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