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