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 "oops/objArrayKlass.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/c2compiler.hpp" 35 #include "opto/callGenerator.hpp" 36 #include "opto/castnode.hpp" 37 #include "opto/cfgnode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/countbitsnode.hpp" 40 #include "opto/intrinsicnode.hpp" 41 #include "opto/idealKit.hpp" 42 #include "opto/mathexactnode.hpp" 43 #include "opto/movenode.hpp" 44 #include "opto/mulnode.hpp" 45 #include "opto/narrowptrnode.hpp" 46 #include "opto/opaquenode.hpp" 47 #include "opto/parse.hpp" 48 #include "opto/runtime.hpp" 49 #include "opto/subnode.hpp" 50 #include "prims/nativeLookup.hpp" 51 #include "runtime/sharedRuntime.hpp" 52 #include "trace/traceMacros.hpp" 53 54 class LibraryIntrinsic : public InlineCallGenerator { 55 // Extend the set of intrinsics known to the runtime: 56 public: 57 private: 58 bool _is_virtual; 59 bool _does_virtual_dispatch; 60 int8_t _predicates_count; // Intrinsic is predicated by several conditions 61 int8_t _last_predicate; // Last generated predicate 62 vmIntrinsics::ID _intrinsic_id; 63 64 public: 65 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 66 : InlineCallGenerator(m), 67 _is_virtual(is_virtual), 68 _does_virtual_dispatch(does_virtual_dispatch), 69 _predicates_count((int8_t)predicates_count), 70 _last_predicate((int8_t)-1), 71 _intrinsic_id(id) 72 { 73 } 74 virtual bool is_intrinsic() const { return true; } 75 virtual bool is_virtual() const { return _is_virtual; } 76 virtual bool is_predicated() const { return _predicates_count > 0; } 77 virtual int predicates_count() const { return _predicates_count; } 78 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 79 virtual JVMState* generate(JVMState* jvms); 80 virtual Node* generate_predicate(JVMState* jvms, int predicate); 81 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 82 }; 83 84 85 // Local helper class for LibraryIntrinsic: 86 class LibraryCallKit : public GraphKit { 87 private: 88 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 89 Node* _result; // the result node, if any 90 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 91 92 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 93 94 public: 95 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 96 : GraphKit(jvms), 97 _intrinsic(intrinsic), 98 _result(NULL) 99 { 100 // Check if this is a root compile. In that case we don't have a caller. 101 if (!jvms->has_method()) { 102 _reexecute_sp = sp(); 103 } else { 104 // Find out how many arguments the interpreter needs when deoptimizing 105 // and save the stack pointer value so it can used by uncommon_trap. 106 // We find the argument count by looking at the declared signature. 107 bool ignored_will_link; 108 ciSignature* declared_signature = NULL; 109 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 110 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 111 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 112 } 113 } 114 115 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 116 117 ciMethod* caller() const { return jvms()->method(); } 118 int bci() const { return jvms()->bci(); } 119 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 120 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 121 ciMethod* callee() const { return _intrinsic->method(); } 122 123 bool try_to_inline(int predicate); 124 Node* try_to_predicate(int predicate); 125 126 void push_result() { 127 // Push the result onto the stack. 128 if (!stopped() && result() != NULL) { 129 BasicType bt = result()->bottom_type()->basic_type(); 130 push_node(bt, result()); 131 } 132 } 133 134 private: 135 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 136 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 137 } 138 139 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 140 void set_result(RegionNode* region, PhiNode* value); 141 Node* result() { return _result; } 142 143 virtual int reexecute_sp() { return _reexecute_sp; } 144 145 // Helper functions to inline natives 146 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 147 Node* generate_slow_guard(Node* test, RegionNode* region); 148 Node* generate_fair_guard(Node* test, RegionNode* region); 149 Node* generate_negative_guard(Node* index, RegionNode* region, 150 // resulting CastII of index: 151 Node* *pos_index = NULL); 152 Node* generate_limit_guard(Node* offset, Node* subseq_length, 153 Node* array_length, 154 RegionNode* region); 155 Node* generate_current_thread(Node* &tls_output); 156 Node* load_mirror_from_klass(Node* klass); 157 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 158 RegionNode* region, int null_path, 159 int offset); 160 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 161 RegionNode* region, int null_path) { 162 int offset = java_lang_Class::klass_offset_in_bytes(); 163 return load_klass_from_mirror_common(mirror, never_see_null, 164 region, null_path, 165 offset); 166 } 167 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 168 RegionNode* region, int null_path) { 169 int offset = java_lang_Class::array_klass_offset_in_bytes(); 170 return load_klass_from_mirror_common(mirror, never_see_null, 171 region, null_path, 172 offset); 173 } 174 Node* generate_access_flags_guard(Node* kls, 175 int modifier_mask, int modifier_bits, 176 RegionNode* region); 177 Node* generate_interface_guard(Node* kls, RegionNode* region); 178 Node* generate_array_guard(Node* kls, RegionNode* region) { 179 return generate_array_guard_common(kls, region, false, false); 180 } 181 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 182 return generate_array_guard_common(kls, region, false, true); 183 } 184 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 185 return generate_array_guard_common(kls, region, true, false); 186 } 187 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 188 return generate_array_guard_common(kls, region, true, true); 189 } 190 Node* generate_array_guard_common(Node* kls, RegionNode* region, 191 bool obj_array, bool not_array); 192 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 193 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 194 bool is_virtual = false, bool is_static = false); 195 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 196 return generate_method_call(method_id, false, true); 197 } 198 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 199 return generate_method_call(method_id, true, false); 200 } 201 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 202 203 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 204 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 205 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 206 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 207 bool inline_string_indexOfChar(); 208 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 209 bool inline_string_toBytesU(); 210 bool inline_string_getCharsU(); 211 bool inline_string_copy(bool compress); 212 bool inline_string_char_access(bool is_store); 213 Node* round_double_node(Node* n); 214 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 215 bool inline_math_native(vmIntrinsics::ID id); 216 bool inline_trig(vmIntrinsics::ID id); 217 bool inline_math(vmIntrinsics::ID id); 218 template <typename OverflowOp> 219 bool inline_math_overflow(Node* arg1, Node* arg2); 220 void inline_math_mathExact(Node* math, Node* test); 221 bool inline_math_addExactI(bool is_increment); 222 bool inline_math_addExactL(bool is_increment); 223 bool inline_math_multiplyExactI(); 224 bool inline_math_multiplyExactL(); 225 bool inline_math_negateExactI(); 226 bool inline_math_negateExactL(); 227 bool inline_math_subtractExactI(bool is_decrement); 228 bool inline_math_subtractExactL(bool is_decrement); 229 bool inline_pow(); 230 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 231 bool inline_min_max(vmIntrinsics::ID id); 232 bool inline_notify(vmIntrinsics::ID id); 233 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 234 // This returns Type::AnyPtr, RawPtr, or OopPtr. 235 int classify_unsafe_addr(Node* &base, Node* &offset); 236 Node* make_unsafe_address(Node* base, Node* offset); 237 // Helper for inline_unsafe_access. 238 // Generates the guards that check whether the result of 239 // Unsafe.getObject should be recorded in an SATB log buffer. 240 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 241 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned); 242 static bool klass_needs_init_guard(Node* kls); 243 bool inline_unsafe_allocate(); 244 bool inline_unsafe_copyMemory(); 245 bool inline_native_currentThread(); 246 #ifdef TRACE_HAVE_INTRINSICS 247 bool inline_native_classID(); 248 bool inline_native_threadID(); 249 #endif 250 bool inline_native_time_funcs(address method, const char* funcName); 251 bool inline_native_isInterrupted(); 252 bool inline_native_Class_query(vmIntrinsics::ID id); 253 bool inline_native_subtype_check(); 254 255 bool inline_native_newArray(); 256 bool inline_native_getLength(); 257 bool inline_array_copyOf(bool is_copyOfRange); 258 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 259 bool inline_objects_checkIndex(); 260 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 261 bool inline_native_clone(bool is_virtual); 262 bool inline_native_Reflection_getCallerClass(); 263 // Helper function for inlining native object hash method 264 bool inline_native_hashcode(bool is_virtual, bool is_static); 265 bool inline_native_getClass(); 266 267 // Helper functions for inlining arraycopy 268 bool inline_arraycopy(); 269 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 270 RegionNode* slow_region); 271 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 272 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 273 274 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 275 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 276 bool inline_unsafe_ordered_store(BasicType type); 277 bool inline_unsafe_fence(vmIntrinsics::ID id); 278 bool inline_fp_conversions(vmIntrinsics::ID id); 279 bool inline_number_methods(vmIntrinsics::ID id); 280 bool inline_reference_get(); 281 bool inline_Class_cast(); 282 bool inline_aescrypt_Block(vmIntrinsics::ID id); 283 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 284 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 285 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 286 Node* inline_counterMode_AESCrypt_predicate(); 287 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 288 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 289 bool inline_ghash_processBlocks(); 290 bool inline_sha_implCompress(vmIntrinsics::ID id); 291 bool inline_digestBase_implCompressMB(int predicate); 292 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 293 bool long_state, address stubAddr, const char *stubName, 294 Node* src_start, Node* ofs, Node* limit); 295 Node* get_state_from_sha_object(Node *sha_object); 296 Node* get_state_from_sha5_object(Node *sha_object); 297 Node* inline_digestBase_implCompressMB_predicate(int predicate); 298 bool inline_encodeISOArray(); 299 bool inline_updateCRC32(); 300 bool inline_updateBytesCRC32(); 301 bool inline_updateByteBufferCRC32(); 302 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 303 bool inline_updateBytesCRC32C(); 304 bool inline_updateDirectByteBufferCRC32C(); 305 bool inline_updateBytesAdler32(); 306 bool inline_updateByteBufferAdler32(); 307 bool inline_multiplyToLen(); 308 bool inline_hasNegatives(); 309 bool inline_squareToLen(); 310 bool inline_mulAdd(); 311 bool inline_montgomeryMultiply(); 312 bool inline_montgomerySquare(); 313 314 bool inline_profileBoolean(); 315 bool inline_isCompileConstant(); 316 }; 317 318 //---------------------------make_vm_intrinsic---------------------------- 319 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 320 vmIntrinsics::ID id = m->intrinsic_id(); 321 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 322 323 if (!m->is_loaded()) { 324 // Do not attempt to inline unloaded methods. 325 return NULL; 326 } 327 328 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 329 bool is_available = false; 330 331 { 332 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 333 // the compiler must transition to '_thread_in_vm' state because both 334 // methods access VM-internal data. 335 VM_ENTRY_MARK; 336 methodHandle mh(THREAD, m->get_Method()); 337 is_available = compiler->is_intrinsic_supported(mh, is_virtual) && 338 !C->directive()->is_intrinsic_disabled(mh) && 339 !vmIntrinsics::is_disabled_by_flags(mh); 340 341 } 342 343 if (is_available) { 344 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 345 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 346 return new LibraryIntrinsic(m, is_virtual, 347 vmIntrinsics::predicates_needed(id), 348 vmIntrinsics::does_virtual_dispatch(id), 349 (vmIntrinsics::ID) id); 350 } else { 351 return NULL; 352 } 353 } 354 355 //----------------------register_library_intrinsics----------------------- 356 // Initialize this file's data structures, for each Compile instance. 357 void Compile::register_library_intrinsics() { 358 // Nothing to do here. 359 } 360 361 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 362 LibraryCallKit kit(jvms, this); 363 Compile* C = kit.C; 364 int nodes = C->unique(); 365 #ifndef PRODUCT 366 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 367 char buf[1000]; 368 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 369 tty->print_cr("Intrinsic %s", str); 370 } 371 #endif 372 ciMethod* callee = kit.callee(); 373 const int bci = kit.bci(); 374 375 // Try to inline the intrinsic. 376 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 377 kit.try_to_inline(_last_predicate)) { 378 if (C->print_intrinsics() || C->print_inlining()) { 379 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 380 } 381 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 382 if (C->log()) { 383 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 384 vmIntrinsics::name_at(intrinsic_id()), 385 (is_virtual() ? " virtual='1'" : ""), 386 C->unique() - nodes); 387 } 388 // Push the result from the inlined method onto the stack. 389 kit.push_result(); 390 C->print_inlining_update(this); 391 return kit.transfer_exceptions_into_jvms(); 392 } 393 394 // The intrinsic bailed out 395 if (C->print_intrinsics() || C->print_inlining()) { 396 if (jvms->has_method()) { 397 // Not a root compile. 398 const char* msg; 399 if (callee->intrinsic_candidate()) { 400 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 401 } else { 402 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 403 : "failed to inline (intrinsic), method not annotated"; 404 } 405 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 406 } else { 407 // Root compile 408 tty->print("Did not generate intrinsic %s%s at bci:%d in", 409 vmIntrinsics::name_at(intrinsic_id()), 410 (is_virtual() ? " (virtual)" : ""), bci); 411 } 412 } 413 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 414 C->print_inlining_update(this); 415 return NULL; 416 } 417 418 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 419 LibraryCallKit kit(jvms, this); 420 Compile* C = kit.C; 421 int nodes = C->unique(); 422 _last_predicate = predicate; 423 #ifndef PRODUCT 424 assert(is_predicated() && predicate < predicates_count(), "sanity"); 425 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 426 char buf[1000]; 427 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 428 tty->print_cr("Predicate for intrinsic %s", str); 429 } 430 #endif 431 ciMethod* callee = kit.callee(); 432 const int bci = kit.bci(); 433 434 Node* slow_ctl = kit.try_to_predicate(predicate); 435 if (!kit.failing()) { 436 if (C->print_intrinsics() || C->print_inlining()) { 437 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 438 } 439 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 440 if (C->log()) { 441 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 442 vmIntrinsics::name_at(intrinsic_id()), 443 (is_virtual() ? " virtual='1'" : ""), 444 C->unique() - nodes); 445 } 446 return slow_ctl; // Could be NULL if the check folds. 447 } 448 449 // The intrinsic bailed out 450 if (C->print_intrinsics() || C->print_inlining()) { 451 if (jvms->has_method()) { 452 // Not a root compile. 453 const char* msg = "failed to generate predicate for intrinsic"; 454 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 455 } else { 456 // Root compile 457 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 458 vmIntrinsics::name_at(intrinsic_id()), 459 (is_virtual() ? " (virtual)" : ""), bci); 460 } 461 } 462 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 463 return NULL; 464 } 465 466 bool LibraryCallKit::try_to_inline(int predicate) { 467 // Handle symbolic names for otherwise undistinguished boolean switches: 468 const bool is_store = true; 469 const bool is_compress = true; 470 const bool is_native_ptr = true; 471 const bool is_static = true; 472 const bool is_volatile = true; 473 474 if (!jvms()->has_method()) { 475 // Root JVMState has a null method. 476 assert(map()->memory()->Opcode() == Op_Parm, ""); 477 // Insert the memory aliasing node 478 set_all_memory(reset_memory()); 479 } 480 assert(merged_memory(), ""); 481 482 483 switch (intrinsic_id()) { 484 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 485 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 486 case vmIntrinsics::_getClass: return inline_native_getClass(); 487 488 case vmIntrinsics::_dsin: 489 case vmIntrinsics::_dcos: 490 case vmIntrinsics::_dtan: 491 case vmIntrinsics::_dabs: 492 case vmIntrinsics::_datan2: 493 case vmIntrinsics::_dsqrt: 494 case vmIntrinsics::_dexp: 495 case vmIntrinsics::_dlog: 496 case vmIntrinsics::_dlog10: 497 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 498 499 case vmIntrinsics::_min: 500 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 501 502 case vmIntrinsics::_notify: 503 case vmIntrinsics::_notifyAll: 504 if (InlineNotify) { 505 return inline_notify(intrinsic_id()); 506 } 507 return false; 508 509 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 510 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 511 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 512 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 513 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 514 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 515 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 516 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 517 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 518 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 519 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 520 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 521 522 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 523 524 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 525 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 526 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 527 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 528 529 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 530 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 531 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 532 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 533 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 534 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 535 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 536 537 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 538 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 539 540 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 541 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 542 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 543 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 544 545 case vmIntrinsics::_compressStringC: 546 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 547 case vmIntrinsics::_inflateStringC: 548 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 549 550 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false); 551 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false); 552 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 553 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 554 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 555 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false); 556 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false); 557 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 558 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 559 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false); 560 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false); 561 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false); 562 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false); 563 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false); 564 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false); 565 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false); 566 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 567 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 568 569 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 570 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 571 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 572 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false); 573 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false); 574 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 575 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 576 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false); 577 578 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false); 579 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false); 580 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false); 581 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false); 582 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false); 583 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 584 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 585 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false); 586 587 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false); 588 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false); 589 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false); 590 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false); 591 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false); 592 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false); 593 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false); 594 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false); 595 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false); 596 597 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false); 598 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false); 599 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false); 600 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false); 601 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false); 602 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false); 603 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false); 604 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false); 605 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false); 606 607 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, true); 608 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, true); 609 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, true); 610 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, true); 611 612 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, true); 613 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, true); 614 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, true); 615 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, true); 616 617 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 618 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 619 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 620 621 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 622 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 623 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 624 625 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 626 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 627 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 628 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 629 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 630 631 case vmIntrinsics::_loadFence: 632 case vmIntrinsics::_storeFence: 633 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 634 635 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 636 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 637 638 #ifdef TRACE_HAVE_INTRINSICS 639 case vmIntrinsics::_classID: return inline_native_classID(); 640 case vmIntrinsics::_threadID: return inline_native_threadID(); 641 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 642 #endif 643 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 644 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 645 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 646 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 647 case vmIntrinsics::_newArray: return inline_native_newArray(); 648 case vmIntrinsics::_getLength: return inline_native_getLength(); 649 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 650 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 651 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 652 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 653 case vmIntrinsics::_Objects_checkIndex: return inline_objects_checkIndex(); 654 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 655 656 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 657 658 case vmIntrinsics::_isInstance: 659 case vmIntrinsics::_getModifiers: 660 case vmIntrinsics::_isInterface: 661 case vmIntrinsics::_isArray: 662 case vmIntrinsics::_isPrimitive: 663 case vmIntrinsics::_getSuperclass: 664 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 665 666 case vmIntrinsics::_floatToRawIntBits: 667 case vmIntrinsics::_floatToIntBits: 668 case vmIntrinsics::_intBitsToFloat: 669 case vmIntrinsics::_doubleToRawLongBits: 670 case vmIntrinsics::_doubleToLongBits: 671 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 672 673 case vmIntrinsics::_numberOfLeadingZeros_i: 674 case vmIntrinsics::_numberOfLeadingZeros_l: 675 case vmIntrinsics::_numberOfTrailingZeros_i: 676 case vmIntrinsics::_numberOfTrailingZeros_l: 677 case vmIntrinsics::_bitCount_i: 678 case vmIntrinsics::_bitCount_l: 679 case vmIntrinsics::_reverseBytes_i: 680 case vmIntrinsics::_reverseBytes_l: 681 case vmIntrinsics::_reverseBytes_s: 682 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 683 684 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 685 686 case vmIntrinsics::_Reference_get: return inline_reference_get(); 687 688 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 689 690 case vmIntrinsics::_aescrypt_encryptBlock: 691 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 692 693 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 694 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 695 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 696 697 case vmIntrinsics::_counterMode_AESCrypt: 698 return inline_counterMode_AESCrypt(intrinsic_id()); 699 700 case vmIntrinsics::_sha_implCompress: 701 case vmIntrinsics::_sha2_implCompress: 702 case vmIntrinsics::_sha5_implCompress: 703 return inline_sha_implCompress(intrinsic_id()); 704 705 case vmIntrinsics::_digestBase_implCompressMB: 706 return inline_digestBase_implCompressMB(predicate); 707 708 case vmIntrinsics::_multiplyToLen: 709 return inline_multiplyToLen(); 710 711 case vmIntrinsics::_squareToLen: 712 return inline_squareToLen(); 713 714 case vmIntrinsics::_mulAdd: 715 return inline_mulAdd(); 716 717 case vmIntrinsics::_montgomeryMultiply: 718 return inline_montgomeryMultiply(); 719 case vmIntrinsics::_montgomerySquare: 720 return inline_montgomerySquare(); 721 722 case vmIntrinsics::_ghash_processBlocks: 723 return inline_ghash_processBlocks(); 724 725 case vmIntrinsics::_encodeISOArray: 726 case vmIntrinsics::_encodeByteISOArray: 727 return inline_encodeISOArray(); 728 729 case vmIntrinsics::_updateCRC32: 730 return inline_updateCRC32(); 731 case vmIntrinsics::_updateBytesCRC32: 732 return inline_updateBytesCRC32(); 733 case vmIntrinsics::_updateByteBufferCRC32: 734 return inline_updateByteBufferCRC32(); 735 736 case vmIntrinsics::_updateBytesCRC32C: 737 return inline_updateBytesCRC32C(); 738 case vmIntrinsics::_updateDirectByteBufferCRC32C: 739 return inline_updateDirectByteBufferCRC32C(); 740 741 case vmIntrinsics::_updateBytesAdler32: 742 return inline_updateBytesAdler32(); 743 case vmIntrinsics::_updateByteBufferAdler32: 744 return inline_updateByteBufferAdler32(); 745 746 case vmIntrinsics::_profileBoolean: 747 return inline_profileBoolean(); 748 case vmIntrinsics::_isCompileConstant: 749 return inline_isCompileConstant(); 750 751 case vmIntrinsics::_hasNegatives: 752 return inline_hasNegatives(); 753 754 default: 755 // If you get here, it may be that someone has added a new intrinsic 756 // to the list in vmSymbols.hpp without implementing it here. 757 #ifndef PRODUCT 758 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 759 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 760 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 761 } 762 #endif 763 return false; 764 } 765 } 766 767 Node* LibraryCallKit::try_to_predicate(int predicate) { 768 if (!jvms()->has_method()) { 769 // Root JVMState has a null method. 770 assert(map()->memory()->Opcode() == Op_Parm, ""); 771 // Insert the memory aliasing node 772 set_all_memory(reset_memory()); 773 } 774 assert(merged_memory(), ""); 775 776 switch (intrinsic_id()) { 777 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 778 return inline_cipherBlockChaining_AESCrypt_predicate(false); 779 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 780 return inline_cipherBlockChaining_AESCrypt_predicate(true); 781 case vmIntrinsics::_counterMode_AESCrypt: 782 return inline_counterMode_AESCrypt_predicate(); 783 case vmIntrinsics::_digestBase_implCompressMB: 784 return inline_digestBase_implCompressMB_predicate(predicate); 785 786 default: 787 // If you get here, it may be that someone has added a new intrinsic 788 // to the list in vmSymbols.hpp without implementing it here. 789 #ifndef PRODUCT 790 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 791 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 792 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 793 } 794 #endif 795 Node* slow_ctl = control(); 796 set_control(top()); // No fast path instrinsic 797 return slow_ctl; 798 } 799 } 800 801 //------------------------------set_result------------------------------- 802 // Helper function for finishing intrinsics. 803 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 804 record_for_igvn(region); 805 set_control(_gvn.transform(region)); 806 set_result( _gvn.transform(value)); 807 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 808 } 809 810 //------------------------------generate_guard--------------------------- 811 // Helper function for generating guarded fast-slow graph structures. 812 // The given 'test', if true, guards a slow path. If the test fails 813 // then a fast path can be taken. (We generally hope it fails.) 814 // In all cases, GraphKit::control() is updated to the fast path. 815 // The returned value represents the control for the slow path. 816 // The return value is never 'top'; it is either a valid control 817 // or NULL if it is obvious that the slow path can never be taken. 818 // Also, if region and the slow control are not NULL, the slow edge 819 // is appended to the region. 820 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 821 if (stopped()) { 822 // Already short circuited. 823 return NULL; 824 } 825 826 // Build an if node and its projections. 827 // If test is true we take the slow path, which we assume is uncommon. 828 if (_gvn.type(test) == TypeInt::ZERO) { 829 // The slow branch is never taken. No need to build this guard. 830 return NULL; 831 } 832 833 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 834 835 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 836 if (if_slow == top()) { 837 // The slow branch is never taken. No need to build this guard. 838 return NULL; 839 } 840 841 if (region != NULL) 842 region->add_req(if_slow); 843 844 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 845 set_control(if_fast); 846 847 return if_slow; 848 } 849 850 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 851 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 852 } 853 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 854 return generate_guard(test, region, PROB_FAIR); 855 } 856 857 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 858 Node* *pos_index) { 859 if (stopped()) 860 return NULL; // already stopped 861 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 862 return NULL; // index is already adequately typed 863 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 864 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 865 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 866 if (is_neg != NULL && pos_index != NULL) { 867 // Emulate effect of Parse::adjust_map_after_if. 868 Node* ccast = new CastIINode(index, TypeInt::POS); 869 ccast->set_req(0, control()); 870 (*pos_index) = _gvn.transform(ccast); 871 } 872 return is_neg; 873 } 874 875 // Make sure that 'position' is a valid limit index, in [0..length]. 876 // There are two equivalent plans for checking this: 877 // A. (offset + copyLength) unsigned<= arrayLength 878 // B. offset <= (arrayLength - copyLength) 879 // We require that all of the values above, except for the sum and 880 // difference, are already known to be non-negative. 881 // Plan A is robust in the face of overflow, if offset and copyLength 882 // are both hugely positive. 883 // 884 // Plan B is less direct and intuitive, but it does not overflow at 885 // all, since the difference of two non-negatives is always 886 // representable. Whenever Java methods must perform the equivalent 887 // check they generally use Plan B instead of Plan A. 888 // For the moment we use Plan A. 889 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 890 Node* subseq_length, 891 Node* array_length, 892 RegionNode* region) { 893 if (stopped()) 894 return NULL; // already stopped 895 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 896 if (zero_offset && subseq_length->eqv_uncast(array_length)) 897 return NULL; // common case of whole-array copy 898 Node* last = subseq_length; 899 if (!zero_offset) // last += offset 900 last = _gvn.transform(new AddINode(last, offset)); 901 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 902 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 903 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 904 return is_over; 905 } 906 907 908 //--------------------------generate_current_thread-------------------- 909 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 910 ciKlass* thread_klass = env()->Thread_klass(); 911 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 912 Node* thread = _gvn.transform(new ThreadLocalNode()); 913 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 914 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 915 tls_output = thread; 916 return threadObj; 917 } 918 919 920 //------------------------------make_string_method_node------------------------ 921 // Helper method for String intrinsic functions. This version is called with 922 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 923 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 924 // containing the lengths of str1 and str2. 925 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 926 Node* result = NULL; 927 switch (opcode) { 928 case Op_StrIndexOf: 929 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 930 str1_start, cnt1, str2_start, cnt2, ae); 931 break; 932 case Op_StrComp: 933 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 934 str1_start, cnt1, str2_start, cnt2, ae); 935 break; 936 case Op_StrEquals: 937 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 938 str1_start, str2_start, cnt1, ae); 939 break; 940 default: 941 ShouldNotReachHere(); 942 return NULL; 943 } 944 945 // All these intrinsics have checks. 946 C->set_has_split_ifs(true); // Has chance for split-if optimization 947 948 return _gvn.transform(result); 949 } 950 951 //------------------------------inline_string_compareTo------------------------ 952 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 953 Node* arg1 = argument(0); 954 Node* arg2 = argument(1); 955 956 // Get start addr and length of first argument 957 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 958 Node* arg1_cnt = load_array_length(arg1); 959 960 // Get start addr and length of second argument 961 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 962 Node* arg2_cnt = load_array_length(arg2); 963 964 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 965 set_result(result); 966 return true; 967 } 968 969 //------------------------------inline_string_equals------------------------ 970 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 971 Node* arg1 = argument(0); 972 Node* arg2 = argument(1); 973 974 // paths (plus control) merge 975 RegionNode* region = new RegionNode(3); 976 Node* phi = new PhiNode(region, TypeInt::BOOL); 977 978 if (!stopped()) { 979 // Get start addr and length of first argument 980 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 981 Node* arg1_cnt = load_array_length(arg1); 982 983 // Get start addr and length of second argument 984 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 985 Node* arg2_cnt = load_array_length(arg2); 986 987 // Check for arg1_cnt != arg2_cnt 988 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 989 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 990 Node* if_ne = generate_slow_guard(bol, NULL); 991 if (if_ne != NULL) { 992 phi->init_req(2, intcon(0)); 993 region->init_req(2, if_ne); 994 } 995 996 // Check for count == 0 is done by assembler code for StrEquals. 997 998 if (!stopped()) { 999 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1000 phi->init_req(1, equals); 1001 region->init_req(1, control()); 1002 } 1003 } 1004 1005 // post merge 1006 set_control(_gvn.transform(region)); 1007 record_for_igvn(region); 1008 1009 set_result(_gvn.transform(phi)); 1010 return true; 1011 } 1012 1013 //------------------------------inline_array_equals---------------------------- 1014 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1015 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1016 Node* arg1 = argument(0); 1017 Node* arg2 = argument(1); 1018 1019 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1020 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1021 return true; 1022 } 1023 1024 //------------------------------inline_hasNegatives------------------------------ 1025 bool LibraryCallKit::inline_hasNegatives() { 1026 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1027 1028 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1029 // no receiver since it is static method 1030 Node* ba = argument(0); 1031 Node* offset = argument(1); 1032 Node* len = argument(2); 1033 1034 RegionNode* bailout = new RegionNode(1); 1035 record_for_igvn(bailout); 1036 1037 // offset must not be negative. 1038 generate_negative_guard(offset, bailout); 1039 1040 // offset + length must not exceed length of ba. 1041 generate_limit_guard(offset, len, load_array_length(ba), bailout); 1042 1043 if (bailout->req() > 1) { 1044 PreserveJVMState pjvms(this); 1045 set_control(_gvn.transform(bailout)); 1046 uncommon_trap(Deoptimization::Reason_intrinsic, 1047 Deoptimization::Action_maybe_recompile); 1048 } 1049 if (!stopped()) { 1050 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1051 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1052 set_result(_gvn.transform(result)); 1053 } 1054 return true; 1055 } 1056 1057 bool LibraryCallKit::inline_objects_checkIndex() { 1058 Node* index = argument(0); 1059 Node* length = argument(1); 1060 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1061 return false; 1062 } 1063 1064 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1065 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1066 1067 { 1068 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1069 uncommon_trap(Deoptimization::Reason_intrinsic, 1070 Deoptimization::Action_make_not_entrant); 1071 } 1072 1073 if (stopped()) { 1074 return false; 1075 } 1076 1077 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1078 BoolTest::mask btest = BoolTest::lt; 1079 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1080 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1081 _gvn.set_type(rc, rc->Value(&_gvn)); 1082 if (!rc_bool->is_Con()) { 1083 record_for_igvn(rc); 1084 } 1085 set_control(_gvn.transform(new IfTrueNode(rc))); 1086 { 1087 PreserveJVMState pjvms(this); 1088 set_control(_gvn.transform(new IfFalseNode(rc))); 1089 uncommon_trap(Deoptimization::Reason_range_check, 1090 Deoptimization::Action_make_not_entrant); 1091 } 1092 1093 if (stopped()) { 1094 return false; 1095 } 1096 1097 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1098 result->set_req(0, control()); 1099 result = _gvn.transform(result); 1100 set_result(result); 1101 replace_in_map(index, result); 1102 return true; 1103 } 1104 1105 //------------------------------inline_string_indexOf------------------------ 1106 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1107 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) { 1108 return false; 1109 } 1110 Node* src = argument(0); 1111 Node* tgt = argument(1); 1112 1113 // Make the merge point 1114 RegionNode* result_rgn = new RegionNode(4); 1115 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1116 1117 // Get start addr and length of source string 1118 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1119 Node* src_count = load_array_length(src); 1120 1121 // Get start addr and length of substring 1122 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1123 Node* tgt_count = load_array_length(tgt); 1124 1125 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1126 // Divide src size by 2 if String is UTF16 encoded 1127 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1128 } 1129 if (ae == StrIntrinsicNode::UU) { 1130 // Divide substring size by 2 if String is UTF16 encoded 1131 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1132 } 1133 1134 // Check for substr count > string count 1135 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1136 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1137 Node* if_gt = generate_slow_guard(bol, NULL); 1138 if (if_gt != NULL) { 1139 result_phi->init_req(2, intcon(-1)); 1140 result_rgn->init_req(2, if_gt); 1141 } 1142 1143 if (!stopped()) { 1144 // Check for substr count == 0 1145 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1146 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1147 Node* if_zero = generate_slow_guard(bol, NULL); 1148 if (if_zero != NULL) { 1149 result_phi->init_req(3, intcon(0)); 1150 result_rgn->init_req(3, if_zero); 1151 } 1152 } 1153 1154 if (!stopped()) { 1155 Node* result = make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1156 result_phi->init_req(1, result); 1157 result_rgn->init_req(1, control()); 1158 } 1159 set_control(_gvn.transform(result_rgn)); 1160 record_for_igvn(result_rgn); 1161 set_result(_gvn.transform(result_phi)); 1162 1163 return true; 1164 } 1165 1166 //-----------------------------inline_string_indexOf----------------------- 1167 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1168 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) { 1169 return false; 1170 } 1171 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1172 Node* src = argument(0); // byte[] 1173 Node* src_count = argument(1); 1174 Node* tgt = argument(2); // byte[] 1175 Node* tgt_count = argument(3); 1176 Node* from_index = argument(4); 1177 1178 // Java code which calls this method has range checks for from_index value. 1179 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1180 1181 // Multiply byte array index by 2 if String is UTF16 encoded 1182 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1183 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1184 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1185 1186 Node* result = make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1187 1188 // The result is index relative to from_index if substring was found, -1 otherwise. 1189 // Generate code which will fold into cmove. 1190 RegionNode* region = new RegionNode(3); 1191 Node* phi = new PhiNode(region, TypeInt::INT); 1192 1193 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1194 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1195 1196 Node* if_lt = generate_slow_guard(bol, NULL); 1197 if (if_lt != NULL) { 1198 // result == -1 1199 phi->init_req(2, result); 1200 region->init_req(2, if_lt); 1201 } 1202 if (!stopped()) { 1203 result = _gvn.transform(new AddINode(result, from_index)); 1204 phi->init_req(1, result); 1205 region->init_req(1, control()); 1206 } 1207 1208 set_control(_gvn.transform(region)); 1209 record_for_igvn(region); 1210 set_result(_gvn.transform(phi)); 1211 1212 return true; 1213 } 1214 1215 //-----------------------------inline_string_indexOfChar----------------------- 1216 bool LibraryCallKit::inline_string_indexOfChar() { 1217 if (!Matcher::has_match_rule(Op_StrIndexOfChar) || !(UseSSE > 4)) { 1218 return false; 1219 } 1220 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1221 Node* src = argument(0); // byte[] 1222 Node* tgt = argument(1); // tgt is int ch 1223 Node* from_index = argument(2); 1224 Node* max = argument(3); 1225 1226 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1227 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1228 1229 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1230 1231 RegionNode* region = new RegionNode(3); 1232 Node* phi = new PhiNode(region, TypeInt::INT); 1233 1234 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1235 C->set_has_split_ifs(true); // Has chance for split-if optimization 1236 _gvn.transform(result); 1237 1238 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1239 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1240 1241 Node* if_lt = generate_slow_guard(bol, NULL); 1242 if (if_lt != NULL) { 1243 // result == -1 1244 phi->init_req(2, result); 1245 region->init_req(2, if_lt); 1246 } 1247 if (!stopped()) { 1248 result = _gvn.transform(new AddINode(result, from_index)); 1249 phi->init_req(1, result); 1250 region->init_req(1, control()); 1251 } 1252 set_control(_gvn.transform(region)); 1253 record_for_igvn(region); 1254 set_result(_gvn.transform(phi)); 1255 1256 return true; 1257 } 1258 //---------------------------inline_string_copy--------------------- 1259 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1260 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1261 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1262 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1263 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1264 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1265 bool LibraryCallKit::inline_string_copy(bool compress) { 1266 int nargs = 5; // 2 oops, 3 ints 1267 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1268 1269 Node* src = argument(0); 1270 Node* src_offset = argument(1); 1271 Node* dst = argument(2); 1272 Node* dst_offset = argument(3); 1273 Node* length = argument(4); 1274 1275 // Check for allocation before we add nodes that would confuse 1276 // tightly_coupled_allocation() 1277 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1278 1279 // Figure out the size and type of the elements we will be copying. 1280 const Type* src_type = src->Value(&_gvn); 1281 const Type* dst_type = dst->Value(&_gvn); 1282 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1283 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1284 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1285 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1286 "Unsupported array types for inline_string_copy"); 1287 1288 // Convert char[] offsets to byte[] offsets 1289 if (compress && src_elem == T_BYTE) { 1290 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1291 } else if (!compress && dst_elem == T_BYTE) { 1292 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1293 } 1294 1295 Node* src_start = array_element_address(src, src_offset, src_elem); 1296 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1297 // 'src_start' points to src array + scaled offset 1298 // 'dst_start' points to dst array + scaled offset 1299 Node* count = NULL; 1300 if (compress) { 1301 count = compress_string(src_start, dst_start, length); 1302 } else { 1303 inflate_string(src_start, dst_start, length); 1304 } 1305 1306 if (alloc != NULL) { 1307 if (alloc->maybe_set_complete(&_gvn)) { 1308 // "You break it, you buy it." 1309 InitializeNode* init = alloc->initialization(); 1310 assert(init->is_complete(), "we just did this"); 1311 init->set_complete_with_arraycopy(); 1312 assert(dst->is_CheckCastPP(), "sanity"); 1313 assert(dst->in(0)->in(0) == init, "dest pinned"); 1314 } 1315 // Do not let stores that initialize this object be reordered with 1316 // a subsequent store that would make this object accessible by 1317 // other threads. 1318 // Record what AllocateNode this StoreStore protects so that 1319 // escape analysis can go from the MemBarStoreStoreNode to the 1320 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1321 // based on the escape status of the AllocateNode. 1322 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1323 } 1324 if (compress) { 1325 set_result(_gvn.transform(count)); 1326 } 1327 return true; 1328 } 1329 1330 #ifdef _LP64 1331 #define XTOP ,top() /*additional argument*/ 1332 #else //_LP64 1333 #define XTOP /*no additional argument*/ 1334 #endif //_LP64 1335 1336 //------------------------inline_string_toBytesU-------------------------- 1337 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1338 bool LibraryCallKit::inline_string_toBytesU() { 1339 // Get the arguments. 1340 Node* value = argument(0); 1341 Node* offset = argument(1); 1342 Node* length = argument(2); 1343 1344 Node* newcopy = NULL; 1345 1346 // Set the original stack and the reexecute bit for the interpreter to reexecute 1347 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1348 { PreserveReexecuteState preexecs(this); 1349 jvms()->set_should_reexecute(true); 1350 1351 // Check if a null path was taken unconditionally. 1352 value = null_check(value); 1353 1354 RegionNode* bailout = new RegionNode(1); 1355 record_for_igvn(bailout); 1356 1357 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1358 generate_negative_guard(length, bailout); 1359 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1360 1361 if (bailout->req() > 1) { 1362 PreserveJVMState pjvms(this); 1363 set_control(_gvn.transform(bailout)); 1364 uncommon_trap(Deoptimization::Reason_intrinsic, 1365 Deoptimization::Action_maybe_recompile); 1366 } 1367 if (stopped()) return true; 1368 1369 // Range checks are done by caller. 1370 1371 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1372 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1373 newcopy = new_array(klass_node, size, 0); // no arguments to push 1374 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1375 1376 // Calculate starting addresses. 1377 Node* src_start = array_element_address(value, offset, T_CHAR); 1378 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1379 1380 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1381 const TypeInt* toffset = gvn().type(offset)->is_int(); 1382 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1383 1384 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1385 const char* copyfunc_name = "arraycopy"; 1386 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1387 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1388 OptoRuntime::fast_arraycopy_Type(), 1389 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1390 src_start, dst_start, ConvI2X(length) XTOP); 1391 // Do not let reads from the cloned object float above the arraycopy. 1392 if (alloc != NULL) { 1393 if (alloc->maybe_set_complete(&_gvn)) { 1394 // "You break it, you buy it." 1395 InitializeNode* init = alloc->initialization(); 1396 assert(init->is_complete(), "we just did this"); 1397 init->set_complete_with_arraycopy(); 1398 assert(newcopy->is_CheckCastPP(), "sanity"); 1399 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1400 } 1401 // Do not let stores that initialize this object be reordered with 1402 // a subsequent store that would make this object accessible by 1403 // other threads. 1404 // Record what AllocateNode this StoreStore protects so that 1405 // escape analysis can go from the MemBarStoreStoreNode to the 1406 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1407 // based on the escape status of the AllocateNode. 1408 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1409 } else { 1410 insert_mem_bar(Op_MemBarCPUOrder); 1411 } 1412 } // original reexecute is set back here 1413 1414 C->set_has_split_ifs(true); // Has chance for split-if optimization 1415 if (!stopped()) { 1416 set_result(newcopy); 1417 } 1418 return true; 1419 } 1420 1421 //------------------------inline_string_getCharsU-------------------------- 1422 // public void StringUTF16.getChars(byte[] value, int srcBegin, int srcEnd, char dst[], int dstBegin) 1423 bool LibraryCallKit::inline_string_getCharsU() { 1424 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1425 1426 // Get the arguments. 1427 Node* value = argument(0); 1428 Node* src_begin = argument(1); 1429 Node* src_end = argument(2); // exclusive offset (i < src_end) 1430 Node* dst = argument(3); 1431 Node* dst_begin = argument(4); 1432 1433 // Check for allocation before we add nodes that would confuse 1434 // tightly_coupled_allocation() 1435 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1436 1437 // Check if a null path was taken unconditionally. 1438 value = null_check(value); 1439 dst = null_check(dst); 1440 if (stopped()) { 1441 return true; 1442 } 1443 1444 // Range checks are done by caller. 1445 1446 // Get length and convert char[] offset to byte[] offset 1447 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1448 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1449 1450 if (!stopped()) { 1451 // Calculate starting addresses. 1452 Node* src_start = array_element_address(value, src_begin, T_BYTE); 1453 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1454 1455 // Check if array addresses are aligned to HeapWordSize 1456 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1457 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1458 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1459 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1460 1461 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1462 const char* copyfunc_name = "arraycopy"; 1463 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1464 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1465 OptoRuntime::fast_arraycopy_Type(), 1466 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1467 src_start, dst_start, ConvI2X(length) XTOP); 1468 // Do not let reads from the cloned object float above the arraycopy. 1469 if (alloc != NULL) { 1470 if (alloc->maybe_set_complete(&_gvn)) { 1471 // "You break it, you buy it." 1472 InitializeNode* init = alloc->initialization(); 1473 assert(init->is_complete(), "we just did this"); 1474 init->set_complete_with_arraycopy(); 1475 assert(dst->is_CheckCastPP(), "sanity"); 1476 assert(dst->in(0)->in(0) == init, "dest pinned"); 1477 } 1478 // Do not let stores that initialize this object be reordered with 1479 // a subsequent store that would make this object accessible by 1480 // other threads. 1481 // Record what AllocateNode this StoreStore protects so that 1482 // escape analysis can go from the MemBarStoreStoreNode to the 1483 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1484 // based on the escape status of the AllocateNode. 1485 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1486 } else { 1487 insert_mem_bar(Op_MemBarCPUOrder); 1488 } 1489 } 1490 1491 C->set_has_split_ifs(true); // Has chance for split-if optimization 1492 return true; 1493 } 1494 1495 //----------------------inline_string_char_access---------------------------- 1496 // Store/Load char to/from byte[] array. 1497 // static void StringUTF16.putChar(byte[] val, int index, int c) 1498 // static char StringUTF16.getChar(byte[] val, int index) 1499 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1500 Node* value = argument(0); 1501 Node* index = argument(1); 1502 Node* ch = is_store ? argument(2) : NULL; 1503 1504 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1505 // correctly requires matched array shapes. 1506 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1507 "sanity: byte[] and char[] bases agree"); 1508 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1509 "sanity: byte[] and char[] scales agree"); 1510 1511 Node* adr = array_element_address(value, index, T_CHAR); 1512 if (is_store) { 1513 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1514 false, false, true /* mismatched */); 1515 } else { 1516 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, MemNode::unordered, 1517 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */); 1518 set_result(ch); 1519 } 1520 return true; 1521 } 1522 1523 //--------------------------round_double_node-------------------------------- 1524 // Round a double node if necessary. 1525 Node* LibraryCallKit::round_double_node(Node* n) { 1526 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1527 n = _gvn.transform(new RoundDoubleNode(0, n)); 1528 return n; 1529 } 1530 1531 //------------------------------inline_math----------------------------------- 1532 // public static double Math.abs(double) 1533 // public static double Math.sqrt(double) 1534 // public static double Math.log(double) 1535 // public static double Math.log10(double) 1536 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1537 Node* arg = round_double_node(argument(0)); 1538 Node* n = NULL; 1539 switch (id) { 1540 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1541 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1542 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1543 default: fatal_unexpected_iid(id); break; 1544 } 1545 set_result(_gvn.transform(n)); 1546 return true; 1547 } 1548 1549 //------------------------------inline_trig---------------------------------- 1550 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1551 // argument reduction which will turn into a fast/slow diamond. 1552 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1553 Node* arg = round_double_node(argument(0)); 1554 Node* n = NULL; 1555 1556 switch (id) { 1557 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1558 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1559 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1560 default: fatal_unexpected_iid(id); break; 1561 } 1562 n = _gvn.transform(n); 1563 1564 // Rounding required? Check for argument reduction! 1565 if (Matcher::strict_fp_requires_explicit_rounding) { 1566 static const double pi_4 = 0.7853981633974483; 1567 static const double neg_pi_4 = -0.7853981633974483; 1568 // pi/2 in 80-bit extended precision 1569 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1570 // -pi/2 in 80-bit extended precision 1571 // 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}; 1572 // Cutoff value for using this argument reduction technique 1573 //static const double pi_2_minus_epsilon = 1.564660403643354; 1574 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1575 1576 // Pseudocode for sin: 1577 // if (x <= Math.PI / 4.0) { 1578 // if (x >= -Math.PI / 4.0) return fsin(x); 1579 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1580 // } else { 1581 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1582 // } 1583 // return StrictMath.sin(x); 1584 1585 // Pseudocode for cos: 1586 // if (x <= Math.PI / 4.0) { 1587 // if (x >= -Math.PI / 4.0) return fcos(x); 1588 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1589 // } else { 1590 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1591 // } 1592 // return StrictMath.cos(x); 1593 1594 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1595 // requires a special machine instruction to load it. Instead we'll try 1596 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1597 // probably do the math inside the SIN encoding. 1598 1599 // Make the merge point 1600 RegionNode* r = new RegionNode(3); 1601 Node* phi = new PhiNode(r, Type::DOUBLE); 1602 1603 // Flatten arg so we need only 1 test 1604 Node *abs = _gvn.transform(new AbsDNode(arg)); 1605 // Node for PI/4 constant 1606 Node *pi4 = makecon(TypeD::make(pi_4)); 1607 // Check PI/4 : abs(arg) 1608 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1609 // Check: If PI/4 < abs(arg) then go slow 1610 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1611 // Branch either way 1612 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1613 set_control(opt_iff(r,iff)); 1614 1615 // Set fast path result 1616 phi->init_req(2, n); 1617 1618 // Slow path - non-blocking leaf call 1619 Node* call = NULL; 1620 switch (id) { 1621 case vmIntrinsics::_dsin: 1622 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1623 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1624 "Sin", NULL, arg, top()); 1625 break; 1626 case vmIntrinsics::_dcos: 1627 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1628 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1629 "Cos", NULL, arg, top()); 1630 break; 1631 case vmIntrinsics::_dtan: 1632 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1633 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1634 "Tan", NULL, arg, top()); 1635 break; 1636 } 1637 assert(control()->in(0) == call, ""); 1638 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1639 r->init_req(1, control()); 1640 phi->init_req(1, slow_result); 1641 1642 // Post-merge 1643 set_control(_gvn.transform(r)); 1644 record_for_igvn(r); 1645 n = _gvn.transform(phi); 1646 1647 C->set_has_split_ifs(true); // Has chance for split-if optimization 1648 } 1649 set_result(n); 1650 return true; 1651 } 1652 1653 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1654 //------------------- 1655 //result=(result.isNaN())? funcAddr():result; 1656 // Check: If isNaN() by checking result!=result? then either trap 1657 // or go to runtime 1658 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1659 // Build the boolean node 1660 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1661 1662 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1663 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1664 // The pow or exp intrinsic returned a NaN, which requires a call 1665 // to the runtime. Recompile with the runtime call. 1666 uncommon_trap(Deoptimization::Reason_intrinsic, 1667 Deoptimization::Action_make_not_entrant); 1668 } 1669 return result; 1670 } else { 1671 // If this inlining ever returned NaN in the past, we compile a call 1672 // to the runtime to properly handle corner cases 1673 1674 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1675 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1676 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1677 1678 if (!if_slow->is_top()) { 1679 RegionNode* result_region = new RegionNode(3); 1680 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1681 1682 result_region->init_req(1, if_fast); 1683 result_val->init_req(1, result); 1684 1685 set_control(if_slow); 1686 1687 const TypePtr* no_memory_effects = NULL; 1688 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1689 no_memory_effects, 1690 x, top(), y, y ? top() : NULL); 1691 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1692 #ifdef ASSERT 1693 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1694 assert(value_top == top(), "second value must be top"); 1695 #endif 1696 1697 result_region->init_req(2, control()); 1698 result_val->init_req(2, value); 1699 set_control(_gvn.transform(result_region)); 1700 return _gvn.transform(result_val); 1701 } else { 1702 return result; 1703 } 1704 } 1705 } 1706 1707 //------------------------------inline_pow------------------------------------- 1708 // Inline power instructions, if possible. 1709 bool LibraryCallKit::inline_pow() { 1710 // Pseudocode for pow 1711 // if (y == 2) { 1712 // return x * x; 1713 // } else { 1714 // if (x <= 0.0) { 1715 // long longy = (long)y; 1716 // if ((double)longy == y) { // if y is long 1717 // if (y + 1 == y) longy = 0; // huge number: even 1718 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1719 // } else { 1720 // result = NaN; 1721 // } 1722 // } else { 1723 // result = DPow(x,y); 1724 // } 1725 // if (result != result)? { 1726 // result = uncommon_trap() or runtime_call(); 1727 // } 1728 // return result; 1729 // } 1730 1731 Node* x = round_double_node(argument(0)); 1732 Node* y = round_double_node(argument(2)); 1733 1734 Node* result = NULL; 1735 1736 Node* const_two_node = makecon(TypeD::make(2.0)); 1737 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1738 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1739 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1740 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1741 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1742 1743 RegionNode* region_node = new RegionNode(3); 1744 region_node->init_req(1, if_true); 1745 1746 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1747 // special case for x^y where y == 2, we can convert it to x * x 1748 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1749 1750 // set control to if_false since we will now process the false branch 1751 set_control(if_false); 1752 1753 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1754 // Short form: skip the fancy tests and just check for NaN result. 1755 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1756 } else { 1757 // If this inlining ever returned NaN in the past, include all 1758 // checks + call to the runtime. 1759 1760 // Set the merge point for If node with condition of (x <= 0.0) 1761 // There are four possible paths to region node and phi node 1762 RegionNode *r = new RegionNode(4); 1763 Node *phi = new PhiNode(r, Type::DOUBLE); 1764 1765 // Build the first if node: if (x <= 0.0) 1766 // Node for 0 constant 1767 Node *zeronode = makecon(TypeD::ZERO); 1768 // Check x:0 1769 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1770 // Check: If (x<=0) then go complex path 1771 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1772 // Branch either way 1773 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1774 // Fast path taken; set region slot 3 1775 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1776 r->init_req(3,fast_taken); // Capture fast-control 1777 1778 // Fast path not-taken, i.e. slow path 1779 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1780 1781 // Set fast path result 1782 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1783 phi->init_req(3, fast_result); 1784 1785 // Complex path 1786 // Build the second if node (if y is long) 1787 // Node for (long)y 1788 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1789 // Node for (double)((long) y) 1790 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1791 // Check (double)((long) y) : y 1792 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1793 // Check if (y isn't long) then go to slow path 1794 1795 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1796 // Branch either way 1797 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1798 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1799 1800 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1801 1802 // Calculate DPow(abs(x), y)*(1 & (long)y) 1803 // Node for constant 1 1804 Node *conone = longcon(1); 1805 // 1& (long)y 1806 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1807 1808 // A huge number is always even. Detect a huge number by checking 1809 // if y + 1 == y and set integer to be tested for parity to 0. 1810 // Required for corner case: 1811 // (long)9.223372036854776E18 = max_jlong 1812 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1813 // max_jlong is odd but 9.223372036854776E18 is even 1814 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1815 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1816 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1817 Node* correctedsign = NULL; 1818 if (ConditionalMoveLimit != 0) { 1819 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1820 } else { 1821 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1822 RegionNode *r = new RegionNode(3); 1823 Node *phi = new PhiNode(r, TypeLong::LONG); 1824 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1825 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1826 phi->init_req(1, signnode); 1827 phi->init_req(2, longcon(0)); 1828 correctedsign = _gvn.transform(phi); 1829 ylong_path = _gvn.transform(r); 1830 record_for_igvn(r); 1831 } 1832 1833 // zero node 1834 Node *conzero = longcon(0); 1835 // Check (1&(long)y)==0? 1836 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1837 // Check if (1&(long)y)!=0?, if so the result is negative 1838 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1839 // abs(x) 1840 Node *absx=_gvn.transform(new AbsDNode(x)); 1841 // abs(x)^y 1842 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1843 // -abs(x)^y 1844 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1845 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1846 Node *signresult = NULL; 1847 if (ConditionalMoveLimit != 0) { 1848 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1849 } else { 1850 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1851 RegionNode *r = new RegionNode(3); 1852 Node *phi = new PhiNode(r, Type::DOUBLE); 1853 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1854 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1855 phi->init_req(1, absxpowy); 1856 phi->init_req(2, negabsxpowy); 1857 signresult = _gvn.transform(phi); 1858 ylong_path = _gvn.transform(r); 1859 record_for_igvn(r); 1860 } 1861 // Set complex path fast result 1862 r->init_req(2, ylong_path); 1863 phi->init_req(2, signresult); 1864 1865 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1866 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1867 r->init_req(1,slow_path); 1868 phi->init_req(1,slow_result); 1869 1870 // Post merge 1871 set_control(_gvn.transform(r)); 1872 record_for_igvn(r); 1873 result = _gvn.transform(phi); 1874 } 1875 1876 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1877 1878 // control from finish_pow_exp is now input to the region node 1879 region_node->set_req(2, control()); 1880 // the result from finish_pow_exp is now input to the phi node 1881 phi_node->init_req(2, result); 1882 set_control(_gvn.transform(region_node)); 1883 record_for_igvn(region_node); 1884 set_result(_gvn.transform(phi_node)); 1885 1886 C->set_has_split_ifs(true); // Has chance for split-if optimization 1887 return true; 1888 } 1889 1890 //------------------------------runtime_math----------------------------- 1891 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1892 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1893 "must be (DD)D or (D)D type"); 1894 1895 // Inputs 1896 Node* a = round_double_node(argument(0)); 1897 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1898 1899 const TypePtr* no_memory_effects = NULL; 1900 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1901 no_memory_effects, 1902 a, top(), b, b ? top() : NULL); 1903 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1904 #ifdef ASSERT 1905 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1906 assert(value_top == top(), "second value must be top"); 1907 #endif 1908 1909 set_result(value); 1910 return true; 1911 } 1912 1913 //------------------------------inline_math_native----------------------------- 1914 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1915 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1916 switch (id) { 1917 // These intrinsics are not properly supported on all hardware 1918 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1919 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1920 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1921 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1922 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1923 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1924 1925 case vmIntrinsics::_dlog: 1926 return StubRoutines::dlog() != NULL ? 1927 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1928 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1929 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1930 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1931 1932 // These intrinsics are supported on all hardware 1933 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1934 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1935 1936 case vmIntrinsics::_dexp: 1937 return StubRoutines::dexp() != NULL ? 1938 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1939 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1940 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1941 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1942 #undef FN_PTR 1943 1944 // These intrinsics are not yet correctly implemented 1945 case vmIntrinsics::_datan2: 1946 return false; 1947 1948 default: 1949 fatal_unexpected_iid(id); 1950 return false; 1951 } 1952 } 1953 1954 static bool is_simple_name(Node* n) { 1955 return (n->req() == 1 // constant 1956 || (n->is_Type() && n->as_Type()->type()->singleton()) 1957 || n->is_Proj() // parameter or return value 1958 || n->is_Phi() // local of some sort 1959 ); 1960 } 1961 1962 //----------------------------inline_notify-----------------------------------* 1963 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1964 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1965 address func; 1966 if (id == vmIntrinsics::_notify) { 1967 func = OptoRuntime::monitor_notify_Java(); 1968 } else { 1969 func = OptoRuntime::monitor_notifyAll_Java(); 1970 } 1971 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1972 make_slow_call_ex(call, env()->Throwable_klass(), false); 1973 return true; 1974 } 1975 1976 1977 //----------------------------inline_min_max----------------------------------- 1978 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1979 set_result(generate_min_max(id, argument(0), argument(1))); 1980 return true; 1981 } 1982 1983 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1984 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1985 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1986 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1987 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1988 1989 { 1990 PreserveJVMState pjvms(this); 1991 PreserveReexecuteState preexecs(this); 1992 jvms()->set_should_reexecute(true); 1993 1994 set_control(slow_path); 1995 set_i_o(i_o()); 1996 1997 uncommon_trap(Deoptimization::Reason_intrinsic, 1998 Deoptimization::Action_none); 1999 } 2000 2001 set_control(fast_path); 2002 set_result(math); 2003 } 2004 2005 template <typename OverflowOp> 2006 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2007 typedef typename OverflowOp::MathOp MathOp; 2008 2009 MathOp* mathOp = new MathOp(arg1, arg2); 2010 Node* operation = _gvn.transform( mathOp ); 2011 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2012 inline_math_mathExact(operation, ofcheck); 2013 return true; 2014 } 2015 2016 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2017 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2018 } 2019 2020 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2021 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2022 } 2023 2024 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2025 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2026 } 2027 2028 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2029 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2030 } 2031 2032 bool LibraryCallKit::inline_math_negateExactI() { 2033 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2034 } 2035 2036 bool LibraryCallKit::inline_math_negateExactL() { 2037 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2038 } 2039 2040 bool LibraryCallKit::inline_math_multiplyExactI() { 2041 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2042 } 2043 2044 bool LibraryCallKit::inline_math_multiplyExactL() { 2045 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2046 } 2047 2048 Node* 2049 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2050 // These are the candidate return value: 2051 Node* xvalue = x0; 2052 Node* yvalue = y0; 2053 2054 if (xvalue == yvalue) { 2055 return xvalue; 2056 } 2057 2058 bool want_max = (id == vmIntrinsics::_max); 2059 2060 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2061 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2062 if (txvalue == NULL || tyvalue == NULL) return top(); 2063 // This is not really necessary, but it is consistent with a 2064 // hypothetical MaxINode::Value method: 2065 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2066 2067 // %%% This folding logic should (ideally) be in a different place. 2068 // Some should be inside IfNode, and there to be a more reliable 2069 // transformation of ?: style patterns into cmoves. We also want 2070 // more powerful optimizations around cmove and min/max. 2071 2072 // Try to find a dominating comparison of these guys. 2073 // It can simplify the index computation for Arrays.copyOf 2074 // and similar uses of System.arraycopy. 2075 // First, compute the normalized version of CmpI(x, y). 2076 int cmp_op = Op_CmpI; 2077 Node* xkey = xvalue; 2078 Node* ykey = yvalue; 2079 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2080 if (ideal_cmpxy->is_Cmp()) { 2081 // E.g., if we have CmpI(length - offset, count), 2082 // it might idealize to CmpI(length, count + offset) 2083 cmp_op = ideal_cmpxy->Opcode(); 2084 xkey = ideal_cmpxy->in(1); 2085 ykey = ideal_cmpxy->in(2); 2086 } 2087 2088 // Start by locating any relevant comparisons. 2089 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2090 Node* cmpxy = NULL; 2091 Node* cmpyx = NULL; 2092 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2093 Node* cmp = start_from->fast_out(k); 2094 if (cmp->outcnt() > 0 && // must have prior uses 2095 cmp->in(0) == NULL && // must be context-independent 2096 cmp->Opcode() == cmp_op) { // right kind of compare 2097 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2098 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2099 } 2100 } 2101 2102 const int NCMPS = 2; 2103 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2104 int cmpn; 2105 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2106 if (cmps[cmpn] != NULL) break; // find a result 2107 } 2108 if (cmpn < NCMPS) { 2109 // Look for a dominating test that tells us the min and max. 2110 int depth = 0; // Limit search depth for speed 2111 Node* dom = control(); 2112 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2113 if (++depth >= 100) break; 2114 Node* ifproj = dom; 2115 if (!ifproj->is_Proj()) continue; 2116 Node* iff = ifproj->in(0); 2117 if (!iff->is_If()) continue; 2118 Node* bol = iff->in(1); 2119 if (!bol->is_Bool()) continue; 2120 Node* cmp = bol->in(1); 2121 if (cmp == NULL) continue; 2122 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2123 if (cmps[cmpn] == cmp) break; 2124 if (cmpn == NCMPS) continue; 2125 BoolTest::mask btest = bol->as_Bool()->_test._test; 2126 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2127 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2128 // At this point, we know that 'x btest y' is true. 2129 switch (btest) { 2130 case BoolTest::eq: 2131 // They are proven equal, so we can collapse the min/max. 2132 // Either value is the answer. Choose the simpler. 2133 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2134 return yvalue; 2135 return xvalue; 2136 case BoolTest::lt: // x < y 2137 case BoolTest::le: // x <= y 2138 return (want_max ? yvalue : xvalue); 2139 case BoolTest::gt: // x > y 2140 case BoolTest::ge: // x >= y 2141 return (want_max ? xvalue : yvalue); 2142 } 2143 } 2144 } 2145 2146 // We failed to find a dominating test. 2147 // Let's pick a test that might GVN with prior tests. 2148 Node* best_bol = NULL; 2149 BoolTest::mask best_btest = BoolTest::illegal; 2150 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2151 Node* cmp = cmps[cmpn]; 2152 if (cmp == NULL) continue; 2153 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2154 Node* bol = cmp->fast_out(j); 2155 if (!bol->is_Bool()) continue; 2156 BoolTest::mask btest = bol->as_Bool()->_test._test; 2157 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2158 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2159 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2160 best_bol = bol->as_Bool(); 2161 best_btest = btest; 2162 } 2163 } 2164 } 2165 2166 Node* answer_if_true = NULL; 2167 Node* answer_if_false = NULL; 2168 switch (best_btest) { 2169 default: 2170 if (cmpxy == NULL) 2171 cmpxy = ideal_cmpxy; 2172 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2173 // and fall through: 2174 case BoolTest::lt: // x < y 2175 case BoolTest::le: // x <= y 2176 answer_if_true = (want_max ? yvalue : xvalue); 2177 answer_if_false = (want_max ? xvalue : yvalue); 2178 break; 2179 case BoolTest::gt: // x > y 2180 case BoolTest::ge: // x >= y 2181 answer_if_true = (want_max ? xvalue : yvalue); 2182 answer_if_false = (want_max ? yvalue : xvalue); 2183 break; 2184 } 2185 2186 jint hi, lo; 2187 if (want_max) { 2188 // We can sharpen the minimum. 2189 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2190 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2191 } else { 2192 // We can sharpen the maximum. 2193 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2194 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2195 } 2196 2197 // Use a flow-free graph structure, to avoid creating excess control edges 2198 // which could hinder other optimizations. 2199 // Since Math.min/max is often used with arraycopy, we want 2200 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2201 Node* cmov = CMoveNode::make(NULL, best_bol, 2202 answer_if_false, answer_if_true, 2203 TypeInt::make(lo, hi, widen)); 2204 2205 return _gvn.transform(cmov); 2206 2207 /* 2208 // This is not as desirable as it may seem, since Min and Max 2209 // nodes do not have a full set of optimizations. 2210 // And they would interfere, anyway, with 'if' optimizations 2211 // and with CMoveI canonical forms. 2212 switch (id) { 2213 case vmIntrinsics::_min: 2214 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2215 case vmIntrinsics::_max: 2216 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2217 default: 2218 ShouldNotReachHere(); 2219 } 2220 */ 2221 } 2222 2223 inline int 2224 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2225 const TypePtr* base_type = TypePtr::NULL_PTR; 2226 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2227 if (base_type == NULL) { 2228 // Unknown type. 2229 return Type::AnyPtr; 2230 } else if (base_type == TypePtr::NULL_PTR) { 2231 // Since this is a NULL+long form, we have to switch to a rawptr. 2232 base = _gvn.transform(new CastX2PNode(offset)); 2233 offset = MakeConX(0); 2234 return Type::RawPtr; 2235 } else if (base_type->base() == Type::RawPtr) { 2236 return Type::RawPtr; 2237 } else if (base_type->isa_oopptr()) { 2238 // Base is never null => always a heap address. 2239 if (base_type->ptr() == TypePtr::NotNull) { 2240 return Type::OopPtr; 2241 } 2242 // Offset is small => always a heap address. 2243 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2244 if (offset_type != NULL && 2245 base_type->offset() == 0 && // (should always be?) 2246 offset_type->_lo >= 0 && 2247 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2248 return Type::OopPtr; 2249 } 2250 // Otherwise, it might either be oop+off or NULL+addr. 2251 return Type::AnyPtr; 2252 } else { 2253 // No information: 2254 return Type::AnyPtr; 2255 } 2256 } 2257 2258 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2259 int kind = classify_unsafe_addr(base, offset); 2260 if (kind == Type::RawPtr) { 2261 return basic_plus_adr(top(), base, offset); 2262 } else { 2263 return basic_plus_adr(base, offset); 2264 } 2265 } 2266 2267 //--------------------------inline_number_methods----------------------------- 2268 // inline int Integer.numberOfLeadingZeros(int) 2269 // inline int Long.numberOfLeadingZeros(long) 2270 // 2271 // inline int Integer.numberOfTrailingZeros(int) 2272 // inline int Long.numberOfTrailingZeros(long) 2273 // 2274 // inline int Integer.bitCount(int) 2275 // inline int Long.bitCount(long) 2276 // 2277 // inline char Character.reverseBytes(char) 2278 // inline short Short.reverseBytes(short) 2279 // inline int Integer.reverseBytes(int) 2280 // inline long Long.reverseBytes(long) 2281 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2282 Node* arg = argument(0); 2283 Node* n = NULL; 2284 switch (id) { 2285 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2286 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2287 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2288 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2289 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2290 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2291 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2292 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2293 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2294 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2295 default: fatal_unexpected_iid(id); break; 2296 } 2297 set_result(_gvn.transform(n)); 2298 return true; 2299 } 2300 2301 //----------------------------inline_unsafe_access---------------------------- 2302 2303 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2304 2305 // Helper that guards and inserts a pre-barrier. 2306 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2307 Node* pre_val, bool need_mem_bar) { 2308 // We could be accessing the referent field of a reference object. If so, when G1 2309 // is enabled, we need to log the value in the referent field in an SATB buffer. 2310 // This routine performs some compile time filters and generates suitable 2311 // runtime filters that guard the pre-barrier code. 2312 // Also add memory barrier for non volatile load from the referent field 2313 // to prevent commoning of loads across safepoint. 2314 if (!UseG1GC && !need_mem_bar) 2315 return; 2316 2317 // Some compile time checks. 2318 2319 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2320 const TypeX* otype = offset->find_intptr_t_type(); 2321 if (otype != NULL && otype->is_con() && 2322 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2323 // Constant offset but not the reference_offset so just return 2324 return; 2325 } 2326 2327 // We only need to generate the runtime guards for instances. 2328 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2329 if (btype != NULL) { 2330 if (btype->isa_aryptr()) { 2331 // Array type so nothing to do 2332 return; 2333 } 2334 2335 const TypeInstPtr* itype = btype->isa_instptr(); 2336 if (itype != NULL) { 2337 // Can the klass of base_oop be statically determined to be 2338 // _not_ a sub-class of Reference and _not_ Object? 2339 ciKlass* klass = itype->klass(); 2340 if ( klass->is_loaded() && 2341 !klass->is_subtype_of(env()->Reference_klass()) && 2342 !env()->Object_klass()->is_subtype_of(klass)) { 2343 return; 2344 } 2345 } 2346 } 2347 2348 // The compile time filters did not reject base_oop/offset so 2349 // we need to generate the following runtime filters 2350 // 2351 // if (offset == java_lang_ref_Reference::_reference_offset) { 2352 // if (instance_of(base, java.lang.ref.Reference)) { 2353 // pre_barrier(_, pre_val, ...); 2354 // } 2355 // } 2356 2357 float likely = PROB_LIKELY( 0.999); 2358 float unlikely = PROB_UNLIKELY(0.999); 2359 2360 IdealKit ideal(this); 2361 #define __ ideal. 2362 2363 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2364 2365 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2366 // Update graphKit memory and control from IdealKit. 2367 sync_kit(ideal); 2368 2369 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2370 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2371 2372 // Update IdealKit memory and control from graphKit. 2373 __ sync_kit(this); 2374 2375 Node* one = __ ConI(1); 2376 // is_instof == 0 if base_oop == NULL 2377 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2378 2379 // Update graphKit from IdeakKit. 2380 sync_kit(ideal); 2381 2382 // Use the pre-barrier to record the value in the referent field 2383 pre_barrier(false /* do_load */, 2384 __ ctrl(), 2385 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2386 pre_val /* pre_val */, 2387 T_OBJECT); 2388 if (need_mem_bar) { 2389 // Add memory barrier to prevent commoning reads from this field 2390 // across safepoint since GC can change its value. 2391 insert_mem_bar(Op_MemBarCPUOrder); 2392 } 2393 // Update IdealKit from graphKit. 2394 __ sync_kit(this); 2395 2396 } __ end_if(); // _ref_type != ref_none 2397 } __ end_if(); // offset == referent_offset 2398 2399 // Final sync IdealKit and GraphKit. 2400 final_sync(ideal); 2401 #undef __ 2402 } 2403 2404 2405 // Interpret Unsafe.fieldOffset cookies correctly: 2406 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2407 2408 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2409 // Attempt to infer a sharper value type from the offset and base type. 2410 ciKlass* sharpened_klass = NULL; 2411 2412 // See if it is an instance field, with an object type. 2413 if (alias_type->field() != NULL) { 2414 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2415 if (alias_type->field()->type()->is_klass()) { 2416 sharpened_klass = alias_type->field()->type()->as_klass(); 2417 } 2418 } 2419 2420 // See if it is a narrow oop array. 2421 if (adr_type->isa_aryptr()) { 2422 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2423 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2424 if (elem_type != NULL) { 2425 sharpened_klass = elem_type->klass(); 2426 } 2427 } 2428 } 2429 2430 // The sharpened class might be unloaded if there is no class loader 2431 // contraint in place. 2432 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2433 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2434 2435 #ifndef PRODUCT 2436 if (C->print_intrinsics() || C->print_inlining()) { 2437 tty->print(" from base type: "); adr_type->dump(); 2438 tty->print(" sharpened value: "); tjp->dump(); 2439 } 2440 #endif 2441 // Sharpen the value type. 2442 return tjp; 2443 } 2444 return NULL; 2445 } 2446 2447 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) { 2448 if (callee()->is_static()) return false; // caller must have the capability! 2449 2450 #ifndef PRODUCT 2451 { 2452 ResourceMark rm; 2453 // Check the signatures. 2454 ciSignature* sig = callee()->signature(); 2455 #ifdef ASSERT 2456 if (!is_store) { 2457 // Object getObject(Object base, int/long offset), etc. 2458 BasicType rtype = sig->return_type()->basic_type(); 2459 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2460 rtype = T_ADDRESS; // it is really a C void* 2461 assert(rtype == type, "getter must return the expected value"); 2462 if (!is_native_ptr) { 2463 assert(sig->count() == 2, "oop getter has 2 arguments"); 2464 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2465 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2466 } else { 2467 assert(sig->count() == 1, "native getter has 1 argument"); 2468 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2469 } 2470 } else { 2471 // void putObject(Object base, int/long offset, Object x), etc. 2472 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2473 if (!is_native_ptr) { 2474 assert(sig->count() == 3, "oop putter has 3 arguments"); 2475 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2476 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2477 } else { 2478 assert(sig->count() == 2, "native putter has 2 arguments"); 2479 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2480 } 2481 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2482 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2483 vtype = T_ADDRESS; // it is really a C void* 2484 assert(vtype == type, "putter must accept the expected value"); 2485 } 2486 #endif // ASSERT 2487 } 2488 #endif //PRODUCT 2489 2490 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2491 2492 Node* receiver = argument(0); // type: oop 2493 2494 // Build address expression. 2495 Node* adr; 2496 Node* heap_base_oop = top(); 2497 Node* offset = top(); 2498 Node* val; 2499 2500 if (!is_native_ptr) { 2501 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2502 Node* base = argument(1); // type: oop 2503 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2504 offset = argument(2); // type: long 2505 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2506 // to be plain byte offsets, which are also the same as those accepted 2507 // by oopDesc::field_base. 2508 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2509 "fieldOffset must be byte-scaled"); 2510 // 32-bit machines ignore the high half! 2511 offset = ConvL2X(offset); 2512 adr = make_unsafe_address(base, offset); 2513 heap_base_oop = base; 2514 val = is_store ? argument(4) : NULL; 2515 } else { 2516 Node* ptr = argument(1); // type: long 2517 ptr = ConvL2X(ptr); // adjust Java long to machine word 2518 adr = make_unsafe_address(NULL, ptr); 2519 val = is_store ? argument(3) : NULL; 2520 } 2521 2522 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2523 2524 // First guess at the value type. 2525 const Type *value_type = Type::get_const_basic_type(type); 2526 2527 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2528 // there was not enough information to nail it down. 2529 Compile::AliasType* alias_type = C->alias_type(adr_type); 2530 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2531 2532 // We will need memory barriers unless we can determine a unique 2533 // alias category for this reference. (Note: If for some reason 2534 // the barriers get omitted and the unsafe reference begins to "pollute" 2535 // the alias analysis of the rest of the graph, either Compile::can_alias 2536 // or Compile::must_alias will throw a diagnostic assert.) 2537 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2538 2539 // If we are reading the value of the referent field of a Reference 2540 // object (either by using Unsafe directly or through reflection) 2541 // then, if G1 is enabled, we need to record the referent in an 2542 // SATB log buffer using the pre-barrier mechanism. 2543 // Also we need to add memory barrier to prevent commoning reads 2544 // from this field across safepoint since GC can change its value. 2545 bool need_read_barrier = !is_native_ptr && !is_store && 2546 offset != top() && heap_base_oop != top(); 2547 2548 if (!is_store && type == T_OBJECT) { 2549 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2550 if (tjp != NULL) { 2551 value_type = tjp; 2552 } 2553 } 2554 2555 receiver = null_check(receiver); 2556 if (stopped()) { 2557 return true; 2558 } 2559 // Heap pointers get a null-check from the interpreter, 2560 // as a courtesy. However, this is not guaranteed by Unsafe, 2561 // and it is not possible to fully distinguish unintended nulls 2562 // from intended ones in this API. 2563 2564 if (is_volatile) { 2565 // We need to emit leading and trailing CPU membars (see below) in 2566 // addition to memory membars when is_volatile. This is a little 2567 // too strong, but avoids the need to insert per-alias-type 2568 // volatile membars (for stores; compare Parse::do_put_xxx), which 2569 // we cannot do effectively here because we probably only have a 2570 // rough approximation of type. 2571 need_mem_bar = true; 2572 // For Stores, place a memory ordering barrier now. 2573 if (is_store) { 2574 insert_mem_bar(Op_MemBarRelease); 2575 } else { 2576 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2577 insert_mem_bar(Op_MemBarVolatile); 2578 } 2579 } 2580 } 2581 2582 // Memory barrier to prevent normal and 'unsafe' accesses from 2583 // bypassing each other. Happens after null checks, so the 2584 // exception paths do not take memory state from the memory barrier, 2585 // so there's no problems making a strong assert about mixing users 2586 // of safe & unsafe memory. 2587 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2588 2589 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM || 2590 alias_type->field() != NULL || alias_type->element() != NULL, "field, array element or unknown"); 2591 bool mismatched = false; 2592 if (alias_type->element() != NULL || alias_type->field() != NULL) { 2593 BasicType bt; 2594 if (alias_type->element() != NULL) { 2595 const Type* element = alias_type->element(); 2596 bt = element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 2597 } else { 2598 bt = alias_type->field()->type()->basic_type(); 2599 } 2600 if (bt == T_ARRAY) { 2601 // accessing an array field with getObject is not a mismatch 2602 bt = T_OBJECT; 2603 } 2604 if (bt != type) { 2605 mismatched = true; 2606 } 2607 } 2608 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2609 2610 if (!is_store) { 2611 Node* p = NULL; 2612 // Try to constant fold a load from a constant field 2613 ciField* field = alias_type->field(); 2614 if (heap_base_oop != top() && 2615 field != NULL && field->is_constant() && field->layout_type() == type) { 2616 // final or stable field 2617 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop); 2618 if (con_type != NULL) { 2619 p = makecon(con_type); 2620 } 2621 } 2622 if (p == NULL) { 2623 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2624 // To be valid, unsafe loads may depend on other conditions than 2625 // the one that guards them: pin the Load node 2626 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched); 2627 // load value 2628 switch (type) { 2629 case T_BOOLEAN: 2630 case T_CHAR: 2631 case T_BYTE: 2632 case T_SHORT: 2633 case T_INT: 2634 case T_LONG: 2635 case T_FLOAT: 2636 case T_DOUBLE: 2637 break; 2638 case T_OBJECT: 2639 if (need_read_barrier) { 2640 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2641 } 2642 break; 2643 case T_ADDRESS: 2644 // Cast to an int type. 2645 p = _gvn.transform(new CastP2XNode(NULL, p)); 2646 p = ConvX2UL(p); 2647 break; 2648 default: 2649 fatal("unexpected type %d: %s", type, type2name(type)); 2650 break; 2651 } 2652 } 2653 // The load node has the control of the preceding MemBarCPUOrder. All 2654 // following nodes will have the control of the MemBarCPUOrder inserted at 2655 // the end of this method. So, pushing the load onto the stack at a later 2656 // point is fine. 2657 set_result(p); 2658 } else { 2659 // place effect of store into memory 2660 switch (type) { 2661 case T_DOUBLE: 2662 val = dstore_rounding(val); 2663 break; 2664 case T_ADDRESS: 2665 // Repackage the long as a pointer. 2666 val = ConvL2X(val); 2667 val = _gvn.transform(new CastX2PNode(val)); 2668 break; 2669 } 2670 2671 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2672 if (type != T_OBJECT ) { 2673 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched); 2674 } else { 2675 // Possibly an oop being stored to Java heap or native memory 2676 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2677 // oop to Java heap. 2678 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2679 } else { 2680 // We can't tell at compile time if we are storing in the Java heap or outside 2681 // of it. So we need to emit code to conditionally do the proper type of 2682 // store. 2683 2684 IdealKit ideal(this); 2685 #define __ ideal. 2686 // QQQ who knows what probability is here?? 2687 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2688 // Sync IdealKit and graphKit. 2689 sync_kit(ideal); 2690 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2691 // Update IdealKit memory. 2692 __ sync_kit(this); 2693 } __ else_(); { 2694 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile, mismatched); 2695 } __ end_if(); 2696 // Final sync IdealKit and GraphKit. 2697 final_sync(ideal); 2698 #undef __ 2699 } 2700 } 2701 } 2702 2703 if (is_volatile) { 2704 if (!is_store) { 2705 insert_mem_bar(Op_MemBarAcquire); 2706 } else { 2707 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2708 insert_mem_bar(Op_MemBarVolatile); 2709 } 2710 } 2711 } 2712 2713 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2714 2715 return true; 2716 } 2717 2718 //----------------------------inline_unsafe_load_store---------------------------- 2719 // This method serves a couple of different customers (depending on LoadStoreKind): 2720 // 2721 // LS_cmpxchg: 2722 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2723 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2724 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2725 // 2726 // LS_xadd: 2727 // public int getAndAddInt( Object o, long offset, int delta) 2728 // public long getAndAddLong(Object o, long offset, long delta) 2729 // 2730 // LS_xchg: 2731 // int getAndSet(Object o, long offset, int newValue) 2732 // long getAndSet(Object o, long offset, long newValue) 2733 // Object getAndSet(Object o, long offset, Object newValue) 2734 // 2735 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2736 // This basic scheme here is the same as inline_unsafe_access, but 2737 // differs in enough details that combining them would make the code 2738 // overly confusing. (This is a true fact! I originally combined 2739 // them, but even I was confused by it!) As much code/comments as 2740 // possible are retained from inline_unsafe_access though to make 2741 // the correspondences clearer. - dl 2742 2743 if (callee()->is_static()) return false; // caller must have the capability! 2744 2745 #ifndef PRODUCT 2746 BasicType rtype; 2747 { 2748 ResourceMark rm; 2749 // Check the signatures. 2750 ciSignature* sig = callee()->signature(); 2751 rtype = sig->return_type()->basic_type(); 2752 if (kind == LS_xadd || kind == LS_xchg) { 2753 // Check the signatures. 2754 #ifdef ASSERT 2755 assert(rtype == type, "get and set must return the expected type"); 2756 assert(sig->count() == 3, "get and set has 3 arguments"); 2757 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2758 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2759 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2760 #endif // ASSERT 2761 } else if (kind == LS_cmpxchg) { 2762 // Check the signatures. 2763 #ifdef ASSERT 2764 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2765 assert(sig->count() == 4, "CAS has 4 arguments"); 2766 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2767 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2768 #endif // ASSERT 2769 } else { 2770 ShouldNotReachHere(); 2771 } 2772 } 2773 #endif //PRODUCT 2774 2775 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2776 2777 // Get arguments: 2778 Node* receiver = NULL; 2779 Node* base = NULL; 2780 Node* offset = NULL; 2781 Node* oldval = NULL; 2782 Node* newval = NULL; 2783 if (kind == LS_cmpxchg) { 2784 const bool two_slot_type = type2size[type] == 2; 2785 receiver = argument(0); // type: oop 2786 base = argument(1); // type: oop 2787 offset = argument(2); // type: long 2788 oldval = argument(4); // type: oop, int, or long 2789 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2790 } else if (kind == LS_xadd || kind == LS_xchg){ 2791 receiver = argument(0); // type: oop 2792 base = argument(1); // type: oop 2793 offset = argument(2); // type: long 2794 oldval = NULL; 2795 newval = argument(4); // type: oop, int, or long 2796 } 2797 2798 // Null check receiver. 2799 receiver = null_check(receiver); 2800 if (stopped()) { 2801 return true; 2802 } 2803 2804 // Build field offset expression. 2805 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2806 // to be plain byte offsets, which are also the same as those accepted 2807 // by oopDesc::field_base. 2808 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2809 // 32-bit machines ignore the high half of long offsets 2810 offset = ConvL2X(offset); 2811 Node* adr = make_unsafe_address(base, offset); 2812 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2813 2814 // For CAS, unlike inline_unsafe_access, there seems no point in 2815 // trying to refine types. Just use the coarse types here. 2816 const Type *value_type = Type::get_const_basic_type(type); 2817 Compile::AliasType* alias_type = C->alias_type(adr_type); 2818 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2819 2820 if (kind == LS_xchg && type == T_OBJECT) { 2821 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2822 if (tjp != NULL) { 2823 value_type = tjp; 2824 } 2825 } 2826 2827 int alias_idx = C->get_alias_index(adr_type); 2828 2829 // Memory-model-wise, a LoadStore acts like a little synchronized 2830 // block, so needs barriers on each side. These don't translate 2831 // into actual barriers on most machines, but we still need rest of 2832 // compiler to respect ordering. 2833 2834 insert_mem_bar(Op_MemBarRelease); 2835 insert_mem_bar(Op_MemBarCPUOrder); 2836 2837 // 4984716: MemBars must be inserted before this 2838 // memory node in order to avoid a false 2839 // dependency which will confuse the scheduler. 2840 Node *mem = memory(alias_idx); 2841 2842 // For now, we handle only those cases that actually exist: ints, 2843 // longs, and Object. Adding others should be straightforward. 2844 Node* load_store = NULL; 2845 switch(type) { 2846 case T_INT: 2847 if (kind == LS_xadd) { 2848 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2849 } else if (kind == LS_xchg) { 2850 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2851 } else if (kind == LS_cmpxchg) { 2852 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2853 } else { 2854 ShouldNotReachHere(); 2855 } 2856 break; 2857 case T_LONG: 2858 if (kind == LS_xadd) { 2859 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2860 } else if (kind == LS_xchg) { 2861 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2862 } else if (kind == LS_cmpxchg) { 2863 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2864 } else { 2865 ShouldNotReachHere(); 2866 } 2867 break; 2868 case T_OBJECT: 2869 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2870 // could be delayed during Parse (for example, in adjust_map_after_if()). 2871 // Execute transformation here to avoid barrier generation in such case. 2872 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2873 newval = _gvn.makecon(TypePtr::NULL_PTR); 2874 2875 // Reference stores need a store barrier. 2876 if (kind == LS_xchg) { 2877 // If pre-barrier must execute before the oop store, old value will require do_load here. 2878 if (!can_move_pre_barrier()) { 2879 pre_barrier(true /* do_load*/, 2880 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2881 NULL /* pre_val*/, 2882 T_OBJECT); 2883 } // Else move pre_barrier to use load_store value, see below. 2884 } else if (kind == LS_cmpxchg) { 2885 // Same as for newval above: 2886 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2887 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2888 } 2889 // The only known value which might get overwritten is oldval. 2890 pre_barrier(false /* do_load */, 2891 control(), NULL, NULL, max_juint, NULL, NULL, 2892 oldval /* pre_val */, 2893 T_OBJECT); 2894 } else { 2895 ShouldNotReachHere(); 2896 } 2897 2898 #ifdef _LP64 2899 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2900 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2901 if (kind == LS_xchg) { 2902 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2903 newval_enc, adr_type, value_type->make_narrowoop())); 2904 } else { 2905 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2906 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2907 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2908 newval_enc, oldval_enc)); 2909 } 2910 } else 2911 #endif 2912 { 2913 if (kind == LS_xchg) { 2914 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2915 } else { 2916 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2917 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2918 } 2919 } 2920 if (kind == LS_cmpxchg) { 2921 // Emit the post barrier only when the actual store happened. 2922 // This makes sense to check only for compareAndSet that can fail to set the value. 2923 // CAS success path is marked more likely since we anticipate this is a performance 2924 // critical path, while CAS failure path can use the penalty for going through unlikely 2925 // path as backoff. Which is still better than doing a store barrier there. 2926 IdealKit ideal(this); 2927 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); { 2928 sync_kit(ideal); 2929 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2930 ideal.sync_kit(this); 2931 } ideal.end_if(); 2932 final_sync(ideal); 2933 } else { 2934 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2935 } 2936 break; 2937 default: 2938 fatal("unexpected type %d: %s", type, type2name(type)); 2939 break; 2940 } 2941 2942 // SCMemProjNodes represent the memory state of a LoadStore. Their 2943 // main role is to prevent LoadStore nodes from being optimized away 2944 // when their results aren't used. 2945 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2946 set_memory(proj, alias_idx); 2947 2948 if (type == T_OBJECT && kind == LS_xchg) { 2949 #ifdef _LP64 2950 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2951 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2952 } 2953 #endif 2954 if (can_move_pre_barrier()) { 2955 // Don't need to load pre_val. The old value is returned by load_store. 2956 // The pre_barrier can execute after the xchg as long as no safepoint 2957 // gets inserted between them. 2958 pre_barrier(false /* do_load */, 2959 control(), NULL, NULL, max_juint, NULL, NULL, 2960 load_store /* pre_val */, 2961 T_OBJECT); 2962 } 2963 } 2964 2965 // Add the trailing membar surrounding the access 2966 insert_mem_bar(Op_MemBarCPUOrder); 2967 insert_mem_bar(Op_MemBarAcquire); 2968 2969 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2970 set_result(load_store); 2971 return true; 2972 } 2973 2974 //----------------------------inline_unsafe_ordered_store---------------------- 2975 // public native void Unsafe.putOrderedObject(Object o, long offset, Object x); 2976 // public native void Unsafe.putOrderedInt(Object o, long offset, int x); 2977 // public native void Unsafe.putOrderedLong(Object o, long offset, long x); 2978 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2979 // This is another variant of inline_unsafe_access, differing in 2980 // that it always issues store-store ("release") barrier and ensures 2981 // store-atomicity (which only matters for "long"). 2982 2983 if (callee()->is_static()) return false; // caller must have the capability! 2984 2985 #ifndef PRODUCT 2986 { 2987 ResourceMark rm; 2988 // Check the signatures. 2989 ciSignature* sig = callee()->signature(); 2990 #ifdef ASSERT 2991 BasicType rtype = sig->return_type()->basic_type(); 2992 assert(rtype == T_VOID, "must return void"); 2993 assert(sig->count() == 3, "has 3 arguments"); 2994 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2995 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2996 #endif // ASSERT 2997 } 2998 #endif //PRODUCT 2999 3000 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3001 3002 // Get arguments: 3003 Node* receiver = argument(0); // type: oop 3004 Node* base = argument(1); // type: oop 3005 Node* offset = argument(2); // type: long 3006 Node* val = argument(4); // type: oop, int, or long 3007 3008 // Null check receiver. 3009 receiver = null_check(receiver); 3010 if (stopped()) { 3011 return true; 3012 } 3013 3014 // Build field offset expression. 3015 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3016 // 32-bit machines ignore the high half of long offsets 3017 offset = ConvL2X(offset); 3018 Node* adr = make_unsafe_address(base, offset); 3019 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3020 const Type *value_type = Type::get_const_basic_type(type); 3021 Compile::AliasType* alias_type = C->alias_type(adr_type); 3022 3023 insert_mem_bar(Op_MemBarRelease); 3024 insert_mem_bar(Op_MemBarCPUOrder); 3025 // Ensure that the store is atomic for longs: 3026 const bool require_atomic_access = true; 3027 Node* store; 3028 if (type == T_OBJECT) // reference stores need a store barrier. 3029 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3030 else { 3031 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3032 } 3033 insert_mem_bar(Op_MemBarCPUOrder); 3034 return true; 3035 } 3036 3037 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3038 // Regardless of form, don't allow previous ld/st to move down, 3039 // then issue acquire, release, or volatile mem_bar. 3040 insert_mem_bar(Op_MemBarCPUOrder); 3041 switch(id) { 3042 case vmIntrinsics::_loadFence: 3043 insert_mem_bar(Op_LoadFence); 3044 return true; 3045 case vmIntrinsics::_storeFence: 3046 insert_mem_bar(Op_StoreFence); 3047 return true; 3048 case vmIntrinsics::_fullFence: 3049 insert_mem_bar(Op_MemBarVolatile); 3050 return true; 3051 default: 3052 fatal_unexpected_iid(id); 3053 return false; 3054 } 3055 } 3056 3057 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3058 if (!kls->is_Con()) { 3059 return true; 3060 } 3061 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3062 if (klsptr == NULL) { 3063 return true; 3064 } 3065 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3066 // don't need a guard for a klass that is already initialized 3067 return !ik->is_initialized(); 3068 } 3069 3070 //----------------------------inline_unsafe_allocate--------------------------- 3071 // public native Object Unsafe.allocateInstance(Class<?> cls); 3072 bool LibraryCallKit::inline_unsafe_allocate() { 3073 if (callee()->is_static()) return false; // caller must have the capability! 3074 3075 null_check_receiver(); // null-check, then ignore 3076 Node* cls = null_check(argument(1)); 3077 if (stopped()) return true; 3078 3079 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3080 kls = null_check(kls); 3081 if (stopped()) return true; // argument was like int.class 3082 3083 Node* test = NULL; 3084 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3085 // Note: The argument might still be an illegal value like 3086 // Serializable.class or Object[].class. The runtime will handle it. 3087 // But we must make an explicit check for initialization. 3088 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3089 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3090 // can generate code to load it as unsigned byte. 3091 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3092 Node* bits = intcon(InstanceKlass::fully_initialized); 3093 test = _gvn.transform(new SubINode(inst, bits)); 3094 // The 'test' is non-zero if we need to take a slow path. 3095 } 3096 3097 Node* obj = new_instance(kls, test); 3098 set_result(obj); 3099 return true; 3100 } 3101 3102 #ifdef TRACE_HAVE_INTRINSICS 3103 /* 3104 * oop -> myklass 3105 * myklass->trace_id |= USED 3106 * return myklass->trace_id & ~0x3 3107 */ 3108 bool LibraryCallKit::inline_native_classID() { 3109 null_check_receiver(); // null-check, then ignore 3110 Node* cls = null_check(argument(1), T_OBJECT); 3111 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3112 kls = null_check(kls, T_OBJECT); 3113 ByteSize offset = TRACE_ID_OFFSET; 3114 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3115 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3116 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3117 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3118 Node* clsused = longcon(0x01l); // set the class bit 3119 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3120 3121 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3122 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3123 set_result(andl); 3124 return true; 3125 } 3126 3127 bool LibraryCallKit::inline_native_threadID() { 3128 Node* tls_ptr = NULL; 3129 Node* cur_thr = generate_current_thread(tls_ptr); 3130 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3131 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3132 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3133 3134 Node* threadid = NULL; 3135 size_t thread_id_size = OSThread::thread_id_size(); 3136 if (thread_id_size == (size_t) BytesPerLong) { 3137 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3138 } else if (thread_id_size == (size_t) BytesPerInt) { 3139 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3140 } else { 3141 ShouldNotReachHere(); 3142 } 3143 set_result(threadid); 3144 return true; 3145 } 3146 #endif 3147 3148 //------------------------inline_native_time_funcs-------------- 3149 // inline code for System.currentTimeMillis() and System.nanoTime() 3150 // these have the same type and signature 3151 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3152 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3153 const TypePtr* no_memory_effects = NULL; 3154 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3155 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3156 #ifdef ASSERT 3157 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3158 assert(value_top == top(), "second value must be top"); 3159 #endif 3160 set_result(value); 3161 return true; 3162 } 3163 3164 //------------------------inline_native_currentThread------------------ 3165 bool LibraryCallKit::inline_native_currentThread() { 3166 Node* junk = NULL; 3167 set_result(generate_current_thread(junk)); 3168 return true; 3169 } 3170 3171 //------------------------inline_native_isInterrupted------------------ 3172 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3173 bool LibraryCallKit::inline_native_isInterrupted() { 3174 // Add a fast path to t.isInterrupted(clear_int): 3175 // (t == Thread.current() && 3176 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3177 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3178 // So, in the common case that the interrupt bit is false, 3179 // we avoid making a call into the VM. Even if the interrupt bit 3180 // is true, if the clear_int argument is false, we avoid the VM call. 3181 // However, if the receiver is not currentThread, we must call the VM, 3182 // because there must be some locking done around the operation. 3183 3184 // We only go to the fast case code if we pass two guards. 3185 // Paths which do not pass are accumulated in the slow_region. 3186 3187 enum { 3188 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3189 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3190 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3191 PATH_LIMIT 3192 }; 3193 3194 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3195 // out of the function. 3196 insert_mem_bar(Op_MemBarCPUOrder); 3197 3198 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3199 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3200 3201 RegionNode* slow_region = new RegionNode(1); 3202 record_for_igvn(slow_region); 3203 3204 // (a) Receiving thread must be the current thread. 3205 Node* rec_thr = argument(0); 3206 Node* tls_ptr = NULL; 3207 Node* cur_thr = generate_current_thread(tls_ptr); 3208 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3209 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3210 3211 generate_slow_guard(bol_thr, slow_region); 3212 3213 // (b) Interrupt bit on TLS must be false. 3214 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3215 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3216 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3217 3218 // Set the control input on the field _interrupted read to prevent it floating up. 3219 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3220 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3221 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3222 3223 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3224 3225 // First fast path: if (!TLS._interrupted) return false; 3226 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3227 result_rgn->init_req(no_int_result_path, false_bit); 3228 result_val->init_req(no_int_result_path, intcon(0)); 3229 3230 // drop through to next case 3231 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3232 3233 #ifndef TARGET_OS_FAMILY_windows 3234 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3235 Node* clr_arg = argument(1); 3236 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3237 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3238 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3239 3240 // Second fast path: ... else if (!clear_int) return true; 3241 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3242 result_rgn->init_req(no_clear_result_path, false_arg); 3243 result_val->init_req(no_clear_result_path, intcon(1)); 3244 3245 // drop through to next case 3246 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3247 #else 3248 // To return true on Windows you must read the _interrupted field 3249 // and check the the event state i.e. take the slow path. 3250 #endif // TARGET_OS_FAMILY_windows 3251 3252 // (d) Otherwise, go to the slow path. 3253 slow_region->add_req(control()); 3254 set_control( _gvn.transform(slow_region)); 3255 3256 if (stopped()) { 3257 // There is no slow path. 3258 result_rgn->init_req(slow_result_path, top()); 3259 result_val->init_req(slow_result_path, top()); 3260 } else { 3261 // non-virtual because it is a private non-static 3262 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3263 3264 Node* slow_val = set_results_for_java_call(slow_call); 3265 // this->control() comes from set_results_for_java_call 3266 3267 Node* fast_io = slow_call->in(TypeFunc::I_O); 3268 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3269 3270 // These two phis are pre-filled with copies of of the fast IO and Memory 3271 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3272 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3273 3274 result_rgn->init_req(slow_result_path, control()); 3275 result_io ->init_req(slow_result_path, i_o()); 3276 result_mem->init_req(slow_result_path, reset_memory()); 3277 result_val->init_req(slow_result_path, slow_val); 3278 3279 set_all_memory(_gvn.transform(result_mem)); 3280 set_i_o( _gvn.transform(result_io)); 3281 } 3282 3283 C->set_has_split_ifs(true); // Has chance for split-if optimization 3284 set_result(result_rgn, result_val); 3285 return true; 3286 } 3287 3288 //---------------------------load_mirror_from_klass---------------------------- 3289 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3290 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3291 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3292 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3293 } 3294 3295 //-----------------------load_klass_from_mirror_common------------------------- 3296 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3297 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3298 // and branch to the given path on the region. 3299 // If never_see_null, take an uncommon trap on null, so we can optimistically 3300 // compile for the non-null case. 3301 // If the region is NULL, force never_see_null = true. 3302 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3303 bool never_see_null, 3304 RegionNode* region, 3305 int null_path, 3306 int offset) { 3307 if (region == NULL) never_see_null = true; 3308 Node* p = basic_plus_adr(mirror, offset); 3309 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3310 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3311 Node* null_ctl = top(); 3312 kls = null_check_oop(kls, &null_ctl, never_see_null); 3313 if (region != NULL) { 3314 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3315 region->init_req(null_path, null_ctl); 3316 } else { 3317 assert(null_ctl == top(), "no loose ends"); 3318 } 3319 return kls; 3320 } 3321 3322 //--------------------(inline_native_Class_query helpers)--------------------- 3323 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3324 // Fall through if (mods & mask) == bits, take the guard otherwise. 3325 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3326 // Branch around if the given klass has the given modifier bit set. 3327 // Like generate_guard, adds a new path onto the region. 3328 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3329 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3330 Node* mask = intcon(modifier_mask); 3331 Node* bits = intcon(modifier_bits); 3332 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3333 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3334 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3335 return generate_fair_guard(bol, region); 3336 } 3337 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3338 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3339 } 3340 3341 //-------------------------inline_native_Class_query------------------- 3342 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3343 const Type* return_type = TypeInt::BOOL; 3344 Node* prim_return_value = top(); // what happens if it's a primitive class? 3345 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3346 bool expect_prim = false; // most of these guys expect to work on refs 3347 3348 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3349 3350 Node* mirror = argument(0); 3351 Node* obj = top(); 3352 3353 switch (id) { 3354 case vmIntrinsics::_isInstance: 3355 // nothing is an instance of a primitive type 3356 prim_return_value = intcon(0); 3357 obj = argument(1); 3358 break; 3359 case vmIntrinsics::_getModifiers: 3360 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3361 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3362 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3363 break; 3364 case vmIntrinsics::_isInterface: 3365 prim_return_value = intcon(0); 3366 break; 3367 case vmIntrinsics::_isArray: 3368 prim_return_value = intcon(0); 3369 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3370 break; 3371 case vmIntrinsics::_isPrimitive: 3372 prim_return_value = intcon(1); 3373 expect_prim = true; // obviously 3374 break; 3375 case vmIntrinsics::_getSuperclass: 3376 prim_return_value = null(); 3377 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3378 break; 3379 case vmIntrinsics::_getClassAccessFlags: 3380 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3381 return_type = TypeInt::INT; // not bool! 6297094 3382 break; 3383 default: 3384 fatal_unexpected_iid(id); 3385 break; 3386 } 3387 3388 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3389 if (mirror_con == NULL) return false; // cannot happen? 3390 3391 #ifndef PRODUCT 3392 if (C->print_intrinsics() || C->print_inlining()) { 3393 ciType* k = mirror_con->java_mirror_type(); 3394 if (k) { 3395 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3396 k->print_name(); 3397 tty->cr(); 3398 } 3399 } 3400 #endif 3401 3402 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3403 RegionNode* region = new RegionNode(PATH_LIMIT); 3404 record_for_igvn(region); 3405 PhiNode* phi = new PhiNode(region, return_type); 3406 3407 // The mirror will never be null of Reflection.getClassAccessFlags, however 3408 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3409 // if it is. See bug 4774291. 3410 3411 // For Reflection.getClassAccessFlags(), the null check occurs in 3412 // the wrong place; see inline_unsafe_access(), above, for a similar 3413 // situation. 3414 mirror = null_check(mirror); 3415 // If mirror or obj is dead, only null-path is taken. 3416 if (stopped()) return true; 3417 3418 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3419 3420 // Now load the mirror's klass metaobject, and null-check it. 3421 // Side-effects region with the control path if the klass is null. 3422 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3423 // If kls is null, we have a primitive mirror. 3424 phi->init_req(_prim_path, prim_return_value); 3425 if (stopped()) { set_result(region, phi); return true; } 3426 bool safe_for_replace = (region->in(_prim_path) == top()); 3427 3428 Node* p; // handy temp 3429 Node* null_ctl; 3430 3431 // Now that we have the non-null klass, we can perform the real query. 3432 // For constant classes, the query will constant-fold in LoadNode::Value. 3433 Node* query_value = top(); 3434 switch (id) { 3435 case vmIntrinsics::_isInstance: 3436 // nothing is an instance of a primitive type 3437 query_value = gen_instanceof(obj, kls, safe_for_replace); 3438 break; 3439 3440 case vmIntrinsics::_getModifiers: 3441 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3442 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3443 break; 3444 3445 case vmIntrinsics::_isInterface: 3446 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3447 if (generate_interface_guard(kls, region) != NULL) 3448 // A guard was added. If the guard is taken, it was an interface. 3449 phi->add_req(intcon(1)); 3450 // If we fall through, it's a plain class. 3451 query_value = intcon(0); 3452 break; 3453 3454 case vmIntrinsics::_isArray: 3455 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3456 if (generate_array_guard(kls, region) != NULL) 3457 // A guard was added. If the guard is taken, it was an array. 3458 phi->add_req(intcon(1)); 3459 // If we fall through, it's a plain class. 3460 query_value = intcon(0); 3461 break; 3462 3463 case vmIntrinsics::_isPrimitive: 3464 query_value = intcon(0); // "normal" path produces false 3465 break; 3466 3467 case vmIntrinsics::_getSuperclass: 3468 // The rules here are somewhat unfortunate, but we can still do better 3469 // with random logic than with a JNI call. 3470 // Interfaces store null or Object as _super, but must report null. 3471 // Arrays store an intermediate super as _super, but must report Object. 3472 // Other types can report the actual _super. 3473 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3474 if (generate_interface_guard(kls, region) != NULL) 3475 // A guard was added. If the guard is taken, it was an interface. 3476 phi->add_req(null()); 3477 if (generate_array_guard(kls, region) != NULL) 3478 // A guard was added. If the guard is taken, it was an array. 3479 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3480 // If we fall through, it's a plain class. Get its _super. 3481 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3482 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3483 null_ctl = top(); 3484 kls = null_check_oop(kls, &null_ctl); 3485 if (null_ctl != top()) { 3486 // If the guard is taken, Object.superClass is null (both klass and mirror). 3487 region->add_req(null_ctl); 3488 phi ->add_req(null()); 3489 } 3490 if (!stopped()) { 3491 query_value = load_mirror_from_klass(kls); 3492 } 3493 break; 3494 3495 case vmIntrinsics::_getClassAccessFlags: 3496 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3497 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3498 break; 3499 3500 default: 3501 fatal_unexpected_iid(id); 3502 break; 3503 } 3504 3505 // Fall-through is the normal case of a query to a real class. 3506 phi->init_req(1, query_value); 3507 region->init_req(1, control()); 3508 3509 C->set_has_split_ifs(true); // Has chance for split-if optimization 3510 set_result(region, phi); 3511 return true; 3512 } 3513 3514 //-------------------------inline_Class_cast------------------- 3515 bool LibraryCallKit::inline_Class_cast() { 3516 Node* mirror = argument(0); // Class 3517 Node* obj = argument(1); 3518 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3519 if (mirror_con == NULL) { 3520 return false; // dead path (mirror->is_top()). 3521 } 3522 if (obj == NULL || obj->is_top()) { 3523 return false; // dead path 3524 } 3525 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3526 3527 // First, see if Class.cast() can be folded statically. 3528 // java_mirror_type() returns non-null for compile-time Class constants. 3529 ciType* tm = mirror_con->java_mirror_type(); 3530 if (tm != NULL && tm->is_klass() && 3531 tp != NULL && tp->klass() != NULL) { 3532 if (!tp->klass()->is_loaded()) { 3533 // Don't use intrinsic when class is not loaded. 3534 return false; 3535 } else { 3536 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3537 if (static_res == Compile::SSC_always_true) { 3538 // isInstance() is true - fold the code. 3539 set_result(obj); 3540 return true; 3541 } else if (static_res == Compile::SSC_always_false) { 3542 // Don't use intrinsic, have to throw ClassCastException. 3543 // If the reference is null, the non-intrinsic bytecode will 3544 // be optimized appropriately. 3545 return false; 3546 } 3547 } 3548 } 3549 3550 // Bailout intrinsic and do normal inlining if exception path is frequent. 3551 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3552 return false; 3553 } 3554 3555 // Generate dynamic checks. 3556 // Class.cast() is java implementation of _checkcast bytecode. 3557 // Do checkcast (Parse::do_checkcast()) optimizations here. 3558 3559 mirror = null_check(mirror); 3560 // If mirror is dead, only null-path is taken. 3561 if (stopped()) { 3562 return true; 3563 } 3564 3565 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3566 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3567 RegionNode* region = new RegionNode(PATH_LIMIT); 3568 record_for_igvn(region); 3569 3570 // Now load the mirror's klass metaobject, and null-check it. 3571 // If kls is null, we have a primitive mirror and 3572 // nothing is an instance of a primitive type. 3573 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3574 3575 Node* res = top(); 3576 if (!stopped()) { 3577 Node* bad_type_ctrl = top(); 3578 // Do checkcast optimizations. 3579 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3580 region->init_req(_bad_type_path, bad_type_ctrl); 3581 } 3582 if (region->in(_prim_path) != top() || 3583 region->in(_bad_type_path) != top()) { 3584 // Let Interpreter throw ClassCastException. 3585 PreserveJVMState pjvms(this); 3586 set_control(_gvn.transform(region)); 3587 uncommon_trap(Deoptimization::Reason_intrinsic, 3588 Deoptimization::Action_maybe_recompile); 3589 } 3590 if (!stopped()) { 3591 set_result(res); 3592 } 3593 return true; 3594 } 3595 3596 3597 //--------------------------inline_native_subtype_check------------------------ 3598 // This intrinsic takes the JNI calls out of the heart of 3599 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3600 bool LibraryCallKit::inline_native_subtype_check() { 3601 // Pull both arguments off the stack. 3602 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3603 args[0] = argument(0); 3604 args[1] = argument(1); 3605 Node* klasses[2]; // corresponding Klasses: superk, subk 3606 klasses[0] = klasses[1] = top(); 3607 3608 enum { 3609 // A full decision tree on {superc is prim, subc is prim}: 3610 _prim_0_path = 1, // {P,N} => false 3611 // {P,P} & superc!=subc => false 3612 _prim_same_path, // {P,P} & superc==subc => true 3613 _prim_1_path, // {N,P} => false 3614 _ref_subtype_path, // {N,N} & subtype check wins => true 3615 _both_ref_path, // {N,N} & subtype check loses => false 3616 PATH_LIMIT 3617 }; 3618 3619 RegionNode* region = new RegionNode(PATH_LIMIT); 3620 Node* phi = new PhiNode(region, TypeInt::BOOL); 3621 record_for_igvn(region); 3622 3623 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3624 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3625 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3626 3627 // First null-check both mirrors and load each mirror's klass metaobject. 3628 int which_arg; 3629 for (which_arg = 0; which_arg <= 1; which_arg++) { 3630 Node* arg = args[which_arg]; 3631 arg = null_check(arg); 3632 if (stopped()) break; 3633 args[which_arg] = arg; 3634 3635 Node* p = basic_plus_adr(arg, class_klass_offset); 3636 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3637 klasses[which_arg] = _gvn.transform(kls); 3638 } 3639 3640 // Having loaded both klasses, test each for null. 3641 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3642 for (which_arg = 0; which_arg <= 1; which_arg++) { 3643 Node* kls = klasses[which_arg]; 3644 Node* null_ctl = top(); 3645 kls = null_check_oop(kls, &null_ctl, never_see_null); 3646 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3647 region->init_req(prim_path, null_ctl); 3648 if (stopped()) break; 3649 klasses[which_arg] = kls; 3650 } 3651 3652 if (!stopped()) { 3653 // now we have two reference types, in klasses[0..1] 3654 Node* subk = klasses[1]; // the argument to isAssignableFrom 3655 Node* superk = klasses[0]; // the receiver 3656 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3657 // now we have a successful reference subtype check 3658 region->set_req(_ref_subtype_path, control()); 3659 } 3660 3661 // If both operands are primitive (both klasses null), then 3662 // we must return true when they are identical primitives. 3663 // It is convenient to test this after the first null klass check. 3664 set_control(region->in(_prim_0_path)); // go back to first null check 3665 if (!stopped()) { 3666 // Since superc is primitive, make a guard for the superc==subc case. 3667 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3668 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3669 generate_guard(bol_eq, region, PROB_FAIR); 3670 if (region->req() == PATH_LIMIT+1) { 3671 // A guard was added. If the added guard is taken, superc==subc. 3672 region->swap_edges(PATH_LIMIT, _prim_same_path); 3673 region->del_req(PATH_LIMIT); 3674 } 3675 region->set_req(_prim_0_path, control()); // Not equal after all. 3676 } 3677 3678 // these are the only paths that produce 'true': 3679 phi->set_req(_prim_same_path, intcon(1)); 3680 phi->set_req(_ref_subtype_path, intcon(1)); 3681 3682 // pull together the cases: 3683 assert(region->req() == PATH_LIMIT, "sane region"); 3684 for (uint i = 1; i < region->req(); i++) { 3685 Node* ctl = region->in(i); 3686 if (ctl == NULL || ctl == top()) { 3687 region->set_req(i, top()); 3688 phi ->set_req(i, top()); 3689 } else if (phi->in(i) == NULL) { 3690 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3691 } 3692 } 3693 3694 set_control(_gvn.transform(region)); 3695 set_result(_gvn.transform(phi)); 3696 return true; 3697 } 3698 3699 //---------------------generate_array_guard_common------------------------ 3700 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3701 bool obj_array, bool not_array) { 3702 3703 if (stopped()) { 3704 return NULL; 3705 } 3706 3707 // If obj_array/non_array==false/false: 3708 // Branch around if the given klass is in fact an array (either obj or prim). 3709 // If obj_array/non_array==false/true: 3710 // Branch around if the given klass is not an array klass of any kind. 3711 // If obj_array/non_array==true/true: 3712 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3713 // If obj_array/non_array==true/false: 3714 // Branch around if the kls is an oop array (Object[] or subtype) 3715 // 3716 // Like generate_guard, adds a new path onto the region. 3717 jint layout_con = 0; 3718 Node* layout_val = get_layout_helper(kls, layout_con); 3719 if (layout_val == NULL) { 3720 bool query = (obj_array 3721 ? Klass::layout_helper_is_objArray(layout_con) 3722 : Klass::layout_helper_is_array(layout_con)); 3723 if (query == not_array) { 3724 return NULL; // never a branch 3725 } else { // always a branch 3726 Node* always_branch = control(); 3727 if (region != NULL) 3728 region->add_req(always_branch); 3729 set_control(top()); 3730 return always_branch; 3731 } 3732 } 3733 // Now test the correct condition. 3734 jint nval = (obj_array 3735 ? ((jint)Klass::_lh_array_tag_type_value 3736 << Klass::_lh_array_tag_shift) 3737 : Klass::_lh_neutral_value); 3738 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3739 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3740 // invert the test if we are looking for a non-array 3741 if (not_array) btest = BoolTest(btest).negate(); 3742 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3743 return generate_fair_guard(bol, region); 3744 } 3745 3746 3747 //-----------------------inline_native_newArray-------------------------- 3748 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3749 bool LibraryCallKit::inline_native_newArray() { 3750 Node* mirror = argument(0); 3751 Node* count_val = argument(1); 3752 3753 mirror = null_check(mirror); 3754 // If mirror or obj is dead, only null-path is taken. 3755 if (stopped()) return true; 3756 3757 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3758 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3759 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3760 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3761 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3762 3763 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3764 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3765 result_reg, _slow_path); 3766 Node* normal_ctl = control(); 3767 Node* no_array_ctl = result_reg->in(_slow_path); 3768 3769 // Generate code for the slow case. We make a call to newArray(). 3770 set_control(no_array_ctl); 3771 if (!stopped()) { 3772 // Either the input type is void.class, or else the 3773 // array klass has not yet been cached. Either the 3774 // ensuing call will throw an exception, or else it 3775 // will cache the array klass for next time. 3776 PreserveJVMState pjvms(this); 3777 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3778 Node* slow_result = set_results_for_java_call(slow_call); 3779 // this->control() comes from set_results_for_java_call 3780 result_reg->set_req(_slow_path, control()); 3781 result_val->set_req(_slow_path, slow_result); 3782 result_io ->set_req(_slow_path, i_o()); 3783 result_mem->set_req(_slow_path, reset_memory()); 3784 } 3785 3786 set_control(normal_ctl); 3787 if (!stopped()) { 3788 // Normal case: The array type has been cached in the java.lang.Class. 3789 // The following call works fine even if the array type is polymorphic. 3790 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3791 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3792 result_reg->init_req(_normal_path, control()); 3793 result_val->init_req(_normal_path, obj); 3794 result_io ->init_req(_normal_path, i_o()); 3795 result_mem->init_req(_normal_path, reset_memory()); 3796 } 3797 3798 // Return the combined state. 3799 set_i_o( _gvn.transform(result_io) ); 3800 set_all_memory( _gvn.transform(result_mem)); 3801 3802 C->set_has_split_ifs(true); // Has chance for split-if optimization 3803 set_result(result_reg, result_val); 3804 return true; 3805 } 3806 3807 //----------------------inline_native_getLength-------------------------- 3808 // public static native int java.lang.reflect.Array.getLength(Object array); 3809 bool LibraryCallKit::inline_native_getLength() { 3810 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3811 3812 Node* array = null_check(argument(0)); 3813 // If array is dead, only null-path is taken. 3814 if (stopped()) return true; 3815 3816 // Deoptimize if it is a non-array. 3817 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3818 3819 if (non_array != NULL) { 3820 PreserveJVMState pjvms(this); 3821 set_control(non_array); 3822 uncommon_trap(Deoptimization::Reason_intrinsic, 3823 Deoptimization::Action_maybe_recompile); 3824 } 3825 3826 // If control is dead, only non-array-path is taken. 3827 if (stopped()) return true; 3828 3829 // The works fine even if the array type is polymorphic. 3830 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3831 Node* result = load_array_length(array); 3832 3833 C->set_has_split_ifs(true); // Has chance for split-if optimization 3834 set_result(result); 3835 return true; 3836 } 3837 3838 //------------------------inline_array_copyOf---------------------------- 3839 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3840 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3841 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3842 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3843 3844 // Get the arguments. 3845 Node* original = argument(0); 3846 Node* start = is_copyOfRange? argument(1): intcon(0); 3847 Node* end = is_copyOfRange? argument(2): argument(1); 3848 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3849 3850 Node* newcopy = NULL; 3851 3852 // Set the original stack and the reexecute bit for the interpreter to reexecute 3853 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3854 { PreserveReexecuteState preexecs(this); 3855 jvms()->set_should_reexecute(true); 3856 3857 array_type_mirror = null_check(array_type_mirror); 3858 original = null_check(original); 3859 3860 // Check if a null path was taken unconditionally. 3861 if (stopped()) return true; 3862 3863 Node* orig_length = load_array_length(original); 3864 3865 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3866 klass_node = null_check(klass_node); 3867 3868 RegionNode* bailout = new RegionNode(1); 3869 record_for_igvn(bailout); 3870 3871 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3872 // Bail out if that is so. 3873 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3874 if (not_objArray != NULL) { 3875 // Improve the klass node's type from the new optimistic assumption: 3876 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3877 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3878 Node* cast = new CastPPNode(klass_node, akls); 3879 cast->init_req(0, control()); 3880 klass_node = _gvn.transform(cast); 3881 } 3882 3883 // Bail out if either start or end is negative. 3884 generate_negative_guard(start, bailout, &start); 3885 generate_negative_guard(end, bailout, &end); 3886 3887 Node* length = end; 3888 if (_gvn.type(start) != TypeInt::ZERO) { 3889 length = _gvn.transform(new SubINode(end, start)); 3890 } 3891 3892 // Bail out if length is negative. 3893 // Without this the new_array would throw 3894 // NegativeArraySizeException but IllegalArgumentException is what 3895 // should be thrown 3896 generate_negative_guard(length, bailout, &length); 3897 3898 if (bailout->req() > 1) { 3899 PreserveJVMState pjvms(this); 3900 set_control(_gvn.transform(bailout)); 3901 uncommon_trap(Deoptimization::Reason_intrinsic, 3902 Deoptimization::Action_maybe_recompile); 3903 } 3904 3905 if (!stopped()) { 3906 // How many elements will we copy from the original? 3907 // The answer is MinI(orig_length - start, length). 3908 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3909 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3910 3911 // Generate a direct call to the right arraycopy function(s). 3912 // We know the copy is disjoint but we might not know if the 3913 // oop stores need checking. 3914 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3915 // This will fail a store-check if x contains any non-nulls. 3916 3917 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3918 // loads/stores but it is legal only if we're sure the 3919 // Arrays.copyOf would succeed. So we need all input arguments 3920 // to the copyOf to be validated, including that the copy to the 3921 // new array won't trigger an ArrayStoreException. That subtype 3922 // check can be optimized if we know something on the type of 3923 // the input array from type speculation. 3924 if (_gvn.type(klass_node)->singleton()) { 3925 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3926 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3927 3928 int test = C->static_subtype_check(superk, subk); 3929 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3930 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3931 if (t_original->speculative_type() != NULL) { 3932 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3933 } 3934 } 3935 } 3936 3937 bool validated = false; 3938 // Reason_class_check rather than Reason_intrinsic because we 3939 // want to intrinsify even if this traps. 3940 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3941 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3942 klass_node); 3943 3944 if (not_subtype_ctrl != top()) { 3945 PreserveJVMState pjvms(this); 3946 set_control(not_subtype_ctrl); 3947 uncommon_trap(Deoptimization::Reason_class_check, 3948 Deoptimization::Action_make_not_entrant); 3949 assert(stopped(), "Should be stopped"); 3950 } 3951 validated = true; 3952 } 3953 3954 if (!stopped()) { 3955 newcopy = new_array(klass_node, length, 0); // no arguments to push 3956 3957 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3958 load_object_klass(original), klass_node); 3959 if (!is_copyOfRange) { 3960 ac->set_copyof(validated); 3961 } else { 3962 ac->set_copyofrange(validated); 3963 } 3964 Node* n = _gvn.transform(ac); 3965 if (n == ac) { 3966 ac->connect_outputs(this); 3967 } else { 3968 assert(validated, "shouldn't transform if all arguments not validated"); 3969 set_all_memory(n); 3970 } 3971 } 3972 } 3973 } // original reexecute is set back here 3974 3975 C->set_has_split_ifs(true); // Has chance for split-if optimization 3976 if (!stopped()) { 3977 set_result(newcopy); 3978 } 3979 return true; 3980 } 3981 3982 3983 //----------------------generate_virtual_guard--------------------------- 3984 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3985 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3986 RegionNode* slow_region) { 3987 ciMethod* method = callee(); 3988 int vtable_index = method->vtable_index(); 3989 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3990 "bad index %d", vtable_index); 3991 // Get the Method* out of the appropriate vtable entry. 3992 int entry_offset = (InstanceKlass::vtable_start_offset() + 3993 vtable_index*vtableEntry::size()) * wordSize + 3994 vtableEntry::method_offset_in_bytes(); 3995 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3996 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3997 3998 // Compare the target method with the expected method (e.g., Object.hashCode). 3999 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4000 4001 Node* native_call = makecon(native_call_addr); 4002 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4003 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4004 4005 return generate_slow_guard(test_native, slow_region); 4006 } 4007 4008 //-----------------------generate_method_call---------------------------- 4009 // Use generate_method_call to make a slow-call to the real 4010 // method if the fast path fails. An alternative would be to 4011 // use a stub like OptoRuntime::slow_arraycopy_Java. 4012 // This only works for expanding the current library call, 4013 // not another intrinsic. (E.g., don't use this for making an 4014 // arraycopy call inside of the copyOf intrinsic.) 4015 CallJavaNode* 4016 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4017 // When compiling the intrinsic method itself, do not use this technique. 4018 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4019 4020 ciMethod* method = callee(); 4021 // ensure the JVMS we have will be correct for this call 4022 guarantee(method_id == method->intrinsic_id(), "must match"); 4023 4024 const TypeFunc* tf = TypeFunc::make(method); 4025 CallJavaNode* slow_call; 4026 if (is_static) { 4027 assert(!is_virtual, ""); 4028 slow_call = new CallStaticJavaNode(C, tf, 4029 SharedRuntime::get_resolve_static_call_stub(), 4030 method, bci()); 4031 } else if (is_virtual) { 4032 null_check_receiver(); 4033 int vtable_index = Method::invalid_vtable_index; 4034 if (UseInlineCaches) { 4035 // Suppress the vtable call 4036 } else { 4037 // hashCode and clone are not a miranda methods, 4038 // so the vtable index is fixed. 4039 // No need to use the linkResolver to get it. 4040 vtable_index = method->vtable_index(); 4041 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4042 "bad index %d", vtable_index); 4043 } 4044 slow_call = new CallDynamicJavaNode(tf, 4045 SharedRuntime::get_resolve_virtual_call_stub(), 4046 method, vtable_index, bci()); 4047 } else { // neither virtual nor static: opt_virtual 4048 null_check_receiver(); 4049 slow_call = new CallStaticJavaNode(C, tf, 4050 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4051 method, bci()); 4052 slow_call->set_optimized_virtual(true); 4053 } 4054 set_arguments_for_java_call(slow_call); 4055 set_edges_for_java_call(slow_call); 4056 return slow_call; 4057 } 4058 4059 4060 /** 4061 * Build special case code for calls to hashCode on an object. This call may 4062 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4063 * slightly different code. 4064 */ 4065 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4066 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4067 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4068 4069 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4070 4071 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4072 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4073 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4074 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4075 Node* obj = NULL; 4076 if (!is_static) { 4077 // Check for hashing null object 4078 obj = null_check_receiver(); 4079 if (stopped()) return true; // unconditionally null 4080 result_reg->init_req(_null_path, top()); 4081 result_val->init_req(_null_path, top()); 4082 } else { 4083 // Do a null check, and return zero if null. 4084 // System.identityHashCode(null) == 0 4085 obj = argument(0); 4086 Node* null_ctl = top(); 4087 obj = null_check_oop(obj, &null_ctl); 4088 result_reg->init_req(_null_path, null_ctl); 4089 result_val->init_req(_null_path, _gvn.intcon(0)); 4090 } 4091 4092 // Unconditionally null? Then return right away. 4093 if (stopped()) { 4094 set_control( result_reg->in(_null_path)); 4095 if (!stopped()) 4096 set_result(result_val->in(_null_path)); 4097 return true; 4098 } 4099 4100 // We only go to the fast case code if we pass a number of guards. The 4101 // paths which do not pass are accumulated in the slow_region. 4102 RegionNode* slow_region = new RegionNode(1); 4103 record_for_igvn(slow_region); 4104 4105 // If this is a virtual call, we generate a funny guard. We pull out 4106 // the vtable entry corresponding to hashCode() from the target object. 4107 // If the target method which we are calling happens to be the native 4108 // Object hashCode() method, we pass the guard. We do not need this 4109 // guard for non-virtual calls -- the caller is known to be the native 4110 // Object hashCode(). 4111 if (is_virtual) { 4112 // After null check, get the object's klass. 4113 Node* obj_klass = load_object_klass(obj); 4114 generate_virtual_guard(obj_klass, slow_region); 4115 } 4116 4117 // Get the header out of the object, use LoadMarkNode when available 4118 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4119 // The control of the load must be NULL. Otherwise, the load can move before 4120 // the null check after castPP removal. 4121 Node* no_ctrl = NULL; 4122 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4123 4124 // Test the header to see if it is unlocked. 4125 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4126 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4127 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4128 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4129 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4130 4131 generate_slow_guard(test_unlocked, slow_region); 4132 4133 // Get the hash value and check to see that it has been properly assigned. 4134 // We depend on hash_mask being at most 32 bits and avoid the use of 4135 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4136 // vm: see markOop.hpp. 4137 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4138 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4139 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4140 // This hack lets the hash bits live anywhere in the mark object now, as long 4141 // as the shift drops the relevant bits into the low 32 bits. Note that 4142 // Java spec says that HashCode is an int so there's no point in capturing 4143 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4144 hshifted_header = ConvX2I(hshifted_header); 4145 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4146 4147 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4148 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4149 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4150 4151 generate_slow_guard(test_assigned, slow_region); 4152 4153 Node* init_mem = reset_memory(); 4154 // fill in the rest of the null path: 4155 result_io ->init_req(_null_path, i_o()); 4156 result_mem->init_req(_null_path, init_mem); 4157 4158 result_val->init_req(_fast_path, hash_val); 4159 result_reg->init_req(_fast_path, control()); 4160 result_io ->init_req(_fast_path, i_o()); 4161 result_mem->init_req(_fast_path, init_mem); 4162 4163 // Generate code for the slow case. We make a call to hashCode(). 4164 set_control(_gvn.transform(slow_region)); 4165 if (!stopped()) { 4166 // No need for PreserveJVMState, because we're using up the present state. 4167 set_all_memory(init_mem); 4168 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4169 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4170 Node* slow_result = set_results_for_java_call(slow_call); 4171 // this->control() comes from set_results_for_java_call 4172 result_reg->init_req(_slow_path, control()); 4173 result_val->init_req(_slow_path, slow_result); 4174 result_io ->set_req(_slow_path, i_o()); 4175 result_mem ->set_req(_slow_path, reset_memory()); 4176 } 4177 4178 // Return the combined state. 4179 set_i_o( _gvn.transform(result_io) ); 4180 set_all_memory( _gvn.transform(result_mem)); 4181 4182 set_result(result_reg, result_val); 4183 return true; 4184 } 4185 4186 //---------------------------inline_native_getClass---------------------------- 4187 // public final native Class<?> java.lang.Object.getClass(); 4188 // 4189 // Build special case code for calls to getClass on an object. 4190 bool LibraryCallKit::inline_native_getClass() { 4191 Node* obj = null_check_receiver(); 4192 if (stopped()) return true; 4193 set_result(load_mirror_from_klass(load_object_klass(obj))); 4194 return true; 4195 } 4196 4197 //-----------------inline_native_Reflection_getCallerClass--------------------- 4198 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4199 // 4200 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4201 // 4202 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4203 // in that it must skip particular security frames and checks for 4204 // caller sensitive methods. 4205 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4206 #ifndef PRODUCT 4207 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4208 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4209 } 4210 #endif 4211 4212 if (!jvms()->has_method()) { 4213 #ifndef PRODUCT 4214 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4215 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4216 } 4217 #endif 4218 return false; 4219 } 4220 4221 // Walk back up the JVM state to find the caller at the required 4222 // depth. 4223 JVMState* caller_jvms = jvms(); 4224 4225 // Cf. JVM_GetCallerClass 4226 // NOTE: Start the loop at depth 1 because the current JVM state does 4227 // not include the Reflection.getCallerClass() frame. 4228 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4229 ciMethod* m = caller_jvms->method(); 4230 switch (n) { 4231 case 0: 4232 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4233 break; 4234 case 1: 4235 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4236 if (!m->caller_sensitive()) { 4237 #ifndef PRODUCT 4238 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4239 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4240 } 4241 #endif 4242 return false; // bail-out; let JVM_GetCallerClass do the work 4243 } 4244 break; 4245 default: 4246 if (!m->is_ignored_by_security_stack_walk()) { 4247 // We have reached the desired frame; return the holder class. 4248 // Acquire method holder as java.lang.Class and push as constant. 4249 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4250 ciInstance* caller_mirror = caller_klass->java_mirror(); 4251 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4252 4253 #ifndef PRODUCT 4254 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4255 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()); 4256 tty->print_cr(" JVM state at this point:"); 4257 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4258 ciMethod* m = jvms()->of_depth(i)->method(); 4259 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4260 } 4261 } 4262 #endif 4263 return true; 4264 } 4265 break; 4266 } 4267 } 4268 4269 #ifndef PRODUCT 4270 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4271 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4272 tty->print_cr(" JVM state at this point:"); 4273 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4274 ciMethod* m = jvms()->of_depth(i)->method(); 4275 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4276 } 4277 } 4278 #endif 4279 4280 return false; // bail-out; let JVM_GetCallerClass do the work 4281 } 4282 4283 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4284 Node* arg = argument(0); 4285 Node* result = NULL; 4286 4287 switch (id) { 4288 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4289 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4290 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4291 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4292 4293 case vmIntrinsics::_doubleToLongBits: { 4294 // two paths (plus control) merge in a wood 4295 RegionNode *r = new RegionNode(3); 4296 Node *phi = new PhiNode(r, TypeLong::LONG); 4297 4298 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4299 // Build the boolean node 4300 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4301 4302 // Branch either way. 4303 // NaN case is less traveled, which makes all the difference. 4304 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4305 Node *opt_isnan = _gvn.transform(ifisnan); 4306 assert( opt_isnan->is_If(), "Expect an IfNode"); 4307 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4308 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4309 4310 set_control(iftrue); 4311 4312 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4313 Node *slow_result = longcon(nan_bits); // return NaN 4314 phi->init_req(1, _gvn.transform( slow_result )); 4315 r->init_req(1, iftrue); 4316 4317 // Else fall through 4318 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4319 set_control(iffalse); 4320 4321 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4322 r->init_req(2, iffalse); 4323 4324 // Post merge 4325 set_control(_gvn.transform(r)); 4326 record_for_igvn(r); 4327 4328 C->set_has_split_ifs(true); // Has chance for split-if optimization 4329 result = phi; 4330 assert(result->bottom_type()->isa_long(), "must be"); 4331 break; 4332 } 4333 4334 case vmIntrinsics::_floatToIntBits: { 4335 // two paths (plus control) merge in a wood 4336 RegionNode *r = new RegionNode(3); 4337 Node *phi = new PhiNode(r, TypeInt::INT); 4338 4339 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4340 // Build the boolean node 4341 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4342 4343 // Branch either way. 4344 // NaN case is less traveled, which makes all the difference. 4345 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4346 Node *opt_isnan = _gvn.transform(ifisnan); 4347 assert( opt_isnan->is_If(), "Expect an IfNode"); 4348 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4349 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4350 4351 set_control(iftrue); 4352 4353 static const jint nan_bits = 0x7fc00000; 4354 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4355 phi->init_req(1, _gvn.transform( slow_result )); 4356 r->init_req(1, iftrue); 4357 4358 // Else fall through 4359 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4360 set_control(iffalse); 4361 4362 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4363 r->init_req(2, iffalse); 4364 4365 // Post merge 4366 set_control(_gvn.transform(r)); 4367 record_for_igvn(r); 4368 4369 C->set_has_split_ifs(true); // Has chance for split-if optimization 4370 result = phi; 4371 assert(result->bottom_type()->isa_int(), "must be"); 4372 break; 4373 } 4374 4375 default: 4376 fatal_unexpected_iid(id); 4377 break; 4378 } 4379 set_result(_gvn.transform(result)); 4380 return true; 4381 } 4382 4383 //----------------------inline_unsafe_copyMemory------------------------- 4384 // public native void Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4385 bool LibraryCallKit::inline_unsafe_copyMemory() { 4386 if (callee()->is_static()) return false; // caller must have the capability! 4387 null_check_receiver(); // null-check receiver 4388 if (stopped()) return true; 4389 4390 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4391 4392 Node* src_ptr = argument(1); // type: oop 4393 Node* src_off = ConvL2X(argument(2)); // type: long 4394 Node* dst_ptr = argument(4); // type: oop 4395 Node* dst_off = ConvL2X(argument(5)); // type: long 4396 Node* size = ConvL2X(argument(7)); // type: long 4397 4398 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4399 "fieldOffset must be byte-scaled"); 4400 4401 Node* src = make_unsafe_address(src_ptr, src_off); 4402 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4403 4404 // Conservatively insert a memory barrier on all memory slices. 4405 // Do not let writes of the copy source or destination float below the copy. 4406 insert_mem_bar(Op_MemBarCPUOrder); 4407 4408 // Call it. Note that the length argument is not scaled. 4409 make_runtime_call(RC_LEAF|RC_NO_FP, 4410 OptoRuntime::fast_arraycopy_Type(), 4411 StubRoutines::unsafe_arraycopy(), 4412 "unsafe_arraycopy", 4413 TypeRawPtr::BOTTOM, 4414 src, dst, size XTOP); 4415 4416 // Do not let reads of the copy destination float above the copy. 4417 insert_mem_bar(Op_MemBarCPUOrder); 4418 4419 return true; 4420 } 4421 4422 //------------------------clone_coping----------------------------------- 4423 // Helper function for inline_native_clone. 4424 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4425 assert(obj_size != NULL, ""); 4426 Node* raw_obj = alloc_obj->in(1); 4427 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4428 4429 AllocateNode* alloc = NULL; 4430 if (ReduceBulkZeroing) { 4431 // We will be completely responsible for initializing this object - 4432 // mark Initialize node as complete. 4433 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4434 // The object was just allocated - there should be no any stores! 4435 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4436 // Mark as complete_with_arraycopy so that on AllocateNode 4437 // expansion, we know this AllocateNode is initialized by an array 4438 // copy and a StoreStore barrier exists after the array copy. 4439 alloc->initialization()->set_complete_with_arraycopy(); 4440 } 4441 4442 // Copy the fastest available way. 4443 // TODO: generate fields copies for small objects instead. 4444 Node* src = obj; 4445 Node* dest = alloc_obj; 4446 Node* size = _gvn.transform(obj_size); 4447 4448 // Exclude the header but include array length to copy by 8 bytes words. 4449 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4450 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4451 instanceOopDesc::base_offset_in_bytes(); 4452 // base_off: 4453 // 8 - 32-bit VM 4454 // 12 - 64-bit VM, compressed klass 4455 // 16 - 64-bit VM, normal klass 4456 if (base_off % BytesPerLong != 0) { 4457 assert(UseCompressedClassPointers, ""); 4458 if (is_array) { 4459 // Exclude length to copy by 8 bytes words. 4460 base_off += sizeof(int); 4461 } else { 4462 // Include klass to copy by 8 bytes words. 4463 base_off = instanceOopDesc::klass_offset_in_bytes(); 4464 } 4465 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4466 } 4467 src = basic_plus_adr(src, base_off); 4468 dest = basic_plus_adr(dest, base_off); 4469 4470 // Compute the length also, if needed: 4471 Node* countx = size; 4472 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4473 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4474 4475 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4476 4477 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4478 ac->set_clonebasic(); 4479 Node* n = _gvn.transform(ac); 4480 if (n == ac) { 4481 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4482 } else { 4483 set_all_memory(n); 4484 } 4485 4486 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4487 if (card_mark) { 4488 assert(!is_array, ""); 4489 // Put in store barrier for any and all oops we are sticking 4490 // into this object. (We could avoid this if we could prove 4491 // that the object type contains no oop fields at all.) 4492 Node* no_particular_value = NULL; 4493 Node* no_particular_field = NULL; 4494 int raw_adr_idx = Compile::AliasIdxRaw; 4495 post_barrier(control(), 4496 memory(raw_adr_type), 4497 alloc_obj, 4498 no_particular_field, 4499 raw_adr_idx, 4500 no_particular_value, 4501 T_OBJECT, 4502 false); 4503 } 4504 4505 // Do not let reads from the cloned object float above the arraycopy. 4506 if (alloc != NULL) { 4507 // Do not let stores that initialize this object be reordered with 4508 // a subsequent store that would make this object accessible by 4509 // other threads. 4510 // Record what AllocateNode this StoreStore protects so that 4511 // escape analysis can go from the MemBarStoreStoreNode to the 4512 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4513 // based on the escape status of the AllocateNode. 4514 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4515 } else { 4516 insert_mem_bar(Op_MemBarCPUOrder); 4517 } 4518 } 4519 4520 //------------------------inline_native_clone---------------------------- 4521 // protected native Object java.lang.Object.clone(); 4522 // 4523 // Here are the simple edge cases: 4524 // null receiver => normal trap 4525 // virtual and clone was overridden => slow path to out-of-line clone 4526 // not cloneable or finalizer => slow path to out-of-line Object.clone 4527 // 4528 // The general case has two steps, allocation and copying. 4529 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4530 // 4531 // Copying also has two cases, oop arrays and everything else. 4532 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4533 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4534 // 4535 // These steps fold up nicely if and when the cloned object's klass 4536 // can be sharply typed as an object array, a type array, or an instance. 4537 // 4538 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4539 PhiNode* result_val; 4540 4541 // Set the reexecute bit for the interpreter to reexecute 4542 // the bytecode that invokes Object.clone if deoptimization happens. 4543 { PreserveReexecuteState preexecs(this); 4544 jvms()->set_should_reexecute(true); 4545 4546 Node* obj = null_check_receiver(); 4547 if (stopped()) return true; 4548 4549 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4550 4551 // If we are going to clone an instance, we need its exact type to 4552 // know the number and types of fields to convert the clone to 4553 // loads/stores. Maybe a speculative type can help us. 4554 if (!obj_type->klass_is_exact() && 4555 obj_type->speculative_type() != NULL && 4556 obj_type->speculative_type()->is_instance_klass()) { 4557 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4558 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4559 !spec_ik->has_injected_fields()) { 4560 ciKlass* k = obj_type->klass(); 4561 if (!k->is_instance_klass() || 4562 k->as_instance_klass()->is_interface() || 4563 k->as_instance_klass()->has_subklass()) { 4564 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4565 } 4566 } 4567 } 4568 4569 Node* obj_klass = load_object_klass(obj); 4570 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4571 const TypeOopPtr* toop = ((tklass != NULL) 4572 ? tklass->as_instance_type() 4573 : TypeInstPtr::NOTNULL); 4574 4575 // Conservatively insert a memory barrier on all memory slices. 4576 // Do not let writes into the original float below the clone. 4577 insert_mem_bar(Op_MemBarCPUOrder); 4578 4579 // paths into result_reg: 4580 enum { 4581 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4582 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4583 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4584 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4585 PATH_LIMIT 4586 }; 4587 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4588 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4589 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4590 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4591 record_for_igvn(result_reg); 4592 4593 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4594 int raw_adr_idx = Compile::AliasIdxRaw; 4595 4596 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4597 if (array_ctl != NULL) { 4598 // It's an array. 4599 PreserveJVMState pjvms(this); 4600 set_control(array_ctl); 4601 Node* obj_length = load_array_length(obj); 4602 Node* obj_size = NULL; 4603 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4604 4605 if (!use_ReduceInitialCardMarks()) { 4606 // If it is an oop array, it requires very special treatment, 4607 // because card marking is required on each card of the array. 4608 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4609 if (is_obja != NULL) { 4610 PreserveJVMState pjvms2(this); 4611 set_control(is_obja); 4612 // Generate a direct call to the right arraycopy function(s). 4613 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4614 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4615 ac->set_cloneoop(); 4616 Node* n = _gvn.transform(ac); 4617 assert(n == ac, "cannot disappear"); 4618 ac->connect_outputs(this); 4619 4620 result_reg->init_req(_objArray_path, control()); 4621 result_val->init_req(_objArray_path, alloc_obj); 4622 result_i_o ->set_req(_objArray_path, i_o()); 4623 result_mem ->set_req(_objArray_path, reset_memory()); 4624 } 4625 } 4626 // Otherwise, there are no card marks to worry about. 4627 // (We can dispense with card marks if we know the allocation 4628 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4629 // causes the non-eden paths to take compensating steps to 4630 // simulate a fresh allocation, so that no further 4631 // card marks are required in compiled code to initialize 4632 // the object.) 4633 4634 if (!stopped()) { 4635 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4636 4637 // Present the results of the copy. 4638 result_reg->init_req(_array_path, control()); 4639 result_val->init_req(_array_path, alloc_obj); 4640 result_i_o ->set_req(_array_path, i_o()); 4641 result_mem ->set_req(_array_path, reset_memory()); 4642 } 4643 } 4644 4645 // We only go to the instance fast case code if we pass a number of guards. 4646 // The paths which do not pass are accumulated in the slow_region. 4647 RegionNode* slow_region = new RegionNode(1); 4648 record_for_igvn(slow_region); 4649 if (!stopped()) { 4650 // It's an instance (we did array above). Make the slow-path tests. 4651 // If this is a virtual call, we generate a funny guard. We grab 4652 // the vtable entry corresponding to clone() from the target object. 4653 // If the target method which we are calling happens to be the 4654 // Object clone() method, we pass the guard. We do not need this 4655 // guard for non-virtual calls; the caller is known to be the native 4656 // Object clone(). 4657 if (is_virtual) { 4658 generate_virtual_guard(obj_klass, slow_region); 4659 } 4660 4661 // The object must be cloneable and must not have a finalizer. 4662 // Both of these conditions may be checked in a single test. 4663 // We could optimize the cloneable test further, but we don't care. 4664 generate_access_flags_guard(obj_klass, 4665 // Test both conditions: 4666 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4667 // Must be cloneable but not finalizer: 4668 JVM_ACC_IS_CLONEABLE, 4669 slow_region); 4670 } 4671 4672 if (!stopped()) { 4673 // It's an instance, and it passed the slow-path tests. 4674 PreserveJVMState pjvms(this); 4675 Node* obj_size = NULL; 4676 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4677 // is reexecuted if deoptimization occurs and there could be problems when merging 4678 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4679 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4680 4681 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4682 4683 // Present the results of the slow call. 4684 result_reg->init_req(_instance_path, control()); 4685 result_val->init_req(_instance_path, alloc_obj); 4686 result_i_o ->set_req(_instance_path, i_o()); 4687 result_mem ->set_req(_instance_path, reset_memory()); 4688 } 4689 4690 // Generate code for the slow case. We make a call to clone(). 4691 set_control(_gvn.transform(slow_region)); 4692 if (!stopped()) { 4693 PreserveJVMState pjvms(this); 4694 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4695 Node* slow_result = set_results_for_java_call(slow_call); 4696 // this->control() comes from set_results_for_java_call 4697 result_reg->init_req(_slow_path, control()); 4698 result_val->init_req(_slow_path, slow_result); 4699 result_i_o ->set_req(_slow_path, i_o()); 4700 result_mem ->set_req(_slow_path, reset_memory()); 4701 } 4702 4703 // Return the combined state. 4704 set_control( _gvn.transform(result_reg)); 4705 set_i_o( _gvn.transform(result_i_o)); 4706 set_all_memory( _gvn.transform(result_mem)); 4707 } // original reexecute is set back here 4708 4709 set_result(_gvn.transform(result_val)); 4710 return true; 4711 } 4712 4713 // If we have a tighly coupled allocation, the arraycopy may take care 4714 // of the array initialization. If one of the guards we insert between 4715 // the allocation and the arraycopy causes a deoptimization, an 4716 // unitialized array will escape the compiled method. To prevent that 4717 // we set the JVM state for uncommon traps between the allocation and 4718 // the arraycopy to the state before the allocation so, in case of 4719 // deoptimization, we'll reexecute the allocation and the 4720 // initialization. 4721 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4722 if (alloc != NULL) { 4723 ciMethod* trap_method = alloc->jvms()->method(); 4724 int trap_bci = alloc->jvms()->bci(); 4725 4726 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4727 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4728 // Make sure there's no store between the allocation and the 4729 // arraycopy otherwise visible side effects could be rexecuted 4730 // in case of deoptimization and cause incorrect execution. 4731 bool no_interfering_store = true; 4732 Node* mem = alloc->in(TypeFunc::Memory); 4733 if (mem->is_MergeMem()) { 4734 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4735 Node* n = mms.memory(); 4736 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4737 assert(n->is_Store(), "what else?"); 4738 no_interfering_store = false; 4739 break; 4740 } 4741 } 4742 } else { 4743 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4744 Node* n = mms.memory(); 4745 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4746 assert(n->is_Store(), "what else?"); 4747 no_interfering_store = false; 4748 break; 4749 } 4750 } 4751 } 4752 4753 if (no_interfering_store) { 4754 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4755 uint size = alloc->req(); 4756 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4757 old_jvms->set_map(sfpt); 4758 for (uint i = 0; i < size; i++) { 4759 sfpt->init_req(i, alloc->in(i)); 4760 } 4761 // re-push array length for deoptimization 4762 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4763 old_jvms->set_sp(old_jvms->sp()+1); 4764 old_jvms->set_monoff(old_jvms->monoff()+1); 4765 old_jvms->set_scloff(old_jvms->scloff()+1); 4766 old_jvms->set_endoff(old_jvms->endoff()+1); 4767 old_jvms->set_should_reexecute(true); 4768 4769 sfpt->set_i_o(map()->i_o()); 4770 sfpt->set_memory(map()->memory()); 4771 sfpt->set_control(map()->control()); 4772 4773 JVMState* saved_jvms = jvms(); 4774 saved_reexecute_sp = _reexecute_sp; 4775 4776 set_jvms(sfpt->jvms()); 4777 _reexecute_sp = jvms()->sp(); 4778 4779 return saved_jvms; 4780 } 4781 } 4782 } 4783 return NULL; 4784 } 4785 4786 // In case of a deoptimization, we restart execution at the 4787 // allocation, allocating a new array. We would leave an uninitialized 4788 // array in the heap that GCs wouldn't expect. Move the allocation 4789 // after the traps so we don't allocate the array if we 4790 // deoptimize. This is possible because tightly_coupled_allocation() 4791 // guarantees there's no observer of the allocated array at this point 4792 // and the control flow is simple enough. 4793 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4794 if (saved_jvms != NULL && !stopped()) { 4795 assert(alloc != NULL, "only with a tightly coupled allocation"); 4796 // restore JVM state to the state at the arraycopy 4797 saved_jvms->map()->set_control(map()->control()); 4798 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4799 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4800 // If we've improved the types of some nodes (null check) while 4801 // emitting the guards, propagate them to the current state 4802 map()->replaced_nodes().apply(saved_jvms->map()); 4803 set_jvms(saved_jvms); 4804 _reexecute_sp = saved_reexecute_sp; 4805 4806 // Remove the allocation from above the guards 4807 CallProjections callprojs; 4808 alloc->extract_projections(&callprojs, true); 4809 InitializeNode* init = alloc->initialization(); 4810 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4811 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4812 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4813 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4814 4815 // move the allocation here (after the guards) 4816 _gvn.hash_delete(alloc); 4817 alloc->set_req(TypeFunc::Control, control()); 4818 alloc->set_req(TypeFunc::I_O, i_o()); 4819 Node *mem = reset_memory(); 4820 set_all_memory(mem); 4821 alloc->set_req(TypeFunc::Memory, mem); 4822 set_control(init->proj_out(TypeFunc::Control)); 4823 set_i_o(callprojs.fallthrough_ioproj); 4824 4825 // Update memory as done in GraphKit::set_output_for_allocation() 4826 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4827 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4828 if (ary_type->isa_aryptr() && length_type != NULL) { 4829 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4830 } 4831 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4832 int elemidx = C->get_alias_index(telemref); 4833 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4834 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4835 4836 Node* allocx = _gvn.transform(alloc); 4837 assert(allocx == alloc, "where has the allocation gone?"); 4838 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4839 4840 _gvn.hash_delete(dest); 4841 dest->set_req(0, control()); 4842 Node* destx = _gvn.transform(dest); 4843 assert(destx == dest, "where has the allocation result gone?"); 4844 } 4845 } 4846 4847 4848 //------------------------------inline_arraycopy----------------------- 4849 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4850 // Object dest, int destPos, 4851 // int length); 4852 bool LibraryCallKit::inline_arraycopy() { 4853 // Get the arguments. 4854 Node* src = argument(0); // type: oop 4855 Node* src_offset = argument(1); // type: int 4856 Node* dest = argument(2); // type: oop 4857 Node* dest_offset = argument(3); // type: int 4858 Node* length = argument(4); // type: int 4859 4860 4861 // Check for allocation before we add nodes that would confuse 4862 // tightly_coupled_allocation() 4863 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4864 4865 int saved_reexecute_sp = -1; 4866 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4867 // See arraycopy_restore_alloc_state() comment 4868 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4869 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4870 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4871 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4872 4873 // The following tests must be performed 4874 // (1) src and dest are arrays. 4875 // (2) src and dest arrays must have elements of the same BasicType 4876 // (3) src and dest must not be null. 4877 // (4) src_offset must not be negative. 4878 // (5) dest_offset must not be negative. 4879 // (6) length must not be negative. 4880 // (7) src_offset + length must not exceed length of src. 4881 // (8) dest_offset + length must not exceed length of dest. 4882 // (9) each element of an oop array must be assignable 4883 4884 // (3) src and dest must not be null. 4885 // always do this here because we need the JVM state for uncommon traps 4886 Node* null_ctl = top(); 4887 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4888 assert(null_ctl->is_top(), "no null control here"); 4889 dest = null_check(dest, T_ARRAY); 4890 4891 if (!can_emit_guards) { 4892 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4893 // guards but the arraycopy node could still take advantage of a 4894 // tightly allocated allocation. tightly_coupled_allocation() is 4895 // called again to make sure it takes the null check above into 4896 // account: the null check is mandatory and if it caused an 4897 // uncommon trap to be emitted then the allocation can't be 4898 // considered tightly coupled in this context. 4899 alloc = tightly_coupled_allocation(dest, NULL); 4900 } 4901 4902 bool validated = false; 4903 4904 const Type* src_type = _gvn.type(src); 4905 const Type* dest_type = _gvn.type(dest); 4906 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4907 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4908 4909 // Do we have the type of src? 4910 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4911 // Do we have the type of dest? 4912 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4913 // Is the type for src from speculation? 4914 bool src_spec = false; 4915 // Is the type for dest from speculation? 4916 bool dest_spec = false; 4917 4918 if ((!has_src || !has_dest) && can_emit_guards) { 4919 // We don't have sufficient type information, let's see if 4920 // speculative types can help. We need to have types for both src 4921 // and dest so that it pays off. 4922 4923 // Do we already have or could we have type information for src 4924 bool could_have_src = has_src; 4925 // Do we already have or could we have type information for dest 4926 bool could_have_dest = has_dest; 4927 4928 ciKlass* src_k = NULL; 4929 if (!has_src) { 4930 src_k = src_type->speculative_type_not_null(); 4931 if (src_k != NULL && src_k->is_array_klass()) { 4932 could_have_src = true; 4933 } 4934 } 4935 4936 ciKlass* dest_k = NULL; 4937 if (!has_dest) { 4938 dest_k = dest_type->speculative_type_not_null(); 4939 if (dest_k != NULL && dest_k->is_array_klass()) { 4940 could_have_dest = true; 4941 } 4942 } 4943 4944 if (could_have_src && could_have_dest) { 4945 // This is going to pay off so emit the required guards 4946 if (!has_src) { 4947 src = maybe_cast_profiled_obj(src, src_k, true); 4948 src_type = _gvn.type(src); 4949 top_src = src_type->isa_aryptr(); 4950 has_src = (top_src != NULL && top_src->klass() != NULL); 4951 src_spec = true; 4952 } 4953 if (!has_dest) { 4954 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4955 dest_type = _gvn.type(dest); 4956 top_dest = dest_type->isa_aryptr(); 4957 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4958 dest_spec = true; 4959 } 4960 } 4961 } 4962 4963 if (has_src && has_dest && can_emit_guards) { 4964 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4965 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4966 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4967 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4968 4969 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4970 // If both arrays are object arrays then having the exact types 4971 // for both will remove the need for a subtype check at runtime 4972 // before the call and may make it possible to pick a faster copy 4973 // routine (without a subtype check on every element) 4974 // Do we have the exact type of src? 4975 bool could_have_src = src_spec; 4976 // Do we have the exact type of dest? 4977 bool could_have_dest = dest_spec; 4978 ciKlass* src_k = top_src->klass(); 4979 ciKlass* dest_k = top_dest->klass(); 4980 if (!src_spec) { 4981 src_k = src_type->speculative_type_not_null(); 4982 if (src_k != NULL && src_k->is_array_klass()) { 4983 could_have_src = true; 4984 } 4985 } 4986 if (!dest_spec) { 4987 dest_k = dest_type->speculative_type_not_null(); 4988 if (dest_k != NULL && dest_k->is_array_klass()) { 4989 could_have_dest = true; 4990 } 4991 } 4992 if (could_have_src && could_have_dest) { 4993 // If we can have both exact types, emit the missing guards 4994 if (could_have_src && !src_spec) { 4995 src = maybe_cast_profiled_obj(src, src_k, true); 4996 } 4997 if (could_have_dest && !dest_spec) { 4998 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4999 } 5000 } 5001 } 5002 } 5003 5004 ciMethod* trap_method = method(); 5005 int trap_bci = bci(); 5006 if (saved_jvms != NULL) { 5007 trap_method = alloc->jvms()->method(); 5008 trap_bci = alloc->jvms()->bci(); 5009 } 5010 5011 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5012 can_emit_guards && 5013 !src->is_top() && !dest->is_top()) { 5014 // validate arguments: enables transformation the ArrayCopyNode 5015 validated = true; 5016 5017 RegionNode* slow_region = new RegionNode(1); 5018 record_for_igvn(slow_region); 5019 5020 // (1) src and dest are arrays. 5021 generate_non_array_guard(load_object_klass(src), slow_region); 5022 generate_non_array_guard(load_object_klass(dest), slow_region); 5023 5024 // (2) src and dest arrays must have elements of the same BasicType 5025 // done at macro expansion or at Ideal transformation time 5026 5027 // (4) src_offset must not be negative. 5028 generate_negative_guard(src_offset, slow_region); 5029 5030 // (5) dest_offset must not be negative. 5031 generate_negative_guard(dest_offset, slow_region); 5032 5033 // (7) src_offset + length must not exceed length of src. 5034 generate_limit_guard(src_offset, length, 5035 load_array_length(src), 5036 slow_region); 5037 5038 // (8) dest_offset + length must not exceed length of dest. 5039 generate_limit_guard(dest_offset, length, 5040 load_array_length(dest), 5041 slow_region); 5042 5043 // (9) each element of an oop array must be assignable 5044 Node* src_klass = load_object_klass(src); 5045 Node* dest_klass = load_object_klass(dest); 5046 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5047 5048 if (not_subtype_ctrl != top()) { 5049 PreserveJVMState pjvms(this); 5050 set_control(not_subtype_ctrl); 5051 uncommon_trap(Deoptimization::Reason_intrinsic, 5052 Deoptimization::Action_make_not_entrant); 5053 assert(stopped(), "Should be stopped"); 5054 } 5055 { 5056 PreserveJVMState pjvms(this); 5057 set_control(_gvn.transform(slow_region)); 5058 uncommon_trap(Deoptimization::Reason_intrinsic, 5059 Deoptimization::Action_make_not_entrant); 5060 assert(stopped(), "Should be stopped"); 5061 } 5062 } 5063 5064 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5065 5066 if (stopped()) { 5067 return true; 5068 } 5069 5070 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5071 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5072 // so the compiler has a chance to eliminate them: during macro expansion, 5073 // we have to set their control (CastPP nodes are eliminated). 5074 load_object_klass(src), load_object_klass(dest), 5075 load_array_length(src), load_array_length(dest)); 5076 5077 ac->set_arraycopy(validated); 5078 5079 Node* n = _gvn.transform(ac); 5080 if (n == ac) { 5081 ac->connect_outputs(this); 5082 } else { 5083 assert(validated, "shouldn't transform if all arguments not validated"); 5084 set_all_memory(n); 5085 } 5086 5087 return true; 5088 } 5089 5090 5091 // Helper function which determines if an arraycopy immediately follows 5092 // an allocation, with no intervening tests or other escapes for the object. 5093 AllocateArrayNode* 5094 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5095 RegionNode* slow_region) { 5096 if (stopped()) return NULL; // no fast path 5097 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5098 5099 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5100 if (alloc == NULL) return NULL; 5101 5102 Node* rawmem = memory(Compile::AliasIdxRaw); 5103 // Is the allocation's memory state untouched? 5104 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5105 // Bail out if there have been raw-memory effects since the allocation. 5106 // (Example: There might have been a call or safepoint.) 5107 return NULL; 5108 } 5109 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5110 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5111 return NULL; 5112 } 5113 5114 // There must be no unexpected observers of this allocation. 5115 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5116 Node* obs = ptr->fast_out(i); 5117 if (obs != this->map()) { 5118 return NULL; 5119 } 5120 } 5121 5122 // This arraycopy must unconditionally follow the allocation of the ptr. 5123 Node* alloc_ctl = ptr->in(0); 5124 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5125 5126 Node* ctl = control(); 5127 while (ctl != alloc_ctl) { 5128 // There may be guards which feed into the slow_region. 5129 // Any other control flow means that we might not get a chance 5130 // to finish initializing the allocated object. 5131 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5132 IfNode* iff = ctl->in(0)->as_If(); 5133 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5134 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5135 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5136 ctl = iff->in(0); // This test feeds the known slow_region. 5137 continue; 5138 } 5139 // One more try: Various low-level checks bottom out in 5140 // uncommon traps. If the debug-info of the trap omits 5141 // any reference to the allocation, as we've already 5142 // observed, then there can be no objection to the trap. 5143 bool found_trap = false; 5144 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5145 Node* obs = not_ctl->fast_out(j); 5146 if (obs->in(0) == not_ctl && obs->is_Call() && 5147 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5148 found_trap = true; break; 5149 } 5150 } 5151 if (found_trap) { 5152 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5153 continue; 5154 } 5155 } 5156 return NULL; 5157 } 5158 5159 // If we get this far, we have an allocation which immediately 5160 // precedes the arraycopy, and we can take over zeroing the new object. 5161 // The arraycopy will finish the initialization, and provide 5162 // a new control state to which we will anchor the destination pointer. 5163 5164 return alloc; 5165 } 5166 5167 //-------------inline_encodeISOArray----------------------------------- 5168 // encode char[] to byte[] in ISO_8859_1 5169 bool LibraryCallKit::inline_encodeISOArray() { 5170 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5171 // no receiver since it is static method 5172 Node *src = argument(0); 5173 Node *src_offset = argument(1); 5174 Node *dst = argument(2); 5175 Node *dst_offset = argument(3); 5176 Node *length = argument(4); 5177 5178 const Type* src_type = src->Value(&_gvn); 5179 const Type* dst_type = dst->Value(&_gvn); 5180 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5181 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5182 if (top_src == NULL || top_src->klass() == NULL || 5183 top_dest == NULL || top_dest->klass() == NULL) { 5184 // failed array check 5185 return false; 5186 } 5187 5188 // Figure out the size and type of the elements we will be copying. 5189 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5190 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5191 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5192 return false; 5193 } 5194 5195 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5196 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5197 // 'src_start' points to src array + scaled offset 5198 // 'dst_start' points to dst array + scaled offset 5199 5200 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5201 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5202 enc = _gvn.transform(enc); 5203 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5204 set_memory(res_mem, mtype); 5205 set_result(enc); 5206 return true; 5207 } 5208 5209 //-------------inline_multiplyToLen----------------------------------- 5210 bool LibraryCallKit::inline_multiplyToLen() { 5211 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5212 5213 address stubAddr = StubRoutines::multiplyToLen(); 5214 if (stubAddr == NULL) { 5215 return false; // Intrinsic's stub is not implemented on this platform 5216 } 5217 const char* stubName = "multiplyToLen"; 5218 5219 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5220 5221 // no receiver because it is a static method 5222 Node* x = argument(0); 5223 Node* xlen = argument(1); 5224 Node* y = argument(2); 5225 Node* ylen = argument(3); 5226 Node* z = argument(4); 5227 5228 const Type* x_type = x->Value(&_gvn); 5229 const Type* y_type = y->Value(&_gvn); 5230 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5231 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5232 if (top_x == NULL || top_x->klass() == NULL || 5233 top_y == NULL || top_y->klass() == NULL) { 5234 // failed array check 5235 return false; 5236 } 5237 5238 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5239 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5240 if (x_elem != T_INT || y_elem != T_INT) { 5241 return false; 5242 } 5243 5244 // Set the original stack and the reexecute bit for the interpreter to reexecute 5245 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5246 // on the return from z array allocation in runtime. 5247 { PreserveReexecuteState preexecs(this); 5248 jvms()->set_should_reexecute(true); 5249 5250 Node* x_start = array_element_address(x, intcon(0), x_elem); 5251 Node* y_start = array_element_address(y, intcon(0), y_elem); 5252 // 'x_start' points to x array + scaled xlen 5253 // 'y_start' points to y array + scaled ylen 5254 5255 // Allocate the result array 5256 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5257 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5258 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5259 5260 IdealKit ideal(this); 5261 5262 #define __ ideal. 5263 Node* one = __ ConI(1); 5264 Node* zero = __ ConI(0); 5265 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5266 __ set(need_alloc, zero); 5267 __ set(z_alloc, z); 5268 __ if_then(z, BoolTest::eq, null()); { 5269 __ increment (need_alloc, one); 5270 } __ else_(); { 5271 // Update graphKit memory and control from IdealKit. 5272 sync_kit(ideal); 5273 Node* zlen_arg = load_array_length(z); 5274 // Update IdealKit memory and control from graphKit. 5275 __ sync_kit(this); 5276 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5277 __ increment (need_alloc, one); 5278 } __ end_if(); 5279 } __ end_if(); 5280 5281 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5282 // Update graphKit memory and control from IdealKit. 5283 sync_kit(ideal); 5284 Node * narr = new_array(klass_node, zlen, 1); 5285 // Update IdealKit memory and control from graphKit. 5286 __ sync_kit(this); 5287 __ set(z_alloc, narr); 5288 } __ end_if(); 5289 5290 sync_kit(ideal); 5291 z = __ value(z_alloc); 5292 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5293 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5294 // Final sync IdealKit and GraphKit. 5295 final_sync(ideal); 5296 #undef __ 5297 5298 Node* z_start = array_element_address(z, intcon(0), T_INT); 5299 5300 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5301 OptoRuntime::multiplyToLen_Type(), 5302 stubAddr, stubName, TypePtr::BOTTOM, 5303 x_start, xlen, y_start, ylen, z_start, zlen); 5304 } // original reexecute is set back here 5305 5306 C->set_has_split_ifs(true); // Has chance for split-if optimization 5307 set_result(z); 5308 return true; 5309 } 5310 5311 //-------------inline_squareToLen------------------------------------ 5312 bool LibraryCallKit::inline_squareToLen() { 5313 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5314 5315 address stubAddr = StubRoutines::squareToLen(); 5316 if (stubAddr == NULL) { 5317 return false; // Intrinsic's stub is not implemented on this platform 5318 } 5319 const char* stubName = "squareToLen"; 5320 5321 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5322 5323 Node* x = argument(0); 5324 Node* len = argument(1); 5325 Node* z = argument(2); 5326 Node* zlen = argument(3); 5327 5328 const Type* x_type = x->Value(&_gvn); 5329 const Type* z_type = z->Value(&_gvn); 5330 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5331 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5332 if (top_x == NULL || top_x->klass() == NULL || 5333 top_z == NULL || top_z->klass() == NULL) { 5334 // failed array check 5335 return false; 5336 } 5337 5338 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5339 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5340 if (x_elem != T_INT || z_elem != T_INT) { 5341 return false; 5342 } 5343 5344 5345 Node* x_start = array_element_address(x, intcon(0), x_elem); 5346 Node* z_start = array_element_address(z, intcon(0), z_elem); 5347 5348 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5349 OptoRuntime::squareToLen_Type(), 5350 stubAddr, stubName, TypePtr::BOTTOM, 5351 x_start, len, z_start, zlen); 5352 5353 set_result(z); 5354 return true; 5355 } 5356 5357 //-------------inline_mulAdd------------------------------------------ 5358 bool LibraryCallKit::inline_mulAdd() { 5359 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5360 5361 address stubAddr = StubRoutines::mulAdd(); 5362 if (stubAddr == NULL) { 5363 return false; // Intrinsic's stub is not implemented on this platform 5364 } 5365 const char* stubName = "mulAdd"; 5366 5367 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5368 5369 Node* out = argument(0); 5370 Node* in = argument(1); 5371 Node* offset = argument(2); 5372 Node* len = argument(3); 5373 Node* k = argument(4); 5374 5375 const Type* out_type = out->Value(&_gvn); 5376 const Type* in_type = in->Value(&_gvn); 5377 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5378 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5379 if (top_out == NULL || top_out->klass() == NULL || 5380 top_in == NULL || top_in->klass() == NULL) { 5381 // failed array check 5382 return false; 5383 } 5384 5385 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5386 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5387 if (out_elem != T_INT || in_elem != T_INT) { 5388 return false; 5389 } 5390 5391 Node* outlen = load_array_length(out); 5392 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5393 Node* out_start = array_element_address(out, intcon(0), out_elem); 5394 Node* in_start = array_element_address(in, intcon(0), in_elem); 5395 5396 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5397 OptoRuntime::mulAdd_Type(), 5398 stubAddr, stubName, TypePtr::BOTTOM, 5399 out_start,in_start, new_offset, len, k); 5400 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5401 set_result(result); 5402 return true; 5403 } 5404 5405 //-------------inline_montgomeryMultiply----------------------------------- 5406 bool LibraryCallKit::inline_montgomeryMultiply() { 5407 address stubAddr = StubRoutines::montgomeryMultiply(); 5408 if (stubAddr == NULL) { 5409 return false; // Intrinsic's stub is not implemented on this platform 5410 } 5411 5412 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5413 const char* stubName = "montgomery_square"; 5414 5415 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5416 5417 Node* a = argument(0); 5418 Node* b = argument(1); 5419 Node* n = argument(2); 5420 Node* len = argument(3); 5421 Node* inv = argument(4); 5422 Node* m = argument(6); 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 const Type* a_type = a->Value(&_gvn); 5484 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5485 const Type* n_type = a->Value(&_gvn); 5486 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5487 const Type* m_type = a->Value(&_gvn); 5488 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5489 if (top_a == NULL || top_a->klass() == NULL || 5490 top_n == NULL || top_n->klass() == NULL || 5491 top_m == NULL || top_m->klass() == NULL) { 5492 // failed array check 5493 return false; 5494 } 5495 5496 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5497 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5498 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5499 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5500 return false; 5501 } 5502 5503 // Make the call 5504 { 5505 Node* a_start = array_element_address(a, intcon(0), a_elem); 5506 Node* n_start = array_element_address(n, intcon(0), n_elem); 5507 Node* m_start = array_element_address(m, intcon(0), m_elem); 5508 5509 Node* call = make_runtime_call(RC_LEAF, 5510 OptoRuntime::montgomerySquare_Type(), 5511 stubAddr, stubName, TypePtr::BOTTOM, 5512 a_start, n_start, len, inv, top(), 5513 m_start); 5514 set_result(m); 5515 } 5516 5517 return true; 5518 } 5519 5520 5521 /** 5522 * Calculate CRC32 for byte. 5523 * int java.util.zip.CRC32.update(int crc, int b) 5524 */ 5525 bool LibraryCallKit::inline_updateCRC32() { 5526 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5527 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5528 // no receiver since it is static method 5529 Node* crc = argument(0); // type: int 5530 Node* b = argument(1); // type: int 5531 5532 /* 5533 * int c = ~ crc; 5534 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5535 * b = b ^ (c >>> 8); 5536 * crc = ~b; 5537 */ 5538 5539 Node* M1 = intcon(-1); 5540 crc = _gvn.transform(new XorINode(crc, M1)); 5541 Node* result = _gvn.transform(new XorINode(crc, b)); 5542 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5543 5544 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5545 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5546 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5547 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5548 5549 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5550 result = _gvn.transform(new XorINode(crc, result)); 5551 result = _gvn.transform(new XorINode(result, M1)); 5552 set_result(result); 5553 return true; 5554 } 5555 5556 /** 5557 * Calculate CRC32 for byte[] array. 5558 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5559 */ 5560 bool LibraryCallKit::inline_updateBytesCRC32() { 5561 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5562 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5563 // no receiver since it is static method 5564 Node* crc = argument(0); // type: int 5565 Node* src = argument(1); // type: oop 5566 Node* offset = argument(2); // type: int 5567 Node* length = argument(3); // type: int 5568 5569 const Type* src_type = src->Value(&_gvn); 5570 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5571 if (top_src == NULL || top_src->klass() == NULL) { 5572 // failed array check 5573 return false; 5574 } 5575 5576 // Figure out the size and type of the elements we will be copying. 5577 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5578 if (src_elem != T_BYTE) { 5579 return false; 5580 } 5581 5582 // 'src_start' points to src array + scaled offset 5583 Node* src_start = array_element_address(src, offset, src_elem); 5584 5585 // We assume that range check is done by caller. 5586 // TODO: generate range check (offset+length < src.length) in debug VM. 5587 5588 // Call the stub. 5589 address stubAddr = StubRoutines::updateBytesCRC32(); 5590 const char *stubName = "updateBytesCRC32"; 5591 5592 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5593 stubAddr, stubName, TypePtr::BOTTOM, 5594 crc, src_start, length); 5595 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5596 set_result(result); 5597 return true; 5598 } 5599 5600 /** 5601 * Calculate CRC32 for ByteBuffer. 5602 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5603 */ 5604 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5605 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5606 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5607 // no receiver since it is static method 5608 Node* crc = argument(0); // type: int 5609 Node* src = argument(1); // type: long 5610 Node* offset = argument(3); // type: int 5611 Node* length = argument(4); // type: int 5612 5613 src = ConvL2X(src); // adjust Java long to machine word 5614 Node* base = _gvn.transform(new CastX2PNode(src)); 5615 offset = ConvI2X(offset); 5616 5617 // 'src_start' points to src array + scaled offset 5618 Node* src_start = basic_plus_adr(top(), base, offset); 5619 5620 // Call the stub. 5621 address stubAddr = StubRoutines::updateBytesCRC32(); 5622 const char *stubName = "updateBytesCRC32"; 5623 5624 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5625 stubAddr, stubName, TypePtr::BOTTOM, 5626 crc, src_start, length); 5627 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5628 set_result(result); 5629 return true; 5630 } 5631 5632 //------------------------------get_table_from_crc32c_class----------------------- 5633 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5634 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5635 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5636 5637 return table; 5638 } 5639 5640 //------------------------------inline_updateBytesCRC32C----------------------- 5641 // 5642 // Calculate CRC32C for byte[] array. 5643 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5644 // 5645 bool LibraryCallKit::inline_updateBytesCRC32C() { 5646 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5647 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5648 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5649 // no receiver since it is a static method 5650 Node* crc = argument(0); // type: int 5651 Node* src = argument(1); // type: oop 5652 Node* offset = argument(2); // type: int 5653 Node* end = argument(3); // type: int 5654 5655 Node* length = _gvn.transform(new SubINode(end, offset)); 5656 5657 const Type* src_type = src->Value(&_gvn); 5658 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5659 if (top_src == NULL || top_src->klass() == NULL) { 5660 // failed array check 5661 return false; 5662 } 5663 5664 // Figure out the size and type of the elements we will be copying. 5665 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5666 if (src_elem != T_BYTE) { 5667 return false; 5668 } 5669 5670 // 'src_start' points to src array + scaled offset 5671 Node* src_start = array_element_address(src, offset, src_elem); 5672 5673 // static final int[] byteTable in class CRC32C 5674 Node* table = get_table_from_crc32c_class(callee()->holder()); 5675 Node* table_start = array_element_address(table, intcon(0), T_INT); 5676 5677 // We assume that range check is done by caller. 5678 // TODO: generate range check (offset+length < src.length) in debug VM. 5679 5680 // Call the stub. 5681 address stubAddr = StubRoutines::updateBytesCRC32C(); 5682 const char *stubName = "updateBytesCRC32C"; 5683 5684 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5685 stubAddr, stubName, TypePtr::BOTTOM, 5686 crc, src_start, length, table_start); 5687 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5688 set_result(result); 5689 return true; 5690 } 5691 5692 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5693 // 5694 // Calculate CRC32C for DirectByteBuffer. 5695 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5696 // 5697 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5698 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5699 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5700 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5701 // no receiver since it is a static method 5702 Node* crc = argument(0); // type: int 5703 Node* src = argument(1); // type: long 5704 Node* offset = argument(3); // type: int 5705 Node* end = argument(4); // type: int 5706 5707 Node* length = _gvn.transform(new SubINode(end, offset)); 5708 5709 src = ConvL2X(src); // adjust Java long to machine word 5710 Node* base = _gvn.transform(new CastX2PNode(src)); 5711 offset = ConvI2X(offset); 5712 5713 // 'src_start' points to src array + scaled offset 5714 Node* src_start = basic_plus_adr(top(), base, offset); 5715 5716 // static final int[] byteTable in class CRC32C 5717 Node* table = get_table_from_crc32c_class(callee()->holder()); 5718 Node* table_start = array_element_address(table, intcon(0), T_INT); 5719 5720 // Call the stub. 5721 address stubAddr = StubRoutines::updateBytesCRC32C(); 5722 const char *stubName = "updateBytesCRC32C"; 5723 5724 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5725 stubAddr, stubName, TypePtr::BOTTOM, 5726 crc, src_start, length, table_start); 5727 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5728 set_result(result); 5729 return true; 5730 } 5731 5732 //------------------------------inline_updateBytesAdler32---------------------- 5733 // 5734 // Calculate Adler32 checksum for byte[] array. 5735 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5736 // 5737 bool LibraryCallKit::inline_updateBytesAdler32() { 5738 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5739 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5740 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5741 // no receiver since it is static method 5742 Node* crc = argument(0); // type: int 5743 Node* src = argument(1); // type: oop 5744 Node* offset = argument(2); // type: int 5745 Node* length = argument(3); // type: int 5746 5747 const Type* src_type = src->Value(&_gvn); 5748 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5749 if (top_src == NULL || top_src->klass() == NULL) { 5750 // failed array check 5751 return false; 5752 } 5753 5754 // Figure out the size and type of the elements we will be copying. 5755 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5756 if (src_elem != T_BYTE) { 5757 return false; 5758 } 5759 5760 // 'src_start' points to src array + scaled offset 5761 Node* src_start = array_element_address(src, offset, src_elem); 5762 5763 // We assume that range check is done by caller. 5764 // TODO: generate range check (offset+length < src.length) in debug VM. 5765 5766 // Call the stub. 5767 address stubAddr = StubRoutines::updateBytesAdler32(); 5768 const char *stubName = "updateBytesAdler32"; 5769 5770 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5771 stubAddr, stubName, TypePtr::BOTTOM, 5772 crc, src_start, length); 5773 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5774 set_result(result); 5775 return true; 5776 } 5777 5778 //------------------------------inline_updateByteBufferAdler32--------------- 5779 // 5780 // Calculate Adler32 checksum for DirectByteBuffer. 5781 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5782 // 5783 bool LibraryCallKit::inline_updateByteBufferAdler32() { 5784 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5785 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5786 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5787 // no receiver since it is static method 5788 Node* crc = argument(0); // type: int 5789 Node* src = argument(1); // type: long 5790 Node* offset = argument(3); // type: int 5791 Node* length = argument(4); // type: int 5792 5793 src = ConvL2X(src); // adjust Java long to machine word 5794 Node* base = _gvn.transform(new CastX2PNode(src)); 5795 offset = ConvI2X(offset); 5796 5797 // 'src_start' points to src array + scaled offset 5798 Node* src_start = basic_plus_adr(top(), base, offset); 5799 5800 // Call the stub. 5801 address stubAddr = StubRoutines::updateBytesAdler32(); 5802 const char *stubName = "updateBytesAdler32"; 5803 5804 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5805 stubAddr, stubName, TypePtr::BOTTOM, 5806 crc, src_start, length); 5807 5808 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5809 set_result(result); 5810 return true; 5811 } 5812 5813 //----------------------------inline_reference_get---------------------------- 5814 // public T java.lang.ref.Reference.get(); 5815 bool LibraryCallKit::inline_reference_get() { 5816 const int referent_offset = java_lang_ref_Reference::referent_offset; 5817 guarantee(referent_offset > 0, "should have already been set"); 5818 5819 // Get the argument: 5820 Node* reference_obj = null_check_receiver(); 5821 if (stopped()) return true; 5822 5823 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5824 5825 ciInstanceKlass* klass = env()->Object_klass(); 5826 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5827 5828 Node* no_ctrl = NULL; 5829 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5830 5831 // Use the pre-barrier to record the value in the referent field 5832 pre_barrier(false /* do_load */, 5833 control(), 5834 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5835 result /* pre_val */, 5836 T_OBJECT); 5837 5838 // Add memory barrier to prevent commoning reads from this field 5839 // across safepoint since GC can change its value. 5840 insert_mem_bar(Op_MemBarCPUOrder); 5841 5842 set_result(result); 5843 return true; 5844 } 5845 5846 5847 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5848 bool is_exact=true, bool is_static=false, 5849 ciInstanceKlass * fromKls=NULL) { 5850 if (fromKls == NULL) { 5851 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5852 assert(tinst != NULL, "obj is null"); 5853 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5854 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5855 fromKls = tinst->klass()->as_instance_klass(); 5856 } else { 5857 assert(is_static, "only for static field access"); 5858 } 5859 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5860 ciSymbol::make(fieldTypeString), 5861 is_static); 5862 5863 assert (field != NULL, "undefined field"); 5864 if (field == NULL) return (Node *) NULL; 5865 5866 if (is_static) { 5867 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5868 fromObj = makecon(tip); 5869 } 5870 5871 // Next code copied from Parse::do_get_xxx(): 5872 5873 // Compute address and memory type. 5874 int offset = field->offset_in_bytes(); 5875 bool is_vol = field->is_volatile(); 5876 ciType* field_klass = field->type(); 5877 assert(field_klass->is_loaded(), "should be loaded"); 5878 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5879 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5880 BasicType bt = field->layout_type(); 5881 5882 // Build the resultant type of the load 5883 const Type *type; 5884 if (bt == T_OBJECT) { 5885 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5886 } else { 5887 type = Type::get_const_basic_type(bt); 5888 } 5889 5890 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5891 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5892 } 5893 // Build the load. 5894 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5895 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5896 // If reference is volatile, prevent following memory ops from 5897 // floating up past the volatile read. Also prevents commoning 5898 // another volatile read. 5899 if (is_vol) { 5900 // Memory barrier includes bogus read of value to force load BEFORE membar 5901 insert_mem_bar(Op_MemBarAcquire, loadedField); 5902 } 5903 return loadedField; 5904 } 5905 5906 5907 //------------------------------inline_aescrypt_Block----------------------- 5908 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5909 address stubAddr = NULL; 5910 const char *stubName; 5911 assert(UseAES, "need AES instruction support"); 5912 5913 switch(id) { 5914 case vmIntrinsics::_aescrypt_encryptBlock: 5915 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5916 stubName = "aescrypt_encryptBlock"; 5917 break; 5918 case vmIntrinsics::_aescrypt_decryptBlock: 5919 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5920 stubName = "aescrypt_decryptBlock"; 5921 break; 5922 } 5923 if (stubAddr == NULL) return false; 5924 5925 Node* aescrypt_object = argument(0); 5926 Node* src = argument(1); 5927 Node* src_offset = argument(2); 5928 Node* dest = argument(3); 5929 Node* dest_offset = argument(4); 5930 5931 // (1) src and dest are arrays. 5932 const Type* src_type = src->Value(&_gvn); 5933 const Type* dest_type = dest->Value(&_gvn); 5934 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5935 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5936 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5937 5938 // for the quick and dirty code we will skip all the checks. 5939 // we are just trying to get the call to be generated. 5940 Node* src_start = src; 5941 Node* dest_start = dest; 5942 if (src_offset != NULL || dest_offset != NULL) { 5943 assert(src_offset != NULL && dest_offset != NULL, ""); 5944 src_start = array_element_address(src, src_offset, T_BYTE); 5945 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5946 } 5947 5948 // now need to get the start of its expanded key array 5949 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5950 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5951 if (k_start == NULL) return false; 5952 5953 if (Matcher::pass_original_key_for_aes()) { 5954 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5955 // compatibility issues between Java key expansion and SPARC crypto instructions 5956 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5957 if (original_k_start == NULL) return false; 5958 5959 // Call the stub. 5960 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5961 stubAddr, stubName, TypePtr::BOTTOM, 5962 src_start, dest_start, k_start, original_k_start); 5963 } else { 5964 // Call the stub. 5965 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5966 stubAddr, stubName, TypePtr::BOTTOM, 5967 src_start, dest_start, k_start); 5968 } 5969 5970 return true; 5971 } 5972 5973 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5974 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5975 address stubAddr = NULL; 5976 const char *stubName = NULL; 5977 5978 assert(UseAES, "need AES instruction support"); 5979 5980 switch(id) { 5981 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5982 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5983 stubName = "cipherBlockChaining_encryptAESCrypt"; 5984 break; 5985 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5986 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5987 stubName = "cipherBlockChaining_decryptAESCrypt"; 5988 break; 5989 } 5990 if (stubAddr == NULL) return false; 5991 5992 Node* cipherBlockChaining_object = argument(0); 5993 Node* src = argument(1); 5994 Node* src_offset = argument(2); 5995 Node* len = argument(3); 5996 Node* dest = argument(4); 5997 Node* dest_offset = argument(5); 5998 5999 // (1) src and dest are arrays. 6000 const Type* src_type = src->Value(&_gvn); 6001 const Type* dest_type = dest->Value(&_gvn); 6002 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6003 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6004 assert (top_src != NULL && top_src->klass() != NULL 6005 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6006 6007 // checks are the responsibility of the caller 6008 Node* src_start = src; 6009 Node* dest_start = dest; 6010 if (src_offset != NULL || dest_offset != NULL) { 6011 assert(src_offset != NULL && dest_offset != NULL, ""); 6012 src_start = array_element_address(src, src_offset, T_BYTE); 6013 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6014 } 6015 6016 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6017 // (because of the predicated logic executed earlier). 6018 // so we cast it here safely. 6019 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6020 6021 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6022 if (embeddedCipherObj == NULL) return false; 6023 6024 // cast it to what we know it will be at runtime 6025 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6026 assert(tinst != NULL, "CBC obj is null"); 6027 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6028 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6029 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6030 6031 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6032 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6033 const TypeOopPtr* xtype = aklass->as_instance_type(); 6034 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6035 aescrypt_object = _gvn.transform(aescrypt_object); 6036 6037 // we need to get the start of the aescrypt_object's expanded key array 6038 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6039 if (k_start == NULL) return false; 6040 6041 // similarly, get the start address of the r vector 6042 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6043 if (objRvec == NULL) return false; 6044 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6045 6046 Node* cbcCrypt; 6047 if (Matcher::pass_original_key_for_aes()) { 6048 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6049 // compatibility issues between Java key expansion and SPARC crypto instructions 6050 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6051 if (original_k_start == NULL) return false; 6052 6053 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6054 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6055 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6056 stubAddr, stubName, TypePtr::BOTTOM, 6057 src_start, dest_start, k_start, r_start, len, original_k_start); 6058 } else { 6059 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6060 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6061 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6062 stubAddr, stubName, TypePtr::BOTTOM, 6063 src_start, dest_start, k_start, r_start, len); 6064 } 6065 6066 // return cipher length (int) 6067 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6068 set_result(retvalue); 6069 return true; 6070 } 6071 6072 //------------------------------inline_counterMode_AESCrypt----------------------- 6073 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6074 assert(UseAES, "need AES instruction support"); 6075 if (!UseCTRAESIntrinsics) return false; 6076 6077 address stubAddr = NULL; 6078 const char *stubName = NULL; 6079 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6080 stubAddr = StubRoutines::counterMode_AESCrypt(); 6081 stubName = "counterMode_AESCrypt"; 6082 } 6083 if (stubAddr == NULL) return false; 6084 6085 Node* counterMode_object = argument(0); 6086 Node* src = argument(1); 6087 Node* src_offset = argument(2); 6088 Node* len = argument(3); 6089 Node* dest = argument(4); 6090 Node* dest_offset = argument(5); 6091 6092 // (1) src and dest are arrays. 6093 const Type* src_type = src->Value(&_gvn); 6094 const Type* dest_type = dest->Value(&_gvn); 6095 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6096 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6097 assert(top_src != NULL && top_src->klass() != NULL 6098 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6099 6100 // checks are the responsibility of the caller 6101 Node* src_start = src; 6102 Node* dest_start = dest; 6103 if (src_offset != NULL || dest_offset != NULL) { 6104 assert(src_offset != NULL && dest_offset != NULL, ""); 6105 src_start = array_element_address(src, src_offset, T_BYTE); 6106 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6107 } 6108 6109 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6110 // (because of the predicated logic executed earlier). 6111 // so we cast it here safely. 6112 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6113 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6114 if (embeddedCipherObj == NULL) return false; 6115 // cast it to what we know it will be at runtime 6116 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6117 assert(tinst != NULL, "CTR obj is null"); 6118 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6119 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6120 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6121 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6122 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6123 const TypeOopPtr* xtype = aklass->as_instance_type(); 6124 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6125 aescrypt_object = _gvn.transform(aescrypt_object); 6126 // we need to get the start of the aescrypt_object's expanded key array 6127 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6128 if (k_start == NULL) return false; 6129 // similarly, get the start address of the r vector 6130 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6131 if (obj_counter == NULL) return false; 6132 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6133 6134 Node* ctrCrypt; 6135 if (Matcher::pass_original_key_for_aes()) { 6136 // no SPARC version for AES/CTR intrinsics now. 6137 return false; 6138 } 6139 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6140 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6141 OptoRuntime::counterMode_aescrypt_Type(), 6142 stubAddr, stubName, TypePtr::BOTTOM, 6143 src_start, dest_start, k_start, cnt_start, len); 6144 6145 // return cipher length (int) 6146 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6147 set_result(retvalue); 6148 return true; 6149 } 6150 6151 //------------------------------get_key_start_from_aescrypt_object----------------------- 6152 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6153 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6154 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6155 if (objAESCryptKey == NULL) return (Node *) NULL; 6156 6157 // now have the array, need to get the start address of the K array 6158 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6159 return k_start; 6160 } 6161 6162 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6163 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6164 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6165 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6166 if (objAESCryptKey == NULL) return (Node *) NULL; 6167 6168 // now have the array, need to get the start address of the lastKey array 6169 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6170 return original_k_start; 6171 } 6172 6173 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6174 // Return node representing slow path of predicate check. 6175 // the pseudo code we want to emulate with this predicate is: 6176 // for encryption: 6177 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6178 // for decryption: 6179 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6180 // note cipher==plain is more conservative than the original java code but that's OK 6181 // 6182 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6183 // The receiver was checked for NULL already. 6184 Node* objCBC = argument(0); 6185 6186 // Load embeddedCipher field of CipherBlockChaining object. 6187 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6188 6189 // get AESCrypt klass for instanceOf check 6190 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6191 // will have same classloader as CipherBlockChaining object 6192 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6193 assert(tinst != NULL, "CBCobj is null"); 6194 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6195 6196 // we want to do an instanceof comparison against the AESCrypt class 6197 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6198 if (!klass_AESCrypt->is_loaded()) { 6199 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6200 Node* ctrl = control(); 6201 set_control(top()); // no regular fast path 6202 return ctrl; 6203 } 6204 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6205 6206 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6207 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6208 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6209 6210 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6211 6212 // for encryption, we are done 6213 if (!decrypting) 6214 return instof_false; // even if it is NULL 6215 6216 // for decryption, we need to add a further check to avoid 6217 // taking the intrinsic path when cipher and plain are the same 6218 // see the original java code for why. 6219 RegionNode* region = new RegionNode(3); 6220 region->init_req(1, instof_false); 6221 Node* src = argument(1); 6222 Node* dest = argument(4); 6223 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6224 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6225 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6226 region->init_req(2, src_dest_conjoint); 6227 6228 record_for_igvn(region); 6229 return _gvn.transform(region); 6230 } 6231 6232 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6233 // Return node representing slow path of predicate check. 6234 // the pseudo code we want to emulate with this predicate is: 6235 // for encryption: 6236 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6237 // for decryption: 6238 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6239 // note cipher==plain is more conservative than the original java code but that's OK 6240 // 6241 6242 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6243 // The receiver was checked for NULL already. 6244 Node* objCTR = argument(0); 6245 6246 // Load embeddedCipher field of CipherBlockChaining object. 6247 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6248 6249 // get AESCrypt klass for instanceOf check 6250 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6251 // will have same classloader as CipherBlockChaining object 6252 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6253 assert(tinst != NULL, "CTRobj is null"); 6254 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6255 6256 // we want to do an instanceof comparison against the AESCrypt class 6257 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6258 if (!klass_AESCrypt->is_loaded()) { 6259 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6260 Node* ctrl = control(); 6261 set_control(top()); // no regular fast path 6262 return ctrl; 6263 } 6264 6265 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6266 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6267 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6268 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6269 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6270 6271 return instof_false; // even if it is NULL 6272 } 6273 6274 //------------------------------inline_ghash_processBlocks 6275 bool LibraryCallKit::inline_ghash_processBlocks() { 6276 address stubAddr; 6277 const char *stubName; 6278 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6279 6280 stubAddr = StubRoutines::ghash_processBlocks(); 6281 stubName = "ghash_processBlocks"; 6282 6283 Node* data = argument(0); 6284 Node* offset = argument(1); 6285 Node* len = argument(2); 6286 Node* state = argument(3); 6287 Node* subkeyH = argument(4); 6288 6289 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6290 assert(state_start, "state is NULL"); 6291 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6292 assert(subkeyH_start, "subkeyH is NULL"); 6293 Node* data_start = array_element_address(data, offset, T_BYTE); 6294 assert(data_start, "data is NULL"); 6295 6296 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6297 OptoRuntime::ghash_processBlocks_Type(), 6298 stubAddr, stubName, TypePtr::BOTTOM, 6299 state_start, subkeyH_start, data_start, len); 6300 return true; 6301 } 6302 6303 //------------------------------inline_sha_implCompress----------------------- 6304 // 6305 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6306 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6307 // 6308 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6309 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6310 // 6311 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6312 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6313 // 6314 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6315 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6316 6317 Node* sha_obj = argument(0); 6318 Node* src = argument(1); // type oop 6319 Node* ofs = argument(2); // type int 6320 6321 const Type* src_type = src->Value(&_gvn); 6322 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6323 if (top_src == NULL || top_src->klass() == NULL) { 6324 // failed array check 6325 return false; 6326 } 6327 // Figure out the size and type of the elements we will be copying. 6328 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6329 if (src_elem != T_BYTE) { 6330 return false; 6331 } 6332 // 'src_start' points to src array + offset 6333 Node* src_start = array_element_address(src, ofs, src_elem); 6334 Node* state = NULL; 6335 address stubAddr; 6336 const char *stubName; 6337 6338 switch(id) { 6339 case vmIntrinsics::_sha_implCompress: 6340 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6341 state = get_state_from_sha_object(sha_obj); 6342 stubAddr = StubRoutines::sha1_implCompress(); 6343 stubName = "sha1_implCompress"; 6344 break; 6345 case vmIntrinsics::_sha2_implCompress: 6346 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6347 state = get_state_from_sha_object(sha_obj); 6348 stubAddr = StubRoutines::sha256_implCompress(); 6349 stubName = "sha256_implCompress"; 6350 break; 6351 case vmIntrinsics::_sha5_implCompress: 6352 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6353 state = get_state_from_sha5_object(sha_obj); 6354 stubAddr = StubRoutines::sha512_implCompress(); 6355 stubName = "sha512_implCompress"; 6356 break; 6357 default: 6358 fatal_unexpected_iid(id); 6359 return false; 6360 } 6361 if (state == NULL) return false; 6362 6363 // Call the stub. 6364 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6365 stubAddr, stubName, TypePtr::BOTTOM, 6366 src_start, state); 6367 6368 return true; 6369 } 6370 6371 //------------------------------inline_digestBase_implCompressMB----------------------- 6372 // 6373 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6374 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6375 // 6376 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6377 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6378 "need SHA1/SHA256/SHA512 instruction support"); 6379 assert((uint)predicate < 3, "sanity"); 6380 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6381 6382 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6383 Node* src = argument(1); // byte[] array 6384 Node* ofs = argument(2); // type int 6385 Node* limit = argument(3); // type int 6386 6387 const Type* src_type = src->Value(&_gvn); 6388 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6389 if (top_src == NULL || top_src->klass() == NULL) { 6390 // failed array check 6391 return false; 6392 } 6393 // Figure out the size and type of the elements we will be copying. 6394 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6395 if (src_elem != T_BYTE) { 6396 return false; 6397 } 6398 // 'src_start' points to src array + offset 6399 Node* src_start = array_element_address(src, ofs, src_elem); 6400 6401 const char* klass_SHA_name = NULL; 6402 const char* stub_name = NULL; 6403 address stub_addr = NULL; 6404 bool long_state = false; 6405 6406 switch (predicate) { 6407 case 0: 6408 if (UseSHA1Intrinsics) { 6409 klass_SHA_name = "sun/security/provider/SHA"; 6410 stub_name = "sha1_implCompressMB"; 6411 stub_addr = StubRoutines::sha1_implCompressMB(); 6412 } 6413 break; 6414 case 1: 6415 if (UseSHA256Intrinsics) { 6416 klass_SHA_name = "sun/security/provider/SHA2"; 6417 stub_name = "sha256_implCompressMB"; 6418 stub_addr = StubRoutines::sha256_implCompressMB(); 6419 } 6420 break; 6421 case 2: 6422 if (UseSHA512Intrinsics) { 6423 klass_SHA_name = "sun/security/provider/SHA5"; 6424 stub_name = "sha512_implCompressMB"; 6425 stub_addr = StubRoutines::sha512_implCompressMB(); 6426 long_state = true; 6427 } 6428 break; 6429 default: 6430 fatal("unknown SHA intrinsic predicate: %d", predicate); 6431 } 6432 if (klass_SHA_name != NULL) { 6433 // get DigestBase klass to lookup for SHA klass 6434 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6435 assert(tinst != NULL, "digestBase_obj is not instance???"); 6436 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6437 6438 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6439 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6440 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6441 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6442 } 6443 return false; 6444 } 6445 //------------------------------inline_sha_implCompressMB----------------------- 6446 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6447 bool long_state, address stubAddr, const char *stubName, 6448 Node* src_start, Node* ofs, Node* limit) { 6449 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6450 const TypeOopPtr* xtype = aklass->as_instance_type(); 6451 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6452 sha_obj = _gvn.transform(sha_obj); 6453 6454 Node* state; 6455 if (long_state) { 6456 state = get_state_from_sha5_object(sha_obj); 6457 } else { 6458 state = get_state_from_sha_object(sha_obj); 6459 } 6460 if (state == NULL) return false; 6461 6462 // Call the stub. 6463 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6464 OptoRuntime::digestBase_implCompressMB_Type(), 6465 stubAddr, stubName, TypePtr::BOTTOM, 6466 src_start, state, ofs, limit); 6467 // return ofs (int) 6468 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6469 set_result(result); 6470 6471 return true; 6472 } 6473 6474 //------------------------------get_state_from_sha_object----------------------- 6475 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6476 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6477 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6478 if (sha_state == NULL) return (Node *) NULL; 6479 6480 // now have the array, need to get the start address of the state array 6481 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6482 return state; 6483 } 6484 6485 //------------------------------get_state_from_sha5_object----------------------- 6486 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6487 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6488 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6489 if (sha_state == NULL) return (Node *) NULL; 6490 6491 // now have the array, need to get the start address of the state array 6492 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6493 return state; 6494 } 6495 6496 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6497 // Return node representing slow path of predicate check. 6498 // the pseudo code we want to emulate with this predicate is: 6499 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6500 // 6501 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6502 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6503 "need SHA1/SHA256/SHA512 instruction support"); 6504 assert((uint)predicate < 3, "sanity"); 6505 6506 // The receiver was checked for NULL already. 6507 Node* digestBaseObj = argument(0); 6508 6509 // get DigestBase klass for instanceOf check 6510 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6511 assert(tinst != NULL, "digestBaseObj is null"); 6512 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6513 6514 const char* klass_SHA_name = NULL; 6515 switch (predicate) { 6516 case 0: 6517 if (UseSHA1Intrinsics) { 6518 // we want to do an instanceof comparison against the SHA class 6519 klass_SHA_name = "sun/security/provider/SHA"; 6520 } 6521 break; 6522 case 1: 6523 if (UseSHA256Intrinsics) { 6524 // we want to do an instanceof comparison against the SHA2 class 6525 klass_SHA_name = "sun/security/provider/SHA2"; 6526 } 6527 break; 6528 case 2: 6529 if (UseSHA512Intrinsics) { 6530 // we want to do an instanceof comparison against the SHA5 class 6531 klass_SHA_name = "sun/security/provider/SHA5"; 6532 } 6533 break; 6534 default: 6535 fatal("unknown SHA intrinsic predicate: %d", predicate); 6536 } 6537 6538 ciKlass* klass_SHA = NULL; 6539 if (klass_SHA_name != NULL) { 6540 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6541 } 6542 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6543 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6544 Node* ctrl = control(); 6545 set_control(top()); // no intrinsic path 6546 return ctrl; 6547 } 6548 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6549 6550 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6551 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6552 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6553 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6554 6555 return instof_false; // even if it is NULL 6556 } 6557 6558 bool LibraryCallKit::inline_profileBoolean() { 6559 Node* counts = argument(1); 6560 const TypeAryPtr* ary = NULL; 6561 ciArray* aobj = NULL; 6562 if (counts->is_Con() 6563 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6564 && (aobj = ary->const_oop()->as_array()) != NULL 6565 && (aobj->length() == 2)) { 6566 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6567 jint false_cnt = aobj->element_value(0).as_int(); 6568 jint true_cnt = aobj->element_value(1).as_int(); 6569 6570 if (C->log() != NULL) { 6571 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6572 false_cnt, true_cnt); 6573 } 6574 6575 if (false_cnt + true_cnt == 0) { 6576 // According to profile, never executed. 6577 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6578 Deoptimization::Action_reinterpret); 6579 return true; 6580 } 6581 6582 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6583 // is a number of each value occurrences. 6584 Node* result = argument(0); 6585 if (false_cnt == 0 || true_cnt == 0) { 6586 // According to profile, one value has been never seen. 6587 int expected_val = (false_cnt == 0) ? 1 : 0; 6588 6589 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6590 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6591 6592 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6593 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6594 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6595 6596 { // Slow path: uncommon trap for never seen value and then reexecute 6597 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6598 // the value has been seen at least once. 6599 PreserveJVMState pjvms(this); 6600 PreserveReexecuteState preexecs(this); 6601 jvms()->set_should_reexecute(true); 6602 6603 set_control(slow_path); 6604 set_i_o(i_o()); 6605 6606 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6607 Deoptimization::Action_reinterpret); 6608 } 6609 // The guard for never seen value enables sharpening of the result and 6610 // returning a constant. It allows to eliminate branches on the same value 6611 // later on. 6612 set_control(fast_path); 6613 result = intcon(expected_val); 6614 } 6615 // Stop profiling. 6616 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6617 // By replacing method body with profile data (represented as ProfileBooleanNode 6618 // on IR level) we effectively disable profiling. 6619 // It enables full speed execution once optimized code is generated. 6620 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6621 C->record_for_igvn(profile); 6622 set_result(profile); 6623 return true; 6624 } else { 6625 // Continue profiling. 6626 // Profile data isn't available at the moment. So, execute method's bytecode version. 6627 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6628 // is compiled and counters aren't available since corresponding MethodHandle 6629 // isn't a compile-time constant. 6630 return false; 6631 } 6632 } 6633 6634 bool LibraryCallKit::inline_isCompileConstant() { 6635 Node* n = argument(0); 6636 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6637 return true; 6638 }