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