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