1 /* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "oops/objArrayKlass.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/c2compiler.hpp" 35 #include "opto/callGenerator.hpp" 36 #include "opto/castnode.hpp" 37 #include "opto/cfgnode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/countbitsnode.hpp" 40 #include "opto/intrinsicnode.hpp" 41 #include "opto/idealKit.hpp" 42 #include "opto/mathexactnode.hpp" 43 #include "opto/movenode.hpp" 44 #include "opto/mulnode.hpp" 45 #include "opto/narrowptrnode.hpp" 46 #include "opto/opaquenode.hpp" 47 #include "opto/parse.hpp" 48 #include "opto/runtime.hpp" 49 #include "opto/subnode.hpp" 50 #include "prims/nativeLookup.hpp" 51 #include "runtime/sharedRuntime.hpp" 52 #include "trace/traceMacros.hpp" 53 54 class LibraryIntrinsic : public InlineCallGenerator { 55 // Extend the set of intrinsics known to the runtime: 56 public: 57 private: 58 bool _is_virtual; 59 bool _does_virtual_dispatch; 60 int8_t _predicates_count; // Intrinsic is predicated by several conditions 61 int8_t _last_predicate; // Last generated predicate 62 vmIntrinsics::ID _intrinsic_id; 63 64 public: 65 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 66 : InlineCallGenerator(m), 67 _is_virtual(is_virtual), 68 _does_virtual_dispatch(does_virtual_dispatch), 69 _predicates_count((int8_t)predicates_count), 70 _last_predicate((int8_t)-1), 71 _intrinsic_id(id) 72 { 73 } 74 virtual bool is_intrinsic() const { return true; } 75 virtual bool is_virtual() const { return _is_virtual; } 76 virtual bool is_predicated() const { return _predicates_count > 0; } 77 virtual int predicates_count() const { return _predicates_count; } 78 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 79 virtual JVMState* generate(JVMState* jvms); 80 virtual Node* generate_predicate(JVMState* jvms, int predicate); 81 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 82 }; 83 84 85 // Local helper class for LibraryIntrinsic: 86 class LibraryCallKit : public GraphKit { 87 private: 88 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 89 Node* _result; // the result node, if any 90 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 91 92 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 93 94 public: 95 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 96 : GraphKit(jvms), 97 _intrinsic(intrinsic), 98 _result(NULL) 99 { 100 // Check if this is a root compile. In that case we don't have a caller. 101 if (!jvms->has_method()) { 102 _reexecute_sp = sp(); 103 } else { 104 // Find out how many arguments the interpreter needs when deoptimizing 105 // and save the stack pointer value so it can used by uncommon_trap. 106 // We find the argument count by looking at the declared signature. 107 bool ignored_will_link; 108 ciSignature* declared_signature = NULL; 109 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 110 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 111 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 112 } 113 } 114 115 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 116 117 ciMethod* caller() const { return jvms()->method(); } 118 int bci() const { return jvms()->bci(); } 119 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 120 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 121 ciMethod* callee() const { return _intrinsic->method(); } 122 123 bool try_to_inline(int predicate); 124 Node* try_to_predicate(int predicate); 125 126 void push_result() { 127 // Push the result onto the stack. 128 if (!stopped() && result() != NULL) { 129 BasicType bt = result()->bottom_type()->basic_type(); 130 push_node(bt, result()); 131 } 132 } 133 134 private: 135 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 136 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 137 } 138 139 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 140 void set_result(RegionNode* region, PhiNode* value); 141 Node* result() { return _result; } 142 143 virtual int reexecute_sp() { return _reexecute_sp; } 144 145 // Helper functions to inline natives 146 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 147 Node* generate_slow_guard(Node* test, RegionNode* region); 148 Node* generate_fair_guard(Node* test, RegionNode* region); 149 Node* generate_negative_guard(Node* index, RegionNode* region, 150 // resulting CastII of index: 151 Node* *pos_index = NULL); 152 Node* generate_limit_guard(Node* offset, Node* subseq_length, 153 Node* array_length, 154 RegionNode* region); 155 Node* generate_current_thread(Node* &tls_output); 156 Node* load_mirror_from_klass(Node* klass); 157 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 158 RegionNode* region, int null_path, 159 int offset); 160 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 161 RegionNode* region, int null_path) { 162 int offset = java_lang_Class::klass_offset_in_bytes(); 163 return load_klass_from_mirror_common(mirror, never_see_null, 164 region, null_path, 165 offset); 166 } 167 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 168 RegionNode* region, int null_path) { 169 int offset = java_lang_Class::array_klass_offset_in_bytes(); 170 return load_klass_from_mirror_common(mirror, never_see_null, 171 region, null_path, 172 offset); 173 } 174 Node* generate_access_flags_guard(Node* kls, 175 int modifier_mask, int modifier_bits, 176 RegionNode* region); 177 Node* generate_interface_guard(Node* kls, RegionNode* region); 178 Node* generate_array_guard(Node* kls, RegionNode* region) { 179 return generate_array_guard_common(kls, region, false, false); 180 } 181 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 182 return generate_array_guard_common(kls, region, false, true); 183 } 184 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 185 return generate_array_guard_common(kls, region, true, false); 186 } 187 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 188 return generate_array_guard_common(kls, region, true, true); 189 } 190 Node* generate_array_guard_common(Node* kls, RegionNode* region, 191 bool obj_array, bool not_array); 192 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 193 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 194 bool is_virtual = false, bool is_static = false); 195 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 196 return generate_method_call(method_id, false, true); 197 } 198 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 199 return generate_method_call(method_id, true, false); 200 } 201 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 202 203 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 204 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 205 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 206 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 207 bool inline_string_indexOfChar(); 208 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 209 bool inline_string_toBytesU(); 210 bool inline_string_getCharsU(); 211 bool inline_string_copy(bool compress); 212 bool inline_string_char_access(bool is_store); 213 Node* round_double_node(Node* n); 214 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 215 bool inline_math_native(vmIntrinsics::ID id); 216 bool inline_trig(vmIntrinsics::ID id); 217 bool inline_math(vmIntrinsics::ID id); 218 template <typename OverflowOp> 219 bool inline_math_overflow(Node* arg1, Node* arg2); 220 void inline_math_mathExact(Node* math, Node* test); 221 bool inline_math_addExactI(bool is_increment); 222 bool inline_math_addExactL(bool is_increment); 223 bool inline_math_multiplyExactI(); 224 bool inline_math_multiplyExactL(); 225 bool inline_math_negateExactI(); 226 bool inline_math_negateExactL(); 227 bool inline_math_subtractExactI(bool is_decrement); 228 bool inline_math_subtractExactL(bool is_decrement); 229 bool inline_pow(); 230 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 231 bool inline_min_max(vmIntrinsics::ID id); 232 bool inline_notify(vmIntrinsics::ID id); 233 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 234 // This returns Type::AnyPtr, RawPtr, or OopPtr. 235 int classify_unsafe_addr(Node* &base, Node* &offset); 236 Node* make_unsafe_address(Node* base, Node* offset); 237 // Helper for inline_unsafe_access. 238 // Generates the guards that check whether the result of 239 // Unsafe.getObject should be recorded in an SATB log buffer. 240 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 241 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned); 242 static bool klass_needs_init_guard(Node* kls); 243 bool inline_unsafe_allocate(); 244 bool inline_unsafe_copyMemory(); 245 bool inline_native_currentThread(); 246 #ifdef TRACE_HAVE_INTRINSICS 247 bool inline_native_classID(); 248 bool inline_native_threadID(); 249 #endif 250 bool inline_native_time_funcs(address method, const char* funcName); 251 bool inline_native_isInterrupted(); 252 bool inline_native_Class_query(vmIntrinsics::ID id); 253 bool inline_native_subtype_check(); 254 255 bool inline_native_newArray(); 256 bool inline_native_getLength(); 257 bool inline_array_copyOf(bool is_copyOfRange); 258 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 259 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 260 bool inline_native_clone(bool is_virtual); 261 bool inline_native_Reflection_getCallerClass(); 262 // Helper function for inlining native object hash method 263 bool inline_native_hashcode(bool is_virtual, bool is_static); 264 bool inline_native_getClass(); 265 266 // Helper functions for inlining arraycopy 267 bool inline_arraycopy(); 268 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 269 RegionNode* slow_region); 270 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 271 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 272 273 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 274 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 275 bool inline_unsafe_ordered_store(BasicType type); 276 bool inline_unsafe_fence(vmIntrinsics::ID id); 277 bool inline_fp_conversions(vmIntrinsics::ID id); 278 bool inline_number_methods(vmIntrinsics::ID id); 279 bool inline_reference_get(); 280 bool inline_Class_cast(); 281 bool inline_aescrypt_Block(vmIntrinsics::ID id); 282 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 283 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 284 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 285 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 286 bool inline_ghash_processBlocks(); 287 bool inline_sha_implCompress(vmIntrinsics::ID id); 288 bool inline_digestBase_implCompressMB(int predicate); 289 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 290 bool long_state, address stubAddr, const char *stubName, 291 Node* src_start, Node* ofs, Node* limit); 292 Node* get_state_from_sha_object(Node *sha_object); 293 Node* get_state_from_sha5_object(Node *sha_object); 294 Node* inline_digestBase_implCompressMB_predicate(int predicate); 295 bool inline_encodeISOArray(); 296 bool inline_updateCRC32(); 297 bool inline_updateBytesCRC32(); 298 bool inline_updateByteBufferCRC32(); 299 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 300 bool inline_updateBytesCRC32C(); 301 bool inline_updateDirectByteBufferCRC32C(); 302 bool inline_updateBytesAdler32(); 303 bool inline_updateByteBufferAdler32(); 304 bool inline_multiplyToLen(); 305 bool inline_hasNegatives(); 306 bool inline_squareToLen(); 307 bool inline_mulAdd(); 308 bool inline_montgomeryMultiply(); 309 bool inline_montgomerySquare(); 310 311 bool inline_profileBoolean(); 312 bool inline_isCompileConstant(); 313 }; 314 315 //---------------------------make_vm_intrinsic---------------------------- 316 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 317 vmIntrinsics::ID id = m->intrinsic_id(); 318 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 319 320 if (!m->is_loaded()) { 321 // Do not attempt to inline unloaded methods. 322 return NULL; 323 } 324 325 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 326 bool is_available = false; 327 328 { 329 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 330 // the compiler must transition to '_thread_in_vm' state because both 331 // methods access VM-internal data. 332 VM_ENTRY_MARK; 333 methodHandle mh(THREAD, m->get_Method()); 334 is_available = compiler->is_intrinsic_supported(mh, is_virtual) && 335 !C->directive()->is_intrinsic_disabled(mh) && 336 !vmIntrinsics::is_disabled_by_flags(mh); 337 338 } 339 340 if (is_available) { 341 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 342 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 343 return new LibraryIntrinsic(m, is_virtual, 344 vmIntrinsics::predicates_needed(id), 345 vmIntrinsics::does_virtual_dispatch(id), 346 (vmIntrinsics::ID) id); 347 } else { 348 return NULL; 349 } 350 } 351 352 //----------------------register_library_intrinsics----------------------- 353 // Initialize this file's data structures, for each Compile instance. 354 void Compile::register_library_intrinsics() { 355 // Nothing to do here. 356 } 357 358 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 359 LibraryCallKit kit(jvms, this); 360 Compile* C = kit.C; 361 int nodes = C->unique(); 362 #ifndef PRODUCT 363 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 364 char buf[1000]; 365 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 366 tty->print_cr("Intrinsic %s", str); 367 } 368 #endif 369 ciMethod* callee = kit.callee(); 370 const int bci = kit.bci(); 371 372 // Try to inline the intrinsic. 373 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 374 kit.try_to_inline(_last_predicate)) { 375 if (C->print_intrinsics() || C->print_inlining()) { 376 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 377 } 378 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 379 if (C->log()) { 380 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 381 vmIntrinsics::name_at(intrinsic_id()), 382 (is_virtual() ? " virtual='1'" : ""), 383 C->unique() - nodes); 384 } 385 // Push the result from the inlined method onto the stack. 386 kit.push_result(); 387 C->print_inlining_update(this); 388 return kit.transfer_exceptions_into_jvms(); 389 } 390 391 // The intrinsic bailed out 392 if (C->print_intrinsics() || C->print_inlining()) { 393 if (jvms->has_method()) { 394 // Not a root compile. 395 const char* msg; 396 if (callee->intrinsic_candidate()) { 397 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 398 } else { 399 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 400 : "failed to inline (intrinsic), method not annotated"; 401 } 402 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 403 } else { 404 // Root compile 405 tty->print("Did not generate intrinsic %s%s at bci:%d in", 406 vmIntrinsics::name_at(intrinsic_id()), 407 (is_virtual() ? " (virtual)" : ""), bci); 408 } 409 } 410 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 411 C->print_inlining_update(this); 412 return NULL; 413 } 414 415 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 416 LibraryCallKit kit(jvms, this); 417 Compile* C = kit.C; 418 int nodes = C->unique(); 419 _last_predicate = predicate; 420 #ifndef PRODUCT 421 assert(is_predicated() && predicate < predicates_count(), "sanity"); 422 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 423 char buf[1000]; 424 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 425 tty->print_cr("Predicate for intrinsic %s", str); 426 } 427 #endif 428 ciMethod* callee = kit.callee(); 429 const int bci = kit.bci(); 430 431 Node* slow_ctl = kit.try_to_predicate(predicate); 432 if (!kit.failing()) { 433 if (C->print_intrinsics() || C->print_inlining()) { 434 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 435 } 436 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 437 if (C->log()) { 438 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 439 vmIntrinsics::name_at(intrinsic_id()), 440 (is_virtual() ? " virtual='1'" : ""), 441 C->unique() - nodes); 442 } 443 return slow_ctl; // Could be NULL if the check folds. 444 } 445 446 // The intrinsic bailed out 447 if (C->print_intrinsics() || C->print_inlining()) { 448 if (jvms->has_method()) { 449 // Not a root compile. 450 const char* msg = "failed to generate predicate for intrinsic"; 451 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 452 } else { 453 // Root compile 454 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 455 vmIntrinsics::name_at(intrinsic_id()), 456 (is_virtual() ? " (virtual)" : ""), bci); 457 } 458 } 459 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 460 return NULL; 461 } 462 463 bool LibraryCallKit::try_to_inline(int predicate) { 464 // Handle symbolic names for otherwise undistinguished boolean switches: 465 const bool is_store = true; 466 const bool is_compress = true; 467 const bool is_native_ptr = true; 468 const bool is_static = true; 469 const bool is_volatile = true; 470 471 if (!jvms()->has_method()) { 472 // Root JVMState has a null method. 473 assert(map()->memory()->Opcode() == Op_Parm, ""); 474 // Insert the memory aliasing node 475 set_all_memory(reset_memory()); 476 } 477 assert(merged_memory(), ""); 478 479 480 switch (intrinsic_id()) { 481 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 482 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 483 case vmIntrinsics::_getClass: return inline_native_getClass(); 484 485 case vmIntrinsics::_dsin: 486 case vmIntrinsics::_dcos: 487 case vmIntrinsics::_dtan: 488 case vmIntrinsics::_dabs: 489 case vmIntrinsics::_datan2: 490 case vmIntrinsics::_dsqrt: 491 case vmIntrinsics::_dexp: 492 case vmIntrinsics::_dlog: 493 case vmIntrinsics::_dlog10: 494 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 495 496 case vmIntrinsics::_min: 497 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 498 499 case vmIntrinsics::_notify: 500 case vmIntrinsics::_notifyAll: 501 if (InlineNotify) { 502 return inline_notify(intrinsic_id()); 503 } 504 return false; 505 506 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 507 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 508 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 509 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 510 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 511 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 512 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 513 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 514 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 515 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 516 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 517 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 518 519 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 520 521 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 522 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 523 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 524 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 525 526 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 527 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 528 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 529 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 530 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 531 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 532 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 533 534 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 535 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 536 537 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 538 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 539 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 540 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 541 542 case vmIntrinsics::_compressStringC: 543 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 544 case vmIntrinsics::_inflateStringC: 545 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 546 547 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false); 548 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false); 549 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 550 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 551 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 552 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false); 553 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false); 554 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 555 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 556 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false); 557 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false); 558 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false); 559 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false); 560 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false); 561 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false); 562 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false); 563 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 564 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 565 566 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 567 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 568 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 569 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false); 570 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false); 571 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 572 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 573 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false); 574 575 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false); 576 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false); 577 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false); 578 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false); 579 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false); 580 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 581 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 582 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false); 583 584 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false); 585 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false); 586 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false); 587 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false); 588 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false); 589 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false); 590 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false); 591 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false); 592 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false); 593 594 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false); 595 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false); 596 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false); 597 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false); 598 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false); 599 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false); 600 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false); 601 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false); 602 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false); 603 604 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, true); 605 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, true); 606 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, true); 607 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, true); 608 609 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, true); 610 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, true); 611 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, true); 612 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, true); 613 614 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 615 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 616 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 617 618 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 619 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 620 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 621 622 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 623 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 624 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 625 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 626 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 627 628 case vmIntrinsics::_loadFence: 629 case vmIntrinsics::_storeFence: 630 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 631 632 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 633 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 634 635 #ifdef TRACE_HAVE_INTRINSICS 636 case vmIntrinsics::_classID: return inline_native_classID(); 637 case vmIntrinsics::_threadID: return inline_native_threadID(); 638 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 639 #endif 640 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 641 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 642 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 643 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 644 case vmIntrinsics::_newArray: return inline_native_newArray(); 645 case vmIntrinsics::_getLength: return inline_native_getLength(); 646 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 647 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 648 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 649 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 650 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 651 652 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 653 654 case vmIntrinsics::_isInstance: 655 case vmIntrinsics::_getModifiers: 656 case vmIntrinsics::_isInterface: 657 case vmIntrinsics::_isArray: 658 case vmIntrinsics::_isPrimitive: 659 case vmIntrinsics::_getSuperclass: 660 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 661 662 case vmIntrinsics::_floatToRawIntBits: 663 case vmIntrinsics::_floatToIntBits: 664 case vmIntrinsics::_intBitsToFloat: 665 case vmIntrinsics::_doubleToRawLongBits: 666 case vmIntrinsics::_doubleToLongBits: 667 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 668 669 case vmIntrinsics::_numberOfLeadingZeros_i: 670 case vmIntrinsics::_numberOfLeadingZeros_l: 671 case vmIntrinsics::_numberOfTrailingZeros_i: 672 case vmIntrinsics::_numberOfTrailingZeros_l: 673 case vmIntrinsics::_bitCount_i: 674 case vmIntrinsics::_bitCount_l: 675 case vmIntrinsics::_reverseBytes_i: 676 case vmIntrinsics::_reverseBytes_l: 677 case vmIntrinsics::_reverseBytes_s: 678 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 679 680 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 681 682 case vmIntrinsics::_Reference_get: return inline_reference_get(); 683 684 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 685 686 case vmIntrinsics::_aescrypt_encryptBlock: 687 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 688 689 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 690 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 691 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 692 693 case vmIntrinsics::_sha_implCompress: 694 case vmIntrinsics::_sha2_implCompress: 695 case vmIntrinsics::_sha5_implCompress: 696 return inline_sha_implCompress(intrinsic_id()); 697 698 case vmIntrinsics::_digestBase_implCompressMB: 699 return inline_digestBase_implCompressMB(predicate); 700 701 case vmIntrinsics::_multiplyToLen: 702 return inline_multiplyToLen(); 703 704 case vmIntrinsics::_squareToLen: 705 return inline_squareToLen(); 706 707 case vmIntrinsics::_mulAdd: 708 return inline_mulAdd(); 709 710 case vmIntrinsics::_montgomeryMultiply: 711 return inline_montgomeryMultiply(); 712 case vmIntrinsics::_montgomerySquare: 713 return inline_montgomerySquare(); 714 715 case vmIntrinsics::_ghash_processBlocks: 716 return inline_ghash_processBlocks(); 717 718 case vmIntrinsics::_encodeISOArray: 719 case vmIntrinsics::_encodeByteISOArray: 720 return inline_encodeISOArray(); 721 722 case vmIntrinsics::_updateCRC32: 723 return inline_updateCRC32(); 724 case vmIntrinsics::_updateBytesCRC32: 725 return inline_updateBytesCRC32(); 726 case vmIntrinsics::_updateByteBufferCRC32: 727 return inline_updateByteBufferCRC32(); 728 729 case vmIntrinsics::_updateBytesCRC32C: 730 return inline_updateBytesCRC32C(); 731 case vmIntrinsics::_updateDirectByteBufferCRC32C: 732 return inline_updateDirectByteBufferCRC32C(); 733 734 case vmIntrinsics::_updateBytesAdler32: 735 return inline_updateBytesAdler32(); 736 case vmIntrinsics::_updateByteBufferAdler32: 737 return inline_updateByteBufferAdler32(); 738 739 case vmIntrinsics::_profileBoolean: 740 return inline_profileBoolean(); 741 case vmIntrinsics::_isCompileConstant: 742 return inline_isCompileConstant(); 743 744 case vmIntrinsics::_hasNegatives: 745 return inline_hasNegatives(); 746 747 default: 748 // If you get here, it may be that someone has added a new intrinsic 749 // to the list in vmSymbols.hpp without implementing it here. 750 #ifndef PRODUCT 751 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 752 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 753 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 754 } 755 #endif 756 return false; 757 } 758 } 759 760 Node* LibraryCallKit::try_to_predicate(int predicate) { 761 if (!jvms()->has_method()) { 762 // Root JVMState has a null method. 763 assert(map()->memory()->Opcode() == Op_Parm, ""); 764 // Insert the memory aliasing node 765 set_all_memory(reset_memory()); 766 } 767 assert(merged_memory(), ""); 768 769 switch (intrinsic_id()) { 770 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 771 return inline_cipherBlockChaining_AESCrypt_predicate(false); 772 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 773 return inline_cipherBlockChaining_AESCrypt_predicate(true); 774 case vmIntrinsics::_digestBase_implCompressMB: 775 return inline_digestBase_implCompressMB_predicate(predicate); 776 777 default: 778 // If you get here, it may be that someone has added a new intrinsic 779 // to the list in vmSymbols.hpp without implementing it here. 780 #ifndef PRODUCT 781 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 782 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 783 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 784 } 785 #endif 786 Node* slow_ctl = control(); 787 set_control(top()); // No fast path instrinsic 788 return slow_ctl; 789 } 790 } 791 792 //------------------------------set_result------------------------------- 793 // Helper function for finishing intrinsics. 794 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 795 record_for_igvn(region); 796 set_control(_gvn.transform(region)); 797 set_result( _gvn.transform(value)); 798 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 799 } 800 801 //------------------------------generate_guard--------------------------- 802 // Helper function for generating guarded fast-slow graph structures. 803 // The given 'test', if true, guards a slow path. If the test fails 804 // then a fast path can be taken. (We generally hope it fails.) 805 // In all cases, GraphKit::control() is updated to the fast path. 806 // The returned value represents the control for the slow path. 807 // The return value is never 'top'; it is either a valid control 808 // or NULL if it is obvious that the slow path can never be taken. 809 // Also, if region and the slow control are not NULL, the slow edge 810 // is appended to the region. 811 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 812 if (stopped()) { 813 // Already short circuited. 814 return NULL; 815 } 816 817 // Build an if node and its projections. 818 // If test is true we take the slow path, which we assume is uncommon. 819 if (_gvn.type(test) == TypeInt::ZERO) { 820 // The slow branch is never taken. No need to build this guard. 821 return NULL; 822 } 823 824 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 825 826 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 827 if (if_slow == top()) { 828 // The slow branch is never taken. No need to build this guard. 829 return NULL; 830 } 831 832 if (region != NULL) 833 region->add_req(if_slow); 834 835 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 836 set_control(if_fast); 837 838 return if_slow; 839 } 840 841 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 842 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 843 } 844 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 845 return generate_guard(test, region, PROB_FAIR); 846 } 847 848 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 849 Node* *pos_index) { 850 if (stopped()) 851 return NULL; // already stopped 852 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 853 return NULL; // index is already adequately typed 854 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 855 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 856 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 857 if (is_neg != NULL && pos_index != NULL) { 858 // Emulate effect of Parse::adjust_map_after_if. 859 Node* ccast = new CastIINode(index, TypeInt::POS); 860 ccast->set_req(0, control()); 861 (*pos_index) = _gvn.transform(ccast); 862 } 863 return is_neg; 864 } 865 866 // Make sure that 'position' is a valid limit index, in [0..length]. 867 // There are two equivalent plans for checking this: 868 // A. (offset + copyLength) unsigned<= arrayLength 869 // B. offset <= (arrayLength - copyLength) 870 // We require that all of the values above, except for the sum and 871 // difference, are already known to be non-negative. 872 // Plan A is robust in the face of overflow, if offset and copyLength 873 // are both hugely positive. 874 // 875 // Plan B is less direct and intuitive, but it does not overflow at 876 // all, since the difference of two non-negatives is always 877 // representable. Whenever Java methods must perform the equivalent 878 // check they generally use Plan B instead of Plan A. 879 // For the moment we use Plan A. 880 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 881 Node* subseq_length, 882 Node* array_length, 883 RegionNode* region) { 884 if (stopped()) 885 return NULL; // already stopped 886 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 887 if (zero_offset && subseq_length->eqv_uncast(array_length)) 888 return NULL; // common case of whole-array copy 889 Node* last = subseq_length; 890 if (!zero_offset) // last += offset 891 last = _gvn.transform(new AddINode(last, offset)); 892 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 893 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 894 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 895 return is_over; 896 } 897 898 899 //--------------------------generate_current_thread-------------------- 900 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 901 ciKlass* thread_klass = env()->Thread_klass(); 902 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 903 Node* thread = _gvn.transform(new ThreadLocalNode()); 904 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 905 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 906 tls_output = thread; 907 return threadObj; 908 } 909 910 911 //------------------------------make_string_method_node------------------------ 912 // Helper method for String intrinsic functions. This version is called with 913 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 914 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 915 // containing the lengths of str1 and str2. 916 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 917 Node* result = NULL; 918 switch (opcode) { 919 case Op_StrIndexOf: 920 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 921 str1_start, cnt1, str2_start, cnt2, ae); 922 break; 923 case Op_StrComp: 924 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 925 str1_start, cnt1, str2_start, cnt2, ae); 926 break; 927 case Op_StrEquals: 928 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 929 str1_start, str2_start, cnt1, ae); 930 break; 931 default: 932 ShouldNotReachHere(); 933 return NULL; 934 } 935 936 // All these intrinsics have checks. 937 C->set_has_split_ifs(true); // Has chance for split-if optimization 938 939 return _gvn.transform(result); 940 } 941 942 //------------------------------inline_string_compareTo------------------------ 943 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 944 Node* arg1 = argument(0); 945 Node* arg2 = argument(1); 946 947 // Get start addr and length of first argument 948 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 949 Node* arg1_cnt = load_array_length(arg1); 950 951 // Get start addr and length of second argument 952 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 953 Node* arg2_cnt = load_array_length(arg2); 954 955 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 956 set_result(result); 957 return true; 958 } 959 960 //------------------------------inline_string_equals------------------------ 961 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 962 Node* arg1 = argument(0); 963 Node* arg2 = argument(1); 964 965 // paths (plus control) merge 966 RegionNode* region = new RegionNode(3); 967 Node* phi = new PhiNode(region, TypeInt::BOOL); 968 969 if (!stopped()) { 970 // Get start addr and length of first argument 971 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 972 Node* arg1_cnt = load_array_length(arg1); 973 974 // Get start addr and length of second argument 975 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 976 Node* arg2_cnt = load_array_length(arg2); 977 978 // Check for arg1_cnt != arg2_cnt 979 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 980 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 981 Node* if_ne = generate_slow_guard(bol, NULL); 982 if (if_ne != NULL) { 983 phi->init_req(2, intcon(0)); 984 region->init_req(2, if_ne); 985 } 986 987 // Check for count == 0 is done by assembler code for StrEquals. 988 989 if (!stopped()) { 990 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 991 phi->init_req(1, equals); 992 region->init_req(1, control()); 993 } 994 } 995 996 // post merge 997 set_control(_gvn.transform(region)); 998 record_for_igvn(region); 999 1000 set_result(_gvn.transform(phi)); 1001 return true; 1002 } 1003 1004 //------------------------------inline_array_equals---------------------------- 1005 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1006 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1007 Node* arg1 = argument(0); 1008 Node* arg2 = argument(1); 1009 1010 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1011 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1012 return true; 1013 } 1014 1015 //------------------------------inline_hasNegatives------------------------------ 1016 bool LibraryCallKit::inline_hasNegatives() { 1017 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1018 1019 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1020 // no receiver since it is static method 1021 Node* ba = argument(0); 1022 Node* offset = argument(1); 1023 Node* len = argument(2); 1024 1025 RegionNode* bailout = new RegionNode(1); 1026 record_for_igvn(bailout); 1027 1028 // offset must not be negative. 1029 generate_negative_guard(offset, bailout); 1030 1031 // offset + length must not exceed length of ba. 1032 generate_limit_guard(offset, len, load_array_length(ba), bailout); 1033 1034 if (bailout->req() > 1) { 1035 PreserveJVMState pjvms(this); 1036 set_control(_gvn.transform(bailout)); 1037 uncommon_trap(Deoptimization::Reason_intrinsic, 1038 Deoptimization::Action_maybe_recompile); 1039 } 1040 if (!stopped()) { 1041 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1042 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1043 set_result(_gvn.transform(result)); 1044 } 1045 return true; 1046 } 1047 1048 //------------------------------inline_string_indexOf------------------------ 1049 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1050 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) { 1051 return false; 1052 } 1053 Node* src = argument(0); 1054 Node* tgt = argument(1); 1055 1056 // Make the merge point 1057 RegionNode* result_rgn = new RegionNode(4); 1058 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1059 1060 // Get start addr and length of source string 1061 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1062 Node* src_count = load_array_length(src); 1063 1064 // Get start addr and length of substring 1065 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1066 Node* tgt_count = load_array_length(tgt); 1067 1068 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1069 // Divide src size by 2 if String is UTF16 encoded 1070 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1071 } 1072 if (ae == StrIntrinsicNode::UU) { 1073 // Divide substring size by 2 if String is UTF16 encoded 1074 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1075 } 1076 1077 // Check for substr count > string count 1078 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1079 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1080 Node* if_gt = generate_slow_guard(bol, NULL); 1081 if (if_gt != NULL) { 1082 result_phi->init_req(2, intcon(-1)); 1083 result_rgn->init_req(2, if_gt); 1084 } 1085 1086 if (!stopped()) { 1087 // Check for substr count == 0 1088 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1089 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1090 Node* if_zero = generate_slow_guard(bol, NULL); 1091 if (if_zero != NULL) { 1092 result_phi->init_req(3, intcon(0)); 1093 result_rgn->init_req(3, if_zero); 1094 } 1095 } 1096 1097 if (!stopped()) { 1098 Node* result = make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1099 result_phi->init_req(1, result); 1100 result_rgn->init_req(1, control()); 1101 } 1102 set_control(_gvn.transform(result_rgn)); 1103 record_for_igvn(result_rgn); 1104 set_result(_gvn.transform(result_phi)); 1105 1106 return true; 1107 } 1108 1109 //-----------------------------inline_string_indexOf----------------------- 1110 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1111 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) { 1112 return false; 1113 } 1114 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1115 Node* src = argument(0); // byte[] 1116 Node* src_count = argument(1); 1117 Node* tgt = argument(2); // byte[] 1118 Node* tgt_count = argument(3); 1119 Node* from_index = argument(4); 1120 1121 // Multiply byte array index by 2 if String is UTF16 encoded 1122 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1123 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1124 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1125 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1126 1127 RegionNode* bailout = new RegionNode(1); 1128 record_for_igvn(bailout); 1129 1130 // Range checks 1131 generate_negative_guard(src_offset, bailout); 1132 generate_negative_guard(src_count, bailout); 1133 generate_negative_guard(tgt_count, bailout); 1134 Node* src_byte_count = _gvn.transform(new LShiftINode(src_count, intcon(1))); 1135 Node* tgt_byte_count = _gvn.transform(new LShiftINode(tgt_count, intcon(1))); 1136 generate_limit_guard(src_offset, (ae == StrIntrinsicNode::LL) ? src_count : src_byte_count, load_array_length(src), bailout); 1137 generate_limit_guard(intcon(0), (ae == StrIntrinsicNode::UU) ? tgt_byte_count : tgt_count, load_array_length(tgt), bailout); 1138 1139 if (bailout->req() > 1) { 1140 PreserveJVMState pjvms(this); 1141 set_control(_gvn.transform(bailout)); 1142 uncommon_trap(Deoptimization::Reason_intrinsic, 1143 Deoptimization::Action_maybe_recompile); 1144 } 1145 if (stopped()) { 1146 return true; 1147 } 1148 1149 Node* result = make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1150 1151 // The result is index relative to from_index if substring was found, -1 otherwise. 1152 // Generate code which will fold into cmove. 1153 RegionNode* region = new RegionNode(3); 1154 Node* phi = new PhiNode(region, TypeInt::INT); 1155 1156 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1157 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1158 1159 Node* if_lt = generate_slow_guard(bol, NULL); 1160 if (if_lt != NULL) { 1161 // result == -1 1162 phi->init_req(2, result); 1163 region->init_req(2, if_lt); 1164 } 1165 if (!stopped()) { 1166 result = _gvn.transform(new AddINode(result, from_index)); 1167 phi->init_req(1, result); 1168 region->init_req(1, control()); 1169 } 1170 1171 set_control(_gvn.transform(region)); 1172 record_for_igvn(region); 1173 set_result(_gvn.transform(phi)); 1174 1175 return true; 1176 } 1177 1178 //-----------------------------inline_string_indexOfChar----------------------- 1179 bool LibraryCallKit::inline_string_indexOfChar() { 1180 if (!Matcher::has_match_rule(Op_StrIndexOfChar) || !(UseSSE > 4)) { 1181 return false; 1182 } 1183 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1184 Node* src = argument(0); // byte[] 1185 Node* tgt = argument(1); // tgt is int ch 1186 Node* from_index = argument(2); 1187 Node* max = argument(3); 1188 1189 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1190 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1191 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1192 1193 RegionNode* bailout = new RegionNode(1); 1194 record_for_igvn(bailout); 1195 1196 // Range checks 1197 generate_negative_guard(src_offset, bailout); 1198 generate_negative_guard(src_count, bailout); 1199 Node* byte_length = _gvn.transform(new LShiftINode(src_count, intcon(1))); 1200 generate_limit_guard(src_offset, byte_length, load_array_length(src), bailout); 1201 1202 if (bailout->req() > 1) { 1203 PreserveJVMState pjvms(this); 1204 set_control(_gvn.transform(bailout)); 1205 uncommon_trap(Deoptimization::Reason_intrinsic, 1206 Deoptimization::Action_maybe_recompile); 1207 } 1208 if (stopped()) { 1209 return true; 1210 } 1211 1212 RegionNode* region = new RegionNode(3); 1213 Node* phi = new PhiNode(region, TypeInt::INT); 1214 1215 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1216 C->set_has_split_ifs(true); // Has chance for split-if optimization 1217 _gvn.transform(result); 1218 1219 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1220 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1221 1222 Node* if_lt = generate_slow_guard(bol, NULL); 1223 if (if_lt != NULL) { 1224 // result == -1 1225 phi->init_req(2, result); 1226 region->init_req(2, if_lt); 1227 } 1228 if (!stopped()) { 1229 result = _gvn.transform(new AddINode(result, from_index)); 1230 phi->init_req(1, result); 1231 region->init_req(1, control()); 1232 } 1233 set_control(_gvn.transform(region)); 1234 record_for_igvn(region); 1235 set_result(_gvn.transform(phi)); 1236 1237 return true; 1238 } 1239 //---------------------------inline_string_copy--------------------- 1240 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1241 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1242 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1243 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1244 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1245 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1246 bool LibraryCallKit::inline_string_copy(bool compress) { 1247 int nargs = 5; // 2 oops, 3 ints 1248 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1249 1250 Node* src = argument(0); 1251 Node* src_offset = argument(1); 1252 Node* dst = argument(2); 1253 Node* dst_offset = argument(3); 1254 Node* length = argument(4); 1255 1256 // Check for allocation before we add nodes that would confuse 1257 // tightly_coupled_allocation() 1258 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1259 1260 // Figure out the size and type of the elements we will be copying. 1261 const Type* src_type = src->Value(&_gvn); 1262 const Type* dst_type = dst->Value(&_gvn); 1263 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1264 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1265 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1266 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1267 "Unsupported array types for inline_string_copy"); 1268 1269 RegionNode* bailout = new RegionNode(1); 1270 record_for_igvn(bailout); 1271 1272 // Range checks 1273 generate_negative_guard(src_offset, bailout); 1274 generate_negative_guard(dst_offset, bailout); 1275 generate_negative_guard(length, bailout); 1276 Node* byte_length = _gvn.transform(new LShiftINode(length, intcon(1))); 1277 generate_limit_guard(src_offset, (compress && src_elem == T_BYTE) ? byte_length : length, load_array_length(src), bailout); 1278 generate_limit_guard(dst_offset, (!compress && dst_elem == T_BYTE) ? byte_length : length, load_array_length(dst), bailout); 1279 1280 if (bailout->req() > 1) { 1281 PreserveJVMState pjvms(this); 1282 set_control(_gvn.transform(bailout)); 1283 uncommon_trap(Deoptimization::Reason_intrinsic, 1284 Deoptimization::Action_maybe_recompile); 1285 } 1286 if (stopped()) { 1287 return true; 1288 } 1289 1290 // Convert char[] offsets to byte[] offsets 1291 if (compress && src_elem == T_BYTE) { 1292 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1293 } else if (!compress && dst_elem == T_BYTE) { 1294 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1295 } 1296 1297 Node* src_start = array_element_address(src, src_offset, src_elem); 1298 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1299 // 'src_start' points to src array + scaled offset 1300 // 'dst_start' points to dst array + scaled offset 1301 Node* count; 1302 if (compress) { 1303 count = compress_string(src_start, dst_start, length); 1304 } else { 1305 inflate_string(src_start, dst_start, length); 1306 } 1307 1308 if (alloc != NULL) { 1309 if (alloc->maybe_set_complete(&_gvn)) { 1310 // "You break it, you buy it." 1311 InitializeNode* init = alloc->initialization(); 1312 assert(init->is_complete(), "we just did this"); 1313 init->set_complete_with_arraycopy(); 1314 assert(dst->is_CheckCastPP(), "sanity"); 1315 assert(dst->in(0)->in(0) == init, "dest pinned"); 1316 } 1317 // Do not let stores that initialize this object be reordered with 1318 // a subsequent store that would make this object accessible by 1319 // other threads. 1320 // Record what AllocateNode this StoreStore protects so that 1321 // escape analysis can go from the MemBarStoreStoreNode to the 1322 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1323 // based on the escape status of the AllocateNode. 1324 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1325 } 1326 if (compress) { 1327 set_result(_gvn.transform(count)); 1328 } 1329 return true; 1330 } 1331 1332 #ifdef _LP64 1333 #define XTOP ,top() /*additional argument*/ 1334 #else //_LP64 1335 #define XTOP /*no additional argument*/ 1336 #endif //_LP64 1337 1338 //------------------------inline_string_toBytesU-------------------------- 1339 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1340 bool LibraryCallKit::inline_string_toBytesU() { 1341 // Get the arguments. 1342 Node* value = argument(0); 1343 Node* offset = argument(1); 1344 Node* length = argument(2); 1345 1346 Node* newcopy = NULL; 1347 1348 // Set the original stack and the reexecute bit for the interpreter to reexecute 1349 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1350 { PreserveReexecuteState preexecs(this); 1351 jvms()->set_should_reexecute(true); 1352 1353 // Check if a null path was taken unconditionally. 1354 value = null_check(value); 1355 1356 RegionNode* bailout = new RegionNode(1); 1357 record_for_igvn(bailout); 1358 1359 // Range checks 1360 generate_negative_guard(length, bailout); 1361 generate_negative_guard(offset, bailout); 1362 generate_limit_guard(offset, length, load_array_length(value), bailout); 1363 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1364 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1365 1366 if (bailout->req() > 1) { 1367 PreserveJVMState pjvms(this); 1368 set_control(_gvn.transform(bailout)); 1369 uncommon_trap(Deoptimization::Reason_intrinsic, 1370 Deoptimization::Action_maybe_recompile); 1371 } 1372 if (stopped()) { 1373 return true; 1374 } 1375 1376 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1377 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1378 newcopy = new_array(klass_node, size, 0); // no arguments to push 1379 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1380 1381 // Calculate starting addresses. 1382 Node* src_start = array_element_address(value, offset, T_CHAR); 1383 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1384 1385 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1386 const TypeInt* toffset = gvn().type(offset)->is_int(); 1387 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1388 1389 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1390 const char* copyfunc_name = "arraycopy"; 1391 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1392 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1393 OptoRuntime::fast_arraycopy_Type(), 1394 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1395 src_start, dst_start, ConvI2X(length) XTOP); 1396 // Do not let reads from the cloned object float above the arraycopy. 1397 if (alloc != NULL) { 1398 if (alloc->maybe_set_complete(&_gvn)) { 1399 // "You break it, you buy it." 1400 InitializeNode* init = alloc->initialization(); 1401 assert(init->is_complete(), "we just did this"); 1402 init->set_complete_with_arraycopy(); 1403 assert(newcopy->is_CheckCastPP(), "sanity"); 1404 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1405 } 1406 // Do not let stores that initialize this object be reordered with 1407 // a subsequent store that would make this object accessible by 1408 // other threads. 1409 // Record what AllocateNode this StoreStore protects so that 1410 // escape analysis can go from the MemBarStoreStoreNode to the 1411 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1412 // based on the escape status of the AllocateNode. 1413 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1414 } else { 1415 insert_mem_bar(Op_MemBarCPUOrder); 1416 } 1417 } // original reexecute is set back here 1418 1419 C->set_has_split_ifs(true); // Has chance for split-if optimization 1420 if (!stopped()) { 1421 set_result(newcopy); 1422 } 1423 return true; 1424 } 1425 1426 //------------------------inline_string_getCharsU-------------------------- 1427 // public void StringUTF16.getChars(byte[] value, int srcBegin, int srcEnd, char dst[], int dstBegin) 1428 bool LibraryCallKit::inline_string_getCharsU() { 1429 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1430 1431 // Get the arguments. 1432 Node* value = argument(0); 1433 Node* src_begin = argument(1); 1434 Node* src_end = argument(2); // exclusive offset (i < src_end) 1435 Node* dst = argument(3); 1436 Node* dst_begin = argument(4); 1437 1438 // Check for allocation before we add nodes that would confuse 1439 // tightly_coupled_allocation() 1440 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1441 1442 // Check if a null path was taken unconditionally. 1443 value = null_check(value); 1444 dst = null_check(dst); 1445 if (stopped()) { 1446 return true; 1447 } 1448 1449 // Get length and convert char[] offset to byte[] offset 1450 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1451 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1452 1453 RegionNode* bailout = new RegionNode(1); 1454 record_for_igvn(bailout); 1455 1456 // Range checks 1457 generate_negative_guard(length, bailout); 1458 generate_negative_guard(src_begin, bailout); 1459 generate_negative_guard(dst_begin, bailout); 1460 Node* byte_length = _gvn.transform(new LShiftINode(length, intcon(1))); 1461 generate_limit_guard(src_begin, byte_length, load_array_length(value), bailout); 1462 generate_limit_guard(dst_begin, length, load_array_length(dst), bailout); 1463 1464 if (bailout->req() > 1) { 1465 PreserveJVMState pjvms(this); 1466 set_control(_gvn.transform(bailout)); 1467 uncommon_trap(Deoptimization::Reason_intrinsic, 1468 Deoptimization::Action_maybe_recompile); 1469 } 1470 if (stopped()) { 1471 return true; 1472 } 1473 1474 if (!stopped()) { 1475 // Calculate starting addresses. 1476 Node* src_start = array_element_address(value, src_begin, T_BYTE); 1477 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1478 1479 // Check if array addresses are aligned to HeapWordSize 1480 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1481 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1482 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1483 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1484 1485 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1486 const char* copyfunc_name = "arraycopy"; 1487 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1488 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1489 OptoRuntime::fast_arraycopy_Type(), 1490 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1491 src_start, dst_start, ConvI2X(length) XTOP); 1492 // Do not let reads from the cloned object float above the arraycopy. 1493 if (alloc != NULL) { 1494 if (alloc->maybe_set_complete(&_gvn)) { 1495 // "You break it, you buy it." 1496 InitializeNode* init = alloc->initialization(); 1497 assert(init->is_complete(), "we just did this"); 1498 init->set_complete_with_arraycopy(); 1499 assert(dst->is_CheckCastPP(), "sanity"); 1500 assert(dst->in(0)->in(0) == init, "dest pinned"); 1501 } 1502 // Do not let stores that initialize this object be reordered with 1503 // a subsequent store that would make this object accessible by 1504 // other threads. 1505 // Record what AllocateNode this StoreStore protects so that 1506 // escape analysis can go from the MemBarStoreStoreNode to the 1507 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1508 // based on the escape status of the AllocateNode. 1509 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1510 } else { 1511 insert_mem_bar(Op_MemBarCPUOrder); 1512 } 1513 } 1514 1515 C->set_has_split_ifs(true); // Has chance for split-if optimization 1516 return true; 1517 } 1518 1519 //----------------------inline_string_char_access---------------------------- 1520 // Store/Load char to/from byte[] array. 1521 // static void StringUTF16.putChar(byte[] val, int index, int c) 1522 // static char StringUTF16.getChar(byte[] val, int index) 1523 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1524 Node* value = argument(0); 1525 Node* index = argument(1); 1526 Node* ch = is_store ? argument(2) : NULL; 1527 1528 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1529 // correctly requires matched array shapes. 1530 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1531 "sanity: byte[] and char[] bases agree"); 1532 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1533 "sanity: byte[] and char[] scales agree"); 1534 1535 Node* adr = array_element_address(value, index, T_CHAR); 1536 if (is_store) { 1537 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered); 1538 } else { 1539 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, MemNode::unordered); 1540 set_result(ch); 1541 } 1542 return true; 1543 } 1544 1545 //--------------------------round_double_node-------------------------------- 1546 // Round a double node if necessary. 1547 Node* LibraryCallKit::round_double_node(Node* n) { 1548 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1549 n = _gvn.transform(new RoundDoubleNode(0, n)); 1550 return n; 1551 } 1552 1553 //------------------------------inline_math----------------------------------- 1554 // public static double Math.abs(double) 1555 // public static double Math.sqrt(double) 1556 // public static double Math.log(double) 1557 // public static double Math.log10(double) 1558 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1559 Node* arg = round_double_node(argument(0)); 1560 Node* n; 1561 switch (id) { 1562 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1563 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1564 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1565 default: fatal_unexpected_iid(id); break; 1566 } 1567 set_result(_gvn.transform(n)); 1568 return true; 1569 } 1570 1571 //------------------------------inline_trig---------------------------------- 1572 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1573 // argument reduction which will turn into a fast/slow diamond. 1574 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1575 Node* arg = round_double_node(argument(0)); 1576 Node* n = NULL; 1577 1578 switch (id) { 1579 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1580 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1581 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1582 default: fatal_unexpected_iid(id); break; 1583 } 1584 n = _gvn.transform(n); 1585 1586 // Rounding required? Check for argument reduction! 1587 if (Matcher::strict_fp_requires_explicit_rounding) { 1588 static const double pi_4 = 0.7853981633974483; 1589 static const double neg_pi_4 = -0.7853981633974483; 1590 // pi/2 in 80-bit extended precision 1591 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1592 // -pi/2 in 80-bit extended precision 1593 // 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}; 1594 // Cutoff value for using this argument reduction technique 1595 //static const double pi_2_minus_epsilon = 1.564660403643354; 1596 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1597 1598 // Pseudocode for sin: 1599 // if (x <= Math.PI / 4.0) { 1600 // if (x >= -Math.PI / 4.0) return fsin(x); 1601 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1602 // } else { 1603 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1604 // } 1605 // return StrictMath.sin(x); 1606 1607 // Pseudocode for cos: 1608 // if (x <= Math.PI / 4.0) { 1609 // if (x >= -Math.PI / 4.0) return fcos(x); 1610 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1611 // } else { 1612 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1613 // } 1614 // return StrictMath.cos(x); 1615 1616 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1617 // requires a special machine instruction to load it. Instead we'll try 1618 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1619 // probably do the math inside the SIN encoding. 1620 1621 // Make the merge point 1622 RegionNode* r = new RegionNode(3); 1623 Node* phi = new PhiNode(r, Type::DOUBLE); 1624 1625 // Flatten arg so we need only 1 test 1626 Node *abs = _gvn.transform(new AbsDNode(arg)); 1627 // Node for PI/4 constant 1628 Node *pi4 = makecon(TypeD::make(pi_4)); 1629 // Check PI/4 : abs(arg) 1630 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1631 // Check: If PI/4 < abs(arg) then go slow 1632 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1633 // Branch either way 1634 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1635 set_control(opt_iff(r,iff)); 1636 1637 // Set fast path result 1638 phi->init_req(2, n); 1639 1640 // Slow path - non-blocking leaf call 1641 Node* call = NULL; 1642 switch (id) { 1643 case vmIntrinsics::_dsin: 1644 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1645 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1646 "Sin", NULL, arg, top()); 1647 break; 1648 case vmIntrinsics::_dcos: 1649 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1650 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1651 "Cos", NULL, arg, top()); 1652 break; 1653 case vmIntrinsics::_dtan: 1654 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1655 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1656 "Tan", NULL, arg, top()); 1657 break; 1658 } 1659 assert(control()->in(0) == call, ""); 1660 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1661 r->init_req(1, control()); 1662 phi->init_req(1, slow_result); 1663 1664 // Post-merge 1665 set_control(_gvn.transform(r)); 1666 record_for_igvn(r); 1667 n = _gvn.transform(phi); 1668 1669 C->set_has_split_ifs(true); // Has chance for split-if optimization 1670 } 1671 set_result(n); 1672 return true; 1673 } 1674 1675 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1676 //------------------- 1677 //result=(result.isNaN())? funcAddr():result; 1678 // Check: If isNaN() by checking result!=result? then either trap 1679 // or go to runtime 1680 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1681 // Build the boolean node 1682 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1683 1684 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1685 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1686 // The pow or exp intrinsic returned a NaN, which requires a call 1687 // to the runtime. Recompile with the runtime call. 1688 uncommon_trap(Deoptimization::Reason_intrinsic, 1689 Deoptimization::Action_make_not_entrant); 1690 } 1691 return result; 1692 } else { 1693 // If this inlining ever returned NaN in the past, we compile a call 1694 // to the runtime to properly handle corner cases 1695 1696 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1697 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1698 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1699 1700 if (!if_slow->is_top()) { 1701 RegionNode* result_region = new RegionNode(3); 1702 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1703 1704 result_region->init_req(1, if_fast); 1705 result_val->init_req(1, result); 1706 1707 set_control(if_slow); 1708 1709 const TypePtr* no_memory_effects = NULL; 1710 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1711 no_memory_effects, 1712 x, top(), y, y ? top() : NULL); 1713 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1714 #ifdef ASSERT 1715 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1716 assert(value_top == top(), "second value must be top"); 1717 #endif 1718 1719 result_region->init_req(2, control()); 1720 result_val->init_req(2, value); 1721 set_control(_gvn.transform(result_region)); 1722 return _gvn.transform(result_val); 1723 } else { 1724 return result; 1725 } 1726 } 1727 } 1728 1729 //------------------------------inline_pow------------------------------------- 1730 // Inline power instructions, if possible. 1731 bool LibraryCallKit::inline_pow() { 1732 // Pseudocode for pow 1733 // if (y == 2) { 1734 // return x * x; 1735 // } else { 1736 // if (x <= 0.0) { 1737 // long longy = (long)y; 1738 // if ((double)longy == y) { // if y is long 1739 // if (y + 1 == y) longy = 0; // huge number: even 1740 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1741 // } else { 1742 // result = NaN; 1743 // } 1744 // } else { 1745 // result = DPow(x,y); 1746 // } 1747 // if (result != result)? { 1748 // result = uncommon_trap() or runtime_call(); 1749 // } 1750 // return result; 1751 // } 1752 1753 Node* x = round_double_node(argument(0)); 1754 Node* y = round_double_node(argument(2)); 1755 1756 Node* result = NULL; 1757 1758 Node* const_two_node = makecon(TypeD::make(2.0)); 1759 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1760 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1761 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1762 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1763 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1764 1765 RegionNode* region_node = new RegionNode(3); 1766 region_node->init_req(1, if_true); 1767 1768 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1769 // special case for x^y where y == 2, we can convert it to x * x 1770 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1771 1772 // set control to if_false since we will now process the false branch 1773 set_control(if_false); 1774 1775 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1776 // Short form: skip the fancy tests and just check for NaN result. 1777 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1778 } else { 1779 // If this inlining ever returned NaN in the past, include all 1780 // checks + call to the runtime. 1781 1782 // Set the merge point for If node with condition of (x <= 0.0) 1783 // There are four possible paths to region node and phi node 1784 RegionNode *r = new RegionNode(4); 1785 Node *phi = new PhiNode(r, Type::DOUBLE); 1786 1787 // Build the first if node: if (x <= 0.0) 1788 // Node for 0 constant 1789 Node *zeronode = makecon(TypeD::ZERO); 1790 // Check x:0 1791 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1792 // Check: If (x<=0) then go complex path 1793 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1794 // Branch either way 1795 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1796 // Fast path taken; set region slot 3 1797 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1798 r->init_req(3,fast_taken); // Capture fast-control 1799 1800 // Fast path not-taken, i.e. slow path 1801 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1802 1803 // Set fast path result 1804 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1805 phi->init_req(3, fast_result); 1806 1807 // Complex path 1808 // Build the second if node (if y is long) 1809 // Node for (long)y 1810 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1811 // Node for (double)((long) y) 1812 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1813 // Check (double)((long) y) : y 1814 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1815 // Check if (y isn't long) then go to slow path 1816 1817 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1818 // Branch either way 1819 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1820 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1821 1822 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1823 1824 // Calculate DPow(abs(x), y)*(1 & (long)y) 1825 // Node for constant 1 1826 Node *conone = longcon(1); 1827 // 1& (long)y 1828 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1829 1830 // A huge number is always even. Detect a huge number by checking 1831 // if y + 1 == y and set integer to be tested for parity to 0. 1832 // Required for corner case: 1833 // (long)9.223372036854776E18 = max_jlong 1834 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1835 // max_jlong is odd but 9.223372036854776E18 is even 1836 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1837 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1838 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1839 Node* correctedsign = NULL; 1840 if (ConditionalMoveLimit != 0) { 1841 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1842 } else { 1843 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1844 RegionNode *r = new RegionNode(3); 1845 Node *phi = new PhiNode(r, TypeLong::LONG); 1846 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1847 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1848 phi->init_req(1, signnode); 1849 phi->init_req(2, longcon(0)); 1850 correctedsign = _gvn.transform(phi); 1851 ylong_path = _gvn.transform(r); 1852 record_for_igvn(r); 1853 } 1854 1855 // zero node 1856 Node *conzero = longcon(0); 1857 // Check (1&(long)y)==0? 1858 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1859 // Check if (1&(long)y)!=0?, if so the result is negative 1860 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1861 // abs(x) 1862 Node *absx=_gvn.transform(new AbsDNode(x)); 1863 // abs(x)^y 1864 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1865 // -abs(x)^y 1866 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1867 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1868 Node *signresult = NULL; 1869 if (ConditionalMoveLimit != 0) { 1870 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1871 } else { 1872 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1873 RegionNode *r = new RegionNode(3); 1874 Node *phi = new PhiNode(r, Type::DOUBLE); 1875 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1876 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1877 phi->init_req(1, absxpowy); 1878 phi->init_req(2, negabsxpowy); 1879 signresult = _gvn.transform(phi); 1880 ylong_path = _gvn.transform(r); 1881 record_for_igvn(r); 1882 } 1883 // Set complex path fast result 1884 r->init_req(2, ylong_path); 1885 phi->init_req(2, signresult); 1886 1887 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1888 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1889 r->init_req(1,slow_path); 1890 phi->init_req(1,slow_result); 1891 1892 // Post merge 1893 set_control(_gvn.transform(r)); 1894 record_for_igvn(r); 1895 result = _gvn.transform(phi); 1896 } 1897 1898 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1899 1900 // control from finish_pow_exp is now input to the region node 1901 region_node->set_req(2, control()); 1902 // the result from finish_pow_exp is now input to the phi node 1903 phi_node->init_req(2, result); 1904 set_control(_gvn.transform(region_node)); 1905 record_for_igvn(region_node); 1906 set_result(_gvn.transform(phi_node)); 1907 1908 C->set_has_split_ifs(true); // Has chance for split-if optimization 1909 return true; 1910 } 1911 1912 //------------------------------runtime_math----------------------------- 1913 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1914 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1915 "must be (DD)D or (D)D type"); 1916 1917 // Inputs 1918 Node* a = round_double_node(argument(0)); 1919 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1920 1921 const TypePtr* no_memory_effects = NULL; 1922 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1923 no_memory_effects, 1924 a, top(), b, b ? top() : NULL); 1925 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1926 #ifdef ASSERT 1927 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1928 assert(value_top == top(), "second value must be top"); 1929 #endif 1930 1931 set_result(value); 1932 return true; 1933 } 1934 1935 //------------------------------inline_math_native----------------------------- 1936 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1937 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1938 switch (id) { 1939 // These intrinsics are not properly supported on all hardware 1940 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1941 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1942 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1943 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1944 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1945 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1946 1947 case vmIntrinsics::_dlog: 1948 return StubRoutines::dlog() != NULL ? 1949 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1950 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1951 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1952 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1953 1954 // These intrinsics are supported on all hardware 1955 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1956 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1957 1958 case vmIntrinsics::_dexp: 1959 return StubRoutines::dexp() != NULL ? 1960 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1961 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1962 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1963 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1964 #undef FN_PTR 1965 1966 // These intrinsics are not yet correctly implemented 1967 case vmIntrinsics::_datan2: 1968 return false; 1969 1970 default: 1971 fatal_unexpected_iid(id); 1972 return false; 1973 } 1974 } 1975 1976 static bool is_simple_name(Node* n) { 1977 return (n->req() == 1 // constant 1978 || (n->is_Type() && n->as_Type()->type()->singleton()) 1979 || n->is_Proj() // parameter or return value 1980 || n->is_Phi() // local of some sort 1981 ); 1982 } 1983 1984 //----------------------------inline_notify-----------------------------------* 1985 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1986 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1987 address func; 1988 if (id == vmIntrinsics::_notify) { 1989 func = OptoRuntime::monitor_notify_Java(); 1990 } else { 1991 func = OptoRuntime::monitor_notifyAll_Java(); 1992 } 1993 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1994 make_slow_call_ex(call, env()->Throwable_klass(), false); 1995 return true; 1996 } 1997 1998 1999 //----------------------------inline_min_max----------------------------------- 2000 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2001 set_result(generate_min_max(id, argument(0), argument(1))); 2002 return true; 2003 } 2004 2005 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2006 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2007 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2008 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2009 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2010 2011 { 2012 PreserveJVMState pjvms(this); 2013 PreserveReexecuteState preexecs(this); 2014 jvms()->set_should_reexecute(true); 2015 2016 set_control(slow_path); 2017 set_i_o(i_o()); 2018 2019 uncommon_trap(Deoptimization::Reason_intrinsic, 2020 Deoptimization::Action_none); 2021 } 2022 2023 set_control(fast_path); 2024 set_result(math); 2025 } 2026 2027 template <typename OverflowOp> 2028 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2029 typedef typename OverflowOp::MathOp MathOp; 2030 2031 MathOp* mathOp = new MathOp(arg1, arg2); 2032 Node* operation = _gvn.transform( mathOp ); 2033 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2034 inline_math_mathExact(operation, ofcheck); 2035 return true; 2036 } 2037 2038 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2039 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2040 } 2041 2042 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2043 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2044 } 2045 2046 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2047 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2048 } 2049 2050 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2051 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2052 } 2053 2054 bool LibraryCallKit::inline_math_negateExactI() { 2055 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2056 } 2057 2058 bool LibraryCallKit::inline_math_negateExactL() { 2059 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2060 } 2061 2062 bool LibraryCallKit::inline_math_multiplyExactI() { 2063 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2064 } 2065 2066 bool LibraryCallKit::inline_math_multiplyExactL() { 2067 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2068 } 2069 2070 Node* 2071 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2072 // These are the candidate return value: 2073 Node* xvalue = x0; 2074 Node* yvalue = y0; 2075 2076 if (xvalue == yvalue) { 2077 return xvalue; 2078 } 2079 2080 bool want_max = (id == vmIntrinsics::_max); 2081 2082 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2083 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2084 if (txvalue == NULL || tyvalue == NULL) return top(); 2085 // This is not really necessary, but it is consistent with a 2086 // hypothetical MaxINode::Value method: 2087 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2088 2089 // %%% This folding logic should (ideally) be in a different place. 2090 // Some should be inside IfNode, and there to be a more reliable 2091 // transformation of ?: style patterns into cmoves. We also want 2092 // more powerful optimizations around cmove and min/max. 2093 2094 // Try to find a dominating comparison of these guys. 2095 // It can simplify the index computation for Arrays.copyOf 2096 // and similar uses of System.arraycopy. 2097 // First, compute the normalized version of CmpI(x, y). 2098 int cmp_op = Op_CmpI; 2099 Node* xkey = xvalue; 2100 Node* ykey = yvalue; 2101 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2102 if (ideal_cmpxy->is_Cmp()) { 2103 // E.g., if we have CmpI(length - offset, count), 2104 // it might idealize to CmpI(length, count + offset) 2105 cmp_op = ideal_cmpxy->Opcode(); 2106 xkey = ideal_cmpxy->in(1); 2107 ykey = ideal_cmpxy->in(2); 2108 } 2109 2110 // Start by locating any relevant comparisons. 2111 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2112 Node* cmpxy = NULL; 2113 Node* cmpyx = NULL; 2114 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2115 Node* cmp = start_from->fast_out(k); 2116 if (cmp->outcnt() > 0 && // must have prior uses 2117 cmp->in(0) == NULL && // must be context-independent 2118 cmp->Opcode() == cmp_op) { // right kind of compare 2119 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2120 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2121 } 2122 } 2123 2124 const int NCMPS = 2; 2125 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2126 int cmpn; 2127 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2128 if (cmps[cmpn] != NULL) break; // find a result 2129 } 2130 if (cmpn < NCMPS) { 2131 // Look for a dominating test that tells us the min and max. 2132 int depth = 0; // Limit search depth for speed 2133 Node* dom = control(); 2134 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2135 if (++depth >= 100) break; 2136 Node* ifproj = dom; 2137 if (!ifproj->is_Proj()) continue; 2138 Node* iff = ifproj->in(0); 2139 if (!iff->is_If()) continue; 2140 Node* bol = iff->in(1); 2141 if (!bol->is_Bool()) continue; 2142 Node* cmp = bol->in(1); 2143 if (cmp == NULL) continue; 2144 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2145 if (cmps[cmpn] == cmp) break; 2146 if (cmpn == NCMPS) continue; 2147 BoolTest::mask btest = bol->as_Bool()->_test._test; 2148 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2149 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2150 // At this point, we know that 'x btest y' is true. 2151 switch (btest) { 2152 case BoolTest::eq: 2153 // They are proven equal, so we can collapse the min/max. 2154 // Either value is the answer. Choose the simpler. 2155 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2156 return yvalue; 2157 return xvalue; 2158 case BoolTest::lt: // x < y 2159 case BoolTest::le: // x <= y 2160 return (want_max ? yvalue : xvalue); 2161 case BoolTest::gt: // x > y 2162 case BoolTest::ge: // x >= y 2163 return (want_max ? xvalue : yvalue); 2164 } 2165 } 2166 } 2167 2168 // We failed to find a dominating test. 2169 // Let's pick a test that might GVN with prior tests. 2170 Node* best_bol = NULL; 2171 BoolTest::mask best_btest = BoolTest::illegal; 2172 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2173 Node* cmp = cmps[cmpn]; 2174 if (cmp == NULL) continue; 2175 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2176 Node* bol = cmp->fast_out(j); 2177 if (!bol->is_Bool()) continue; 2178 BoolTest::mask btest = bol->as_Bool()->_test._test; 2179 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2180 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2181 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2182 best_bol = bol->as_Bool(); 2183 best_btest = btest; 2184 } 2185 } 2186 } 2187 2188 Node* answer_if_true = NULL; 2189 Node* answer_if_false = NULL; 2190 switch (best_btest) { 2191 default: 2192 if (cmpxy == NULL) 2193 cmpxy = ideal_cmpxy; 2194 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2195 // and fall through: 2196 case BoolTest::lt: // x < y 2197 case BoolTest::le: // x <= y 2198 answer_if_true = (want_max ? yvalue : xvalue); 2199 answer_if_false = (want_max ? xvalue : yvalue); 2200 break; 2201 case BoolTest::gt: // x > y 2202 case BoolTest::ge: // x >= y 2203 answer_if_true = (want_max ? xvalue : yvalue); 2204 answer_if_false = (want_max ? yvalue : xvalue); 2205 break; 2206 } 2207 2208 jint hi, lo; 2209 if (want_max) { 2210 // We can sharpen the minimum. 2211 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2212 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2213 } else { 2214 // We can sharpen the maximum. 2215 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2216 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2217 } 2218 2219 // Use a flow-free graph structure, to avoid creating excess control edges 2220 // which could hinder other optimizations. 2221 // Since Math.min/max is often used with arraycopy, we want 2222 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2223 Node* cmov = CMoveNode::make(NULL, best_bol, 2224 answer_if_false, answer_if_true, 2225 TypeInt::make(lo, hi, widen)); 2226 2227 return _gvn.transform(cmov); 2228 2229 /* 2230 // This is not as desirable as it may seem, since Min and Max 2231 // nodes do not have a full set of optimizations. 2232 // And they would interfere, anyway, with 'if' optimizations 2233 // and with CMoveI canonical forms. 2234 switch (id) { 2235 case vmIntrinsics::_min: 2236 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2237 case vmIntrinsics::_max: 2238 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2239 default: 2240 ShouldNotReachHere(); 2241 } 2242 */ 2243 } 2244 2245 inline int 2246 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2247 const TypePtr* base_type = TypePtr::NULL_PTR; 2248 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2249 if (base_type == NULL) { 2250 // Unknown type. 2251 return Type::AnyPtr; 2252 } else if (base_type == TypePtr::NULL_PTR) { 2253 // Since this is a NULL+long form, we have to switch to a rawptr. 2254 base = _gvn.transform(new CastX2PNode(offset)); 2255 offset = MakeConX(0); 2256 return Type::RawPtr; 2257 } else if (base_type->base() == Type::RawPtr) { 2258 return Type::RawPtr; 2259 } else if (base_type->isa_oopptr()) { 2260 // Base is never null => always a heap address. 2261 if (base_type->ptr() == TypePtr::NotNull) { 2262 return Type::OopPtr; 2263 } 2264 // Offset is small => always a heap address. 2265 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2266 if (offset_type != NULL && 2267 base_type->offset() == 0 && // (should always be?) 2268 offset_type->_lo >= 0 && 2269 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2270 return Type::OopPtr; 2271 } 2272 // Otherwise, it might either be oop+off or NULL+addr. 2273 return Type::AnyPtr; 2274 } else { 2275 // No information: 2276 return Type::AnyPtr; 2277 } 2278 } 2279 2280 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2281 int kind = classify_unsafe_addr(base, offset); 2282 if (kind == Type::RawPtr) { 2283 return basic_plus_adr(top(), base, offset); 2284 } else { 2285 return basic_plus_adr(base, offset); 2286 } 2287 } 2288 2289 //--------------------------inline_number_methods----------------------------- 2290 // inline int Integer.numberOfLeadingZeros(int) 2291 // inline int Long.numberOfLeadingZeros(long) 2292 // 2293 // inline int Integer.numberOfTrailingZeros(int) 2294 // inline int Long.numberOfTrailingZeros(long) 2295 // 2296 // inline int Integer.bitCount(int) 2297 // inline int Long.bitCount(long) 2298 // 2299 // inline char Character.reverseBytes(char) 2300 // inline short Short.reverseBytes(short) 2301 // inline int Integer.reverseBytes(int) 2302 // inline long Long.reverseBytes(long) 2303 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2304 Node* arg = argument(0); 2305 Node* n; 2306 switch (id) { 2307 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2308 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2309 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2310 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2311 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2312 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2313 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2314 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2315 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2316 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2317 default: fatal_unexpected_iid(id); break; 2318 } 2319 set_result(_gvn.transform(n)); 2320 return true; 2321 } 2322 2323 //----------------------------inline_unsafe_access---------------------------- 2324 2325 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2326 2327 // Helper that guards and inserts a pre-barrier. 2328 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2329 Node* pre_val, bool need_mem_bar) { 2330 // We could be accessing the referent field of a reference object. If so, when G1 2331 // is enabled, we need to log the value in the referent field in an SATB buffer. 2332 // This routine performs some compile time filters and generates suitable 2333 // runtime filters that guard the pre-barrier code. 2334 // Also add memory barrier for non volatile load from the referent field 2335 // to prevent commoning of loads across safepoint. 2336 if (!UseG1GC && !need_mem_bar) 2337 return; 2338 2339 // Some compile time checks. 2340 2341 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2342 const TypeX* otype = offset->find_intptr_t_type(); 2343 if (otype != NULL && otype->is_con() && 2344 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2345 // Constant offset but not the reference_offset so just return 2346 return; 2347 } 2348 2349 // We only need to generate the runtime guards for instances. 2350 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2351 if (btype != NULL) { 2352 if (btype->isa_aryptr()) { 2353 // Array type so nothing to do 2354 return; 2355 } 2356 2357 const TypeInstPtr* itype = btype->isa_instptr(); 2358 if (itype != NULL) { 2359 // Can the klass of base_oop be statically determined to be 2360 // _not_ a sub-class of Reference and _not_ Object? 2361 ciKlass* klass = itype->klass(); 2362 if ( klass->is_loaded() && 2363 !klass->is_subtype_of(env()->Reference_klass()) && 2364 !env()->Object_klass()->is_subtype_of(klass)) { 2365 return; 2366 } 2367 } 2368 } 2369 2370 // The compile time filters did not reject base_oop/offset so 2371 // we need to generate the following runtime filters 2372 // 2373 // if (offset == java_lang_ref_Reference::_reference_offset) { 2374 // if (instance_of(base, java.lang.ref.Reference)) { 2375 // pre_barrier(_, pre_val, ...); 2376 // } 2377 // } 2378 2379 float likely = PROB_LIKELY( 0.999); 2380 float unlikely = PROB_UNLIKELY(0.999); 2381 2382 IdealKit ideal(this); 2383 #define __ ideal. 2384 2385 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2386 2387 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2388 // Update graphKit memory and control from IdealKit. 2389 sync_kit(ideal); 2390 2391 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2392 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2393 2394 // Update IdealKit memory and control from graphKit. 2395 __ sync_kit(this); 2396 2397 Node* one = __ ConI(1); 2398 // is_instof == 0 if base_oop == NULL 2399 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2400 2401 // Update graphKit from IdeakKit. 2402 sync_kit(ideal); 2403 2404 // Use the pre-barrier to record the value in the referent field 2405 pre_barrier(false /* do_load */, 2406 __ ctrl(), 2407 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2408 pre_val /* pre_val */, 2409 T_OBJECT); 2410 if (need_mem_bar) { 2411 // Add memory barrier to prevent commoning reads from this field 2412 // across safepoint since GC can change its value. 2413 insert_mem_bar(Op_MemBarCPUOrder); 2414 } 2415 // Update IdealKit from graphKit. 2416 __ sync_kit(this); 2417 2418 } __ end_if(); // _ref_type != ref_none 2419 } __ end_if(); // offset == referent_offset 2420 2421 // Final sync IdealKit and GraphKit. 2422 final_sync(ideal); 2423 #undef __ 2424 } 2425 2426 2427 // Interpret Unsafe.fieldOffset cookies correctly: 2428 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2429 2430 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2431 // Attempt to infer a sharper value type from the offset and base type. 2432 ciKlass* sharpened_klass = NULL; 2433 2434 // See if it is an instance field, with an object type. 2435 if (alias_type->field() != NULL) { 2436 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2437 if (alias_type->field()->type()->is_klass()) { 2438 sharpened_klass = alias_type->field()->type()->as_klass(); 2439 } 2440 } 2441 2442 // See if it is a narrow oop array. 2443 if (adr_type->isa_aryptr()) { 2444 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2445 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2446 if (elem_type != NULL) { 2447 sharpened_klass = elem_type->klass(); 2448 } 2449 } 2450 } 2451 2452 // The sharpened class might be unloaded if there is no class loader 2453 // contraint in place. 2454 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2455 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2456 2457 #ifndef PRODUCT 2458 if (C->print_intrinsics() || C->print_inlining()) { 2459 tty->print(" from base type: "); adr_type->dump(); 2460 tty->print(" sharpened value: "); tjp->dump(); 2461 } 2462 #endif 2463 // Sharpen the value type. 2464 return tjp; 2465 } 2466 return NULL; 2467 } 2468 2469 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) { 2470 if (callee()->is_static()) return false; // caller must have the capability! 2471 2472 #ifndef PRODUCT 2473 { 2474 ResourceMark rm; 2475 // Check the signatures. 2476 ciSignature* sig = callee()->signature(); 2477 #ifdef ASSERT 2478 if (!is_store) { 2479 // Object getObject(Object base, int/long offset), etc. 2480 BasicType rtype = sig->return_type()->basic_type(); 2481 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2482 rtype = T_ADDRESS; // it is really a C void* 2483 assert(rtype == type, "getter must return the expected value"); 2484 if (!is_native_ptr) { 2485 assert(sig->count() == 2, "oop getter has 2 arguments"); 2486 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2487 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2488 } else { 2489 assert(sig->count() == 1, "native getter has 1 argument"); 2490 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2491 } 2492 } else { 2493 // void putObject(Object base, int/long offset, Object x), etc. 2494 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2495 if (!is_native_ptr) { 2496 assert(sig->count() == 3, "oop putter has 3 arguments"); 2497 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2498 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2499 } else { 2500 assert(sig->count() == 2, "native putter has 2 arguments"); 2501 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2502 } 2503 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2504 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2505 vtype = T_ADDRESS; // it is really a C void* 2506 assert(vtype == type, "putter must accept the expected value"); 2507 } 2508 #endif // ASSERT 2509 } 2510 #endif //PRODUCT 2511 2512 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2513 2514 Node* receiver = argument(0); // type: oop 2515 2516 // Build address expression. 2517 Node* adr; 2518 Node* heap_base_oop = top(); 2519 Node* offset = top(); 2520 Node* val; 2521 2522 if (!is_native_ptr) { 2523 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2524 Node* base = argument(1); // type: oop 2525 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2526 offset = argument(2); // type: long 2527 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2528 // to be plain byte offsets, which are also the same as those accepted 2529 // by oopDesc::field_base. 2530 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2531 "fieldOffset must be byte-scaled"); 2532 // 32-bit machines ignore the high half! 2533 offset = ConvL2X(offset); 2534 adr = make_unsafe_address(base, offset); 2535 heap_base_oop = base; 2536 val = is_store ? argument(4) : NULL; 2537 } else { 2538 Node* ptr = argument(1); // type: long 2539 ptr = ConvL2X(ptr); // adjust Java long to machine word 2540 adr = make_unsafe_address(NULL, ptr); 2541 val = is_store ? argument(3) : NULL; 2542 } 2543 2544 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2545 2546 // First guess at the value type. 2547 const Type *value_type = Type::get_const_basic_type(type); 2548 2549 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2550 // there was not enough information to nail it down. 2551 Compile::AliasType* alias_type = C->alias_type(adr_type); 2552 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2553 2554 // We will need memory barriers unless we can determine a unique 2555 // alias category for this reference. (Note: If for some reason 2556 // the barriers get omitted and the unsafe reference begins to "pollute" 2557 // the alias analysis of the rest of the graph, either Compile::can_alias 2558 // or Compile::must_alias will throw a diagnostic assert.) 2559 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2560 2561 // If we are reading the value of the referent field of a Reference 2562 // object (either by using Unsafe directly or through reflection) 2563 // then, if G1 is enabled, we need to record the referent in an 2564 // SATB log buffer using the pre-barrier mechanism. 2565 // Also we need to add memory barrier to prevent commoning reads 2566 // from this field across safepoint since GC can change its value. 2567 bool need_read_barrier = !is_native_ptr && !is_store && 2568 offset != top() && heap_base_oop != top(); 2569 2570 if (!is_store && type == T_OBJECT) { 2571 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2572 if (tjp != NULL) { 2573 value_type = tjp; 2574 } 2575 } 2576 2577 receiver = null_check(receiver); 2578 if (stopped()) { 2579 return true; 2580 } 2581 // Heap pointers get a null-check from the interpreter, 2582 // as a courtesy. However, this is not guaranteed by Unsafe, 2583 // and it is not possible to fully distinguish unintended nulls 2584 // from intended ones in this API. 2585 2586 if (is_volatile) { 2587 // We need to emit leading and trailing CPU membars (see below) in 2588 // addition to memory membars when is_volatile. This is a little 2589 // too strong, but avoids the need to insert per-alias-type 2590 // volatile membars (for stores; compare Parse::do_put_xxx), which 2591 // we cannot do effectively here because we probably only have a 2592 // rough approximation of type. 2593 need_mem_bar = true; 2594 // For Stores, place a memory ordering barrier now. 2595 if (is_store) { 2596 insert_mem_bar(Op_MemBarRelease); 2597 } else { 2598 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2599 insert_mem_bar(Op_MemBarVolatile); 2600 } 2601 } 2602 } 2603 2604 // Memory barrier to prevent normal and 'unsafe' accesses from 2605 // bypassing each other. Happens after null checks, so the 2606 // exception paths do not take memory state from the memory barrier, 2607 // so there's no problems making a strong assert about mixing users 2608 // of safe & unsafe memory. 2609 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2610 2611 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM || 2612 alias_type->field() != NULL || alias_type->element() != NULL, "field, array element or unknown"); 2613 bool mismatched = false; 2614 if (alias_type->element() != NULL || alias_type->field() != NULL) { 2615 BasicType bt; 2616 if (alias_type->element() != NULL) { 2617 const Type* element = alias_type->element(); 2618 bt = element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 2619 } else { 2620 bt = alias_type->field()->type()->basic_type(); 2621 } 2622 if (bt != type) { 2623 mismatched = true; 2624 } 2625 } 2626 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2627 2628 if (!is_store) { 2629 Node* p = NULL; 2630 // Try to constant fold a load from a constant field 2631 ciField* field = alias_type->field(); 2632 if (heap_base_oop != top() && 2633 field != NULL && field->is_constant() && field->layout_type() == type) { 2634 // final or stable field 2635 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop); 2636 if (con_type != NULL) { 2637 p = makecon(con_type); 2638 } 2639 } 2640 if (p == NULL) { 2641 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2642 // To be valid, unsafe loads may depend on other conditions than 2643 // the one that guards them: pin the Load node 2644 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched); 2645 // load value 2646 switch (type) { 2647 case T_BOOLEAN: 2648 case T_CHAR: 2649 case T_BYTE: 2650 case T_SHORT: 2651 case T_INT: 2652 case T_LONG: 2653 case T_FLOAT: 2654 case T_DOUBLE: 2655 break; 2656 case T_OBJECT: 2657 if (need_read_barrier) { 2658 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2659 } 2660 break; 2661 case T_ADDRESS: 2662 // Cast to an int type. 2663 p = _gvn.transform(new CastP2XNode(NULL, p)); 2664 p = ConvX2UL(p); 2665 break; 2666 default: 2667 fatal("unexpected type %d: %s", type, type2name(type)); 2668 break; 2669 } 2670 } 2671 // The load node has the control of the preceding MemBarCPUOrder. All 2672 // following nodes will have the control of the MemBarCPUOrder inserted at 2673 // the end of this method. So, pushing the load onto the stack at a later 2674 // point is fine. 2675 set_result(p); 2676 } else { 2677 // place effect of store into memory 2678 switch (type) { 2679 case T_DOUBLE: 2680 val = dstore_rounding(val); 2681 break; 2682 case T_ADDRESS: 2683 // Repackage the long as a pointer. 2684 val = ConvL2X(val); 2685 val = _gvn.transform(new CastX2PNode(val)); 2686 break; 2687 } 2688 2689 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2690 if (type != T_OBJECT ) { 2691 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched); 2692 } else { 2693 // Possibly an oop being stored to Java heap or native memory 2694 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2695 // oop to Java heap. 2696 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2697 } else { 2698 // We can't tell at compile time if we are storing in the Java heap or outside 2699 // of it. So we need to emit code to conditionally do the proper type of 2700 // store. 2701 2702 IdealKit ideal(this); 2703 #define __ ideal. 2704 // QQQ who knows what probability is here?? 2705 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2706 // Sync IdealKit and graphKit. 2707 sync_kit(ideal); 2708 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2709 // Update IdealKit memory. 2710 __ sync_kit(this); 2711 } __ else_(); { 2712 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile, mismatched); 2713 } __ end_if(); 2714 // Final sync IdealKit and GraphKit. 2715 final_sync(ideal); 2716 #undef __ 2717 } 2718 } 2719 } 2720 2721 if (is_volatile) { 2722 if (!is_store) { 2723 insert_mem_bar(Op_MemBarAcquire); 2724 } else { 2725 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2726 insert_mem_bar(Op_MemBarVolatile); 2727 } 2728 } 2729 } 2730 2731 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2732 2733 return true; 2734 } 2735 2736 //----------------------------inline_unsafe_load_store---------------------------- 2737 // This method serves a couple of different customers (depending on LoadStoreKind): 2738 // 2739 // LS_cmpxchg: 2740 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2741 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2742 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2743 // 2744 // LS_xadd: 2745 // public int getAndAddInt( Object o, long offset, int delta) 2746 // public long getAndAddLong(Object o, long offset, long delta) 2747 // 2748 // LS_xchg: 2749 // int getAndSet(Object o, long offset, int newValue) 2750 // long getAndSet(Object o, long offset, long newValue) 2751 // Object getAndSet(Object o, long offset, Object newValue) 2752 // 2753 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2754 // This basic scheme here is the same as inline_unsafe_access, but 2755 // differs in enough details that combining them would make the code 2756 // overly confusing. (This is a true fact! I originally combined 2757 // them, but even I was confused by it!) As much code/comments as 2758 // possible are retained from inline_unsafe_access though to make 2759 // the correspondences clearer. - dl 2760 2761 if (callee()->is_static()) return false; // caller must have the capability! 2762 2763 #ifndef PRODUCT 2764 BasicType rtype; 2765 { 2766 ResourceMark rm; 2767 // Check the signatures. 2768 ciSignature* sig = callee()->signature(); 2769 rtype = sig->return_type()->basic_type(); 2770 if (kind == LS_xadd || kind == LS_xchg) { 2771 // Check the signatures. 2772 #ifdef ASSERT 2773 assert(rtype == type, "get and set must return the expected type"); 2774 assert(sig->count() == 3, "get and set has 3 arguments"); 2775 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2776 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2777 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2778 #endif // ASSERT 2779 } else if (kind == LS_cmpxchg) { 2780 // Check the signatures. 2781 #ifdef ASSERT 2782 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2783 assert(sig->count() == 4, "CAS has 4 arguments"); 2784 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2785 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2786 #endif // ASSERT 2787 } else { 2788 ShouldNotReachHere(); 2789 } 2790 } 2791 #endif //PRODUCT 2792 2793 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2794 2795 // Get arguments: 2796 Node* receiver = NULL; 2797 Node* base = NULL; 2798 Node* offset = NULL; 2799 Node* oldval = NULL; 2800 Node* newval = NULL; 2801 if (kind == LS_cmpxchg) { 2802 const bool two_slot_type = type2size[type] == 2; 2803 receiver = argument(0); // type: oop 2804 base = argument(1); // type: oop 2805 offset = argument(2); // type: long 2806 oldval = argument(4); // type: oop, int, or long 2807 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2808 } else if (kind == LS_xadd || kind == LS_xchg){ 2809 receiver = argument(0); // type: oop 2810 base = argument(1); // type: oop 2811 offset = argument(2); // type: long 2812 oldval = NULL; 2813 newval = argument(4); // type: oop, int, or long 2814 } 2815 2816 // Null check receiver. 2817 receiver = null_check(receiver); 2818 if (stopped()) { 2819 return true; 2820 } 2821 2822 // Build field offset expression. 2823 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2824 // to be plain byte offsets, which are also the same as those accepted 2825 // by oopDesc::field_base. 2826 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2827 // 32-bit machines ignore the high half of long offsets 2828 offset = ConvL2X(offset); 2829 Node* adr = make_unsafe_address(base, offset); 2830 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2831 2832 // For CAS, unlike inline_unsafe_access, there seems no point in 2833 // trying to refine types. Just use the coarse types here. 2834 const Type *value_type = Type::get_const_basic_type(type); 2835 Compile::AliasType* alias_type = C->alias_type(adr_type); 2836 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2837 2838 if (kind == LS_xchg && type == T_OBJECT) { 2839 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2840 if (tjp != NULL) { 2841 value_type = tjp; 2842 } 2843 } 2844 2845 int alias_idx = C->get_alias_index(adr_type); 2846 2847 // Memory-model-wise, a LoadStore acts like a little synchronized 2848 // block, so needs barriers on each side. These don't translate 2849 // into actual barriers on most machines, but we still need rest of 2850 // compiler to respect ordering. 2851 2852 insert_mem_bar(Op_MemBarRelease); 2853 insert_mem_bar(Op_MemBarCPUOrder); 2854 2855 // 4984716: MemBars must be inserted before this 2856 // memory node in order to avoid a false 2857 // dependency which will confuse the scheduler. 2858 Node *mem = memory(alias_idx); 2859 2860 // For now, we handle only those cases that actually exist: ints, 2861 // longs, and Object. Adding others should be straightforward. 2862 Node* load_store; 2863 switch(type) { 2864 case T_INT: 2865 if (kind == LS_xadd) { 2866 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2867 } else if (kind == LS_xchg) { 2868 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2869 } else if (kind == LS_cmpxchg) { 2870 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2871 } else { 2872 ShouldNotReachHere(); 2873 } 2874 break; 2875 case T_LONG: 2876 if (kind == LS_xadd) { 2877 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2878 } else if (kind == LS_xchg) { 2879 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2880 } else if (kind == LS_cmpxchg) { 2881 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2882 } else { 2883 ShouldNotReachHere(); 2884 } 2885 break; 2886 case T_OBJECT: 2887 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2888 // could be delayed during Parse (for example, in adjust_map_after_if()). 2889 // Execute transformation here to avoid barrier generation in such case. 2890 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2891 newval = _gvn.makecon(TypePtr::NULL_PTR); 2892 2893 // Reference stores need a store barrier. 2894 if (kind == LS_xchg) { 2895 // If pre-barrier must execute before the oop store, old value will require do_load here. 2896 if (!can_move_pre_barrier()) { 2897 pre_barrier(true /* do_load*/, 2898 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2899 NULL /* pre_val*/, 2900 T_OBJECT); 2901 } // Else move pre_barrier to use load_store value, see below. 2902 } else if (kind == LS_cmpxchg) { 2903 // Same as for newval above: 2904 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2905 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2906 } 2907 // The only known value which might get overwritten is oldval. 2908 pre_barrier(false /* do_load */, 2909 control(), NULL, NULL, max_juint, NULL, NULL, 2910 oldval /* pre_val */, 2911 T_OBJECT); 2912 } else { 2913 ShouldNotReachHere(); 2914 } 2915 2916 #ifdef _LP64 2917 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2918 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2919 if (kind == LS_xchg) { 2920 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2921 newval_enc, adr_type, value_type->make_narrowoop())); 2922 } else { 2923 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2924 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2925 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2926 newval_enc, oldval_enc)); 2927 } 2928 } else 2929 #endif 2930 { 2931 if (kind == LS_xchg) { 2932 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2933 } else { 2934 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2935 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2936 } 2937 } 2938 if (kind == LS_cmpxchg) { 2939 // Emit the post barrier only when the actual store happened. 2940 // This makes sense to check only for compareAndSet that can fail to set the value. 2941 // CAS success path is marked more likely since we anticipate this is a performance 2942 // critical path, while CAS failure path can use the penalty for going through unlikely 2943 // path as backoff. Which is still better than doing a store barrier there. 2944 IdealKit ideal(this); 2945 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); { 2946 sync_kit(ideal); 2947 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2948 ideal.sync_kit(this); 2949 } ideal.end_if(); 2950 final_sync(ideal); 2951 } else { 2952 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2953 } 2954 break; 2955 default: 2956 fatal("unexpected type %d: %s", type, type2name(type)); 2957 break; 2958 } 2959 2960 // SCMemProjNodes represent the memory state of a LoadStore. Their 2961 // main role is to prevent LoadStore nodes from being optimized away 2962 // when their results aren't used. 2963 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2964 set_memory(proj, alias_idx); 2965 2966 if (type == T_OBJECT && kind == LS_xchg) { 2967 #ifdef _LP64 2968 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2969 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2970 } 2971 #endif 2972 if (can_move_pre_barrier()) { 2973 // Don't need to load pre_val. The old value is returned by load_store. 2974 // The pre_barrier can execute after the xchg as long as no safepoint 2975 // gets inserted between them. 2976 pre_barrier(false /* do_load */, 2977 control(), NULL, NULL, max_juint, NULL, NULL, 2978 load_store /* pre_val */, 2979 T_OBJECT); 2980 } 2981 } 2982 2983 // Add the trailing membar surrounding the access 2984 insert_mem_bar(Op_MemBarCPUOrder); 2985 insert_mem_bar(Op_MemBarAcquire); 2986 2987 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2988 set_result(load_store); 2989 return true; 2990 } 2991 2992 //----------------------------inline_unsafe_ordered_store---------------------- 2993 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 2994 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 2995 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 2996 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2997 // This is another variant of inline_unsafe_access, differing in 2998 // that it always issues store-store ("release") barrier and ensures 2999 // store-atomicity (which only matters for "long"). 3000 3001 if (callee()->is_static()) return false; // caller must have the capability! 3002 3003 #ifndef PRODUCT 3004 { 3005 ResourceMark rm; 3006 // Check the signatures. 3007 ciSignature* sig = callee()->signature(); 3008 #ifdef ASSERT 3009 BasicType rtype = sig->return_type()->basic_type(); 3010 assert(rtype == T_VOID, "must return void"); 3011 assert(sig->count() == 3, "has 3 arguments"); 3012 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3013 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3014 #endif // ASSERT 3015 } 3016 #endif //PRODUCT 3017 3018 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3019 3020 // Get arguments: 3021 Node* receiver = argument(0); // type: oop 3022 Node* base = argument(1); // type: oop 3023 Node* offset = argument(2); // type: long 3024 Node* val = argument(4); // type: oop, int, or long 3025 3026 // Null check receiver. 3027 receiver = null_check(receiver); 3028 if (stopped()) { 3029 return true; 3030 } 3031 3032 // Build field offset expression. 3033 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3034 // 32-bit machines ignore the high half of long offsets 3035 offset = ConvL2X(offset); 3036 Node* adr = make_unsafe_address(base, offset); 3037 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3038 const Type *value_type = Type::get_const_basic_type(type); 3039 Compile::AliasType* alias_type = C->alias_type(adr_type); 3040 3041 insert_mem_bar(Op_MemBarRelease); 3042 insert_mem_bar(Op_MemBarCPUOrder); 3043 // Ensure that the store is atomic for longs: 3044 const bool require_atomic_access = true; 3045 Node* store; 3046 if (type == T_OBJECT) // reference stores need a store barrier. 3047 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3048 else { 3049 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3050 } 3051 insert_mem_bar(Op_MemBarCPUOrder); 3052 return true; 3053 } 3054 3055 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3056 // Regardless of form, don't allow previous ld/st to move down, 3057 // then issue acquire, release, or volatile mem_bar. 3058 insert_mem_bar(Op_MemBarCPUOrder); 3059 switch(id) { 3060 case vmIntrinsics::_loadFence: 3061 insert_mem_bar(Op_LoadFence); 3062 return true; 3063 case vmIntrinsics::_storeFence: 3064 insert_mem_bar(Op_StoreFence); 3065 return true; 3066 case vmIntrinsics::_fullFence: 3067 insert_mem_bar(Op_MemBarVolatile); 3068 return true; 3069 default: 3070 fatal_unexpected_iid(id); 3071 return false; 3072 } 3073 } 3074 3075 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3076 if (!kls->is_Con()) { 3077 return true; 3078 } 3079 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3080 if (klsptr == NULL) { 3081 return true; 3082 } 3083 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3084 // don't need a guard for a klass that is already initialized 3085 return !ik->is_initialized(); 3086 } 3087 3088 //----------------------------inline_unsafe_allocate--------------------------- 3089 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3090 bool LibraryCallKit::inline_unsafe_allocate() { 3091 if (callee()->is_static()) return false; // caller must have the capability! 3092 3093 null_check_receiver(); // null-check, then ignore 3094 Node* cls = null_check(argument(1)); 3095 if (stopped()) return true; 3096 3097 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3098 kls = null_check(kls); 3099 if (stopped()) return true; // argument was like int.class 3100 3101 Node* test = NULL; 3102 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3103 // Note: The argument might still be an illegal value like 3104 // Serializable.class or Object[].class. The runtime will handle it. 3105 // But we must make an explicit check for initialization. 3106 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3107 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3108 // can generate code to load it as unsigned byte. 3109 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3110 Node* bits = intcon(InstanceKlass::fully_initialized); 3111 test = _gvn.transform(new SubINode(inst, bits)); 3112 // The 'test' is non-zero if we need to take a slow path. 3113 } 3114 3115 Node* obj = new_instance(kls, test); 3116 set_result(obj); 3117 return true; 3118 } 3119 3120 #ifdef TRACE_HAVE_INTRINSICS 3121 /* 3122 * oop -> myklass 3123 * myklass->trace_id |= USED 3124 * return myklass->trace_id & ~0x3 3125 */ 3126 bool LibraryCallKit::inline_native_classID() { 3127 null_check_receiver(); // null-check, then ignore 3128 Node* cls = null_check(argument(1), T_OBJECT); 3129 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3130 kls = null_check(kls, T_OBJECT); 3131 ByteSize offset = TRACE_ID_OFFSET; 3132 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3133 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3134 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3135 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3136 Node* clsused = longcon(0x01l); // set the class bit 3137 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3138 3139 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3140 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3141 set_result(andl); 3142 return true; 3143 } 3144 3145 bool LibraryCallKit::inline_native_threadID() { 3146 Node* tls_ptr = NULL; 3147 Node* cur_thr = generate_current_thread(tls_ptr); 3148 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3149 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3150 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3151 3152 Node* threadid = NULL; 3153 size_t thread_id_size = OSThread::thread_id_size(); 3154 if (thread_id_size == (size_t) BytesPerLong) { 3155 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3156 } else if (thread_id_size == (size_t) BytesPerInt) { 3157 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3158 } else { 3159 ShouldNotReachHere(); 3160 } 3161 set_result(threadid); 3162 return true; 3163 } 3164 #endif 3165 3166 //------------------------inline_native_time_funcs-------------- 3167 // inline code for System.currentTimeMillis() and System.nanoTime() 3168 // these have the same type and signature 3169 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3170 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3171 const TypePtr* no_memory_effects = NULL; 3172 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3173 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3174 #ifdef ASSERT 3175 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3176 assert(value_top == top(), "second value must be top"); 3177 #endif 3178 set_result(value); 3179 return true; 3180 } 3181 3182 //------------------------inline_native_currentThread------------------ 3183 bool LibraryCallKit::inline_native_currentThread() { 3184 Node* junk = NULL; 3185 set_result(generate_current_thread(junk)); 3186 return true; 3187 } 3188 3189 //------------------------inline_native_isInterrupted------------------ 3190 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3191 bool LibraryCallKit::inline_native_isInterrupted() { 3192 // Add a fast path to t.isInterrupted(clear_int): 3193 // (t == Thread.current() && 3194 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3195 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3196 // So, in the common case that the interrupt bit is false, 3197 // we avoid making a call into the VM. Even if the interrupt bit 3198 // is true, if the clear_int argument is false, we avoid the VM call. 3199 // However, if the receiver is not currentThread, we must call the VM, 3200 // because there must be some locking done around the operation. 3201 3202 // We only go to the fast case code if we pass two guards. 3203 // Paths which do not pass are accumulated in the slow_region. 3204 3205 enum { 3206 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3207 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3208 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3209 PATH_LIMIT 3210 }; 3211 3212 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3213 // out of the function. 3214 insert_mem_bar(Op_MemBarCPUOrder); 3215 3216 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3217 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3218 3219 RegionNode* slow_region = new RegionNode(1); 3220 record_for_igvn(slow_region); 3221 3222 // (a) Receiving thread must be the current thread. 3223 Node* rec_thr = argument(0); 3224 Node* tls_ptr = NULL; 3225 Node* cur_thr = generate_current_thread(tls_ptr); 3226 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3227 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3228 3229 generate_slow_guard(bol_thr, slow_region); 3230 3231 // (b) Interrupt bit on TLS must be false. 3232 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3233 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3234 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3235 3236 // Set the control input on the field _interrupted read to prevent it floating up. 3237 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3238 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3239 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3240 3241 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3242 3243 // First fast path: if (!TLS._interrupted) return false; 3244 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3245 result_rgn->init_req(no_int_result_path, false_bit); 3246 result_val->init_req(no_int_result_path, intcon(0)); 3247 3248 // drop through to next case 3249 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3250 3251 #ifndef TARGET_OS_FAMILY_windows 3252 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3253 Node* clr_arg = argument(1); 3254 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3255 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3256 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3257 3258 // Second fast path: ... else if (!clear_int) return true; 3259 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3260 result_rgn->init_req(no_clear_result_path, false_arg); 3261 result_val->init_req(no_clear_result_path, intcon(1)); 3262 3263 // drop through to next case 3264 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3265 #else 3266 // To return true on Windows you must read the _interrupted field 3267 // and check the the event state i.e. take the slow path. 3268 #endif // TARGET_OS_FAMILY_windows 3269 3270 // (d) Otherwise, go to the slow path. 3271 slow_region->add_req(control()); 3272 set_control( _gvn.transform(slow_region)); 3273 3274 if (stopped()) { 3275 // There is no slow path. 3276 result_rgn->init_req(slow_result_path, top()); 3277 result_val->init_req(slow_result_path, top()); 3278 } else { 3279 // non-virtual because it is a private non-static 3280 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3281 3282 Node* slow_val = set_results_for_java_call(slow_call); 3283 // this->control() comes from set_results_for_java_call 3284 3285 Node* fast_io = slow_call->in(TypeFunc::I_O); 3286 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3287 3288 // These two phis are pre-filled with copies of of the fast IO and Memory 3289 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3290 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3291 3292 result_rgn->init_req(slow_result_path, control()); 3293 result_io ->init_req(slow_result_path, i_o()); 3294 result_mem->init_req(slow_result_path, reset_memory()); 3295 result_val->init_req(slow_result_path, slow_val); 3296 3297 set_all_memory(_gvn.transform(result_mem)); 3298 set_i_o( _gvn.transform(result_io)); 3299 } 3300 3301 C->set_has_split_ifs(true); // Has chance for split-if optimization 3302 set_result(result_rgn, result_val); 3303 return true; 3304 } 3305 3306 //---------------------------load_mirror_from_klass---------------------------- 3307 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3308 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3309 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3310 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3311 } 3312 3313 //-----------------------load_klass_from_mirror_common------------------------- 3314 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3315 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3316 // and branch to the given path on the region. 3317 // If never_see_null, take an uncommon trap on null, so we can optimistically 3318 // compile for the non-null case. 3319 // If the region is NULL, force never_see_null = true. 3320 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3321 bool never_see_null, 3322 RegionNode* region, 3323 int null_path, 3324 int offset) { 3325 if (region == NULL) never_see_null = true; 3326 Node* p = basic_plus_adr(mirror, offset); 3327 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3328 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3329 Node* null_ctl = top(); 3330 kls = null_check_oop(kls, &null_ctl, never_see_null); 3331 if (region != NULL) { 3332 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3333 region->init_req(null_path, null_ctl); 3334 } else { 3335 assert(null_ctl == top(), "no loose ends"); 3336 } 3337 return kls; 3338 } 3339 3340 //--------------------(inline_native_Class_query helpers)--------------------- 3341 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3342 // Fall through if (mods & mask) == bits, take the guard otherwise. 3343 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3344 // Branch around if the given klass has the given modifier bit set. 3345 // Like generate_guard, adds a new path onto the region. 3346 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3347 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3348 Node* mask = intcon(modifier_mask); 3349 Node* bits = intcon(modifier_bits); 3350 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3351 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3352 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3353 return generate_fair_guard(bol, region); 3354 } 3355 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3356 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3357 } 3358 3359 //-------------------------inline_native_Class_query------------------- 3360 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3361 const Type* return_type = TypeInt::BOOL; 3362 Node* prim_return_value = top(); // what happens if it's a primitive class? 3363 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3364 bool expect_prim = false; // most of these guys expect to work on refs 3365 3366 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3367 3368 Node* mirror = argument(0); 3369 Node* obj = top(); 3370 3371 switch (id) { 3372 case vmIntrinsics::_isInstance: 3373 // nothing is an instance of a primitive type 3374 prim_return_value = intcon(0); 3375 obj = argument(1); 3376 break; 3377 case vmIntrinsics::_getModifiers: 3378 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3379 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3380 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3381 break; 3382 case vmIntrinsics::_isInterface: 3383 prim_return_value = intcon(0); 3384 break; 3385 case vmIntrinsics::_isArray: 3386 prim_return_value = intcon(0); 3387 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3388 break; 3389 case vmIntrinsics::_isPrimitive: 3390 prim_return_value = intcon(1); 3391 expect_prim = true; // obviously 3392 break; 3393 case vmIntrinsics::_getSuperclass: 3394 prim_return_value = null(); 3395 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3396 break; 3397 case vmIntrinsics::_getClassAccessFlags: 3398 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3399 return_type = TypeInt::INT; // not bool! 6297094 3400 break; 3401 default: 3402 fatal_unexpected_iid(id); 3403 break; 3404 } 3405 3406 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3407 if (mirror_con == NULL) return false; // cannot happen? 3408 3409 #ifndef PRODUCT 3410 if (C->print_intrinsics() || C->print_inlining()) { 3411 ciType* k = mirror_con->java_mirror_type(); 3412 if (k) { 3413 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3414 k->print_name(); 3415 tty->cr(); 3416 } 3417 } 3418 #endif 3419 3420 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3421 RegionNode* region = new RegionNode(PATH_LIMIT); 3422 record_for_igvn(region); 3423 PhiNode* phi = new PhiNode(region, return_type); 3424 3425 // The mirror will never be null of Reflection.getClassAccessFlags, however 3426 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3427 // if it is. See bug 4774291. 3428 3429 // For Reflection.getClassAccessFlags(), the null check occurs in 3430 // the wrong place; see inline_unsafe_access(), above, for a similar 3431 // situation. 3432 mirror = null_check(mirror); 3433 // If mirror or obj is dead, only null-path is taken. 3434 if (stopped()) return true; 3435 3436 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3437 3438 // Now load the mirror's klass metaobject, and null-check it. 3439 // Side-effects region with the control path if the klass is null. 3440 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3441 // If kls is null, we have a primitive mirror. 3442 phi->init_req(_prim_path, prim_return_value); 3443 if (stopped()) { set_result(region, phi); return true; } 3444 bool safe_for_replace = (region->in(_prim_path) == top()); 3445 3446 Node* p; // handy temp 3447 Node* null_ctl; 3448 3449 // Now that we have the non-null klass, we can perform the real query. 3450 // For constant classes, the query will constant-fold in LoadNode::Value. 3451 Node* query_value = top(); 3452 switch (id) { 3453 case vmIntrinsics::_isInstance: 3454 // nothing is an instance of a primitive type 3455 query_value = gen_instanceof(obj, kls, safe_for_replace); 3456 break; 3457 3458 case vmIntrinsics::_getModifiers: 3459 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3460 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3461 break; 3462 3463 case vmIntrinsics::_isInterface: 3464 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3465 if (generate_interface_guard(kls, region) != NULL) 3466 // A guard was added. If the guard is taken, it was an interface. 3467 phi->add_req(intcon(1)); 3468 // If we fall through, it's a plain class. 3469 query_value = intcon(0); 3470 break; 3471 3472 case vmIntrinsics::_isArray: 3473 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3474 if (generate_array_guard(kls, region) != NULL) 3475 // A guard was added. If the guard is taken, it was an array. 3476 phi->add_req(intcon(1)); 3477 // If we fall through, it's a plain class. 3478 query_value = intcon(0); 3479 break; 3480 3481 case vmIntrinsics::_isPrimitive: 3482 query_value = intcon(0); // "normal" path produces false 3483 break; 3484 3485 case vmIntrinsics::_getSuperclass: 3486 // The rules here are somewhat unfortunate, but we can still do better 3487 // with random logic than with a JNI call. 3488 // Interfaces store null or Object as _super, but must report null. 3489 // Arrays store an intermediate super as _super, but must report Object. 3490 // Other types can report the actual _super. 3491 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3492 if (generate_interface_guard(kls, region) != NULL) 3493 // A guard was added. If the guard is taken, it was an interface. 3494 phi->add_req(null()); 3495 if (generate_array_guard(kls, region) != NULL) 3496 // A guard was added. If the guard is taken, it was an array. 3497 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3498 // If we fall through, it's a plain class. Get its _super. 3499 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3500 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3501 null_ctl = top(); 3502 kls = null_check_oop(kls, &null_ctl); 3503 if (null_ctl != top()) { 3504 // If the guard is taken, Object.superClass is null (both klass and mirror). 3505 region->add_req(null_ctl); 3506 phi ->add_req(null()); 3507 } 3508 if (!stopped()) { 3509 query_value = load_mirror_from_klass(kls); 3510 } 3511 break; 3512 3513 case vmIntrinsics::_getClassAccessFlags: 3514 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3515 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3516 break; 3517 3518 default: 3519 fatal_unexpected_iid(id); 3520 break; 3521 } 3522 3523 // Fall-through is the normal case of a query to a real class. 3524 phi->init_req(1, query_value); 3525 region->init_req(1, control()); 3526 3527 C->set_has_split_ifs(true); // Has chance for split-if optimization 3528 set_result(region, phi); 3529 return true; 3530 } 3531 3532 //-------------------------inline_Class_cast------------------- 3533 bool LibraryCallKit::inline_Class_cast() { 3534 Node* mirror = argument(0); // Class 3535 Node* obj = argument(1); 3536 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3537 if (mirror_con == NULL) { 3538 return false; // dead path (mirror->is_top()). 3539 } 3540 if (obj == NULL || obj->is_top()) { 3541 return false; // dead path 3542 } 3543 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3544 3545 // First, see if Class.cast() can be folded statically. 3546 // java_mirror_type() returns non-null for compile-time Class constants. 3547 ciType* tm = mirror_con->java_mirror_type(); 3548 if (tm != NULL && tm->is_klass() && 3549 tp != NULL && tp->klass() != NULL) { 3550 if (!tp->klass()->is_loaded()) { 3551 // Don't use intrinsic when class is not loaded. 3552 return false; 3553 } else { 3554 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3555 if (static_res == Compile::SSC_always_true) { 3556 // isInstance() is true - fold the code. 3557 set_result(obj); 3558 return true; 3559 } else if (static_res == Compile::SSC_always_false) { 3560 // Don't use intrinsic, have to throw ClassCastException. 3561 // If the reference is null, the non-intrinsic bytecode will 3562 // be optimized appropriately. 3563 return false; 3564 } 3565 } 3566 } 3567 3568 // Bailout intrinsic and do normal inlining if exception path is frequent. 3569 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3570 return false; 3571 } 3572 3573 // Generate dynamic checks. 3574 // Class.cast() is java implementation of _checkcast bytecode. 3575 // Do checkcast (Parse::do_checkcast()) optimizations here. 3576 3577 mirror = null_check(mirror); 3578 // If mirror is dead, only null-path is taken. 3579 if (stopped()) { 3580 return true; 3581 } 3582 3583 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3584 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3585 RegionNode* region = new RegionNode(PATH_LIMIT); 3586 record_for_igvn(region); 3587 3588 // Now load the mirror's klass metaobject, and null-check it. 3589 // If kls is null, we have a primitive mirror and 3590 // nothing is an instance of a primitive type. 3591 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3592 3593 Node* res = top(); 3594 if (!stopped()) { 3595 Node* bad_type_ctrl = top(); 3596 // Do checkcast optimizations. 3597 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3598 region->init_req(_bad_type_path, bad_type_ctrl); 3599 } 3600 if (region->in(_prim_path) != top() || 3601 region->in(_bad_type_path) != top()) { 3602 // Let Interpreter throw ClassCastException. 3603 PreserveJVMState pjvms(this); 3604 set_control(_gvn.transform(region)); 3605 uncommon_trap(Deoptimization::Reason_intrinsic, 3606 Deoptimization::Action_maybe_recompile); 3607 } 3608 if (!stopped()) { 3609 set_result(res); 3610 } 3611 return true; 3612 } 3613 3614 3615 //--------------------------inline_native_subtype_check------------------------ 3616 // This intrinsic takes the JNI calls out of the heart of 3617 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3618 bool LibraryCallKit::inline_native_subtype_check() { 3619 // Pull both arguments off the stack. 3620 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3621 args[0] = argument(0); 3622 args[1] = argument(1); 3623 Node* klasses[2]; // corresponding Klasses: superk, subk 3624 klasses[0] = klasses[1] = top(); 3625 3626 enum { 3627 // A full decision tree on {superc is prim, subc is prim}: 3628 _prim_0_path = 1, // {P,N} => false 3629 // {P,P} & superc!=subc => false 3630 _prim_same_path, // {P,P} & superc==subc => true 3631 _prim_1_path, // {N,P} => false 3632 _ref_subtype_path, // {N,N} & subtype check wins => true 3633 _both_ref_path, // {N,N} & subtype check loses => false 3634 PATH_LIMIT 3635 }; 3636 3637 RegionNode* region = new RegionNode(PATH_LIMIT); 3638 Node* phi = new PhiNode(region, TypeInt::BOOL); 3639 record_for_igvn(region); 3640 3641 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3642 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3643 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3644 3645 // First null-check both mirrors and load each mirror's klass metaobject. 3646 int which_arg; 3647 for (which_arg = 0; which_arg <= 1; which_arg++) { 3648 Node* arg = args[which_arg]; 3649 arg = null_check(arg); 3650 if (stopped()) break; 3651 args[which_arg] = arg; 3652 3653 Node* p = basic_plus_adr(arg, class_klass_offset); 3654 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3655 klasses[which_arg] = _gvn.transform(kls); 3656 } 3657 3658 // Having loaded both klasses, test each for null. 3659 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3660 for (which_arg = 0; which_arg <= 1; which_arg++) { 3661 Node* kls = klasses[which_arg]; 3662 Node* null_ctl = top(); 3663 kls = null_check_oop(kls, &null_ctl, never_see_null); 3664 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3665 region->init_req(prim_path, null_ctl); 3666 if (stopped()) break; 3667 klasses[which_arg] = kls; 3668 } 3669 3670 if (!stopped()) { 3671 // now we have two reference types, in klasses[0..1] 3672 Node* subk = klasses[1]; // the argument to isAssignableFrom 3673 Node* superk = klasses[0]; // the receiver 3674 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3675 // now we have a successful reference subtype check 3676 region->set_req(_ref_subtype_path, control()); 3677 } 3678 3679 // If both operands are primitive (both klasses null), then 3680 // we must return true when they are identical primitives. 3681 // It is convenient to test this after the first null klass check. 3682 set_control(region->in(_prim_0_path)); // go back to first null check 3683 if (!stopped()) { 3684 // Since superc is primitive, make a guard for the superc==subc case. 3685 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3686 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3687 generate_guard(bol_eq, region, PROB_FAIR); 3688 if (region->req() == PATH_LIMIT+1) { 3689 // A guard was added. If the added guard is taken, superc==subc. 3690 region->swap_edges(PATH_LIMIT, _prim_same_path); 3691 region->del_req(PATH_LIMIT); 3692 } 3693 region->set_req(_prim_0_path, control()); // Not equal after all. 3694 } 3695 3696 // these are the only paths that produce 'true': 3697 phi->set_req(_prim_same_path, intcon(1)); 3698 phi->set_req(_ref_subtype_path, intcon(1)); 3699 3700 // pull together the cases: 3701 assert(region->req() == PATH_LIMIT, "sane region"); 3702 for (uint i = 1; i < region->req(); i++) { 3703 Node* ctl = region->in(i); 3704 if (ctl == NULL || ctl == top()) { 3705 region->set_req(i, top()); 3706 phi ->set_req(i, top()); 3707 } else if (phi->in(i) == NULL) { 3708 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3709 } 3710 } 3711 3712 set_control(_gvn.transform(region)); 3713 set_result(_gvn.transform(phi)); 3714 return true; 3715 } 3716 3717 //---------------------generate_array_guard_common------------------------ 3718 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3719 bool obj_array, bool not_array) { 3720 3721 if (stopped()) { 3722 return NULL; 3723 } 3724 3725 // If obj_array/non_array==false/false: 3726 // Branch around if the given klass is in fact an array (either obj or prim). 3727 // If obj_array/non_array==false/true: 3728 // Branch around if the given klass is not an array klass of any kind. 3729 // If obj_array/non_array==true/true: 3730 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3731 // If obj_array/non_array==true/false: 3732 // Branch around if the kls is an oop array (Object[] or subtype) 3733 // 3734 // Like generate_guard, adds a new path onto the region. 3735 jint layout_con = 0; 3736 Node* layout_val = get_layout_helper(kls, layout_con); 3737 if (layout_val == NULL) { 3738 bool query = (obj_array 3739 ? Klass::layout_helper_is_objArray(layout_con) 3740 : Klass::layout_helper_is_array(layout_con)); 3741 if (query == not_array) { 3742 return NULL; // never a branch 3743 } else { // always a branch 3744 Node* always_branch = control(); 3745 if (region != NULL) 3746 region->add_req(always_branch); 3747 set_control(top()); 3748 return always_branch; 3749 } 3750 } 3751 // Now test the correct condition. 3752 jint nval = (obj_array 3753 ? ((jint)Klass::_lh_array_tag_type_value 3754 << Klass::_lh_array_tag_shift) 3755 : Klass::_lh_neutral_value); 3756 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3757 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3758 // invert the test if we are looking for a non-array 3759 if (not_array) btest = BoolTest(btest).negate(); 3760 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3761 return generate_fair_guard(bol, region); 3762 } 3763 3764 3765 //-----------------------inline_native_newArray-------------------------- 3766 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3767 bool LibraryCallKit::inline_native_newArray() { 3768 Node* mirror = argument(0); 3769 Node* count_val = argument(1); 3770 3771 mirror = null_check(mirror); 3772 // If mirror or obj is dead, only null-path is taken. 3773 if (stopped()) return true; 3774 3775 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3776 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3777 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3778 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3779 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3780 3781 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3782 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3783 result_reg, _slow_path); 3784 Node* normal_ctl = control(); 3785 Node* no_array_ctl = result_reg->in(_slow_path); 3786 3787 // Generate code for the slow case. We make a call to newArray(). 3788 set_control(no_array_ctl); 3789 if (!stopped()) { 3790 // Either the input type is void.class, or else the 3791 // array klass has not yet been cached. Either the 3792 // ensuing call will throw an exception, or else it 3793 // will cache the array klass for next time. 3794 PreserveJVMState pjvms(this); 3795 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3796 Node* slow_result = set_results_for_java_call(slow_call); 3797 // this->control() comes from set_results_for_java_call 3798 result_reg->set_req(_slow_path, control()); 3799 result_val->set_req(_slow_path, slow_result); 3800 result_io ->set_req(_slow_path, i_o()); 3801 result_mem->set_req(_slow_path, reset_memory()); 3802 } 3803 3804 set_control(normal_ctl); 3805 if (!stopped()) { 3806 // Normal case: The array type has been cached in the java.lang.Class. 3807 // The following call works fine even if the array type is polymorphic. 3808 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3809 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3810 result_reg->init_req(_normal_path, control()); 3811 result_val->init_req(_normal_path, obj); 3812 result_io ->init_req(_normal_path, i_o()); 3813 result_mem->init_req(_normal_path, reset_memory()); 3814 } 3815 3816 // Return the combined state. 3817 set_i_o( _gvn.transform(result_io) ); 3818 set_all_memory( _gvn.transform(result_mem)); 3819 3820 C->set_has_split_ifs(true); // Has chance for split-if optimization 3821 set_result(result_reg, result_val); 3822 return true; 3823 } 3824 3825 //----------------------inline_native_getLength-------------------------- 3826 // public static native int java.lang.reflect.Array.getLength(Object array); 3827 bool LibraryCallKit::inline_native_getLength() { 3828 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3829 3830 Node* array = null_check(argument(0)); 3831 // If array is dead, only null-path is taken. 3832 if (stopped()) return true; 3833 3834 // Deoptimize if it is a non-array. 3835 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3836 3837 if (non_array != NULL) { 3838 PreserveJVMState pjvms(this); 3839 set_control(non_array); 3840 uncommon_trap(Deoptimization::Reason_intrinsic, 3841 Deoptimization::Action_maybe_recompile); 3842 } 3843 3844 // If control is dead, only non-array-path is taken. 3845 if (stopped()) return true; 3846 3847 // The works fine even if the array type is polymorphic. 3848 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3849 Node* result = load_array_length(array); 3850 3851 C->set_has_split_ifs(true); // Has chance for split-if optimization 3852 set_result(result); 3853 return true; 3854 } 3855 3856 //------------------------inline_array_copyOf---------------------------- 3857 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3858 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3859 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3860 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3861 3862 // Get the arguments. 3863 Node* original = argument(0); 3864 Node* start = is_copyOfRange? argument(1): intcon(0); 3865 Node* end = is_copyOfRange? argument(2): argument(1); 3866 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3867 3868 Node* newcopy; 3869 3870 // Set the original stack and the reexecute bit for the interpreter to reexecute 3871 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3872 { PreserveReexecuteState preexecs(this); 3873 jvms()->set_should_reexecute(true); 3874 3875 array_type_mirror = null_check(array_type_mirror); 3876 original = null_check(original); 3877 3878 // Check if a null path was taken unconditionally. 3879 if (stopped()) return true; 3880 3881 Node* orig_length = load_array_length(original); 3882 3883 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3884 klass_node = null_check(klass_node); 3885 3886 RegionNode* bailout = new RegionNode(1); 3887 record_for_igvn(bailout); 3888 3889 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3890 // Bail out if that is so. 3891 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3892 if (not_objArray != NULL) { 3893 // Improve the klass node's type from the new optimistic assumption: 3894 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3895 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3896 Node* cast = new CastPPNode(klass_node, akls); 3897 cast->init_req(0, control()); 3898 klass_node = _gvn.transform(cast); 3899 } 3900 3901 // Bail out if either start or end is negative. 3902 generate_negative_guard(start, bailout, &start); 3903 generate_negative_guard(end, bailout, &end); 3904 3905 Node* length = end; 3906 if (_gvn.type(start) != TypeInt::ZERO) { 3907 length = _gvn.transform(new SubINode(end, start)); 3908 } 3909 3910 // Bail out if length is negative. 3911 // Without this the new_array would throw 3912 // NegativeArraySizeException but IllegalArgumentException is what 3913 // should be thrown 3914 generate_negative_guard(length, bailout, &length); 3915 3916 if (bailout->req() > 1) { 3917 PreserveJVMState pjvms(this); 3918 set_control(_gvn.transform(bailout)); 3919 uncommon_trap(Deoptimization::Reason_intrinsic, 3920 Deoptimization::Action_maybe_recompile); 3921 } 3922 3923 if (!stopped()) { 3924 // How many elements will we copy from the original? 3925 // The answer is MinI(orig_length - start, length). 3926 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3927 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3928 3929 // Generate a direct call to the right arraycopy function(s). 3930 // We know the copy is disjoint but we might not know if the 3931 // oop stores need checking. 3932 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3933 // This will fail a store-check if x contains any non-nulls. 3934 3935 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3936 // loads/stores but it is legal only if we're sure the 3937 // Arrays.copyOf would succeed. So we need all input arguments 3938 // to the copyOf to be validated, including that the copy to the 3939 // new array won't trigger an ArrayStoreException. That subtype 3940 // check can be optimized if we know something on the type of 3941 // the input array from type speculation. 3942 if (_gvn.type(klass_node)->singleton()) { 3943 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3944 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3945 3946 int test = C->static_subtype_check(superk, subk); 3947 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3948 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3949 if (t_original->speculative_type() != NULL) { 3950 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3951 } 3952 } 3953 } 3954 3955 bool validated = false; 3956 // Reason_class_check rather than Reason_intrinsic because we 3957 // want to intrinsify even if this traps. 3958 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3959 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3960 klass_node); 3961 3962 if (not_subtype_ctrl != top()) { 3963 PreserveJVMState pjvms(this); 3964 set_control(not_subtype_ctrl); 3965 uncommon_trap(Deoptimization::Reason_class_check, 3966 Deoptimization::Action_make_not_entrant); 3967 assert(stopped(), "Should be stopped"); 3968 } 3969 validated = true; 3970 } 3971 3972 if (!stopped()) { 3973 newcopy = new_array(klass_node, length, 0); // no arguments to push 3974 3975 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3976 load_object_klass(original), klass_node); 3977 if (!is_copyOfRange) { 3978 ac->set_copyof(validated); 3979 } else { 3980 ac->set_copyofrange(validated); 3981 } 3982 Node* n = _gvn.transform(ac); 3983 if (n == ac) { 3984 ac->connect_outputs(this); 3985 } else { 3986 assert(validated, "shouldn't transform if all arguments not validated"); 3987 set_all_memory(n); 3988 } 3989 } 3990 } 3991 } // original reexecute is set back here 3992 3993 C->set_has_split_ifs(true); // Has chance for split-if optimization 3994 if (!stopped()) { 3995 set_result(newcopy); 3996 } 3997 return true; 3998 } 3999 4000 4001 //----------------------generate_virtual_guard--------------------------- 4002 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4003 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4004 RegionNode* slow_region) { 4005 ciMethod* method = callee(); 4006 int vtable_index = method->vtable_index(); 4007 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4008 "bad index %d", vtable_index); 4009 // Get the Method* out of the appropriate vtable entry. 4010 int entry_offset = (InstanceKlass::vtable_start_offset() + 4011 vtable_index*vtableEntry::size()) * wordSize + 4012 vtableEntry::method_offset_in_bytes(); 4013 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4014 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4015 4016 // Compare the target method with the expected method (e.g., Object.hashCode). 4017 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4018 4019 Node* native_call = makecon(native_call_addr); 4020 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4021 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4022 4023 return generate_slow_guard(test_native, slow_region); 4024 } 4025 4026 //-----------------------generate_method_call---------------------------- 4027 // Use generate_method_call to make a slow-call to the real 4028 // method if the fast path fails. An alternative would be to 4029 // use a stub like OptoRuntime::slow_arraycopy_Java. 4030 // This only works for expanding the current library call, 4031 // not another intrinsic. (E.g., don't use this for making an 4032 // arraycopy call inside of the copyOf intrinsic.) 4033 CallJavaNode* 4034 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4035 // When compiling the intrinsic method itself, do not use this technique. 4036 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4037 4038 ciMethod* method = callee(); 4039 // ensure the JVMS we have will be correct for this call 4040 guarantee(method_id == method->intrinsic_id(), "must match"); 4041 4042 const TypeFunc* tf = TypeFunc::make(method); 4043 CallJavaNode* slow_call; 4044 if (is_static) { 4045 assert(!is_virtual, ""); 4046 slow_call = new CallStaticJavaNode(C, tf, 4047 SharedRuntime::get_resolve_static_call_stub(), 4048 method, bci()); 4049 } else if (is_virtual) { 4050 null_check_receiver(); 4051 int vtable_index = Method::invalid_vtable_index; 4052 if (UseInlineCaches) { 4053 // Suppress the vtable call 4054 } else { 4055 // hashCode and clone are not a miranda methods, 4056 // so the vtable index is fixed. 4057 // No need to use the linkResolver to get it. 4058 vtable_index = method->vtable_index(); 4059 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4060 "bad index %d", vtable_index); 4061 } 4062 slow_call = new CallDynamicJavaNode(tf, 4063 SharedRuntime::get_resolve_virtual_call_stub(), 4064 method, vtable_index, bci()); 4065 } else { // neither virtual nor static: opt_virtual 4066 null_check_receiver(); 4067 slow_call = new CallStaticJavaNode(C, tf, 4068 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4069 method, bci()); 4070 slow_call->set_optimized_virtual(true); 4071 } 4072 set_arguments_for_java_call(slow_call); 4073 set_edges_for_java_call(slow_call); 4074 return slow_call; 4075 } 4076 4077 4078 /** 4079 * Build special case code for calls to hashCode on an object. This call may 4080 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4081 * slightly different code. 4082 */ 4083 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4084 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4085 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4086 4087 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4088 4089 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4090 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4091 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4092 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4093 Node* obj = NULL; 4094 if (!is_static) { 4095 // Check for hashing null object 4096 obj = null_check_receiver(); 4097 if (stopped()) return true; // unconditionally null 4098 result_reg->init_req(_null_path, top()); 4099 result_val->init_req(_null_path, top()); 4100 } else { 4101 // Do a null check, and return zero if null. 4102 // System.identityHashCode(null) == 0 4103 obj = argument(0); 4104 Node* null_ctl = top(); 4105 obj = null_check_oop(obj, &null_ctl); 4106 result_reg->init_req(_null_path, null_ctl); 4107 result_val->init_req(_null_path, _gvn.intcon(0)); 4108 } 4109 4110 // Unconditionally null? Then return right away. 4111 if (stopped()) { 4112 set_control( result_reg->in(_null_path)); 4113 if (!stopped()) 4114 set_result(result_val->in(_null_path)); 4115 return true; 4116 } 4117 4118 // We only go to the fast case code if we pass a number of guards. The 4119 // paths which do not pass are accumulated in the slow_region. 4120 RegionNode* slow_region = new RegionNode(1); 4121 record_for_igvn(slow_region); 4122 4123 // If this is a virtual call, we generate a funny guard. We pull out 4124 // the vtable entry corresponding to hashCode() from the target object. 4125 // If the target method which we are calling happens to be the native 4126 // Object hashCode() method, we pass the guard. We do not need this 4127 // guard for non-virtual calls -- the caller is known to be the native 4128 // Object hashCode(). 4129 if (is_virtual) { 4130 // After null check, get the object's klass. 4131 Node* obj_klass = load_object_klass(obj); 4132 generate_virtual_guard(obj_klass, slow_region); 4133 } 4134 4135 // Get the header out of the object, use LoadMarkNode when available 4136 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4137 // The control of the load must be NULL. Otherwise, the load can move before 4138 // the null check after castPP removal. 4139 Node* no_ctrl = NULL; 4140 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4141 4142 // Test the header to see if it is unlocked. 4143 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4144 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4145 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4146 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4147 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4148 4149 generate_slow_guard(test_unlocked, slow_region); 4150 4151 // Get the hash value and check to see that it has been properly assigned. 4152 // We depend on hash_mask being at most 32 bits and avoid the use of 4153 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4154 // vm: see markOop.hpp. 4155 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4156 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4157 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4158 // This hack lets the hash bits live anywhere in the mark object now, as long 4159 // as the shift drops the relevant bits into the low 32 bits. Note that 4160 // Java spec says that HashCode is an int so there's no point in capturing 4161 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4162 hshifted_header = ConvX2I(hshifted_header); 4163 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4164 4165 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4166 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4167 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4168 4169 generate_slow_guard(test_assigned, slow_region); 4170 4171 Node* init_mem = reset_memory(); 4172 // fill in the rest of the null path: 4173 result_io ->init_req(_null_path, i_o()); 4174 result_mem->init_req(_null_path, init_mem); 4175 4176 result_val->init_req(_fast_path, hash_val); 4177 result_reg->init_req(_fast_path, control()); 4178 result_io ->init_req(_fast_path, i_o()); 4179 result_mem->init_req(_fast_path, init_mem); 4180 4181 // Generate code for the slow case. We make a call to hashCode(). 4182 set_control(_gvn.transform(slow_region)); 4183 if (!stopped()) { 4184 // No need for PreserveJVMState, because we're using up the present state. 4185 set_all_memory(init_mem); 4186 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4187 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4188 Node* slow_result = set_results_for_java_call(slow_call); 4189 // this->control() comes from set_results_for_java_call 4190 result_reg->init_req(_slow_path, control()); 4191 result_val->init_req(_slow_path, slow_result); 4192 result_io ->set_req(_slow_path, i_o()); 4193 result_mem ->set_req(_slow_path, reset_memory()); 4194 } 4195 4196 // Return the combined state. 4197 set_i_o( _gvn.transform(result_io) ); 4198 set_all_memory( _gvn.transform(result_mem)); 4199 4200 set_result(result_reg, result_val); 4201 return true; 4202 } 4203 4204 //---------------------------inline_native_getClass---------------------------- 4205 // public final native Class<?> java.lang.Object.getClass(); 4206 // 4207 // Build special case code for calls to getClass on an object. 4208 bool LibraryCallKit::inline_native_getClass() { 4209 Node* obj = null_check_receiver(); 4210 if (stopped()) return true; 4211 set_result(load_mirror_from_klass(load_object_klass(obj))); 4212 return true; 4213 } 4214 4215 //-----------------inline_native_Reflection_getCallerClass--------------------- 4216 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4217 // 4218 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4219 // 4220 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4221 // in that it must skip particular security frames and checks for 4222 // caller sensitive methods. 4223 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4224 #ifndef PRODUCT 4225 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4226 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4227 } 4228 #endif 4229 4230 if (!jvms()->has_method()) { 4231 #ifndef PRODUCT 4232 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4233 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4234 } 4235 #endif 4236 return false; 4237 } 4238 4239 // Walk back up the JVM state to find the caller at the required 4240 // depth. 4241 JVMState* caller_jvms = jvms(); 4242 4243 // Cf. JVM_GetCallerClass 4244 // NOTE: Start the loop at depth 1 because the current JVM state does 4245 // not include the Reflection.getCallerClass() frame. 4246 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4247 ciMethod* m = caller_jvms->method(); 4248 switch (n) { 4249 case 0: 4250 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4251 break; 4252 case 1: 4253 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4254 if (!m->caller_sensitive()) { 4255 #ifndef PRODUCT 4256 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4257 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4258 } 4259 #endif 4260 return false; // bail-out; let JVM_GetCallerClass do the work 4261 } 4262 break; 4263 default: 4264 if (!m->is_ignored_by_security_stack_walk()) { 4265 // We have reached the desired frame; return the holder class. 4266 // Acquire method holder as java.lang.Class and push as constant. 4267 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4268 ciInstance* caller_mirror = caller_klass->java_mirror(); 4269 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4270 4271 #ifndef PRODUCT 4272 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4273 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()); 4274 tty->print_cr(" JVM state at this point:"); 4275 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4276 ciMethod* m = jvms()->of_depth(i)->method(); 4277 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4278 } 4279 } 4280 #endif 4281 return true; 4282 } 4283 break; 4284 } 4285 } 4286 4287 #ifndef PRODUCT 4288 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4289 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4290 tty->print_cr(" JVM state at this point:"); 4291 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4292 ciMethod* m = jvms()->of_depth(i)->method(); 4293 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4294 } 4295 } 4296 #endif 4297 4298 return false; // bail-out; let JVM_GetCallerClass do the work 4299 } 4300 4301 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4302 Node* arg = argument(0); 4303 Node* result; 4304 4305 switch (id) { 4306 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4307 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4308 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4309 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4310 4311 case vmIntrinsics::_doubleToLongBits: { 4312 // two paths (plus control) merge in a wood 4313 RegionNode *r = new RegionNode(3); 4314 Node *phi = new PhiNode(r, TypeLong::LONG); 4315 4316 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4317 // Build the boolean node 4318 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4319 4320 // Branch either way. 4321 // NaN case is less traveled, which makes all the difference. 4322 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4323 Node *opt_isnan = _gvn.transform(ifisnan); 4324 assert( opt_isnan->is_If(), "Expect an IfNode"); 4325 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4326 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4327 4328 set_control(iftrue); 4329 4330 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4331 Node *slow_result = longcon(nan_bits); // return NaN 4332 phi->init_req(1, _gvn.transform( slow_result )); 4333 r->init_req(1, iftrue); 4334 4335 // Else fall through 4336 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4337 set_control(iffalse); 4338 4339 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4340 r->init_req(2, iffalse); 4341 4342 // Post merge 4343 set_control(_gvn.transform(r)); 4344 record_for_igvn(r); 4345 4346 C->set_has_split_ifs(true); // Has chance for split-if optimization 4347 result = phi; 4348 assert(result->bottom_type()->isa_long(), "must be"); 4349 break; 4350 } 4351 4352 case vmIntrinsics::_floatToIntBits: { 4353 // two paths (plus control) merge in a wood 4354 RegionNode *r = new RegionNode(3); 4355 Node *phi = new PhiNode(r, TypeInt::INT); 4356 4357 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4358 // Build the boolean node 4359 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4360 4361 // Branch either way. 4362 // NaN case is less traveled, which makes all the difference. 4363 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4364 Node *opt_isnan = _gvn.transform(ifisnan); 4365 assert( opt_isnan->is_If(), "Expect an IfNode"); 4366 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4367 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4368 4369 set_control(iftrue); 4370 4371 static const jint nan_bits = 0x7fc00000; 4372 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4373 phi->init_req(1, _gvn.transform( slow_result )); 4374 r->init_req(1, iftrue); 4375 4376 // Else fall through 4377 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4378 set_control(iffalse); 4379 4380 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4381 r->init_req(2, iffalse); 4382 4383 // Post merge 4384 set_control(_gvn.transform(r)); 4385 record_for_igvn(r); 4386 4387 C->set_has_split_ifs(true); // Has chance for split-if optimization 4388 result = phi; 4389 assert(result->bottom_type()->isa_int(), "must be"); 4390 break; 4391 } 4392 4393 default: 4394 fatal_unexpected_iid(id); 4395 break; 4396 } 4397 set_result(_gvn.transform(result)); 4398 return true; 4399 } 4400 4401 //----------------------inline_unsafe_copyMemory------------------------- 4402 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4403 bool LibraryCallKit::inline_unsafe_copyMemory() { 4404 if (callee()->is_static()) return false; // caller must have the capability! 4405 null_check_receiver(); // null-check receiver 4406 if (stopped()) return true; 4407 4408 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4409 4410 Node* src_ptr = argument(1); // type: oop 4411 Node* src_off = ConvL2X(argument(2)); // type: long 4412 Node* dst_ptr = argument(4); // type: oop 4413 Node* dst_off = ConvL2X(argument(5)); // type: long 4414 Node* size = ConvL2X(argument(7)); // type: long 4415 4416 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4417 "fieldOffset must be byte-scaled"); 4418 4419 Node* src = make_unsafe_address(src_ptr, src_off); 4420 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4421 4422 // Conservatively insert a memory barrier on all memory slices. 4423 // Do not let writes of the copy source or destination float below the copy. 4424 insert_mem_bar(Op_MemBarCPUOrder); 4425 4426 // Call it. Note that the length argument is not scaled. 4427 make_runtime_call(RC_LEAF|RC_NO_FP, 4428 OptoRuntime::fast_arraycopy_Type(), 4429 StubRoutines::unsafe_arraycopy(), 4430 "unsafe_arraycopy", 4431 TypeRawPtr::BOTTOM, 4432 src, dst, size XTOP); 4433 4434 // Do not let reads of the copy destination float above the copy. 4435 insert_mem_bar(Op_MemBarCPUOrder); 4436 4437 return true; 4438 } 4439 4440 //------------------------clone_coping----------------------------------- 4441 // Helper function for inline_native_clone. 4442 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4443 assert(obj_size != NULL, ""); 4444 Node* raw_obj = alloc_obj->in(1); 4445 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4446 4447 AllocateNode* alloc = NULL; 4448 if (ReduceBulkZeroing) { 4449 // We will be completely responsible for initializing this object - 4450 // mark Initialize node as complete. 4451 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4452 // The object was just allocated - there should be no any stores! 4453 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4454 // Mark as complete_with_arraycopy so that on AllocateNode 4455 // expansion, we know this AllocateNode is initialized by an array 4456 // copy and a StoreStore barrier exists after the array copy. 4457 alloc->initialization()->set_complete_with_arraycopy(); 4458 } 4459 4460 // Copy the fastest available way. 4461 // TODO: generate fields copies for small objects instead. 4462 Node* src = obj; 4463 Node* dest = alloc_obj; 4464 Node* size = _gvn.transform(obj_size); 4465 4466 // Exclude the header but include array length to copy by 8 bytes words. 4467 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4468 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4469 instanceOopDesc::base_offset_in_bytes(); 4470 // base_off: 4471 // 8 - 32-bit VM 4472 // 12 - 64-bit VM, compressed klass 4473 // 16 - 64-bit VM, normal klass 4474 if (base_off % BytesPerLong != 0) { 4475 assert(UseCompressedClassPointers, ""); 4476 if (is_array) { 4477 // Exclude length to copy by 8 bytes words. 4478 base_off += sizeof(int); 4479 } else { 4480 // Include klass to copy by 8 bytes words. 4481 base_off = instanceOopDesc::klass_offset_in_bytes(); 4482 } 4483 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4484 } 4485 src = basic_plus_adr(src, base_off); 4486 dest = basic_plus_adr(dest, base_off); 4487 4488 // Compute the length also, if needed: 4489 Node* countx = size; 4490 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4491 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4492 4493 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4494 4495 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4496 ac->set_clonebasic(); 4497 Node* n = _gvn.transform(ac); 4498 if (n == ac) { 4499 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4500 } else { 4501 set_all_memory(n); 4502 } 4503 4504 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4505 if (card_mark) { 4506 assert(!is_array, ""); 4507 // Put in store barrier for any and all oops we are sticking 4508 // into this object. (We could avoid this if we could prove 4509 // that the object type contains no oop fields at all.) 4510 Node* no_particular_value = NULL; 4511 Node* no_particular_field = NULL; 4512 int raw_adr_idx = Compile::AliasIdxRaw; 4513 post_barrier(control(), 4514 memory(raw_adr_type), 4515 alloc_obj, 4516 no_particular_field, 4517 raw_adr_idx, 4518 no_particular_value, 4519 T_OBJECT, 4520 false); 4521 } 4522 4523 // Do not let reads from the cloned object float above the arraycopy. 4524 if (alloc != NULL) { 4525 // Do not let stores that initialize this object be reordered with 4526 // a subsequent store that would make this object accessible by 4527 // other threads. 4528 // Record what AllocateNode this StoreStore protects so that 4529 // escape analysis can go from the MemBarStoreStoreNode to the 4530 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4531 // based on the escape status of the AllocateNode. 4532 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4533 } else { 4534 insert_mem_bar(Op_MemBarCPUOrder); 4535 } 4536 } 4537 4538 //------------------------inline_native_clone---------------------------- 4539 // protected native Object java.lang.Object.clone(); 4540 // 4541 // Here are the simple edge cases: 4542 // null receiver => normal trap 4543 // virtual and clone was overridden => slow path to out-of-line clone 4544 // not cloneable or finalizer => slow path to out-of-line Object.clone 4545 // 4546 // The general case has two steps, allocation and copying. 4547 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4548 // 4549 // Copying also has two cases, oop arrays and everything else. 4550 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4551 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4552 // 4553 // These steps fold up nicely if and when the cloned object's klass 4554 // can be sharply typed as an object array, a type array, or an instance. 4555 // 4556 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4557 PhiNode* result_val; 4558 4559 // Set the reexecute bit for the interpreter to reexecute 4560 // the bytecode that invokes Object.clone if deoptimization happens. 4561 { PreserveReexecuteState preexecs(this); 4562 jvms()->set_should_reexecute(true); 4563 4564 Node* obj = null_check_receiver(); 4565 if (stopped()) return true; 4566 4567 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4568 4569 // If we are going to clone an instance, we need its exact type to 4570 // know the number and types of fields to convert the clone to 4571 // loads/stores. Maybe a speculative type can help us. 4572 if (!obj_type->klass_is_exact() && 4573 obj_type->speculative_type() != NULL && 4574 obj_type->speculative_type()->is_instance_klass()) { 4575 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4576 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4577 !spec_ik->has_injected_fields()) { 4578 ciKlass* k = obj_type->klass(); 4579 if (!k->is_instance_klass() || 4580 k->as_instance_klass()->is_interface() || 4581 k->as_instance_klass()->has_subklass()) { 4582 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4583 } 4584 } 4585 } 4586 4587 Node* obj_klass = load_object_klass(obj); 4588 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4589 const TypeOopPtr* toop = ((tklass != NULL) 4590 ? tklass->as_instance_type() 4591 : TypeInstPtr::NOTNULL); 4592 4593 // Conservatively insert a memory barrier on all memory slices. 4594 // Do not let writes into the original float below the clone. 4595 insert_mem_bar(Op_MemBarCPUOrder); 4596 4597 // paths into result_reg: 4598 enum { 4599 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4600 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4601 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4602 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4603 PATH_LIMIT 4604 }; 4605 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4606 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4607 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4608 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4609 record_for_igvn(result_reg); 4610 4611 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4612 int raw_adr_idx = Compile::AliasIdxRaw; 4613 4614 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4615 if (array_ctl != NULL) { 4616 // It's an array. 4617 PreserveJVMState pjvms(this); 4618 set_control(array_ctl); 4619 Node* obj_length = load_array_length(obj); 4620 Node* obj_size = NULL; 4621 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4622 4623 if (!use_ReduceInitialCardMarks()) { 4624 // If it is an oop array, it requires very special treatment, 4625 // because card marking is required on each card of the array. 4626 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4627 if (is_obja != NULL) { 4628 PreserveJVMState pjvms2(this); 4629 set_control(is_obja); 4630 // Generate a direct call to the right arraycopy function(s). 4631 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4632 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4633 ac->set_cloneoop(); 4634 Node* n = _gvn.transform(ac); 4635 assert(n == ac, "cannot disappear"); 4636 ac->connect_outputs(this); 4637 4638 result_reg->init_req(_objArray_path, control()); 4639 result_val->init_req(_objArray_path, alloc_obj); 4640 result_i_o ->set_req(_objArray_path, i_o()); 4641 result_mem ->set_req(_objArray_path, reset_memory()); 4642 } 4643 } 4644 // Otherwise, there are no card marks to worry about. 4645 // (We can dispense with card marks if we know the allocation 4646 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4647 // causes the non-eden paths to take compensating steps to 4648 // simulate a fresh allocation, so that no further 4649 // card marks are required in compiled code to initialize 4650 // the object.) 4651 4652 if (!stopped()) { 4653 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4654 4655 // Present the results of the copy. 4656 result_reg->init_req(_array_path, control()); 4657 result_val->init_req(_array_path, alloc_obj); 4658 result_i_o ->set_req(_array_path, i_o()); 4659 result_mem ->set_req(_array_path, reset_memory()); 4660 } 4661 } 4662 4663 // We only go to the instance fast case code if we pass a number of guards. 4664 // The paths which do not pass are accumulated in the slow_region. 4665 RegionNode* slow_region = new RegionNode(1); 4666 record_for_igvn(slow_region); 4667 if (!stopped()) { 4668 // It's an instance (we did array above). Make the slow-path tests. 4669 // If this is a virtual call, we generate a funny guard. We grab 4670 // the vtable entry corresponding to clone() from the target object. 4671 // If the target method which we are calling happens to be the 4672 // Object clone() method, we pass the guard. We do not need this 4673 // guard for non-virtual calls; the caller is known to be the native 4674 // Object clone(). 4675 if (is_virtual) { 4676 generate_virtual_guard(obj_klass, slow_region); 4677 } 4678 4679 // The object must be cloneable and must not have a finalizer. 4680 // Both of these conditions may be checked in a single test. 4681 // We could optimize the cloneable test further, but we don't care. 4682 generate_access_flags_guard(obj_klass, 4683 // Test both conditions: 4684 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4685 // Must be cloneable but not finalizer: 4686 JVM_ACC_IS_CLONEABLE, 4687 slow_region); 4688 } 4689 4690 if (!stopped()) { 4691 // It's an instance, and it passed the slow-path tests. 4692 PreserveJVMState pjvms(this); 4693 Node* obj_size = NULL; 4694 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4695 // is reexecuted if deoptimization occurs and there could be problems when merging 4696 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4697 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4698 4699 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4700 4701 // Present the results of the slow call. 4702 result_reg->init_req(_instance_path, control()); 4703 result_val->init_req(_instance_path, alloc_obj); 4704 result_i_o ->set_req(_instance_path, i_o()); 4705 result_mem ->set_req(_instance_path, reset_memory()); 4706 } 4707 4708 // Generate code for the slow case. We make a call to clone(). 4709 set_control(_gvn.transform(slow_region)); 4710 if (!stopped()) { 4711 PreserveJVMState pjvms(this); 4712 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4713 Node* slow_result = set_results_for_java_call(slow_call); 4714 // this->control() comes from set_results_for_java_call 4715 result_reg->init_req(_slow_path, control()); 4716 result_val->init_req(_slow_path, slow_result); 4717 result_i_o ->set_req(_slow_path, i_o()); 4718 result_mem ->set_req(_slow_path, reset_memory()); 4719 } 4720 4721 // Return the combined state. 4722 set_control( _gvn.transform(result_reg)); 4723 set_i_o( _gvn.transform(result_i_o)); 4724 set_all_memory( _gvn.transform(result_mem)); 4725 } // original reexecute is set back here 4726 4727 set_result(_gvn.transform(result_val)); 4728 return true; 4729 } 4730 4731 // If we have a tighly coupled allocation, the arraycopy may take care 4732 // of the array initialization. If one of the guards we insert between 4733 // the allocation and the arraycopy causes a deoptimization, an 4734 // unitialized array will escape the compiled method. To prevent that 4735 // we set the JVM state for uncommon traps between the allocation and 4736 // the arraycopy to the state before the allocation so, in case of 4737 // deoptimization, we'll reexecute the allocation and the 4738 // initialization. 4739 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4740 if (alloc != NULL) { 4741 ciMethod* trap_method = alloc->jvms()->method(); 4742 int trap_bci = alloc->jvms()->bci(); 4743 4744 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4745 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4746 // Make sure there's no store between the allocation and the 4747 // arraycopy otherwise visible side effects could be rexecuted 4748 // in case of deoptimization and cause incorrect execution. 4749 bool no_interfering_store = true; 4750 Node* mem = alloc->in(TypeFunc::Memory); 4751 if (mem->is_MergeMem()) { 4752 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4753 Node* n = mms.memory(); 4754 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4755 assert(n->is_Store(), "what else?"); 4756 no_interfering_store = false; 4757 break; 4758 } 4759 } 4760 } else { 4761 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4762 Node* n = mms.memory(); 4763 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4764 assert(n->is_Store(), "what else?"); 4765 no_interfering_store = false; 4766 break; 4767 } 4768 } 4769 } 4770 4771 if (no_interfering_store) { 4772 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4773 uint size = alloc->req(); 4774 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4775 old_jvms->set_map(sfpt); 4776 for (uint i = 0; i < size; i++) { 4777 sfpt->init_req(i, alloc->in(i)); 4778 } 4779 // re-push array length for deoptimization 4780 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4781 old_jvms->set_sp(old_jvms->sp()+1); 4782 old_jvms->set_monoff(old_jvms->monoff()+1); 4783 old_jvms->set_scloff(old_jvms->scloff()+1); 4784 old_jvms->set_endoff(old_jvms->endoff()+1); 4785 old_jvms->set_should_reexecute(true); 4786 4787 sfpt->set_i_o(map()->i_o()); 4788 sfpt->set_memory(map()->memory()); 4789 sfpt->set_control(map()->control()); 4790 4791 JVMState* saved_jvms = jvms(); 4792 saved_reexecute_sp = _reexecute_sp; 4793 4794 set_jvms(sfpt->jvms()); 4795 _reexecute_sp = jvms()->sp(); 4796 4797 return saved_jvms; 4798 } 4799 } 4800 } 4801 return NULL; 4802 } 4803 4804 // In case of a deoptimization, we restart execution at the 4805 // allocation, allocating a new array. We would leave an uninitialized 4806 // array in the heap that GCs wouldn't expect. Move the allocation 4807 // after the traps so we don't allocate the array if we 4808 // deoptimize. This is possible because tightly_coupled_allocation() 4809 // guarantees there's no observer of the allocated array at this point 4810 // and the control flow is simple enough. 4811 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4812 if (saved_jvms != NULL && !stopped()) { 4813 assert(alloc != NULL, "only with a tightly coupled allocation"); 4814 // restore JVM state to the state at the arraycopy 4815 saved_jvms->map()->set_control(map()->control()); 4816 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4817 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4818 // If we've improved the types of some nodes (null check) while 4819 // emitting the guards, propagate them to the current state 4820 map()->replaced_nodes().apply(saved_jvms->map()); 4821 set_jvms(saved_jvms); 4822 _reexecute_sp = saved_reexecute_sp; 4823 4824 // Remove the allocation from above the guards 4825 CallProjections callprojs; 4826 alloc->extract_projections(&callprojs, true); 4827 InitializeNode* init = alloc->initialization(); 4828 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4829 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4830 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4831 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4832 4833 // move the allocation here (after the guards) 4834 _gvn.hash_delete(alloc); 4835 alloc->set_req(TypeFunc::Control, control()); 4836 alloc->set_req(TypeFunc::I_O, i_o()); 4837 Node *mem = reset_memory(); 4838 set_all_memory(mem); 4839 alloc->set_req(TypeFunc::Memory, mem); 4840 set_control(init->proj_out(TypeFunc::Control)); 4841 set_i_o(callprojs.fallthrough_ioproj); 4842 4843 // Update memory as done in GraphKit::set_output_for_allocation() 4844 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4845 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4846 if (ary_type->isa_aryptr() && length_type != NULL) { 4847 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4848 } 4849 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4850 int elemidx = C->get_alias_index(telemref); 4851 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4852 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4853 4854 Node* allocx = _gvn.transform(alloc); 4855 assert(allocx == alloc, "where has the allocation gone?"); 4856 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4857 4858 _gvn.hash_delete(dest); 4859 dest->set_req(0, control()); 4860 Node* destx = _gvn.transform(dest); 4861 assert(destx == dest, "where has the allocation result gone?"); 4862 } 4863 } 4864 4865 4866 //------------------------------inline_arraycopy----------------------- 4867 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4868 // Object dest, int destPos, 4869 // int length); 4870 bool LibraryCallKit::inline_arraycopy() { 4871 // Get the arguments. 4872 Node* src = argument(0); // type: oop 4873 Node* src_offset = argument(1); // type: int 4874 Node* dest = argument(2); // type: oop 4875 Node* dest_offset = argument(3); // type: int 4876 Node* length = argument(4); // type: int 4877 4878 4879 // Check for allocation before we add nodes that would confuse 4880 // tightly_coupled_allocation() 4881 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4882 4883 int saved_reexecute_sp = -1; 4884 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4885 // See arraycopy_restore_alloc_state() comment 4886 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4887 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4888 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4889 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4890 4891 // The following tests must be performed 4892 // (1) src and dest are arrays. 4893 // (2) src and dest arrays must have elements of the same BasicType 4894 // (3) src and dest must not be null. 4895 // (4) src_offset must not be negative. 4896 // (5) dest_offset must not be negative. 4897 // (6) length must not be negative. 4898 // (7) src_offset + length must not exceed length of src. 4899 // (8) dest_offset + length must not exceed length of dest. 4900 // (9) each element of an oop array must be assignable 4901 4902 // (3) src and dest must not be null. 4903 // always do this here because we need the JVM state for uncommon traps 4904 Node* null_ctl = top(); 4905 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4906 assert(null_ctl->is_top(), "no null control here"); 4907 dest = null_check(dest, T_ARRAY); 4908 4909 if (!can_emit_guards) { 4910 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4911 // guards but the arraycopy node could still take advantage of a 4912 // tightly allocated allocation. tightly_coupled_allocation() is 4913 // called again to make sure it takes the null check above into 4914 // account: the null check is mandatory and if it caused an 4915 // uncommon trap to be emitted then the allocation can't be 4916 // considered tightly coupled in this context. 4917 alloc = tightly_coupled_allocation(dest, NULL); 4918 } 4919 4920 bool validated = false; 4921 4922 const Type* src_type = _gvn.type(src); 4923 const Type* dest_type = _gvn.type(dest); 4924 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4925 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4926 4927 // Do we have the type of src? 4928 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4929 // Do we have the type of dest? 4930 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4931 // Is the type for src from speculation? 4932 bool src_spec = false; 4933 // Is the type for dest from speculation? 4934 bool dest_spec = false; 4935 4936 if ((!has_src || !has_dest) && can_emit_guards) { 4937 // We don't have sufficient type information, let's see if 4938 // speculative types can help. We need to have types for both src 4939 // and dest so that it pays off. 4940 4941 // Do we already have or could we have type information for src 4942 bool could_have_src = has_src; 4943 // Do we already have or could we have type information for dest 4944 bool could_have_dest = has_dest; 4945 4946 ciKlass* src_k = NULL; 4947 if (!has_src) { 4948 src_k = src_type->speculative_type_not_null(); 4949 if (src_k != NULL && src_k->is_array_klass()) { 4950 could_have_src = true; 4951 } 4952 } 4953 4954 ciKlass* dest_k = NULL; 4955 if (!has_dest) { 4956 dest_k = dest_type->speculative_type_not_null(); 4957 if (dest_k != NULL && dest_k->is_array_klass()) { 4958 could_have_dest = true; 4959 } 4960 } 4961 4962 if (could_have_src && could_have_dest) { 4963 // This is going to pay off so emit the required guards 4964 if (!has_src) { 4965 src = maybe_cast_profiled_obj(src, src_k, true); 4966 src_type = _gvn.type(src); 4967 top_src = src_type->isa_aryptr(); 4968 has_src = (top_src != NULL && top_src->klass() != NULL); 4969 src_spec = true; 4970 } 4971 if (!has_dest) { 4972 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4973 dest_type = _gvn.type(dest); 4974 top_dest = dest_type->isa_aryptr(); 4975 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4976 dest_spec = true; 4977 } 4978 } 4979 } 4980 4981 if (has_src && has_dest && can_emit_guards) { 4982 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4983 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4984 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4985 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4986 4987 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4988 // If both arrays are object arrays then having the exact types 4989 // for both will remove the need for a subtype check at runtime 4990 // before the call and may make it possible to pick a faster copy 4991 // routine (without a subtype check on every element) 4992 // Do we have the exact type of src? 4993 bool could_have_src = src_spec; 4994 // Do we have the exact type of dest? 4995 bool could_have_dest = dest_spec; 4996 ciKlass* src_k = top_src->klass(); 4997 ciKlass* dest_k = top_dest->klass(); 4998 if (!src_spec) { 4999 src_k = src_type->speculative_type_not_null(); 5000 if (src_k != NULL && src_k->is_array_klass()) { 5001 could_have_src = true; 5002 } 5003 } 5004 if (!dest_spec) { 5005 dest_k = dest_type->speculative_type_not_null(); 5006 if (dest_k != NULL && dest_k->is_array_klass()) { 5007 could_have_dest = true; 5008 } 5009 } 5010 if (could_have_src && could_have_dest) { 5011 // If we can have both exact types, emit the missing guards 5012 if (could_have_src && !src_spec) { 5013 src = maybe_cast_profiled_obj(src, src_k, true); 5014 } 5015 if (could_have_dest && !dest_spec) { 5016 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5017 } 5018 } 5019 } 5020 } 5021 5022 ciMethod* trap_method = method(); 5023 int trap_bci = bci(); 5024 if (saved_jvms != NULL) { 5025 trap_method = alloc->jvms()->method(); 5026 trap_bci = alloc->jvms()->bci(); 5027 } 5028 5029 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5030 can_emit_guards && 5031 !src->is_top() && !dest->is_top()) { 5032 // validate arguments: enables transformation the ArrayCopyNode 5033 validated = true; 5034 5035 RegionNode* slow_region = new RegionNode(1); 5036 record_for_igvn(slow_region); 5037 5038 // (1) src and dest are arrays. 5039 generate_non_array_guard(load_object_klass(src), slow_region); 5040 generate_non_array_guard(load_object_klass(dest), slow_region); 5041 5042 // (2) src and dest arrays must have elements of the same BasicType 5043 // done at macro expansion or at Ideal transformation time 5044 5045 // (4) src_offset must not be negative. 5046 generate_negative_guard(src_offset, slow_region); 5047 5048 // (5) dest_offset must not be negative. 5049 generate_negative_guard(dest_offset, slow_region); 5050 5051 // (7) src_offset + length must not exceed length of src. 5052 generate_limit_guard(src_offset, length, 5053 load_array_length(src), 5054 slow_region); 5055 5056 // (8) dest_offset + length must not exceed length of dest. 5057 generate_limit_guard(dest_offset, length, 5058 load_array_length(dest), 5059 slow_region); 5060 5061 // (9) each element of an oop array must be assignable 5062 Node* src_klass = load_object_klass(src); 5063 Node* dest_klass = load_object_klass(dest); 5064 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5065 5066 if (not_subtype_ctrl != top()) { 5067 PreserveJVMState pjvms(this); 5068 set_control(not_subtype_ctrl); 5069 uncommon_trap(Deoptimization::Reason_intrinsic, 5070 Deoptimization::Action_make_not_entrant); 5071 assert(stopped(), "Should be stopped"); 5072 } 5073 { 5074 PreserveJVMState pjvms(this); 5075 set_control(_gvn.transform(slow_region)); 5076 uncommon_trap(Deoptimization::Reason_intrinsic, 5077 Deoptimization::Action_make_not_entrant); 5078 assert(stopped(), "Should be stopped"); 5079 } 5080 } 5081 5082 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5083 5084 if (stopped()) { 5085 return true; 5086 } 5087 5088 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5089 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5090 // so the compiler has a chance to eliminate them: during macro expansion, 5091 // we have to set their control (CastPP nodes are eliminated). 5092 load_object_klass(src), load_object_klass(dest), 5093 load_array_length(src), load_array_length(dest)); 5094 5095 ac->set_arraycopy(validated); 5096 5097 Node* n = _gvn.transform(ac); 5098 if (n == ac) { 5099 ac->connect_outputs(this); 5100 } else { 5101 assert(validated, "shouldn't transform if all arguments not validated"); 5102 set_all_memory(n); 5103 } 5104 5105 return true; 5106 } 5107 5108 5109 // Helper function which determines if an arraycopy immediately follows 5110 // an allocation, with no intervening tests or other escapes for the object. 5111 AllocateArrayNode* 5112 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5113 RegionNode* slow_region) { 5114 if (stopped()) return NULL; // no fast path 5115 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5116 5117 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5118 if (alloc == NULL) return NULL; 5119 5120 Node* rawmem = memory(Compile::AliasIdxRaw); 5121 // Is the allocation's memory state untouched? 5122 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5123 // Bail out if there have been raw-memory effects since the allocation. 5124 // (Example: There might have been a call or safepoint.) 5125 return NULL; 5126 } 5127 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5128 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5129 return NULL; 5130 } 5131 5132 // There must be no unexpected observers of this allocation. 5133 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5134 Node* obs = ptr->fast_out(i); 5135 if (obs != this->map()) { 5136 return NULL; 5137 } 5138 } 5139 5140 // This arraycopy must unconditionally follow the allocation of the ptr. 5141 Node* alloc_ctl = ptr->in(0); 5142 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5143 5144 Node* ctl = control(); 5145 while (ctl != alloc_ctl) { 5146 // There may be guards which feed into the slow_region. 5147 // Any other control flow means that we might not get a chance 5148 // to finish initializing the allocated object. 5149 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5150 IfNode* iff = ctl->in(0)->as_If(); 5151 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5152 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5153 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5154 ctl = iff->in(0); // This test feeds the known slow_region. 5155 continue; 5156 } 5157 // One more try: Various low-level checks bottom out in 5158 // uncommon traps. If the debug-info of the trap omits 5159 // any reference to the allocation, as we've already 5160 // observed, then there can be no objection to the trap. 5161 bool found_trap = false; 5162 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5163 Node* obs = not_ctl->fast_out(j); 5164 if (obs->in(0) == not_ctl && obs->is_Call() && 5165 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5166 found_trap = true; break; 5167 } 5168 } 5169 if (found_trap) { 5170 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5171 continue; 5172 } 5173 } 5174 return NULL; 5175 } 5176 5177 // If we get this far, we have an allocation which immediately 5178 // precedes the arraycopy, and we can take over zeroing the new object. 5179 // The arraycopy will finish the initialization, and provide 5180 // a new control state to which we will anchor the destination pointer. 5181 5182 return alloc; 5183 } 5184 5185 //-------------inline_encodeISOArray----------------------------------- 5186 // encode char[] to byte[] in ISO_8859_1 5187 bool LibraryCallKit::inline_encodeISOArray() { 5188 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5189 // no receiver since it is static method 5190 Node *src = argument(0); 5191 Node *src_offset = argument(1); 5192 Node *dst = argument(2); 5193 Node *dst_offset = argument(3); 5194 Node *length = argument(4); 5195 5196 const Type* src_type = src->Value(&_gvn); 5197 const Type* dst_type = dst->Value(&_gvn); 5198 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5199 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5200 if (top_src == NULL || top_src->klass() == NULL || 5201 top_dest == NULL || top_dest->klass() == NULL) { 5202 // failed array check 5203 return false; 5204 } 5205 5206 // Figure out the size and type of the elements we will be copying. 5207 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5208 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5209 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5210 return false; 5211 } 5212 5213 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5214 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5215 // 'src_start' points to src array + scaled offset 5216 // 'dst_start' points to dst array + scaled offset 5217 5218 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5219 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5220 enc = _gvn.transform(enc); 5221 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5222 set_memory(res_mem, mtype); 5223 set_result(enc); 5224 return true; 5225 } 5226 5227 //-------------inline_multiplyToLen----------------------------------- 5228 bool LibraryCallKit::inline_multiplyToLen() { 5229 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5230 5231 address stubAddr = StubRoutines::multiplyToLen(); 5232 if (stubAddr == NULL) { 5233 return false; // Intrinsic's stub is not implemented on this platform 5234 } 5235 const char* stubName = "multiplyToLen"; 5236 5237 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5238 5239 // no receiver because it is a static method 5240 Node* x = argument(0); 5241 Node* xlen = argument(1); 5242 Node* y = argument(2); 5243 Node* ylen = argument(3); 5244 Node* z = argument(4); 5245 5246 const Type* x_type = x->Value(&_gvn); 5247 const Type* y_type = y->Value(&_gvn); 5248 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5249 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5250 if (top_x == NULL || top_x->klass() == NULL || 5251 top_y == NULL || top_y->klass() == NULL) { 5252 // failed array check 5253 return false; 5254 } 5255 5256 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5257 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5258 if (x_elem != T_INT || y_elem != T_INT) { 5259 return false; 5260 } 5261 5262 // Set the original stack and the reexecute bit for the interpreter to reexecute 5263 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5264 // on the return from z array allocation in runtime. 5265 { PreserveReexecuteState preexecs(this); 5266 jvms()->set_should_reexecute(true); 5267 5268 Node* x_start = array_element_address(x, intcon(0), x_elem); 5269 Node* y_start = array_element_address(y, intcon(0), y_elem); 5270 // 'x_start' points to x array + scaled xlen 5271 // 'y_start' points to y array + scaled ylen 5272 5273 // Allocate the result array 5274 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5275 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5276 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5277 5278 IdealKit ideal(this); 5279 5280 #define __ ideal. 5281 Node* one = __ ConI(1); 5282 Node* zero = __ ConI(0); 5283 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5284 __ set(need_alloc, zero); 5285 __ set(z_alloc, z); 5286 __ if_then(z, BoolTest::eq, null()); { 5287 __ increment (need_alloc, one); 5288 } __ else_(); { 5289 // Update graphKit memory and control from IdealKit. 5290 sync_kit(ideal); 5291 Node* zlen_arg = load_array_length(z); 5292 // Update IdealKit memory and control from graphKit. 5293 __ sync_kit(this); 5294 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5295 __ increment (need_alloc, one); 5296 } __ end_if(); 5297 } __ end_if(); 5298 5299 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5300 // Update graphKit memory and control from IdealKit. 5301 sync_kit(ideal); 5302 Node * narr = new_array(klass_node, zlen, 1); 5303 // Update IdealKit memory and control from graphKit. 5304 __ sync_kit(this); 5305 __ set(z_alloc, narr); 5306 } __ end_if(); 5307 5308 sync_kit(ideal); 5309 z = __ value(z_alloc); 5310 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5311 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5312 // Final sync IdealKit and GraphKit. 5313 final_sync(ideal); 5314 #undef __ 5315 5316 Node* z_start = array_element_address(z, intcon(0), T_INT); 5317 5318 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5319 OptoRuntime::multiplyToLen_Type(), 5320 stubAddr, stubName, TypePtr::BOTTOM, 5321 x_start, xlen, y_start, ylen, z_start, zlen); 5322 } // original reexecute is set back here 5323 5324 C->set_has_split_ifs(true); // Has chance for split-if optimization 5325 set_result(z); 5326 return true; 5327 } 5328 5329 //-------------inline_squareToLen------------------------------------ 5330 bool LibraryCallKit::inline_squareToLen() { 5331 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5332 5333 address stubAddr = StubRoutines::squareToLen(); 5334 if (stubAddr == NULL) { 5335 return false; // Intrinsic's stub is not implemented on this platform 5336 } 5337 const char* stubName = "squareToLen"; 5338 5339 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5340 5341 Node* x = argument(0); 5342 Node* len = argument(1); 5343 Node* z = argument(2); 5344 Node* zlen = argument(3); 5345 5346 const Type* x_type = x->Value(&_gvn); 5347 const Type* z_type = z->Value(&_gvn); 5348 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5349 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5350 if (top_x == NULL || top_x->klass() == NULL || 5351 top_z == NULL || top_z->klass() == NULL) { 5352 // failed array check 5353 return false; 5354 } 5355 5356 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5357 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5358 if (x_elem != T_INT || z_elem != T_INT) { 5359 return false; 5360 } 5361 5362 5363 Node* x_start = array_element_address(x, intcon(0), x_elem); 5364 Node* z_start = array_element_address(z, intcon(0), z_elem); 5365 5366 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5367 OptoRuntime::squareToLen_Type(), 5368 stubAddr, stubName, TypePtr::BOTTOM, 5369 x_start, len, z_start, zlen); 5370 5371 set_result(z); 5372 return true; 5373 } 5374 5375 //-------------inline_mulAdd------------------------------------------ 5376 bool LibraryCallKit::inline_mulAdd() { 5377 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5378 5379 address stubAddr = StubRoutines::mulAdd(); 5380 if (stubAddr == NULL) { 5381 return false; // Intrinsic's stub is not implemented on this platform 5382 } 5383 const char* stubName = "mulAdd"; 5384 5385 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5386 5387 Node* out = argument(0); 5388 Node* in = argument(1); 5389 Node* offset = argument(2); 5390 Node* len = argument(3); 5391 Node* k = argument(4); 5392 5393 const Type* out_type = out->Value(&_gvn); 5394 const Type* in_type = in->Value(&_gvn); 5395 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5396 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5397 if (top_out == NULL || top_out->klass() == NULL || 5398 top_in == NULL || top_in->klass() == NULL) { 5399 // failed array check 5400 return false; 5401 } 5402 5403 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5404 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5405 if (out_elem != T_INT || in_elem != T_INT) { 5406 return false; 5407 } 5408 5409 Node* outlen = load_array_length(out); 5410 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5411 Node* out_start = array_element_address(out, intcon(0), out_elem); 5412 Node* in_start = array_element_address(in, intcon(0), in_elem); 5413 5414 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5415 OptoRuntime::mulAdd_Type(), 5416 stubAddr, stubName, TypePtr::BOTTOM, 5417 out_start,in_start, new_offset, len, k); 5418 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5419 set_result(result); 5420 return true; 5421 } 5422 5423 //-------------inline_montgomeryMultiply----------------------------------- 5424 bool LibraryCallKit::inline_montgomeryMultiply() { 5425 address stubAddr = StubRoutines::montgomeryMultiply(); 5426 if (stubAddr == NULL) { 5427 return false; // Intrinsic's stub is not implemented on this platform 5428 } 5429 5430 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5431 const char* stubName = "montgomery_square"; 5432 5433 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5434 5435 Node* a = argument(0); 5436 Node* b = argument(1); 5437 Node* n = argument(2); 5438 Node* len = argument(3); 5439 Node* inv = argument(4); 5440 Node* m = argument(6); 5441 5442 const Type* a_type = a->Value(&_gvn); 5443 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5444 const Type* b_type = b->Value(&_gvn); 5445 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5446 const Type* n_type = a->Value(&_gvn); 5447 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5448 const Type* m_type = a->Value(&_gvn); 5449 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5450 if (top_a == NULL || top_a->klass() == NULL || 5451 top_b == NULL || top_b->klass() == NULL || 5452 top_n == NULL || top_n->klass() == NULL || 5453 top_m == NULL || top_m->klass() == NULL) { 5454 // failed array check 5455 return false; 5456 } 5457 5458 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5459 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5460 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5461 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5462 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5463 return false; 5464 } 5465 5466 // Make the call 5467 { 5468 Node* a_start = array_element_address(a, intcon(0), a_elem); 5469 Node* b_start = array_element_address(b, intcon(0), b_elem); 5470 Node* n_start = array_element_address(n, intcon(0), n_elem); 5471 Node* m_start = array_element_address(m, intcon(0), m_elem); 5472 5473 Node* call = make_runtime_call(RC_LEAF, 5474 OptoRuntime::montgomeryMultiply_Type(), 5475 stubAddr, stubName, TypePtr::BOTTOM, 5476 a_start, b_start, n_start, len, inv, top(), 5477 m_start); 5478 set_result(m); 5479 } 5480 5481 return true; 5482 } 5483 5484 bool LibraryCallKit::inline_montgomerySquare() { 5485 address stubAddr = StubRoutines::montgomerySquare(); 5486 if (stubAddr == NULL) { 5487 return false; // Intrinsic's stub is not implemented on this platform 5488 } 5489 5490 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5491 const char* stubName = "montgomery_square"; 5492 5493 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5494 5495 Node* a = argument(0); 5496 Node* n = argument(1); 5497 Node* len = argument(2); 5498 Node* inv = argument(3); 5499 Node* m = argument(5); 5500 5501 const Type* a_type = a->Value(&_gvn); 5502 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5503 const Type* n_type = a->Value(&_gvn); 5504 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5505 const Type* m_type = a->Value(&_gvn); 5506 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5507 if (top_a == NULL || top_a->klass() == NULL || 5508 top_n == NULL || top_n->klass() == NULL || 5509 top_m == NULL || top_m->klass() == NULL) { 5510 // failed array check 5511 return false; 5512 } 5513 5514 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5515 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5516 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5517 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5518 return false; 5519 } 5520 5521 // Make the call 5522 { 5523 Node* a_start = array_element_address(a, intcon(0), a_elem); 5524 Node* n_start = array_element_address(n, intcon(0), n_elem); 5525 Node* m_start = array_element_address(m, intcon(0), m_elem); 5526 5527 Node* call = make_runtime_call(RC_LEAF, 5528 OptoRuntime::montgomerySquare_Type(), 5529 stubAddr, stubName, TypePtr::BOTTOM, 5530 a_start, n_start, len, inv, top(), 5531 m_start); 5532 set_result(m); 5533 } 5534 5535 return true; 5536 } 5537 5538 5539 /** 5540 * Calculate CRC32 for byte. 5541 * int java.util.zip.CRC32.update(int crc, int b) 5542 */ 5543 bool LibraryCallKit::inline_updateCRC32() { 5544 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5545 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5546 // no receiver since it is static method 5547 Node* crc = argument(0); // type: int 5548 Node* b = argument(1); // type: int 5549 5550 /* 5551 * int c = ~ crc; 5552 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5553 * b = b ^ (c >>> 8); 5554 * crc = ~b; 5555 */ 5556 5557 Node* M1 = intcon(-1); 5558 crc = _gvn.transform(new XorINode(crc, M1)); 5559 Node* result = _gvn.transform(new XorINode(crc, b)); 5560 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5561 5562 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5563 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5564 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5565 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5566 5567 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5568 result = _gvn.transform(new XorINode(crc, result)); 5569 result = _gvn.transform(new XorINode(result, M1)); 5570 set_result(result); 5571 return true; 5572 } 5573 5574 /** 5575 * Calculate CRC32 for byte[] array. 5576 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5577 */ 5578 bool LibraryCallKit::inline_updateBytesCRC32() { 5579 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5580 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5581 // no receiver since it is static method 5582 Node* crc = argument(0); // type: int 5583 Node* src = argument(1); // type: oop 5584 Node* offset = argument(2); // type: int 5585 Node* length = argument(3); // type: int 5586 5587 const Type* src_type = src->Value(&_gvn); 5588 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5589 if (top_src == NULL || top_src->klass() == NULL) { 5590 // failed array check 5591 return false; 5592 } 5593 5594 // Figure out the size and type of the elements we will be copying. 5595 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5596 if (src_elem != T_BYTE) { 5597 return false; 5598 } 5599 5600 // 'src_start' points to src array + scaled offset 5601 Node* src_start = array_element_address(src, offset, src_elem); 5602 5603 // We assume that range check is done by caller. 5604 // TODO: generate range check (offset+length < src.length) in debug VM. 5605 5606 // Call the stub. 5607 address stubAddr = StubRoutines::updateBytesCRC32(); 5608 const char *stubName = "updateBytesCRC32"; 5609 5610 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5611 stubAddr, stubName, TypePtr::BOTTOM, 5612 crc, src_start, length); 5613 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5614 set_result(result); 5615 return true; 5616 } 5617 5618 /** 5619 * Calculate CRC32 for ByteBuffer. 5620 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5621 */ 5622 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5623 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5624 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5625 // no receiver since it is static method 5626 Node* crc = argument(0); // type: int 5627 Node* src = argument(1); // type: long 5628 Node* offset = argument(3); // type: int 5629 Node* length = argument(4); // type: int 5630 5631 src = ConvL2X(src); // adjust Java long to machine word 5632 Node* base = _gvn.transform(new CastX2PNode(src)); 5633 offset = ConvI2X(offset); 5634 5635 // 'src_start' points to src array + scaled offset 5636 Node* src_start = basic_plus_adr(top(), base, offset); 5637 5638 // Call the stub. 5639 address stubAddr = StubRoutines::updateBytesCRC32(); 5640 const char *stubName = "updateBytesCRC32"; 5641 5642 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5643 stubAddr, stubName, TypePtr::BOTTOM, 5644 crc, src_start, length); 5645 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5646 set_result(result); 5647 return true; 5648 } 5649 5650 //------------------------------get_table_from_crc32c_class----------------------- 5651 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5652 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5653 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5654 5655 return table; 5656 } 5657 5658 //------------------------------inline_updateBytesCRC32C----------------------- 5659 // 5660 // Calculate CRC32C for byte[] array. 5661 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5662 // 5663 bool LibraryCallKit::inline_updateBytesCRC32C() { 5664 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5665 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5666 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5667 // no receiver since it is a static method 5668 Node* crc = argument(0); // type: int 5669 Node* src = argument(1); // type: oop 5670 Node* offset = argument(2); // type: int 5671 Node* end = argument(3); // type: int 5672 5673 Node* length = _gvn.transform(new SubINode(end, offset)); 5674 5675 const Type* src_type = src->Value(&_gvn); 5676 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5677 if (top_src == NULL || top_src->klass() == NULL) { 5678 // failed array check 5679 return false; 5680 } 5681 5682 // Figure out the size and type of the elements we will be copying. 5683 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5684 if (src_elem != T_BYTE) { 5685 return false; 5686 } 5687 5688 // 'src_start' points to src array + scaled offset 5689 Node* src_start = array_element_address(src, offset, src_elem); 5690 5691 // static final int[] byteTable in class CRC32C 5692 Node* table = get_table_from_crc32c_class(callee()->holder()); 5693 Node* table_start = array_element_address(table, intcon(0), T_INT); 5694 5695 // We assume that range check is done by caller. 5696 // TODO: generate range check (offset+length < src.length) in debug VM. 5697 5698 // Call the stub. 5699 address stubAddr = StubRoutines::updateBytesCRC32C(); 5700 const char *stubName = "updateBytesCRC32C"; 5701 5702 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5703 stubAddr, stubName, TypePtr::BOTTOM, 5704 crc, src_start, length, table_start); 5705 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5706 set_result(result); 5707 return true; 5708 } 5709 5710 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5711 // 5712 // Calculate CRC32C for DirectByteBuffer. 5713 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5714 // 5715 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5716 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5717 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5718 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5719 // no receiver since it is a static method 5720 Node* crc = argument(0); // type: int 5721 Node* src = argument(1); // type: long 5722 Node* offset = argument(3); // type: int 5723 Node* end = argument(4); // type: int 5724 5725 Node* length = _gvn.transform(new SubINode(end, offset)); 5726 5727 src = ConvL2X(src); // adjust Java long to machine word 5728 Node* base = _gvn.transform(new CastX2PNode(src)); 5729 offset = ConvI2X(offset); 5730 5731 // 'src_start' points to src array + scaled offset 5732 Node* src_start = basic_plus_adr(top(), base, offset); 5733 5734 // static final int[] byteTable in class CRC32C 5735 Node* table = get_table_from_crc32c_class(callee()->holder()); 5736 Node* table_start = array_element_address(table, intcon(0), T_INT); 5737 5738 // Call the stub. 5739 address stubAddr = StubRoutines::updateBytesCRC32C(); 5740 const char *stubName = "updateBytesCRC32C"; 5741 5742 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5743 stubAddr, stubName, TypePtr::BOTTOM, 5744 crc, src_start, length, table_start); 5745 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5746 set_result(result); 5747 return true; 5748 } 5749 5750 //------------------------------inline_updateBytesAdler32---------------------- 5751 // 5752 // Calculate Adler32 checksum for byte[] array. 5753 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5754 // 5755 bool LibraryCallKit::inline_updateBytesAdler32() { 5756 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5757 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5758 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5759 // no receiver since it is static method 5760 Node* crc = argument(0); // type: int 5761 Node* src = argument(1); // type: oop 5762 Node* offset = argument(2); // type: int 5763 Node* length = argument(3); // type: int 5764 5765 const Type* src_type = src->Value(&_gvn); 5766 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5767 if (top_src == NULL || top_src->klass() == NULL) { 5768 // failed array check 5769 return false; 5770 } 5771 5772 // Figure out the size and type of the elements we will be copying. 5773 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5774 if (src_elem != T_BYTE) { 5775 return false; 5776 } 5777 5778 // 'src_start' points to src array + scaled offset 5779 Node* src_start = array_element_address(src, offset, src_elem); 5780 5781 // We assume that range check is done by caller. 5782 // TODO: generate range check (offset+length < src.length) in debug VM. 5783 5784 // Call the stub. 5785 address stubAddr = StubRoutines::updateBytesAdler32(); 5786 const char *stubName = "updateBytesAdler32"; 5787 5788 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5789 stubAddr, stubName, TypePtr::BOTTOM, 5790 crc, src_start, length); 5791 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5792 set_result(result); 5793 return true; 5794 } 5795 5796 //------------------------------inline_updateByteBufferAdler32--------------- 5797 // 5798 // Calculate Adler32 checksum for DirectByteBuffer. 5799 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5800 // 5801 bool LibraryCallKit::inline_updateByteBufferAdler32() { 5802 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5803 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5804 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5805 // no receiver since it is static method 5806 Node* crc = argument(0); // type: int 5807 Node* src = argument(1); // type: long 5808 Node* offset = argument(3); // type: int 5809 Node* length = argument(4); // type: int 5810 5811 src = ConvL2X(src); // adjust Java long to machine word 5812 Node* base = _gvn.transform(new CastX2PNode(src)); 5813 offset = ConvI2X(offset); 5814 5815 // 'src_start' points to src array + scaled offset 5816 Node* src_start = basic_plus_adr(top(), base, offset); 5817 5818 // Call the stub. 5819 address stubAddr = StubRoutines::updateBytesAdler32(); 5820 const char *stubName = "updateBytesAdler32"; 5821 5822 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5823 stubAddr, stubName, TypePtr::BOTTOM, 5824 crc, src_start, length); 5825 5826 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5827 set_result(result); 5828 return true; 5829 } 5830 5831 //----------------------------inline_reference_get---------------------------- 5832 // public T java.lang.ref.Reference.get(); 5833 bool LibraryCallKit::inline_reference_get() { 5834 const int referent_offset = java_lang_ref_Reference::referent_offset; 5835 guarantee(referent_offset > 0, "should have already been set"); 5836 5837 // Get the argument: 5838 Node* reference_obj = null_check_receiver(); 5839 if (stopped()) return true; 5840 5841 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5842 5843 ciInstanceKlass* klass = env()->Object_klass(); 5844 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5845 5846 Node* no_ctrl = NULL; 5847 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5848 5849 // Use the pre-barrier to record the value in the referent field 5850 pre_barrier(false /* do_load */, 5851 control(), 5852 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5853 result /* pre_val */, 5854 T_OBJECT); 5855 5856 // Add memory barrier to prevent commoning reads from this field 5857 // across safepoint since GC can change its value. 5858 insert_mem_bar(Op_MemBarCPUOrder); 5859 5860 set_result(result); 5861 return true; 5862 } 5863 5864 5865 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5866 bool is_exact=true, bool is_static=false, 5867 ciInstanceKlass * fromKls=NULL) { 5868 if (fromKls == NULL) { 5869 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5870 assert(tinst != NULL, "obj is null"); 5871 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5872 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5873 fromKls = tinst->klass()->as_instance_klass(); 5874 } else { 5875 assert(is_static, "only for static field access"); 5876 } 5877 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5878 ciSymbol::make(fieldTypeString), 5879 is_static); 5880 5881 assert (field != NULL, "undefined field"); 5882 if (field == NULL) return (Node *) NULL; 5883 5884 if (is_static) { 5885 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5886 fromObj = makecon(tip); 5887 } 5888 5889 // Next code copied from Parse::do_get_xxx(): 5890 5891 // Compute address and memory type. 5892 int offset = field->offset_in_bytes(); 5893 bool is_vol = field->is_volatile(); 5894 ciType* field_klass = field->type(); 5895 assert(field_klass->is_loaded(), "should be loaded"); 5896 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5897 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5898 BasicType bt = field->layout_type(); 5899 5900 // Build the resultant type of the load 5901 const Type *type; 5902 if (bt == T_OBJECT) { 5903 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5904 } else { 5905 type = Type::get_const_basic_type(bt); 5906 } 5907 5908 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5909 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5910 } 5911 // Build the load. 5912 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5913 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5914 // If reference is volatile, prevent following memory ops from 5915 // floating up past the volatile read. Also prevents commoning 5916 // another volatile read. 5917 if (is_vol) { 5918 // Memory barrier includes bogus read of value to force load BEFORE membar 5919 insert_mem_bar(Op_MemBarAcquire, loadedField); 5920 } 5921 return loadedField; 5922 } 5923 5924 5925 //------------------------------inline_aescrypt_Block----------------------- 5926 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5927 address stubAddr; 5928 const char *stubName; 5929 assert(UseAES, "need AES instruction support"); 5930 5931 switch(id) { 5932 case vmIntrinsics::_aescrypt_encryptBlock: 5933 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5934 stubName = "aescrypt_encryptBlock"; 5935 break; 5936 case vmIntrinsics::_aescrypt_decryptBlock: 5937 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5938 stubName = "aescrypt_decryptBlock"; 5939 break; 5940 } 5941 if (stubAddr == NULL) return false; 5942 5943 Node* aescrypt_object = argument(0); 5944 Node* src = argument(1); 5945 Node* src_offset = argument(2); 5946 Node* dest = argument(3); 5947 Node* dest_offset = argument(4); 5948 5949 // (1) src and dest are arrays. 5950 const Type* src_type = src->Value(&_gvn); 5951 const Type* dest_type = dest->Value(&_gvn); 5952 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5953 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5954 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5955 5956 // for the quick and dirty code we will skip all the checks. 5957 // we are just trying to get the call to be generated. 5958 Node* src_start = src; 5959 Node* dest_start = dest; 5960 if (src_offset != NULL || dest_offset != NULL) { 5961 assert(src_offset != NULL && dest_offset != NULL, ""); 5962 src_start = array_element_address(src, src_offset, T_BYTE); 5963 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5964 } 5965 5966 // now need to get the start of its expanded key array 5967 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5968 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5969 if (k_start == NULL) return false; 5970 5971 if (Matcher::pass_original_key_for_aes()) { 5972 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5973 // compatibility issues between Java key expansion and SPARC crypto instructions 5974 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5975 if (original_k_start == NULL) return false; 5976 5977 // Call the stub. 5978 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5979 stubAddr, stubName, TypePtr::BOTTOM, 5980 src_start, dest_start, k_start, original_k_start); 5981 } else { 5982 // Call the stub. 5983 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5984 stubAddr, stubName, TypePtr::BOTTOM, 5985 src_start, dest_start, k_start); 5986 } 5987 5988 return true; 5989 } 5990 5991 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5992 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5993 address stubAddr; 5994 const char *stubName; 5995 5996 assert(UseAES, "need AES instruction support"); 5997 5998 switch(id) { 5999 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6000 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6001 stubName = "cipherBlockChaining_encryptAESCrypt"; 6002 break; 6003 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6004 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6005 stubName = "cipherBlockChaining_decryptAESCrypt"; 6006 break; 6007 } 6008 if (stubAddr == NULL) return false; 6009 6010 Node* cipherBlockChaining_object = argument(0); 6011 Node* src = argument(1); 6012 Node* src_offset = argument(2); 6013 Node* len = argument(3); 6014 Node* dest = argument(4); 6015 Node* dest_offset = argument(5); 6016 6017 // (1) src and dest are arrays. 6018 const Type* src_type = src->Value(&_gvn); 6019 const Type* dest_type = dest->Value(&_gvn); 6020 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6021 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6022 assert (top_src != NULL && top_src->klass() != NULL 6023 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6024 6025 // checks are the responsibility of the caller 6026 Node* src_start = src; 6027 Node* dest_start = dest; 6028 if (src_offset != NULL || dest_offset != NULL) { 6029 assert(src_offset != NULL && dest_offset != NULL, ""); 6030 src_start = array_element_address(src, src_offset, T_BYTE); 6031 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6032 } 6033 6034 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6035 // (because of the predicated logic executed earlier). 6036 // so we cast it here safely. 6037 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6038 6039 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6040 if (embeddedCipherObj == NULL) return false; 6041 6042 // cast it to what we know it will be at runtime 6043 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6044 assert(tinst != NULL, "CBC obj is null"); 6045 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6046 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6047 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6048 6049 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6050 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6051 const TypeOopPtr* xtype = aklass->as_instance_type(); 6052 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6053 aescrypt_object = _gvn.transform(aescrypt_object); 6054 6055 // we need to get the start of the aescrypt_object's expanded key array 6056 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6057 if (k_start == NULL) return false; 6058 6059 // similarly, get the start address of the r vector 6060 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6061 if (objRvec == NULL) return false; 6062 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6063 6064 Node* cbcCrypt; 6065 if (Matcher::pass_original_key_for_aes()) { 6066 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6067 // compatibility issues between Java key expansion and SPARC crypto instructions 6068 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6069 if (original_k_start == NULL) return false; 6070 6071 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6072 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6073 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6074 stubAddr, stubName, TypePtr::BOTTOM, 6075 src_start, dest_start, k_start, r_start, len, original_k_start); 6076 } else { 6077 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6078 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6079 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6080 stubAddr, stubName, TypePtr::BOTTOM, 6081 src_start, dest_start, k_start, r_start, len); 6082 } 6083 6084 // return cipher length (int) 6085 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6086 set_result(retvalue); 6087 return true; 6088 } 6089 6090 //------------------------------get_key_start_from_aescrypt_object----------------------- 6091 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6092 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6093 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6094 if (objAESCryptKey == NULL) return (Node *) NULL; 6095 6096 // now have the array, need to get the start address of the K array 6097 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6098 return k_start; 6099 } 6100 6101 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6102 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6103 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6104 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6105 if (objAESCryptKey == NULL) return (Node *) NULL; 6106 6107 // now have the array, need to get the start address of the lastKey array 6108 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6109 return original_k_start; 6110 } 6111 6112 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6113 // Return node representing slow path of predicate check. 6114 // the pseudo code we want to emulate with this predicate is: 6115 // for encryption: 6116 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6117 // for decryption: 6118 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6119 // note cipher==plain is more conservative than the original java code but that's OK 6120 // 6121 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6122 // The receiver was checked for NULL already. 6123 Node* objCBC = argument(0); 6124 6125 // Load embeddedCipher field of CipherBlockChaining object. 6126 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6127 6128 // get AESCrypt klass for instanceOf check 6129 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6130 // will have same classloader as CipherBlockChaining object 6131 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6132 assert(tinst != NULL, "CBCobj is null"); 6133 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6134 6135 // we want to do an instanceof comparison against the AESCrypt class 6136 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6137 if (!klass_AESCrypt->is_loaded()) { 6138 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6139 Node* ctrl = control(); 6140 set_control(top()); // no regular fast path 6141 return ctrl; 6142 } 6143 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6144 6145 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6146 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6147 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6148 6149 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6150 6151 // for encryption, we are done 6152 if (!decrypting) 6153 return instof_false; // even if it is NULL 6154 6155 // for decryption, we need to add a further check to avoid 6156 // taking the intrinsic path when cipher and plain are the same 6157 // see the original java code for why. 6158 RegionNode* region = new RegionNode(3); 6159 region->init_req(1, instof_false); 6160 Node* src = argument(1); 6161 Node* dest = argument(4); 6162 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6163 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6164 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6165 region->init_req(2, src_dest_conjoint); 6166 6167 record_for_igvn(region); 6168 return _gvn.transform(region); 6169 } 6170 6171 //------------------------------inline_ghash_processBlocks 6172 bool LibraryCallKit::inline_ghash_processBlocks() { 6173 address stubAddr; 6174 const char *stubName; 6175 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6176 6177 stubAddr = StubRoutines::ghash_processBlocks(); 6178 stubName = "ghash_processBlocks"; 6179 6180 Node* data = argument(0); 6181 Node* offset = argument(1); 6182 Node* len = argument(2); 6183 Node* state = argument(3); 6184 Node* subkeyH = argument(4); 6185 6186 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6187 assert(state_start, "state is NULL"); 6188 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6189 assert(subkeyH_start, "subkeyH is NULL"); 6190 Node* data_start = array_element_address(data, offset, T_BYTE); 6191 assert(data_start, "data is NULL"); 6192 6193 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6194 OptoRuntime::ghash_processBlocks_Type(), 6195 stubAddr, stubName, TypePtr::BOTTOM, 6196 state_start, subkeyH_start, data_start, len); 6197 return true; 6198 } 6199 6200 //------------------------------inline_sha_implCompress----------------------- 6201 // 6202 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6203 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6204 // 6205 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6206 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6207 // 6208 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6209 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6210 // 6211 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6212 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6213 6214 Node* sha_obj = argument(0); 6215 Node* src = argument(1); // type oop 6216 Node* ofs = argument(2); // type int 6217 6218 const Type* src_type = src->Value(&_gvn); 6219 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6220 if (top_src == NULL || top_src->klass() == NULL) { 6221 // failed array check 6222 return false; 6223 } 6224 // Figure out the size and type of the elements we will be copying. 6225 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6226 if (src_elem != T_BYTE) { 6227 return false; 6228 } 6229 // 'src_start' points to src array + offset 6230 Node* src_start = array_element_address(src, ofs, src_elem); 6231 Node* state = NULL; 6232 address stubAddr; 6233 const char *stubName; 6234 6235 switch(id) { 6236 case vmIntrinsics::_sha_implCompress: 6237 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6238 state = get_state_from_sha_object(sha_obj); 6239 stubAddr = StubRoutines::sha1_implCompress(); 6240 stubName = "sha1_implCompress"; 6241 break; 6242 case vmIntrinsics::_sha2_implCompress: 6243 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6244 state = get_state_from_sha_object(sha_obj); 6245 stubAddr = StubRoutines::sha256_implCompress(); 6246 stubName = "sha256_implCompress"; 6247 break; 6248 case vmIntrinsics::_sha5_implCompress: 6249 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6250 state = get_state_from_sha5_object(sha_obj); 6251 stubAddr = StubRoutines::sha512_implCompress(); 6252 stubName = "sha512_implCompress"; 6253 break; 6254 default: 6255 fatal_unexpected_iid(id); 6256 return false; 6257 } 6258 if (state == NULL) return false; 6259 6260 // Call the stub. 6261 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6262 stubAddr, stubName, TypePtr::BOTTOM, 6263 src_start, state); 6264 6265 return true; 6266 } 6267 6268 //------------------------------inline_digestBase_implCompressMB----------------------- 6269 // 6270 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6271 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6272 // 6273 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6274 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6275 "need SHA1/SHA256/SHA512 instruction support"); 6276 assert((uint)predicate < 3, "sanity"); 6277 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6278 6279 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6280 Node* src = argument(1); // byte[] array 6281 Node* ofs = argument(2); // type int 6282 Node* limit = argument(3); // type int 6283 6284 const Type* src_type = src->Value(&_gvn); 6285 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6286 if (top_src == NULL || top_src->klass() == NULL) { 6287 // failed array check 6288 return false; 6289 } 6290 // Figure out the size and type of the elements we will be copying. 6291 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6292 if (src_elem != T_BYTE) { 6293 return false; 6294 } 6295 // 'src_start' points to src array + offset 6296 Node* src_start = array_element_address(src, ofs, src_elem); 6297 6298 const char* klass_SHA_name = NULL; 6299 const char* stub_name = NULL; 6300 address stub_addr = NULL; 6301 bool long_state = false; 6302 6303 switch (predicate) { 6304 case 0: 6305 if (UseSHA1Intrinsics) { 6306 klass_SHA_name = "sun/security/provider/SHA"; 6307 stub_name = "sha1_implCompressMB"; 6308 stub_addr = StubRoutines::sha1_implCompressMB(); 6309 } 6310 break; 6311 case 1: 6312 if (UseSHA256Intrinsics) { 6313 klass_SHA_name = "sun/security/provider/SHA2"; 6314 stub_name = "sha256_implCompressMB"; 6315 stub_addr = StubRoutines::sha256_implCompressMB(); 6316 } 6317 break; 6318 case 2: 6319 if (UseSHA512Intrinsics) { 6320 klass_SHA_name = "sun/security/provider/SHA5"; 6321 stub_name = "sha512_implCompressMB"; 6322 stub_addr = StubRoutines::sha512_implCompressMB(); 6323 long_state = true; 6324 } 6325 break; 6326 default: 6327 fatal("unknown SHA intrinsic predicate: %d", predicate); 6328 } 6329 if (klass_SHA_name != NULL) { 6330 // get DigestBase klass to lookup for SHA klass 6331 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6332 assert(tinst != NULL, "digestBase_obj is not instance???"); 6333 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6334 6335 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6336 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6337 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6338 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6339 } 6340 return false; 6341 } 6342 //------------------------------inline_sha_implCompressMB----------------------- 6343 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6344 bool long_state, address stubAddr, const char *stubName, 6345 Node* src_start, Node* ofs, Node* limit) { 6346 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6347 const TypeOopPtr* xtype = aklass->as_instance_type(); 6348 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6349 sha_obj = _gvn.transform(sha_obj); 6350 6351 Node* state; 6352 if (long_state) { 6353 state = get_state_from_sha5_object(sha_obj); 6354 } else { 6355 state = get_state_from_sha_object(sha_obj); 6356 } 6357 if (state == NULL) return false; 6358 6359 // Call the stub. 6360 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6361 OptoRuntime::digestBase_implCompressMB_Type(), 6362 stubAddr, stubName, TypePtr::BOTTOM, 6363 src_start, state, ofs, limit); 6364 // return ofs (int) 6365 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6366 set_result(result); 6367 6368 return true; 6369 } 6370 6371 //------------------------------get_state_from_sha_object----------------------- 6372 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6373 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6374 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6375 if (sha_state == NULL) return (Node *) NULL; 6376 6377 // now have the array, need to get the start address of the state array 6378 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6379 return state; 6380 } 6381 6382 //------------------------------get_state_from_sha5_object----------------------- 6383 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6384 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6385 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6386 if (sha_state == NULL) return (Node *) NULL; 6387 6388 // now have the array, need to get the start address of the state array 6389 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6390 return state; 6391 } 6392 6393 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6394 // Return node representing slow path of predicate check. 6395 // the pseudo code we want to emulate with this predicate is: 6396 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6397 // 6398 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6399 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6400 "need SHA1/SHA256/SHA512 instruction support"); 6401 assert((uint)predicate < 3, "sanity"); 6402 6403 // The receiver was checked for NULL already. 6404 Node* digestBaseObj = argument(0); 6405 6406 // get DigestBase klass for instanceOf check 6407 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6408 assert(tinst != NULL, "digestBaseObj is null"); 6409 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6410 6411 const char* klass_SHA_name = NULL; 6412 switch (predicate) { 6413 case 0: 6414 if (UseSHA1Intrinsics) { 6415 // we want to do an instanceof comparison against the SHA class 6416 klass_SHA_name = "sun/security/provider/SHA"; 6417 } 6418 break; 6419 case 1: 6420 if (UseSHA256Intrinsics) { 6421 // we want to do an instanceof comparison against the SHA2 class 6422 klass_SHA_name = "sun/security/provider/SHA2"; 6423 } 6424 break; 6425 case 2: 6426 if (UseSHA512Intrinsics) { 6427 // we want to do an instanceof comparison against the SHA5 class 6428 klass_SHA_name = "sun/security/provider/SHA5"; 6429 } 6430 break; 6431 default: 6432 fatal("unknown SHA intrinsic predicate: %d", predicate); 6433 } 6434 6435 ciKlass* klass_SHA = NULL; 6436 if (klass_SHA_name != NULL) { 6437 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6438 } 6439 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6440 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6441 Node* ctrl = control(); 6442 set_control(top()); // no intrinsic path 6443 return ctrl; 6444 } 6445 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6446 6447 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6448 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6449 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6450 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6451 6452 return instof_false; // even if it is NULL 6453 } 6454 6455 bool LibraryCallKit::inline_profileBoolean() { 6456 Node* counts = argument(1); 6457 const TypeAryPtr* ary = NULL; 6458 ciArray* aobj = NULL; 6459 if (counts->is_Con() 6460 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6461 && (aobj = ary->const_oop()->as_array()) != NULL 6462 && (aobj->length() == 2)) { 6463 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6464 jint false_cnt = aobj->element_value(0).as_int(); 6465 jint true_cnt = aobj->element_value(1).as_int(); 6466 6467 if (C->log() != NULL) { 6468 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6469 false_cnt, true_cnt); 6470 } 6471 6472 if (false_cnt + true_cnt == 0) { 6473 // According to profile, never executed. 6474 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6475 Deoptimization::Action_reinterpret); 6476 return true; 6477 } 6478 6479 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6480 // is a number of each value occurrences. 6481 Node* result = argument(0); 6482 if (false_cnt == 0 || true_cnt == 0) { 6483 // According to profile, one value has been never seen. 6484 int expected_val = (false_cnt == 0) ? 1 : 0; 6485 6486 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6487 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6488 6489 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6490 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6491 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6492 6493 { // Slow path: uncommon trap for never seen value and then reexecute 6494 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6495 // the value has been seen at least once. 6496 PreserveJVMState pjvms(this); 6497 PreserveReexecuteState preexecs(this); 6498 jvms()->set_should_reexecute(true); 6499 6500 set_control(slow_path); 6501 set_i_o(i_o()); 6502 6503 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6504 Deoptimization::Action_reinterpret); 6505 } 6506 // The guard for never seen value enables sharpening of the result and 6507 // returning a constant. It allows to eliminate branches on the same value 6508 // later on. 6509 set_control(fast_path); 6510 result = intcon(expected_val); 6511 } 6512 // Stop profiling. 6513 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6514 // By replacing method body with profile data (represented as ProfileBooleanNode 6515 // on IR level) we effectively disable profiling. 6516 // It enables full speed execution once optimized code is generated. 6517 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6518 C->record_for_igvn(profile); 6519 set_result(profile); 6520 return true; 6521 } else { 6522 // Continue profiling. 6523 // Profile data isn't available at the moment. So, execute method's bytecode version. 6524 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6525 // is compiled and counters aren't available since corresponding MethodHandle 6526 // isn't a compile-time constant. 6527 return false; 6528 } 6529 } 6530 6531 bool LibraryCallKit::inline_isCompileConstant() { 6532 Node* n = argument(0); 6533 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6534 return true; 6535 }