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