 6242 Node* state = NULL; 6243 address stubAddr; 6244 const char *stubName; 6245 6246 switch(id) { 6247 case vmIntrinsics::_sha_implCompress: 6248 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6249 state = get_state_from_sha_object(sha_obj); 6250 stubAddr = StubRoutines::sha1_implCompress(); 6251 stubName = "sha1_implCompress"; 6252 break; 6253 case vmIntrinsics::_sha2_implCompress: 6254 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6255 state = get_state_from_sha_object(sha_obj); 6256 stubAddr = StubRoutines::sha256_implCompress(); 6257 stubName = "sha256_implCompress"; 6258 break; 6259 case vmIntrinsics::_sha5_implCompress: 6260 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6261 state = get_state_from_sha5_object(sha_obj); 6262 stubAddr = StubRoutines::sha512_implCompress(); 6263 stubName = "sha512_implCompress"; 6264 break; 6265 default: 6266 fatal_unexpected_iid(id); 6267 return false; 6268 } 6269 if (state == NULL) return false; 6270 6271 // Call the stub. 6272 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6273 stubAddr, stubName, TypePtr::BOTTOM, 6274 src_start, state); 6275 6276 return true; 6277 } 6278 6279 //------------------------------inline_digestBase_implCompressMB----------------------- 6280 // 6281 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6282 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6283 // 6284 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6285 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6286 "need SHA1/SHA256/SHA512 instruction support"); 6287 assert((uint)predicate < 3, "sanity"); 6288 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6289 6290 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6291 Node* src = argument(1); // byte[] array 6292 Node* ofs = argument(2); // type int 6293 Node* limit = argument(3); // type int 6294 6295 const Type* src_type = src->Value(&_gvn); 6296 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6297 if (top_src == NULL || top_src->klass() == NULL) { 6298 // failed array check 6299 return false; 6300 } 6301 // Figure out the size and type of the elements we will be copying. 6302 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6303 if (src_elem != T_BYTE) { 6304 return false; 6305 } 6306 // 'src_start' points to src array + offset 6307 Node* src_start = array_element_address(src, ofs, src_elem); 6308 6309 const char* klass_SHA_name = NULL; 6310 const char* stub_name = NULL; 6311 address stub_addr = NULL; 6312 bool long_state = false; 6313 6314 switch (predicate) { 6315 case 0: 6316 if (UseSHA1Intrinsics) { 6317 klass_SHA_name = "sun/security/provider/SHA"; 6318 stub_name = "sha1_implCompressMB"; 6319 stub_addr = StubRoutines::sha1_implCompressMB(); 6320 } 6321 break; 6322 case 1: 6323 if (UseSHA256Intrinsics) { 6324 klass_SHA_name = "sun/security/provider/SHA2"; 6325 stub_name = "sha256_implCompressMB"; 6326 stub_addr = StubRoutines::sha256_implCompressMB(); 6327 } 6328 break; 6329 case 2: 6330 if (UseSHA512Intrinsics) { 6331 klass_SHA_name = "sun/security/provider/SHA5"; 6332 stub_name = "sha512_implCompressMB"; 6333 stub_addr = StubRoutines::sha512_implCompressMB(); 6334 long_state = true; 6335 } 6336 break; 6337 default: 6338 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6339 } 6340 if (klass_SHA_name != NULL) { 6341 // get DigestBase klass to lookup for SHA klass 6342 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6343 assert(tinst != NULL, "digestBase_obj is not instance???"); 6344 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6345 6346 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6347 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6348 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6349 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6350 } 6351 return false; 6352 } 6353 //------------------------------inline_sha_implCompressMB----------------------- 6354 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6355 bool long_state, address stubAddr, const char *stubName, 6356 Node* src_start, Node* ofs, Node* limit) { 6357 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6358 const TypeOopPtr* xtype = aklass->as_instance_type(); 6359 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6360 sha_obj = _gvn.transform(sha_obj); 6361 6362 Node* state; 6363 if (long_state) { 6364 state = get_state_from_sha5_object(sha_obj); 6365 } else { 6366 state = get_state_from_sha_object(sha_obj); 6367 } 6368 if (state == NULL) return false; 6369 6370 // Call the stub. 6371 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6372 OptoRuntime::digestBase_implCompressMB_Type(), 6373 stubAddr, stubName, TypePtr::BOTTOM, 6374 src_start, state, ofs, limit); 6375 // return ofs (int) 6376 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6377 set_result(result); 6378 6379 return true; 6380 } 6381 6382 //------------------------------get_state_from_sha_object----------------------- 6383 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6384 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6385 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6386 if (sha_state == NULL) return (Node *) NULL; 6387 6388 // now have the array, need to get the start address of the state array 6389 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6390 return state; 6391 } 6392 6393 //------------------------------get_state_from_sha5_object----------------------- 6394 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6395 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6396 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6397 if (sha_state == NULL) return (Node *) NULL; 6398 6399 // now have the array, need to get the start address of the state array 6400 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6401 return state; 6402 } 6403 6404 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6405 // Return node representing slow path of predicate check. 6406 // the pseudo code we want to emulate with this predicate is: 6407 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6408 // 6409 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6410 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6411 "need SHA1/SHA256/SHA512 instruction support"); 6412 assert((uint)predicate < 3, "sanity"); 6413 6414 // The receiver was checked for NULL already. 6415 Node* digestBaseObj = argument(0); 6416 6417 // get DigestBase klass for instanceOf check 6418 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6419 assert(tinst != NULL, "digestBaseObj is null"); 6420 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6421 6422 const char* klass_SHA_name = NULL; 6423 switch (predicate) { 6424 case 0: 6425 if (UseSHA1Intrinsics) { 6426 // we want to do an instanceof comparison against the SHA class 6427 klass_SHA_name = "sun/security/provider/SHA"; 6428 } 6429 break; 6430 case 1: 6431 if (UseSHA256Intrinsics) { 6432 // we want to do an instanceof comparison against the SHA2 class 6433 klass_SHA_name = "sun/security/provider/SHA2"; 6434 } 6435 break; 6436 case 2: 6437 if (UseSHA512Intrinsics) { 6438 // we want to do an instanceof comparison against the SHA5 class 6439 klass_SHA_name = "sun/security/provider/SHA5"; 6440 } 6441 break; 6442 default: 6443 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6444 } 6445 6446 ciKlass* klass_SHA = NULL; 6447 if (klass_SHA_name != NULL) { 6448 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6449 } 6450 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6451 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6452 Node* ctrl = control(); 6453 set_control(top()); // no intrinsic path 6454 return ctrl; 6455 } 6456 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6457 6458 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6459 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6460 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6461 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6462 6463 return instof_false; // even if it is NULL 6464 } 6465 6466 bool LibraryCallKit::inline_profileBoolean() { 6467 Node* counts = argument(1); 6468 const TypeAryPtr* ary = NULL; 6469 ciArray* aobj = NULL; 6470 if (counts->is_Con() 6471 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6472 && (aobj = ary->const_oop()->as_array()) != NULL 6473 && (aobj->length() == 2)) { 6474 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6475 jint false_cnt = aobj->element_value(0).as_int(); 6476 jint true_cnt = aobj->element_value(1).as_int(); 6477 6478 if (C->log() != NULL) { 6479 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6480 false_cnt, true_cnt); 6481 } 6482 6483 if (false_cnt + true_cnt == 0) { 6484 // According to profile, never executed. 6485 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6486 Deoptimization::Action_reinterpret); 6487 return true; 6488 } 6489 6490 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6491 // is a number of each value occurrences. 6492 Node* result = argument(0); 6493 if (false_cnt == 0 || true_cnt == 0) { 6494 // According to profile, one value has been never seen. 6495 int expected_val = (false_cnt == 0) ? 1 : 0; 6496 6497 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6498 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6499 6500 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6501 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6502 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6503 6504 { // Slow path: uncommon trap for never seen value and then reexecute 6505 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6506 // the value has been seen at least once. 6507 PreserveJVMState pjvms(this); 6508 PreserveReexecuteState preexecs(this); 6509 jvms()->set_should_reexecute(true); 6510 6511 set_control(slow_path); 6512 set_i_o(i_o()); 6513 6514 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6515 Deoptimization::Action_reinterpret); 6516 } 6517 // The guard for never seen value enables sharpening of the result and 6518 // returning a constant. It allows to eliminate branches on the same value 6519 // later on. 6520 set_control(fast_path); 6521 result = intcon(expected_val); 6522 } 6523 // Stop profiling. 6524 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6525 // By replacing method body with profile data (represented as ProfileBooleanNode 6526 // on IR level) we effectively disable profiling. 6527 // It enables full speed execution once optimized code is generated. 6528 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6529 C->record_for_igvn(profile); 6530 set_result(profile); 6531 return true; 6532 } else { 6533 // Continue profiling. 6534 // Profile data isn't available at the moment. So, execute method's bytecode version. 6535 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6536 // is compiled and counters aren't available since corresponding MethodHandle 6537 // isn't a compile-time constant. 6538 return false; 6539 } 6540 } 6541 6542 bool LibraryCallKit::inline_isCompileConstant() { 6543 Node* n = argument(0); 6544 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6545 return true; 6546 }