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