1 /* 2 * Copyright (c) 1999, 2020, 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 "ci/ciUtilities.inline.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "classfile/vmSymbols.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/compileLog.hpp" 32 #include "gc/shared/barrierSet.hpp" 33 #include "jfr/support/jfrIntrinsics.hpp" 34 #include "memory/resourceArea.hpp" 35 #include "oops/klass.inline.hpp" 36 #include "oops/objArrayKlass.hpp" 37 #include "opto/addnode.hpp" 38 #include "opto/arraycopynode.hpp" 39 #include "opto/c2compiler.hpp" 40 #include "opto/callGenerator.hpp" 41 #include "opto/castnode.hpp" 42 #include "opto/cfgnode.hpp" 43 #include "opto/convertnode.hpp" 44 #include "opto/countbitsnode.hpp" 45 #include "opto/intrinsicnode.hpp" 46 #include "opto/idealKit.hpp" 47 #include "opto/mathexactnode.hpp" 48 #include "opto/movenode.hpp" 49 #include "opto/mulnode.hpp" 50 #include "opto/narrowptrnode.hpp" 51 #include "opto/opaquenode.hpp" 52 #include "opto/parse.hpp" 53 #include "opto/runtime.hpp" 54 #include "opto/rootnode.hpp" 55 #include "opto/subnode.hpp" 56 #include "prims/nativeLookup.hpp" 57 #include "prims/unsafe.hpp" 58 #include "runtime/objectMonitor.hpp" 59 #include "runtime/sharedRuntime.hpp" 60 #include "utilities/macros.hpp" 61 #include "utilities/powerOfTwo.hpp" 62 63 class LibraryIntrinsic : public InlineCallGenerator { 64 // Extend the set of intrinsics known to the runtime: 65 public: 66 private: 67 bool _is_virtual; 68 bool _does_virtual_dispatch; 69 int8_t _predicates_count; // Intrinsic is predicated by several conditions 70 int8_t _last_predicate; // Last generated predicate 71 vmIntrinsics::ID _intrinsic_id; 72 73 public: 74 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 75 : InlineCallGenerator(m), 76 _is_virtual(is_virtual), 77 _does_virtual_dispatch(does_virtual_dispatch), 78 _predicates_count((int8_t)predicates_count), 79 _last_predicate((int8_t)-1), 80 _intrinsic_id(id) 81 { 82 } 83 virtual bool is_intrinsic() const { return true; } 84 virtual bool is_virtual() const { return _is_virtual; } 85 virtual bool is_predicated() const { return _predicates_count > 0; } 86 virtual int predicates_count() const { return _predicates_count; } 87 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 88 virtual JVMState* generate(JVMState* jvms); 89 virtual Node* generate_predicate(JVMState* jvms, int predicate); 90 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 91 }; 92 93 94 // Local helper class for LibraryIntrinsic: 95 class LibraryCallKit : public GraphKit { 96 private: 97 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 98 Node* _result; // the result node, if any 99 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 100 101 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type); 102 103 public: 104 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 105 : GraphKit(jvms), 106 _intrinsic(intrinsic), 107 _result(NULL) 108 { 109 // Check if this is a root compile. In that case we don't have a caller. 110 if (!jvms->has_method()) { 111 _reexecute_sp = sp(); 112 } else { 113 // Find out how many arguments the interpreter needs when deoptimizing 114 // and save the stack pointer value so it can used by uncommon_trap. 115 // We find the argument count by looking at the declared signature. 116 bool ignored_will_link; 117 ciSignature* declared_signature = NULL; 118 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 119 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 120 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 121 } 122 } 123 124 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 125 126 ciMethod* caller() const { return jvms()->method(); } 127 int bci() const { return jvms()->bci(); } 128 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 129 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 130 ciMethod* callee() const { return _intrinsic->method(); } 131 132 bool try_to_inline(int predicate); 133 Node* try_to_predicate(int predicate); 134 135 void push_result() { 136 // Push the result onto the stack. 137 if (!stopped() && result() != NULL) { 138 BasicType bt = result()->bottom_type()->basic_type(); 139 push_node(bt, result()); 140 } 141 } 142 143 private: 144 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 145 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 146 } 147 148 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 149 void set_result(RegionNode* region, PhiNode* value); 150 Node* result() { return _result; } 151 152 virtual int reexecute_sp() { return _reexecute_sp; } 153 154 // Helper functions to inline natives 155 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 156 Node* generate_slow_guard(Node* test, RegionNode* region); 157 Node* generate_fair_guard(Node* test, RegionNode* region); 158 Node* generate_negative_guard(Node* index, RegionNode* region, 159 // resulting CastII of index: 160 Node* *pos_index = NULL); 161 Node* generate_limit_guard(Node* offset, Node* subseq_length, 162 Node* array_length, 163 RegionNode* region); 164 void generate_string_range_check(Node* array, Node* offset, 165 Node* length, bool char_count); 166 Node* generate_current_thread(Node* &tls_output); 167 Node* load_mirror_from_klass(Node* klass); 168 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 169 RegionNode* region, int null_path, 170 int offset); 171 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 172 RegionNode* region, int null_path) { 173 int offset = java_lang_Class::klass_offset(); 174 return load_klass_from_mirror_common(mirror, never_see_null, 175 region, null_path, 176 offset); 177 } 178 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 179 RegionNode* region, int null_path) { 180 int offset = java_lang_Class::array_klass_offset(); 181 return load_klass_from_mirror_common(mirror, never_see_null, 182 region, null_path, 183 offset); 184 } 185 Node* generate_access_flags_guard(Node* kls, 186 int modifier_mask, int modifier_bits, 187 RegionNode* region); 188 Node* generate_interface_guard(Node* kls, RegionNode* region); 189 Node* generate_hidden_class_guard(Node* kls, RegionNode* region); 190 Node* generate_array_guard(Node* kls, RegionNode* region) { 191 return generate_array_guard_common(kls, region, false, false); 192 } 193 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 194 return generate_array_guard_common(kls, region, false, true); 195 } 196 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 197 return generate_array_guard_common(kls, region, true, false); 198 } 199 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 200 return generate_array_guard_common(kls, region, true, true); 201 } 202 Node* generate_array_guard_common(Node* kls, RegionNode* region, 203 bool obj_array, bool not_array); 204 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 205 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 206 bool is_virtual = false, bool is_static = false); 207 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 208 return generate_method_call(method_id, false, true); 209 } 210 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 211 return generate_method_call(method_id, true, false); 212 } 213 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 214 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 215 216 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 217 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 218 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 219 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 220 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 221 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); 222 bool inline_string_indexOfChar(); 223 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 224 bool inline_string_toBytesU(); 225 bool inline_string_getCharsU(); 226 bool inline_string_copy(bool compress); 227 bool inline_string_char_access(bool is_store); 228 Node* round_double_node(Node* n); 229 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 230 bool inline_math_native(vmIntrinsics::ID id); 231 bool inline_math(vmIntrinsics::ID id); 232 bool inline_double_math(vmIntrinsics::ID id); 233 template <typename OverflowOp> 234 bool inline_math_overflow(Node* arg1, Node* arg2); 235 void inline_math_mathExact(Node* math, Node* test); 236 bool inline_math_addExactI(bool is_increment); 237 bool inline_math_addExactL(bool is_increment); 238 bool inline_math_multiplyExactI(); 239 bool inline_math_multiplyExactL(); 240 bool inline_math_multiplyHigh(); 241 bool inline_math_negateExactI(); 242 bool inline_math_negateExactL(); 243 bool inline_math_subtractExactI(bool is_decrement); 244 bool inline_math_subtractExactL(bool is_decrement); 245 bool inline_min_max(vmIntrinsics::ID id); 246 bool inline_notify(vmIntrinsics::ID id); 247 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 248 // This returns Type::AnyPtr, RawPtr, or OopPtr. 249 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type); 250 Node* make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type = T_ILLEGAL, bool can_cast = false); 251 252 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind; 253 DecoratorSet mo_decorator_for_access_kind(AccessKind kind); 254 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned); 255 static bool klass_needs_init_guard(Node* kls); 256 bool inline_unsafe_allocate(); 257 bool inline_unsafe_newArray(bool uninitialized); 258 bool inline_unsafe_writeback0(); 259 bool inline_unsafe_writebackSync0(bool is_pre); 260 bool inline_unsafe_copyMemory(); 261 bool inline_native_currentThread(); 262 263 bool inline_native_time_funcs(address method, const char* funcName); 264 #ifdef JFR_HAVE_INTRINSICS 265 bool inline_native_classID(); 266 bool inline_native_getEventWriter(); 267 #endif 268 bool inline_native_Class_query(vmIntrinsics::ID id); 269 bool inline_native_subtype_check(); 270 bool inline_native_getLength(); 271 bool inline_array_copyOf(bool is_copyOfRange); 272 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 273 bool inline_preconditions_checkIndex(); 274 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array); 275 bool inline_native_clone(bool is_virtual); 276 bool inline_native_Reflection_getCallerClass(); 277 // Helper function for inlining native object hash method 278 bool inline_native_hashcode(bool is_virtual, bool is_static); 279 bool inline_native_getClass(); 280 281 // Helper functions for inlining arraycopy 282 bool inline_arraycopy(); 283 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 284 RegionNode* slow_region); 285 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 286 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp, 287 uint new_idx); 288 289 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind; 290 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind); 291 bool inline_unsafe_fence(vmIntrinsics::ID id); 292 bool inline_onspinwait(); 293 bool inline_fp_conversions(vmIntrinsics::ID id); 294 bool inline_number_methods(vmIntrinsics::ID id); 295 bool inline_reference_get(); 296 bool inline_Class_cast(); 297 bool inline_aescrypt_Block(vmIntrinsics::ID id); 298 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 299 bool inline_electronicCodeBook_AESCrypt(vmIntrinsics::ID id); 300 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 301 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 302 Node* inline_electronicCodeBook_AESCrypt_predicate(bool decrypting); 303 Node* inline_counterMode_AESCrypt_predicate(); 304 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 305 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 306 bool inline_ghash_processBlocks(); 307 bool inline_base64_encodeBlock(); 308 bool inline_base64_decodeBlock(); 309 bool inline_digestBase_implCompress(vmIntrinsics::ID id); 310 bool inline_sha_implCompress(vmIntrinsics::ID id); 311 bool inline_digestBase_implCompressMB(int predicate); 312 bool inline_digestBase_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass, 313 bool long_state, address stubAddr, const char *stubName, 314 Node* src_start, Node* ofs, Node* limit); 315 Node* get_state_from_digest_object(Node *digestBase_object); 316 Node* get_long_state_from_digest_object(Node *digestBase_object); 317 Node* inline_digestBase_implCompressMB_predicate(int predicate); 318 bool inline_encodeISOArray(); 319 bool inline_updateCRC32(); 320 bool inline_updateBytesCRC32(); 321 bool inline_updateByteBufferCRC32(); 322 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 323 bool inline_updateBytesCRC32C(); 324 bool inline_updateDirectByteBufferCRC32C(); 325 bool inline_updateBytesAdler32(); 326 bool inline_updateByteBufferAdler32(); 327 bool inline_multiplyToLen(); 328 bool inline_hasNegatives(); 329 bool inline_squareToLen(); 330 bool inline_mulAdd(); 331 bool inline_montgomeryMultiply(); 332 bool inline_montgomerySquare(); 333 bool inline_bigIntegerShift(bool isRightShift); 334 bool inline_vectorizedMismatch(); 335 bool inline_fma(vmIntrinsics::ID id); 336 bool inline_character_compare(vmIntrinsics::ID id); 337 bool inline_fp_min_max(vmIntrinsics::ID id); 338 339 bool inline_profileBoolean(); 340 bool inline_isCompileConstant(); 341 void clear_upper_avx() { 342 #ifdef X86 343 if (UseAVX >= 2) { 344 C->set_clear_upper_avx(true); 345 } 346 #endif 347 } 348 }; 349 350 //---------------------------make_vm_intrinsic---------------------------- 351 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 352 vmIntrinsics::ID id = m->intrinsic_id(); 353 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 354 355 if (!m->is_loaded()) { 356 // Do not attempt to inline unloaded methods. 357 return NULL; 358 } 359 360 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 361 bool is_available = false; 362 363 { 364 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 365 // the compiler must transition to '_thread_in_vm' state because both 366 // methods access VM-internal data. 367 VM_ENTRY_MARK; 368 methodHandle mh(THREAD, m->get_Method()); 369 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) && 370 !C->directive()->is_intrinsic_disabled(mh) && 371 !vmIntrinsics::is_disabled_by_flags(mh); 372 373 } 374 375 if (is_available) { 376 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 377 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 378 return new LibraryIntrinsic(m, is_virtual, 379 vmIntrinsics::predicates_needed(id), 380 vmIntrinsics::does_virtual_dispatch(id), 381 (vmIntrinsics::ID) id); 382 } else { 383 return NULL; 384 } 385 } 386 387 //----------------------register_library_intrinsics----------------------- 388 // Initialize this file's data structures, for each Compile instance. 389 void Compile::register_library_intrinsics() { 390 // Nothing to do here. 391 } 392 393 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 394 LibraryCallKit kit(jvms, this); 395 Compile* C = kit.C; 396 int nodes = C->unique(); 397 #ifndef PRODUCT 398 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 399 char buf[1000]; 400 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 401 tty->print_cr("Intrinsic %s", str); 402 } 403 #endif 404 ciMethod* callee = kit.callee(); 405 const int bci = kit.bci(); 406 407 // Try to inline the intrinsic. 408 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 409 kit.try_to_inline(_last_predicate)) { 410 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)" 411 : "(intrinsic)"; 412 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 413 if (C->print_intrinsics() || C->print_inlining()) { 414 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 415 } 416 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 417 if (C->log()) { 418 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 419 vmIntrinsics::name_at(intrinsic_id()), 420 (is_virtual() ? " virtual='1'" : ""), 421 C->unique() - nodes); 422 } 423 // Push the result from the inlined method onto the stack. 424 kit.push_result(); 425 C->print_inlining_update(this); 426 return kit.transfer_exceptions_into_jvms(); 427 } 428 429 // The intrinsic bailed out 430 if (jvms->has_method()) { 431 // Not a root compile. 432 const char* msg; 433 if (callee->intrinsic_candidate()) { 434 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 435 } else { 436 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 437 : "failed to inline (intrinsic), method not annotated"; 438 } 439 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg); 440 if (C->print_intrinsics() || C->print_inlining()) { 441 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 442 } 443 } else { 444 // Root compile 445 ResourceMark rm; 446 stringStream msg_stream; 447 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 448 vmIntrinsics::name_at(intrinsic_id()), 449 is_virtual() ? " (virtual)" : "", bci); 450 const char *msg = msg_stream.as_string(); 451 log_debug(jit, inlining)("%s", msg); 452 if (C->print_intrinsics() || C->print_inlining()) { 453 tty->print("%s", msg); 454 } 455 } 456 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 457 C->print_inlining_update(this); 458 return NULL; 459 } 460 461 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 462 LibraryCallKit kit(jvms, this); 463 Compile* C = kit.C; 464 int nodes = C->unique(); 465 _last_predicate = predicate; 466 #ifndef PRODUCT 467 assert(is_predicated() && predicate < predicates_count(), "sanity"); 468 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 469 char buf[1000]; 470 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 471 tty->print_cr("Predicate for intrinsic %s", str); 472 } 473 #endif 474 ciMethod* callee = kit.callee(); 475 const int bci = kit.bci(); 476 477 Node* slow_ctl = kit.try_to_predicate(predicate); 478 if (!kit.failing()) { 479 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)" 480 : "(intrinsic, predicate)"; 481 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 482 if (C->print_intrinsics() || C->print_inlining()) { 483 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 484 } 485 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 486 if (C->log()) { 487 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 488 vmIntrinsics::name_at(intrinsic_id()), 489 (is_virtual() ? " virtual='1'" : ""), 490 C->unique() - nodes); 491 } 492 return slow_ctl; // Could be NULL if the check folds. 493 } 494 495 // The intrinsic bailed out 496 if (jvms->has_method()) { 497 // Not a root compile. 498 const char* msg = "failed to generate predicate for intrinsic"; 499 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg); 500 if (C->print_intrinsics() || C->print_inlining()) { 501 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 502 } 503 } else { 504 // Root compile 505 ResourceMark rm; 506 stringStream msg_stream; 507 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 508 vmIntrinsics::name_at(intrinsic_id()), 509 is_virtual() ? " (virtual)" : "", bci); 510 const char *msg = msg_stream.as_string(); 511 log_debug(jit, inlining)("%s", msg); 512 if (C->print_intrinsics() || C->print_inlining()) { 513 C->print_inlining_stream()->print("%s", msg); 514 } 515 } 516 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 517 return NULL; 518 } 519 520 bool LibraryCallKit::try_to_inline(int predicate) { 521 // Handle symbolic names for otherwise undistinguished boolean switches: 522 const bool is_store = true; 523 const bool is_compress = true; 524 const bool is_static = true; 525 const bool is_volatile = true; 526 527 if (!jvms()->has_method()) { 528 // Root JVMState has a null method. 529 assert(map()->memory()->Opcode() == Op_Parm, ""); 530 // Insert the memory aliasing node 531 set_all_memory(reset_memory()); 532 } 533 assert(merged_memory(), ""); 534 535 536 switch (intrinsic_id()) { 537 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 538 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 539 case vmIntrinsics::_getClass: return inline_native_getClass(); 540 541 case vmIntrinsics::_ceil: 542 case vmIntrinsics::_floor: 543 case vmIntrinsics::_rint: 544 case vmIntrinsics::_dsin: 545 case vmIntrinsics::_dcos: 546 case vmIntrinsics::_dtan: 547 case vmIntrinsics::_dabs: 548 case vmIntrinsics::_fabs: 549 case vmIntrinsics::_iabs: 550 case vmIntrinsics::_labs: 551 case vmIntrinsics::_datan2: 552 case vmIntrinsics::_dsqrt: 553 case vmIntrinsics::_dexp: 554 case vmIntrinsics::_dlog: 555 case vmIntrinsics::_dlog10: 556 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 557 558 case vmIntrinsics::_min: 559 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 560 561 case vmIntrinsics::_notify: 562 case vmIntrinsics::_notifyAll: 563 return inline_notify(intrinsic_id()); 564 565 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 566 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 567 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 568 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 569 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 570 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 571 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 572 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 573 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh(); 574 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 575 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 576 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 577 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 578 579 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 580 581 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 582 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 583 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 584 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 585 586 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 587 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 588 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 589 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 590 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 591 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 592 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 593 594 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 595 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 596 597 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 598 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 599 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 600 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 601 602 case vmIntrinsics::_compressStringC: 603 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 604 case vmIntrinsics::_inflateStringC: 605 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 606 607 case vmIntrinsics::_getReference: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 608 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 609 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 610 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 611 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 612 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 613 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 614 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 615 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 616 617 case vmIntrinsics::_putReference: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 618 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 619 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 620 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 621 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 622 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 623 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 624 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 625 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 626 627 case vmIntrinsics::_getReferenceVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 628 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 629 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 630 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 631 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 632 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 633 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 634 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 635 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 636 637 case vmIntrinsics::_putReferenceVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 638 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 639 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 640 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 641 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 642 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 643 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 644 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 645 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 646 647 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 648 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 649 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 650 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 651 652 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 653 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 654 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 655 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 656 657 case vmIntrinsics::_getReferenceAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 658 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 659 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 660 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 661 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 662 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 663 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 664 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 665 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 666 667 case vmIntrinsics::_putReferenceRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 668 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 669 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 670 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 671 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 672 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 673 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 674 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 675 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 676 677 case vmIntrinsics::_getReferenceOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 678 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 679 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 680 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 681 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 682 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 683 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 684 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 685 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 686 687 case vmIntrinsics::_putReferenceOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 688 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 689 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 690 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 691 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 692 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 693 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 694 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 695 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 696 697 case vmIntrinsics::_compareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 698 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); 699 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); 700 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 701 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 702 703 case vmIntrinsics::_weakCompareAndSetReferencePlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 704 case vmIntrinsics::_weakCompareAndSetReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 705 case vmIntrinsics::_weakCompareAndSetReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 706 case vmIntrinsics::_weakCompareAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 707 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); 708 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); 709 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); 710 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); 711 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); 712 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); 713 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); 714 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); 715 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 716 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 717 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 718 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 719 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 720 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 721 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 722 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 723 724 case vmIntrinsics::_compareAndExchangeReference: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 725 case vmIntrinsics::_compareAndExchangeReferenceAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 726 case vmIntrinsics::_compareAndExchangeReferenceRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 727 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); 728 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); 729 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); 730 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); 731 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); 732 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); 733 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 734 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 735 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 736 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 737 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 738 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 739 740 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); 741 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); 742 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 743 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 744 745 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); 746 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); 747 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 748 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 749 case vmIntrinsics::_getAndSetReference: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 750 751 case vmIntrinsics::_loadFence: 752 case vmIntrinsics::_storeFence: 753 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 754 755 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 756 757 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 758 759 #ifdef JFR_HAVE_INTRINSICS 760 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime"); 761 case vmIntrinsics::_getClassId: return inline_native_classID(); 762 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter(); 763 #endif 764 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 765 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 766 case vmIntrinsics::_writeback0: return inline_unsafe_writeback0(); 767 case vmIntrinsics::_writebackPreSync0: return inline_unsafe_writebackSync0(true); 768 case vmIntrinsics::_writebackPostSync0: return inline_unsafe_writebackSync0(false); 769 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 770 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 771 case vmIntrinsics::_getLength: return inline_native_getLength(); 772 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 773 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 774 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 775 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 776 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 777 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 778 779 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 780 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 781 782 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 783 784 case vmIntrinsics::_isInstance: 785 case vmIntrinsics::_getModifiers: 786 case vmIntrinsics::_isInterface: 787 case vmIntrinsics::_isArray: 788 case vmIntrinsics::_isPrimitive: 789 case vmIntrinsics::_isHidden: 790 case vmIntrinsics::_getSuperclass: 791 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 792 793 case vmIntrinsics::_floatToRawIntBits: 794 case vmIntrinsics::_floatToIntBits: 795 case vmIntrinsics::_intBitsToFloat: 796 case vmIntrinsics::_doubleToRawLongBits: 797 case vmIntrinsics::_doubleToLongBits: 798 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 799 800 case vmIntrinsics::_numberOfLeadingZeros_i: 801 case vmIntrinsics::_numberOfLeadingZeros_l: 802 case vmIntrinsics::_numberOfTrailingZeros_i: 803 case vmIntrinsics::_numberOfTrailingZeros_l: 804 case vmIntrinsics::_bitCount_i: 805 case vmIntrinsics::_bitCount_l: 806 case vmIntrinsics::_reverseBytes_i: 807 case vmIntrinsics::_reverseBytes_l: 808 case vmIntrinsics::_reverseBytes_s: 809 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 810 811 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 812 813 case vmIntrinsics::_Reference_get: return inline_reference_get(); 814 815 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 816 817 case vmIntrinsics::_aescrypt_encryptBlock: 818 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 819 820 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 821 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 822 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 823 824 case vmIntrinsics::_electronicCodeBook_encryptAESCrypt: 825 case vmIntrinsics::_electronicCodeBook_decryptAESCrypt: 826 return inline_electronicCodeBook_AESCrypt(intrinsic_id()); 827 828 case vmIntrinsics::_counterMode_AESCrypt: 829 return inline_counterMode_AESCrypt(intrinsic_id()); 830 831 case vmIntrinsics::_md5_implCompress: 832 case vmIntrinsics::_sha_implCompress: 833 case vmIntrinsics::_sha2_implCompress: 834 case vmIntrinsics::_sha5_implCompress: 835 return inline_digestBase_implCompress(intrinsic_id()); 836 837 case vmIntrinsics::_digestBase_implCompressMB: 838 return inline_digestBase_implCompressMB(predicate); 839 840 case vmIntrinsics::_multiplyToLen: 841 return inline_multiplyToLen(); 842 843 case vmIntrinsics::_squareToLen: 844 return inline_squareToLen(); 845 846 case vmIntrinsics::_mulAdd: 847 return inline_mulAdd(); 848 849 case vmIntrinsics::_montgomeryMultiply: 850 return inline_montgomeryMultiply(); 851 case vmIntrinsics::_montgomerySquare: 852 return inline_montgomerySquare(); 853 854 case vmIntrinsics::_bigIntegerRightShiftWorker: 855 return inline_bigIntegerShift(true); 856 case vmIntrinsics::_bigIntegerLeftShiftWorker: 857 return inline_bigIntegerShift(false); 858 859 case vmIntrinsics::_vectorizedMismatch: 860 return inline_vectorizedMismatch(); 861 862 case vmIntrinsics::_ghash_processBlocks: 863 return inline_ghash_processBlocks(); 864 case vmIntrinsics::_base64_encodeBlock: 865 return inline_base64_encodeBlock(); 866 case vmIntrinsics::_base64_decodeBlock: 867 return inline_base64_decodeBlock(); 868 869 case vmIntrinsics::_encodeISOArray: 870 case vmIntrinsics::_encodeByteISOArray: 871 return inline_encodeISOArray(); 872 873 case vmIntrinsics::_updateCRC32: 874 return inline_updateCRC32(); 875 case vmIntrinsics::_updateBytesCRC32: 876 return inline_updateBytesCRC32(); 877 case vmIntrinsics::_updateByteBufferCRC32: 878 return inline_updateByteBufferCRC32(); 879 880 case vmIntrinsics::_updateBytesCRC32C: 881 return inline_updateBytesCRC32C(); 882 case vmIntrinsics::_updateDirectByteBufferCRC32C: 883 return inline_updateDirectByteBufferCRC32C(); 884 885 case vmIntrinsics::_updateBytesAdler32: 886 return inline_updateBytesAdler32(); 887 case vmIntrinsics::_updateByteBufferAdler32: 888 return inline_updateByteBufferAdler32(); 889 890 case vmIntrinsics::_profileBoolean: 891 return inline_profileBoolean(); 892 case vmIntrinsics::_isCompileConstant: 893 return inline_isCompileConstant(); 894 895 case vmIntrinsics::_hasNegatives: 896 return inline_hasNegatives(); 897 898 case vmIntrinsics::_fmaD: 899 case vmIntrinsics::_fmaF: 900 return inline_fma(intrinsic_id()); 901 902 case vmIntrinsics::_isDigit: 903 case vmIntrinsics::_isLowerCase: 904 case vmIntrinsics::_isUpperCase: 905 case vmIntrinsics::_isWhitespace: 906 return inline_character_compare(intrinsic_id()); 907 908 case vmIntrinsics::_maxF: 909 case vmIntrinsics::_minF: 910 case vmIntrinsics::_maxD: 911 case vmIntrinsics::_minD: 912 return inline_fp_min_max(intrinsic_id()); 913 914 default: 915 // If you get here, it may be that someone has added a new intrinsic 916 // to the list in vmSymbols.hpp without implementing it here. 917 #ifndef PRODUCT 918 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 919 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 920 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 921 } 922 #endif 923 return false; 924 } 925 } 926 927 Node* LibraryCallKit::try_to_predicate(int predicate) { 928 if (!jvms()->has_method()) { 929 // Root JVMState has a null method. 930 assert(map()->memory()->Opcode() == Op_Parm, ""); 931 // Insert the memory aliasing node 932 set_all_memory(reset_memory()); 933 } 934 assert(merged_memory(), ""); 935 936 switch (intrinsic_id()) { 937 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 938 return inline_cipherBlockChaining_AESCrypt_predicate(false); 939 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 940 return inline_cipherBlockChaining_AESCrypt_predicate(true); 941 case vmIntrinsics::_electronicCodeBook_encryptAESCrypt: 942 return inline_electronicCodeBook_AESCrypt_predicate(false); 943 case vmIntrinsics::_electronicCodeBook_decryptAESCrypt: 944 return inline_electronicCodeBook_AESCrypt_predicate(true); 945 case vmIntrinsics::_counterMode_AESCrypt: 946 return inline_counterMode_AESCrypt_predicate(); 947 case vmIntrinsics::_digestBase_implCompressMB: 948 return inline_digestBase_implCompressMB_predicate(predicate); 949 950 default: 951 // If you get here, it may be that someone has added a new intrinsic 952 // to the list in vmSymbols.hpp without implementing it here. 953 #ifndef PRODUCT 954 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 955 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 956 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 957 } 958 #endif 959 Node* slow_ctl = control(); 960 set_control(top()); // No fast path instrinsic 961 return slow_ctl; 962 } 963 } 964 965 //------------------------------set_result------------------------------- 966 // Helper function for finishing intrinsics. 967 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 968 record_for_igvn(region); 969 set_control(_gvn.transform(region)); 970 set_result( _gvn.transform(value)); 971 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 972 } 973 974 //------------------------------generate_guard--------------------------- 975 // Helper function for generating guarded fast-slow graph structures. 976 // The given 'test', if true, guards a slow path. If the test fails 977 // then a fast path can be taken. (We generally hope it fails.) 978 // In all cases, GraphKit::control() is updated to the fast path. 979 // The returned value represents the control for the slow path. 980 // The return value is never 'top'; it is either a valid control 981 // or NULL if it is obvious that the slow path can never be taken. 982 // Also, if region and the slow control are not NULL, the slow edge 983 // is appended to the region. 984 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 985 if (stopped()) { 986 // Already short circuited. 987 return NULL; 988 } 989 990 // Build an if node and its projections. 991 // If test is true we take the slow path, which we assume is uncommon. 992 if (_gvn.type(test) == TypeInt::ZERO) { 993 // The slow branch is never taken. No need to build this guard. 994 return NULL; 995 } 996 997 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 998 999 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1000 if (if_slow == top()) { 1001 // The slow branch is never taken. No need to build this guard. 1002 return NULL; 1003 } 1004 1005 if (region != NULL) 1006 region->add_req(if_slow); 1007 1008 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1009 set_control(if_fast); 1010 1011 return if_slow; 1012 } 1013 1014 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1015 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1016 } 1017 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1018 return generate_guard(test, region, PROB_FAIR); 1019 } 1020 1021 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1022 Node* *pos_index) { 1023 if (stopped()) 1024 return NULL; // already stopped 1025 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1026 return NULL; // index is already adequately typed 1027 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1028 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1029 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1030 if (is_neg != NULL && pos_index != NULL) { 1031 // Emulate effect of Parse::adjust_map_after_if. 1032 Node* ccast = new CastIINode(index, TypeInt::POS); 1033 ccast->set_req(0, control()); 1034 (*pos_index) = _gvn.transform(ccast); 1035 } 1036 return is_neg; 1037 } 1038 1039 // Make sure that 'position' is a valid limit index, in [0..length]. 1040 // There are two equivalent plans for checking this: 1041 // A. (offset + copyLength) unsigned<= arrayLength 1042 // B. offset <= (arrayLength - copyLength) 1043 // We require that all of the values above, except for the sum and 1044 // difference, are already known to be non-negative. 1045 // Plan A is robust in the face of overflow, if offset and copyLength 1046 // are both hugely positive. 1047 // 1048 // Plan B is less direct and intuitive, but it does not overflow at 1049 // all, since the difference of two non-negatives is always 1050 // representable. Whenever Java methods must perform the equivalent 1051 // check they generally use Plan B instead of Plan A. 1052 // For the moment we use Plan A. 1053 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1054 Node* subseq_length, 1055 Node* array_length, 1056 RegionNode* region) { 1057 if (stopped()) 1058 return NULL; // already stopped 1059 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1060 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1061 return NULL; // common case of whole-array copy 1062 Node* last = subseq_length; 1063 if (!zero_offset) // last += offset 1064 last = _gvn.transform(new AddINode(last, offset)); 1065 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1066 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1067 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1068 return is_over; 1069 } 1070 1071 // Emit range checks for the given String.value byte array 1072 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 1073 if (stopped()) { 1074 return; // already stopped 1075 } 1076 RegionNode* bailout = new RegionNode(1); 1077 record_for_igvn(bailout); 1078 if (char_count) { 1079 // Convert char count to byte count 1080 count = _gvn.transform(new LShiftINode(count, intcon(1))); 1081 } 1082 1083 // Offset and count must not be negative 1084 generate_negative_guard(offset, bailout); 1085 generate_negative_guard(count, bailout); 1086 // Offset + count must not exceed length of array 1087 generate_limit_guard(offset, count, load_array_length(array), bailout); 1088 1089 if (bailout->req() > 1) { 1090 PreserveJVMState pjvms(this); 1091 set_control(_gvn.transform(bailout)); 1092 uncommon_trap(Deoptimization::Reason_intrinsic, 1093 Deoptimization::Action_maybe_recompile); 1094 } 1095 } 1096 1097 //--------------------------generate_current_thread-------------------- 1098 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1099 ciKlass* thread_klass = env()->Thread_klass(); 1100 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1101 Node* thread = _gvn.transform(new ThreadLocalNode()); 1102 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1103 tls_output = thread; 1104 Node* thread_obj_handle = LoadNode::make(_gvn, NULL, immutable_memory(), p, p->bottom_type()->is_ptr(), TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 1105 thread_obj_handle = _gvn.transform(thread_obj_handle); 1106 return access_load(thread_obj_handle, thread_type, T_OBJECT, IN_NATIVE | C2_IMMUTABLE_MEMORY); 1107 } 1108 1109 1110 //------------------------------make_string_method_node------------------------ 1111 // Helper method for String intrinsic functions. This version is called with 1112 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 1113 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 1114 // containing the lengths of str1 and str2. 1115 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 1116 Node* result = NULL; 1117 switch (opcode) { 1118 case Op_StrIndexOf: 1119 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 1120 str1_start, cnt1, str2_start, cnt2, ae); 1121 break; 1122 case Op_StrComp: 1123 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 1124 str1_start, cnt1, str2_start, cnt2, ae); 1125 break; 1126 case Op_StrEquals: 1127 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 1128 // Use the constant length if there is one because optimized match rule may exist. 1129 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 1130 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 1131 break; 1132 default: 1133 ShouldNotReachHere(); 1134 return NULL; 1135 } 1136 1137 // All these intrinsics have checks. 1138 C->set_has_split_ifs(true); // Has chance for split-if optimization 1139 clear_upper_avx(); 1140 1141 return _gvn.transform(result); 1142 } 1143 1144 //------------------------------inline_string_compareTo------------------------ 1145 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 1146 Node* arg1 = argument(0); 1147 Node* arg2 = argument(1); 1148 1149 arg1 = must_be_not_null(arg1, true); 1150 arg2 = must_be_not_null(arg2, true); 1151 1152 // Get start addr and length of first argument 1153 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1154 Node* arg1_cnt = load_array_length(arg1); 1155 1156 // Get start addr and length of second argument 1157 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1158 Node* arg2_cnt = load_array_length(arg2); 1159 1160 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1161 set_result(result); 1162 return true; 1163 } 1164 1165 //------------------------------inline_string_equals------------------------ 1166 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1167 Node* arg1 = argument(0); 1168 Node* arg2 = argument(1); 1169 1170 // paths (plus control) merge 1171 RegionNode* region = new RegionNode(3); 1172 Node* phi = new PhiNode(region, TypeInt::BOOL); 1173 1174 if (!stopped()) { 1175 1176 arg1 = must_be_not_null(arg1, true); 1177 arg2 = must_be_not_null(arg2, true); 1178 1179 // Get start addr and length of first argument 1180 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1181 Node* arg1_cnt = load_array_length(arg1); 1182 1183 // Get start addr and length of second argument 1184 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1185 Node* arg2_cnt = load_array_length(arg2); 1186 1187 // Check for arg1_cnt != arg2_cnt 1188 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1189 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1190 Node* if_ne = generate_slow_guard(bol, NULL); 1191 if (if_ne != NULL) { 1192 phi->init_req(2, intcon(0)); 1193 region->init_req(2, if_ne); 1194 } 1195 1196 // Check for count == 0 is done by assembler code for StrEquals. 1197 1198 if (!stopped()) { 1199 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1200 phi->init_req(1, equals); 1201 region->init_req(1, control()); 1202 } 1203 } 1204 1205 // post merge 1206 set_control(_gvn.transform(region)); 1207 record_for_igvn(region); 1208 1209 set_result(_gvn.transform(phi)); 1210 return true; 1211 } 1212 1213 //------------------------------inline_array_equals---------------------------- 1214 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1215 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1216 Node* arg1 = argument(0); 1217 Node* arg2 = argument(1); 1218 1219 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1220 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1221 clear_upper_avx(); 1222 1223 return true; 1224 } 1225 1226 //------------------------------inline_hasNegatives------------------------------ 1227 bool LibraryCallKit::inline_hasNegatives() { 1228 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1229 return false; 1230 } 1231 1232 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1233 // no receiver since it is static method 1234 Node* ba = argument(0); 1235 Node* offset = argument(1); 1236 Node* len = argument(2); 1237 1238 ba = must_be_not_null(ba, true); 1239 1240 // Range checks 1241 generate_string_range_check(ba, offset, len, false); 1242 if (stopped()) { 1243 return true; 1244 } 1245 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1246 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1247 set_result(_gvn.transform(result)); 1248 return true; 1249 } 1250 1251 bool LibraryCallKit::inline_preconditions_checkIndex() { 1252 Node* index = argument(0); 1253 Node* length = argument(1); 1254 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1255 return false; 1256 } 1257 1258 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1259 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1260 1261 { 1262 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1263 uncommon_trap(Deoptimization::Reason_intrinsic, 1264 Deoptimization::Action_make_not_entrant); 1265 } 1266 1267 if (stopped()) { 1268 return false; 1269 } 1270 1271 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1272 BoolTest::mask btest = BoolTest::lt; 1273 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1274 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1275 _gvn.set_type(rc, rc->Value(&_gvn)); 1276 if (!rc_bool->is_Con()) { 1277 record_for_igvn(rc); 1278 } 1279 set_control(_gvn.transform(new IfTrueNode(rc))); 1280 { 1281 PreserveJVMState pjvms(this); 1282 set_control(_gvn.transform(new IfFalseNode(rc))); 1283 uncommon_trap(Deoptimization::Reason_range_check, 1284 Deoptimization::Action_make_not_entrant); 1285 } 1286 1287 if (stopped()) { 1288 return false; 1289 } 1290 1291 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1292 result->set_req(0, control()); 1293 result = _gvn.transform(result); 1294 set_result(result); 1295 replace_in_map(index, result); 1296 clear_upper_avx(); 1297 return true; 1298 } 1299 1300 //------------------------------inline_string_indexOf------------------------ 1301 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1302 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1303 return false; 1304 } 1305 Node* src = argument(0); 1306 Node* tgt = argument(1); 1307 1308 // Make the merge point 1309 RegionNode* result_rgn = new RegionNode(4); 1310 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1311 1312 src = must_be_not_null(src, true); 1313 tgt = must_be_not_null(tgt, true); 1314 1315 // Get start addr and length of source string 1316 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1317 Node* src_count = load_array_length(src); 1318 1319 // Get start addr and length of substring 1320 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1321 Node* tgt_count = load_array_length(tgt); 1322 1323 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1324 // Divide src size by 2 if String is UTF16 encoded 1325 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1326 } 1327 if (ae == StrIntrinsicNode::UU) { 1328 // Divide substring size by 2 if String is UTF16 encoded 1329 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1330 } 1331 1332 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1333 if (result != NULL) { 1334 result_phi->init_req(3, result); 1335 result_rgn->init_req(3, control()); 1336 } 1337 set_control(_gvn.transform(result_rgn)); 1338 record_for_igvn(result_rgn); 1339 set_result(_gvn.transform(result_phi)); 1340 1341 return true; 1342 } 1343 1344 //-----------------------------inline_string_indexOf----------------------- 1345 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1346 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1347 return false; 1348 } 1349 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1350 return false; 1351 } 1352 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1353 Node* src = argument(0); // byte[] 1354 Node* src_count = argument(1); // char count 1355 Node* tgt = argument(2); // byte[] 1356 Node* tgt_count = argument(3); // char count 1357 Node* from_index = argument(4); // char index 1358 1359 src = must_be_not_null(src, true); 1360 tgt = must_be_not_null(tgt, true); 1361 1362 // Multiply byte array index by 2 if String is UTF16 encoded 1363 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1364 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1365 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1366 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1367 1368 // Range checks 1369 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1370 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1371 if (stopped()) { 1372 return true; 1373 } 1374 1375 RegionNode* region = new RegionNode(5); 1376 Node* phi = new PhiNode(region, TypeInt::INT); 1377 1378 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1379 if (result != NULL) { 1380 // The result is index relative to from_index if substring was found, -1 otherwise. 1381 // Generate code which will fold into cmove. 1382 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1383 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1384 1385 Node* if_lt = generate_slow_guard(bol, NULL); 1386 if (if_lt != NULL) { 1387 // result == -1 1388 phi->init_req(3, result); 1389 region->init_req(3, if_lt); 1390 } 1391 if (!stopped()) { 1392 result = _gvn.transform(new AddINode(result, from_index)); 1393 phi->init_req(4, result); 1394 region->init_req(4, control()); 1395 } 1396 } 1397 1398 set_control(_gvn.transform(region)); 1399 record_for_igvn(region); 1400 set_result(_gvn.transform(phi)); 1401 clear_upper_avx(); 1402 1403 return true; 1404 } 1405 1406 // Create StrIndexOfNode with fast path checks 1407 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1408 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1409 // Check for substr count > string count 1410 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1411 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1412 Node* if_gt = generate_slow_guard(bol, NULL); 1413 if (if_gt != NULL) { 1414 phi->init_req(1, intcon(-1)); 1415 region->init_req(1, if_gt); 1416 } 1417 if (!stopped()) { 1418 // Check for substr count == 0 1419 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1420 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1421 Node* if_zero = generate_slow_guard(bol, NULL); 1422 if (if_zero != NULL) { 1423 phi->init_req(2, intcon(0)); 1424 region->init_req(2, if_zero); 1425 } 1426 } 1427 if (!stopped()) { 1428 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1429 } 1430 return NULL; 1431 } 1432 1433 //-----------------------------inline_string_indexOfChar----------------------- 1434 bool LibraryCallKit::inline_string_indexOfChar() { 1435 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1436 return false; 1437 } 1438 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1439 return false; 1440 } 1441 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1442 Node* src = argument(0); // byte[] 1443 Node* tgt = argument(1); // tgt is int ch 1444 Node* from_index = argument(2); 1445 Node* max = argument(3); 1446 1447 src = must_be_not_null(src, true); 1448 1449 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1450 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1451 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1452 1453 // Range checks 1454 generate_string_range_check(src, src_offset, src_count, true); 1455 if (stopped()) { 1456 return true; 1457 } 1458 1459 RegionNode* region = new RegionNode(3); 1460 Node* phi = new PhiNode(region, TypeInt::INT); 1461 1462 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1463 C->set_has_split_ifs(true); // Has chance for split-if optimization 1464 _gvn.transform(result); 1465 1466 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1467 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1468 1469 Node* if_lt = generate_slow_guard(bol, NULL); 1470 if (if_lt != NULL) { 1471 // result == -1 1472 phi->init_req(2, result); 1473 region->init_req(2, if_lt); 1474 } 1475 if (!stopped()) { 1476 result = _gvn.transform(new AddINode(result, from_index)); 1477 phi->init_req(1, result); 1478 region->init_req(1, control()); 1479 } 1480 set_control(_gvn.transform(region)); 1481 record_for_igvn(region); 1482 set_result(_gvn.transform(phi)); 1483 1484 return true; 1485 } 1486 //---------------------------inline_string_copy--------------------- 1487 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1488 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1489 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1490 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1491 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1492 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1493 bool LibraryCallKit::inline_string_copy(bool compress) { 1494 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1495 return false; 1496 } 1497 int nargs = 5; // 2 oops, 3 ints 1498 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1499 1500 Node* src = argument(0); 1501 Node* src_offset = argument(1); 1502 Node* dst = argument(2); 1503 Node* dst_offset = argument(3); 1504 Node* length = argument(4); 1505 1506 // Check for allocation before we add nodes that would confuse 1507 // tightly_coupled_allocation() 1508 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1509 1510 // Figure out the size and type of the elements we will be copying. 1511 const Type* src_type = src->Value(&_gvn); 1512 const Type* dst_type = dst->Value(&_gvn); 1513 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1514 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1515 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1516 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1517 "Unsupported array types for inline_string_copy"); 1518 1519 src = must_be_not_null(src, true); 1520 dst = must_be_not_null(dst, true); 1521 1522 // Convert char[] offsets to byte[] offsets 1523 bool convert_src = (compress && src_elem == T_BYTE); 1524 bool convert_dst = (!compress && dst_elem == T_BYTE); 1525 if (convert_src) { 1526 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1527 } else if (convert_dst) { 1528 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1529 } 1530 1531 // Range checks 1532 generate_string_range_check(src, src_offset, length, convert_src); 1533 generate_string_range_check(dst, dst_offset, length, convert_dst); 1534 if (stopped()) { 1535 return true; 1536 } 1537 1538 Node* src_start = array_element_address(src, src_offset, src_elem); 1539 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1540 // 'src_start' points to src array + scaled offset 1541 // 'dst_start' points to dst array + scaled offset 1542 Node* count = NULL; 1543 if (compress) { 1544 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1545 } else { 1546 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1547 } 1548 1549 if (alloc != NULL) { 1550 if (alloc->maybe_set_complete(&_gvn)) { 1551 // "You break it, you buy it." 1552 InitializeNode* init = alloc->initialization(); 1553 assert(init->is_complete(), "we just did this"); 1554 init->set_complete_with_arraycopy(); 1555 assert(dst->is_CheckCastPP(), "sanity"); 1556 assert(dst->in(0)->in(0) == init, "dest pinned"); 1557 } 1558 // Do not let stores that initialize this object be reordered with 1559 // a subsequent store that would make this object accessible by 1560 // other threads. 1561 // Record what AllocateNode this StoreStore protects so that 1562 // escape analysis can go from the MemBarStoreStoreNode to the 1563 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1564 // based on the escape status of the AllocateNode. 1565 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1566 } 1567 if (compress) { 1568 set_result(_gvn.transform(count)); 1569 } 1570 clear_upper_avx(); 1571 1572 return true; 1573 } 1574 1575 #ifdef _LP64 1576 #define XTOP ,top() /*additional argument*/ 1577 #else //_LP64 1578 #define XTOP /*no additional argument*/ 1579 #endif //_LP64 1580 1581 //------------------------inline_string_toBytesU-------------------------- 1582 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1583 bool LibraryCallKit::inline_string_toBytesU() { 1584 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1585 return false; 1586 } 1587 // Get the arguments. 1588 Node* value = argument(0); 1589 Node* offset = argument(1); 1590 Node* length = argument(2); 1591 1592 Node* newcopy = NULL; 1593 1594 // Set the original stack and the reexecute bit for the interpreter to reexecute 1595 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1596 { PreserveReexecuteState preexecs(this); 1597 jvms()->set_should_reexecute(true); 1598 1599 // Check if a null path was taken unconditionally. 1600 value = null_check(value); 1601 1602 RegionNode* bailout = new RegionNode(1); 1603 record_for_igvn(bailout); 1604 1605 // Range checks 1606 generate_negative_guard(offset, bailout); 1607 generate_negative_guard(length, bailout); 1608 generate_limit_guard(offset, length, load_array_length(value), bailout); 1609 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1610 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1611 1612 if (bailout->req() > 1) { 1613 PreserveJVMState pjvms(this); 1614 set_control(_gvn.transform(bailout)); 1615 uncommon_trap(Deoptimization::Reason_intrinsic, 1616 Deoptimization::Action_maybe_recompile); 1617 } 1618 if (stopped()) { 1619 return true; 1620 } 1621 1622 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1623 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1624 newcopy = new_array(klass_node, size, 0); // no arguments to push 1625 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1626 1627 // Calculate starting addresses. 1628 Node* src_start = array_element_address(value, offset, T_CHAR); 1629 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1630 1631 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1632 const TypeInt* toffset = gvn().type(offset)->is_int(); 1633 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1634 1635 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1636 const char* copyfunc_name = "arraycopy"; 1637 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1638 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1639 OptoRuntime::fast_arraycopy_Type(), 1640 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1641 src_start, dst_start, ConvI2X(length) XTOP); 1642 // Do not let reads from the cloned object float above the arraycopy. 1643 if (alloc != NULL) { 1644 if (alloc->maybe_set_complete(&_gvn)) { 1645 // "You break it, you buy it." 1646 InitializeNode* init = alloc->initialization(); 1647 assert(init->is_complete(), "we just did this"); 1648 init->set_complete_with_arraycopy(); 1649 assert(newcopy->is_CheckCastPP(), "sanity"); 1650 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1651 } 1652 // Do not let stores that initialize this object be reordered with 1653 // a subsequent store that would make this object accessible by 1654 // other threads. 1655 // Record what AllocateNode this StoreStore protects so that 1656 // escape analysis can go from the MemBarStoreStoreNode to the 1657 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1658 // based on the escape status of the AllocateNode. 1659 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1660 } else { 1661 insert_mem_bar(Op_MemBarCPUOrder); 1662 } 1663 } // original reexecute is set back here 1664 1665 C->set_has_split_ifs(true); // Has chance for split-if optimization 1666 if (!stopped()) { 1667 set_result(newcopy); 1668 } 1669 clear_upper_avx(); 1670 1671 return true; 1672 } 1673 1674 //------------------------inline_string_getCharsU-------------------------- 1675 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1676 bool LibraryCallKit::inline_string_getCharsU() { 1677 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1678 return false; 1679 } 1680 1681 // Get the arguments. 1682 Node* src = argument(0); 1683 Node* src_begin = argument(1); 1684 Node* src_end = argument(2); // exclusive offset (i < src_end) 1685 Node* dst = argument(3); 1686 Node* dst_begin = argument(4); 1687 1688 // Check for allocation before we add nodes that would confuse 1689 // tightly_coupled_allocation() 1690 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1691 1692 // Check if a null path was taken unconditionally. 1693 src = null_check(src); 1694 dst = null_check(dst); 1695 if (stopped()) { 1696 return true; 1697 } 1698 1699 // Get length and convert char[] offset to byte[] offset 1700 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1701 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1702 1703 // Range checks 1704 generate_string_range_check(src, src_begin, length, true); 1705 generate_string_range_check(dst, dst_begin, length, false); 1706 if (stopped()) { 1707 return true; 1708 } 1709 1710 if (!stopped()) { 1711 // Calculate starting addresses. 1712 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1713 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1714 1715 // Check if array addresses are aligned to HeapWordSize 1716 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1717 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1718 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1719 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1720 1721 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1722 const char* copyfunc_name = "arraycopy"; 1723 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1724 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1725 OptoRuntime::fast_arraycopy_Type(), 1726 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1727 src_start, dst_start, ConvI2X(length) XTOP); 1728 // Do not let reads from the cloned object float above the arraycopy. 1729 if (alloc != NULL) { 1730 if (alloc->maybe_set_complete(&_gvn)) { 1731 // "You break it, you buy it." 1732 InitializeNode* init = alloc->initialization(); 1733 assert(init->is_complete(), "we just did this"); 1734 init->set_complete_with_arraycopy(); 1735 assert(dst->is_CheckCastPP(), "sanity"); 1736 assert(dst->in(0)->in(0) == init, "dest pinned"); 1737 } 1738 // Do not let stores that initialize this object be reordered with 1739 // a subsequent store that would make this object accessible by 1740 // other threads. 1741 // Record what AllocateNode this StoreStore protects so that 1742 // escape analysis can go from the MemBarStoreStoreNode to the 1743 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1744 // based on the escape status of the AllocateNode. 1745 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1746 } else { 1747 insert_mem_bar(Op_MemBarCPUOrder); 1748 } 1749 } 1750 1751 C->set_has_split_ifs(true); // Has chance for split-if optimization 1752 return true; 1753 } 1754 1755 //----------------------inline_string_char_access---------------------------- 1756 // Store/Load char to/from byte[] array. 1757 // static void StringUTF16.putChar(byte[] val, int index, int c) 1758 // static char StringUTF16.getChar(byte[] val, int index) 1759 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1760 Node* value = argument(0); 1761 Node* index = argument(1); 1762 Node* ch = is_store ? argument(2) : NULL; 1763 1764 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1765 // correctly requires matched array shapes. 1766 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1767 "sanity: byte[] and char[] bases agree"); 1768 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1769 "sanity: byte[] and char[] scales agree"); 1770 1771 // Bail when getChar over constants is requested: constant folding would 1772 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1773 // Java method would constant fold nicely instead. 1774 if (!is_store && value->is_Con() && index->is_Con()) { 1775 return false; 1776 } 1777 1778 value = must_be_not_null(value, true); 1779 1780 Node* adr = array_element_address(value, index, T_CHAR); 1781 if (adr->is_top()) { 1782 return false; 1783 } 1784 if (is_store) { 1785 access_store_at(value, adr, TypeAryPtr::BYTES, ch, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED); 1786 } else { 1787 ch = access_load_at(value, adr, TypeAryPtr::BYTES, TypeInt::CHAR, T_CHAR, IN_HEAP | MO_UNORDERED | C2_MISMATCHED | C2_CONTROL_DEPENDENT_LOAD); 1788 set_result(ch); 1789 } 1790 return true; 1791 } 1792 1793 //--------------------------round_double_node-------------------------------- 1794 // Round a double node if necessary. 1795 Node* LibraryCallKit::round_double_node(Node* n) { 1796 if (Matcher::strict_fp_requires_explicit_rounding) { 1797 #ifdef IA32 1798 if (UseSSE < 2) { 1799 n = _gvn.transform(new RoundDoubleNode(NULL, n)); 1800 } 1801 #else 1802 Unimplemented(); 1803 #endif // IA32 1804 } 1805 return n; 1806 } 1807 1808 //------------------------------inline_math----------------------------------- 1809 // public static double Math.abs(double) 1810 // public static double Math.sqrt(double) 1811 // public static double Math.log(double) 1812 // public static double Math.log10(double) 1813 bool LibraryCallKit::inline_double_math(vmIntrinsics::ID id) { 1814 Node* arg = round_double_node(argument(0)); 1815 Node* n = NULL; 1816 switch (id) { 1817 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1818 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1819 case vmIntrinsics::_ceil: n = RoundDoubleModeNode::make(_gvn, arg, RoundDoubleModeNode::rmode_ceil); break; 1820 case vmIntrinsics::_floor: n = RoundDoubleModeNode::make(_gvn, arg, RoundDoubleModeNode::rmode_floor); break; 1821 case vmIntrinsics::_rint: n = RoundDoubleModeNode::make(_gvn, arg, RoundDoubleModeNode::rmode_rint); break; 1822 default: fatal_unexpected_iid(id); break; 1823 } 1824 set_result(_gvn.transform(n)); 1825 return true; 1826 } 1827 1828 //------------------------------inline_math----------------------------------- 1829 // public static float Math.abs(float) 1830 // public static int Math.abs(int) 1831 // public static long Math.abs(long) 1832 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1833 Node* arg = argument(0); 1834 Node* n = NULL; 1835 switch (id) { 1836 case vmIntrinsics::_fabs: n = new AbsFNode( arg); break; 1837 case vmIntrinsics::_iabs: n = new AbsINode( arg); break; 1838 case vmIntrinsics::_labs: n = new AbsLNode( arg); break; 1839 default: fatal_unexpected_iid(id); break; 1840 } 1841 set_result(_gvn.transform(n)); 1842 return true; 1843 } 1844 1845 //------------------------------runtime_math----------------------------- 1846 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1847 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1848 "must be (DD)D or (D)D type"); 1849 1850 // Inputs 1851 Node* a = round_double_node(argument(0)); 1852 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1853 1854 const TypePtr* no_memory_effects = NULL; 1855 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1856 no_memory_effects, 1857 a, top(), b, b ? top() : NULL); 1858 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1859 #ifdef ASSERT 1860 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1861 assert(value_top == top(), "second value must be top"); 1862 #endif 1863 1864 set_result(value); 1865 return true; 1866 } 1867 1868 //------------------------------inline_math_native----------------------------- 1869 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1870 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1871 switch (id) { 1872 // These intrinsics are not properly supported on all hardware 1873 case vmIntrinsics::_dsin: 1874 return StubRoutines::dsin() != NULL ? 1875 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1876 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1877 case vmIntrinsics::_dcos: 1878 return StubRoutines::dcos() != NULL ? 1879 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1880 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1881 case vmIntrinsics::_dtan: 1882 return StubRoutines::dtan() != NULL ? 1883 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1884 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1885 case vmIntrinsics::_dlog: 1886 return StubRoutines::dlog() != NULL ? 1887 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1888 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1889 case vmIntrinsics::_dlog10: 1890 return StubRoutines::dlog10() != NULL ? 1891 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1892 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1893 1894 // These intrinsics are supported on all hardware 1895 case vmIntrinsics::_ceil: 1896 case vmIntrinsics::_floor: 1897 case vmIntrinsics::_rint: return Matcher::match_rule_supported(Op_RoundDoubleMode) ? inline_double_math(id) : false; 1898 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_double_math(id) : false; 1899 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_double_math(id) : false; 1900 case vmIntrinsics::_fabs: return Matcher::match_rule_supported(Op_AbsF) ? inline_math(id) : false; 1901 case vmIntrinsics::_iabs: return Matcher::match_rule_supported(Op_AbsI) ? inline_math(id) : false; 1902 case vmIntrinsics::_labs: return Matcher::match_rule_supported(Op_AbsL) ? inline_math(id) : false; 1903 1904 case vmIntrinsics::_dexp: 1905 return StubRoutines::dexp() != NULL ? 1906 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1907 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1908 case vmIntrinsics::_dpow: { 1909 Node* exp = round_double_node(argument(2)); 1910 const TypeD* d = _gvn.type(exp)->isa_double_constant(); 1911 if (d != NULL && d->getd() == 2.0) { 1912 // Special case: pow(x, 2.0) => x * x 1913 Node* base = round_double_node(argument(0)); 1914 set_result(_gvn.transform(new MulDNode(base, base))); 1915 return true; 1916 } 1917 return StubRoutines::dpow() != NULL ? 1918 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1919 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1920 } 1921 #undef FN_PTR 1922 1923 // These intrinsics are not yet correctly implemented 1924 case vmIntrinsics::_datan2: 1925 return false; 1926 1927 default: 1928 fatal_unexpected_iid(id); 1929 return false; 1930 } 1931 } 1932 1933 static bool is_simple_name(Node* n) { 1934 return (n->req() == 1 // constant 1935 || (n->is_Type() && n->as_Type()->type()->singleton()) 1936 || n->is_Proj() // parameter or return value 1937 || n->is_Phi() // local of some sort 1938 ); 1939 } 1940 1941 //----------------------------inline_notify-----------------------------------* 1942 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1943 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1944 address func; 1945 if (id == vmIntrinsics::_notify) { 1946 func = OptoRuntime::monitor_notify_Java(); 1947 } else { 1948 func = OptoRuntime::monitor_notifyAll_Java(); 1949 } 1950 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1951 make_slow_call_ex(call, env()->Throwable_klass(), false); 1952 return true; 1953 } 1954 1955 1956 //----------------------------inline_min_max----------------------------------- 1957 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1958 set_result(generate_min_max(id, argument(0), argument(1))); 1959 return true; 1960 } 1961 1962 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1963 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1964 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1965 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1966 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1967 1968 { 1969 PreserveJVMState pjvms(this); 1970 PreserveReexecuteState preexecs(this); 1971 jvms()->set_should_reexecute(true); 1972 1973 set_control(slow_path); 1974 set_i_o(i_o()); 1975 1976 uncommon_trap(Deoptimization::Reason_intrinsic, 1977 Deoptimization::Action_none); 1978 } 1979 1980 set_control(fast_path); 1981 set_result(math); 1982 } 1983 1984 template <typename OverflowOp> 1985 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1986 typedef typename OverflowOp::MathOp MathOp; 1987 1988 MathOp* mathOp = new MathOp(arg1, arg2); 1989 Node* operation = _gvn.transform( mathOp ); 1990 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1991 inline_math_mathExact(operation, ofcheck); 1992 return true; 1993 } 1994 1995 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1996 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1997 } 1998 1999 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2000 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2001 } 2002 2003 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2004 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2005 } 2006 2007 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2008 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2009 } 2010 2011 bool LibraryCallKit::inline_math_negateExactI() { 2012 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2013 } 2014 2015 bool LibraryCallKit::inline_math_negateExactL() { 2016 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2017 } 2018 2019 bool LibraryCallKit::inline_math_multiplyExactI() { 2020 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2021 } 2022 2023 bool LibraryCallKit::inline_math_multiplyExactL() { 2024 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2025 } 2026 2027 bool LibraryCallKit::inline_math_multiplyHigh() { 2028 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2)))); 2029 return true; 2030 } 2031 2032 Node* 2033 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2034 // These are the candidate return value: 2035 Node* xvalue = x0; 2036 Node* yvalue = y0; 2037 2038 if (xvalue == yvalue) { 2039 return xvalue; 2040 } 2041 2042 bool want_max = (id == vmIntrinsics::_max); 2043 2044 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2045 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2046 if (txvalue == NULL || tyvalue == NULL) return top(); 2047 // This is not really necessary, but it is consistent with a 2048 // hypothetical MaxINode::Value method: 2049 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2050 2051 // %%% This folding logic should (ideally) be in a different place. 2052 // Some should be inside IfNode, and there to be a more reliable 2053 // transformation of ?: style patterns into cmoves. We also want 2054 // more powerful optimizations around cmove and min/max. 2055 2056 // Try to find a dominating comparison of these guys. 2057 // It can simplify the index computation for Arrays.copyOf 2058 // and similar uses of System.arraycopy. 2059 // First, compute the normalized version of CmpI(x, y). 2060 int cmp_op = Op_CmpI; 2061 Node* xkey = xvalue; 2062 Node* ykey = yvalue; 2063 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2064 if (ideal_cmpxy->is_Cmp()) { 2065 // E.g., if we have CmpI(length - offset, count), 2066 // it might idealize to CmpI(length, count + offset) 2067 cmp_op = ideal_cmpxy->Opcode(); 2068 xkey = ideal_cmpxy->in(1); 2069 ykey = ideal_cmpxy->in(2); 2070 } 2071 2072 // Start by locating any relevant comparisons. 2073 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2074 Node* cmpxy = NULL; 2075 Node* cmpyx = NULL; 2076 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2077 Node* cmp = start_from->fast_out(k); 2078 if (cmp->outcnt() > 0 && // must have prior uses 2079 cmp->in(0) == NULL && // must be context-independent 2080 cmp->Opcode() == cmp_op) { // right kind of compare 2081 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2082 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2083 } 2084 } 2085 2086 const int NCMPS = 2; 2087 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2088 int cmpn; 2089 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2090 if (cmps[cmpn] != NULL) break; // find a result 2091 } 2092 if (cmpn < NCMPS) { 2093 // Look for a dominating test that tells us the min and max. 2094 int depth = 0; // Limit search depth for speed 2095 Node* dom = control(); 2096 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2097 if (++depth >= 100) break; 2098 Node* ifproj = dom; 2099 if (!ifproj->is_Proj()) continue; 2100 Node* iff = ifproj->in(0); 2101 if (!iff->is_If()) continue; 2102 Node* bol = iff->in(1); 2103 if (!bol->is_Bool()) continue; 2104 Node* cmp = bol->in(1); 2105 if (cmp == NULL) continue; 2106 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2107 if (cmps[cmpn] == cmp) break; 2108 if (cmpn == NCMPS) continue; 2109 BoolTest::mask btest = bol->as_Bool()->_test._test; 2110 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2111 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2112 // At this point, we know that 'x btest y' is true. 2113 switch (btest) { 2114 case BoolTest::eq: 2115 // They are proven equal, so we can collapse the min/max. 2116 // Either value is the answer. Choose the simpler. 2117 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2118 return yvalue; 2119 return xvalue; 2120 case BoolTest::lt: // x < y 2121 case BoolTest::le: // x <= y 2122 return (want_max ? yvalue : xvalue); 2123 case BoolTest::gt: // x > y 2124 case BoolTest::ge: // x >= y 2125 return (want_max ? xvalue : yvalue); 2126 default: 2127 break; 2128 } 2129 } 2130 } 2131 2132 // We failed to find a dominating test. 2133 // Let's pick a test that might GVN with prior tests. 2134 Node* best_bol = NULL; 2135 BoolTest::mask best_btest = BoolTest::illegal; 2136 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2137 Node* cmp = cmps[cmpn]; 2138 if (cmp == NULL) continue; 2139 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2140 Node* bol = cmp->fast_out(j); 2141 if (!bol->is_Bool()) continue; 2142 BoolTest::mask btest = bol->as_Bool()->_test._test; 2143 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2144 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2145 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2146 best_bol = bol->as_Bool(); 2147 best_btest = btest; 2148 } 2149 } 2150 } 2151 2152 Node* answer_if_true = NULL; 2153 Node* answer_if_false = NULL; 2154 switch (best_btest) { 2155 default: 2156 if (cmpxy == NULL) 2157 cmpxy = ideal_cmpxy; 2158 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2159 // and fall through: 2160 case BoolTest::lt: // x < y 2161 case BoolTest::le: // x <= y 2162 answer_if_true = (want_max ? yvalue : xvalue); 2163 answer_if_false = (want_max ? xvalue : yvalue); 2164 break; 2165 case BoolTest::gt: // x > y 2166 case BoolTest::ge: // x >= y 2167 answer_if_true = (want_max ? xvalue : yvalue); 2168 answer_if_false = (want_max ? yvalue : xvalue); 2169 break; 2170 } 2171 2172 jint hi, lo; 2173 if (want_max) { 2174 // We can sharpen the minimum. 2175 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2176 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2177 } else { 2178 // We can sharpen the maximum. 2179 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2180 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2181 } 2182 2183 // Use a flow-free graph structure, to avoid creating excess control edges 2184 // which could hinder other optimizations. 2185 // Since Math.min/max is often used with arraycopy, we want 2186 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2187 Node* cmov = CMoveNode::make(NULL, best_bol, 2188 answer_if_false, answer_if_true, 2189 TypeInt::make(lo, hi, widen)); 2190 2191 return _gvn.transform(cmov); 2192 2193 /* 2194 // This is not as desirable as it may seem, since Min and Max 2195 // nodes do not have a full set of optimizations. 2196 // And they would interfere, anyway, with 'if' optimizations 2197 // and with CMoveI canonical forms. 2198 switch (id) { 2199 case vmIntrinsics::_min: 2200 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2201 case vmIntrinsics::_max: 2202 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2203 default: 2204 ShouldNotReachHere(); 2205 } 2206 */ 2207 } 2208 2209 inline int 2210 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) { 2211 const TypePtr* base_type = TypePtr::NULL_PTR; 2212 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2213 if (base_type == NULL) { 2214 // Unknown type. 2215 return Type::AnyPtr; 2216 } else if (base_type == TypePtr::NULL_PTR) { 2217 // Since this is a NULL+long form, we have to switch to a rawptr. 2218 base = _gvn.transform(new CastX2PNode(offset)); 2219 offset = MakeConX(0); 2220 return Type::RawPtr; 2221 } else if (base_type->base() == Type::RawPtr) { 2222 return Type::RawPtr; 2223 } else if (base_type->isa_oopptr()) { 2224 // Base is never null => always a heap address. 2225 if (!TypePtr::NULL_PTR->higher_equal(base_type)) { 2226 return Type::OopPtr; 2227 } 2228 // Offset is small => always a heap address. 2229 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2230 if (offset_type != NULL && 2231 base_type->offset() == 0 && // (should always be?) 2232 offset_type->_lo >= 0 && 2233 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2234 return Type::OopPtr; 2235 } else if (type == T_OBJECT) { 2236 // off heap access to an oop doesn't make any sense. Has to be on 2237 // heap. 2238 return Type::OopPtr; 2239 } 2240 // Otherwise, it might either be oop+off or NULL+addr. 2241 return Type::AnyPtr; 2242 } else { 2243 // No information: 2244 return Type::AnyPtr; 2245 } 2246 } 2247 2248 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, DecoratorSet decorators, BasicType type, bool can_cast) { 2249 Node* uncasted_base = base; 2250 int kind = classify_unsafe_addr(uncasted_base, offset, type); 2251 if (kind == Type::RawPtr) { 2252 return basic_plus_adr(top(), uncasted_base, offset); 2253 } else if (kind == Type::AnyPtr) { 2254 assert(base == uncasted_base, "unexpected base change"); 2255 if (can_cast) { 2256 if (!_gvn.type(base)->speculative_maybe_null() && 2257 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2258 // According to profiling, this access is always on 2259 // heap. Casting the base to not null and thus avoiding membars 2260 // around the access should allow better optimizations 2261 Node* null_ctl = top(); 2262 base = null_check_oop(base, &null_ctl, true, true, true); 2263 assert(null_ctl->is_top(), "no null control here"); 2264 return basic_plus_adr(base, offset); 2265 } else if (_gvn.type(base)->speculative_always_null() && 2266 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2267 // According to profiling, this access is always off 2268 // heap. 2269 base = null_assert(base); 2270 Node* raw_base = _gvn.transform(new CastX2PNode(offset)); 2271 offset = MakeConX(0); 2272 return basic_plus_adr(top(), raw_base, offset); 2273 } 2274 } 2275 // We don't know if it's an on heap or off heap access. Fall back 2276 // to raw memory access. 2277 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM)); 2278 return basic_plus_adr(top(), raw, offset); 2279 } else { 2280 assert(base == uncasted_base, "unexpected base change"); 2281 // We know it's an on heap access so base can't be null 2282 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) { 2283 base = must_be_not_null(base, true); 2284 } 2285 return basic_plus_adr(base, offset); 2286 } 2287 } 2288 2289 //--------------------------inline_number_methods----------------------------- 2290 // inline int Integer.numberOfLeadingZeros(int) 2291 // inline int Long.numberOfLeadingZeros(long) 2292 // 2293 // inline int Integer.numberOfTrailingZeros(int) 2294 // inline int Long.numberOfTrailingZeros(long) 2295 // 2296 // inline int Integer.bitCount(int) 2297 // inline int Long.bitCount(long) 2298 // 2299 // inline char Character.reverseBytes(char) 2300 // inline short Short.reverseBytes(short) 2301 // inline int Integer.reverseBytes(int) 2302 // inline long Long.reverseBytes(long) 2303 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2304 Node* arg = argument(0); 2305 Node* n = NULL; 2306 switch (id) { 2307 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2308 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2309 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2310 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2311 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2312 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2313 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2314 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2315 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2316 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2317 default: fatal_unexpected_iid(id); break; 2318 } 2319 set_result(_gvn.transform(n)); 2320 return true; 2321 } 2322 2323 //----------------------------inline_unsafe_access---------------------------- 2324 2325 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2326 // Attempt to infer a sharper value type from the offset and base type. 2327 ciKlass* sharpened_klass = NULL; 2328 2329 // See if it is an instance field, with an object type. 2330 if (alias_type->field() != NULL) { 2331 if (alias_type->field()->type()->is_klass()) { 2332 sharpened_klass = alias_type->field()->type()->as_klass(); 2333 } 2334 } 2335 2336 // See if it is a narrow oop array. 2337 if (adr_type->isa_aryptr()) { 2338 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2339 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2340 if (elem_type != NULL) { 2341 sharpened_klass = elem_type->klass(); 2342 } 2343 } 2344 } 2345 2346 // The sharpened class might be unloaded if there is no class loader 2347 // contraint in place. 2348 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2349 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2350 2351 #ifndef PRODUCT 2352 if (C->print_intrinsics() || C->print_inlining()) { 2353 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2354 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2355 } 2356 #endif 2357 // Sharpen the value type. 2358 return tjp; 2359 } 2360 return NULL; 2361 } 2362 2363 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) { 2364 switch (kind) { 2365 case Relaxed: 2366 return MO_UNORDERED; 2367 case Opaque: 2368 return MO_RELAXED; 2369 case Acquire: 2370 return MO_ACQUIRE; 2371 case Release: 2372 return MO_RELEASE; 2373 case Volatile: 2374 return MO_SEQ_CST; 2375 default: 2376 ShouldNotReachHere(); 2377 return 0; 2378 } 2379 } 2380 2381 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2382 if (callee()->is_static()) return false; // caller must have the capability! 2383 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2384 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2385 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2386 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2387 2388 if (is_reference_type(type)) { 2389 decorators |= ON_UNKNOWN_OOP_REF; 2390 } 2391 2392 if (unaligned) { 2393 decorators |= C2_UNALIGNED; 2394 } 2395 2396 #ifndef PRODUCT 2397 { 2398 ResourceMark rm; 2399 // Check the signatures. 2400 ciSignature* sig = callee()->signature(); 2401 #ifdef ASSERT 2402 if (!is_store) { 2403 // Object getReference(Object base, int/long offset), etc. 2404 BasicType rtype = sig->return_type()->basic_type(); 2405 assert(rtype == type, "getter must return the expected value"); 2406 assert(sig->count() == 2, "oop getter has 2 arguments"); 2407 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2408 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2409 } else { 2410 // void putReference(Object base, int/long offset, Object x), etc. 2411 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2412 assert(sig->count() == 3, "oop putter has 3 arguments"); 2413 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2414 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2415 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2416 assert(vtype == type, "putter must accept the expected value"); 2417 } 2418 #endif // ASSERT 2419 } 2420 #endif //PRODUCT 2421 2422 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2423 2424 Node* receiver = argument(0); // type: oop 2425 2426 // Build address expression. 2427 Node* adr; 2428 Node* heap_base_oop = top(); 2429 Node* offset = top(); 2430 Node* val; 2431 2432 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2433 Node* base = argument(1); // type: oop 2434 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2435 offset = argument(2); // type: long 2436 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2437 // to be plain byte offsets, which are also the same as those accepted 2438 // by oopDesc::field_addr. 2439 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2440 "fieldOffset must be byte-scaled"); 2441 // 32-bit machines ignore the high half! 2442 offset = ConvL2X(offset); 2443 adr = make_unsafe_address(base, offset, is_store ? ACCESS_WRITE : ACCESS_READ, type, kind == Relaxed); 2444 2445 if (_gvn.type(base)->isa_ptr() == TypePtr::NULL_PTR) { 2446 if (type != T_OBJECT) { 2447 decorators |= IN_NATIVE; // off-heap primitive access 2448 } else { 2449 return false; // off-heap oop accesses are not supported 2450 } 2451 } else { 2452 heap_base_oop = base; // on-heap or mixed access 2453 } 2454 2455 // Can base be NULL? Otherwise, always on-heap access. 2456 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(base)); 2457 2458 if (!can_access_non_heap) { 2459 decorators |= IN_HEAP; 2460 } 2461 2462 val = is_store ? argument(4) : NULL; 2463 2464 const TypePtr* adr_type = _gvn.type(adr)->isa_ptr(); 2465 if (adr_type == TypePtr::NULL_PTR) { 2466 return false; // off-heap access with zero address 2467 } 2468 2469 // Try to categorize the address. 2470 Compile::AliasType* alias_type = C->alias_type(adr_type); 2471 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2472 2473 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2474 alias_type->adr_type() == TypeAryPtr::RANGE) { 2475 return false; // not supported 2476 } 2477 2478 bool mismatched = false; 2479 BasicType bt = alias_type->basic_type(); 2480 if (bt != T_ILLEGAL) { 2481 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2482 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2483 // Alias type doesn't differentiate between byte[] and boolean[]). 2484 // Use address type to get the element type. 2485 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2486 } 2487 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2488 // accessing an array field with getReference is not a mismatch 2489 bt = T_OBJECT; 2490 } 2491 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2492 // Don't intrinsify mismatched object accesses 2493 return false; 2494 } 2495 mismatched = (bt != type); 2496 } else if (alias_type->adr_type()->isa_oopptr()) { 2497 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched 2498 } 2499 2500 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2501 2502 if (mismatched) { 2503 decorators |= C2_MISMATCHED; 2504 } 2505 2506 // First guess at the value type. 2507 const Type *value_type = Type::get_const_basic_type(type); 2508 2509 // Figure out the memory ordering. 2510 decorators |= mo_decorator_for_access_kind(kind); 2511 2512 if (!is_store && type == T_OBJECT) { 2513 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2514 if (tjp != NULL) { 2515 value_type = tjp; 2516 } 2517 } 2518 2519 receiver = null_check(receiver); 2520 if (stopped()) { 2521 return true; 2522 } 2523 // Heap pointers get a null-check from the interpreter, 2524 // as a courtesy. However, this is not guaranteed by Unsafe, 2525 // and it is not possible to fully distinguish unintended nulls 2526 // from intended ones in this API. 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() && field != NULL && field->is_constant() && !mismatched) { 2533 // final or stable field 2534 p = make_constant_from_field(field, heap_base_oop); 2535 } 2536 2537 if (p == NULL) { // Could not constant fold the load 2538 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators); 2539 // Normalize the value returned by getBoolean in the following cases 2540 if (type == T_BOOLEAN && 2541 (mismatched || 2542 heap_base_oop == top() || // - heap_base_oop is NULL or 2543 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL 2544 // and the unsafe access is made to large offset 2545 // (i.e., larger than the maximum offset necessary for any 2546 // field access) 2547 ) { 2548 IdealKit ideal = IdealKit(this); 2549 #define __ ideal. 2550 IdealVariable normalized_result(ideal); 2551 __ declarations_done(); 2552 __ set(normalized_result, p); 2553 __ if_then(p, BoolTest::ne, ideal.ConI(0)); 2554 __ set(normalized_result, ideal.ConI(1)); 2555 ideal.end_if(); 2556 final_sync(ideal); 2557 p = __ value(normalized_result); 2558 #undef __ 2559 } 2560 } 2561 if (type == T_ADDRESS) { 2562 p = gvn().transform(new CastP2XNode(NULL, p)); 2563 p = ConvX2UL(p); 2564 } 2565 // The load node has the control of the preceding MemBarCPUOrder. All 2566 // following nodes will have the control of the MemBarCPUOrder inserted at 2567 // the end of this method. So, pushing the load onto the stack at a later 2568 // point is fine. 2569 set_result(p); 2570 } else { 2571 if (bt == T_ADDRESS) { 2572 // Repackage the long as a pointer. 2573 val = ConvL2X(val); 2574 val = gvn().transform(new CastX2PNode(val)); 2575 } 2576 access_store_at(heap_base_oop, adr, adr_type, val, value_type, type, decorators); 2577 } 2578 2579 return true; 2580 } 2581 2582 //----------------------------inline_unsafe_load_store---------------------------- 2583 // This method serves a couple of different customers (depending on LoadStoreKind): 2584 // 2585 // LS_cmp_swap: 2586 // 2587 // boolean compareAndSetReference(Object o, long offset, Object expected, Object x); 2588 // boolean compareAndSetInt( Object o, long offset, int expected, int x); 2589 // boolean compareAndSetLong( Object o, long offset, long expected, long x); 2590 // 2591 // LS_cmp_swap_weak: 2592 // 2593 // boolean weakCompareAndSetReference( Object o, long offset, Object expected, Object x); 2594 // boolean weakCompareAndSetReferencePlain( Object o, long offset, Object expected, Object x); 2595 // boolean weakCompareAndSetReferenceAcquire(Object o, long offset, Object expected, Object x); 2596 // boolean weakCompareAndSetReferenceRelease(Object o, long offset, Object expected, Object x); 2597 // 2598 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x); 2599 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x); 2600 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x); 2601 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x); 2602 // 2603 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x); 2604 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x); 2605 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x); 2606 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x); 2607 // 2608 // LS_cmp_exchange: 2609 // 2610 // Object compareAndExchangeReferenceVolatile(Object o, long offset, Object expected, Object x); 2611 // Object compareAndExchangeReferenceAcquire( Object o, long offset, Object expected, Object x); 2612 // Object compareAndExchangeReferenceRelease( Object o, long offset, Object expected, Object x); 2613 // 2614 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2615 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2616 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2617 // 2618 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2619 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2620 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2621 // 2622 // LS_get_add: 2623 // 2624 // int getAndAddInt( Object o, long offset, int delta) 2625 // long getAndAddLong(Object o, long offset, long delta) 2626 // 2627 // LS_get_set: 2628 // 2629 // int getAndSet(Object o, long offset, int newValue) 2630 // long getAndSet(Object o, long offset, long newValue) 2631 // Object getAndSet(Object o, long offset, Object newValue) 2632 // 2633 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2634 // This basic scheme here is the same as inline_unsafe_access, but 2635 // differs in enough details that combining them would make the code 2636 // overly confusing. (This is a true fact! I originally combined 2637 // them, but even I was confused by it!) As much code/comments as 2638 // possible are retained from inline_unsafe_access though to make 2639 // the correspondences clearer. - dl 2640 2641 if (callee()->is_static()) return false; // caller must have the capability! 2642 2643 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2644 decorators |= mo_decorator_for_access_kind(access_kind); 2645 2646 #ifndef PRODUCT 2647 BasicType rtype; 2648 { 2649 ResourceMark rm; 2650 // Check the signatures. 2651 ciSignature* sig = callee()->signature(); 2652 rtype = sig->return_type()->basic_type(); 2653 switch(kind) { 2654 case LS_get_add: 2655 case LS_get_set: { 2656 // Check the signatures. 2657 #ifdef ASSERT 2658 assert(rtype == type, "get and set must return the expected type"); 2659 assert(sig->count() == 3, "get and set has 3 arguments"); 2660 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2661 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2662 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2663 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2664 #endif // ASSERT 2665 break; 2666 } 2667 case LS_cmp_swap: 2668 case LS_cmp_swap_weak: { 2669 // Check the signatures. 2670 #ifdef ASSERT 2671 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2672 assert(sig->count() == 4, "CAS has 4 arguments"); 2673 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2674 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2675 #endif // ASSERT 2676 break; 2677 } 2678 case LS_cmp_exchange: { 2679 // Check the signatures. 2680 #ifdef ASSERT 2681 assert(rtype == type, "CAS must return the expected type"); 2682 assert(sig->count() == 4, "CAS has 4 arguments"); 2683 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2684 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2685 #endif // ASSERT 2686 break; 2687 } 2688 default: 2689 ShouldNotReachHere(); 2690 } 2691 } 2692 #endif //PRODUCT 2693 2694 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2695 2696 // Get arguments: 2697 Node* receiver = NULL; 2698 Node* base = NULL; 2699 Node* offset = NULL; 2700 Node* oldval = NULL; 2701 Node* newval = NULL; 2702 switch(kind) { 2703 case LS_cmp_swap: 2704 case LS_cmp_swap_weak: 2705 case LS_cmp_exchange: { 2706 const bool two_slot_type = type2size[type] == 2; 2707 receiver = argument(0); // type: oop 2708 base = argument(1); // type: oop 2709 offset = argument(2); // type: long 2710 oldval = argument(4); // type: oop, int, or long 2711 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2712 break; 2713 } 2714 case LS_get_add: 2715 case LS_get_set: { 2716 receiver = argument(0); // type: oop 2717 base = argument(1); // type: oop 2718 offset = argument(2); // type: long 2719 oldval = NULL; 2720 newval = argument(4); // type: oop, int, or long 2721 break; 2722 } 2723 default: 2724 ShouldNotReachHere(); 2725 } 2726 2727 // Build field offset expression. 2728 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2729 // to be plain byte offsets, which are also the same as those accepted 2730 // by oopDesc::field_addr. 2731 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2732 // 32-bit machines ignore the high half of long offsets 2733 offset = ConvL2X(offset); 2734 Node* adr = make_unsafe_address(base, offset, ACCESS_WRITE | ACCESS_READ, type, false); 2735 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2736 2737 Compile::AliasType* alias_type = C->alias_type(adr_type); 2738 BasicType bt = alias_type->basic_type(); 2739 if (bt != T_ILLEGAL && 2740 (is_reference_type(bt) != (type == T_OBJECT))) { 2741 // Don't intrinsify mismatched object accesses. 2742 return false; 2743 } 2744 2745 // For CAS, unlike inline_unsafe_access, there seems no point in 2746 // trying to refine types. Just use the coarse types here. 2747 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2748 const Type *value_type = Type::get_const_basic_type(type); 2749 2750 switch (kind) { 2751 case LS_get_set: 2752 case LS_cmp_exchange: { 2753 if (type == T_OBJECT) { 2754 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2755 if (tjp != NULL) { 2756 value_type = tjp; 2757 } 2758 } 2759 break; 2760 } 2761 case LS_cmp_swap: 2762 case LS_cmp_swap_weak: 2763 case LS_get_add: 2764 break; 2765 default: 2766 ShouldNotReachHere(); 2767 } 2768 2769 // Null check receiver. 2770 receiver = null_check(receiver); 2771 if (stopped()) { 2772 return true; 2773 } 2774 2775 int alias_idx = C->get_alias_index(adr_type); 2776 2777 if (is_reference_type(type)) { 2778 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF; 2779 2780 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2781 // could be delayed during Parse (for example, in adjust_map_after_if()). 2782 // Execute transformation here to avoid barrier generation in such case. 2783 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2784 newval = _gvn.makecon(TypePtr::NULL_PTR); 2785 2786 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) { 2787 // Refine the value to a null constant, when it is known to be null 2788 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2789 } 2790 } 2791 2792 Node* result = NULL; 2793 switch (kind) { 2794 case LS_cmp_exchange: { 2795 result = access_atomic_cmpxchg_val_at(base, adr, adr_type, alias_idx, 2796 oldval, newval, value_type, type, decorators); 2797 break; 2798 } 2799 case LS_cmp_swap_weak: 2800 decorators |= C2_WEAK_CMPXCHG; 2801 case LS_cmp_swap: { 2802 result = access_atomic_cmpxchg_bool_at(base, adr, adr_type, alias_idx, 2803 oldval, newval, value_type, type, decorators); 2804 break; 2805 } 2806 case LS_get_set: { 2807 result = access_atomic_xchg_at(base, adr, adr_type, alias_idx, 2808 newval, value_type, type, decorators); 2809 break; 2810 } 2811 case LS_get_add: { 2812 result = access_atomic_add_at(base, adr, adr_type, alias_idx, 2813 newval, value_type, type, decorators); 2814 break; 2815 } 2816 default: 2817 ShouldNotReachHere(); 2818 } 2819 2820 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2821 set_result(result); 2822 return true; 2823 } 2824 2825 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2826 // Regardless of form, don't allow previous ld/st to move down, 2827 // then issue acquire, release, or volatile mem_bar. 2828 insert_mem_bar(Op_MemBarCPUOrder); 2829 switch(id) { 2830 case vmIntrinsics::_loadFence: 2831 insert_mem_bar(Op_LoadFence); 2832 return true; 2833 case vmIntrinsics::_storeFence: 2834 insert_mem_bar(Op_StoreFence); 2835 return true; 2836 case vmIntrinsics::_fullFence: 2837 insert_mem_bar(Op_MemBarVolatile); 2838 return true; 2839 default: 2840 fatal_unexpected_iid(id); 2841 return false; 2842 } 2843 } 2844 2845 bool LibraryCallKit::inline_onspinwait() { 2846 insert_mem_bar(Op_OnSpinWait); 2847 return true; 2848 } 2849 2850 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 2851 if (!kls->is_Con()) { 2852 return true; 2853 } 2854 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 2855 if (klsptr == NULL) { 2856 return true; 2857 } 2858 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 2859 // don't need a guard for a klass that is already initialized 2860 return !ik->is_initialized(); 2861 } 2862 2863 //----------------------------inline_unsafe_writeback0------------------------- 2864 // public native void Unsafe.writeback0(long address) 2865 bool LibraryCallKit::inline_unsafe_writeback0() { 2866 if (!Matcher::has_match_rule(Op_CacheWB)) { 2867 return false; 2868 } 2869 #ifndef PRODUCT 2870 assert(Matcher::has_match_rule(Op_CacheWBPreSync), "found match rule for CacheWB but not CacheWBPreSync"); 2871 assert(Matcher::has_match_rule(Op_CacheWBPostSync), "found match rule for CacheWB but not CacheWBPostSync"); 2872 ciSignature* sig = callee()->signature(); 2873 assert(sig->type_at(0)->basic_type() == T_LONG, "Unsafe_writeback0 address is long!"); 2874 #endif 2875 null_check_receiver(); // null-check, then ignore 2876 Node *addr = argument(1); 2877 addr = new CastX2PNode(addr); 2878 addr = _gvn.transform(addr); 2879 Node *flush = new CacheWBNode(control(), memory(TypeRawPtr::BOTTOM), addr); 2880 flush = _gvn.transform(flush); 2881 set_memory(flush, TypeRawPtr::BOTTOM); 2882 return true; 2883 } 2884 2885 //----------------------------inline_unsafe_writeback0------------------------- 2886 // public native void Unsafe.writeback0(long address) 2887 bool LibraryCallKit::inline_unsafe_writebackSync0(bool is_pre) { 2888 if (is_pre && !Matcher::has_match_rule(Op_CacheWBPreSync)) { 2889 return false; 2890 } 2891 if (!is_pre && !Matcher::has_match_rule(Op_CacheWBPostSync)) { 2892 return false; 2893 } 2894 #ifndef PRODUCT 2895 assert(Matcher::has_match_rule(Op_CacheWB), 2896 (is_pre ? "found match rule for CacheWBPreSync but not CacheWB" 2897 : "found match rule for CacheWBPostSync but not CacheWB")); 2898 2899 #endif 2900 null_check_receiver(); // null-check, then ignore 2901 Node *sync; 2902 if (is_pre) { 2903 sync = new CacheWBPreSyncNode(control(), memory(TypeRawPtr::BOTTOM)); 2904 } else { 2905 sync = new CacheWBPostSyncNode(control(), memory(TypeRawPtr::BOTTOM)); 2906 } 2907 sync = _gvn.transform(sync); 2908 set_memory(sync, TypeRawPtr::BOTTOM); 2909 return true; 2910 } 2911 2912 //----------------------------inline_unsafe_allocate--------------------------- 2913 // public native Object Unsafe.allocateInstance(Class<?> cls); 2914 bool LibraryCallKit::inline_unsafe_allocate() { 2915 if (callee()->is_static()) return false; // caller must have the capability! 2916 2917 null_check_receiver(); // null-check, then ignore 2918 Node* cls = null_check(argument(1)); 2919 if (stopped()) return true; 2920 2921 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2922 kls = null_check(kls); 2923 if (stopped()) return true; // argument was like int.class 2924 2925 Node* test = NULL; 2926 if (LibraryCallKit::klass_needs_init_guard(kls)) { 2927 // Note: The argument might still be an illegal value like 2928 // Serializable.class or Object[].class. The runtime will handle it. 2929 // But we must make an explicit check for initialization. 2930 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 2931 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 2932 // can generate code to load it as unsigned byte. 2933 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 2934 Node* bits = intcon(InstanceKlass::fully_initialized); 2935 test = _gvn.transform(new SubINode(inst, bits)); 2936 // The 'test' is non-zero if we need to take a slow path. 2937 } 2938 2939 Node* obj = new_instance(kls, test); 2940 set_result(obj); 2941 return true; 2942 } 2943 2944 //------------------------inline_native_time_funcs-------------- 2945 // inline code for System.currentTimeMillis() and System.nanoTime() 2946 // these have the same type and signature 2947 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 2948 const TypeFunc* tf = OptoRuntime::void_long_Type(); 2949 const TypePtr* no_memory_effects = NULL; 2950 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2951 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 2952 #ifdef ASSERT 2953 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 2954 assert(value_top == top(), "second value must be top"); 2955 #endif 2956 set_result(value); 2957 return true; 2958 } 2959 2960 #ifdef JFR_HAVE_INTRINSICS 2961 2962 /* 2963 * oop -> myklass 2964 * myklass->trace_id |= USED 2965 * return myklass->trace_id & ~0x3 2966 */ 2967 bool LibraryCallKit::inline_native_classID() { 2968 Node* cls = null_check(argument(0), T_OBJECT); 2969 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2970 kls = null_check(kls, T_OBJECT); 2971 2972 ByteSize offset = KLASS_TRACE_ID_OFFSET; 2973 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 2974 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 2975 2976 Node* clsused = longcon(0x01l); // set the class bit 2977 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 2978 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 2979 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 2980 2981 #ifdef TRACE_ID_META_BITS 2982 Node* mbits = longcon(~TRACE_ID_META_BITS); 2983 tvalue = _gvn.transform(new AndLNode(tvalue, mbits)); 2984 #endif 2985 #ifdef TRACE_ID_SHIFT 2986 Node* cbits = intcon(TRACE_ID_SHIFT); 2987 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits)); 2988 #endif 2989 2990 set_result(tvalue); 2991 return true; 2992 2993 } 2994 2995 bool LibraryCallKit::inline_native_getEventWriter() { 2996 Node* tls_ptr = _gvn.transform(new ThreadLocalNode()); 2997 2998 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr, 2999 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)); 3000 3001 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered); 3002 3003 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) ); 3004 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) ); 3005 3006 IfNode* iff_jobj_null = 3007 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN); 3008 3009 enum { _normal_path = 1, 3010 _null_path = 2, 3011 PATH_LIMIT }; 3012 3013 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3014 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM); 3015 3016 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null)); 3017 result_rgn->init_req(_null_path, jobj_is_null); 3018 result_val->init_req(_null_path, null()); 3019 3020 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null)); 3021 set_control(jobj_is_not_null); 3022 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT, 3023 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD); 3024 result_rgn->init_req(_normal_path, control()); 3025 result_val->init_req(_normal_path, res); 3026 3027 set_result(result_rgn, result_val); 3028 3029 return true; 3030 } 3031 3032 #endif // JFR_HAVE_INTRINSICS 3033 3034 //------------------------inline_native_currentThread------------------ 3035 bool LibraryCallKit::inline_native_currentThread() { 3036 Node* junk = NULL; 3037 set_result(generate_current_thread(junk)); 3038 return true; 3039 } 3040 3041 //---------------------------load_mirror_from_klass---------------------------- 3042 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3043 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3044 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3045 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3046 // mirror = ((OopHandle)mirror)->resolve(); 3047 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); 3048 } 3049 3050 //-----------------------load_klass_from_mirror_common------------------------- 3051 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3052 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3053 // and branch to the given path on the region. 3054 // If never_see_null, take an uncommon trap on null, so we can optimistically 3055 // compile for the non-null case. 3056 // If the region is NULL, force never_see_null = true. 3057 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3058 bool never_see_null, 3059 RegionNode* region, 3060 int null_path, 3061 int offset) { 3062 if (region == NULL) never_see_null = true; 3063 Node* p = basic_plus_adr(mirror, offset); 3064 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3065 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3066 Node* null_ctl = top(); 3067 kls = null_check_oop(kls, &null_ctl, never_see_null); 3068 if (region != NULL) { 3069 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3070 region->init_req(null_path, null_ctl); 3071 } else { 3072 assert(null_ctl == top(), "no loose ends"); 3073 } 3074 return kls; 3075 } 3076 3077 //--------------------(inline_native_Class_query helpers)--------------------- 3078 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3079 // Fall through if (mods & mask) == bits, take the guard otherwise. 3080 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3081 // Branch around if the given klass has the given modifier bit set. 3082 // Like generate_guard, adds a new path onto the region. 3083 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3084 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3085 Node* mask = intcon(modifier_mask); 3086 Node* bits = intcon(modifier_bits); 3087 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3088 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3089 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3090 return generate_fair_guard(bol, region); 3091 } 3092 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3093 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3094 } 3095 Node* LibraryCallKit::generate_hidden_class_guard(Node* kls, RegionNode* region) { 3096 return generate_access_flags_guard(kls, JVM_ACC_IS_HIDDEN_CLASS, 0, region); 3097 } 3098 3099 //-------------------------inline_native_Class_query------------------- 3100 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3101 const Type* return_type = TypeInt::BOOL; 3102 Node* prim_return_value = top(); // what happens if it's a primitive class? 3103 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3104 bool expect_prim = false; // most of these guys expect to work on refs 3105 3106 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3107 3108 Node* mirror = argument(0); 3109 Node* obj = top(); 3110 3111 switch (id) { 3112 case vmIntrinsics::_isInstance: 3113 // nothing is an instance of a primitive type 3114 prim_return_value = intcon(0); 3115 obj = argument(1); 3116 break; 3117 case vmIntrinsics::_getModifiers: 3118 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3119 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3120 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3121 break; 3122 case vmIntrinsics::_isInterface: 3123 prim_return_value = intcon(0); 3124 break; 3125 case vmIntrinsics::_isArray: 3126 prim_return_value = intcon(0); 3127 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3128 break; 3129 case vmIntrinsics::_isPrimitive: 3130 prim_return_value = intcon(1); 3131 expect_prim = true; // obviously 3132 break; 3133 case vmIntrinsics::_isHidden: 3134 prim_return_value = intcon(0); 3135 break; 3136 case vmIntrinsics::_getSuperclass: 3137 prim_return_value = null(); 3138 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3139 break; 3140 case vmIntrinsics::_getClassAccessFlags: 3141 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3142 return_type = TypeInt::INT; // not bool! 6297094 3143 break; 3144 default: 3145 fatal_unexpected_iid(id); 3146 break; 3147 } 3148 3149 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3150 if (mirror_con == NULL) return false; // cannot happen? 3151 3152 #ifndef PRODUCT 3153 if (C->print_intrinsics() || C->print_inlining()) { 3154 ciType* k = mirror_con->java_mirror_type(); 3155 if (k) { 3156 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3157 k->print_name(); 3158 tty->cr(); 3159 } 3160 } 3161 #endif 3162 3163 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3164 RegionNode* region = new RegionNode(PATH_LIMIT); 3165 record_for_igvn(region); 3166 PhiNode* phi = new PhiNode(region, return_type); 3167 3168 // The mirror will never be null of Reflection.getClassAccessFlags, however 3169 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3170 // if it is. See bug 4774291. 3171 3172 // For Reflection.getClassAccessFlags(), the null check occurs in 3173 // the wrong place; see inline_unsafe_access(), above, for a similar 3174 // situation. 3175 mirror = null_check(mirror); 3176 // If mirror or obj is dead, only null-path is taken. 3177 if (stopped()) return true; 3178 3179 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3180 3181 // Now load the mirror's klass metaobject, and null-check it. 3182 // Side-effects region with the control path if the klass is null. 3183 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3184 // If kls is null, we have a primitive mirror. 3185 phi->init_req(_prim_path, prim_return_value); 3186 if (stopped()) { set_result(region, phi); return true; } 3187 bool safe_for_replace = (region->in(_prim_path) == top()); 3188 3189 Node* p; // handy temp 3190 Node* null_ctl; 3191 3192 // Now that we have the non-null klass, we can perform the real query. 3193 // For constant classes, the query will constant-fold in LoadNode::Value. 3194 Node* query_value = top(); 3195 switch (id) { 3196 case vmIntrinsics::_isInstance: 3197 // nothing is an instance of a primitive type 3198 query_value = gen_instanceof(obj, kls, safe_for_replace); 3199 break; 3200 3201 case vmIntrinsics::_getModifiers: 3202 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3203 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3204 break; 3205 3206 case vmIntrinsics::_isInterface: 3207 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3208 if (generate_interface_guard(kls, region) != NULL) 3209 // A guard was added. If the guard is taken, it was an interface. 3210 phi->add_req(intcon(1)); 3211 // If we fall through, it's a plain class. 3212 query_value = intcon(0); 3213 break; 3214 3215 case vmIntrinsics::_isArray: 3216 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3217 if (generate_array_guard(kls, region) != NULL) 3218 // A guard was added. If the guard is taken, it was an array. 3219 phi->add_req(intcon(1)); 3220 // If we fall through, it's a plain class. 3221 query_value = intcon(0); 3222 break; 3223 3224 case vmIntrinsics::_isPrimitive: 3225 query_value = intcon(0); // "normal" path produces false 3226 break; 3227 3228 case vmIntrinsics::_isHidden: 3229 // (To verify this code sequence, check the asserts in JVM_IsHiddenClass.) 3230 if (generate_hidden_class_guard(kls, region) != NULL) 3231 // A guard was added. If the guard is taken, it was an hidden class. 3232 phi->add_req(intcon(1)); 3233 // If we fall through, it's a plain class. 3234 query_value = intcon(0); 3235 break; 3236 3237 3238 case vmIntrinsics::_getSuperclass: 3239 // The rules here are somewhat unfortunate, but we can still do better 3240 // with random logic than with a JNI call. 3241 // Interfaces store null or Object as _super, but must report null. 3242 // Arrays store an intermediate super as _super, but must report Object. 3243 // Other types can report the actual _super. 3244 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3245 if (generate_interface_guard(kls, region) != NULL) 3246 // A guard was added. If the guard is taken, it was an interface. 3247 phi->add_req(null()); 3248 if (generate_array_guard(kls, region) != NULL) 3249 // A guard was added. If the guard is taken, it was an array. 3250 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3251 // If we fall through, it's a plain class. Get its _super. 3252 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3253 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3254 null_ctl = top(); 3255 kls = null_check_oop(kls, &null_ctl); 3256 if (null_ctl != top()) { 3257 // If the guard is taken, Object.superClass is null (both klass and mirror). 3258 region->add_req(null_ctl); 3259 phi ->add_req(null()); 3260 } 3261 if (!stopped()) { 3262 query_value = load_mirror_from_klass(kls); 3263 } 3264 break; 3265 3266 case vmIntrinsics::_getClassAccessFlags: 3267 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3268 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3269 break; 3270 3271 default: 3272 fatal_unexpected_iid(id); 3273 break; 3274 } 3275 3276 // Fall-through is the normal case of a query to a real class. 3277 phi->init_req(1, query_value); 3278 region->init_req(1, control()); 3279 3280 C->set_has_split_ifs(true); // Has chance for split-if optimization 3281 set_result(region, phi); 3282 return true; 3283 } 3284 3285 //-------------------------inline_Class_cast------------------- 3286 bool LibraryCallKit::inline_Class_cast() { 3287 Node* mirror = argument(0); // Class 3288 Node* obj = argument(1); 3289 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3290 if (mirror_con == NULL) { 3291 return false; // dead path (mirror->is_top()). 3292 } 3293 if (obj == NULL || obj->is_top()) { 3294 return false; // dead path 3295 } 3296 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3297 3298 // First, see if Class.cast() can be folded statically. 3299 // java_mirror_type() returns non-null for compile-time Class constants. 3300 ciType* tm = mirror_con->java_mirror_type(); 3301 if (tm != NULL && tm->is_klass() && 3302 tp != NULL && tp->klass() != NULL) { 3303 if (!tp->klass()->is_loaded()) { 3304 // Don't use intrinsic when class is not loaded. 3305 return false; 3306 } else { 3307 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3308 if (static_res == Compile::SSC_always_true) { 3309 // isInstance() is true - fold the code. 3310 set_result(obj); 3311 return true; 3312 } else if (static_res == Compile::SSC_always_false) { 3313 // Don't use intrinsic, have to throw ClassCastException. 3314 // If the reference is null, the non-intrinsic bytecode will 3315 // be optimized appropriately. 3316 return false; 3317 } 3318 } 3319 } 3320 3321 // Bailout intrinsic and do normal inlining if exception path is frequent. 3322 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3323 return false; 3324 } 3325 3326 // Generate dynamic checks. 3327 // Class.cast() is java implementation of _checkcast bytecode. 3328 // Do checkcast (Parse::do_checkcast()) optimizations here. 3329 3330 mirror = null_check(mirror); 3331 // If mirror is dead, only null-path is taken. 3332 if (stopped()) { 3333 return true; 3334 } 3335 3336 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3337 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3338 RegionNode* region = new RegionNode(PATH_LIMIT); 3339 record_for_igvn(region); 3340 3341 // Now load the mirror's klass metaobject, and null-check it. 3342 // If kls is null, we have a primitive mirror and 3343 // nothing is an instance of a primitive type. 3344 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3345 3346 Node* res = top(); 3347 if (!stopped()) { 3348 Node* bad_type_ctrl = top(); 3349 // Do checkcast optimizations. 3350 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3351 region->init_req(_bad_type_path, bad_type_ctrl); 3352 } 3353 if (region->in(_prim_path) != top() || 3354 region->in(_bad_type_path) != top()) { 3355 // Let Interpreter throw ClassCastException. 3356 PreserveJVMState pjvms(this); 3357 set_control(_gvn.transform(region)); 3358 uncommon_trap(Deoptimization::Reason_intrinsic, 3359 Deoptimization::Action_maybe_recompile); 3360 } 3361 if (!stopped()) { 3362 set_result(res); 3363 } 3364 return true; 3365 } 3366 3367 3368 //--------------------------inline_native_subtype_check------------------------ 3369 // This intrinsic takes the JNI calls out of the heart of 3370 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3371 bool LibraryCallKit::inline_native_subtype_check() { 3372 // Pull both arguments off the stack. 3373 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3374 args[0] = argument(0); 3375 args[1] = argument(1); 3376 Node* klasses[2]; // corresponding Klasses: superk, subk 3377 klasses[0] = klasses[1] = top(); 3378 3379 enum { 3380 // A full decision tree on {superc is prim, subc is prim}: 3381 _prim_0_path = 1, // {P,N} => false 3382 // {P,P} & superc!=subc => false 3383 _prim_same_path, // {P,P} & superc==subc => true 3384 _prim_1_path, // {N,P} => false 3385 _ref_subtype_path, // {N,N} & subtype check wins => true 3386 _both_ref_path, // {N,N} & subtype check loses => false 3387 PATH_LIMIT 3388 }; 3389 3390 RegionNode* region = new RegionNode(PATH_LIMIT); 3391 Node* phi = new PhiNode(region, TypeInt::BOOL); 3392 record_for_igvn(region); 3393 3394 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3395 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3396 int class_klass_offset = java_lang_Class::klass_offset(); 3397 3398 // First null-check both mirrors and load each mirror's klass metaobject. 3399 int which_arg; 3400 for (which_arg = 0; which_arg <= 1; which_arg++) { 3401 Node* arg = args[which_arg]; 3402 arg = null_check(arg); 3403 if (stopped()) break; 3404 args[which_arg] = arg; 3405 3406 Node* p = basic_plus_adr(arg, class_klass_offset); 3407 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3408 klasses[which_arg] = _gvn.transform(kls); 3409 } 3410 3411 // Having loaded both klasses, test each for null. 3412 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3413 for (which_arg = 0; which_arg <= 1; which_arg++) { 3414 Node* kls = klasses[which_arg]; 3415 Node* null_ctl = top(); 3416 kls = null_check_oop(kls, &null_ctl, never_see_null); 3417 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3418 region->init_req(prim_path, null_ctl); 3419 if (stopped()) break; 3420 klasses[which_arg] = kls; 3421 } 3422 3423 if (!stopped()) { 3424 // now we have two reference types, in klasses[0..1] 3425 Node* subk = klasses[1]; // the argument to isAssignableFrom 3426 Node* superk = klasses[0]; // the receiver 3427 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3428 // now we have a successful reference subtype check 3429 region->set_req(_ref_subtype_path, control()); 3430 } 3431 3432 // If both operands are primitive (both klasses null), then 3433 // we must return true when they are identical primitives. 3434 // It is convenient to test this after the first null klass check. 3435 set_control(region->in(_prim_0_path)); // go back to first null check 3436 if (!stopped()) { 3437 // Since superc is primitive, make a guard for the superc==subc case. 3438 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3439 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3440 generate_guard(bol_eq, region, PROB_FAIR); 3441 if (region->req() == PATH_LIMIT+1) { 3442 // A guard was added. If the added guard is taken, superc==subc. 3443 region->swap_edges(PATH_LIMIT, _prim_same_path); 3444 region->del_req(PATH_LIMIT); 3445 } 3446 region->set_req(_prim_0_path, control()); // Not equal after all. 3447 } 3448 3449 // these are the only paths that produce 'true': 3450 phi->set_req(_prim_same_path, intcon(1)); 3451 phi->set_req(_ref_subtype_path, intcon(1)); 3452 3453 // pull together the cases: 3454 assert(region->req() == PATH_LIMIT, "sane region"); 3455 for (uint i = 1; i < region->req(); i++) { 3456 Node* ctl = region->in(i); 3457 if (ctl == NULL || ctl == top()) { 3458 region->set_req(i, top()); 3459 phi ->set_req(i, top()); 3460 } else if (phi->in(i) == NULL) { 3461 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3462 } 3463 } 3464 3465 set_control(_gvn.transform(region)); 3466 set_result(_gvn.transform(phi)); 3467 return true; 3468 } 3469 3470 //---------------------generate_array_guard_common------------------------ 3471 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3472 bool obj_array, bool not_array) { 3473 3474 if (stopped()) { 3475 return NULL; 3476 } 3477 3478 // If obj_array/non_array==false/false: 3479 // Branch around if the given klass is in fact an array (either obj or prim). 3480 // If obj_array/non_array==false/true: 3481 // Branch around if the given klass is not an array klass of any kind. 3482 // If obj_array/non_array==true/true: 3483 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3484 // If obj_array/non_array==true/false: 3485 // Branch around if the kls is an oop array (Object[] or subtype) 3486 // 3487 // Like generate_guard, adds a new path onto the region. 3488 jint layout_con = 0; 3489 Node* layout_val = get_layout_helper(kls, layout_con); 3490 if (layout_val == NULL) { 3491 bool query = (obj_array 3492 ? Klass::layout_helper_is_objArray(layout_con) 3493 : Klass::layout_helper_is_array(layout_con)); 3494 if (query == not_array) { 3495 return NULL; // never a branch 3496 } else { // always a branch 3497 Node* always_branch = control(); 3498 if (region != NULL) 3499 region->add_req(always_branch); 3500 set_control(top()); 3501 return always_branch; 3502 } 3503 } 3504 // Now test the correct condition. 3505 jint nval = (obj_array 3506 ? (jint)(Klass::_lh_array_tag_type_value 3507 << Klass::_lh_array_tag_shift) 3508 : Klass::_lh_neutral_value); 3509 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3510 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3511 // invert the test if we are looking for a non-array 3512 if (not_array) btest = BoolTest(btest).negate(); 3513 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3514 return generate_fair_guard(bol, region); 3515 } 3516 3517 3518 //-----------------------inline_native_newArray-------------------------- 3519 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3520 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 3521 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 3522 Node* mirror; 3523 Node* count_val; 3524 if (uninitialized) { 3525 mirror = argument(1); 3526 count_val = argument(2); 3527 } else { 3528 mirror = argument(0); 3529 count_val = argument(1); 3530 } 3531 3532 mirror = null_check(mirror); 3533 // If mirror or obj is dead, only null-path is taken. 3534 if (stopped()) return true; 3535 3536 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3537 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3538 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3539 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3540 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3541 3542 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3543 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3544 result_reg, _slow_path); 3545 Node* normal_ctl = control(); 3546 Node* no_array_ctl = result_reg->in(_slow_path); 3547 3548 // Generate code for the slow case. We make a call to newArray(). 3549 set_control(no_array_ctl); 3550 if (!stopped()) { 3551 // Either the input type is void.class, or else the 3552 // array klass has not yet been cached. Either the 3553 // ensuing call will throw an exception, or else it 3554 // will cache the array klass for next time. 3555 PreserveJVMState pjvms(this); 3556 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3557 Node* slow_result = set_results_for_java_call(slow_call); 3558 // this->control() comes from set_results_for_java_call 3559 result_reg->set_req(_slow_path, control()); 3560 result_val->set_req(_slow_path, slow_result); 3561 result_io ->set_req(_slow_path, i_o()); 3562 result_mem->set_req(_slow_path, reset_memory()); 3563 } 3564 3565 set_control(normal_ctl); 3566 if (!stopped()) { 3567 // Normal case: The array type has been cached in the java.lang.Class. 3568 // The following call works fine even if the array type is polymorphic. 3569 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3570 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3571 result_reg->init_req(_normal_path, control()); 3572 result_val->init_req(_normal_path, obj); 3573 result_io ->init_req(_normal_path, i_o()); 3574 result_mem->init_req(_normal_path, reset_memory()); 3575 3576 if (uninitialized) { 3577 // Mark the allocation so that zeroing is skipped 3578 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 3579 alloc->maybe_set_complete(&_gvn); 3580 } 3581 } 3582 3583 // Return the combined state. 3584 set_i_o( _gvn.transform(result_io) ); 3585 set_all_memory( _gvn.transform(result_mem)); 3586 3587 C->set_has_split_ifs(true); // Has chance for split-if optimization 3588 set_result(result_reg, result_val); 3589 return true; 3590 } 3591 3592 //----------------------inline_native_getLength-------------------------- 3593 // public static native int java.lang.reflect.Array.getLength(Object array); 3594 bool LibraryCallKit::inline_native_getLength() { 3595 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3596 3597 Node* array = null_check(argument(0)); 3598 // If array is dead, only null-path is taken. 3599 if (stopped()) return true; 3600 3601 // Deoptimize if it is a non-array. 3602 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3603 3604 if (non_array != NULL) { 3605 PreserveJVMState pjvms(this); 3606 set_control(non_array); 3607 uncommon_trap(Deoptimization::Reason_intrinsic, 3608 Deoptimization::Action_maybe_recompile); 3609 } 3610 3611 // If control is dead, only non-array-path is taken. 3612 if (stopped()) return true; 3613 3614 // The works fine even if the array type is polymorphic. 3615 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3616 Node* result = load_array_length(array); 3617 3618 C->set_has_split_ifs(true); // Has chance for split-if optimization 3619 set_result(result); 3620 return true; 3621 } 3622 3623 //------------------------inline_array_copyOf---------------------------- 3624 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3625 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3626 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3627 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3628 3629 // Get the arguments. 3630 Node* original = argument(0); 3631 Node* start = is_copyOfRange? argument(1): intcon(0); 3632 Node* end = is_copyOfRange? argument(2): argument(1); 3633 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3634 3635 Node* newcopy = NULL; 3636 3637 // Set the original stack and the reexecute bit for the interpreter to reexecute 3638 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3639 { PreserveReexecuteState preexecs(this); 3640 jvms()->set_should_reexecute(true); 3641 3642 array_type_mirror = null_check(array_type_mirror); 3643 original = null_check(original); 3644 3645 // Check if a null path was taken unconditionally. 3646 if (stopped()) return true; 3647 3648 Node* orig_length = load_array_length(original); 3649 3650 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3651 klass_node = null_check(klass_node); 3652 3653 RegionNode* bailout = new RegionNode(1); 3654 record_for_igvn(bailout); 3655 3656 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3657 // Bail out if that is so. 3658 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3659 if (not_objArray != NULL) { 3660 // Improve the klass node's type from the new optimistic assumption: 3661 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3662 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3663 Node* cast = new CastPPNode(klass_node, akls); 3664 cast->init_req(0, control()); 3665 klass_node = _gvn.transform(cast); 3666 } 3667 3668 // Bail out if either start or end is negative. 3669 generate_negative_guard(start, bailout, &start); 3670 generate_negative_guard(end, bailout, &end); 3671 3672 Node* length = end; 3673 if (_gvn.type(start) != TypeInt::ZERO) { 3674 length = _gvn.transform(new SubINode(end, start)); 3675 } 3676 3677 // Bail out if length is negative. 3678 // Without this the new_array would throw 3679 // NegativeArraySizeException but IllegalArgumentException is what 3680 // should be thrown 3681 generate_negative_guard(length, bailout, &length); 3682 3683 if (bailout->req() > 1) { 3684 PreserveJVMState pjvms(this); 3685 set_control(_gvn.transform(bailout)); 3686 uncommon_trap(Deoptimization::Reason_intrinsic, 3687 Deoptimization::Action_maybe_recompile); 3688 } 3689 3690 if (!stopped()) { 3691 // How many elements will we copy from the original? 3692 // The answer is MinI(orig_length - start, length). 3693 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3694 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3695 3696 // Generate a direct call to the right arraycopy function(s). 3697 // We know the copy is disjoint but we might not know if the 3698 // oop stores need checking. 3699 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3700 // This will fail a store-check if x contains any non-nulls. 3701 3702 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3703 // loads/stores but it is legal only if we're sure the 3704 // Arrays.copyOf would succeed. So we need all input arguments 3705 // to the copyOf to be validated, including that the copy to the 3706 // new array won't trigger an ArrayStoreException. That subtype 3707 // check can be optimized if we know something on the type of 3708 // the input array from type speculation. 3709 if (_gvn.type(klass_node)->singleton()) { 3710 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3711 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3712 3713 int test = C->static_subtype_check(superk, subk); 3714 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3715 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3716 if (t_original->speculative_type() != NULL) { 3717 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3718 } 3719 } 3720 } 3721 3722 bool validated = false; 3723 // Reason_class_check rather than Reason_intrinsic because we 3724 // want to intrinsify even if this traps. 3725 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3726 Node* not_subtype_ctrl = gen_subtype_check(original, klass_node); 3727 3728 if (not_subtype_ctrl != top()) { 3729 PreserveJVMState pjvms(this); 3730 set_control(not_subtype_ctrl); 3731 uncommon_trap(Deoptimization::Reason_class_check, 3732 Deoptimization::Action_make_not_entrant); 3733 assert(stopped(), "Should be stopped"); 3734 } 3735 validated = true; 3736 } 3737 3738 if (!stopped()) { 3739 newcopy = new_array(klass_node, length, 0); // no arguments to push 3740 3741 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false, 3742 load_object_klass(original), klass_node); 3743 if (!is_copyOfRange) { 3744 ac->set_copyof(validated); 3745 } else { 3746 ac->set_copyofrange(validated); 3747 } 3748 Node* n = _gvn.transform(ac); 3749 if (n == ac) { 3750 ac->connect_outputs(this); 3751 } else { 3752 assert(validated, "shouldn't transform if all arguments not validated"); 3753 set_all_memory(n); 3754 } 3755 } 3756 } 3757 } // original reexecute is set back here 3758 3759 C->set_has_split_ifs(true); // Has chance for split-if optimization 3760 if (!stopped()) { 3761 set_result(newcopy); 3762 } 3763 return true; 3764 } 3765 3766 3767 //----------------------generate_virtual_guard--------------------------- 3768 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3769 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3770 RegionNode* slow_region) { 3771 ciMethod* method = callee(); 3772 int vtable_index = method->vtable_index(); 3773 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3774 "bad index %d", vtable_index); 3775 // Get the Method* out of the appropriate vtable entry. 3776 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 3777 vtable_index*vtableEntry::size_in_bytes() + 3778 vtableEntry::method_offset_in_bytes(); 3779 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3780 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3781 3782 // Compare the target method with the expected method (e.g., Object.hashCode). 3783 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3784 3785 Node* native_call = makecon(native_call_addr); 3786 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3787 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3788 3789 return generate_slow_guard(test_native, slow_region); 3790 } 3791 3792 //-----------------------generate_method_call---------------------------- 3793 // Use generate_method_call to make a slow-call to the real 3794 // method if the fast path fails. An alternative would be to 3795 // use a stub like OptoRuntime::slow_arraycopy_Java. 3796 // This only works for expanding the current library call, 3797 // not another intrinsic. (E.g., don't use this for making an 3798 // arraycopy call inside of the copyOf intrinsic.) 3799 CallJavaNode* 3800 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3801 // When compiling the intrinsic method itself, do not use this technique. 3802 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3803 3804 ciMethod* method = callee(); 3805 // ensure the JVMS we have will be correct for this call 3806 guarantee(method_id == method->intrinsic_id(), "must match"); 3807 3808 const TypeFunc* tf = TypeFunc::make(method); 3809 CallJavaNode* slow_call; 3810 if (is_static) { 3811 assert(!is_virtual, ""); 3812 slow_call = new CallStaticJavaNode(C, tf, 3813 SharedRuntime::get_resolve_static_call_stub(), 3814 method, bci()); 3815 } else if (is_virtual) { 3816 null_check_receiver(); 3817 int vtable_index = Method::invalid_vtable_index; 3818 if (UseInlineCaches) { 3819 // Suppress the vtable call 3820 } else { 3821 // hashCode and clone are not a miranda methods, 3822 // so the vtable index is fixed. 3823 // No need to use the linkResolver to get it. 3824 vtable_index = method->vtable_index(); 3825 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3826 "bad index %d", vtable_index); 3827 } 3828 slow_call = new CallDynamicJavaNode(tf, 3829 SharedRuntime::get_resolve_virtual_call_stub(), 3830 method, vtable_index, bci()); 3831 } else { // neither virtual nor static: opt_virtual 3832 null_check_receiver(); 3833 slow_call = new CallStaticJavaNode(C, tf, 3834 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3835 method, bci()); 3836 slow_call->set_optimized_virtual(true); 3837 } 3838 if (CallGenerator::is_inlined_method_handle_intrinsic(this->method(), bci(), callee())) { 3839 // To be able to issue a direct call (optimized virtual or virtual) 3840 // and skip a call to MH.linkTo*/invokeBasic adapter, additional information 3841 // about the method being invoked should be attached to the call site to 3842 // make resolution logic work (see SharedRuntime::resolve_{virtual,opt_virtual}_call_C). 3843 slow_call->set_override_symbolic_info(true); 3844 } 3845 set_arguments_for_java_call(slow_call); 3846 set_edges_for_java_call(slow_call); 3847 return slow_call; 3848 } 3849 3850 3851 /** 3852 * Build special case code for calls to hashCode on an object. This call may 3853 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 3854 * slightly different code. 3855 */ 3856 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3857 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3858 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3859 3860 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3861 3862 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3863 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 3864 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3865 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3866 Node* obj = NULL; 3867 if (!is_static) { 3868 // Check for hashing null object 3869 obj = null_check_receiver(); 3870 if (stopped()) return true; // unconditionally null 3871 result_reg->init_req(_null_path, top()); 3872 result_val->init_req(_null_path, top()); 3873 } else { 3874 // Do a null check, and return zero if null. 3875 // System.identityHashCode(null) == 0 3876 obj = argument(0); 3877 Node* null_ctl = top(); 3878 obj = null_check_oop(obj, &null_ctl); 3879 result_reg->init_req(_null_path, null_ctl); 3880 result_val->init_req(_null_path, _gvn.intcon(0)); 3881 } 3882 3883 // Unconditionally null? Then return right away. 3884 if (stopped()) { 3885 set_control( result_reg->in(_null_path)); 3886 if (!stopped()) 3887 set_result(result_val->in(_null_path)); 3888 return true; 3889 } 3890 3891 // We only go to the fast case code if we pass a number of guards. The 3892 // paths which do not pass are accumulated in the slow_region. 3893 RegionNode* slow_region = new RegionNode(1); 3894 record_for_igvn(slow_region); 3895 3896 // If this is a virtual call, we generate a funny guard. We pull out 3897 // the vtable entry corresponding to hashCode() from the target object. 3898 // If the target method which we are calling happens to be the native 3899 // Object hashCode() method, we pass the guard. We do not need this 3900 // guard for non-virtual calls -- the caller is known to be the native 3901 // Object hashCode(). 3902 if (is_virtual) { 3903 // After null check, get the object's klass. 3904 Node* obj_klass = load_object_klass(obj); 3905 generate_virtual_guard(obj_klass, slow_region); 3906 } 3907 3908 // Get the header out of the object, use LoadMarkNode when available 3909 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3910 // The control of the load must be NULL. Otherwise, the load can move before 3911 // the null check after castPP removal. 3912 Node* no_ctrl = NULL; 3913 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3914 3915 // Test the header to see if it is unlocked. 3916 Node *lock_mask = _gvn.MakeConX(markWord::biased_lock_mask_in_place); 3917 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 3918 Node *unlocked_val = _gvn.MakeConX(markWord::unlocked_value); 3919 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 3920 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 3921 3922 generate_slow_guard(test_unlocked, slow_region); 3923 3924 // Get the hash value and check to see that it has been properly assigned. 3925 // We depend on hash_mask being at most 32 bits and avoid the use of 3926 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3927 // vm: see markWord.hpp. 3928 Node *hash_mask = _gvn.intcon(markWord::hash_mask); 3929 Node *hash_shift = _gvn.intcon(markWord::hash_shift); 3930 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 3931 // This hack lets the hash bits live anywhere in the mark object now, as long 3932 // as the shift drops the relevant bits into the low 32 bits. Note that 3933 // Java spec says that HashCode is an int so there's no point in capturing 3934 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3935 hshifted_header = ConvX2I(hshifted_header); 3936 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 3937 3938 Node *no_hash_val = _gvn.intcon(markWord::no_hash); 3939 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 3940 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 3941 3942 generate_slow_guard(test_assigned, slow_region); 3943 3944 Node* init_mem = reset_memory(); 3945 // fill in the rest of the null path: 3946 result_io ->init_req(_null_path, i_o()); 3947 result_mem->init_req(_null_path, init_mem); 3948 3949 result_val->init_req(_fast_path, hash_val); 3950 result_reg->init_req(_fast_path, control()); 3951 result_io ->init_req(_fast_path, i_o()); 3952 result_mem->init_req(_fast_path, init_mem); 3953 3954 // Generate code for the slow case. We make a call to hashCode(). 3955 set_control(_gvn.transform(slow_region)); 3956 if (!stopped()) { 3957 // No need for PreserveJVMState, because we're using up the present state. 3958 set_all_memory(init_mem); 3959 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 3960 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3961 Node* slow_result = set_results_for_java_call(slow_call); 3962 // this->control() comes from set_results_for_java_call 3963 result_reg->init_req(_slow_path, control()); 3964 result_val->init_req(_slow_path, slow_result); 3965 result_io ->set_req(_slow_path, i_o()); 3966 result_mem ->set_req(_slow_path, reset_memory()); 3967 } 3968 3969 // Return the combined state. 3970 set_i_o( _gvn.transform(result_io) ); 3971 set_all_memory( _gvn.transform(result_mem)); 3972 3973 set_result(result_reg, result_val); 3974 return true; 3975 } 3976 3977 //---------------------------inline_native_getClass---------------------------- 3978 // public final native Class<?> java.lang.Object.getClass(); 3979 // 3980 // Build special case code for calls to getClass on an object. 3981 bool LibraryCallKit::inline_native_getClass() { 3982 Node* obj = null_check_receiver(); 3983 if (stopped()) return true; 3984 set_result(load_mirror_from_klass(load_object_klass(obj))); 3985 return true; 3986 } 3987 3988 //-----------------inline_native_Reflection_getCallerClass--------------------- 3989 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 3990 // 3991 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3992 // 3993 // NOTE: This code must perform the same logic as JVM_GetCallerClass 3994 // in that it must skip particular security frames and checks for 3995 // caller sensitive methods. 3996 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3997 #ifndef PRODUCT 3998 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 3999 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4000 } 4001 #endif 4002 4003 if (!jvms()->has_method()) { 4004 #ifndef PRODUCT 4005 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4006 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4007 } 4008 #endif 4009 return false; 4010 } 4011 4012 // Walk back up the JVM state to find the caller at the required 4013 // depth. 4014 JVMState* caller_jvms = jvms(); 4015 4016 // Cf. JVM_GetCallerClass 4017 // NOTE: Start the loop at depth 1 because the current JVM state does 4018 // not include the Reflection.getCallerClass() frame. 4019 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4020 ciMethod* m = caller_jvms->method(); 4021 switch (n) { 4022 case 0: 4023 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4024 break; 4025 case 1: 4026 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4027 if (!m->caller_sensitive()) { 4028 #ifndef PRODUCT 4029 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4030 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4031 } 4032 #endif 4033 return false; // bail-out; let JVM_GetCallerClass do the work 4034 } 4035 break; 4036 default: 4037 if (!m->is_ignored_by_security_stack_walk()) { 4038 // We have reached the desired frame; return the holder class. 4039 // Acquire method holder as java.lang.Class and push as constant. 4040 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4041 ciInstance* caller_mirror = caller_klass->java_mirror(); 4042 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4043 4044 #ifndef PRODUCT 4045 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4046 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()); 4047 tty->print_cr(" JVM state at this point:"); 4048 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4049 ciMethod* m = jvms()->of_depth(i)->method(); 4050 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4051 } 4052 } 4053 #endif 4054 return true; 4055 } 4056 break; 4057 } 4058 } 4059 4060 #ifndef PRODUCT 4061 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4062 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4063 tty->print_cr(" JVM state at this point:"); 4064 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4065 ciMethod* m = jvms()->of_depth(i)->method(); 4066 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4067 } 4068 } 4069 #endif 4070 4071 return false; // bail-out; let JVM_GetCallerClass do the work 4072 } 4073 4074 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4075 Node* arg = argument(0); 4076 Node* result = NULL; 4077 4078 switch (id) { 4079 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4080 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4081 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4082 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4083 4084 case vmIntrinsics::_doubleToLongBits: { 4085 // two paths (plus control) merge in a wood 4086 RegionNode *r = new RegionNode(3); 4087 Node *phi = new PhiNode(r, TypeLong::LONG); 4088 4089 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4090 // Build the boolean node 4091 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4092 4093 // Branch either way. 4094 // NaN case is less traveled, which makes all the difference. 4095 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4096 Node *opt_isnan = _gvn.transform(ifisnan); 4097 assert( opt_isnan->is_If(), "Expect an IfNode"); 4098 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4099 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4100 4101 set_control(iftrue); 4102 4103 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4104 Node *slow_result = longcon(nan_bits); // return NaN 4105 phi->init_req(1, _gvn.transform( slow_result )); 4106 r->init_req(1, iftrue); 4107 4108 // Else fall through 4109 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4110 set_control(iffalse); 4111 4112 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4113 r->init_req(2, iffalse); 4114 4115 // Post merge 4116 set_control(_gvn.transform(r)); 4117 record_for_igvn(r); 4118 4119 C->set_has_split_ifs(true); // Has chance for split-if optimization 4120 result = phi; 4121 assert(result->bottom_type()->isa_long(), "must be"); 4122 break; 4123 } 4124 4125 case vmIntrinsics::_floatToIntBits: { 4126 // two paths (plus control) merge in a wood 4127 RegionNode *r = new RegionNode(3); 4128 Node *phi = new PhiNode(r, TypeInt::INT); 4129 4130 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4131 // Build the boolean node 4132 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4133 4134 // Branch either way. 4135 // NaN case is less traveled, which makes all the difference. 4136 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4137 Node *opt_isnan = _gvn.transform(ifisnan); 4138 assert( opt_isnan->is_If(), "Expect an IfNode"); 4139 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4140 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4141 4142 set_control(iftrue); 4143 4144 static const jint nan_bits = 0x7fc00000; 4145 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4146 phi->init_req(1, _gvn.transform( slow_result )); 4147 r->init_req(1, iftrue); 4148 4149 // Else fall through 4150 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4151 set_control(iffalse); 4152 4153 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4154 r->init_req(2, iffalse); 4155 4156 // Post merge 4157 set_control(_gvn.transform(r)); 4158 record_for_igvn(r); 4159 4160 C->set_has_split_ifs(true); // Has chance for split-if optimization 4161 result = phi; 4162 assert(result->bottom_type()->isa_int(), "must be"); 4163 break; 4164 } 4165 4166 default: 4167 fatal_unexpected_iid(id); 4168 break; 4169 } 4170 set_result(_gvn.transform(result)); 4171 return true; 4172 } 4173 4174 //----------------------inline_unsafe_copyMemory------------------------- 4175 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4176 bool LibraryCallKit::inline_unsafe_copyMemory() { 4177 if (callee()->is_static()) return false; // caller must have the capability! 4178 null_check_receiver(); // null-check receiver 4179 if (stopped()) return true; 4180 4181 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4182 4183 Node* src_ptr = argument(1); // type: oop 4184 Node* src_off = ConvL2X(argument(2)); // type: long 4185 Node* dst_ptr = argument(4); // type: oop 4186 Node* dst_off = ConvL2X(argument(5)); // type: long 4187 Node* size = ConvL2X(argument(7)); // type: long 4188 4189 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4190 "fieldOffset must be byte-scaled"); 4191 4192 Node* src = make_unsafe_address(src_ptr, src_off, ACCESS_READ); 4193 Node* dst = make_unsafe_address(dst_ptr, dst_off, ACCESS_WRITE); 4194 4195 // Conservatively insert a memory barrier on all memory slices. 4196 // Do not let writes of the copy source or destination float below the copy. 4197 insert_mem_bar(Op_MemBarCPUOrder); 4198 4199 Node* thread = _gvn.transform(new ThreadLocalNode()); 4200 Node* doing_unsafe_access_addr = basic_plus_adr(top(), thread, in_bytes(JavaThread::doing_unsafe_access_offset())); 4201 BasicType doing_unsafe_access_bt = T_BYTE; 4202 assert((sizeof(bool) * CHAR_BIT) == 8, "not implemented"); 4203 4204 // update volatile field 4205 store_to_memory(control(), doing_unsafe_access_addr, intcon(1), doing_unsafe_access_bt, Compile::AliasIdxRaw, MemNode::unordered); 4206 4207 // Call it. Note that the length argument is not scaled. 4208 make_runtime_call(RC_LEAF|RC_NO_FP, 4209 OptoRuntime::fast_arraycopy_Type(), 4210 StubRoutines::unsafe_arraycopy(), 4211 "unsafe_arraycopy", 4212 TypeRawPtr::BOTTOM, 4213 src, dst, size XTOP); 4214 4215 store_to_memory(control(), doing_unsafe_access_addr, intcon(0), doing_unsafe_access_bt, Compile::AliasIdxRaw, MemNode::unordered); 4216 4217 // Do not let reads of the copy destination float above the copy. 4218 insert_mem_bar(Op_MemBarCPUOrder); 4219 4220 return true; 4221 } 4222 4223 //------------------------clone_coping----------------------------------- 4224 // Helper function for inline_native_clone. 4225 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) { 4226 assert(obj_size != NULL, ""); 4227 Node* raw_obj = alloc_obj->in(1); 4228 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4229 4230 AllocateNode* alloc = NULL; 4231 if (ReduceBulkZeroing) { 4232 // We will be completely responsible for initializing this object - 4233 // mark Initialize node as complete. 4234 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4235 // The object was just allocated - there should be no any stores! 4236 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4237 // Mark as complete_with_arraycopy so that on AllocateNode 4238 // expansion, we know this AllocateNode is initialized by an array 4239 // copy and a StoreStore barrier exists after the array copy. 4240 alloc->initialization()->set_complete_with_arraycopy(); 4241 } 4242 4243 Node* size = _gvn.transform(obj_size); 4244 access_clone(obj, alloc_obj, size, is_array); 4245 4246 // Do not let reads from the cloned object float above the arraycopy. 4247 if (alloc != NULL) { 4248 // Do not let stores that initialize this object be reordered with 4249 // a subsequent store that would make this object accessible by 4250 // other threads. 4251 // Record what AllocateNode this StoreStore protects so that 4252 // escape analysis can go from the MemBarStoreStoreNode to the 4253 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4254 // based on the escape status of the AllocateNode. 4255 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 4256 } else { 4257 insert_mem_bar(Op_MemBarCPUOrder); 4258 } 4259 } 4260 4261 //------------------------inline_native_clone---------------------------- 4262 // protected native Object java.lang.Object.clone(); 4263 // 4264 // Here are the simple edge cases: 4265 // null receiver => normal trap 4266 // virtual and clone was overridden => slow path to out-of-line clone 4267 // not cloneable or finalizer => slow path to out-of-line Object.clone 4268 // 4269 // The general case has two steps, allocation and copying. 4270 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4271 // 4272 // Copying also has two cases, oop arrays and everything else. 4273 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4274 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4275 // 4276 // These steps fold up nicely if and when the cloned object's klass 4277 // can be sharply typed as an object array, a type array, or an instance. 4278 // 4279 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4280 PhiNode* result_val; 4281 4282 // Set the reexecute bit for the interpreter to reexecute 4283 // the bytecode that invokes Object.clone if deoptimization happens. 4284 { PreserveReexecuteState preexecs(this); 4285 jvms()->set_should_reexecute(true); 4286 4287 Node* obj = null_check_receiver(); 4288 if (stopped()) return true; 4289 4290 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4291 4292 // If we are going to clone an instance, we need its exact type to 4293 // know the number and types of fields to convert the clone to 4294 // loads/stores. Maybe a speculative type can help us. 4295 if (!obj_type->klass_is_exact() && 4296 obj_type->speculative_type() != NULL && 4297 obj_type->speculative_type()->is_instance_klass()) { 4298 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4299 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4300 !spec_ik->has_injected_fields()) { 4301 ciKlass* k = obj_type->klass(); 4302 if (!k->is_instance_klass() || 4303 k->as_instance_klass()->is_interface() || 4304 k->as_instance_klass()->has_subklass()) { 4305 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4306 } 4307 } 4308 } 4309 4310 // Conservatively insert a memory barrier on all memory slices. 4311 // Do not let writes into the original float below the clone. 4312 insert_mem_bar(Op_MemBarCPUOrder); 4313 4314 // paths into result_reg: 4315 enum { 4316 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4317 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4318 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4319 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4320 PATH_LIMIT 4321 }; 4322 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4323 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4324 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4325 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4326 record_for_igvn(result_reg); 4327 4328 Node* obj_klass = load_object_klass(obj); 4329 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4330 if (array_ctl != NULL) { 4331 // It's an array. 4332 PreserveJVMState pjvms(this); 4333 set_control(array_ctl); 4334 Node* obj_length = load_array_length(obj); 4335 Node* obj_size = NULL; 4336 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size, /*deoptimize_on_exception=*/true); 4337 4338 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4339 if (bs->array_copy_requires_gc_barriers(true, T_OBJECT, true, BarrierSetC2::Parsing)) { 4340 // If it is an oop array, it requires very special treatment, 4341 // because gc barriers are required when accessing the array. 4342 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4343 if (is_obja != NULL) { 4344 PreserveJVMState pjvms2(this); 4345 set_control(is_obja); 4346 // Generate a direct call to the right arraycopy function(s). 4347 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4348 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false); 4349 ac->set_clone_oop_array(); 4350 Node* n = _gvn.transform(ac); 4351 assert(n == ac, "cannot disappear"); 4352 ac->connect_outputs(this, /*deoptimize_on_exception=*/true); 4353 4354 result_reg->init_req(_objArray_path, control()); 4355 result_val->init_req(_objArray_path, alloc_obj); 4356 result_i_o ->set_req(_objArray_path, i_o()); 4357 result_mem ->set_req(_objArray_path, reset_memory()); 4358 } 4359 } 4360 // Otherwise, there are no barriers to worry about. 4361 // (We can dispense with card marks if we know the allocation 4362 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4363 // causes the non-eden paths to take compensating steps to 4364 // simulate a fresh allocation, so that no further 4365 // card marks are required in compiled code to initialize 4366 // the object.) 4367 4368 if (!stopped()) { 4369 copy_to_clone(obj, alloc_obj, obj_size, true); 4370 4371 // Present the results of the copy. 4372 result_reg->init_req(_array_path, control()); 4373 result_val->init_req(_array_path, alloc_obj); 4374 result_i_o ->set_req(_array_path, i_o()); 4375 result_mem ->set_req(_array_path, reset_memory()); 4376 } 4377 } 4378 4379 // We only go to the instance fast case code if we pass a number of guards. 4380 // The paths which do not pass are accumulated in the slow_region. 4381 RegionNode* slow_region = new RegionNode(1); 4382 record_for_igvn(slow_region); 4383 if (!stopped()) { 4384 // It's an instance (we did array above). Make the slow-path tests. 4385 // If this is a virtual call, we generate a funny guard. We grab 4386 // the vtable entry corresponding to clone() from the target object. 4387 // If the target method which we are calling happens to be the 4388 // Object clone() method, we pass the guard. We do not need this 4389 // guard for non-virtual calls; the caller is known to be the native 4390 // Object clone(). 4391 if (is_virtual) { 4392 generate_virtual_guard(obj_klass, slow_region); 4393 } 4394 4395 // The object must be easily cloneable and must not have a finalizer. 4396 // Both of these conditions may be checked in a single test. 4397 // We could optimize the test further, but we don't care. 4398 generate_access_flags_guard(obj_klass, 4399 // Test both conditions: 4400 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4401 // Must be cloneable but not finalizer: 4402 JVM_ACC_IS_CLONEABLE_FAST, 4403 slow_region); 4404 } 4405 4406 if (!stopped()) { 4407 // It's an instance, and it passed the slow-path tests. 4408 PreserveJVMState pjvms(this); 4409 Node* obj_size = NULL; 4410 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4411 // is reexecuted if deoptimization occurs and there could be problems when merging 4412 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4413 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4414 4415 copy_to_clone(obj, alloc_obj, obj_size, false); 4416 4417 // Present the results of the slow call. 4418 result_reg->init_req(_instance_path, control()); 4419 result_val->init_req(_instance_path, alloc_obj); 4420 result_i_o ->set_req(_instance_path, i_o()); 4421 result_mem ->set_req(_instance_path, reset_memory()); 4422 } 4423 4424 // Generate code for the slow case. We make a call to clone(). 4425 set_control(_gvn.transform(slow_region)); 4426 if (!stopped()) { 4427 PreserveJVMState pjvms(this); 4428 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4429 // We need to deoptimize on exception (see comment above) 4430 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true); 4431 // this->control() comes from set_results_for_java_call 4432 result_reg->init_req(_slow_path, control()); 4433 result_val->init_req(_slow_path, slow_result); 4434 result_i_o ->set_req(_slow_path, i_o()); 4435 result_mem ->set_req(_slow_path, reset_memory()); 4436 } 4437 4438 // Return the combined state. 4439 set_control( _gvn.transform(result_reg)); 4440 set_i_o( _gvn.transform(result_i_o)); 4441 set_all_memory( _gvn.transform(result_mem)); 4442 } // original reexecute is set back here 4443 4444 set_result(_gvn.transform(result_val)); 4445 return true; 4446 } 4447 4448 // If we have a tightly coupled allocation, the arraycopy may take care 4449 // of the array initialization. If one of the guards we insert between 4450 // the allocation and the arraycopy causes a deoptimization, an 4451 // unitialized array will escape the compiled method. To prevent that 4452 // we set the JVM state for uncommon traps between the allocation and 4453 // the arraycopy to the state before the allocation so, in case of 4454 // deoptimization, we'll reexecute the allocation and the 4455 // initialization. 4456 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4457 if (alloc != NULL) { 4458 ciMethod* trap_method = alloc->jvms()->method(); 4459 int trap_bci = alloc->jvms()->bci(); 4460 4461 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 4462 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4463 // Make sure there's no store between the allocation and the 4464 // arraycopy otherwise visible side effects could be rexecuted 4465 // in case of deoptimization and cause incorrect execution. 4466 bool no_interfering_store = true; 4467 Node* mem = alloc->in(TypeFunc::Memory); 4468 if (mem->is_MergeMem()) { 4469 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4470 Node* n = mms.memory(); 4471 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4472 assert(n->is_Store(), "what else?"); 4473 no_interfering_store = false; 4474 break; 4475 } 4476 } 4477 } else { 4478 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4479 Node* n = mms.memory(); 4480 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4481 assert(n->is_Store(), "what else?"); 4482 no_interfering_store = false; 4483 break; 4484 } 4485 } 4486 } 4487 4488 if (no_interfering_store) { 4489 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4490 uint size = alloc->req(); 4491 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4492 old_jvms->set_map(sfpt); 4493 for (uint i = 0; i < size; i++) { 4494 sfpt->init_req(i, alloc->in(i)); 4495 } 4496 // re-push array length for deoptimization 4497 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4498 old_jvms->set_sp(old_jvms->sp()+1); 4499 old_jvms->set_monoff(old_jvms->monoff()+1); 4500 old_jvms->set_scloff(old_jvms->scloff()+1); 4501 old_jvms->set_endoff(old_jvms->endoff()+1); 4502 old_jvms->set_should_reexecute(true); 4503 4504 sfpt->set_i_o(map()->i_o()); 4505 sfpt->set_memory(map()->memory()); 4506 sfpt->set_control(map()->control()); 4507 4508 JVMState* saved_jvms = jvms(); 4509 saved_reexecute_sp = _reexecute_sp; 4510 4511 set_jvms(sfpt->jvms()); 4512 _reexecute_sp = jvms()->sp(); 4513 4514 return saved_jvms; 4515 } 4516 } 4517 } 4518 return NULL; 4519 } 4520 4521 // In case of a deoptimization, we restart execution at the 4522 // allocation, allocating a new array. We would leave an uninitialized 4523 // array in the heap that GCs wouldn't expect. Move the allocation 4524 // after the traps so we don't allocate the array if we 4525 // deoptimize. This is possible because tightly_coupled_allocation() 4526 // guarantees there's no observer of the allocated array at this point 4527 // and the control flow is simple enough. 4528 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, 4529 int saved_reexecute_sp, uint new_idx) { 4530 if (saved_jvms != NULL && !stopped()) { 4531 assert(alloc != NULL, "only with a tightly coupled allocation"); 4532 // restore JVM state to the state at the arraycopy 4533 saved_jvms->map()->set_control(map()->control()); 4534 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4535 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4536 // If we've improved the types of some nodes (null check) while 4537 // emitting the guards, propagate them to the current state 4538 map()->replaced_nodes().apply(saved_jvms->map(), new_idx); 4539 set_jvms(saved_jvms); 4540 _reexecute_sp = saved_reexecute_sp; 4541 4542 // Remove the allocation from above the guards 4543 CallProjections callprojs; 4544 alloc->extract_projections(&callprojs, true); 4545 InitializeNode* init = alloc->initialization(); 4546 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4547 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4548 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4549 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4550 4551 // move the allocation here (after the guards) 4552 _gvn.hash_delete(alloc); 4553 alloc->set_req(TypeFunc::Control, control()); 4554 alloc->set_req(TypeFunc::I_O, i_o()); 4555 Node *mem = reset_memory(); 4556 set_all_memory(mem); 4557 alloc->set_req(TypeFunc::Memory, mem); 4558 set_control(init->proj_out_or_null(TypeFunc::Control)); 4559 set_i_o(callprojs.fallthrough_ioproj); 4560 4561 // Update memory as done in GraphKit::set_output_for_allocation() 4562 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4563 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4564 if (ary_type->isa_aryptr() && length_type != NULL) { 4565 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4566 } 4567 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4568 int elemidx = C->get_alias_index(telemref); 4569 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw); 4570 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx); 4571 4572 Node* allocx = _gvn.transform(alloc); 4573 assert(allocx == alloc, "where has the allocation gone?"); 4574 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4575 4576 _gvn.hash_delete(dest); 4577 dest->set_req(0, control()); 4578 Node* destx = _gvn.transform(dest); 4579 assert(destx == dest, "where has the allocation result gone?"); 4580 } 4581 } 4582 4583 4584 //------------------------------inline_arraycopy----------------------- 4585 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4586 // Object dest, int destPos, 4587 // int length); 4588 bool LibraryCallKit::inline_arraycopy() { 4589 // Get the arguments. 4590 Node* src = argument(0); // type: oop 4591 Node* src_offset = argument(1); // type: int 4592 Node* dest = argument(2); // type: oop 4593 Node* dest_offset = argument(3); // type: int 4594 Node* length = argument(4); // type: int 4595 4596 uint new_idx = C->unique(); 4597 4598 // Check for allocation before we add nodes that would confuse 4599 // tightly_coupled_allocation() 4600 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4601 4602 int saved_reexecute_sp = -1; 4603 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4604 // See arraycopy_restore_alloc_state() comment 4605 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4606 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4607 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards 4608 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4609 4610 // The following tests must be performed 4611 // (1) src and dest are arrays. 4612 // (2) src and dest arrays must have elements of the same BasicType 4613 // (3) src and dest must not be null. 4614 // (4) src_offset must not be negative. 4615 // (5) dest_offset must not be negative. 4616 // (6) length must not be negative. 4617 // (7) src_offset + length must not exceed length of src. 4618 // (8) dest_offset + length must not exceed length of dest. 4619 // (9) each element of an oop array must be assignable 4620 4621 // (3) src and dest must not be null. 4622 // always do this here because we need the JVM state for uncommon traps 4623 Node* null_ctl = top(); 4624 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4625 assert(null_ctl->is_top(), "no null control here"); 4626 dest = null_check(dest, T_ARRAY); 4627 4628 if (!can_emit_guards) { 4629 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4630 // guards but the arraycopy node could still take advantage of a 4631 // tightly allocated allocation. tightly_coupled_allocation() is 4632 // called again to make sure it takes the null check above into 4633 // account: the null check is mandatory and if it caused an 4634 // uncommon trap to be emitted then the allocation can't be 4635 // considered tightly coupled in this context. 4636 alloc = tightly_coupled_allocation(dest, NULL); 4637 } 4638 4639 bool validated = false; 4640 4641 const Type* src_type = _gvn.type(src); 4642 const Type* dest_type = _gvn.type(dest); 4643 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4644 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4645 4646 // Do we have the type of src? 4647 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4648 // Do we have the type of dest? 4649 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4650 // Is the type for src from speculation? 4651 bool src_spec = false; 4652 // Is the type for dest from speculation? 4653 bool dest_spec = false; 4654 4655 if ((!has_src || !has_dest) && can_emit_guards) { 4656 // We don't have sufficient type information, let's see if 4657 // speculative types can help. We need to have types for both src 4658 // and dest so that it pays off. 4659 4660 // Do we already have or could we have type information for src 4661 bool could_have_src = has_src; 4662 // Do we already have or could we have type information for dest 4663 bool could_have_dest = has_dest; 4664 4665 ciKlass* src_k = NULL; 4666 if (!has_src) { 4667 src_k = src_type->speculative_type_not_null(); 4668 if (src_k != NULL && src_k->is_array_klass()) { 4669 could_have_src = true; 4670 } 4671 } 4672 4673 ciKlass* dest_k = NULL; 4674 if (!has_dest) { 4675 dest_k = dest_type->speculative_type_not_null(); 4676 if (dest_k != NULL && dest_k->is_array_klass()) { 4677 could_have_dest = true; 4678 } 4679 } 4680 4681 if (could_have_src && could_have_dest) { 4682 // This is going to pay off so emit the required guards 4683 if (!has_src) { 4684 src = maybe_cast_profiled_obj(src, src_k, true); 4685 src_type = _gvn.type(src); 4686 top_src = src_type->isa_aryptr(); 4687 has_src = (top_src != NULL && top_src->klass() != NULL); 4688 src_spec = true; 4689 } 4690 if (!has_dest) { 4691 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4692 dest_type = _gvn.type(dest); 4693 top_dest = dest_type->isa_aryptr(); 4694 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4695 dest_spec = true; 4696 } 4697 } 4698 } 4699 4700 if (has_src && has_dest && can_emit_guards) { 4701 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4702 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4703 if (is_reference_type(src_elem)) src_elem = T_OBJECT; 4704 if (is_reference_type(dest_elem)) dest_elem = T_OBJECT; 4705 4706 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4707 // If both arrays are object arrays then having the exact types 4708 // for both will remove the need for a subtype check at runtime 4709 // before the call and may make it possible to pick a faster copy 4710 // routine (without a subtype check on every element) 4711 // Do we have the exact type of src? 4712 bool could_have_src = src_spec; 4713 // Do we have the exact type of dest? 4714 bool could_have_dest = dest_spec; 4715 ciKlass* src_k = top_src->klass(); 4716 ciKlass* dest_k = top_dest->klass(); 4717 if (!src_spec) { 4718 src_k = src_type->speculative_type_not_null(); 4719 if (src_k != NULL && src_k->is_array_klass()) { 4720 could_have_src = true; 4721 } 4722 } 4723 if (!dest_spec) { 4724 dest_k = dest_type->speculative_type_not_null(); 4725 if (dest_k != NULL && dest_k->is_array_klass()) { 4726 could_have_dest = true; 4727 } 4728 } 4729 if (could_have_src && could_have_dest) { 4730 // If we can have both exact types, emit the missing guards 4731 if (could_have_src && !src_spec) { 4732 src = maybe_cast_profiled_obj(src, src_k, true); 4733 } 4734 if (could_have_dest && !dest_spec) { 4735 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4736 } 4737 } 4738 } 4739 } 4740 4741 ciMethod* trap_method = method(); 4742 int trap_bci = bci(); 4743 if (saved_jvms != NULL) { 4744 trap_method = alloc->jvms()->method(); 4745 trap_bci = alloc->jvms()->bci(); 4746 } 4747 4748 bool negative_length_guard_generated = false; 4749 4750 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 4751 can_emit_guards && 4752 !src->is_top() && !dest->is_top()) { 4753 // validate arguments: enables transformation the ArrayCopyNode 4754 validated = true; 4755 4756 RegionNode* slow_region = new RegionNode(1); 4757 record_for_igvn(slow_region); 4758 4759 // (1) src and dest are arrays. 4760 generate_non_array_guard(load_object_klass(src), slow_region); 4761 generate_non_array_guard(load_object_klass(dest), slow_region); 4762 4763 // (2) src and dest arrays must have elements of the same BasicType 4764 // done at macro expansion or at Ideal transformation time 4765 4766 // (4) src_offset must not be negative. 4767 generate_negative_guard(src_offset, slow_region); 4768 4769 // (5) dest_offset must not be negative. 4770 generate_negative_guard(dest_offset, slow_region); 4771 4772 // (7) src_offset + length must not exceed length of src. 4773 generate_limit_guard(src_offset, length, 4774 load_array_length(src), 4775 slow_region); 4776 4777 // (8) dest_offset + length must not exceed length of dest. 4778 generate_limit_guard(dest_offset, length, 4779 load_array_length(dest), 4780 slow_region); 4781 4782 // (6) length must not be negative. 4783 // This is also checked in generate_arraycopy() during macro expansion, but 4784 // we also have to check it here for the case where the ArrayCopyNode will 4785 // be eliminated by Escape Analysis. 4786 if (EliminateAllocations) { 4787 generate_negative_guard(length, slow_region); 4788 negative_length_guard_generated = true; 4789 } 4790 4791 // (9) each element of an oop array must be assignable 4792 Node* dest_klass = load_object_klass(dest); 4793 if (src != dest) { 4794 Node* not_subtype_ctrl = gen_subtype_check(src, dest_klass); 4795 4796 if (not_subtype_ctrl != top()) { 4797 PreserveJVMState pjvms(this); 4798 set_control(not_subtype_ctrl); 4799 uncommon_trap(Deoptimization::Reason_intrinsic, 4800 Deoptimization::Action_make_not_entrant); 4801 assert(stopped(), "Should be stopped"); 4802 } 4803 } 4804 { 4805 PreserveJVMState pjvms(this); 4806 set_control(_gvn.transform(slow_region)); 4807 uncommon_trap(Deoptimization::Reason_intrinsic, 4808 Deoptimization::Action_make_not_entrant); 4809 assert(stopped(), "Should be stopped"); 4810 } 4811 4812 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr(); 4813 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass()); 4814 src = _gvn.transform(new CheckCastPPNode(control(), src, toop)); 4815 } 4816 4817 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx); 4818 4819 if (stopped()) { 4820 return true; 4821 } 4822 4823 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated, 4824 // Create LoadRange and LoadKlass nodes for use during macro expansion here 4825 // so the compiler has a chance to eliminate them: during macro expansion, 4826 // we have to set their control (CastPP nodes are eliminated). 4827 load_object_klass(src), load_object_klass(dest), 4828 load_array_length(src), load_array_length(dest)); 4829 4830 ac->set_arraycopy(validated); 4831 4832 Node* n = _gvn.transform(ac); 4833 if (n == ac) { 4834 ac->connect_outputs(this); 4835 } else { 4836 assert(validated, "shouldn't transform if all arguments not validated"); 4837 set_all_memory(n); 4838 } 4839 clear_upper_avx(); 4840 4841 4842 return true; 4843 } 4844 4845 4846 // Helper function which determines if an arraycopy immediately follows 4847 // an allocation, with no intervening tests or other escapes for the object. 4848 AllocateArrayNode* 4849 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4850 RegionNode* slow_region) { 4851 if (stopped()) return NULL; // no fast path 4852 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4853 4854 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4855 if (alloc == NULL) return NULL; 4856 4857 Node* rawmem = memory(Compile::AliasIdxRaw); 4858 // Is the allocation's memory state untouched? 4859 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4860 // Bail out if there have been raw-memory effects since the allocation. 4861 // (Example: There might have been a call or safepoint.) 4862 return NULL; 4863 } 4864 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4865 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4866 return NULL; 4867 } 4868 4869 // There must be no unexpected observers of this allocation. 4870 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4871 Node* obs = ptr->fast_out(i); 4872 if (obs != this->map()) { 4873 return NULL; 4874 } 4875 } 4876 4877 // This arraycopy must unconditionally follow the allocation of the ptr. 4878 Node* alloc_ctl = ptr->in(0); 4879 Node* ctl = control(); 4880 while (ctl != alloc_ctl) { 4881 // There may be guards which feed into the slow_region. 4882 // Any other control flow means that we might not get a chance 4883 // to finish initializing the allocated object. 4884 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4885 IfNode* iff = ctl->in(0)->as_If(); 4886 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con); 4887 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4888 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4889 ctl = iff->in(0); // This test feeds the known slow_region. 4890 continue; 4891 } 4892 // One more try: Various low-level checks bottom out in 4893 // uncommon traps. If the debug-info of the trap omits 4894 // any reference to the allocation, as we've already 4895 // observed, then there can be no objection to the trap. 4896 bool found_trap = false; 4897 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 4898 Node* obs = not_ctl->fast_out(j); 4899 if (obs->in(0) == not_ctl && obs->is_Call() && 4900 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 4901 found_trap = true; break; 4902 } 4903 } 4904 if (found_trap) { 4905 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 4906 continue; 4907 } 4908 } 4909 return NULL; 4910 } 4911 4912 // If we get this far, we have an allocation which immediately 4913 // precedes the arraycopy, and we can take over zeroing the new object. 4914 // The arraycopy will finish the initialization, and provide 4915 // a new control state to which we will anchor the destination pointer. 4916 4917 return alloc; 4918 } 4919 4920 //-------------inline_encodeISOArray----------------------------------- 4921 // encode char[] to byte[] in ISO_8859_1 4922 bool LibraryCallKit::inline_encodeISOArray() { 4923 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 4924 // no receiver since it is static method 4925 Node *src = argument(0); 4926 Node *src_offset = argument(1); 4927 Node *dst = argument(2); 4928 Node *dst_offset = argument(3); 4929 Node *length = argument(4); 4930 4931 src = must_be_not_null(src, true); 4932 dst = must_be_not_null(dst, true); 4933 4934 const Type* src_type = src->Value(&_gvn); 4935 const Type* dst_type = dst->Value(&_gvn); 4936 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4937 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 4938 if (top_src == NULL || top_src->klass() == NULL || 4939 top_dest == NULL || top_dest->klass() == NULL) { 4940 // failed array check 4941 return false; 4942 } 4943 4944 // Figure out the size and type of the elements we will be copying. 4945 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4946 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4947 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 4948 return false; 4949 } 4950 4951 Node* src_start = array_element_address(src, src_offset, T_CHAR); 4952 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 4953 // 'src_start' points to src array + scaled offset 4954 // 'dst_start' points to dst array + scaled offset 4955 4956 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 4957 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 4958 enc = _gvn.transform(enc); 4959 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 4960 set_memory(res_mem, mtype); 4961 set_result(enc); 4962 clear_upper_avx(); 4963 4964 return true; 4965 } 4966 4967 //-------------inline_multiplyToLen----------------------------------- 4968 bool LibraryCallKit::inline_multiplyToLen() { 4969 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 4970 4971 address stubAddr = StubRoutines::multiplyToLen(); 4972 if (stubAddr == NULL) { 4973 return false; // Intrinsic's stub is not implemented on this platform 4974 } 4975 const char* stubName = "multiplyToLen"; 4976 4977 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 4978 4979 // no receiver because it is a static method 4980 Node* x = argument(0); 4981 Node* xlen = argument(1); 4982 Node* y = argument(2); 4983 Node* ylen = argument(3); 4984 Node* z = argument(4); 4985 4986 x = must_be_not_null(x, true); 4987 y = must_be_not_null(y, true); 4988 4989 const Type* x_type = x->Value(&_gvn); 4990 const Type* y_type = y->Value(&_gvn); 4991 const TypeAryPtr* top_x = x_type->isa_aryptr(); 4992 const TypeAryPtr* top_y = y_type->isa_aryptr(); 4993 if (top_x == NULL || top_x->klass() == NULL || 4994 top_y == NULL || top_y->klass() == NULL) { 4995 // failed array check 4996 return false; 4997 } 4998 4999 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5000 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5001 if (x_elem != T_INT || y_elem != T_INT) { 5002 return false; 5003 } 5004 5005 // Set the original stack and the reexecute bit for the interpreter to reexecute 5006 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5007 // on the return from z array allocation in runtime. 5008 { PreserveReexecuteState preexecs(this); 5009 jvms()->set_should_reexecute(true); 5010 5011 Node* x_start = array_element_address(x, intcon(0), x_elem); 5012 Node* y_start = array_element_address(y, intcon(0), y_elem); 5013 // 'x_start' points to x array + scaled xlen 5014 // 'y_start' points to y array + scaled ylen 5015 5016 // Allocate the result array 5017 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5018 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5019 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5020 5021 IdealKit ideal(this); 5022 5023 #define __ ideal. 5024 Node* one = __ ConI(1); 5025 Node* zero = __ ConI(0); 5026 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5027 __ set(need_alloc, zero); 5028 __ set(z_alloc, z); 5029 __ if_then(z, BoolTest::eq, null()); { 5030 __ increment (need_alloc, one); 5031 } __ else_(); { 5032 // Update graphKit memory and control from IdealKit. 5033 sync_kit(ideal); 5034 Node *cast = new CastPPNode(z, TypePtr::NOTNULL); 5035 cast->init_req(0, control()); 5036 _gvn.set_type(cast, cast->bottom_type()); 5037 C->record_for_igvn(cast); 5038 5039 Node* zlen_arg = load_array_length(cast); 5040 // Update IdealKit memory and control from graphKit. 5041 __ sync_kit(this); 5042 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5043 __ increment (need_alloc, one); 5044 } __ end_if(); 5045 } __ end_if(); 5046 5047 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5048 // Update graphKit memory and control from IdealKit. 5049 sync_kit(ideal); 5050 Node * narr = new_array(klass_node, zlen, 1); 5051 // Update IdealKit memory and control from graphKit. 5052 __ sync_kit(this); 5053 __ set(z_alloc, narr); 5054 } __ end_if(); 5055 5056 sync_kit(ideal); 5057 z = __ value(z_alloc); 5058 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5059 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5060 // Final sync IdealKit and GraphKit. 5061 final_sync(ideal); 5062 #undef __ 5063 5064 Node* z_start = array_element_address(z, intcon(0), T_INT); 5065 5066 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5067 OptoRuntime::multiplyToLen_Type(), 5068 stubAddr, stubName, TypePtr::BOTTOM, 5069 x_start, xlen, y_start, ylen, z_start, zlen); 5070 } // original reexecute is set back here 5071 5072 C->set_has_split_ifs(true); // Has chance for split-if optimization 5073 set_result(z); 5074 return true; 5075 } 5076 5077 //-------------inline_squareToLen------------------------------------ 5078 bool LibraryCallKit::inline_squareToLen() { 5079 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5080 5081 address stubAddr = StubRoutines::squareToLen(); 5082 if (stubAddr == NULL) { 5083 return false; // Intrinsic's stub is not implemented on this platform 5084 } 5085 const char* stubName = "squareToLen"; 5086 5087 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5088 5089 Node* x = argument(0); 5090 Node* len = argument(1); 5091 Node* z = argument(2); 5092 Node* zlen = argument(3); 5093 5094 x = must_be_not_null(x, true); 5095 z = must_be_not_null(z, true); 5096 5097 const Type* x_type = x->Value(&_gvn); 5098 const Type* z_type = z->Value(&_gvn); 5099 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5100 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5101 if (top_x == NULL || top_x->klass() == NULL || 5102 top_z == NULL || top_z->klass() == NULL) { 5103 // failed array check 5104 return false; 5105 } 5106 5107 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5108 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5109 if (x_elem != T_INT || z_elem != T_INT) { 5110 return false; 5111 } 5112 5113 5114 Node* x_start = array_element_address(x, intcon(0), x_elem); 5115 Node* z_start = array_element_address(z, intcon(0), z_elem); 5116 5117 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5118 OptoRuntime::squareToLen_Type(), 5119 stubAddr, stubName, TypePtr::BOTTOM, 5120 x_start, len, z_start, zlen); 5121 5122 set_result(z); 5123 return true; 5124 } 5125 5126 //-------------inline_mulAdd------------------------------------------ 5127 bool LibraryCallKit::inline_mulAdd() { 5128 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5129 5130 address stubAddr = StubRoutines::mulAdd(); 5131 if (stubAddr == NULL) { 5132 return false; // Intrinsic's stub is not implemented on this platform 5133 } 5134 const char* stubName = "mulAdd"; 5135 5136 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5137 5138 Node* out = argument(0); 5139 Node* in = argument(1); 5140 Node* offset = argument(2); 5141 Node* len = argument(3); 5142 Node* k = argument(4); 5143 5144 out = must_be_not_null(out, true); 5145 5146 const Type* out_type = out->Value(&_gvn); 5147 const Type* in_type = in->Value(&_gvn); 5148 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5149 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5150 if (top_out == NULL || top_out->klass() == NULL || 5151 top_in == NULL || top_in->klass() == NULL) { 5152 // failed array check 5153 return false; 5154 } 5155 5156 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5157 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5158 if (out_elem != T_INT || in_elem != T_INT) { 5159 return false; 5160 } 5161 5162 Node* outlen = load_array_length(out); 5163 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5164 Node* out_start = array_element_address(out, intcon(0), out_elem); 5165 Node* in_start = array_element_address(in, intcon(0), in_elem); 5166 5167 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5168 OptoRuntime::mulAdd_Type(), 5169 stubAddr, stubName, TypePtr::BOTTOM, 5170 out_start,in_start, new_offset, len, k); 5171 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5172 set_result(result); 5173 return true; 5174 } 5175 5176 //-------------inline_montgomeryMultiply----------------------------------- 5177 bool LibraryCallKit::inline_montgomeryMultiply() { 5178 address stubAddr = StubRoutines::montgomeryMultiply(); 5179 if (stubAddr == NULL) { 5180 return false; // Intrinsic's stub is not implemented on this platform 5181 } 5182 5183 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5184 const char* stubName = "montgomery_multiply"; 5185 5186 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5187 5188 Node* a = argument(0); 5189 Node* b = argument(1); 5190 Node* n = argument(2); 5191 Node* len = argument(3); 5192 Node* inv = argument(4); 5193 Node* m = argument(6); 5194 5195 const Type* a_type = a->Value(&_gvn); 5196 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5197 const Type* b_type = b->Value(&_gvn); 5198 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5199 const Type* n_type = a->Value(&_gvn); 5200 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5201 const Type* m_type = a->Value(&_gvn); 5202 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5203 if (top_a == NULL || top_a->klass() == NULL || 5204 top_b == NULL || top_b->klass() == NULL || 5205 top_n == NULL || top_n->klass() == NULL || 5206 top_m == NULL || top_m->klass() == NULL) { 5207 // failed array check 5208 return false; 5209 } 5210 5211 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5212 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5213 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5214 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5215 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5216 return false; 5217 } 5218 5219 // Make the call 5220 { 5221 Node* a_start = array_element_address(a, intcon(0), a_elem); 5222 Node* b_start = array_element_address(b, intcon(0), b_elem); 5223 Node* n_start = array_element_address(n, intcon(0), n_elem); 5224 Node* m_start = array_element_address(m, intcon(0), m_elem); 5225 5226 Node* call = make_runtime_call(RC_LEAF, 5227 OptoRuntime::montgomeryMultiply_Type(), 5228 stubAddr, stubName, TypePtr::BOTTOM, 5229 a_start, b_start, n_start, len, inv, top(), 5230 m_start); 5231 set_result(m); 5232 } 5233 5234 return true; 5235 } 5236 5237 bool LibraryCallKit::inline_montgomerySquare() { 5238 address stubAddr = StubRoutines::montgomerySquare(); 5239 if (stubAddr == NULL) { 5240 return false; // Intrinsic's stub is not implemented on this platform 5241 } 5242 5243 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5244 const char* stubName = "montgomery_square"; 5245 5246 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5247 5248 Node* a = argument(0); 5249 Node* n = argument(1); 5250 Node* len = argument(2); 5251 Node* inv = argument(3); 5252 Node* m = argument(5); 5253 5254 const Type* a_type = a->Value(&_gvn); 5255 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5256 const Type* n_type = a->Value(&_gvn); 5257 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5258 const Type* m_type = a->Value(&_gvn); 5259 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5260 if (top_a == NULL || top_a->klass() == NULL || 5261 top_n == NULL || top_n->klass() == NULL || 5262 top_m == NULL || top_m->klass() == NULL) { 5263 // failed array check 5264 return false; 5265 } 5266 5267 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5268 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5269 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5270 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5271 return false; 5272 } 5273 5274 // Make the call 5275 { 5276 Node* a_start = array_element_address(a, intcon(0), a_elem); 5277 Node* n_start = array_element_address(n, intcon(0), n_elem); 5278 Node* m_start = array_element_address(m, intcon(0), m_elem); 5279 5280 Node* call = make_runtime_call(RC_LEAF, 5281 OptoRuntime::montgomerySquare_Type(), 5282 stubAddr, stubName, TypePtr::BOTTOM, 5283 a_start, n_start, len, inv, top(), 5284 m_start); 5285 set_result(m); 5286 } 5287 5288 return true; 5289 } 5290 5291 bool LibraryCallKit::inline_bigIntegerShift(bool isRightShift) { 5292 address stubAddr = NULL; 5293 const char* stubName = NULL; 5294 5295 stubAddr = isRightShift? StubRoutines::bigIntegerRightShift(): StubRoutines::bigIntegerLeftShift(); 5296 if (stubAddr == NULL) { 5297 return false; // Intrinsic's stub is not implemented on this platform 5298 } 5299 5300 stubName = isRightShift? "bigIntegerRightShiftWorker" : "bigIntegerLeftShiftWorker"; 5301 5302 assert(callee()->signature()->size() == 5, "expected 5 arguments"); 5303 5304 Node* newArr = argument(0); 5305 Node* oldArr = argument(1); 5306 Node* newIdx = argument(2); 5307 Node* shiftCount = argument(3); 5308 Node* numIter = argument(4); 5309 5310 const Type* newArr_type = newArr->Value(&_gvn); 5311 const TypeAryPtr* top_newArr = newArr_type->isa_aryptr(); 5312 const Type* oldArr_type = oldArr->Value(&_gvn); 5313 const TypeAryPtr* top_oldArr = oldArr_type->isa_aryptr(); 5314 if (top_newArr == NULL || top_newArr->klass() == NULL || top_oldArr == NULL 5315 || top_oldArr->klass() == NULL) { 5316 return false; 5317 } 5318 5319 BasicType newArr_elem = newArr_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5320 BasicType oldArr_elem = oldArr_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5321 if (newArr_elem != T_INT || oldArr_elem != T_INT) { 5322 return false; 5323 } 5324 5325 // Make the call 5326 { 5327 Node* newArr_start = array_element_address(newArr, intcon(0), newArr_elem); 5328 Node* oldArr_start = array_element_address(oldArr, intcon(0), oldArr_elem); 5329 5330 Node* call = make_runtime_call(RC_LEAF, 5331 OptoRuntime::bigIntegerShift_Type(), 5332 stubAddr, 5333 stubName, 5334 TypePtr::BOTTOM, 5335 newArr_start, 5336 oldArr_start, 5337 newIdx, 5338 shiftCount, 5339 numIter); 5340 } 5341 5342 return true; 5343 } 5344 5345 //-------------inline_vectorizedMismatch------------------------------ 5346 bool LibraryCallKit::inline_vectorizedMismatch() { 5347 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5348 5349 address stubAddr = StubRoutines::vectorizedMismatch(); 5350 if (stubAddr == NULL) { 5351 return false; // Intrinsic's stub is not implemented on this platform 5352 } 5353 const char* stubName = "vectorizedMismatch"; 5354 int size_l = callee()->signature()->size(); 5355 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5356 5357 Node* obja = argument(0); 5358 Node* aoffset = argument(1); 5359 Node* objb = argument(3); 5360 Node* boffset = argument(4); 5361 Node* length = argument(6); 5362 Node* scale = argument(7); 5363 5364 const Type* a_type = obja->Value(&_gvn); 5365 const Type* b_type = objb->Value(&_gvn); 5366 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5367 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5368 if (top_a == NULL || top_a->klass() == NULL || 5369 top_b == NULL || top_b->klass() == NULL) { 5370 // failed array check 5371 return false; 5372 } 5373 5374 Node* call; 5375 jvms()->set_should_reexecute(true); 5376 5377 Node* obja_adr = make_unsafe_address(obja, aoffset, ACCESS_READ); 5378 Node* objb_adr = make_unsafe_address(objb, boffset, ACCESS_READ); 5379 5380 call = make_runtime_call(RC_LEAF, 5381 OptoRuntime::vectorizedMismatch_Type(), 5382 stubAddr, stubName, TypePtr::BOTTOM, 5383 obja_adr, objb_adr, length, scale); 5384 5385 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5386 set_result(result); 5387 return true; 5388 } 5389 5390 /** 5391 * Calculate CRC32 for byte. 5392 * int java.util.zip.CRC32.update(int crc, int b) 5393 */ 5394 bool LibraryCallKit::inline_updateCRC32() { 5395 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5396 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5397 // no receiver since it is static method 5398 Node* crc = argument(0); // type: int 5399 Node* b = argument(1); // type: int 5400 5401 /* 5402 * int c = ~ crc; 5403 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5404 * b = b ^ (c >>> 8); 5405 * crc = ~b; 5406 */ 5407 5408 Node* M1 = intcon(-1); 5409 crc = _gvn.transform(new XorINode(crc, M1)); 5410 Node* result = _gvn.transform(new XorINode(crc, b)); 5411 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5412 5413 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5414 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5415 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5416 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5417 5418 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5419 result = _gvn.transform(new XorINode(crc, result)); 5420 result = _gvn.transform(new XorINode(result, M1)); 5421 set_result(result); 5422 return true; 5423 } 5424 5425 /** 5426 * Calculate CRC32 for byte[] array. 5427 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5428 */ 5429 bool LibraryCallKit::inline_updateBytesCRC32() { 5430 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5431 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5432 // no receiver since it is static method 5433 Node* crc = argument(0); // type: int 5434 Node* src = argument(1); // type: oop 5435 Node* offset = argument(2); // type: int 5436 Node* length = argument(3); // type: int 5437 5438 const Type* src_type = src->Value(&_gvn); 5439 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5440 if (top_src == NULL || top_src->klass() == NULL) { 5441 // failed array check 5442 return false; 5443 } 5444 5445 // Figure out the size and type of the elements we will be copying. 5446 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5447 if (src_elem != T_BYTE) { 5448 return false; 5449 } 5450 5451 // 'src_start' points to src array + scaled offset 5452 src = must_be_not_null(src, true); 5453 Node* src_start = array_element_address(src, offset, src_elem); 5454 5455 // We assume that range check is done by caller. 5456 // TODO: generate range check (offset+length < src.length) in debug VM. 5457 5458 // Call the stub. 5459 address stubAddr = StubRoutines::updateBytesCRC32(); 5460 const char *stubName = "updateBytesCRC32"; 5461 5462 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5463 stubAddr, stubName, TypePtr::BOTTOM, 5464 crc, src_start, length); 5465 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5466 set_result(result); 5467 return true; 5468 } 5469 5470 /** 5471 * Calculate CRC32 for ByteBuffer. 5472 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5473 */ 5474 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5475 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5476 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5477 // no receiver since it is static method 5478 Node* crc = argument(0); // type: int 5479 Node* src = argument(1); // type: long 5480 Node* offset = argument(3); // type: int 5481 Node* length = argument(4); // type: int 5482 5483 src = ConvL2X(src); // adjust Java long to machine word 5484 Node* base = _gvn.transform(new CastX2PNode(src)); 5485 offset = ConvI2X(offset); 5486 5487 // 'src_start' points to src array + scaled offset 5488 Node* src_start = basic_plus_adr(top(), base, offset); 5489 5490 // Call the stub. 5491 address stubAddr = StubRoutines::updateBytesCRC32(); 5492 const char *stubName = "updateBytesCRC32"; 5493 5494 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5495 stubAddr, stubName, TypePtr::BOTTOM, 5496 crc, src_start, length); 5497 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5498 set_result(result); 5499 return true; 5500 } 5501 5502 //------------------------------get_table_from_crc32c_class----------------------- 5503 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5504 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5505 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5506 5507 return table; 5508 } 5509 5510 //------------------------------inline_updateBytesCRC32C----------------------- 5511 // 5512 // Calculate CRC32C for byte[] array. 5513 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5514 // 5515 bool LibraryCallKit::inline_updateBytesCRC32C() { 5516 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5517 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5518 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5519 // no receiver since it is a static method 5520 Node* crc = argument(0); // type: int 5521 Node* src = argument(1); // type: oop 5522 Node* offset = argument(2); // type: int 5523 Node* end = argument(3); // type: int 5524 5525 Node* length = _gvn.transform(new SubINode(end, offset)); 5526 5527 const Type* src_type = src->Value(&_gvn); 5528 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5529 if (top_src == NULL || top_src->klass() == NULL) { 5530 // failed array check 5531 return false; 5532 } 5533 5534 // Figure out the size and type of the elements we will be copying. 5535 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5536 if (src_elem != T_BYTE) { 5537 return false; 5538 } 5539 5540 // 'src_start' points to src array + scaled offset 5541 src = must_be_not_null(src, true); 5542 Node* src_start = array_element_address(src, offset, src_elem); 5543 5544 // static final int[] byteTable in class CRC32C 5545 Node* table = get_table_from_crc32c_class(callee()->holder()); 5546 table = must_be_not_null(table, true); 5547 Node* table_start = array_element_address(table, intcon(0), T_INT); 5548 5549 // We assume that range check is done by caller. 5550 // TODO: generate range check (offset+length < src.length) in debug VM. 5551 5552 // Call the stub. 5553 address stubAddr = StubRoutines::updateBytesCRC32C(); 5554 const char *stubName = "updateBytesCRC32C"; 5555 5556 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5557 stubAddr, stubName, TypePtr::BOTTOM, 5558 crc, src_start, length, table_start); 5559 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5560 set_result(result); 5561 return true; 5562 } 5563 5564 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5565 // 5566 // Calculate CRC32C for DirectByteBuffer. 5567 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5568 // 5569 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5570 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5571 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5572 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5573 // no receiver since it is a static method 5574 Node* crc = argument(0); // type: int 5575 Node* src = argument(1); // type: long 5576 Node* offset = argument(3); // type: int 5577 Node* end = argument(4); // type: int 5578 5579 Node* length = _gvn.transform(new SubINode(end, offset)); 5580 5581 src = ConvL2X(src); // adjust Java long to machine word 5582 Node* base = _gvn.transform(new CastX2PNode(src)); 5583 offset = ConvI2X(offset); 5584 5585 // 'src_start' points to src array + scaled offset 5586 Node* src_start = basic_plus_adr(top(), base, offset); 5587 5588 // static final int[] byteTable in class CRC32C 5589 Node* table = get_table_from_crc32c_class(callee()->holder()); 5590 table = must_be_not_null(table, true); 5591 Node* table_start = array_element_address(table, intcon(0), T_INT); 5592 5593 // Call the stub. 5594 address stubAddr = StubRoutines::updateBytesCRC32C(); 5595 const char *stubName = "updateBytesCRC32C"; 5596 5597 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5598 stubAddr, stubName, TypePtr::BOTTOM, 5599 crc, src_start, length, table_start); 5600 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5601 set_result(result); 5602 return true; 5603 } 5604 5605 //------------------------------inline_updateBytesAdler32---------------------- 5606 // 5607 // Calculate Adler32 checksum for byte[] array. 5608 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5609 // 5610 bool LibraryCallKit::inline_updateBytesAdler32() { 5611 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5612 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5613 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5614 // no receiver since it is static method 5615 Node* crc = argument(0); // type: int 5616 Node* src = argument(1); // type: oop 5617 Node* offset = argument(2); // type: int 5618 Node* length = argument(3); // type: int 5619 5620 const Type* src_type = src->Value(&_gvn); 5621 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5622 if (top_src == NULL || top_src->klass() == NULL) { 5623 // failed array check 5624 return false; 5625 } 5626 5627 // Figure out the size and type of the elements we will be copying. 5628 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5629 if (src_elem != T_BYTE) { 5630 return false; 5631 } 5632 5633 // 'src_start' points to src array + scaled offset 5634 Node* src_start = array_element_address(src, offset, src_elem); 5635 5636 // We assume that range check is done by caller. 5637 // TODO: generate range check (offset+length < src.length) in debug VM. 5638 5639 // Call the stub. 5640 address stubAddr = StubRoutines::updateBytesAdler32(); 5641 const char *stubName = "updateBytesAdler32"; 5642 5643 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5644 stubAddr, stubName, TypePtr::BOTTOM, 5645 crc, src_start, length); 5646 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5647 set_result(result); 5648 return true; 5649 } 5650 5651 //------------------------------inline_updateByteBufferAdler32--------------- 5652 // 5653 // Calculate Adler32 checksum for DirectByteBuffer. 5654 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5655 // 5656 bool LibraryCallKit::inline_updateByteBufferAdler32() { 5657 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5658 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5659 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5660 // no receiver since it is static method 5661 Node* crc = argument(0); // type: int 5662 Node* src = argument(1); // type: long 5663 Node* offset = argument(3); // type: int 5664 Node* length = argument(4); // type: int 5665 5666 src = ConvL2X(src); // adjust Java long to machine word 5667 Node* base = _gvn.transform(new CastX2PNode(src)); 5668 offset = ConvI2X(offset); 5669 5670 // 'src_start' points to src array + scaled offset 5671 Node* src_start = basic_plus_adr(top(), base, offset); 5672 5673 // Call the stub. 5674 address stubAddr = StubRoutines::updateBytesAdler32(); 5675 const char *stubName = "updateBytesAdler32"; 5676 5677 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5678 stubAddr, stubName, TypePtr::BOTTOM, 5679 crc, src_start, length); 5680 5681 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5682 set_result(result); 5683 return true; 5684 } 5685 5686 //----------------------------inline_reference_get---------------------------- 5687 // public T java.lang.ref.Reference.get(); 5688 bool LibraryCallKit::inline_reference_get() { 5689 const int referent_offset = java_lang_ref_Reference::referent_offset(); 5690 5691 // Get the argument: 5692 Node* reference_obj = null_check_receiver(); 5693 if (stopped()) return true; 5694 5695 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr(); 5696 assert(tinst != NULL, "obj is null"); 5697 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5698 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass(); 5699 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"), 5700 ciSymbol::make("Ljava/lang/Object;"), 5701 false); 5702 assert (field != NULL, "undefined field"); 5703 5704 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5705 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5706 5707 ciInstanceKlass* klass = env()->Object_klass(); 5708 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5709 5710 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF; 5711 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators); 5712 // Add memory barrier to prevent commoning reads from this field 5713 // across safepoint since GC can change its value. 5714 insert_mem_bar(Op_MemBarCPUOrder); 5715 5716 set_result(result); 5717 return true; 5718 } 5719 5720 5721 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5722 bool is_exact=true, bool is_static=false, 5723 ciInstanceKlass * fromKls=NULL) { 5724 if (fromKls == NULL) { 5725 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5726 assert(tinst != NULL, "obj is null"); 5727 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5728 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5729 fromKls = tinst->klass()->as_instance_klass(); 5730 } else { 5731 assert(is_static, "only for static field access"); 5732 } 5733 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5734 ciSymbol::make(fieldTypeString), 5735 is_static); 5736 5737 assert (field != NULL, "undefined field"); 5738 if (field == NULL) return (Node *) NULL; 5739 5740 if (is_static) { 5741 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5742 fromObj = makecon(tip); 5743 } 5744 5745 // Next code copied from Parse::do_get_xxx(): 5746 5747 // Compute address and memory type. 5748 int offset = field->offset_in_bytes(); 5749 bool is_vol = field->is_volatile(); 5750 ciType* field_klass = field->type(); 5751 assert(field_klass->is_loaded(), "should be loaded"); 5752 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5753 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5754 BasicType bt = field->layout_type(); 5755 5756 // Build the resultant type of the load 5757 const Type *type; 5758 if (bt == T_OBJECT) { 5759 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5760 } else { 5761 type = Type::get_const_basic_type(bt); 5762 } 5763 5764 DecoratorSet decorators = IN_HEAP; 5765 5766 if (is_vol) { 5767 decorators |= MO_SEQ_CST; 5768 } 5769 5770 return access_load_at(fromObj, adr, adr_type, type, bt, decorators); 5771 } 5772 5773 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5774 bool is_exact = true, bool is_static = false, 5775 ciInstanceKlass * fromKls = NULL) { 5776 if (fromKls == NULL) { 5777 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5778 assert(tinst != NULL, "obj is null"); 5779 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5780 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5781 fromKls = tinst->klass()->as_instance_klass(); 5782 } 5783 else { 5784 assert(is_static, "only for static field access"); 5785 } 5786 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5787 ciSymbol::make(fieldTypeString), 5788 is_static); 5789 5790 assert(field != NULL, "undefined field"); 5791 assert(!field->is_volatile(), "not defined for volatile fields"); 5792 5793 if (is_static) { 5794 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5795 fromObj = makecon(tip); 5796 } 5797 5798 // Next code copied from Parse::do_get_xxx(): 5799 5800 // Compute address and memory type. 5801 int offset = field->offset_in_bytes(); 5802 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5803 5804 return adr; 5805 } 5806 5807 //------------------------------inline_aescrypt_Block----------------------- 5808 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5809 address stubAddr = NULL; 5810 const char *stubName; 5811 assert(UseAES, "need AES instruction support"); 5812 5813 switch(id) { 5814 case vmIntrinsics::_aescrypt_encryptBlock: 5815 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5816 stubName = "aescrypt_encryptBlock"; 5817 break; 5818 case vmIntrinsics::_aescrypt_decryptBlock: 5819 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5820 stubName = "aescrypt_decryptBlock"; 5821 break; 5822 default: 5823 break; 5824 } 5825 if (stubAddr == NULL) return false; 5826 5827 Node* aescrypt_object = argument(0); 5828 Node* src = argument(1); 5829 Node* src_offset = argument(2); 5830 Node* dest = argument(3); 5831 Node* dest_offset = argument(4); 5832 5833 src = must_be_not_null(src, true); 5834 dest = must_be_not_null(dest, true); 5835 5836 // (1) src and dest are arrays. 5837 const Type* src_type = src->Value(&_gvn); 5838 const Type* dest_type = dest->Value(&_gvn); 5839 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5840 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5841 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5842 5843 // for the quick and dirty code we will skip all the checks. 5844 // we are just trying to get the call to be generated. 5845 Node* src_start = src; 5846 Node* dest_start = dest; 5847 if (src_offset != NULL || dest_offset != NULL) { 5848 assert(src_offset != NULL && dest_offset != NULL, ""); 5849 src_start = array_element_address(src, src_offset, T_BYTE); 5850 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5851 } 5852 5853 // now need to get the start of its expanded key array 5854 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5855 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5856 if (k_start == NULL) return false; 5857 5858 if (Matcher::pass_original_key_for_aes()) { 5859 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5860 // compatibility issues between Java key expansion and SPARC crypto instructions 5861 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5862 if (original_k_start == NULL) return false; 5863 5864 // Call the stub. 5865 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5866 stubAddr, stubName, TypePtr::BOTTOM, 5867 src_start, dest_start, k_start, original_k_start); 5868 } else { 5869 // Call the stub. 5870 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5871 stubAddr, stubName, TypePtr::BOTTOM, 5872 src_start, dest_start, k_start); 5873 } 5874 5875 return true; 5876 } 5877 5878 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5879 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5880 address stubAddr = NULL; 5881 const char *stubName = NULL; 5882 5883 assert(UseAES, "need AES instruction support"); 5884 5885 switch(id) { 5886 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5887 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5888 stubName = "cipherBlockChaining_encryptAESCrypt"; 5889 break; 5890 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5891 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5892 stubName = "cipherBlockChaining_decryptAESCrypt"; 5893 break; 5894 default: 5895 break; 5896 } 5897 if (stubAddr == NULL) return false; 5898 5899 Node* cipherBlockChaining_object = argument(0); 5900 Node* src = argument(1); 5901 Node* src_offset = argument(2); 5902 Node* len = argument(3); 5903 Node* dest = argument(4); 5904 Node* dest_offset = argument(5); 5905 5906 src = must_be_not_null(src, false); 5907 dest = must_be_not_null(dest, false); 5908 5909 // (1) src and dest are arrays. 5910 const Type* src_type = src->Value(&_gvn); 5911 const Type* dest_type = dest->Value(&_gvn); 5912 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5913 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5914 assert (top_src != NULL && top_src->klass() != NULL 5915 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5916 5917 // checks are the responsibility of the caller 5918 Node* src_start = src; 5919 Node* dest_start = dest; 5920 if (src_offset != NULL || dest_offset != NULL) { 5921 assert(src_offset != NULL && dest_offset != NULL, ""); 5922 src_start = array_element_address(src, src_offset, T_BYTE); 5923 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5924 } 5925 5926 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5927 // (because of the predicated logic executed earlier). 5928 // so we cast it here safely. 5929 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5930 5931 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5932 if (embeddedCipherObj == NULL) return false; 5933 5934 // cast it to what we know it will be at runtime 5935 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5936 assert(tinst != NULL, "CBC obj is null"); 5937 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5938 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5939 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5940 5941 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5942 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5943 const TypeOopPtr* xtype = aklass->as_instance_type(); 5944 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5945 aescrypt_object = _gvn.transform(aescrypt_object); 5946 5947 // we need to get the start of the aescrypt_object's expanded key array 5948 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5949 if (k_start == NULL) return false; 5950 5951 // similarly, get the start address of the r vector 5952 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5953 if (objRvec == NULL) return false; 5954 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5955 5956 Node* cbcCrypt; 5957 if (Matcher::pass_original_key_for_aes()) { 5958 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5959 // compatibility issues between Java key expansion and SPARC crypto instructions 5960 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5961 if (original_k_start == NULL) return false; 5962 5963 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5964 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5965 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5966 stubAddr, stubName, TypePtr::BOTTOM, 5967 src_start, dest_start, k_start, r_start, len, original_k_start); 5968 } else { 5969 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5970 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5971 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5972 stubAddr, stubName, TypePtr::BOTTOM, 5973 src_start, dest_start, k_start, r_start, len); 5974 } 5975 5976 // return cipher length (int) 5977 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5978 set_result(retvalue); 5979 return true; 5980 } 5981 5982 //------------------------------inline_electronicCodeBook_AESCrypt----------------------- 5983 bool LibraryCallKit::inline_electronicCodeBook_AESCrypt(vmIntrinsics::ID id) { 5984 address stubAddr = NULL; 5985 const char *stubName = NULL; 5986 5987 assert(UseAES, "need AES instruction support"); 5988 5989 switch (id) { 5990 case vmIntrinsics::_electronicCodeBook_encryptAESCrypt: 5991 stubAddr = StubRoutines::electronicCodeBook_encryptAESCrypt(); 5992 stubName = "electronicCodeBook_encryptAESCrypt"; 5993 break; 5994 case vmIntrinsics::_electronicCodeBook_decryptAESCrypt: 5995 stubAddr = StubRoutines::electronicCodeBook_decryptAESCrypt(); 5996 stubName = "electronicCodeBook_decryptAESCrypt"; 5997 break; 5998 default: 5999 break; 6000 } 6001 6002 if (stubAddr == NULL) return false; 6003 6004 Node* electronicCodeBook_object = argument(0); 6005 Node* src = argument(1); 6006 Node* src_offset = argument(2); 6007 Node* len = argument(3); 6008 Node* dest = argument(4); 6009 Node* dest_offset = argument(5); 6010 6011 // (1) src and dest are arrays. 6012 const Type* src_type = src->Value(&_gvn); 6013 const Type* dest_type = dest->Value(&_gvn); 6014 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6015 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6016 assert(top_src != NULL && top_src->klass() != NULL 6017 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6018 6019 // checks are the responsibility of the caller 6020 Node* src_start = src; 6021 Node* dest_start = dest; 6022 if (src_offset != NULL || dest_offset != NULL) { 6023 assert(src_offset != NULL && dest_offset != NULL, ""); 6024 src_start = array_element_address(src, src_offset, T_BYTE); 6025 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6026 } 6027 6028 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6029 // (because of the predicated logic executed earlier). 6030 // so we cast it here safely. 6031 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6032 6033 Node* embeddedCipherObj = load_field_from_object(electronicCodeBook_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6034 if (embeddedCipherObj == NULL) return false; 6035 6036 // cast it to what we know it will be at runtime 6037 const TypeInstPtr* tinst = _gvn.type(electronicCodeBook_object)->isa_instptr(); 6038 assert(tinst != NULL, "ECB obj is null"); 6039 assert(tinst->klass()->is_loaded(), "ECB obj is not loaded"); 6040 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6041 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6042 6043 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6044 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6045 const TypeOopPtr* xtype = aklass->as_instance_type(); 6046 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6047 aescrypt_object = _gvn.transform(aescrypt_object); 6048 6049 // we need to get the start of the aescrypt_object's expanded key array 6050 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6051 if (k_start == NULL) return false; 6052 6053 Node* ecbCrypt; 6054 if (Matcher::pass_original_key_for_aes()) { 6055 // no SPARC version for AES/ECB intrinsics now. 6056 return false; 6057 } 6058 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6059 ecbCrypt = make_runtime_call(RC_LEAF | RC_NO_FP, 6060 OptoRuntime::electronicCodeBook_aescrypt_Type(), 6061 stubAddr, stubName, TypePtr::BOTTOM, 6062 src_start, dest_start, k_start, len); 6063 6064 // return cipher length (int) 6065 Node* retvalue = _gvn.transform(new ProjNode(ecbCrypt, TypeFunc::Parms)); 6066 set_result(retvalue); 6067 return true; 6068 } 6069 6070 //------------------------------inline_counterMode_AESCrypt----------------------- 6071 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6072 assert(UseAES, "need AES instruction support"); 6073 if (!UseAESCTRIntrinsics) return false; 6074 6075 address stubAddr = NULL; 6076 const char *stubName = NULL; 6077 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6078 stubAddr = StubRoutines::counterMode_AESCrypt(); 6079 stubName = "counterMode_AESCrypt"; 6080 } 6081 if (stubAddr == NULL) return false; 6082 6083 Node* counterMode_object = argument(0); 6084 Node* src = argument(1); 6085 Node* src_offset = argument(2); 6086 Node* len = argument(3); 6087 Node* dest = argument(4); 6088 Node* dest_offset = argument(5); 6089 6090 // (1) src and dest are arrays. 6091 const Type* src_type = src->Value(&_gvn); 6092 const Type* dest_type = dest->Value(&_gvn); 6093 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6094 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6095 assert(top_src != NULL && top_src->klass() != NULL && 6096 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6097 6098 // checks are the responsibility of the caller 6099 Node* src_start = src; 6100 Node* dest_start = dest; 6101 if (src_offset != NULL || dest_offset != NULL) { 6102 assert(src_offset != NULL && dest_offset != NULL, ""); 6103 src_start = array_element_address(src, src_offset, T_BYTE); 6104 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6105 } 6106 6107 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6108 // (because of the predicated logic executed earlier). 6109 // so we cast it here safely. 6110 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6111 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6112 if (embeddedCipherObj == NULL) return false; 6113 // cast it to what we know it will be at runtime 6114 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6115 assert(tinst != NULL, "CTR obj is null"); 6116 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6117 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6118 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6119 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6120 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6121 const TypeOopPtr* xtype = aklass->as_instance_type(); 6122 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6123 aescrypt_object = _gvn.transform(aescrypt_object); 6124 // we need to get the start of the aescrypt_object's expanded key array 6125 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6126 if (k_start == NULL) return false; 6127 // similarly, get the start address of the r vector 6128 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6129 if (obj_counter == NULL) return false; 6130 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6131 6132 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6133 if (saved_encCounter == NULL) return false; 6134 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6135 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6136 6137 Node* ctrCrypt; 6138 if (Matcher::pass_original_key_for_aes()) { 6139 // no SPARC version for AES/CTR intrinsics now. 6140 return false; 6141 } 6142 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6143 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6144 OptoRuntime::counterMode_aescrypt_Type(), 6145 stubAddr, stubName, TypePtr::BOTTOM, 6146 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6147 6148 // return cipher length (int) 6149 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6150 set_result(retvalue); 6151 return true; 6152 } 6153 6154 //------------------------------get_key_start_from_aescrypt_object----------------------- 6155 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6156 #if defined(PPC64) || defined(S390) 6157 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6158 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6159 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6160 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6161 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6162 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6163 if (objSessionK == NULL) { 6164 return (Node *) NULL; 6165 } 6166 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6167 #else 6168 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6169 #endif // PPC64 6170 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6171 if (objAESCryptKey == NULL) return (Node *) NULL; 6172 6173 // now have the array, need to get the start address of the K array 6174 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6175 return k_start; 6176 } 6177 6178 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6179 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6180 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6181 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6182 if (objAESCryptKey == NULL) return (Node *) NULL; 6183 6184 // now have the array, need to get the start address of the lastKey array 6185 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6186 return original_k_start; 6187 } 6188 6189 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6190 // Return node representing slow path of predicate check. 6191 // the pseudo code we want to emulate with this predicate is: 6192 // for encryption: 6193 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6194 // for decryption: 6195 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6196 // note cipher==plain is more conservative than the original java code but that's OK 6197 // 6198 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6199 // The receiver was checked for NULL already. 6200 Node* objCBC = argument(0); 6201 6202 Node* src = argument(1); 6203 Node* dest = argument(4); 6204 6205 // Load embeddedCipher field of CipherBlockChaining object. 6206 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6207 6208 // get AESCrypt klass for instanceOf check 6209 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6210 // will have same classloader as CipherBlockChaining object 6211 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6212 assert(tinst != NULL, "CBCobj is null"); 6213 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6214 6215 // we want to do an instanceof comparison against the AESCrypt class 6216 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6217 if (!klass_AESCrypt->is_loaded()) { 6218 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6219 Node* ctrl = control(); 6220 set_control(top()); // no regular fast path 6221 return ctrl; 6222 } 6223 6224 src = must_be_not_null(src, true); 6225 dest = must_be_not_null(dest, true); 6226 6227 // Resolve oops to stable for CmpP below. 6228 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6229 6230 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6231 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6232 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6233 6234 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6235 6236 // for encryption, we are done 6237 if (!decrypting) 6238 return instof_false; // even if it is NULL 6239 6240 // for decryption, we need to add a further check to avoid 6241 // taking the intrinsic path when cipher and plain are the same 6242 // see the original java code for why. 6243 RegionNode* region = new RegionNode(3); 6244 region->init_req(1, instof_false); 6245 6246 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6247 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6248 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6249 region->init_req(2, src_dest_conjoint); 6250 6251 record_for_igvn(region); 6252 return _gvn.transform(region); 6253 } 6254 6255 //----------------------------inline_electronicCodeBook_AESCrypt_predicate---------------------------- 6256 // Return node representing slow path of predicate check. 6257 // the pseudo code we want to emulate with this predicate is: 6258 // for encryption: 6259 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6260 // for decryption: 6261 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6262 // note cipher==plain is more conservative than the original java code but that's OK 6263 // 6264 Node* LibraryCallKit::inline_electronicCodeBook_AESCrypt_predicate(bool decrypting) { 6265 // The receiver was checked for NULL already. 6266 Node* objECB = argument(0); 6267 6268 // Load embeddedCipher field of ElectronicCodeBook object. 6269 Node* embeddedCipherObj = load_field_from_object(objECB, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6270 6271 // get AESCrypt klass for instanceOf check 6272 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6273 // will have same classloader as ElectronicCodeBook object 6274 const TypeInstPtr* tinst = _gvn.type(objECB)->isa_instptr(); 6275 assert(tinst != NULL, "ECBobj is null"); 6276 assert(tinst->klass()->is_loaded(), "ECBobj is not loaded"); 6277 6278 // we want to do an instanceof comparison against the AESCrypt class 6279 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6280 if (!klass_AESCrypt->is_loaded()) { 6281 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6282 Node* ctrl = control(); 6283 set_control(top()); // no regular fast path 6284 return ctrl; 6285 } 6286 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6287 6288 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6289 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6290 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6291 6292 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6293 6294 // for encryption, we are done 6295 if (!decrypting) 6296 return instof_false; // even if it is NULL 6297 6298 // for decryption, we need to add a further check to avoid 6299 // taking the intrinsic path when cipher and plain are the same 6300 // see the original java code for why. 6301 RegionNode* region = new RegionNode(3); 6302 region->init_req(1, instof_false); 6303 Node* src = argument(1); 6304 Node* dest = argument(4); 6305 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6306 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6307 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6308 region->init_req(2, src_dest_conjoint); 6309 6310 record_for_igvn(region); 6311 return _gvn.transform(region); 6312 } 6313 6314 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6315 // Return node representing slow path of predicate check. 6316 // the pseudo code we want to emulate with this predicate is: 6317 // for encryption: 6318 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6319 // for decryption: 6320 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6321 // note cipher==plain is more conservative than the original java code but that's OK 6322 // 6323 6324 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6325 // The receiver was checked for NULL already. 6326 Node* objCTR = argument(0); 6327 6328 // Load embeddedCipher field of CipherBlockChaining object. 6329 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6330 6331 // get AESCrypt klass for instanceOf check 6332 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6333 // will have same classloader as CipherBlockChaining object 6334 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6335 assert(tinst != NULL, "CTRobj is null"); 6336 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6337 6338 // we want to do an instanceof comparison against the AESCrypt class 6339 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6340 if (!klass_AESCrypt->is_loaded()) { 6341 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6342 Node* ctrl = control(); 6343 set_control(top()); // no regular fast path 6344 return ctrl; 6345 } 6346 6347 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6348 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6349 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6350 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6351 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6352 6353 return instof_false; // even if it is NULL 6354 } 6355 6356 //------------------------------inline_ghash_processBlocks 6357 bool LibraryCallKit::inline_ghash_processBlocks() { 6358 address stubAddr; 6359 const char *stubName; 6360 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6361 6362 stubAddr = StubRoutines::ghash_processBlocks(); 6363 stubName = "ghash_processBlocks"; 6364 6365 Node* data = argument(0); 6366 Node* offset = argument(1); 6367 Node* len = argument(2); 6368 Node* state = argument(3); 6369 Node* subkeyH = argument(4); 6370 6371 state = must_be_not_null(state, true); 6372 subkeyH = must_be_not_null(subkeyH, true); 6373 data = must_be_not_null(data, true); 6374 6375 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6376 assert(state_start, "state is NULL"); 6377 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6378 assert(subkeyH_start, "subkeyH is NULL"); 6379 Node* data_start = array_element_address(data, offset, T_BYTE); 6380 assert(data_start, "data is NULL"); 6381 6382 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6383 OptoRuntime::ghash_processBlocks_Type(), 6384 stubAddr, stubName, TypePtr::BOTTOM, 6385 state_start, subkeyH_start, data_start, len); 6386 return true; 6387 } 6388 6389 bool LibraryCallKit::inline_base64_encodeBlock() { 6390 address stubAddr; 6391 const char *stubName; 6392 assert(UseBASE64Intrinsics, "need Base64 intrinsics support"); 6393 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters"); 6394 stubAddr = StubRoutines::base64_encodeBlock(); 6395 stubName = "encodeBlock"; 6396 6397 if (!stubAddr) return false; 6398 Node* base64obj = argument(0); 6399 Node* src = argument(1); 6400 Node* offset = argument(2); 6401 Node* len = argument(3); 6402 Node* dest = argument(4); 6403 Node* dp = argument(5); 6404 Node* isURL = argument(6); 6405 6406 src = must_be_not_null(src, true); 6407 dest = must_be_not_null(dest, true); 6408 6409 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 6410 assert(src_start, "source array is NULL"); 6411 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE); 6412 assert(dest_start, "destination array is NULL"); 6413 6414 Node* base64 = make_runtime_call(RC_LEAF, 6415 OptoRuntime::base64_encodeBlock_Type(), 6416 stubAddr, stubName, TypePtr::BOTTOM, 6417 src_start, offset, len, dest_start, dp, isURL); 6418 return true; 6419 } 6420 6421 bool LibraryCallKit::inline_base64_decodeBlock() { 6422 address stubAddr; 6423 const char *stubName; 6424 assert(UseBASE64Intrinsics, "need Base64 intrinsics support"); 6425 assert(callee()->signature()->size() == 5, "base64_decodeBlock has 5 parameters"); 6426 stubAddr = StubRoutines::base64_decodeBlock(); 6427 stubName = "decodeBlock"; 6428 6429 if (!stubAddr) return false; 6430 Node* base64obj = argument(0); 6431 Node* src = argument(1); 6432 Node* offset = argument(2); 6433 Node* len = argument(3); 6434 Node* dest = argument(4); 6435 Node* isURL = argument(5); 6436 6437 src = must_be_not_null(src, true); 6438 dest = must_be_not_null(dest, true); 6439 6440 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 6441 assert(src_start, "source array is NULL"); 6442 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE); 6443 assert(dest_start, "destination array is NULL"); 6444 6445 Node* call = make_runtime_call(RC_LEAF, 6446 OptoRuntime::base64_decodeBlock_Type(), 6447 stubAddr, stubName, TypePtr::BOTTOM, 6448 src_start, offset, len, dest_start, isURL); 6449 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6450 set_result(result); 6451 return true; 6452 } 6453 6454 //------------------------------inline_digestBase_implCompress----------------------- 6455 // 6456 // Calculate MD5 for single-block byte[] array. 6457 // void com.sun.security.provider.MD5.implCompress(byte[] buf, int ofs) 6458 // 6459 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6460 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6461 // 6462 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6463 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6464 // 6465 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6466 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6467 // 6468 bool LibraryCallKit::inline_digestBase_implCompress(vmIntrinsics::ID id) { 6469 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6470 6471 Node* digestBase_obj = argument(0); 6472 Node* src = argument(1); // type oop 6473 Node* ofs = argument(2); // type int 6474 6475 const Type* src_type = src->Value(&_gvn); 6476 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6477 if (top_src == NULL || top_src->klass() == NULL) { 6478 // failed array check 6479 return false; 6480 } 6481 // Figure out the size and type of the elements we will be copying. 6482 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6483 if (src_elem != T_BYTE) { 6484 return false; 6485 } 6486 // 'src_start' points to src array + offset 6487 src = must_be_not_null(src, true); 6488 Node* src_start = array_element_address(src, ofs, src_elem); 6489 Node* state = NULL; 6490 address stubAddr; 6491 const char *stubName; 6492 6493 switch(id) { 6494 case vmIntrinsics::_md5_implCompress: 6495 assert(UseMD5Intrinsics, "need MD5 instruction support"); 6496 state = get_state_from_digest_object(digestBase_obj); 6497 stubAddr = StubRoutines::md5_implCompress(); 6498 stubName = "md5_implCompress"; 6499 break; 6500 case vmIntrinsics::_sha_implCompress: 6501 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6502 state = get_state_from_digest_object(digestBase_obj); 6503 stubAddr = StubRoutines::sha1_implCompress(); 6504 stubName = "sha1_implCompress"; 6505 break; 6506 case vmIntrinsics::_sha2_implCompress: 6507 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6508 state = get_state_from_digest_object(digestBase_obj); 6509 stubAddr = StubRoutines::sha256_implCompress(); 6510 stubName = "sha256_implCompress"; 6511 break; 6512 case vmIntrinsics::_sha5_implCompress: 6513 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6514 state = get_long_state_from_digest_object(digestBase_obj); 6515 stubAddr = StubRoutines::sha512_implCompress(); 6516 stubName = "sha512_implCompress"; 6517 break; 6518 default: 6519 fatal_unexpected_iid(id); 6520 return false; 6521 } 6522 if (state == NULL) return false; 6523 6524 assert(stubAddr != NULL, "Stub is generated"); 6525 if (stubAddr == NULL) return false; 6526 6527 // Call the stub. 6528 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::digestBase_implCompress_Type(), 6529 stubAddr, stubName, TypePtr::BOTTOM, 6530 src_start, state); 6531 6532 return true; 6533 } 6534 6535 //------------------------------inline_digestBase_implCompressMB----------------------- 6536 // 6537 // Calculate MD5/SHA/SHA2/SHA5 for multi-block byte[] array. 6538 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6539 // 6540 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6541 assert(UseMD5Intrinsics || UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6542 "need MD5/SHA1/SHA256/SHA512 instruction support"); 6543 assert((uint)predicate < 4, "sanity"); 6544 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6545 6546 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6547 Node* src = argument(1); // byte[] array 6548 Node* ofs = argument(2); // type int 6549 Node* limit = argument(3); // type int 6550 6551 const Type* src_type = src->Value(&_gvn); 6552 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6553 if (top_src == NULL || top_src->klass() == NULL) { 6554 // failed array check 6555 return false; 6556 } 6557 // Figure out the size and type of the elements we will be copying. 6558 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6559 if (src_elem != T_BYTE) { 6560 return false; 6561 } 6562 // 'src_start' points to src array + offset 6563 src = must_be_not_null(src, false); 6564 Node* src_start = array_element_address(src, ofs, src_elem); 6565 6566 const char* klass_digestBase_name = NULL; 6567 const char* stub_name = NULL; 6568 address stub_addr = NULL; 6569 bool long_state = false; 6570 6571 switch (predicate) { 6572 case 0: 6573 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_md5_implCompress)) { 6574 klass_digestBase_name = "sun/security/provider/MD5"; 6575 stub_name = "md5_implCompressMB"; 6576 stub_addr = StubRoutines::md5_implCompressMB(); 6577 } 6578 break; 6579 case 1: 6580 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_sha_implCompress)) { 6581 klass_digestBase_name = "sun/security/provider/SHA"; 6582 stub_name = "sha1_implCompressMB"; 6583 stub_addr = StubRoutines::sha1_implCompressMB(); 6584 } 6585 break; 6586 case 2: 6587 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_sha2_implCompress)) { 6588 klass_digestBase_name = "sun/security/provider/SHA2"; 6589 stub_name = "sha256_implCompressMB"; 6590 stub_addr = StubRoutines::sha256_implCompressMB(); 6591 } 6592 break; 6593 case 3: 6594 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_sha5_implCompress)) { 6595 klass_digestBase_name = "sun/security/provider/SHA5"; 6596 stub_name = "sha512_implCompressMB"; 6597 stub_addr = StubRoutines::sha512_implCompressMB(); 6598 long_state = true; 6599 } 6600 break; 6601 default: 6602 fatal("unknown DigestBase intrinsic predicate: %d", predicate); 6603 } 6604 if (klass_digestBase_name != NULL) { 6605 assert(stub_addr != NULL, "Stub is generated"); 6606 if (stub_addr == NULL) return false; 6607 6608 // get DigestBase klass to lookup for SHA klass 6609 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6610 assert(tinst != NULL, "digestBase_obj is not instance???"); 6611 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6612 6613 ciKlass* klass_digestBase = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_digestBase_name)); 6614 assert(klass_digestBase->is_loaded(), "predicate checks that this class is loaded"); 6615 ciInstanceKlass* instklass_digestBase = klass_digestBase->as_instance_klass(); 6616 return inline_digestBase_implCompressMB(digestBase_obj, instklass_digestBase, long_state, stub_addr, stub_name, src_start, ofs, limit); 6617 } 6618 return false; 6619 } 6620 6621 //------------------------------inline_digestBase_implCompressMB----------------------- 6622 bool LibraryCallKit::inline_digestBase_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_digestBase, 6623 bool long_state, address stubAddr, const char *stubName, 6624 Node* src_start, Node* ofs, Node* limit) { 6625 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_digestBase); 6626 const TypeOopPtr* xtype = aklass->as_instance_type(); 6627 Node* digest_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6628 digest_obj = _gvn.transform(digest_obj); 6629 6630 Node* state; 6631 if (long_state) { 6632 state = get_long_state_from_digest_object(digest_obj); 6633 } else { 6634 state = get_state_from_digest_object(digest_obj); 6635 } 6636 if (state == NULL) return false; 6637 6638 // Call the stub. 6639 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6640 OptoRuntime::digestBase_implCompressMB_Type(), 6641 stubAddr, stubName, TypePtr::BOTTOM, 6642 src_start, state, ofs, limit); 6643 // return ofs (int) 6644 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6645 set_result(result); 6646 6647 return true; 6648 } 6649 6650 //------------------------------get_state_from_digest_object----------------------- 6651 Node * LibraryCallKit::get_state_from_digest_object(Node *digest_object) { 6652 Node* digest_state = load_field_from_object(digest_object, "state", "[I", /*is_exact*/ false); 6653 assert (digest_state != NULL, "wrong version of sun.security.provider.MD5/SHA/SHA2"); 6654 if (digest_state == NULL) return (Node *) NULL; 6655 6656 // now have the array, need to get the start address of the state array 6657 Node* state = array_element_address(digest_state, intcon(0), T_INT); 6658 return state; 6659 } 6660 6661 //------------------------------get_long_state_from_digest_object----------------------- 6662 Node * LibraryCallKit::get_long_state_from_digest_object(Node *digest_object) { 6663 Node* digest_state = load_field_from_object(digest_object, "state", "[J", /*is_exact*/ false); 6664 assert (digest_state != NULL, "wrong version of sun.security.provider.SHA5"); 6665 if (digest_state == NULL) return (Node *) NULL; 6666 6667 // now have the array, need to get the start address of the state array 6668 Node* state = array_element_address(digest_state, intcon(0), T_LONG); 6669 return state; 6670 } 6671 6672 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6673 // Return node representing slow path of predicate check. 6674 // the pseudo code we want to emulate with this predicate is: 6675 // if (digestBaseObj instanceof MD5/SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6676 // 6677 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6678 assert(UseMD5Intrinsics || UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6679 "need MD5/SHA1/SHA256/SHA512 instruction support"); 6680 assert((uint)predicate < 4, "sanity"); 6681 6682 // The receiver was checked for NULL already. 6683 Node* digestBaseObj = argument(0); 6684 6685 // get DigestBase klass for instanceOf check 6686 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6687 assert(tinst != NULL, "digestBaseObj is null"); 6688 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6689 6690 const char* klass_name = NULL; 6691 switch (predicate) { 6692 case 0: 6693 if (UseMD5Intrinsics) { 6694 // we want to do an instanceof comparison against the MD5 class 6695 klass_name = "sun/security/provider/MD5"; 6696 } 6697 break; 6698 case 1: 6699 if (UseSHA1Intrinsics) { 6700 // we want to do an instanceof comparison against the SHA class 6701 klass_name = "sun/security/provider/SHA"; 6702 } 6703 break; 6704 case 2: 6705 if (UseSHA256Intrinsics) { 6706 // we want to do an instanceof comparison against the SHA2 class 6707 klass_name = "sun/security/provider/SHA2"; 6708 } 6709 break; 6710 case 3: 6711 if (UseSHA512Intrinsics) { 6712 // we want to do an instanceof comparison against the SHA5 class 6713 klass_name = "sun/security/provider/SHA5"; 6714 } 6715 break; 6716 default: 6717 fatal("unknown SHA intrinsic predicate: %d", predicate); 6718 } 6719 6720 ciKlass* klass = NULL; 6721 if (klass_name != NULL) { 6722 klass = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_name)); 6723 } 6724 if ((klass == NULL) || !klass->is_loaded()) { 6725 // if none of MD5/SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6726 Node* ctrl = control(); 6727 set_control(top()); // no intrinsic path 6728 return ctrl; 6729 } 6730 ciInstanceKlass* instklass = klass->as_instance_klass(); 6731 6732 Node* instof = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass))); 6733 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6734 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6735 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6736 6737 return instof_false; // even if it is NULL 6738 } 6739 6740 //-------------inline_fma----------------------------------- 6741 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) { 6742 Node *a = NULL; 6743 Node *b = NULL; 6744 Node *c = NULL; 6745 Node* result = NULL; 6746 switch (id) { 6747 case vmIntrinsics::_fmaD: 6748 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each."); 6749 // no receiver since it is static method 6750 a = round_double_node(argument(0)); 6751 b = round_double_node(argument(2)); 6752 c = round_double_node(argument(4)); 6753 result = _gvn.transform(new FmaDNode(control(), a, b, c)); 6754 break; 6755 case vmIntrinsics::_fmaF: 6756 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each."); 6757 a = argument(0); 6758 b = argument(1); 6759 c = argument(2); 6760 result = _gvn.transform(new FmaFNode(control(), a, b, c)); 6761 break; 6762 default: 6763 fatal_unexpected_iid(id); break; 6764 } 6765 set_result(result); 6766 return true; 6767 } 6768 6769 bool LibraryCallKit::inline_character_compare(vmIntrinsics::ID id) { 6770 // argument(0) is receiver 6771 Node* codePoint = argument(1); 6772 Node* n = NULL; 6773 6774 switch (id) { 6775 case vmIntrinsics::_isDigit : 6776 n = new DigitNode(control(), codePoint); 6777 break; 6778 case vmIntrinsics::_isLowerCase : 6779 n = new LowerCaseNode(control(), codePoint); 6780 break; 6781 case vmIntrinsics::_isUpperCase : 6782 n = new UpperCaseNode(control(), codePoint); 6783 break; 6784 case vmIntrinsics::_isWhitespace : 6785 n = new WhitespaceNode(control(), codePoint); 6786 break; 6787 default: 6788 fatal_unexpected_iid(id); 6789 } 6790 6791 set_result(_gvn.transform(n)); 6792 return true; 6793 } 6794 6795 //------------------------------inline_fp_min_max------------------------------ 6796 bool LibraryCallKit::inline_fp_min_max(vmIntrinsics::ID id) { 6797 /* DISABLED BECAUSE METHOD DATA ISN'T COLLECTED PER CALL-SITE, SEE JDK-8015416. 6798 6799 // The intrinsic should be used only when the API branches aren't predictable, 6800 // the last one performing the most important comparison. The following heuristic 6801 // uses the branch statistics to eventually bail out if necessary. 6802 6803 ciMethodData *md = callee()->method_data(); 6804 6805 if ( md != NULL && md->is_mature() && md->invocation_count() > 0 ) { 6806 ciCallProfile cp = caller()->call_profile_at_bci(bci()); 6807 6808 if ( ((double)cp.count()) / ((double)md->invocation_count()) < 0.8 ) { 6809 // Bail out if the call-site didn't contribute enough to the statistics. 6810 return false; 6811 } 6812 6813 uint taken = 0, not_taken = 0; 6814 6815 for (ciProfileData *p = md->first_data(); md->is_valid(p); p = md->next_data(p)) { 6816 if (p->is_BranchData()) { 6817 taken = ((ciBranchData*)p)->taken(); 6818 not_taken = ((ciBranchData*)p)->not_taken(); 6819 } 6820 } 6821 6822 double balance = (((double)taken) - ((double)not_taken)) / ((double)md->invocation_count()); 6823 balance = balance < 0 ? -balance : balance; 6824 if ( balance > 0.2 ) { 6825 // Bail out if the most important branch is predictable enough. 6826 return false; 6827 } 6828 } 6829 */ 6830 6831 Node *a = NULL; 6832 Node *b = NULL; 6833 Node *n = NULL; 6834 switch (id) { 6835 case vmIntrinsics::_maxF: 6836 case vmIntrinsics::_minF: 6837 assert(callee()->signature()->size() == 2, "minF/maxF has 2 parameters of size 1 each."); 6838 a = argument(0); 6839 b = argument(1); 6840 break; 6841 case vmIntrinsics::_maxD: 6842 case vmIntrinsics::_minD: 6843 assert(callee()->signature()->size() == 4, "minD/maxD has 2 parameters of size 2 each."); 6844 a = round_double_node(argument(0)); 6845 b = round_double_node(argument(2)); 6846 break; 6847 default: 6848 fatal_unexpected_iid(id); 6849 break; 6850 } 6851 switch (id) { 6852 case vmIntrinsics::_maxF: n = new MaxFNode(a, b); break; 6853 case vmIntrinsics::_minF: n = new MinFNode(a, b); break; 6854 case vmIntrinsics::_maxD: n = new MaxDNode(a, b); break; 6855 case vmIntrinsics::_minD: n = new MinDNode(a, b); break; 6856 default: fatal_unexpected_iid(id); break; 6857 } 6858 set_result(_gvn.transform(n)); 6859 return true; 6860 } 6861 6862 bool LibraryCallKit::inline_profileBoolean() { 6863 Node* counts = argument(1); 6864 const TypeAryPtr* ary = NULL; 6865 ciArray* aobj = NULL; 6866 if (counts->is_Con() 6867 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6868 && (aobj = ary->const_oop()->as_array()) != NULL 6869 && (aobj->length() == 2)) { 6870 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6871 jint false_cnt = aobj->element_value(0).as_int(); 6872 jint true_cnt = aobj->element_value(1).as_int(); 6873 6874 if (C->log() != NULL) { 6875 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6876 false_cnt, true_cnt); 6877 } 6878 6879 if (false_cnt + true_cnt == 0) { 6880 // According to profile, never executed. 6881 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6882 Deoptimization::Action_reinterpret); 6883 return true; 6884 } 6885 6886 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6887 // is a number of each value occurrences. 6888 Node* result = argument(0); 6889 if (false_cnt == 0 || true_cnt == 0) { 6890 // According to profile, one value has been never seen. 6891 int expected_val = (false_cnt == 0) ? 1 : 0; 6892 6893 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6894 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6895 6896 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6897 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6898 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6899 6900 { // Slow path: uncommon trap for never seen value and then reexecute 6901 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6902 // the value has been seen at least once. 6903 PreserveJVMState pjvms(this); 6904 PreserveReexecuteState preexecs(this); 6905 jvms()->set_should_reexecute(true); 6906 6907 set_control(slow_path); 6908 set_i_o(i_o()); 6909 6910 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6911 Deoptimization::Action_reinterpret); 6912 } 6913 // The guard for never seen value enables sharpening of the result and 6914 // returning a constant. It allows to eliminate branches on the same value 6915 // later on. 6916 set_control(fast_path); 6917 result = intcon(expected_val); 6918 } 6919 // Stop profiling. 6920 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6921 // By replacing method body with profile data (represented as ProfileBooleanNode 6922 // on IR level) we effectively disable profiling. 6923 // It enables full speed execution once optimized code is generated. 6924 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6925 C->record_for_igvn(profile); 6926 set_result(profile); 6927 return true; 6928 } else { 6929 // Continue profiling. 6930 // Profile data isn't available at the moment. So, execute method's bytecode version. 6931 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6932 // is compiled and counters aren't available since corresponding MethodHandle 6933 // isn't a compile-time constant. 6934 return false; 6935 } 6936 } 6937 6938 bool LibraryCallKit::inline_isCompileConstant() { 6939 Node* n = argument(0); 6940 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6941 return true; 6942 }