1 /* 2 * Copyright (c) 1999, 2018, 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/objArrayKlass.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/arraycopynode.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/castnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/convertnode.hpp" 43 #include "opto/countbitsnode.hpp" 44 #include "opto/intrinsicnode.hpp" 45 #include "opto/idealKit.hpp" 46 #include "opto/mathexactnode.hpp" 47 #include "opto/movenode.hpp" 48 #include "opto/mulnode.hpp" 49 #include "opto/narrowptrnode.hpp" 50 #include "opto/opaquenode.hpp" 51 #include "opto/parse.hpp" 52 #include "opto/runtime.hpp" 53 #include "opto/rootnode.hpp" 54 #include "opto/subnode.hpp" 55 #include "opto/vectornode.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 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_in_bytes(); 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_in_bytes(); 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_array_guard(Node* kls, RegionNode* region) { 190 return generate_array_guard_common(kls, region, false, false); 191 } 192 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 193 return generate_array_guard_common(kls, region, false, true); 194 } 195 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 196 return generate_array_guard_common(kls, region, true, false); 197 } 198 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 199 return generate_array_guard_common(kls, region, true, true); 200 } 201 Node* generate_array_guard_common(Node* kls, RegionNode* region, 202 bool obj_array, bool not_array); 203 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 204 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 205 bool is_virtual = false, bool is_static = false); 206 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 207 return generate_method_call(method_id, false, true); 208 } 209 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 210 return generate_method_call(method_id, true, false); 211 } 212 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 213 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 214 215 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 216 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 217 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 218 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 219 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 220 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); 221 bool inline_string_indexOfChar(); 222 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 223 bool inline_string_toBytesU(); 224 bool inline_string_getCharsU(); 225 bool inline_string_copy(bool compress); 226 bool inline_string_char_access(bool is_store); 227 Node* round_double_node(Node* n); 228 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 229 bool inline_math_native(vmIntrinsics::ID id); 230 bool inline_math(vmIntrinsics::ID id); 231 template <typename OverflowOp> 232 bool inline_math_overflow(Node* arg1, Node* arg2); 233 void inline_math_mathExact(Node* math, Node* test); 234 bool inline_math_addExactI(bool is_increment); 235 bool inline_math_addExactL(bool is_increment); 236 bool inline_math_multiplyExactI(); 237 bool inline_math_multiplyExactL(); 238 bool inline_math_multiplyHigh(); 239 bool inline_math_negateExactI(); 240 bool inline_math_negateExactL(); 241 bool inline_math_subtractExactI(bool is_decrement); 242 bool inline_math_subtractExactL(bool is_decrement); 243 bool inline_min_max(vmIntrinsics::ID id); 244 bool inline_notify(vmIntrinsics::ID id); 245 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 246 // This returns Type::AnyPtr, RawPtr, or OopPtr. 247 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type); 248 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false); 249 250 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind; 251 DecoratorSet mo_decorator_for_access_kind(AccessKind kind); 252 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned); 253 static bool klass_needs_init_guard(Node* kls); 254 bool inline_unsafe_allocate(); 255 bool inline_unsafe_newArray(bool uninitialized); 256 bool inline_unsafe_copyMemory(); 257 bool inline_native_currentThread(); 258 259 bool inline_native_time_funcs(address method, const char* funcName); 260 #ifdef JFR_HAVE_INTRINSICS 261 bool inline_native_classID(); 262 bool inline_native_getEventWriter(); 263 #endif 264 bool inline_native_isInterrupted(); 265 bool inline_native_Class_query(vmIntrinsics::ID id); 266 bool inline_native_subtype_check(); 267 bool inline_native_getLength(); 268 bool inline_array_copyOf(bool is_copyOfRange); 269 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 270 bool inline_preconditions_checkIndex(); 271 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array); 272 bool inline_native_clone(bool is_virtual); 273 bool inline_native_Reflection_getCallerClass(); 274 // Helper function for inlining native object hash method 275 bool inline_native_hashcode(bool is_virtual, bool is_static); 276 bool inline_native_getClass(); 277 278 // Helper functions for inlining arraycopy 279 bool inline_arraycopy(); 280 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 281 RegionNode* slow_region); 282 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 283 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp, 284 uint new_idx); 285 286 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind; 287 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind); 288 bool inline_unsafe_fence(vmIntrinsics::ID id); 289 bool inline_onspinwait(); 290 bool inline_fp_conversions(vmIntrinsics::ID id); 291 bool inline_number_methods(vmIntrinsics::ID id); 292 bool inline_reference_get(); 293 bool inline_Class_cast(); 294 bool inline_aescrypt_Block(vmIntrinsics::ID id); 295 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 296 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 297 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 298 Node* inline_counterMode_AESCrypt_predicate(); 299 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 300 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 301 bool inline_ghash_processBlocks(); 302 bool inline_base64_encodeBlock(); 303 bool inline_sha_implCompress(vmIntrinsics::ID id); 304 bool inline_digestBase_implCompressMB(int predicate); 305 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 306 bool long_state, address stubAddr, const char *stubName, 307 Node* src_start, Node* ofs, Node* limit); 308 Node* get_state_from_sha_object(Node *sha_object); 309 Node* get_state_from_sha5_object(Node *sha_object); 310 Node* inline_digestBase_implCompressMB_predicate(int predicate); 311 bool inline_encodeISOArray(); 312 bool inline_updateCRC32(); 313 bool inline_updateBytesCRC32(); 314 bool inline_updateByteBufferCRC32(); 315 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 316 bool inline_updateBytesCRC32C(); 317 bool inline_updateDirectByteBufferCRC32C(); 318 bool inline_updateBytesAdler32(); 319 bool inline_updateByteBufferAdler32(); 320 bool inline_multiplyToLen(); 321 bool inline_hasNegatives(); 322 bool inline_squareToLen(); 323 bool inline_mulAdd(); 324 bool inline_montgomeryMultiply(); 325 bool inline_montgomerySquare(); 326 bool inline_vectorizedMismatch(); 327 bool inline_fma(vmIntrinsics::ID id); 328 329 bool inline_profileBoolean(); 330 bool inline_isCompileConstant(); 331 332 // Vector API support 333 bool inline_vector_nary_operation(int n); 334 bool inline_vector_broadcast_coerced(); 335 bool inline_vector_mem_operation(bool is_store); 336 bool inline_vector_reduction(); 337 bool inline_vector_test(); 338 bool inline_vector_blend(); 339 bool inline_vector_compare(); 340 bool inline_vector_broadcast_int(); 341 bool inline_vector_cast_reinterpret(bool is_cast); 342 bool inline_vector_extract(); 343 bool inline_vector_insert(); 344 Node* box_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem); 345 Node* unbox_vector(Node* in, const TypeInstPtr* vbox_type, BasicType bt, int num_elem); 346 Node* shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem); 347 Node* gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2); 348 void set_vector_result(Node* result, bool set_res = true); 349 350 void clear_upper_avx() { 351 #ifdef X86 352 if (UseAVX >= 2) { 353 C->set_clear_upper_avx(true); 354 } 355 #endif 356 } 357 }; 358 359 //---------------------------make_vm_intrinsic---------------------------- 360 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 361 vmIntrinsics::ID id = m->intrinsic_id(); 362 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 363 364 if (!m->is_loaded()) { 365 // Do not attempt to inline unloaded methods. 366 return NULL; 367 } 368 369 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 370 bool is_available = false; 371 372 { 373 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 374 // the compiler must transition to '_thread_in_vm' state because both 375 // methods access VM-internal data. 376 VM_ENTRY_MARK; 377 methodHandle mh(THREAD, m->get_Method()); 378 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) && 379 !C->directive()->is_intrinsic_disabled(mh) && 380 !vmIntrinsics::is_disabled_by_flags(mh); 381 382 } 383 384 if (is_available) { 385 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 386 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 387 return new LibraryIntrinsic(m, is_virtual, 388 vmIntrinsics::predicates_needed(id), 389 vmIntrinsics::does_virtual_dispatch(id), 390 (vmIntrinsics::ID) id); 391 } else { 392 return NULL; 393 } 394 } 395 396 //----------------------register_library_intrinsics----------------------- 397 // Initialize this file's data structures, for each Compile instance. 398 void Compile::register_library_intrinsics() { 399 // Nothing to do here. 400 } 401 402 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 403 LibraryCallKit kit(jvms, this); 404 Compile* C = kit.C; 405 int nodes = C->unique(); 406 #ifndef PRODUCT 407 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 408 char buf[1000]; 409 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 410 tty->print_cr("Intrinsic %s", str); 411 } 412 #endif 413 ciMethod* callee = kit.callee(); 414 const int bci = kit.bci(); 415 416 // Try to inline the intrinsic. 417 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 418 kit.try_to_inline(_last_predicate)) { 419 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)" 420 : "(intrinsic)"; 421 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 422 if (C->print_intrinsics() || C->print_inlining()) { 423 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 424 } 425 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 426 if (C->log()) { 427 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 428 vmIntrinsics::name_at(intrinsic_id()), 429 (is_virtual() ? " virtual='1'" : ""), 430 C->unique() - nodes); 431 } 432 // Push the result from the inlined method onto the stack. 433 kit.push_result(); 434 C->print_inlining_update(this); 435 return kit.transfer_exceptions_into_jvms(); 436 } 437 438 // The intrinsic bailed out 439 if (jvms->has_method()) { 440 // Not a root compile. 441 const char* msg; 442 if (callee->intrinsic_candidate()) { 443 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 444 } else { 445 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 446 : "failed to inline (intrinsic), method not annotated"; 447 } 448 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg); 449 if (C->print_intrinsics() || C->print_inlining()) { 450 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 451 } 452 } else { 453 // Root compile 454 ResourceMark rm; 455 stringStream msg_stream; 456 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 457 vmIntrinsics::name_at(intrinsic_id()), 458 is_virtual() ? " (virtual)" : "", bci); 459 const char *msg = msg_stream.as_string(); 460 log_debug(jit, inlining)("%s", msg); 461 if (C->print_intrinsics() || C->print_inlining()) { 462 tty->print("%s", msg); 463 } 464 } 465 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 466 C->print_inlining_update(this); 467 return NULL; 468 } 469 470 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 471 LibraryCallKit kit(jvms, this); 472 Compile* C = kit.C; 473 int nodes = C->unique(); 474 _last_predicate = predicate; 475 #ifndef PRODUCT 476 assert(is_predicated() && predicate < predicates_count(), "sanity"); 477 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 478 char buf[1000]; 479 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 480 tty->print_cr("Predicate for intrinsic %s", str); 481 } 482 #endif 483 ciMethod* callee = kit.callee(); 484 const int bci = kit.bci(); 485 486 Node* slow_ctl = kit.try_to_predicate(predicate); 487 if (!kit.failing()) { 488 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)" 489 : "(intrinsic, predicate)"; 490 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg); 491 if (C->print_intrinsics() || C->print_inlining()) { 492 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg); 493 } 494 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 495 if (C->log()) { 496 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 497 vmIntrinsics::name_at(intrinsic_id()), 498 (is_virtual() ? " virtual='1'" : ""), 499 C->unique() - nodes); 500 } 501 return slow_ctl; // Could be NULL if the check folds. 502 } 503 504 // The intrinsic bailed out 505 if (jvms->has_method()) { 506 // Not a root compile. 507 const char* msg = "failed to generate predicate for intrinsic"; 508 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg); 509 if (C->print_intrinsics() || C->print_inlining()) { 510 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 511 } 512 } else { 513 // Root compile 514 ResourceMark rm; 515 stringStream msg_stream; 516 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in", 517 vmIntrinsics::name_at(intrinsic_id()), 518 is_virtual() ? " (virtual)" : "", bci); 519 const char *msg = msg_stream.as_string(); 520 log_debug(jit, inlining)("%s", msg); 521 if (C->print_intrinsics() || C->print_inlining()) { 522 C->print_inlining_stream()->print("%s", msg); 523 } 524 } 525 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 526 return NULL; 527 } 528 529 bool LibraryCallKit::try_to_inline(int predicate) { 530 // Handle symbolic names for otherwise undistinguished boolean switches: 531 const bool is_store = true; 532 const bool is_compress = true; 533 const bool is_static = true; 534 const bool is_volatile = true; 535 536 if (!jvms()->has_method()) { 537 // Root JVMState has a null method. 538 assert(map()->memory()->Opcode() == Op_Parm, ""); 539 // Insert the memory aliasing node 540 set_all_memory(reset_memory()); 541 } 542 assert(merged_memory(), ""); 543 544 switch (intrinsic_id()) { 545 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 546 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 547 case vmIntrinsics::_getClass: return inline_native_getClass(); 548 549 case vmIntrinsics::_dsin: 550 case vmIntrinsics::_dcos: 551 case vmIntrinsics::_dtan: 552 case vmIntrinsics::_dabs: 553 case vmIntrinsics::_datan2: 554 case vmIntrinsics::_dsqrt: 555 case vmIntrinsics::_dexp: 556 case vmIntrinsics::_dlog: 557 case vmIntrinsics::_dlog10: 558 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 559 560 case vmIntrinsics::_min: 561 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 562 563 case vmIntrinsics::_notify: 564 case vmIntrinsics::_notifyAll: 565 if (ObjectMonitor::Knob_InlineNotify) { 566 return inline_notify(intrinsic_id()); 567 } 568 return false; 569 570 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 571 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 572 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 573 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 574 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 575 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 576 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 577 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 578 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh(); 579 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 580 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 581 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 582 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 583 584 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 585 586 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 587 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 588 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 589 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 590 591 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 592 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 593 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 594 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 595 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 596 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 597 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 598 599 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 600 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 601 602 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 603 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 604 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 605 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 606 607 case vmIntrinsics::_compressStringC: 608 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 609 case vmIntrinsics::_inflateStringC: 610 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 611 612 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 613 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 614 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 615 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 616 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 617 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 618 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 619 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 620 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 621 622 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 623 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 624 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 625 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 626 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 627 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 628 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 629 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 630 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 631 632 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 633 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 634 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 635 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 636 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 637 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 638 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 639 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 640 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 641 642 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 643 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 644 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 645 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 646 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 647 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 648 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 649 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 650 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 651 652 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 653 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 654 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 655 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 656 657 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 658 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 659 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 660 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 661 662 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 663 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 664 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 665 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 666 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 667 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 668 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 669 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 670 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 671 672 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 673 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 674 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 675 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 676 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 677 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 678 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 679 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 680 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 681 682 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 683 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 684 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 685 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 686 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 687 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 688 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 689 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 690 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 691 692 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 693 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 694 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 695 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 696 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 697 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 698 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 699 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 700 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 701 702 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 703 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); 704 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); 705 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 706 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 707 708 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 709 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 710 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 711 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 712 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); 713 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); 714 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); 715 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); 716 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); 717 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); 718 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); 719 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); 720 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 721 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 722 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 723 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 724 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 725 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 726 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 727 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 728 729 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 730 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 731 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 732 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); 733 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); 734 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); 735 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); 736 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); 737 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); 738 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 739 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 740 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 741 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 742 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 743 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 744 745 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); 746 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); 747 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 748 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 749 750 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); 751 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); 752 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 753 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 754 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 755 756 case vmIntrinsics::_loadFence: 757 case vmIntrinsics::_storeFence: 758 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 759 760 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 761 762 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 763 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 764 765 #ifdef JFR_HAVE_INTRINSICS 766 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime"); 767 case vmIntrinsics::_getClassId: return inline_native_classID(); 768 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter(); 769 #endif 770 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 771 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 772 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 773 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 774 case vmIntrinsics::_getLength: return inline_native_getLength(); 775 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 776 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 777 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 778 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 779 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 780 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 781 782 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 783 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 784 785 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 786 787 case vmIntrinsics::_isInstance: 788 case vmIntrinsics::_getModifiers: 789 case vmIntrinsics::_isInterface: 790 case vmIntrinsics::_isArray: 791 case vmIntrinsics::_isPrimitive: 792 case vmIntrinsics::_getSuperclass: 793 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 794 795 case vmIntrinsics::_floatToRawIntBits: 796 case vmIntrinsics::_floatToIntBits: 797 case vmIntrinsics::_intBitsToFloat: 798 case vmIntrinsics::_doubleToRawLongBits: 799 case vmIntrinsics::_doubleToLongBits: 800 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 801 802 case vmIntrinsics::_numberOfLeadingZeros_i: 803 case vmIntrinsics::_numberOfLeadingZeros_l: 804 case vmIntrinsics::_numberOfTrailingZeros_i: 805 case vmIntrinsics::_numberOfTrailingZeros_l: 806 case vmIntrinsics::_bitCount_i: 807 case vmIntrinsics::_bitCount_l: 808 case vmIntrinsics::_reverseBytes_i: 809 case vmIntrinsics::_reverseBytes_l: 810 case vmIntrinsics::_reverseBytes_s: 811 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 812 813 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 814 815 case vmIntrinsics::_Reference_get: return inline_reference_get(); 816 817 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 818 819 case vmIntrinsics::_aescrypt_encryptBlock: 820 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 821 822 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 823 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 824 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 825 826 case vmIntrinsics::_counterMode_AESCrypt: 827 return inline_counterMode_AESCrypt(intrinsic_id()); 828 829 case vmIntrinsics::_sha_implCompress: 830 case vmIntrinsics::_sha2_implCompress: 831 case vmIntrinsics::_sha5_implCompress: 832 return inline_sha_implCompress(intrinsic_id()); 833 834 case vmIntrinsics::_digestBase_implCompressMB: 835 return inline_digestBase_implCompressMB(predicate); 836 837 case vmIntrinsics::_multiplyToLen: 838 return inline_multiplyToLen(); 839 840 case vmIntrinsics::_squareToLen: 841 return inline_squareToLen(); 842 843 case vmIntrinsics::_mulAdd: 844 return inline_mulAdd(); 845 846 case vmIntrinsics::_montgomeryMultiply: 847 return inline_montgomeryMultiply(); 848 case vmIntrinsics::_montgomerySquare: 849 return inline_montgomerySquare(); 850 851 case vmIntrinsics::_vectorizedMismatch: 852 return inline_vectorizedMismatch(); 853 854 case vmIntrinsics::_ghash_processBlocks: 855 return inline_ghash_processBlocks(); 856 case vmIntrinsics::_base64_encodeBlock: 857 return inline_base64_encodeBlock(); 858 859 case vmIntrinsics::_encodeISOArray: 860 case vmIntrinsics::_encodeByteISOArray: 861 return inline_encodeISOArray(); 862 863 case vmIntrinsics::_updateCRC32: 864 return inline_updateCRC32(); 865 case vmIntrinsics::_updateBytesCRC32: 866 return inline_updateBytesCRC32(); 867 case vmIntrinsics::_updateByteBufferCRC32: 868 return inline_updateByteBufferCRC32(); 869 870 case vmIntrinsics::_updateBytesCRC32C: 871 return inline_updateBytesCRC32C(); 872 case vmIntrinsics::_updateDirectByteBufferCRC32C: 873 return inline_updateDirectByteBufferCRC32C(); 874 875 case vmIntrinsics::_updateBytesAdler32: 876 return inline_updateBytesAdler32(); 877 case vmIntrinsics::_updateByteBufferAdler32: 878 return inline_updateByteBufferAdler32(); 879 880 case vmIntrinsics::_profileBoolean: 881 return inline_profileBoolean(); 882 case vmIntrinsics::_isCompileConstant: 883 return inline_isCompileConstant(); 884 885 case vmIntrinsics::_hasNegatives: 886 return inline_hasNegatives(); 887 888 case vmIntrinsics::_fmaD: 889 case vmIntrinsics::_fmaF: 890 return inline_fma(intrinsic_id()); 891 892 case vmIntrinsics::_VectorUnaryOp: 893 return inline_vector_nary_operation(1); 894 case vmIntrinsics::_VectorBinaryOp: 895 return inline_vector_nary_operation(2); 896 case vmIntrinsics::_VectorTernaryOp: 897 return inline_vector_nary_operation(3); 898 899 case vmIntrinsics::_VectorBroadcastCoerced: 900 return inline_vector_broadcast_coerced(); 901 case vmIntrinsics::_VectorLoadOp: 902 return inline_vector_mem_operation(/*is_store=*/false); 903 case vmIntrinsics::_VectorStoreOp: 904 return inline_vector_mem_operation(/*is_store=*/true); 905 case vmIntrinsics::_VectorReductionCoerced: 906 return inline_vector_reduction(); 907 case vmIntrinsics::_VectorTest: 908 return inline_vector_test(); 909 case vmIntrinsics::_VectorBlend: 910 return inline_vector_blend(); 911 case vmIntrinsics::_VectorCompare: 912 return inline_vector_compare(); 913 case vmIntrinsics::_VectorBroadcastInt: 914 return inline_vector_broadcast_int(); 915 case vmIntrinsics::_VectorReinterpret: 916 return inline_vector_cast_reinterpret(/*is_cast*/ false); 917 case vmIntrinsics::_VectorCast: 918 return inline_vector_cast_reinterpret(/*is_cast*/ true); 919 case vmIntrinsics::_VectorInsert: 920 return inline_vector_insert(); 921 case vmIntrinsics::_VectorExtract: 922 return inline_vector_extract(); 923 924 default: 925 // If you get here, it may be that someone has added a new intrinsic 926 // to the list in vmSymbols.hpp without implementing it here. 927 #ifndef PRODUCT 928 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 929 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 930 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 931 } 932 #endif 933 return false; 934 } 935 } 936 937 Node* LibraryCallKit::try_to_predicate(int predicate) { 938 if (!jvms()->has_method()) { 939 // Root JVMState has a null method. 940 assert(map()->memory()->Opcode() == Op_Parm, ""); 941 // Insert the memory aliasing node 942 set_all_memory(reset_memory()); 943 } 944 assert(merged_memory(), ""); 945 946 switch (intrinsic_id()) { 947 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 948 return inline_cipherBlockChaining_AESCrypt_predicate(false); 949 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 950 return inline_cipherBlockChaining_AESCrypt_predicate(true); 951 case vmIntrinsics::_counterMode_AESCrypt: 952 return inline_counterMode_AESCrypt_predicate(); 953 case vmIntrinsics::_digestBase_implCompressMB: 954 return inline_digestBase_implCompressMB_predicate(predicate); 955 956 default: 957 // If you get here, it may be that someone has added a new intrinsic 958 // to the list in vmSymbols.hpp without implementing it here. 959 #ifndef PRODUCT 960 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 961 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 962 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 963 } 964 #endif 965 Node* slow_ctl = control(); 966 set_control(top()); // No fast path instrinsic 967 return slow_ctl; 968 } 969 } 970 971 //------------------------------set_result------------------------------- 972 // Helper function for finishing intrinsics. 973 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 974 record_for_igvn(region); 975 set_control(_gvn.transform(region)); 976 set_result( _gvn.transform(value)); 977 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 978 } 979 980 //------------------------------generate_guard--------------------------- 981 // Helper function for generating guarded fast-slow graph structures. 982 // The given 'test', if true, guards a slow path. If the test fails 983 // then a fast path can be taken. (We generally hope it fails.) 984 // In all cases, GraphKit::control() is updated to the fast path. 985 // The returned value represents the control for the slow path. 986 // The return value is never 'top'; it is either a valid control 987 // or NULL if it is obvious that the slow path can never be taken. 988 // Also, if region and the slow control are not NULL, the slow edge 989 // is appended to the region. 990 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 991 if (stopped()) { 992 // Already short circuited. 993 return NULL; 994 } 995 996 // Build an if node and its projections. 997 // If test is true we take the slow path, which we assume is uncommon. 998 if (_gvn.type(test) == TypeInt::ZERO) { 999 // The slow branch is never taken. No need to build this guard. 1000 return NULL; 1001 } 1002 1003 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1004 1005 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1006 if (if_slow == top()) { 1007 // The slow branch is never taken. No need to build this guard. 1008 return NULL; 1009 } 1010 1011 if (region != NULL) 1012 region->add_req(if_slow); 1013 1014 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1015 set_control(if_fast); 1016 1017 return if_slow; 1018 } 1019 1020 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1021 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1022 } 1023 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1024 return generate_guard(test, region, PROB_FAIR); 1025 } 1026 1027 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1028 Node* *pos_index) { 1029 if (stopped()) 1030 return NULL; // already stopped 1031 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1032 return NULL; // index is already adequately typed 1033 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1034 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1035 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1036 if (is_neg != NULL && pos_index != NULL) { 1037 // Emulate effect of Parse::adjust_map_after_if. 1038 Node* ccast = new CastIINode(index, TypeInt::POS); 1039 ccast->set_req(0, control()); 1040 (*pos_index) = _gvn.transform(ccast); 1041 } 1042 return is_neg; 1043 } 1044 1045 // Make sure that 'position' is a valid limit index, in [0..length]. 1046 // There are two equivalent plans for checking this: 1047 // A. (offset + copyLength) unsigned<= arrayLength 1048 // B. offset <= (arrayLength - copyLength) 1049 // We require that all of the values above, except for the sum and 1050 // difference, are already known to be non-negative. 1051 // Plan A is robust in the face of overflow, if offset and copyLength 1052 // are both hugely positive. 1053 // 1054 // Plan B is less direct and intuitive, but it does not overflow at 1055 // all, since the difference of two non-negatives is always 1056 // representable. Whenever Java methods must perform the equivalent 1057 // check they generally use Plan B instead of Plan A. 1058 // For the moment we use Plan A. 1059 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1060 Node* subseq_length, 1061 Node* array_length, 1062 RegionNode* region) { 1063 if (stopped()) 1064 return NULL; // already stopped 1065 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1066 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1067 return NULL; // common case of whole-array copy 1068 Node* last = subseq_length; 1069 if (!zero_offset) // last += offset 1070 last = _gvn.transform(new AddINode(last, offset)); 1071 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1072 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1073 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1074 return is_over; 1075 } 1076 1077 // Emit range checks for the given String.value byte array 1078 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 1079 if (stopped()) { 1080 return; // already stopped 1081 } 1082 RegionNode* bailout = new RegionNode(1); 1083 record_for_igvn(bailout); 1084 if (char_count) { 1085 // Convert char count to byte count 1086 count = _gvn.transform(new LShiftINode(count, intcon(1))); 1087 } 1088 1089 // Offset and count must not be negative 1090 generate_negative_guard(offset, bailout); 1091 generate_negative_guard(count, bailout); 1092 // Offset + count must not exceed length of array 1093 generate_limit_guard(offset, count, load_array_length(array), bailout); 1094 1095 if (bailout->req() > 1) { 1096 PreserveJVMState pjvms(this); 1097 set_control(_gvn.transform(bailout)); 1098 uncommon_trap(Deoptimization::Reason_intrinsic, 1099 Deoptimization::Action_maybe_recompile); 1100 } 1101 } 1102 1103 //--------------------------generate_current_thread-------------------- 1104 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1105 ciKlass* thread_klass = env()->Thread_klass(); 1106 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1107 Node* thread = _gvn.transform(new ThreadLocalNode()); 1108 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1109 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1110 tls_output = thread; 1111 return threadObj; 1112 } 1113 1114 1115 //------------------------------make_string_method_node------------------------ 1116 // Helper method for String intrinsic functions. This version is called with 1117 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 1118 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 1119 // containing the lengths of str1 and str2. 1120 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 1121 Node* result = NULL; 1122 switch (opcode) { 1123 case Op_StrIndexOf: 1124 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 1125 str1_start, cnt1, str2_start, cnt2, ae); 1126 break; 1127 case Op_StrComp: 1128 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 1129 str1_start, cnt1, str2_start, cnt2, ae); 1130 break; 1131 case Op_StrEquals: 1132 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 1133 // Use the constant length if there is one because optimized match rule may exist. 1134 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 1135 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 1136 break; 1137 default: 1138 ShouldNotReachHere(); 1139 return NULL; 1140 } 1141 1142 // All these intrinsics have checks. 1143 C->set_has_split_ifs(true); // Has chance for split-if optimization 1144 clear_upper_avx(); 1145 1146 return _gvn.transform(result); 1147 } 1148 1149 //------------------------------inline_string_compareTo------------------------ 1150 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 1151 Node* arg1 = argument(0); 1152 Node* arg2 = argument(1); 1153 1154 arg1 = must_be_not_null(arg1, true); 1155 arg2 = must_be_not_null(arg2, true); 1156 1157 // Get start addr and length of first argument 1158 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1159 Node* arg1_cnt = load_array_length(arg1); 1160 1161 // Get start addr and length of second argument 1162 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1163 Node* arg2_cnt = load_array_length(arg2); 1164 1165 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1166 set_result(result); 1167 return true; 1168 } 1169 1170 //------------------------------inline_string_equals------------------------ 1171 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1172 Node* arg1 = argument(0); 1173 Node* arg2 = argument(1); 1174 1175 // paths (plus control) merge 1176 RegionNode* region = new RegionNode(3); 1177 Node* phi = new PhiNode(region, TypeInt::BOOL); 1178 1179 if (!stopped()) { 1180 1181 arg1 = must_be_not_null(arg1, true); 1182 arg2 = must_be_not_null(arg2, true); 1183 1184 // Get start addr and length of first argument 1185 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1186 Node* arg1_cnt = load_array_length(arg1); 1187 1188 // Get start addr and length of second argument 1189 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1190 Node* arg2_cnt = load_array_length(arg2); 1191 1192 // Check for arg1_cnt != arg2_cnt 1193 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1194 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1195 Node* if_ne = generate_slow_guard(bol, NULL); 1196 if (if_ne != NULL) { 1197 phi->init_req(2, intcon(0)); 1198 region->init_req(2, if_ne); 1199 } 1200 1201 // Check for count == 0 is done by assembler code for StrEquals. 1202 1203 if (!stopped()) { 1204 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1205 phi->init_req(1, equals); 1206 region->init_req(1, control()); 1207 } 1208 } 1209 1210 // post merge 1211 set_control(_gvn.transform(region)); 1212 record_for_igvn(region); 1213 1214 set_result(_gvn.transform(phi)); 1215 return true; 1216 } 1217 1218 //------------------------------inline_array_equals---------------------------- 1219 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1220 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1221 Node* arg1 = argument(0); 1222 Node* arg2 = argument(1); 1223 1224 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1225 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1226 clear_upper_avx(); 1227 1228 return true; 1229 } 1230 1231 //------------------------------inline_hasNegatives------------------------------ 1232 bool LibraryCallKit::inline_hasNegatives() { 1233 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1234 return false; 1235 } 1236 1237 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1238 // no receiver since it is static method 1239 Node* ba = argument(0); 1240 Node* offset = argument(1); 1241 Node* len = argument(2); 1242 1243 ba = must_be_not_null(ba, true); 1244 1245 // Range checks 1246 generate_string_range_check(ba, offset, len, false); 1247 if (stopped()) { 1248 return true; 1249 } 1250 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1251 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1252 set_result(_gvn.transform(result)); 1253 return true; 1254 } 1255 1256 bool LibraryCallKit::inline_preconditions_checkIndex() { 1257 Node* index = argument(0); 1258 Node* length = argument(1); 1259 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1260 return false; 1261 } 1262 1263 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1264 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1265 1266 { 1267 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1268 uncommon_trap(Deoptimization::Reason_intrinsic, 1269 Deoptimization::Action_make_not_entrant); 1270 } 1271 1272 if (stopped()) { 1273 return false; 1274 } 1275 1276 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1277 BoolTest::mask btest = BoolTest::lt; 1278 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1279 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1280 _gvn.set_type(rc, rc->Value(&_gvn)); 1281 if (!rc_bool->is_Con()) { 1282 record_for_igvn(rc); 1283 } 1284 set_control(_gvn.transform(new IfTrueNode(rc))); 1285 { 1286 PreserveJVMState pjvms(this); 1287 set_control(_gvn.transform(new IfFalseNode(rc))); 1288 uncommon_trap(Deoptimization::Reason_range_check, 1289 Deoptimization::Action_make_not_entrant); 1290 } 1291 1292 if (stopped()) { 1293 return false; 1294 } 1295 1296 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1297 result->set_req(0, control()); 1298 result = _gvn.transform(result); 1299 set_result(result); 1300 replace_in_map(index, result); 1301 clear_upper_avx(); 1302 return true; 1303 } 1304 1305 //------------------------------inline_string_indexOf------------------------ 1306 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1307 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1308 return false; 1309 } 1310 Node* src = argument(0); 1311 Node* tgt = argument(1); 1312 1313 // Make the merge point 1314 RegionNode* result_rgn = new RegionNode(4); 1315 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1316 1317 src = must_be_not_null(src, true); 1318 tgt = must_be_not_null(tgt, true); 1319 1320 // Get start addr and length of source string 1321 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1322 Node* src_count = load_array_length(src); 1323 1324 // Get start addr and length of substring 1325 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1326 Node* tgt_count = load_array_length(tgt); 1327 1328 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1329 // Divide src size by 2 if String is UTF16 encoded 1330 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1331 } 1332 if (ae == StrIntrinsicNode::UU) { 1333 // Divide substring size by 2 if String is UTF16 encoded 1334 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1335 } 1336 1337 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1338 if (result != NULL) { 1339 result_phi->init_req(3, result); 1340 result_rgn->init_req(3, control()); 1341 } 1342 set_control(_gvn.transform(result_rgn)); 1343 record_for_igvn(result_rgn); 1344 set_result(_gvn.transform(result_phi)); 1345 1346 return true; 1347 } 1348 1349 //-----------------------------inline_string_indexOf----------------------- 1350 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1351 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1352 return false; 1353 } 1354 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1355 return false; 1356 } 1357 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1358 Node* src = argument(0); // byte[] 1359 Node* src_count = argument(1); // char count 1360 Node* tgt = argument(2); // byte[] 1361 Node* tgt_count = argument(3); // char count 1362 Node* from_index = argument(4); // char index 1363 1364 src = must_be_not_null(src, true); 1365 tgt = must_be_not_null(tgt, true); 1366 1367 // Multiply byte array index by 2 if String is UTF16 encoded 1368 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1369 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1370 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1371 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1372 1373 // Range checks 1374 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1375 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1376 if (stopped()) { 1377 return true; 1378 } 1379 1380 RegionNode* region = new RegionNode(5); 1381 Node* phi = new PhiNode(region, TypeInt::INT); 1382 1383 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1384 if (result != NULL) { 1385 // The result is index relative to from_index if substring was found, -1 otherwise. 1386 // Generate code which will fold into cmove. 1387 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1388 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1389 1390 Node* if_lt = generate_slow_guard(bol, NULL); 1391 if (if_lt != NULL) { 1392 // result == -1 1393 phi->init_req(3, result); 1394 region->init_req(3, if_lt); 1395 } 1396 if (!stopped()) { 1397 result = _gvn.transform(new AddINode(result, from_index)); 1398 phi->init_req(4, result); 1399 region->init_req(4, control()); 1400 } 1401 } 1402 1403 set_control(_gvn.transform(region)); 1404 record_for_igvn(region); 1405 set_result(_gvn.transform(phi)); 1406 clear_upper_avx(); 1407 1408 return true; 1409 } 1410 1411 // Create StrIndexOfNode with fast path checks 1412 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1413 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1414 // Check for substr count > string count 1415 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1416 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1417 Node* if_gt = generate_slow_guard(bol, NULL); 1418 if (if_gt != NULL) { 1419 phi->init_req(1, intcon(-1)); 1420 region->init_req(1, if_gt); 1421 } 1422 if (!stopped()) { 1423 // Check for substr count == 0 1424 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1425 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1426 Node* if_zero = generate_slow_guard(bol, NULL); 1427 if (if_zero != NULL) { 1428 phi->init_req(2, intcon(0)); 1429 region->init_req(2, if_zero); 1430 } 1431 } 1432 if (!stopped()) { 1433 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1434 } 1435 return NULL; 1436 } 1437 1438 //-----------------------------inline_string_indexOfChar----------------------- 1439 bool LibraryCallKit::inline_string_indexOfChar() { 1440 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1441 return false; 1442 } 1443 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1444 return false; 1445 } 1446 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1447 Node* src = argument(0); // byte[] 1448 Node* tgt = argument(1); // tgt is int ch 1449 Node* from_index = argument(2); 1450 Node* max = argument(3); 1451 1452 src = must_be_not_null(src, true); 1453 1454 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1455 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1456 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1457 1458 // Range checks 1459 generate_string_range_check(src, src_offset, src_count, true); 1460 if (stopped()) { 1461 return true; 1462 } 1463 1464 RegionNode* region = new RegionNode(3); 1465 Node* phi = new PhiNode(region, TypeInt::INT); 1466 1467 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1468 C->set_has_split_ifs(true); // Has chance for split-if optimization 1469 _gvn.transform(result); 1470 1471 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1472 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1473 1474 Node* if_lt = generate_slow_guard(bol, NULL); 1475 if (if_lt != NULL) { 1476 // result == -1 1477 phi->init_req(2, result); 1478 region->init_req(2, if_lt); 1479 } 1480 if (!stopped()) { 1481 result = _gvn.transform(new AddINode(result, from_index)); 1482 phi->init_req(1, result); 1483 region->init_req(1, control()); 1484 } 1485 set_control(_gvn.transform(region)); 1486 record_for_igvn(region); 1487 set_result(_gvn.transform(phi)); 1488 1489 return true; 1490 } 1491 //---------------------------inline_string_copy--------------------- 1492 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1493 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1494 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1495 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1496 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1497 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1498 bool LibraryCallKit::inline_string_copy(bool compress) { 1499 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1500 return false; 1501 } 1502 int nargs = 5; // 2 oops, 3 ints 1503 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1504 1505 Node* src = argument(0); 1506 Node* src_offset = argument(1); 1507 Node* dst = argument(2); 1508 Node* dst_offset = argument(3); 1509 Node* length = argument(4); 1510 1511 // Check for allocation before we add nodes that would confuse 1512 // tightly_coupled_allocation() 1513 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1514 1515 // Figure out the size and type of the elements we will be copying. 1516 const Type* src_type = src->Value(&_gvn); 1517 const Type* dst_type = dst->Value(&_gvn); 1518 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1519 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1520 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1521 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1522 "Unsupported array types for inline_string_copy"); 1523 1524 src = must_be_not_null(src, true); 1525 dst = must_be_not_null(dst, true); 1526 1527 // Convert char[] offsets to byte[] offsets 1528 bool convert_src = (compress && src_elem == T_BYTE); 1529 bool convert_dst = (!compress && dst_elem == T_BYTE); 1530 if (convert_src) { 1531 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1532 } else if (convert_dst) { 1533 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1534 } 1535 1536 // Range checks 1537 generate_string_range_check(src, src_offset, length, convert_src); 1538 generate_string_range_check(dst, dst_offset, length, convert_dst); 1539 if (stopped()) { 1540 return true; 1541 } 1542 1543 Node* src_start = array_element_address(src, src_offset, src_elem); 1544 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1545 // 'src_start' points to src array + scaled offset 1546 // 'dst_start' points to dst array + scaled offset 1547 Node* count = NULL; 1548 if (compress) { 1549 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1550 } else { 1551 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1552 } 1553 1554 if (alloc != NULL) { 1555 if (alloc->maybe_set_complete(&_gvn)) { 1556 // "You break it, you buy it." 1557 InitializeNode* init = alloc->initialization(); 1558 assert(init->is_complete(), "we just did this"); 1559 init->set_complete_with_arraycopy(); 1560 assert(dst->is_CheckCastPP(), "sanity"); 1561 assert(dst->in(0)->in(0) == init, "dest pinned"); 1562 } 1563 // Do not let stores that initialize this object be reordered with 1564 // a subsequent store that would make this object accessible by 1565 // other threads. 1566 // Record what AllocateNode this StoreStore protects so that 1567 // escape analysis can go from the MemBarStoreStoreNode to the 1568 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1569 // based on the escape status of the AllocateNode. 1570 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1571 } 1572 if (compress) { 1573 set_result(_gvn.transform(count)); 1574 } 1575 clear_upper_avx(); 1576 1577 return true; 1578 } 1579 1580 #ifdef _LP64 1581 #define XTOP ,top() /*additional argument*/ 1582 #else //_LP64 1583 #define XTOP /*no additional argument*/ 1584 #endif //_LP64 1585 1586 //------------------------inline_string_toBytesU-------------------------- 1587 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1588 bool LibraryCallKit::inline_string_toBytesU() { 1589 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1590 return false; 1591 } 1592 // Get the arguments. 1593 Node* value = argument(0); 1594 Node* offset = argument(1); 1595 Node* length = argument(2); 1596 1597 Node* newcopy = NULL; 1598 1599 // Set the original stack and the reexecute bit for the interpreter to reexecute 1600 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1601 { PreserveReexecuteState preexecs(this); 1602 jvms()->set_should_reexecute(true); 1603 1604 // Check if a null path was taken unconditionally. 1605 value = null_check(value); 1606 1607 RegionNode* bailout = new RegionNode(1); 1608 record_for_igvn(bailout); 1609 1610 // Range checks 1611 generate_negative_guard(offset, bailout); 1612 generate_negative_guard(length, bailout); 1613 generate_limit_guard(offset, length, load_array_length(value), bailout); 1614 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1615 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1616 1617 if (bailout->req() > 1) { 1618 PreserveJVMState pjvms(this); 1619 set_control(_gvn.transform(bailout)); 1620 uncommon_trap(Deoptimization::Reason_intrinsic, 1621 Deoptimization::Action_maybe_recompile); 1622 } 1623 if (stopped()) { 1624 return true; 1625 } 1626 1627 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1628 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1629 newcopy = new_array(klass_node, size, 0); // no arguments to push 1630 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1631 1632 // Calculate starting addresses. 1633 Node* src_start = array_element_address(value, offset, T_CHAR); 1634 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1635 1636 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1637 const TypeInt* toffset = gvn().type(offset)->is_int(); 1638 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1639 1640 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1641 const char* copyfunc_name = "arraycopy"; 1642 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1643 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1644 OptoRuntime::fast_arraycopy_Type(), 1645 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1646 src_start, dst_start, ConvI2X(length) XTOP); 1647 // Do not let reads from the cloned object float above the arraycopy. 1648 if (alloc != NULL) { 1649 if (alloc->maybe_set_complete(&_gvn)) { 1650 // "You break it, you buy it." 1651 InitializeNode* init = alloc->initialization(); 1652 assert(init->is_complete(), "we just did this"); 1653 init->set_complete_with_arraycopy(); 1654 assert(newcopy->is_CheckCastPP(), "sanity"); 1655 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1656 } 1657 // Do not let stores that initialize this object be reordered with 1658 // a subsequent store that would make this object accessible by 1659 // other threads. 1660 // Record what AllocateNode this StoreStore protects so that 1661 // escape analysis can go from the MemBarStoreStoreNode to the 1662 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1663 // based on the escape status of the AllocateNode. 1664 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1665 } else { 1666 insert_mem_bar(Op_MemBarCPUOrder); 1667 } 1668 } // original reexecute is set back here 1669 1670 C->set_has_split_ifs(true); // Has chance for split-if optimization 1671 if (!stopped()) { 1672 set_result(newcopy); 1673 } 1674 clear_upper_avx(); 1675 1676 return true; 1677 } 1678 1679 //------------------------inline_string_getCharsU-------------------------- 1680 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1681 bool LibraryCallKit::inline_string_getCharsU() { 1682 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1683 return false; 1684 } 1685 1686 // Get the arguments. 1687 Node* src = argument(0); 1688 Node* src_begin = argument(1); 1689 Node* src_end = argument(2); // exclusive offset (i < src_end) 1690 Node* dst = argument(3); 1691 Node* dst_begin = argument(4); 1692 1693 // Check for allocation before we add nodes that would confuse 1694 // tightly_coupled_allocation() 1695 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1696 1697 // Check if a null path was taken unconditionally. 1698 src = null_check(src); 1699 dst = null_check(dst); 1700 if (stopped()) { 1701 return true; 1702 } 1703 1704 // Get length and convert char[] offset to byte[] offset 1705 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1706 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1707 1708 // Range checks 1709 generate_string_range_check(src, src_begin, length, true); 1710 generate_string_range_check(dst, dst_begin, length, false); 1711 if (stopped()) { 1712 return true; 1713 } 1714 1715 if (!stopped()) { 1716 // Calculate starting addresses. 1717 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1718 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1719 1720 // Check if array addresses are aligned to HeapWordSize 1721 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1722 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1723 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1724 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1725 1726 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1727 const char* copyfunc_name = "arraycopy"; 1728 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1729 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1730 OptoRuntime::fast_arraycopy_Type(), 1731 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1732 src_start, dst_start, ConvI2X(length) XTOP); 1733 // Do not let reads from the cloned object float above the arraycopy. 1734 if (alloc != NULL) { 1735 if (alloc->maybe_set_complete(&_gvn)) { 1736 // "You break it, you buy it." 1737 InitializeNode* init = alloc->initialization(); 1738 assert(init->is_complete(), "we just did this"); 1739 init->set_complete_with_arraycopy(); 1740 assert(dst->is_CheckCastPP(), "sanity"); 1741 assert(dst->in(0)->in(0) == init, "dest pinned"); 1742 } 1743 // Do not let stores that initialize this object be reordered with 1744 // a subsequent store that would make this object accessible by 1745 // other threads. 1746 // Record what AllocateNode this StoreStore protects so that 1747 // escape analysis can go from the MemBarStoreStoreNode to the 1748 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1749 // based on the escape status of the AllocateNode. 1750 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 1751 } else { 1752 insert_mem_bar(Op_MemBarCPUOrder); 1753 } 1754 } 1755 1756 C->set_has_split_ifs(true); // Has chance for split-if optimization 1757 return true; 1758 } 1759 1760 //----------------------inline_string_char_access---------------------------- 1761 // Store/Load char to/from byte[] array. 1762 // static void StringUTF16.putChar(byte[] val, int index, int c) 1763 // static char StringUTF16.getChar(byte[] val, int index) 1764 bool LibraryCallKit::inline_string_char_access(bool is_store) { 1765 Node* value = argument(0); 1766 Node* index = argument(1); 1767 Node* ch = is_store ? argument(2) : NULL; 1768 1769 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1770 // correctly requires matched array shapes. 1771 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1772 "sanity: byte[] and char[] bases agree"); 1773 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1774 "sanity: byte[] and char[] scales agree"); 1775 1776 // Bail when getChar over constants is requested: constant folding would 1777 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1778 // Java method would constant fold nicely instead. 1779 if (!is_store && value->is_Con() && index->is_Con()) { 1780 return false; 1781 } 1782 1783 value = must_be_not_null(value, true); 1784 1785 Node* adr = array_element_address(value, index, T_CHAR); 1786 if (adr->is_top()) { 1787 return false; 1788 } 1789 if (is_store) { 1790 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1791 false, false, true /* mismatched */); 1792 } else { 1793 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1794 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */); 1795 set_result(ch); 1796 } 1797 return true; 1798 } 1799 1800 //--------------------------round_double_node-------------------------------- 1801 // Round a double node if necessary. 1802 Node* LibraryCallKit::round_double_node(Node* n) { 1803 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1804 n = _gvn.transform(new RoundDoubleNode(0, n)); 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_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 default: fatal_unexpected_iid(id); break; 1820 } 1821 set_result(_gvn.transform(n)); 1822 return true; 1823 } 1824 1825 //------------------------------runtime_math----------------------------- 1826 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1827 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1828 "must be (DD)D or (D)D type"); 1829 1830 // Inputs 1831 Node* a = round_double_node(argument(0)); 1832 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1833 1834 const TypePtr* no_memory_effects = NULL; 1835 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1836 no_memory_effects, 1837 a, top(), b, b ? top() : NULL); 1838 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1839 #ifdef ASSERT 1840 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1841 assert(value_top == top(), "second value must be top"); 1842 #endif 1843 1844 set_result(value); 1845 return true; 1846 } 1847 1848 //------------------------------inline_math_native----------------------------- 1849 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1850 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1851 switch (id) { 1852 // These intrinsics are not properly supported on all hardware 1853 case vmIntrinsics::_dsin: 1854 return StubRoutines::dsin() != NULL ? 1855 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1856 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1857 case vmIntrinsics::_dcos: 1858 return StubRoutines::dcos() != NULL ? 1859 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1860 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1861 case vmIntrinsics::_dtan: 1862 return StubRoutines::dtan() != NULL ? 1863 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1864 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1865 case vmIntrinsics::_dlog: 1866 return StubRoutines::dlog() != NULL ? 1867 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1868 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1869 case vmIntrinsics::_dlog10: 1870 return StubRoutines::dlog10() != NULL ? 1871 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1872 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1873 1874 // These intrinsics are supported on all hardware 1875 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1876 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1877 1878 case vmIntrinsics::_dexp: 1879 return StubRoutines::dexp() != NULL ? 1880 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1881 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1882 case vmIntrinsics::_dpow: { 1883 Node* exp = round_double_node(argument(2)); 1884 const TypeD* d = _gvn.type(exp)->isa_double_constant(); 1885 if (d != NULL && d->getd() == 2.0) { 1886 // Special case: pow(x, 2.0) => x * x 1887 Node* base = round_double_node(argument(0)); 1888 set_result(_gvn.transform(new MulDNode(base, base))); 1889 return true; 1890 } 1891 return StubRoutines::dpow() != NULL ? 1892 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1893 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1894 } 1895 #undef FN_PTR 1896 1897 // These intrinsics are not yet correctly implemented 1898 case vmIntrinsics::_datan2: 1899 return false; 1900 1901 default: 1902 fatal_unexpected_iid(id); 1903 return false; 1904 } 1905 } 1906 1907 static bool is_simple_name(Node* n) { 1908 return (n->req() == 1 // constant 1909 || (n->is_Type() && n->as_Type()->type()->singleton()) 1910 || n->is_Proj() // parameter or return value 1911 || n->is_Phi() // local of some sort 1912 ); 1913 } 1914 1915 //----------------------------inline_notify-----------------------------------* 1916 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1917 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1918 address func; 1919 if (id == vmIntrinsics::_notify) { 1920 func = OptoRuntime::monitor_notify_Java(); 1921 } else { 1922 func = OptoRuntime::monitor_notifyAll_Java(); 1923 } 1924 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1925 make_slow_call_ex(call, env()->Throwable_klass(), false); 1926 return true; 1927 } 1928 1929 1930 //----------------------------inline_min_max----------------------------------- 1931 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1932 set_result(generate_min_max(id, argument(0), argument(1))); 1933 return true; 1934 } 1935 1936 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1937 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1938 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1939 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1940 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1941 1942 { 1943 PreserveJVMState pjvms(this); 1944 PreserveReexecuteState preexecs(this); 1945 jvms()->set_should_reexecute(true); 1946 1947 set_control(slow_path); 1948 set_i_o(i_o()); 1949 1950 uncommon_trap(Deoptimization::Reason_intrinsic, 1951 Deoptimization::Action_none); 1952 } 1953 1954 set_control(fast_path); 1955 set_result(math); 1956 } 1957 1958 template <typename OverflowOp> 1959 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1960 typedef typename OverflowOp::MathOp MathOp; 1961 1962 MathOp* mathOp = new MathOp(arg1, arg2); 1963 Node* operation = _gvn.transform( mathOp ); 1964 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1965 inline_math_mathExact(operation, ofcheck); 1966 return true; 1967 } 1968 1969 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1970 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1971 } 1972 1973 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1974 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1975 } 1976 1977 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1978 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1979 } 1980 1981 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1982 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1983 } 1984 1985 bool LibraryCallKit::inline_math_negateExactI() { 1986 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1987 } 1988 1989 bool LibraryCallKit::inline_math_negateExactL() { 1990 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 1991 } 1992 1993 bool LibraryCallKit::inline_math_multiplyExactI() { 1994 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 1995 } 1996 1997 bool LibraryCallKit::inline_math_multiplyExactL() { 1998 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 1999 } 2000 2001 bool LibraryCallKit::inline_math_multiplyHigh() { 2002 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2)))); 2003 return true; 2004 } 2005 2006 Node* 2007 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2008 // These are the candidate return value: 2009 Node* xvalue = x0; 2010 Node* yvalue = y0; 2011 2012 if (xvalue == yvalue) { 2013 return xvalue; 2014 } 2015 2016 bool want_max = (id == vmIntrinsics::_max); 2017 2018 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2019 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2020 if (txvalue == NULL || tyvalue == NULL) return top(); 2021 // This is not really necessary, but it is consistent with a 2022 // hypothetical MaxINode::Value method: 2023 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2024 2025 // %%% This folding logic should (ideally) be in a different place. 2026 // Some should be inside IfNode, and there to be a more reliable 2027 // transformation of ?: style patterns into cmoves. We also want 2028 // more powerful optimizations around cmove and min/max. 2029 2030 // Try to find a dominating comparison of these guys. 2031 // It can simplify the index computation for Arrays.copyOf 2032 // and similar uses of System.arraycopy. 2033 // First, compute the normalized version of CmpI(x, y). 2034 int cmp_op = Op_CmpI; 2035 Node* xkey = xvalue; 2036 Node* ykey = yvalue; 2037 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2038 if (ideal_cmpxy->is_Cmp()) { 2039 // E.g., if we have CmpI(length - offset, count), 2040 // it might idealize to CmpI(length, count + offset) 2041 cmp_op = ideal_cmpxy->Opcode(); 2042 xkey = ideal_cmpxy->in(1); 2043 ykey = ideal_cmpxy->in(2); 2044 } 2045 2046 // Start by locating any relevant comparisons. 2047 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2048 Node* cmpxy = NULL; 2049 Node* cmpyx = NULL; 2050 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2051 Node* cmp = start_from->fast_out(k); 2052 if (cmp->outcnt() > 0 && // must have prior uses 2053 cmp->in(0) == NULL && // must be context-independent 2054 cmp->Opcode() == cmp_op) { // right kind of compare 2055 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2056 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2057 } 2058 } 2059 2060 const int NCMPS = 2; 2061 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2062 int cmpn; 2063 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2064 if (cmps[cmpn] != NULL) break; // find a result 2065 } 2066 if (cmpn < NCMPS) { 2067 // Look for a dominating test that tells us the min and max. 2068 int depth = 0; // Limit search depth for speed 2069 Node* dom = control(); 2070 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2071 if (++depth >= 100) break; 2072 Node* ifproj = dom; 2073 if (!ifproj->is_Proj()) continue; 2074 Node* iff = ifproj->in(0); 2075 if (!iff->is_If()) continue; 2076 Node* bol = iff->in(1); 2077 if (!bol->is_Bool()) continue; 2078 Node* cmp = bol->in(1); 2079 if (cmp == NULL) continue; 2080 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2081 if (cmps[cmpn] == cmp) break; 2082 if (cmpn == NCMPS) continue; 2083 BoolTest::mask btest = bol->as_Bool()->_test._test; 2084 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2085 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2086 // At this point, we know that 'x btest y' is true. 2087 switch (btest) { 2088 case BoolTest::eq: 2089 // They are proven equal, so we can collapse the min/max. 2090 // Either value is the answer. Choose the simpler. 2091 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2092 return yvalue; 2093 return xvalue; 2094 case BoolTest::lt: // x < y 2095 case BoolTest::le: // x <= y 2096 return (want_max ? yvalue : xvalue); 2097 case BoolTest::gt: // x > y 2098 case BoolTest::ge: // x >= y 2099 return (want_max ? xvalue : yvalue); 2100 default: 2101 break; 2102 } 2103 } 2104 } 2105 2106 // We failed to find a dominating test. 2107 // Let's pick a test that might GVN with prior tests. 2108 Node* best_bol = NULL; 2109 BoolTest::mask best_btest = BoolTest::illegal; 2110 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2111 Node* cmp = cmps[cmpn]; 2112 if (cmp == NULL) continue; 2113 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2114 Node* bol = cmp->fast_out(j); 2115 if (!bol->is_Bool()) continue; 2116 BoolTest::mask btest = bol->as_Bool()->_test._test; 2117 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2118 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2119 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2120 best_bol = bol->as_Bool(); 2121 best_btest = btest; 2122 } 2123 } 2124 } 2125 2126 Node* answer_if_true = NULL; 2127 Node* answer_if_false = NULL; 2128 switch (best_btest) { 2129 default: 2130 if (cmpxy == NULL) 2131 cmpxy = ideal_cmpxy; 2132 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2133 // and fall through: 2134 case BoolTest::lt: // x < y 2135 case BoolTest::le: // x <= y 2136 answer_if_true = (want_max ? yvalue : xvalue); 2137 answer_if_false = (want_max ? xvalue : yvalue); 2138 break; 2139 case BoolTest::gt: // x > y 2140 case BoolTest::ge: // x >= y 2141 answer_if_true = (want_max ? xvalue : yvalue); 2142 answer_if_false = (want_max ? yvalue : xvalue); 2143 break; 2144 } 2145 2146 jint hi, lo; 2147 if (want_max) { 2148 // We can sharpen the minimum. 2149 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2150 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2151 } else { 2152 // We can sharpen the maximum. 2153 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2154 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2155 } 2156 2157 // Use a flow-free graph structure, to avoid creating excess control edges 2158 // which could hinder other optimizations. 2159 // Since Math.min/max is often used with arraycopy, we want 2160 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2161 Node* cmov = CMoveNode::make(NULL, best_bol, 2162 answer_if_false, answer_if_true, 2163 TypeInt::make(lo, hi, widen)); 2164 2165 return _gvn.transform(cmov); 2166 2167 /* 2168 // This is not as desirable as it may seem, since Min and Max 2169 // nodes do not have a full set of optimizations. 2170 // And they would interfere, anyway, with 'if' optimizations 2171 // and with CMoveI canonical forms. 2172 switch (id) { 2173 case vmIntrinsics::_min: 2174 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2175 case vmIntrinsics::_max: 2176 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2177 default: 2178 ShouldNotReachHere(); 2179 } 2180 */ 2181 } 2182 2183 inline int 2184 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) { 2185 const TypePtr* base_type = TypePtr::NULL_PTR; 2186 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2187 if (base_type == NULL) { 2188 // Unknown type. 2189 return Type::AnyPtr; 2190 } else if (base_type == TypePtr::NULL_PTR) { 2191 // Since this is a NULL+long form, we have to switch to a rawptr. 2192 base = _gvn.transform(new CastX2PNode(offset)); 2193 offset = MakeConX(0); 2194 return Type::RawPtr; 2195 } else if (base_type->base() == Type::RawPtr) { 2196 return Type::RawPtr; 2197 } else if (base_type->isa_oopptr()) { 2198 // Base is never null => always a heap address. 2199 if (!TypePtr::NULL_PTR->higher_equal(base_type)) { 2200 return Type::OopPtr; 2201 } 2202 // Offset is small => always a heap address. 2203 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2204 if (offset_type != NULL && 2205 base_type->offset() == 0 && // (should always be?) 2206 offset_type->_lo >= 0 && 2207 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2208 return Type::OopPtr; 2209 } else if (type == T_OBJECT) { 2210 // off heap access to an oop doesn't make any sense. Has to be on 2211 // heap. 2212 return Type::OopPtr; 2213 } 2214 // Otherwise, it might either be oop+off or NULL+addr. 2215 return Type::AnyPtr; 2216 } else { 2217 // No information: 2218 return Type::AnyPtr; 2219 } 2220 } 2221 2222 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) { 2223 Node* uncasted_base = base; 2224 int kind = classify_unsafe_addr(uncasted_base, offset, type); 2225 if (kind == Type::RawPtr) { 2226 return basic_plus_adr(top(), uncasted_base, offset); 2227 } else if (kind == Type::AnyPtr) { 2228 assert(base == uncasted_base, "unexpected base change"); 2229 if (can_cast) { 2230 if (!_gvn.type(base)->speculative_maybe_null() && 2231 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2232 // According to profiling, this access is always on 2233 // heap. Casting the base to not null and thus avoiding membars 2234 // around the access should allow better optimizations 2235 Node* null_ctl = top(); 2236 base = null_check_oop(base, &null_ctl, true, true, true); 2237 assert(null_ctl->is_top(), "no null control here"); 2238 return basic_plus_adr(base, offset); 2239 } else if (_gvn.type(base)->speculative_always_null() && 2240 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2241 // According to profiling, this access is always off 2242 // heap. 2243 base = null_assert(base); 2244 Node* raw_base = _gvn.transform(new CastX2PNode(offset)); 2245 offset = MakeConX(0); 2246 return basic_plus_adr(top(), raw_base, offset); 2247 } 2248 } 2249 // We don't know if it's an on heap or off heap access. Fall back 2250 // to raw memory access. 2251 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM)); 2252 return basic_plus_adr(top(), raw, offset); 2253 } else { 2254 assert(base == uncasted_base, "unexpected base change"); 2255 // We know it's an on heap access so base can't be null 2256 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) { 2257 base = must_be_not_null(base, true); 2258 } 2259 return basic_plus_adr(base, offset); 2260 } 2261 } 2262 2263 //--------------------------inline_number_methods----------------------------- 2264 // inline int Integer.numberOfLeadingZeros(int) 2265 // inline int Long.numberOfLeadingZeros(long) 2266 // 2267 // inline int Integer.numberOfTrailingZeros(int) 2268 // inline int Long.numberOfTrailingZeros(long) 2269 // 2270 // inline int Integer.bitCount(int) 2271 // inline int Long.bitCount(long) 2272 // 2273 // inline char Character.reverseBytes(char) 2274 // inline short Short.reverseBytes(short) 2275 // inline int Integer.reverseBytes(int) 2276 // inline long Long.reverseBytes(long) 2277 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2278 Node* arg = argument(0); 2279 Node* n = NULL; 2280 switch (id) { 2281 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2282 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2283 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2284 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2285 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2286 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2287 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2288 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2289 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2290 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2291 default: fatal_unexpected_iid(id); break; 2292 } 2293 set_result(_gvn.transform(n)); 2294 return true; 2295 } 2296 2297 //----------------------------inline_unsafe_access---------------------------- 2298 2299 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2300 // Attempt to infer a sharper value type from the offset and base type. 2301 ciKlass* sharpened_klass = NULL; 2302 2303 // See if it is an instance field, with an object type. 2304 if (alias_type->field() != NULL) { 2305 if (alias_type->field()->type()->is_klass()) { 2306 sharpened_klass = alias_type->field()->type()->as_klass(); 2307 } 2308 } 2309 2310 // See if it is a narrow oop array. 2311 if (adr_type->isa_aryptr()) { 2312 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2313 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2314 if (elem_type != NULL) { 2315 sharpened_klass = elem_type->klass(); 2316 } 2317 } 2318 } 2319 2320 // The sharpened class might be unloaded if there is no class loader 2321 // contraint in place. 2322 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2323 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2324 2325 #ifndef PRODUCT 2326 if (C->print_intrinsics() || C->print_inlining()) { 2327 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2328 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2329 } 2330 #endif 2331 // Sharpen the value type. 2332 return tjp; 2333 } 2334 return NULL; 2335 } 2336 2337 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) { 2338 switch (kind) { 2339 case Relaxed: 2340 return MO_UNORDERED; 2341 case Opaque: 2342 return MO_RELAXED; 2343 case Acquire: 2344 return MO_ACQUIRE; 2345 case Release: 2346 return MO_RELEASE; 2347 case Volatile: 2348 return MO_SEQ_CST; 2349 default: 2350 ShouldNotReachHere(); 2351 return 0; 2352 } 2353 } 2354 2355 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2356 if (callee()->is_static()) return false; // caller must have the capability! 2357 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2358 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2359 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2360 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2361 2362 if (type == T_OBJECT || type == T_ARRAY) { 2363 decorators |= ON_UNKNOWN_OOP_REF; 2364 } 2365 2366 if (unaligned) { 2367 decorators |= C2_UNALIGNED; 2368 } 2369 2370 #ifndef PRODUCT 2371 { 2372 ResourceMark rm; 2373 // Check the signatures. 2374 ciSignature* sig = callee()->signature(); 2375 #ifdef ASSERT 2376 if (!is_store) { 2377 // Object getObject(Object base, int/long offset), etc. 2378 BasicType rtype = sig->return_type()->basic_type(); 2379 assert(rtype == type, "getter must return the expected value"); 2380 assert(sig->count() == 2, "oop getter has 2 arguments"); 2381 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2382 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2383 } else { 2384 // void putObject(Object base, int/long offset, Object x), etc. 2385 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2386 assert(sig->count() == 3, "oop putter has 3 arguments"); 2387 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2388 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2389 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2390 assert(vtype == type, "putter must accept the expected value"); 2391 } 2392 #endif // ASSERT 2393 } 2394 #endif //PRODUCT 2395 2396 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2397 2398 Node* receiver = argument(0); // type: oop 2399 2400 // Build address expression. 2401 Node* adr; 2402 Node* heap_base_oop = top(); 2403 Node* offset = top(); 2404 Node* val; 2405 2406 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2407 Node* base = argument(1); // type: oop 2408 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2409 offset = argument(2); // type: long 2410 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2411 // to be plain byte offsets, which are also the same as those accepted 2412 // by oopDesc::field_addr. 2413 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2414 "fieldOffset must be byte-scaled"); 2415 // 32-bit machines ignore the high half! 2416 offset = ConvL2X(offset); 2417 adr = make_unsafe_address(base, offset, type, kind == Relaxed); 2418 2419 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) { 2420 heap_base_oop = base; 2421 } else if (type == T_OBJECT) { 2422 return false; // off-heap oop accesses are not supported 2423 } 2424 2425 // Can base be NULL? Otherwise, always on-heap access. 2426 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop)); 2427 2428 if (!can_access_non_heap) { 2429 decorators |= IN_HEAP; 2430 } 2431 2432 val = is_store ? argument(4) : NULL; 2433 2434 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2435 2436 // Try to categorize the address. 2437 Compile::AliasType* alias_type = C->alias_type(adr_type); 2438 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2439 2440 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2441 alias_type->adr_type() == TypeAryPtr::RANGE) { 2442 return false; // not supported 2443 } 2444 2445 bool mismatched = false; 2446 BasicType bt = alias_type->basic_type(); 2447 if (bt != T_ILLEGAL) { 2448 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2449 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2450 // Alias type doesn't differentiate between byte[] and boolean[]). 2451 // Use address type to get the element type. 2452 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2453 } 2454 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2455 // accessing an array field with getObject is not a mismatch 2456 bt = T_OBJECT; 2457 } 2458 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2459 // Don't intrinsify mismatched object accesses 2460 return false; 2461 } 2462 mismatched = (bt != type); 2463 } else if (alias_type->adr_type()->isa_oopptr()) { 2464 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched 2465 } 2466 2467 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2468 2469 if (mismatched) { 2470 decorators |= C2_MISMATCHED; 2471 } 2472 2473 // First guess at the value type. 2474 const Type *value_type = Type::get_const_basic_type(type); 2475 2476 // Figure out the memory ordering. 2477 decorators |= mo_decorator_for_access_kind(kind); 2478 2479 if (!is_store && type == T_OBJECT) { 2480 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2481 if (tjp != NULL) { 2482 value_type = tjp; 2483 } 2484 } 2485 2486 receiver = null_check(receiver); 2487 if (stopped()) { 2488 return true; 2489 } 2490 // Heap pointers get a null-check from the interpreter, 2491 // as a courtesy. However, this is not guaranteed by Unsafe, 2492 // and it is not possible to fully distinguish unintended nulls 2493 // from intended ones in this API. 2494 2495 if (!is_store) { 2496 Node* p = NULL; 2497 // Try to constant fold a load from a constant field 2498 ciField* field = alias_type->field(); 2499 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { 2500 // final or stable field 2501 p = make_constant_from_field(field, heap_base_oop); 2502 } 2503 2504 if (p == NULL) { // Could not constant fold the load 2505 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators); 2506 // Normalize the value returned by getBoolean in the following cases 2507 if (type == T_BOOLEAN && 2508 (mismatched || 2509 heap_base_oop == top() || // - heap_base_oop is NULL or 2510 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL 2511 // and the unsafe access is made to large offset 2512 // (i.e., larger than the maximum offset necessary for any 2513 // field access) 2514 ) { 2515 IdealKit ideal = IdealKit(this); 2516 #define __ ideal. 2517 IdealVariable normalized_result(ideal); 2518 __ declarations_done(); 2519 __ set(normalized_result, p); 2520 __ if_then(p, BoolTest::ne, ideal.ConI(0)); 2521 __ set(normalized_result, ideal.ConI(1)); 2522 ideal.end_if(); 2523 final_sync(ideal); 2524 p = __ value(normalized_result); 2525 #undef __ 2526 } 2527 } 2528 if (type == T_ADDRESS) { 2529 p = gvn().transform(new CastP2XNode(NULL, p)); 2530 p = ConvX2UL(p); 2531 } 2532 // The load node has the control of the preceding MemBarCPUOrder. All 2533 // following nodes will have the control of the MemBarCPUOrder inserted at 2534 // the end of this method. So, pushing the load onto the stack at a later 2535 // point is fine. 2536 set_result(p); 2537 } else { 2538 if (bt == T_ADDRESS) { 2539 // Repackage the long as a pointer. 2540 val = ConvL2X(val); 2541 val = gvn().transform(new CastX2PNode(val)); 2542 } 2543 access_store_at(control(), heap_base_oop, adr, adr_type, val, value_type, type, decorators); 2544 } 2545 2546 return true; 2547 } 2548 2549 //----------------------------inline_unsafe_load_store---------------------------- 2550 // This method serves a couple of different customers (depending on LoadStoreKind): 2551 // 2552 // LS_cmp_swap: 2553 // 2554 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x); 2555 // boolean compareAndSetInt( Object o, long offset, int expected, int x); 2556 // boolean compareAndSetLong( Object o, long offset, long expected, long x); 2557 // 2558 // LS_cmp_swap_weak: 2559 // 2560 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x); 2561 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x); 2562 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x); 2563 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x); 2564 // 2565 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x); 2566 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x); 2567 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x); 2568 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x); 2569 // 2570 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x); 2571 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x); 2572 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x); 2573 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x); 2574 // 2575 // LS_cmp_exchange: 2576 // 2577 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x); 2578 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x); 2579 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x); 2580 // 2581 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2582 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2583 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2584 // 2585 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2586 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2587 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2588 // 2589 // LS_get_add: 2590 // 2591 // int getAndAddInt( Object o, long offset, int delta) 2592 // long getAndAddLong(Object o, long offset, long delta) 2593 // 2594 // LS_get_set: 2595 // 2596 // int getAndSet(Object o, long offset, int newValue) 2597 // long getAndSet(Object o, long offset, long newValue) 2598 // Object getAndSet(Object o, long offset, Object newValue) 2599 // 2600 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2601 // This basic scheme here is the same as inline_unsafe_access, but 2602 // differs in enough details that combining them would make the code 2603 // overly confusing. (This is a true fact! I originally combined 2604 // them, but even I was confused by it!) As much code/comments as 2605 // possible are retained from inline_unsafe_access though to make 2606 // the correspondences clearer. - dl 2607 2608 if (callee()->is_static()) return false; // caller must have the capability! 2609 2610 DecoratorSet decorators = C2_UNSAFE_ACCESS; 2611 decorators |= mo_decorator_for_access_kind(access_kind); 2612 2613 #ifndef PRODUCT 2614 BasicType rtype; 2615 { 2616 ResourceMark rm; 2617 // Check the signatures. 2618 ciSignature* sig = callee()->signature(); 2619 rtype = sig->return_type()->basic_type(); 2620 switch(kind) { 2621 case LS_get_add: 2622 case LS_get_set: { 2623 // Check the signatures. 2624 #ifdef ASSERT 2625 assert(rtype == type, "get and set must return the expected type"); 2626 assert(sig->count() == 3, "get and set has 3 arguments"); 2627 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2628 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2629 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2630 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2631 #endif // ASSERT 2632 break; 2633 } 2634 case LS_cmp_swap: 2635 case LS_cmp_swap_weak: { 2636 // Check the signatures. 2637 #ifdef ASSERT 2638 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2639 assert(sig->count() == 4, "CAS has 4 arguments"); 2640 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2641 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2642 #endif // ASSERT 2643 break; 2644 } 2645 case LS_cmp_exchange: { 2646 // Check the signatures. 2647 #ifdef ASSERT 2648 assert(rtype == type, "CAS must return the expected type"); 2649 assert(sig->count() == 4, "CAS has 4 arguments"); 2650 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2651 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2652 #endif // ASSERT 2653 break; 2654 } 2655 default: 2656 ShouldNotReachHere(); 2657 } 2658 } 2659 #endif //PRODUCT 2660 2661 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2662 2663 // Get arguments: 2664 Node* receiver = NULL; 2665 Node* base = NULL; 2666 Node* offset = NULL; 2667 Node* oldval = NULL; 2668 Node* newval = NULL; 2669 switch(kind) { 2670 case LS_cmp_swap: 2671 case LS_cmp_swap_weak: 2672 case LS_cmp_exchange: { 2673 const bool two_slot_type = type2size[type] == 2; 2674 receiver = argument(0); // type: oop 2675 base = argument(1); // type: oop 2676 offset = argument(2); // type: long 2677 oldval = argument(4); // type: oop, int, or long 2678 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2679 break; 2680 } 2681 case LS_get_add: 2682 case LS_get_set: { 2683 receiver = argument(0); // type: oop 2684 base = argument(1); // type: oop 2685 offset = argument(2); // type: long 2686 oldval = NULL; 2687 newval = argument(4); // type: oop, int, or long 2688 break; 2689 } 2690 default: 2691 ShouldNotReachHere(); 2692 } 2693 2694 // Build field offset expression. 2695 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2696 // to be plain byte offsets, which are also the same as those accepted 2697 // by oopDesc::field_addr. 2698 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2699 // 32-bit machines ignore the high half of long offsets 2700 offset = ConvL2X(offset); 2701 Node* adr = make_unsafe_address(base, offset, type, false); 2702 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2703 2704 Compile::AliasType* alias_type = C->alias_type(adr_type); 2705 BasicType bt = alias_type->basic_type(); 2706 if (bt != T_ILLEGAL && 2707 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2708 // Don't intrinsify mismatched object accesses. 2709 return false; 2710 } 2711 2712 // For CAS, unlike inline_unsafe_access, there seems no point in 2713 // trying to refine types. Just use the coarse types here. 2714 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2715 const Type *value_type = Type::get_const_basic_type(type); 2716 2717 switch (kind) { 2718 case LS_get_set: 2719 case LS_cmp_exchange: { 2720 if (type == T_OBJECT) { 2721 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2722 if (tjp != NULL) { 2723 value_type = tjp; 2724 } 2725 } 2726 break; 2727 } 2728 case LS_cmp_swap: 2729 case LS_cmp_swap_weak: 2730 case LS_get_add: 2731 break; 2732 default: 2733 ShouldNotReachHere(); 2734 } 2735 2736 // Null check receiver. 2737 receiver = null_check(receiver); 2738 if (stopped()) { 2739 return true; 2740 } 2741 2742 int alias_idx = C->get_alias_index(adr_type); 2743 2744 if (type == T_OBJECT || type == T_ARRAY) { 2745 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF; 2746 2747 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2748 // could be delayed during Parse (for example, in adjust_map_after_if()). 2749 // Execute transformation here to avoid barrier generation in such case. 2750 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2751 newval = _gvn.makecon(TypePtr::NULL_PTR); 2752 2753 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) { 2754 // Refine the value to a null constant, when it is known to be null 2755 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2756 } 2757 } 2758 2759 Node* result = NULL; 2760 switch (kind) { 2761 case LS_cmp_exchange: { 2762 result = access_atomic_cmpxchg_val_at(control(), base, adr, adr_type, alias_idx, 2763 oldval, newval, value_type, type, decorators); 2764 break; 2765 } 2766 case LS_cmp_swap_weak: 2767 decorators |= C2_WEAK_CMPXCHG; 2768 case LS_cmp_swap: { 2769 result = access_atomic_cmpxchg_bool_at(control(), base, adr, adr_type, alias_idx, 2770 oldval, newval, value_type, type, decorators); 2771 break; 2772 } 2773 case LS_get_set: { 2774 result = access_atomic_xchg_at(control(), base, adr, adr_type, alias_idx, 2775 newval, value_type, type, decorators); 2776 break; 2777 } 2778 case LS_get_add: { 2779 result = access_atomic_add_at(control(), base, adr, adr_type, alias_idx, 2780 newval, value_type, type, decorators); 2781 break; 2782 } 2783 default: 2784 ShouldNotReachHere(); 2785 } 2786 2787 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2788 set_result(result); 2789 return true; 2790 } 2791 2792 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 2793 // Regardless of form, don't allow previous ld/st to move down, 2794 // then issue acquire, release, or volatile mem_bar. 2795 insert_mem_bar(Op_MemBarCPUOrder); 2796 switch(id) { 2797 case vmIntrinsics::_loadFence: 2798 insert_mem_bar(Op_LoadFence); 2799 return true; 2800 case vmIntrinsics::_storeFence: 2801 insert_mem_bar(Op_StoreFence); 2802 return true; 2803 case vmIntrinsics::_fullFence: 2804 insert_mem_bar(Op_MemBarVolatile); 2805 return true; 2806 default: 2807 fatal_unexpected_iid(id); 2808 return false; 2809 } 2810 } 2811 2812 bool LibraryCallKit::inline_onspinwait() { 2813 insert_mem_bar(Op_OnSpinWait); 2814 return true; 2815 } 2816 2817 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 2818 if (!kls->is_Con()) { 2819 return true; 2820 } 2821 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 2822 if (klsptr == NULL) { 2823 return true; 2824 } 2825 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 2826 // don't need a guard for a klass that is already initialized 2827 return !ik->is_initialized(); 2828 } 2829 2830 //----------------------------inline_unsafe_allocate--------------------------- 2831 // public native Object Unsafe.allocateInstance(Class<?> cls); 2832 bool LibraryCallKit::inline_unsafe_allocate() { 2833 if (callee()->is_static()) return false; // caller must have the capability! 2834 2835 null_check_receiver(); // null-check, then ignore 2836 Node* cls = null_check(argument(1)); 2837 if (stopped()) return true; 2838 2839 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2840 kls = null_check(kls); 2841 if (stopped()) return true; // argument was like int.class 2842 2843 Node* test = NULL; 2844 if (LibraryCallKit::klass_needs_init_guard(kls)) { 2845 // Note: The argument might still be an illegal value like 2846 // Serializable.class or Object[].class. The runtime will handle it. 2847 // But we must make an explicit check for initialization. 2848 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 2849 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 2850 // can generate code to load it as unsigned byte. 2851 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 2852 Node* bits = intcon(InstanceKlass::fully_initialized); 2853 test = _gvn.transform(new SubINode(inst, bits)); 2854 // The 'test' is non-zero if we need to take a slow path. 2855 } 2856 2857 Node* obj = new_instance(kls, test); 2858 set_result(obj); 2859 return true; 2860 } 2861 2862 //------------------------inline_native_time_funcs-------------- 2863 // inline code for System.currentTimeMillis() and System.nanoTime() 2864 // these have the same type and signature 2865 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 2866 const TypeFunc* tf = OptoRuntime::void_long_Type(); 2867 const TypePtr* no_memory_effects = NULL; 2868 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2869 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 2870 #ifdef ASSERT 2871 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 2872 assert(value_top == top(), "second value must be top"); 2873 #endif 2874 set_result(value); 2875 return true; 2876 } 2877 2878 #ifdef JFR_HAVE_INTRINSICS 2879 2880 /* 2881 * oop -> myklass 2882 * myklass->trace_id |= USED 2883 * return myklass->trace_id & ~0x3 2884 */ 2885 bool LibraryCallKit::inline_native_classID() { 2886 Node* cls = null_check(argument(0), T_OBJECT); 2887 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 2888 kls = null_check(kls, T_OBJECT); 2889 2890 ByteSize offset = KLASS_TRACE_ID_OFFSET; 2891 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 2892 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 2893 2894 Node* clsused = longcon(0x01l); // set the class bit 2895 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 2896 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 2897 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 2898 2899 #ifdef TRACE_ID_META_BITS 2900 Node* mbits = longcon(~TRACE_ID_META_BITS); 2901 tvalue = _gvn.transform(new AndLNode(tvalue, mbits)); 2902 #endif 2903 #ifdef TRACE_ID_SHIFT 2904 Node* cbits = intcon(TRACE_ID_SHIFT); 2905 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits)); 2906 #endif 2907 2908 set_result(tvalue); 2909 return true; 2910 2911 } 2912 2913 bool LibraryCallKit::inline_native_getEventWriter() { 2914 Node* tls_ptr = _gvn.transform(new ThreadLocalNode()); 2915 2916 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr, 2917 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR)); 2918 2919 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered); 2920 2921 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) ); 2922 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) ); 2923 2924 IfNode* iff_jobj_null = 2925 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN); 2926 2927 enum { _normal_path = 1, 2928 _null_path = 2, 2929 PATH_LIMIT }; 2930 2931 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 2932 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM); 2933 2934 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null)); 2935 result_rgn->init_req(_null_path, jobj_is_null); 2936 result_val->init_req(_null_path, null()); 2937 2938 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null)); 2939 set_control(jobj_is_not_null); 2940 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT, 2941 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD); 2942 result_rgn->init_req(_normal_path, control()); 2943 result_val->init_req(_normal_path, res); 2944 2945 set_result(result_rgn, result_val); 2946 2947 return true; 2948 } 2949 2950 #endif // JFR_HAVE_INTRINSICS 2951 2952 //------------------------inline_native_currentThread------------------ 2953 bool LibraryCallKit::inline_native_currentThread() { 2954 Node* junk = NULL; 2955 set_result(generate_current_thread(junk)); 2956 return true; 2957 } 2958 2959 //------------------------inline_native_isInterrupted------------------ 2960 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 2961 bool LibraryCallKit::inline_native_isInterrupted() { 2962 // Add a fast path to t.isInterrupted(clear_int): 2963 // (t == Thread.current() && 2964 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 2965 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 2966 // So, in the common case that the interrupt bit is false, 2967 // we avoid making a call into the VM. Even if the interrupt bit 2968 // is true, if the clear_int argument is false, we avoid the VM call. 2969 // However, if the receiver is not currentThread, we must call the VM, 2970 // because there must be some locking done around the operation. 2971 2972 // We only go to the fast case code if we pass two guards. 2973 // Paths which do not pass are accumulated in the slow_region. 2974 2975 enum { 2976 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 2977 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 2978 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 2979 PATH_LIMIT 2980 }; 2981 2982 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 2983 // out of the function. 2984 insert_mem_bar(Op_MemBarCPUOrder); 2985 2986 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 2987 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 2988 2989 RegionNode* slow_region = new RegionNode(1); 2990 record_for_igvn(slow_region); 2991 2992 // (a) Receiving thread must be the current thread. 2993 Node* rec_thr = argument(0); 2994 Node* tls_ptr = NULL; 2995 Node* cur_thr = generate_current_thread(tls_ptr); 2996 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 2997 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 2998 2999 generate_slow_guard(bol_thr, slow_region); 3000 3001 // (b) Interrupt bit on TLS must be false. 3002 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3003 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3004 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3005 3006 // Set the control input on the field _interrupted read to prevent it floating up. 3007 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3008 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3009 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3010 3011 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3012 3013 // First fast path: if (!TLS._interrupted) return false; 3014 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3015 result_rgn->init_req(no_int_result_path, false_bit); 3016 result_val->init_req(no_int_result_path, intcon(0)); 3017 3018 // drop through to next case 3019 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3020 3021 #ifndef _WINDOWS 3022 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3023 Node* clr_arg = argument(1); 3024 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3025 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3026 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3027 3028 // Second fast path: ... else if (!clear_int) return true; 3029 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3030 result_rgn->init_req(no_clear_result_path, false_arg); 3031 result_val->init_req(no_clear_result_path, intcon(1)); 3032 3033 // drop through to next case 3034 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3035 #else 3036 // To return true on Windows you must read the _interrupted field 3037 // and check the event state i.e. take the slow path. 3038 #endif // _WINDOWS 3039 3040 // (d) Otherwise, go to the slow path. 3041 slow_region->add_req(control()); 3042 set_control( _gvn.transform(slow_region)); 3043 3044 if (stopped()) { 3045 // There is no slow path. 3046 result_rgn->init_req(slow_result_path, top()); 3047 result_val->init_req(slow_result_path, top()); 3048 } else { 3049 // non-virtual because it is a private non-static 3050 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3051 3052 Node* slow_val = set_results_for_java_call(slow_call); 3053 // this->control() comes from set_results_for_java_call 3054 3055 Node* fast_io = slow_call->in(TypeFunc::I_O); 3056 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3057 3058 // These two phis are pre-filled with copies of of the fast IO and Memory 3059 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3060 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3061 3062 result_rgn->init_req(slow_result_path, control()); 3063 result_io ->init_req(slow_result_path, i_o()); 3064 result_mem->init_req(slow_result_path, reset_memory()); 3065 result_val->init_req(slow_result_path, slow_val); 3066 3067 set_all_memory(_gvn.transform(result_mem)); 3068 set_i_o( _gvn.transform(result_io)); 3069 } 3070 3071 C->set_has_split_ifs(true); // Has chance for split-if optimization 3072 set_result(result_rgn, result_val); 3073 return true; 3074 } 3075 3076 //---------------------------load_mirror_from_klass---------------------------- 3077 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3078 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3079 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3080 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3081 // mirror = ((OopHandle)mirror)->resolve(); 3082 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); 3083 } 3084 3085 //-----------------------load_klass_from_mirror_common------------------------- 3086 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3087 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3088 // and branch to the given path on the region. 3089 // If never_see_null, take an uncommon trap on null, so we can optimistically 3090 // compile for the non-null case. 3091 // If the region is NULL, force never_see_null = true. 3092 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3093 bool never_see_null, 3094 RegionNode* region, 3095 int null_path, 3096 int offset) { 3097 if (region == NULL) never_see_null = true; 3098 Node* p = basic_plus_adr(mirror, offset); 3099 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3100 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3101 Node* null_ctl = top(); 3102 kls = null_check_oop(kls, &null_ctl, never_see_null); 3103 if (region != NULL) { 3104 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3105 region->init_req(null_path, null_ctl); 3106 } else { 3107 assert(null_ctl == top(), "no loose ends"); 3108 } 3109 return kls; 3110 } 3111 3112 //--------------------(inline_native_Class_query helpers)--------------------- 3113 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3114 // Fall through if (mods & mask) == bits, take the guard otherwise. 3115 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3116 // Branch around if the given klass has the given modifier bit set. 3117 // Like generate_guard, adds a new path onto the region. 3118 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3119 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3120 Node* mask = intcon(modifier_mask); 3121 Node* bits = intcon(modifier_bits); 3122 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3123 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3124 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3125 return generate_fair_guard(bol, region); 3126 } 3127 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3128 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3129 } 3130 3131 //-------------------------inline_native_Class_query------------------- 3132 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3133 const Type* return_type = TypeInt::BOOL; 3134 Node* prim_return_value = top(); // what happens if it's a primitive class? 3135 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3136 bool expect_prim = false; // most of these guys expect to work on refs 3137 3138 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3139 3140 Node* mirror = argument(0); 3141 Node* obj = top(); 3142 3143 switch (id) { 3144 case vmIntrinsics::_isInstance: 3145 // nothing is an instance of a primitive type 3146 prim_return_value = intcon(0); 3147 obj = argument(1); 3148 break; 3149 case vmIntrinsics::_getModifiers: 3150 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3151 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3152 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3153 break; 3154 case vmIntrinsics::_isInterface: 3155 prim_return_value = intcon(0); 3156 break; 3157 case vmIntrinsics::_isArray: 3158 prim_return_value = intcon(0); 3159 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3160 break; 3161 case vmIntrinsics::_isPrimitive: 3162 prim_return_value = intcon(1); 3163 expect_prim = true; // obviously 3164 break; 3165 case vmIntrinsics::_getSuperclass: 3166 prim_return_value = null(); 3167 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3168 break; 3169 case vmIntrinsics::_getClassAccessFlags: 3170 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3171 return_type = TypeInt::INT; // not bool! 6297094 3172 break; 3173 default: 3174 fatal_unexpected_iid(id); 3175 break; 3176 } 3177 3178 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3179 if (mirror_con == NULL) return false; // cannot happen? 3180 3181 #ifndef PRODUCT 3182 if (C->print_intrinsics() || C->print_inlining()) { 3183 ciType* k = mirror_con->java_mirror_type(); 3184 if (k) { 3185 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3186 k->print_name(); 3187 tty->cr(); 3188 } 3189 } 3190 #endif 3191 3192 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3193 RegionNode* region = new RegionNode(PATH_LIMIT); 3194 record_for_igvn(region); 3195 PhiNode* phi = new PhiNode(region, return_type); 3196 3197 // The mirror will never be null of Reflection.getClassAccessFlags, however 3198 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3199 // if it is. See bug 4774291. 3200 3201 // For Reflection.getClassAccessFlags(), the null check occurs in 3202 // the wrong place; see inline_unsafe_access(), above, for a similar 3203 // situation. 3204 mirror = null_check(mirror); 3205 // If mirror or obj is dead, only null-path is taken. 3206 if (stopped()) return true; 3207 3208 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3209 3210 // Now load the mirror's klass metaobject, and null-check it. 3211 // Side-effects region with the control path if the klass is null. 3212 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3213 // If kls is null, we have a primitive mirror. 3214 phi->init_req(_prim_path, prim_return_value); 3215 if (stopped()) { set_result(region, phi); return true; } 3216 bool safe_for_replace = (region->in(_prim_path) == top()); 3217 3218 Node* p; // handy temp 3219 Node* null_ctl; 3220 3221 // Now that we have the non-null klass, we can perform the real query. 3222 // For constant classes, the query will constant-fold in LoadNode::Value. 3223 Node* query_value = top(); 3224 switch (id) { 3225 case vmIntrinsics::_isInstance: 3226 // nothing is an instance of a primitive type 3227 query_value = gen_instanceof(obj, kls, safe_for_replace); 3228 break; 3229 3230 case vmIntrinsics::_getModifiers: 3231 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3232 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3233 break; 3234 3235 case vmIntrinsics::_isInterface: 3236 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3237 if (generate_interface_guard(kls, region) != NULL) 3238 // A guard was added. If the guard is taken, it was an interface. 3239 phi->add_req(intcon(1)); 3240 // If we fall through, it's a plain class. 3241 query_value = intcon(0); 3242 break; 3243 3244 case vmIntrinsics::_isArray: 3245 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3246 if (generate_array_guard(kls, region) != NULL) 3247 // A guard was added. If the guard is taken, it was an array. 3248 phi->add_req(intcon(1)); 3249 // If we fall through, it's a plain class. 3250 query_value = intcon(0); 3251 break; 3252 3253 case vmIntrinsics::_isPrimitive: 3254 query_value = intcon(0); // "normal" path produces false 3255 break; 3256 3257 case vmIntrinsics::_getSuperclass: 3258 // The rules here are somewhat unfortunate, but we can still do better 3259 // with random logic than with a JNI call. 3260 // Interfaces store null or Object as _super, but must report null. 3261 // Arrays store an intermediate super as _super, but must report Object. 3262 // Other types can report the actual _super. 3263 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3264 if (generate_interface_guard(kls, region) != NULL) 3265 // A guard was added. If the guard is taken, it was an interface. 3266 phi->add_req(null()); 3267 if (generate_array_guard(kls, region) != NULL) 3268 // A guard was added. If the guard is taken, it was an array. 3269 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3270 // If we fall through, it's a plain class. Get its _super. 3271 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3272 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3273 null_ctl = top(); 3274 kls = null_check_oop(kls, &null_ctl); 3275 if (null_ctl != top()) { 3276 // If the guard is taken, Object.superClass is null (both klass and mirror). 3277 region->add_req(null_ctl); 3278 phi ->add_req(null()); 3279 } 3280 if (!stopped()) { 3281 query_value = load_mirror_from_klass(kls); 3282 } 3283 break; 3284 3285 case vmIntrinsics::_getClassAccessFlags: 3286 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3287 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3288 break; 3289 3290 default: 3291 fatal_unexpected_iid(id); 3292 break; 3293 } 3294 3295 // Fall-through is the normal case of a query to a real class. 3296 phi->init_req(1, query_value); 3297 region->init_req(1, control()); 3298 3299 C->set_has_split_ifs(true); // Has chance for split-if optimization 3300 set_result(region, phi); 3301 return true; 3302 } 3303 3304 //-------------------------inline_Class_cast------------------- 3305 bool LibraryCallKit::inline_Class_cast() { 3306 Node* mirror = argument(0); // Class 3307 Node* obj = argument(1); 3308 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3309 if (mirror_con == NULL) { 3310 return false; // dead path (mirror->is_top()). 3311 } 3312 if (obj == NULL || obj->is_top()) { 3313 return false; // dead path 3314 } 3315 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3316 3317 // First, see if Class.cast() can be folded statically. 3318 // java_mirror_type() returns non-null for compile-time Class constants. 3319 ciType* tm = mirror_con->java_mirror_type(); 3320 if (tm != NULL && tm->is_klass() && 3321 tp != NULL && tp->klass() != NULL) { 3322 if (!tp->klass()->is_loaded()) { 3323 // Don't use intrinsic when class is not loaded. 3324 return false; 3325 } else { 3326 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3327 if (static_res == Compile::SSC_always_true) { 3328 // isInstance() is true - fold the code. 3329 set_result(obj); 3330 return true; 3331 } else if (static_res == Compile::SSC_always_false) { 3332 // Don't use intrinsic, have to throw ClassCastException. 3333 // If the reference is null, the non-intrinsic bytecode will 3334 // be optimized appropriately. 3335 return false; 3336 } 3337 } 3338 } 3339 3340 // Bailout intrinsic and do normal inlining if exception path is frequent. 3341 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3342 return false; 3343 } 3344 3345 // Generate dynamic checks. 3346 // Class.cast() is java implementation of _checkcast bytecode. 3347 // Do checkcast (Parse::do_checkcast()) optimizations here. 3348 3349 mirror = null_check(mirror); 3350 // If mirror is dead, only null-path is taken. 3351 if (stopped()) { 3352 return true; 3353 } 3354 3355 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3356 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3357 RegionNode* region = new RegionNode(PATH_LIMIT); 3358 record_for_igvn(region); 3359 3360 // Now load the mirror's klass metaobject, and null-check it. 3361 // If kls is null, we have a primitive mirror and 3362 // nothing is an instance of a primitive type. 3363 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3364 3365 Node* res = top(); 3366 if (!stopped()) { 3367 Node* bad_type_ctrl = top(); 3368 // Do checkcast optimizations. 3369 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3370 region->init_req(_bad_type_path, bad_type_ctrl); 3371 } 3372 if (region->in(_prim_path) != top() || 3373 region->in(_bad_type_path) != top()) { 3374 // Let Interpreter throw ClassCastException. 3375 PreserveJVMState pjvms(this); 3376 set_control(_gvn.transform(region)); 3377 uncommon_trap(Deoptimization::Reason_intrinsic, 3378 Deoptimization::Action_maybe_recompile); 3379 } 3380 if (!stopped()) { 3381 set_result(res); 3382 } 3383 return true; 3384 } 3385 3386 3387 //--------------------------inline_native_subtype_check------------------------ 3388 // This intrinsic takes the JNI calls out of the heart of 3389 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3390 bool LibraryCallKit::inline_native_subtype_check() { 3391 // Pull both arguments off the stack. 3392 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3393 args[0] = argument(0); 3394 args[1] = argument(1); 3395 Node* klasses[2]; // corresponding Klasses: superk, subk 3396 klasses[0] = klasses[1] = top(); 3397 3398 enum { 3399 // A full decision tree on {superc is prim, subc is prim}: 3400 _prim_0_path = 1, // {P,N} => false 3401 // {P,P} & superc!=subc => false 3402 _prim_same_path, // {P,P} & superc==subc => true 3403 _prim_1_path, // {N,P} => false 3404 _ref_subtype_path, // {N,N} & subtype check wins => true 3405 _both_ref_path, // {N,N} & subtype check loses => false 3406 PATH_LIMIT 3407 }; 3408 3409 RegionNode* region = new RegionNode(PATH_LIMIT); 3410 Node* phi = new PhiNode(region, TypeInt::BOOL); 3411 record_for_igvn(region); 3412 3413 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3414 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3415 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3416 3417 // First null-check both mirrors and load each mirror's klass metaobject. 3418 int which_arg; 3419 for (which_arg = 0; which_arg <= 1; which_arg++) { 3420 Node* arg = args[which_arg]; 3421 arg = null_check(arg); 3422 if (stopped()) break; 3423 args[which_arg] = arg; 3424 3425 Node* p = basic_plus_adr(arg, class_klass_offset); 3426 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3427 klasses[which_arg] = _gvn.transform(kls); 3428 } 3429 3430 // Having loaded both klasses, test each for null. 3431 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3432 for (which_arg = 0; which_arg <= 1; which_arg++) { 3433 Node* kls = klasses[which_arg]; 3434 Node* null_ctl = top(); 3435 kls = null_check_oop(kls, &null_ctl, never_see_null); 3436 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3437 region->init_req(prim_path, null_ctl); 3438 if (stopped()) break; 3439 klasses[which_arg] = kls; 3440 } 3441 3442 if (!stopped()) { 3443 // now we have two reference types, in klasses[0..1] 3444 Node* subk = klasses[1]; // the argument to isAssignableFrom 3445 Node* superk = klasses[0]; // the receiver 3446 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3447 // now we have a successful reference subtype check 3448 region->set_req(_ref_subtype_path, control()); 3449 } 3450 3451 // If both operands are primitive (both klasses null), then 3452 // we must return true when they are identical primitives. 3453 // It is convenient to test this after the first null klass check. 3454 set_control(region->in(_prim_0_path)); // go back to first null check 3455 if (!stopped()) { 3456 // Since superc is primitive, make a guard for the superc==subc case. 3457 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3458 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3459 generate_guard(bol_eq, region, PROB_FAIR); 3460 if (region->req() == PATH_LIMIT+1) { 3461 // A guard was added. If the added guard is taken, superc==subc. 3462 region->swap_edges(PATH_LIMIT, _prim_same_path); 3463 region->del_req(PATH_LIMIT); 3464 } 3465 region->set_req(_prim_0_path, control()); // Not equal after all. 3466 } 3467 3468 // these are the only paths that produce 'true': 3469 phi->set_req(_prim_same_path, intcon(1)); 3470 phi->set_req(_ref_subtype_path, intcon(1)); 3471 3472 // pull together the cases: 3473 assert(region->req() == PATH_LIMIT, "sane region"); 3474 for (uint i = 1; i < region->req(); i++) { 3475 Node* ctl = region->in(i); 3476 if (ctl == NULL || ctl == top()) { 3477 region->set_req(i, top()); 3478 phi ->set_req(i, top()); 3479 } else if (phi->in(i) == NULL) { 3480 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3481 } 3482 } 3483 3484 set_control(_gvn.transform(region)); 3485 set_result(_gvn.transform(phi)); 3486 return true; 3487 } 3488 3489 //---------------------generate_array_guard_common------------------------ 3490 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3491 bool obj_array, bool not_array) { 3492 3493 if (stopped()) { 3494 return NULL; 3495 } 3496 3497 // If obj_array/non_array==false/false: 3498 // Branch around if the given klass is in fact an array (either obj or prim). 3499 // If obj_array/non_array==false/true: 3500 // Branch around if the given klass is not an array klass of any kind. 3501 // If obj_array/non_array==true/true: 3502 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3503 // If obj_array/non_array==true/false: 3504 // Branch around if the kls is an oop array (Object[] or subtype) 3505 // 3506 // Like generate_guard, adds a new path onto the region. 3507 jint layout_con = 0; 3508 Node* layout_val = get_layout_helper(kls, layout_con); 3509 if (layout_val == NULL) { 3510 bool query = (obj_array 3511 ? Klass::layout_helper_is_objArray(layout_con) 3512 : Klass::layout_helper_is_array(layout_con)); 3513 if (query == not_array) { 3514 return NULL; // never a branch 3515 } else { // always a branch 3516 Node* always_branch = control(); 3517 if (region != NULL) 3518 region->add_req(always_branch); 3519 set_control(top()); 3520 return always_branch; 3521 } 3522 } 3523 // Now test the correct condition. 3524 jint nval = (obj_array 3525 ? (jint)(Klass::_lh_array_tag_type_value 3526 << Klass::_lh_array_tag_shift) 3527 : Klass::_lh_neutral_value); 3528 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3529 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3530 // invert the test if we are looking for a non-array 3531 if (not_array) btest = BoolTest(btest).negate(); 3532 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3533 return generate_fair_guard(bol, region); 3534 } 3535 3536 3537 //-----------------------inline_native_newArray-------------------------- 3538 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3539 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 3540 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 3541 Node* mirror; 3542 Node* count_val; 3543 if (uninitialized) { 3544 mirror = argument(1); 3545 count_val = argument(2); 3546 } else { 3547 mirror = argument(0); 3548 count_val = argument(1); 3549 } 3550 3551 mirror = null_check(mirror); 3552 // If mirror or obj is dead, only null-path is taken. 3553 if (stopped()) return true; 3554 3555 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3556 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3557 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3558 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3559 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3560 3561 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3562 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3563 result_reg, _slow_path); 3564 Node* normal_ctl = control(); 3565 Node* no_array_ctl = result_reg->in(_slow_path); 3566 3567 // Generate code for the slow case. We make a call to newArray(). 3568 set_control(no_array_ctl); 3569 if (!stopped()) { 3570 // Either the input type is void.class, or else the 3571 // array klass has not yet been cached. Either the 3572 // ensuing call will throw an exception, or else it 3573 // will cache the array klass for next time. 3574 PreserveJVMState pjvms(this); 3575 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3576 Node* slow_result = set_results_for_java_call(slow_call); 3577 // this->control() comes from set_results_for_java_call 3578 result_reg->set_req(_slow_path, control()); 3579 result_val->set_req(_slow_path, slow_result); 3580 result_io ->set_req(_slow_path, i_o()); 3581 result_mem->set_req(_slow_path, reset_memory()); 3582 } 3583 3584 set_control(normal_ctl); 3585 if (!stopped()) { 3586 // Normal case: The array type has been cached in the java.lang.Class. 3587 // The following call works fine even if the array type is polymorphic. 3588 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3589 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3590 result_reg->init_req(_normal_path, control()); 3591 result_val->init_req(_normal_path, obj); 3592 result_io ->init_req(_normal_path, i_o()); 3593 result_mem->init_req(_normal_path, reset_memory()); 3594 3595 if (uninitialized) { 3596 // Mark the allocation so that zeroing is skipped 3597 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 3598 alloc->maybe_set_complete(&_gvn); 3599 } 3600 } 3601 3602 // Return the combined state. 3603 set_i_o( _gvn.transform(result_io) ); 3604 set_all_memory( _gvn.transform(result_mem)); 3605 3606 C->set_has_split_ifs(true); // Has chance for split-if optimization 3607 set_result(result_reg, result_val); 3608 return true; 3609 } 3610 3611 //----------------------inline_native_getLength-------------------------- 3612 // public static native int java.lang.reflect.Array.getLength(Object array); 3613 bool LibraryCallKit::inline_native_getLength() { 3614 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3615 3616 Node* array = null_check(argument(0)); 3617 // If array is dead, only null-path is taken. 3618 if (stopped()) return true; 3619 3620 // Deoptimize if it is a non-array. 3621 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3622 3623 if (non_array != NULL) { 3624 PreserveJVMState pjvms(this); 3625 set_control(non_array); 3626 uncommon_trap(Deoptimization::Reason_intrinsic, 3627 Deoptimization::Action_maybe_recompile); 3628 } 3629 3630 // If control is dead, only non-array-path is taken. 3631 if (stopped()) return true; 3632 3633 // The works fine even if the array type is polymorphic. 3634 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3635 Node* result = load_array_length(array); 3636 3637 C->set_has_split_ifs(true); // Has chance for split-if optimization 3638 set_result(result); 3639 return true; 3640 } 3641 3642 //------------------------inline_array_copyOf---------------------------- 3643 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3644 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3645 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3646 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3647 3648 // Get the arguments. 3649 Node* original = argument(0); 3650 Node* start = is_copyOfRange? argument(1): intcon(0); 3651 Node* end = is_copyOfRange? argument(2): argument(1); 3652 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3653 3654 Node* newcopy = NULL; 3655 3656 // Set the original stack and the reexecute bit for the interpreter to reexecute 3657 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3658 { PreserveReexecuteState preexecs(this); 3659 jvms()->set_should_reexecute(true); 3660 3661 array_type_mirror = null_check(array_type_mirror); 3662 original = null_check(original); 3663 3664 // Check if a null path was taken unconditionally. 3665 if (stopped()) return true; 3666 3667 Node* orig_length = load_array_length(original); 3668 3669 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3670 klass_node = null_check(klass_node); 3671 3672 RegionNode* bailout = new RegionNode(1); 3673 record_for_igvn(bailout); 3674 3675 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3676 // Bail out if that is so. 3677 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3678 if (not_objArray != NULL) { 3679 // Improve the klass node's type from the new optimistic assumption: 3680 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3681 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3682 Node* cast = new CastPPNode(klass_node, akls); 3683 cast->init_req(0, control()); 3684 klass_node = _gvn.transform(cast); 3685 } 3686 3687 // Bail out if either start or end is negative. 3688 generate_negative_guard(start, bailout, &start); 3689 generate_negative_guard(end, bailout, &end); 3690 3691 Node* length = end; 3692 if (_gvn.type(start) != TypeInt::ZERO) { 3693 length = _gvn.transform(new SubINode(end, start)); 3694 } 3695 3696 // Bail out if length is negative. 3697 // Without this the new_array would throw 3698 // NegativeArraySizeException but IllegalArgumentException is what 3699 // should be thrown 3700 generate_negative_guard(length, bailout, &length); 3701 3702 if (bailout->req() > 1) { 3703 PreserveJVMState pjvms(this); 3704 set_control(_gvn.transform(bailout)); 3705 uncommon_trap(Deoptimization::Reason_intrinsic, 3706 Deoptimization::Action_maybe_recompile); 3707 } 3708 3709 if (!stopped()) { 3710 // How many elements will we copy from the original? 3711 // The answer is MinI(orig_length - start, length). 3712 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3713 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3714 3715 // Generate a direct call to the right arraycopy function(s). 3716 // We know the copy is disjoint but we might not know if the 3717 // oop stores need checking. 3718 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3719 // This will fail a store-check if x contains any non-nulls. 3720 3721 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3722 // loads/stores but it is legal only if we're sure the 3723 // Arrays.copyOf would succeed. So we need all input arguments 3724 // to the copyOf to be validated, including that the copy to the 3725 // new array won't trigger an ArrayStoreException. That subtype 3726 // check can be optimized if we know something on the type of 3727 // the input array from type speculation. 3728 if (_gvn.type(klass_node)->singleton()) { 3729 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3730 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3731 3732 int test = C->static_subtype_check(superk, subk); 3733 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3734 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3735 if (t_original->speculative_type() != NULL) { 3736 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3737 } 3738 } 3739 } 3740 3741 bool validated = false; 3742 // Reason_class_check rather than Reason_intrinsic because we 3743 // want to intrinsify even if this traps. 3744 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3745 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3746 klass_node); 3747 3748 if (not_subtype_ctrl != top()) { 3749 PreserveJVMState pjvms(this); 3750 set_control(not_subtype_ctrl); 3751 uncommon_trap(Deoptimization::Reason_class_check, 3752 Deoptimization::Action_make_not_entrant); 3753 assert(stopped(), "Should be stopped"); 3754 } 3755 validated = true; 3756 } 3757 3758 if (!stopped()) { 3759 newcopy = new_array(klass_node, length, 0); // no arguments to push 3760 3761 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false, 3762 load_object_klass(original), klass_node); 3763 if (!is_copyOfRange) { 3764 ac->set_copyof(validated); 3765 } else { 3766 ac->set_copyofrange(validated); 3767 } 3768 Node* n = _gvn.transform(ac); 3769 if (n == ac) { 3770 ac->connect_outputs(this); 3771 } else { 3772 assert(validated, "shouldn't transform if all arguments not validated"); 3773 set_all_memory(n); 3774 } 3775 } 3776 } 3777 } // original reexecute is set back here 3778 3779 C->set_has_split_ifs(true); // Has chance for split-if optimization 3780 if (!stopped()) { 3781 set_result(newcopy); 3782 } 3783 return true; 3784 } 3785 3786 3787 //----------------------generate_virtual_guard--------------------------- 3788 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3789 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3790 RegionNode* slow_region) { 3791 ciMethod* method = callee(); 3792 int vtable_index = method->vtable_index(); 3793 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3794 "bad index %d", vtable_index); 3795 // Get the Method* out of the appropriate vtable entry. 3796 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 3797 vtable_index*vtableEntry::size_in_bytes() + 3798 vtableEntry::method_offset_in_bytes(); 3799 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3800 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3801 3802 // Compare the target method with the expected method (e.g., Object.hashCode). 3803 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3804 3805 Node* native_call = makecon(native_call_addr); 3806 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3807 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3808 3809 return generate_slow_guard(test_native, slow_region); 3810 } 3811 3812 //-----------------------generate_method_call---------------------------- 3813 // Use generate_method_call to make a slow-call to the real 3814 // method if the fast path fails. An alternative would be to 3815 // use a stub like OptoRuntime::slow_arraycopy_Java. 3816 // This only works for expanding the current library call, 3817 // not another intrinsic. (E.g., don't use this for making an 3818 // arraycopy call inside of the copyOf intrinsic.) 3819 CallJavaNode* 3820 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3821 // When compiling the intrinsic method itself, do not use this technique. 3822 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3823 3824 ciMethod* method = callee(); 3825 // ensure the JVMS we have will be correct for this call 3826 guarantee(method_id == method->intrinsic_id(), "must match"); 3827 3828 const TypeFunc* tf = TypeFunc::make(method); 3829 CallJavaNode* slow_call; 3830 if (is_static) { 3831 assert(!is_virtual, ""); 3832 slow_call = new CallStaticJavaNode(C, tf, 3833 SharedRuntime::get_resolve_static_call_stub(), 3834 method, bci()); 3835 } else if (is_virtual) { 3836 null_check_receiver(); 3837 int vtable_index = Method::invalid_vtable_index; 3838 if (UseInlineCaches) { 3839 // Suppress the vtable call 3840 } else { 3841 // hashCode and clone are not a miranda methods, 3842 // so the vtable index is fixed. 3843 // No need to use the linkResolver to get it. 3844 vtable_index = method->vtable_index(); 3845 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3846 "bad index %d", vtable_index); 3847 } 3848 slow_call = new CallDynamicJavaNode(tf, 3849 SharedRuntime::get_resolve_virtual_call_stub(), 3850 method, vtable_index, bci()); 3851 } else { // neither virtual nor static: opt_virtual 3852 null_check_receiver(); 3853 slow_call = new CallStaticJavaNode(C, tf, 3854 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3855 method, bci()); 3856 slow_call->set_optimized_virtual(true); 3857 } 3858 set_arguments_for_java_call(slow_call); 3859 set_edges_for_java_call(slow_call); 3860 return slow_call; 3861 } 3862 3863 3864 /** 3865 * Build special case code for calls to hashCode on an object. This call may 3866 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 3867 * slightly different code. 3868 */ 3869 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3870 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3871 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3872 3873 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3874 3875 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3876 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 3877 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3878 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3879 Node* obj = NULL; 3880 if (!is_static) { 3881 // Check for hashing null object 3882 obj = null_check_receiver(); 3883 if (stopped()) return true; // unconditionally null 3884 result_reg->init_req(_null_path, top()); 3885 result_val->init_req(_null_path, top()); 3886 } else { 3887 // Do a null check, and return zero if null. 3888 // System.identityHashCode(null) == 0 3889 obj = argument(0); 3890 Node* null_ctl = top(); 3891 obj = null_check_oop(obj, &null_ctl); 3892 result_reg->init_req(_null_path, null_ctl); 3893 result_val->init_req(_null_path, _gvn.intcon(0)); 3894 } 3895 3896 // Unconditionally null? Then return right away. 3897 if (stopped()) { 3898 set_control( result_reg->in(_null_path)); 3899 if (!stopped()) 3900 set_result(result_val->in(_null_path)); 3901 return true; 3902 } 3903 3904 // We only go to the fast case code if we pass a number of guards. The 3905 // paths which do not pass are accumulated in the slow_region. 3906 RegionNode* slow_region = new RegionNode(1); 3907 record_for_igvn(slow_region); 3908 3909 // If this is a virtual call, we generate a funny guard. We pull out 3910 // the vtable entry corresponding to hashCode() from the target object. 3911 // If the target method which we are calling happens to be the native 3912 // Object hashCode() method, we pass the guard. We do not need this 3913 // guard for non-virtual calls -- the caller is known to be the native 3914 // Object hashCode(). 3915 if (is_virtual) { 3916 // After null check, get the object's klass. 3917 Node* obj_klass = load_object_klass(obj); 3918 generate_virtual_guard(obj_klass, slow_region); 3919 } 3920 3921 // Get the header out of the object, use LoadMarkNode when available 3922 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3923 // The control of the load must be NULL. Otherwise, the load can move before 3924 // the null check after castPP removal. 3925 Node* no_ctrl = NULL; 3926 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3927 3928 // Test the header to see if it is unlocked. 3929 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3930 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 3931 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3932 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 3933 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 3934 3935 generate_slow_guard(test_unlocked, slow_region); 3936 3937 // Get the hash value and check to see that it has been properly assigned. 3938 // We depend on hash_mask being at most 32 bits and avoid the use of 3939 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3940 // vm: see markOop.hpp. 3941 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3942 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3943 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 3944 // This hack lets the hash bits live anywhere in the mark object now, as long 3945 // as the shift drops the relevant bits into the low 32 bits. Note that 3946 // Java spec says that HashCode is an int so there's no point in capturing 3947 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3948 hshifted_header = ConvX2I(hshifted_header); 3949 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 3950 3951 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3952 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 3953 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 3954 3955 generate_slow_guard(test_assigned, slow_region); 3956 3957 Node* init_mem = reset_memory(); 3958 // fill in the rest of the null path: 3959 result_io ->init_req(_null_path, i_o()); 3960 result_mem->init_req(_null_path, init_mem); 3961 3962 result_val->init_req(_fast_path, hash_val); 3963 result_reg->init_req(_fast_path, control()); 3964 result_io ->init_req(_fast_path, i_o()); 3965 result_mem->init_req(_fast_path, init_mem); 3966 3967 // Generate code for the slow case. We make a call to hashCode(). 3968 set_control(_gvn.transform(slow_region)); 3969 if (!stopped()) { 3970 // No need for PreserveJVMState, because we're using up the present state. 3971 set_all_memory(init_mem); 3972 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 3973 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3974 Node* slow_result = set_results_for_java_call(slow_call); 3975 // this->control() comes from set_results_for_java_call 3976 result_reg->init_req(_slow_path, control()); 3977 result_val->init_req(_slow_path, slow_result); 3978 result_io ->set_req(_slow_path, i_o()); 3979 result_mem ->set_req(_slow_path, reset_memory()); 3980 } 3981 3982 // Return the combined state. 3983 set_i_o( _gvn.transform(result_io) ); 3984 set_all_memory( _gvn.transform(result_mem)); 3985 3986 set_result(result_reg, result_val); 3987 return true; 3988 } 3989 3990 //---------------------------inline_native_getClass---------------------------- 3991 // public final native Class<?> java.lang.Object.getClass(); 3992 // 3993 // Build special case code for calls to getClass on an object. 3994 bool LibraryCallKit::inline_native_getClass() { 3995 Node* obj = null_check_receiver(); 3996 if (stopped()) return true; 3997 set_result(load_mirror_from_klass(load_object_klass(obj))); 3998 return true; 3999 } 4000 4001 //-----------------inline_native_Reflection_getCallerClass--------------------- 4002 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4003 // 4004 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4005 // 4006 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4007 // in that it must skip particular security frames and checks for 4008 // caller sensitive methods. 4009 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4010 #ifndef PRODUCT 4011 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4012 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4013 } 4014 #endif 4015 4016 if (!jvms()->has_method()) { 4017 #ifndef PRODUCT 4018 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4019 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4020 } 4021 #endif 4022 return false; 4023 } 4024 4025 // Walk back up the JVM state to find the caller at the required 4026 // depth. 4027 JVMState* caller_jvms = jvms(); 4028 4029 // Cf. JVM_GetCallerClass 4030 // NOTE: Start the loop at depth 1 because the current JVM state does 4031 // not include the Reflection.getCallerClass() frame. 4032 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4033 ciMethod* m = caller_jvms->method(); 4034 switch (n) { 4035 case 0: 4036 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4037 break; 4038 case 1: 4039 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4040 if (!m->caller_sensitive()) { 4041 #ifndef PRODUCT 4042 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4043 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4044 } 4045 #endif 4046 return false; // bail-out; let JVM_GetCallerClass do the work 4047 } 4048 break; 4049 default: 4050 if (!m->is_ignored_by_security_stack_walk()) { 4051 // We have reached the desired frame; return the holder class. 4052 // Acquire method holder as java.lang.Class and push as constant. 4053 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4054 ciInstance* caller_mirror = caller_klass->java_mirror(); 4055 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4056 4057 #ifndef PRODUCT 4058 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4059 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()); 4060 tty->print_cr(" JVM state at this point:"); 4061 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4062 ciMethod* m = jvms()->of_depth(i)->method(); 4063 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4064 } 4065 } 4066 #endif 4067 return true; 4068 } 4069 break; 4070 } 4071 } 4072 4073 #ifndef PRODUCT 4074 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4075 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4076 tty->print_cr(" JVM state at this point:"); 4077 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4078 ciMethod* m = jvms()->of_depth(i)->method(); 4079 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4080 } 4081 } 4082 #endif 4083 4084 return false; // bail-out; let JVM_GetCallerClass do the work 4085 } 4086 4087 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4088 Node* arg = argument(0); 4089 Node* result = NULL; 4090 4091 switch (id) { 4092 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4093 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4094 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4095 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4096 4097 case vmIntrinsics::_doubleToLongBits: { 4098 // two paths (plus control) merge in a wood 4099 RegionNode *r = new RegionNode(3); 4100 Node *phi = new PhiNode(r, TypeLong::LONG); 4101 4102 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4103 // Build the boolean node 4104 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4105 4106 // Branch either way. 4107 // NaN case is less traveled, which makes all the difference. 4108 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4109 Node *opt_isnan = _gvn.transform(ifisnan); 4110 assert( opt_isnan->is_If(), "Expect an IfNode"); 4111 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4112 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4113 4114 set_control(iftrue); 4115 4116 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4117 Node *slow_result = longcon(nan_bits); // return NaN 4118 phi->init_req(1, _gvn.transform( slow_result )); 4119 r->init_req(1, iftrue); 4120 4121 // Else fall through 4122 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4123 set_control(iffalse); 4124 4125 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4126 r->init_req(2, iffalse); 4127 4128 // Post merge 4129 set_control(_gvn.transform(r)); 4130 record_for_igvn(r); 4131 4132 C->set_has_split_ifs(true); // Has chance for split-if optimization 4133 result = phi; 4134 assert(result->bottom_type()->isa_long(), "must be"); 4135 break; 4136 } 4137 4138 case vmIntrinsics::_floatToIntBits: { 4139 // two paths (plus control) merge in a wood 4140 RegionNode *r = new RegionNode(3); 4141 Node *phi = new PhiNode(r, TypeInt::INT); 4142 4143 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4144 // Build the boolean node 4145 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4146 4147 // Branch either way. 4148 // NaN case is less traveled, which makes all the difference. 4149 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4150 Node *opt_isnan = _gvn.transform(ifisnan); 4151 assert( opt_isnan->is_If(), "Expect an IfNode"); 4152 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4153 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4154 4155 set_control(iftrue); 4156 4157 static const jint nan_bits = 0x7fc00000; 4158 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4159 phi->init_req(1, _gvn.transform( slow_result )); 4160 r->init_req(1, iftrue); 4161 4162 // Else fall through 4163 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4164 set_control(iffalse); 4165 4166 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4167 r->init_req(2, iffalse); 4168 4169 // Post merge 4170 set_control(_gvn.transform(r)); 4171 record_for_igvn(r); 4172 4173 C->set_has_split_ifs(true); // Has chance for split-if optimization 4174 result = phi; 4175 assert(result->bottom_type()->isa_int(), "must be"); 4176 break; 4177 } 4178 4179 default: 4180 fatal_unexpected_iid(id); 4181 break; 4182 } 4183 set_result(_gvn.transform(result)); 4184 return true; 4185 } 4186 4187 //----------------------inline_unsafe_copyMemory------------------------- 4188 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4189 bool LibraryCallKit::inline_unsafe_copyMemory() { 4190 if (callee()->is_static()) return false; // caller must have the capability! 4191 null_check_receiver(); // null-check receiver 4192 if (stopped()) return true; 4193 4194 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4195 4196 Node* src_ptr = argument(1); // type: oop 4197 Node* src_off = ConvL2X(argument(2)); // type: long 4198 Node* dst_ptr = argument(4); // type: oop 4199 Node* dst_off = ConvL2X(argument(5)); // type: long 4200 Node* size = ConvL2X(argument(7)); // type: long 4201 4202 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4203 "fieldOffset must be byte-scaled"); 4204 4205 Node* src = make_unsafe_address(src_ptr, src_off); 4206 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4207 4208 // Conservatively insert a memory barrier on all memory slices. 4209 // Do not let writes of the copy source or destination float below the copy. 4210 insert_mem_bar(Op_MemBarCPUOrder); 4211 4212 // Call it. Note that the length argument is not scaled. 4213 make_runtime_call(RC_LEAF|RC_NO_FP, 4214 OptoRuntime::fast_arraycopy_Type(), 4215 StubRoutines::unsafe_arraycopy(), 4216 "unsafe_arraycopy", 4217 TypeRawPtr::BOTTOM, 4218 src, dst, size XTOP); 4219 4220 // Do not let reads of the copy destination float above the copy. 4221 insert_mem_bar(Op_MemBarCPUOrder); 4222 4223 return true; 4224 } 4225 4226 //------------------------clone_coping----------------------------------- 4227 // Helper function for inline_native_clone. 4228 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) { 4229 assert(obj_size != NULL, ""); 4230 Node* raw_obj = alloc_obj->in(1); 4231 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4232 4233 AllocateNode* alloc = NULL; 4234 if (ReduceBulkZeroing) { 4235 // We will be completely responsible for initializing this object - 4236 // mark Initialize node as complete. 4237 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4238 // The object was just allocated - there should be no any stores! 4239 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4240 // Mark as complete_with_arraycopy so that on AllocateNode 4241 // expansion, we know this AllocateNode is initialized by an array 4242 // copy and a StoreStore barrier exists after the array copy. 4243 alloc->initialization()->set_complete_with_arraycopy(); 4244 } 4245 4246 // Copy the fastest available way. 4247 // TODO: generate fields copies for small objects instead. 4248 Node* size = _gvn.transform(obj_size); 4249 4250 access_clone(control(), obj, alloc_obj, size, is_array); 4251 4252 // Do not let reads from the cloned object float above the arraycopy. 4253 if (alloc != NULL) { 4254 // Do not let stores that initialize this object be reordered with 4255 // a subsequent store that would make this object accessible by 4256 // other threads. 4257 // Record what AllocateNode this StoreStore protects so that 4258 // escape analysis can go from the MemBarStoreStoreNode to the 4259 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4260 // based on the escape status of the AllocateNode. 4261 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 4262 } else { 4263 insert_mem_bar(Op_MemBarCPUOrder); 4264 } 4265 } 4266 4267 //------------------------inline_native_clone---------------------------- 4268 // protected native Object java.lang.Object.clone(); 4269 // 4270 // Here are the simple edge cases: 4271 // null receiver => normal trap 4272 // virtual and clone was overridden => slow path to out-of-line clone 4273 // not cloneable or finalizer => slow path to out-of-line Object.clone 4274 // 4275 // The general case has two steps, allocation and copying. 4276 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4277 // 4278 // Copying also has two cases, oop arrays and everything else. 4279 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4280 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4281 // 4282 // These steps fold up nicely if and when the cloned object's klass 4283 // can be sharply typed as an object array, a type array, or an instance. 4284 // 4285 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4286 PhiNode* result_val; 4287 4288 // Set the reexecute bit for the interpreter to reexecute 4289 // the bytecode that invokes Object.clone if deoptimization happens. 4290 { PreserveReexecuteState preexecs(this); 4291 jvms()->set_should_reexecute(true); 4292 4293 Node* obj = null_check_receiver(); 4294 if (stopped()) return true; 4295 4296 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4297 4298 // If we are going to clone an instance, we need its exact type to 4299 // know the number and types of fields to convert the clone to 4300 // loads/stores. Maybe a speculative type can help us. 4301 if (!obj_type->klass_is_exact() && 4302 obj_type->speculative_type() != NULL && 4303 obj_type->speculative_type()->is_instance_klass()) { 4304 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4305 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4306 !spec_ik->has_injected_fields()) { 4307 ciKlass* k = obj_type->klass(); 4308 if (!k->is_instance_klass() || 4309 k->as_instance_klass()->is_interface() || 4310 k->as_instance_klass()->has_subklass()) { 4311 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4312 } 4313 } 4314 } 4315 4316 Node* obj_klass = load_object_klass(obj); 4317 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4318 const TypeOopPtr* toop = ((tklass != NULL) 4319 ? tklass->as_instance_type() 4320 : TypeInstPtr::NOTNULL); 4321 4322 // Conservatively insert a memory barrier on all memory slices. 4323 // Do not let writes into the original float below the clone. 4324 insert_mem_bar(Op_MemBarCPUOrder); 4325 4326 // paths into result_reg: 4327 enum { 4328 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4329 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4330 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4331 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4332 PATH_LIMIT 4333 }; 4334 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4335 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4336 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4337 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4338 record_for_igvn(result_reg); 4339 4340 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4341 if (array_ctl != NULL) { 4342 // It's an array. 4343 PreserveJVMState pjvms(this); 4344 set_control(array_ctl); 4345 Node* obj_length = load_array_length(obj); 4346 Node* obj_size = NULL; 4347 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4348 4349 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4350 if (bs->array_copy_requires_gc_barriers(T_OBJECT)) { 4351 // If it is an oop array, it requires very special treatment, 4352 // because gc barriers are required when accessing the array. 4353 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4354 if (is_obja != NULL) { 4355 PreserveJVMState pjvms2(this); 4356 set_control(is_obja); 4357 // Generate a direct call to the right arraycopy function(s). 4358 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4359 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false); 4360 ac->set_cloneoop(); 4361 Node* n = _gvn.transform(ac); 4362 assert(n == ac, "cannot disappear"); 4363 ac->connect_outputs(this); 4364 4365 result_reg->init_req(_objArray_path, control()); 4366 result_val->init_req(_objArray_path, alloc_obj); 4367 result_i_o ->set_req(_objArray_path, i_o()); 4368 result_mem ->set_req(_objArray_path, reset_memory()); 4369 } 4370 } 4371 // Otherwise, there are no barriers to worry about. 4372 // (We can dispense with card marks if we know the allocation 4373 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4374 // causes the non-eden paths to take compensating steps to 4375 // simulate a fresh allocation, so that no further 4376 // card marks are required in compiled code to initialize 4377 // the object.) 4378 4379 if (!stopped()) { 4380 copy_to_clone(obj, alloc_obj, obj_size, true); 4381 4382 // Present the results of the copy. 4383 result_reg->init_req(_array_path, control()); 4384 result_val->init_req(_array_path, alloc_obj); 4385 result_i_o ->set_req(_array_path, i_o()); 4386 result_mem ->set_req(_array_path, reset_memory()); 4387 } 4388 } 4389 4390 // We only go to the instance fast case code if we pass a number of guards. 4391 // The paths which do not pass are accumulated in the slow_region. 4392 RegionNode* slow_region = new RegionNode(1); 4393 record_for_igvn(slow_region); 4394 if (!stopped()) { 4395 // It's an instance (we did array above). Make the slow-path tests. 4396 // If this is a virtual call, we generate a funny guard. We grab 4397 // the vtable entry corresponding to clone() from the target object. 4398 // If the target method which we are calling happens to be the 4399 // Object clone() method, we pass the guard. We do not need this 4400 // guard for non-virtual calls; the caller is known to be the native 4401 // Object clone(). 4402 if (is_virtual) { 4403 generate_virtual_guard(obj_klass, slow_region); 4404 } 4405 4406 // The object must be easily cloneable and must not have a finalizer. 4407 // Both of these conditions may be checked in a single test. 4408 // We could optimize the test further, but we don't care. 4409 generate_access_flags_guard(obj_klass, 4410 // Test both conditions: 4411 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4412 // Must be cloneable but not finalizer: 4413 JVM_ACC_IS_CLONEABLE_FAST, 4414 slow_region); 4415 } 4416 4417 if (!stopped()) { 4418 // It's an instance, and it passed the slow-path tests. 4419 PreserveJVMState pjvms(this); 4420 Node* obj_size = NULL; 4421 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4422 // is reexecuted if deoptimization occurs and there could be problems when merging 4423 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4424 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4425 4426 copy_to_clone(obj, alloc_obj, obj_size, false); 4427 4428 // Present the results of the slow call. 4429 result_reg->init_req(_instance_path, control()); 4430 result_val->init_req(_instance_path, alloc_obj); 4431 result_i_o ->set_req(_instance_path, i_o()); 4432 result_mem ->set_req(_instance_path, reset_memory()); 4433 } 4434 4435 // Generate code for the slow case. We make a call to clone(). 4436 set_control(_gvn.transform(slow_region)); 4437 if (!stopped()) { 4438 PreserveJVMState pjvms(this); 4439 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4440 // We need to deoptimize on exception (see comment above) 4441 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true); 4442 // this->control() comes from set_results_for_java_call 4443 result_reg->init_req(_slow_path, control()); 4444 result_val->init_req(_slow_path, slow_result); 4445 result_i_o ->set_req(_slow_path, i_o()); 4446 result_mem ->set_req(_slow_path, reset_memory()); 4447 } 4448 4449 // Return the combined state. 4450 set_control( _gvn.transform(result_reg)); 4451 set_i_o( _gvn.transform(result_i_o)); 4452 set_all_memory( _gvn.transform(result_mem)); 4453 } // original reexecute is set back here 4454 4455 set_result(_gvn.transform(result_val)); 4456 return true; 4457 } 4458 4459 // If we have a tighly coupled allocation, the arraycopy may take care 4460 // of the array initialization. If one of the guards we insert between 4461 // the allocation and the arraycopy causes a deoptimization, an 4462 // unitialized array will escape the compiled method. To prevent that 4463 // we set the JVM state for uncommon traps between the allocation and 4464 // the arraycopy to the state before the allocation so, in case of 4465 // deoptimization, we'll reexecute the allocation and the 4466 // initialization. 4467 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4468 if (alloc != NULL) { 4469 ciMethod* trap_method = alloc->jvms()->method(); 4470 int trap_bci = alloc->jvms()->bci(); 4471 4472 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4473 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4474 // Make sure there's no store between the allocation and the 4475 // arraycopy otherwise visible side effects could be rexecuted 4476 // in case of deoptimization and cause incorrect execution. 4477 bool no_interfering_store = true; 4478 Node* mem = alloc->in(TypeFunc::Memory); 4479 if (mem->is_MergeMem()) { 4480 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4481 Node* n = mms.memory(); 4482 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4483 assert(n->is_Store(), "what else?"); 4484 no_interfering_store = false; 4485 break; 4486 } 4487 } 4488 } else { 4489 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4490 Node* n = mms.memory(); 4491 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4492 assert(n->is_Store(), "what else?"); 4493 no_interfering_store = false; 4494 break; 4495 } 4496 } 4497 } 4498 4499 if (no_interfering_store) { 4500 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4501 uint size = alloc->req(); 4502 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4503 old_jvms->set_map(sfpt); 4504 for (uint i = 0; i < size; i++) { 4505 sfpt->init_req(i, alloc->in(i)); 4506 } 4507 // re-push array length for deoptimization 4508 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4509 old_jvms->set_sp(old_jvms->sp()+1); 4510 old_jvms->set_monoff(old_jvms->monoff()+1); 4511 old_jvms->set_scloff(old_jvms->scloff()+1); 4512 old_jvms->set_endoff(old_jvms->endoff()+1); 4513 old_jvms->set_should_reexecute(true); 4514 4515 sfpt->set_i_o(map()->i_o()); 4516 sfpt->set_memory(map()->memory()); 4517 sfpt->set_control(map()->control()); 4518 4519 JVMState* saved_jvms = jvms(); 4520 saved_reexecute_sp = _reexecute_sp; 4521 4522 set_jvms(sfpt->jvms()); 4523 _reexecute_sp = jvms()->sp(); 4524 4525 return saved_jvms; 4526 } 4527 } 4528 } 4529 return NULL; 4530 } 4531 4532 // In case of a deoptimization, we restart execution at the 4533 // allocation, allocating a new array. We would leave an uninitialized 4534 // array in the heap that GCs wouldn't expect. Move the allocation 4535 // after the traps so we don't allocate the array if we 4536 // deoptimize. This is possible because tightly_coupled_allocation() 4537 // guarantees there's no observer of the allocated array at this point 4538 // and the control flow is simple enough. 4539 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, 4540 int saved_reexecute_sp, uint new_idx) { 4541 if (saved_jvms != NULL && !stopped()) { 4542 assert(alloc != NULL, "only with a tightly coupled allocation"); 4543 // restore JVM state to the state at the arraycopy 4544 saved_jvms->map()->set_control(map()->control()); 4545 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4546 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4547 // If we've improved the types of some nodes (null check) while 4548 // emitting the guards, propagate them to the current state 4549 map()->replaced_nodes().apply(saved_jvms->map(), new_idx); 4550 set_jvms(saved_jvms); 4551 _reexecute_sp = saved_reexecute_sp; 4552 4553 // Remove the allocation from above the guards 4554 CallProjections callprojs; 4555 alloc->extract_projections(&callprojs, true); 4556 InitializeNode* init = alloc->initialization(); 4557 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4558 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4559 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4560 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4561 4562 // move the allocation here (after the guards) 4563 _gvn.hash_delete(alloc); 4564 alloc->set_req(TypeFunc::Control, control()); 4565 alloc->set_req(TypeFunc::I_O, i_o()); 4566 Node *mem = reset_memory(); 4567 set_all_memory(mem); 4568 alloc->set_req(TypeFunc::Memory, mem); 4569 set_control(init->proj_out_or_null(TypeFunc::Control)); 4570 set_i_o(callprojs.fallthrough_ioproj); 4571 4572 // Update memory as done in GraphKit::set_output_for_allocation() 4573 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4574 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4575 if (ary_type->isa_aryptr() && length_type != NULL) { 4576 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4577 } 4578 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4579 int elemidx = C->get_alias_index(telemref); 4580 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw); 4581 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx); 4582 4583 Node* allocx = _gvn.transform(alloc); 4584 assert(allocx == alloc, "where has the allocation gone?"); 4585 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4586 4587 _gvn.hash_delete(dest); 4588 dest->set_req(0, control()); 4589 Node* destx = _gvn.transform(dest); 4590 assert(destx == dest, "where has the allocation result gone?"); 4591 } 4592 } 4593 4594 4595 //------------------------------inline_arraycopy----------------------- 4596 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4597 // Object dest, int destPos, 4598 // int length); 4599 bool LibraryCallKit::inline_arraycopy() { 4600 // Get the arguments. 4601 Node* src = argument(0); // type: oop 4602 Node* src_offset = argument(1); // type: int 4603 Node* dest = argument(2); // type: oop 4604 Node* dest_offset = argument(3); // type: int 4605 Node* length = argument(4); // type: int 4606 4607 uint new_idx = C->unique(); 4608 4609 // Check for allocation before we add nodes that would confuse 4610 // tightly_coupled_allocation() 4611 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4612 4613 int saved_reexecute_sp = -1; 4614 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4615 // See arraycopy_restore_alloc_state() comment 4616 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4617 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4618 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards 4619 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4620 4621 // The following tests must be performed 4622 // (1) src and dest are arrays. 4623 // (2) src and dest arrays must have elements of the same BasicType 4624 // (3) src and dest must not be null. 4625 // (4) src_offset must not be negative. 4626 // (5) dest_offset must not be negative. 4627 // (6) length must not be negative. 4628 // (7) src_offset + length must not exceed length of src. 4629 // (8) dest_offset + length must not exceed length of dest. 4630 // (9) each element of an oop array must be assignable 4631 4632 // (3) src and dest must not be null. 4633 // always do this here because we need the JVM state for uncommon traps 4634 Node* null_ctl = top(); 4635 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4636 assert(null_ctl->is_top(), "no null control here"); 4637 dest = null_check(dest, T_ARRAY); 4638 4639 if (!can_emit_guards) { 4640 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4641 // guards but the arraycopy node could still take advantage of a 4642 // tightly allocated allocation. tightly_coupled_allocation() is 4643 // called again to make sure it takes the null check above into 4644 // account: the null check is mandatory and if it caused an 4645 // uncommon trap to be emitted then the allocation can't be 4646 // considered tightly coupled in this context. 4647 alloc = tightly_coupled_allocation(dest, NULL); 4648 } 4649 4650 bool validated = false; 4651 4652 const Type* src_type = _gvn.type(src); 4653 const Type* dest_type = _gvn.type(dest); 4654 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4655 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4656 4657 // Do we have the type of src? 4658 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4659 // Do we have the type of dest? 4660 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4661 // Is the type for src from speculation? 4662 bool src_spec = false; 4663 // Is the type for dest from speculation? 4664 bool dest_spec = false; 4665 4666 if ((!has_src || !has_dest) && can_emit_guards) { 4667 // We don't have sufficient type information, let's see if 4668 // speculative types can help. We need to have types for both src 4669 // and dest so that it pays off. 4670 4671 // Do we already have or could we have type information for src 4672 bool could_have_src = has_src; 4673 // Do we already have or could we have type information for dest 4674 bool could_have_dest = has_dest; 4675 4676 ciKlass* src_k = NULL; 4677 if (!has_src) { 4678 src_k = src_type->speculative_type_not_null(); 4679 if (src_k != NULL && src_k->is_array_klass()) { 4680 could_have_src = true; 4681 } 4682 } 4683 4684 ciKlass* dest_k = NULL; 4685 if (!has_dest) { 4686 dest_k = dest_type->speculative_type_not_null(); 4687 if (dest_k != NULL && dest_k->is_array_klass()) { 4688 could_have_dest = true; 4689 } 4690 } 4691 4692 if (could_have_src && could_have_dest) { 4693 // This is going to pay off so emit the required guards 4694 if (!has_src) { 4695 src = maybe_cast_profiled_obj(src, src_k, true); 4696 src_type = _gvn.type(src); 4697 top_src = src_type->isa_aryptr(); 4698 has_src = (top_src != NULL && top_src->klass() != NULL); 4699 src_spec = true; 4700 } 4701 if (!has_dest) { 4702 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4703 dest_type = _gvn.type(dest); 4704 top_dest = dest_type->isa_aryptr(); 4705 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4706 dest_spec = true; 4707 } 4708 } 4709 } 4710 4711 if (has_src && has_dest && can_emit_guards) { 4712 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4713 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4714 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4715 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4716 4717 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4718 // If both arrays are object arrays then having the exact types 4719 // for both will remove the need for a subtype check at runtime 4720 // before the call and may make it possible to pick a faster copy 4721 // routine (without a subtype check on every element) 4722 // Do we have the exact type of src? 4723 bool could_have_src = src_spec; 4724 // Do we have the exact type of dest? 4725 bool could_have_dest = dest_spec; 4726 ciKlass* src_k = top_src->klass(); 4727 ciKlass* dest_k = top_dest->klass(); 4728 if (!src_spec) { 4729 src_k = src_type->speculative_type_not_null(); 4730 if (src_k != NULL && src_k->is_array_klass()) { 4731 could_have_src = true; 4732 } 4733 } 4734 if (!dest_spec) { 4735 dest_k = dest_type->speculative_type_not_null(); 4736 if (dest_k != NULL && dest_k->is_array_klass()) { 4737 could_have_dest = true; 4738 } 4739 } 4740 if (could_have_src && could_have_dest) { 4741 // If we can have both exact types, emit the missing guards 4742 if (could_have_src && !src_spec) { 4743 src = maybe_cast_profiled_obj(src, src_k, true); 4744 } 4745 if (could_have_dest && !dest_spec) { 4746 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4747 } 4748 } 4749 } 4750 } 4751 4752 ciMethod* trap_method = method(); 4753 int trap_bci = bci(); 4754 if (saved_jvms != NULL) { 4755 trap_method = alloc->jvms()->method(); 4756 trap_bci = alloc->jvms()->bci(); 4757 } 4758 4759 bool negative_length_guard_generated = false; 4760 4761 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 4762 can_emit_guards && 4763 !src->is_top() && !dest->is_top()) { 4764 // validate arguments: enables transformation the ArrayCopyNode 4765 validated = true; 4766 4767 RegionNode* slow_region = new RegionNode(1); 4768 record_for_igvn(slow_region); 4769 4770 // (1) src and dest are arrays. 4771 generate_non_array_guard(load_object_klass(src), slow_region); 4772 generate_non_array_guard(load_object_klass(dest), slow_region); 4773 4774 // (2) src and dest arrays must have elements of the same BasicType 4775 // done at macro expansion or at Ideal transformation time 4776 4777 // (4) src_offset must not be negative. 4778 generate_negative_guard(src_offset, slow_region); 4779 4780 // (5) dest_offset must not be negative. 4781 generate_negative_guard(dest_offset, slow_region); 4782 4783 // (7) src_offset + length must not exceed length of src. 4784 generate_limit_guard(src_offset, length, 4785 load_array_length(src), 4786 slow_region); 4787 4788 // (8) dest_offset + length must not exceed length of dest. 4789 generate_limit_guard(dest_offset, length, 4790 load_array_length(dest), 4791 slow_region); 4792 4793 // (6) length must not be negative. 4794 // This is also checked in generate_arraycopy() during macro expansion, but 4795 // we also have to check it here for the case where the ArrayCopyNode will 4796 // be eliminated by Escape Analysis. 4797 if (EliminateAllocations) { 4798 generate_negative_guard(length, slow_region); 4799 negative_length_guard_generated = true; 4800 } 4801 4802 // (9) each element of an oop array must be assignable 4803 Node* src_klass = load_object_klass(src); 4804 Node* dest_klass = load_object_klass(dest); 4805 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4806 4807 if (not_subtype_ctrl != top()) { 4808 PreserveJVMState pjvms(this); 4809 set_control(not_subtype_ctrl); 4810 uncommon_trap(Deoptimization::Reason_intrinsic, 4811 Deoptimization::Action_make_not_entrant); 4812 assert(stopped(), "Should be stopped"); 4813 } 4814 { 4815 PreserveJVMState pjvms(this); 4816 set_control(_gvn.transform(slow_region)); 4817 uncommon_trap(Deoptimization::Reason_intrinsic, 4818 Deoptimization::Action_make_not_entrant); 4819 assert(stopped(), "Should be stopped"); 4820 } 4821 4822 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr(); 4823 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass()); 4824 src = _gvn.transform(new CheckCastPPNode(control(), src, toop)); 4825 } 4826 4827 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx); 4828 4829 if (stopped()) { 4830 return true; 4831 } 4832 4833 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated, 4834 // Create LoadRange and LoadKlass nodes for use during macro expansion here 4835 // so the compiler has a chance to eliminate them: during macro expansion, 4836 // we have to set their control (CastPP nodes are eliminated). 4837 load_object_klass(src), load_object_klass(dest), 4838 load_array_length(src), load_array_length(dest)); 4839 4840 ac->set_arraycopy(validated); 4841 4842 Node* n = _gvn.transform(ac); 4843 if (n == ac) { 4844 ac->connect_outputs(this); 4845 } else { 4846 assert(validated, "shouldn't transform if all arguments not validated"); 4847 set_all_memory(n); 4848 } 4849 clear_upper_avx(); 4850 4851 4852 return true; 4853 } 4854 4855 4856 // Helper function which determines if an arraycopy immediately follows 4857 // an allocation, with no intervening tests or other escapes for the object. 4858 AllocateArrayNode* 4859 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4860 RegionNode* slow_region) { 4861 if (stopped()) return NULL; // no fast path 4862 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4863 4864 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4865 if (alloc == NULL) return NULL; 4866 4867 Node* rawmem = memory(Compile::AliasIdxRaw); 4868 // Is the allocation's memory state untouched? 4869 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4870 // Bail out if there have been raw-memory effects since the allocation. 4871 // (Example: There might have been a call or safepoint.) 4872 return NULL; 4873 } 4874 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4875 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4876 return NULL; 4877 } 4878 4879 // There must be no unexpected observers of this allocation. 4880 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4881 Node* obs = ptr->fast_out(i); 4882 if (obs != this->map()) { 4883 return NULL; 4884 } 4885 } 4886 4887 // This arraycopy must unconditionally follow the allocation of the ptr. 4888 Node* alloc_ctl = ptr->in(0); 4889 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4890 4891 Node* ctl = control(); 4892 while (ctl != alloc_ctl) { 4893 // There may be guards which feed into the slow_region. 4894 // Any other control flow means that we might not get a chance 4895 // to finish initializing the allocated object. 4896 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4897 IfNode* iff = ctl->in(0)->as_If(); 4898 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con); 4899 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4900 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4901 ctl = iff->in(0); // This test feeds the known slow_region. 4902 continue; 4903 } 4904 // One more try: Various low-level checks bottom out in 4905 // uncommon traps. If the debug-info of the trap omits 4906 // any reference to the allocation, as we've already 4907 // observed, then there can be no objection to the trap. 4908 bool found_trap = false; 4909 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 4910 Node* obs = not_ctl->fast_out(j); 4911 if (obs->in(0) == not_ctl && obs->is_Call() && 4912 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 4913 found_trap = true; break; 4914 } 4915 } 4916 if (found_trap) { 4917 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 4918 continue; 4919 } 4920 } 4921 return NULL; 4922 } 4923 4924 // If we get this far, we have an allocation which immediately 4925 // precedes the arraycopy, and we can take over zeroing the new object. 4926 // The arraycopy will finish the initialization, and provide 4927 // a new control state to which we will anchor the destination pointer. 4928 4929 return alloc; 4930 } 4931 4932 //-------------inline_encodeISOArray----------------------------------- 4933 // encode char[] to byte[] in ISO_8859_1 4934 bool LibraryCallKit::inline_encodeISOArray() { 4935 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 4936 // no receiver since it is static method 4937 Node *src = argument(0); 4938 Node *src_offset = argument(1); 4939 Node *dst = argument(2); 4940 Node *dst_offset = argument(3); 4941 Node *length = argument(4); 4942 4943 src = must_be_not_null(src, true); 4944 dst = must_be_not_null(dst, true); 4945 4946 const Type* src_type = src->Value(&_gvn); 4947 const Type* dst_type = dst->Value(&_gvn); 4948 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4949 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 4950 if (top_src == NULL || top_src->klass() == NULL || 4951 top_dest == NULL || top_dest->klass() == NULL) { 4952 // failed array check 4953 return false; 4954 } 4955 4956 // Figure out the size and type of the elements we will be copying. 4957 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4958 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4959 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 4960 return false; 4961 } 4962 4963 Node* src_start = array_element_address(src, src_offset, T_CHAR); 4964 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 4965 // 'src_start' points to src array + scaled offset 4966 // 'dst_start' points to dst array + scaled offset 4967 4968 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 4969 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 4970 enc = _gvn.transform(enc); 4971 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 4972 set_memory(res_mem, mtype); 4973 set_result(enc); 4974 clear_upper_avx(); 4975 4976 return true; 4977 } 4978 4979 //-------------inline_multiplyToLen----------------------------------- 4980 bool LibraryCallKit::inline_multiplyToLen() { 4981 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 4982 4983 address stubAddr = StubRoutines::multiplyToLen(); 4984 if (stubAddr == NULL) { 4985 return false; // Intrinsic's stub is not implemented on this platform 4986 } 4987 const char* stubName = "multiplyToLen"; 4988 4989 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 4990 4991 // no receiver because it is a static method 4992 Node* x = argument(0); 4993 Node* xlen = argument(1); 4994 Node* y = argument(2); 4995 Node* ylen = argument(3); 4996 Node* z = argument(4); 4997 4998 x = must_be_not_null(x, true); 4999 y = must_be_not_null(y, true); 5000 5001 const Type* x_type = x->Value(&_gvn); 5002 const Type* y_type = y->Value(&_gvn); 5003 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5004 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5005 if (top_x == NULL || top_x->klass() == NULL || 5006 top_y == NULL || top_y->klass() == NULL) { 5007 // failed array check 5008 return false; 5009 } 5010 5011 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5012 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5013 if (x_elem != T_INT || y_elem != T_INT) { 5014 return false; 5015 } 5016 5017 // Set the original stack and the reexecute bit for the interpreter to reexecute 5018 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5019 // on the return from z array allocation in runtime. 5020 { PreserveReexecuteState preexecs(this); 5021 jvms()->set_should_reexecute(true); 5022 5023 Node* x_start = array_element_address(x, intcon(0), x_elem); 5024 Node* y_start = array_element_address(y, intcon(0), y_elem); 5025 // 'x_start' points to x array + scaled xlen 5026 // 'y_start' points to y array + scaled ylen 5027 5028 // Allocate the result array 5029 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5030 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5031 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5032 5033 IdealKit ideal(this); 5034 5035 #define __ ideal. 5036 Node* one = __ ConI(1); 5037 Node* zero = __ ConI(0); 5038 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5039 __ set(need_alloc, zero); 5040 __ set(z_alloc, z); 5041 __ if_then(z, BoolTest::eq, null()); { 5042 __ increment (need_alloc, one); 5043 } __ else_(); { 5044 // Update graphKit memory and control from IdealKit. 5045 sync_kit(ideal); 5046 Node *cast = new CastPPNode(z, TypePtr::NOTNULL); 5047 cast->init_req(0, control()); 5048 _gvn.set_type(cast, cast->bottom_type()); 5049 C->record_for_igvn(cast); 5050 5051 Node* zlen_arg = load_array_length(cast); 5052 // Update IdealKit memory and control from graphKit. 5053 __ sync_kit(this); 5054 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5055 __ increment (need_alloc, one); 5056 } __ end_if(); 5057 } __ end_if(); 5058 5059 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5060 // Update graphKit memory and control from IdealKit. 5061 sync_kit(ideal); 5062 Node * narr = new_array(klass_node, zlen, 1); 5063 // Update IdealKit memory and control from graphKit. 5064 __ sync_kit(this); 5065 __ set(z_alloc, narr); 5066 } __ end_if(); 5067 5068 sync_kit(ideal); 5069 z = __ value(z_alloc); 5070 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5071 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5072 // Final sync IdealKit and GraphKit. 5073 final_sync(ideal); 5074 #undef __ 5075 5076 Node* z_start = array_element_address(z, intcon(0), T_INT); 5077 5078 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5079 OptoRuntime::multiplyToLen_Type(), 5080 stubAddr, stubName, TypePtr::BOTTOM, 5081 x_start, xlen, y_start, ylen, z_start, zlen); 5082 } // original reexecute is set back here 5083 5084 C->set_has_split_ifs(true); // Has chance for split-if optimization 5085 set_result(z); 5086 return true; 5087 } 5088 5089 //-------------inline_squareToLen------------------------------------ 5090 bool LibraryCallKit::inline_squareToLen() { 5091 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5092 5093 address stubAddr = StubRoutines::squareToLen(); 5094 if (stubAddr == NULL) { 5095 return false; // Intrinsic's stub is not implemented on this platform 5096 } 5097 const char* stubName = "squareToLen"; 5098 5099 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5100 5101 Node* x = argument(0); 5102 Node* len = argument(1); 5103 Node* z = argument(2); 5104 Node* zlen = argument(3); 5105 5106 x = must_be_not_null(x, true); 5107 z = must_be_not_null(z, true); 5108 5109 const Type* x_type = x->Value(&_gvn); 5110 const Type* z_type = z->Value(&_gvn); 5111 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5112 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5113 if (top_x == NULL || top_x->klass() == NULL || 5114 top_z == NULL || top_z->klass() == NULL) { 5115 // failed array check 5116 return false; 5117 } 5118 5119 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5120 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5121 if (x_elem != T_INT || z_elem != T_INT) { 5122 return false; 5123 } 5124 5125 5126 Node* x_start = array_element_address(x, intcon(0), x_elem); 5127 Node* z_start = array_element_address(z, intcon(0), z_elem); 5128 5129 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5130 OptoRuntime::squareToLen_Type(), 5131 stubAddr, stubName, TypePtr::BOTTOM, 5132 x_start, len, z_start, zlen); 5133 5134 set_result(z); 5135 return true; 5136 } 5137 5138 //-------------inline_mulAdd------------------------------------------ 5139 bool LibraryCallKit::inline_mulAdd() { 5140 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5141 5142 address stubAddr = StubRoutines::mulAdd(); 5143 if (stubAddr == NULL) { 5144 return false; // Intrinsic's stub is not implemented on this platform 5145 } 5146 const char* stubName = "mulAdd"; 5147 5148 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5149 5150 Node* out = argument(0); 5151 Node* in = argument(1); 5152 Node* offset = argument(2); 5153 Node* len = argument(3); 5154 Node* k = argument(4); 5155 5156 out = must_be_not_null(out, true); 5157 5158 const Type* out_type = out->Value(&_gvn); 5159 const Type* in_type = in->Value(&_gvn); 5160 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5161 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5162 if (top_out == NULL || top_out->klass() == NULL || 5163 top_in == NULL || top_in->klass() == NULL) { 5164 // failed array check 5165 return false; 5166 } 5167 5168 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5169 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5170 if (out_elem != T_INT || in_elem != T_INT) { 5171 return false; 5172 } 5173 5174 Node* outlen = load_array_length(out); 5175 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5176 Node* out_start = array_element_address(out, intcon(0), out_elem); 5177 Node* in_start = array_element_address(in, intcon(0), in_elem); 5178 5179 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5180 OptoRuntime::mulAdd_Type(), 5181 stubAddr, stubName, TypePtr::BOTTOM, 5182 out_start,in_start, new_offset, len, k); 5183 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5184 set_result(result); 5185 return true; 5186 } 5187 5188 //-------------inline_montgomeryMultiply----------------------------------- 5189 bool LibraryCallKit::inline_montgomeryMultiply() { 5190 address stubAddr = StubRoutines::montgomeryMultiply(); 5191 if (stubAddr == NULL) { 5192 return false; // Intrinsic's stub is not implemented on this platform 5193 } 5194 5195 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5196 const char* stubName = "montgomery_multiply"; 5197 5198 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5199 5200 Node* a = argument(0); 5201 Node* b = argument(1); 5202 Node* n = argument(2); 5203 Node* len = argument(3); 5204 Node* inv = argument(4); 5205 Node* m = argument(6); 5206 5207 const Type* a_type = a->Value(&_gvn); 5208 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5209 const Type* b_type = b->Value(&_gvn); 5210 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5211 const Type* n_type = a->Value(&_gvn); 5212 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5213 const Type* m_type = a->Value(&_gvn); 5214 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5215 if (top_a == NULL || top_a->klass() == NULL || 5216 top_b == NULL || top_b->klass() == NULL || 5217 top_n == NULL || top_n->klass() == NULL || 5218 top_m == NULL || top_m->klass() == NULL) { 5219 // failed array check 5220 return false; 5221 } 5222 5223 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5224 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5225 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5226 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5227 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5228 return false; 5229 } 5230 5231 // Make the call 5232 { 5233 Node* a_start = array_element_address(a, intcon(0), a_elem); 5234 Node* b_start = array_element_address(b, intcon(0), b_elem); 5235 Node* n_start = array_element_address(n, intcon(0), n_elem); 5236 Node* m_start = array_element_address(m, intcon(0), m_elem); 5237 5238 Node* call = make_runtime_call(RC_LEAF, 5239 OptoRuntime::montgomeryMultiply_Type(), 5240 stubAddr, stubName, TypePtr::BOTTOM, 5241 a_start, b_start, n_start, len, inv, top(), 5242 m_start); 5243 set_result(m); 5244 } 5245 5246 return true; 5247 } 5248 5249 bool LibraryCallKit::inline_montgomerySquare() { 5250 address stubAddr = StubRoutines::montgomerySquare(); 5251 if (stubAddr == NULL) { 5252 return false; // Intrinsic's stub is not implemented on this platform 5253 } 5254 5255 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5256 const char* stubName = "montgomery_square"; 5257 5258 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5259 5260 Node* a = argument(0); 5261 Node* n = argument(1); 5262 Node* len = argument(2); 5263 Node* inv = argument(3); 5264 Node* m = argument(5); 5265 5266 const Type* a_type = a->Value(&_gvn); 5267 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5268 const Type* n_type = a->Value(&_gvn); 5269 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5270 const Type* m_type = a->Value(&_gvn); 5271 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5272 if (top_a == NULL || top_a->klass() == NULL || 5273 top_n == NULL || top_n->klass() == NULL || 5274 top_m == NULL || top_m->klass() == NULL) { 5275 // failed array check 5276 return false; 5277 } 5278 5279 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5280 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5281 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5282 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5283 return false; 5284 } 5285 5286 // Make the call 5287 { 5288 Node* a_start = array_element_address(a, intcon(0), a_elem); 5289 Node* n_start = array_element_address(n, intcon(0), n_elem); 5290 Node* m_start = array_element_address(m, intcon(0), m_elem); 5291 5292 Node* call = make_runtime_call(RC_LEAF, 5293 OptoRuntime::montgomerySquare_Type(), 5294 stubAddr, stubName, TypePtr::BOTTOM, 5295 a_start, n_start, len, inv, top(), 5296 m_start); 5297 set_result(m); 5298 } 5299 5300 return true; 5301 } 5302 5303 //-------------inline_vectorizedMismatch------------------------------ 5304 bool LibraryCallKit::inline_vectorizedMismatch() { 5305 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5306 5307 address stubAddr = StubRoutines::vectorizedMismatch(); 5308 if (stubAddr == NULL) { 5309 return false; // Intrinsic's stub is not implemented on this platform 5310 } 5311 const char* stubName = "vectorizedMismatch"; 5312 int size_l = callee()->signature()->size(); 5313 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5314 5315 Node* obja = argument(0); 5316 Node* aoffset = argument(1); 5317 Node* objb = argument(3); 5318 Node* boffset = argument(4); 5319 Node* length = argument(6); 5320 Node* scale = argument(7); 5321 5322 const Type* a_type = obja->Value(&_gvn); 5323 const Type* b_type = objb->Value(&_gvn); 5324 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5325 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5326 if (top_a == NULL || top_a->klass() == NULL || 5327 top_b == NULL || top_b->klass() == NULL) { 5328 // failed array check 5329 return false; 5330 } 5331 5332 Node* call; 5333 jvms()->set_should_reexecute(true); 5334 5335 Node* obja_adr = make_unsafe_address(obja, aoffset); 5336 Node* objb_adr = make_unsafe_address(objb, boffset); 5337 5338 call = make_runtime_call(RC_LEAF, 5339 OptoRuntime::vectorizedMismatch_Type(), 5340 stubAddr, stubName, TypePtr::BOTTOM, 5341 obja_adr, objb_adr, length, scale); 5342 5343 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5344 set_result(result); 5345 return true; 5346 } 5347 5348 /** 5349 * Calculate CRC32 for byte. 5350 * int java.util.zip.CRC32.update(int crc, int b) 5351 */ 5352 bool LibraryCallKit::inline_updateCRC32() { 5353 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5354 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5355 // no receiver since it is static method 5356 Node* crc = argument(0); // type: int 5357 Node* b = argument(1); // type: int 5358 5359 /* 5360 * int c = ~ crc; 5361 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5362 * b = b ^ (c >>> 8); 5363 * crc = ~b; 5364 */ 5365 5366 Node* M1 = intcon(-1); 5367 crc = _gvn.transform(new XorINode(crc, M1)); 5368 Node* result = _gvn.transform(new XorINode(crc, b)); 5369 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5370 5371 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5372 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5373 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5374 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5375 5376 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5377 result = _gvn.transform(new XorINode(crc, result)); 5378 result = _gvn.transform(new XorINode(result, M1)); 5379 set_result(result); 5380 return true; 5381 } 5382 5383 /** 5384 * Calculate CRC32 for byte[] array. 5385 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5386 */ 5387 bool LibraryCallKit::inline_updateBytesCRC32() { 5388 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5389 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5390 // no receiver since it is static method 5391 Node* crc = argument(0); // type: int 5392 Node* src = argument(1); // type: oop 5393 Node* offset = argument(2); // type: int 5394 Node* length = argument(3); // type: int 5395 5396 const Type* src_type = src->Value(&_gvn); 5397 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5398 if (top_src == NULL || top_src->klass() == NULL) { 5399 // failed array check 5400 return false; 5401 } 5402 5403 // Figure out the size and type of the elements we will be copying. 5404 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5405 if (src_elem != T_BYTE) { 5406 return false; 5407 } 5408 5409 // 'src_start' points to src array + scaled offset 5410 src = must_be_not_null(src, true); 5411 Node* src_start = array_element_address(src, offset, src_elem); 5412 5413 // We assume that range check is done by caller. 5414 // TODO: generate range check (offset+length < src.length) in debug VM. 5415 5416 // Call the stub. 5417 address stubAddr = StubRoutines::updateBytesCRC32(); 5418 const char *stubName = "updateBytesCRC32"; 5419 5420 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5421 stubAddr, stubName, TypePtr::BOTTOM, 5422 crc, src_start, length); 5423 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5424 set_result(result); 5425 return true; 5426 } 5427 5428 /** 5429 * Calculate CRC32 for ByteBuffer. 5430 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5431 */ 5432 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5433 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5434 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5435 // no receiver since it is static method 5436 Node* crc = argument(0); // type: int 5437 Node* src = argument(1); // type: long 5438 Node* offset = argument(3); // type: int 5439 Node* length = argument(4); // type: int 5440 5441 src = ConvL2X(src); // adjust Java long to machine word 5442 Node* base = _gvn.transform(new CastX2PNode(src)); 5443 offset = ConvI2X(offset); 5444 5445 // 'src_start' points to src array + scaled offset 5446 Node* src_start = basic_plus_adr(top(), base, offset); 5447 5448 // Call the stub. 5449 address stubAddr = StubRoutines::updateBytesCRC32(); 5450 const char *stubName = "updateBytesCRC32"; 5451 5452 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5453 stubAddr, stubName, TypePtr::BOTTOM, 5454 crc, src_start, length); 5455 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5456 set_result(result); 5457 return true; 5458 } 5459 5460 //------------------------------get_table_from_crc32c_class----------------------- 5461 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5462 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5463 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5464 5465 return table; 5466 } 5467 5468 //------------------------------inline_updateBytesCRC32C----------------------- 5469 // 5470 // Calculate CRC32C for byte[] array. 5471 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5472 // 5473 bool LibraryCallKit::inline_updateBytesCRC32C() { 5474 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5475 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5476 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5477 // no receiver since it is a static method 5478 Node* crc = argument(0); // type: int 5479 Node* src = argument(1); // type: oop 5480 Node* offset = argument(2); // type: int 5481 Node* end = argument(3); // type: int 5482 5483 Node* length = _gvn.transform(new SubINode(end, offset)); 5484 5485 const Type* src_type = src->Value(&_gvn); 5486 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5487 if (top_src == NULL || top_src->klass() == NULL) { 5488 // failed array check 5489 return false; 5490 } 5491 5492 // Figure out the size and type of the elements we will be copying. 5493 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5494 if (src_elem != T_BYTE) { 5495 return false; 5496 } 5497 5498 // 'src_start' points to src array + scaled offset 5499 src = must_be_not_null(src, true); 5500 Node* src_start = array_element_address(src, offset, src_elem); 5501 5502 // static final int[] byteTable in class CRC32C 5503 Node* table = get_table_from_crc32c_class(callee()->holder()); 5504 table = must_be_not_null(table, true); 5505 Node* table_start = array_element_address(table, intcon(0), T_INT); 5506 5507 // We assume that range check is done by caller. 5508 // TODO: generate range check (offset+length < src.length) in debug VM. 5509 5510 // Call the stub. 5511 address stubAddr = StubRoutines::updateBytesCRC32C(); 5512 const char *stubName = "updateBytesCRC32C"; 5513 5514 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5515 stubAddr, stubName, TypePtr::BOTTOM, 5516 crc, src_start, length, table_start); 5517 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5518 set_result(result); 5519 return true; 5520 } 5521 5522 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5523 // 5524 // Calculate CRC32C for DirectByteBuffer. 5525 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5526 // 5527 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5528 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5529 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5530 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5531 // no receiver since it is a static method 5532 Node* crc = argument(0); // type: int 5533 Node* src = argument(1); // type: long 5534 Node* offset = argument(3); // type: int 5535 Node* end = argument(4); // type: int 5536 5537 Node* length = _gvn.transform(new SubINode(end, offset)); 5538 5539 src = ConvL2X(src); // adjust Java long to machine word 5540 Node* base = _gvn.transform(new CastX2PNode(src)); 5541 offset = ConvI2X(offset); 5542 5543 // 'src_start' points to src array + scaled offset 5544 Node* src_start = basic_plus_adr(top(), base, offset); 5545 5546 // static final int[] byteTable in class CRC32C 5547 Node* table = get_table_from_crc32c_class(callee()->holder()); 5548 table = must_be_not_null(table, true); 5549 Node* table_start = array_element_address(table, intcon(0), T_INT); 5550 5551 // Call the stub. 5552 address stubAddr = StubRoutines::updateBytesCRC32C(); 5553 const char *stubName = "updateBytesCRC32C"; 5554 5555 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5556 stubAddr, stubName, TypePtr::BOTTOM, 5557 crc, src_start, length, table_start); 5558 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5559 set_result(result); 5560 return true; 5561 } 5562 5563 //------------------------------inline_updateBytesAdler32---------------------- 5564 // 5565 // Calculate Adler32 checksum for byte[] array. 5566 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5567 // 5568 bool LibraryCallKit::inline_updateBytesAdler32() { 5569 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5570 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5571 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5572 // no receiver since it is static method 5573 Node* crc = argument(0); // type: int 5574 Node* src = argument(1); // type: oop 5575 Node* offset = argument(2); // type: int 5576 Node* length = argument(3); // type: int 5577 5578 const Type* src_type = src->Value(&_gvn); 5579 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5580 if (top_src == NULL || top_src->klass() == NULL) { 5581 // failed array check 5582 return false; 5583 } 5584 5585 // Figure out the size and type of the elements we will be copying. 5586 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5587 if (src_elem != T_BYTE) { 5588 return false; 5589 } 5590 5591 // 'src_start' points to src array + scaled offset 5592 Node* src_start = array_element_address(src, offset, src_elem); 5593 5594 // We assume that range check is done by caller. 5595 // TODO: generate range check (offset+length < src.length) in debug VM. 5596 5597 // Call the stub. 5598 address stubAddr = StubRoutines::updateBytesAdler32(); 5599 const char *stubName = "updateBytesAdler32"; 5600 5601 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5602 stubAddr, stubName, TypePtr::BOTTOM, 5603 crc, src_start, length); 5604 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5605 set_result(result); 5606 return true; 5607 } 5608 5609 //------------------------------inline_updateByteBufferAdler32--------------- 5610 // 5611 // Calculate Adler32 checksum for DirectByteBuffer. 5612 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5613 // 5614 bool LibraryCallKit::inline_updateByteBufferAdler32() { 5615 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5616 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5617 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5618 // no receiver since it is static method 5619 Node* crc = argument(0); // type: int 5620 Node* src = argument(1); // type: long 5621 Node* offset = argument(3); // type: int 5622 Node* length = argument(4); // type: int 5623 5624 src = ConvL2X(src); // adjust Java long to machine word 5625 Node* base = _gvn.transform(new CastX2PNode(src)); 5626 offset = ConvI2X(offset); 5627 5628 // 'src_start' points to src array + scaled offset 5629 Node* src_start = basic_plus_adr(top(), base, offset); 5630 5631 // Call the stub. 5632 address stubAddr = StubRoutines::updateBytesAdler32(); 5633 const char *stubName = "updateBytesAdler32"; 5634 5635 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5636 stubAddr, stubName, TypePtr::BOTTOM, 5637 crc, src_start, length); 5638 5639 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5640 set_result(result); 5641 return true; 5642 } 5643 5644 //----------------------------inline_reference_get---------------------------- 5645 // public T java.lang.ref.Reference.get(); 5646 bool LibraryCallKit::inline_reference_get() { 5647 const int referent_offset = java_lang_ref_Reference::referent_offset; 5648 guarantee(referent_offset > 0, "should have already been set"); 5649 5650 // Get the argument: 5651 Node* reference_obj = null_check_receiver(); 5652 if (stopped()) return true; 5653 5654 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr(); 5655 assert(tinst != NULL, "obj is null"); 5656 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5657 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass(); 5658 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"), 5659 ciSymbol::make("Ljava/lang/Object;"), 5660 false); 5661 assert (field != NULL, "undefined field"); 5662 5663 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5664 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5665 5666 ciInstanceKlass* klass = env()->Object_klass(); 5667 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5668 5669 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF; 5670 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators); 5671 // Add memory barrier to prevent commoning reads from this field 5672 // across safepoint since GC can change its value. 5673 insert_mem_bar(Op_MemBarCPUOrder); 5674 5675 set_result(result); 5676 return true; 5677 } 5678 5679 5680 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5681 bool is_exact=true, bool is_static=false, 5682 ciInstanceKlass * fromKls=NULL) { 5683 if (fromKls == NULL) { 5684 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5685 assert(tinst != NULL, "obj is null"); 5686 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5687 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5688 fromKls = tinst->klass()->as_instance_klass(); 5689 } else { 5690 assert(is_static, "only for static field access"); 5691 } 5692 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5693 ciSymbol::make(fieldTypeString), 5694 is_static); 5695 5696 assert (field != NULL, "undefined field"); 5697 if (field == NULL) return (Node *) NULL; 5698 5699 if (is_static) { 5700 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5701 fromObj = makecon(tip); 5702 } 5703 5704 // Next code copied from Parse::do_get_xxx(): 5705 5706 // Compute address and memory type. 5707 int offset = field->offset_in_bytes(); 5708 bool is_vol = field->is_volatile(); 5709 ciType* field_klass = field->type(); 5710 assert(field_klass->is_loaded(), "should be loaded"); 5711 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5712 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5713 BasicType bt = field->layout_type(); 5714 5715 // Build the resultant type of the load 5716 const Type *type; 5717 if (bt == T_OBJECT) { 5718 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5719 } else { 5720 type = Type::get_const_basic_type(bt); 5721 } 5722 5723 DecoratorSet decorators = IN_HEAP; 5724 5725 if (is_vol) { 5726 decorators |= MO_SEQ_CST; 5727 } 5728 5729 return access_load_at(fromObj, adr, adr_type, type, bt, decorators); 5730 } 5731 5732 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5733 bool is_exact = true, bool is_static = false, 5734 ciInstanceKlass * fromKls = NULL) { 5735 if (fromKls == NULL) { 5736 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5737 assert(tinst != NULL, "obj is null"); 5738 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5739 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5740 fromKls = tinst->klass()->as_instance_klass(); 5741 } 5742 else { 5743 assert(is_static, "only for static field access"); 5744 } 5745 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5746 ciSymbol::make(fieldTypeString), 5747 is_static); 5748 5749 assert(field != NULL, "undefined field"); 5750 assert(!field->is_volatile(), "not defined for volatile fields"); 5751 5752 if (is_static) { 5753 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5754 fromObj = makecon(tip); 5755 } 5756 5757 // Next code copied from Parse::do_get_xxx(): 5758 5759 // Compute address and memory type. 5760 int offset = field->offset_in_bytes(); 5761 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5762 5763 return adr; 5764 } 5765 5766 //------------------------------inline_aescrypt_Block----------------------- 5767 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5768 address stubAddr = NULL; 5769 const char *stubName; 5770 assert(UseAES, "need AES instruction support"); 5771 5772 switch(id) { 5773 case vmIntrinsics::_aescrypt_encryptBlock: 5774 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5775 stubName = "aescrypt_encryptBlock"; 5776 break; 5777 case vmIntrinsics::_aescrypt_decryptBlock: 5778 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5779 stubName = "aescrypt_decryptBlock"; 5780 break; 5781 default: 5782 break; 5783 } 5784 if (stubAddr == NULL) return false; 5785 5786 Node* aescrypt_object = argument(0); 5787 Node* src = argument(1); 5788 Node* src_offset = argument(2); 5789 Node* dest = argument(3); 5790 Node* dest_offset = argument(4); 5791 5792 src = must_be_not_null(src, true); 5793 dest = must_be_not_null(dest, true); 5794 5795 // (1) src and dest are arrays. 5796 const Type* src_type = src->Value(&_gvn); 5797 const Type* dest_type = dest->Value(&_gvn); 5798 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5799 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5800 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5801 5802 // for the quick and dirty code we will skip all the checks. 5803 // we are just trying to get the call to be generated. 5804 Node* src_start = src; 5805 Node* dest_start = dest; 5806 if (src_offset != NULL || dest_offset != NULL) { 5807 assert(src_offset != NULL && dest_offset != NULL, ""); 5808 src_start = array_element_address(src, src_offset, T_BYTE); 5809 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5810 } 5811 5812 // now need to get the start of its expanded key array 5813 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5814 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5815 if (k_start == NULL) return false; 5816 5817 if (Matcher::pass_original_key_for_aes()) { 5818 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5819 // compatibility issues between Java key expansion and SPARC crypto instructions 5820 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5821 if (original_k_start == NULL) return false; 5822 5823 // Call the stub. 5824 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5825 stubAddr, stubName, TypePtr::BOTTOM, 5826 src_start, dest_start, k_start, original_k_start); 5827 } else { 5828 // Call the stub. 5829 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5830 stubAddr, stubName, TypePtr::BOTTOM, 5831 src_start, dest_start, k_start); 5832 } 5833 5834 return true; 5835 } 5836 5837 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5838 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5839 address stubAddr = NULL; 5840 const char *stubName = NULL; 5841 5842 assert(UseAES, "need AES instruction support"); 5843 5844 switch(id) { 5845 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5846 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5847 stubName = "cipherBlockChaining_encryptAESCrypt"; 5848 break; 5849 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5850 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5851 stubName = "cipherBlockChaining_decryptAESCrypt"; 5852 break; 5853 default: 5854 break; 5855 } 5856 if (stubAddr == NULL) return false; 5857 5858 Node* cipherBlockChaining_object = argument(0); 5859 Node* src = argument(1); 5860 Node* src_offset = argument(2); 5861 Node* len = argument(3); 5862 Node* dest = argument(4); 5863 Node* dest_offset = argument(5); 5864 5865 src = must_be_not_null(src, false); 5866 dest = must_be_not_null(dest, false); 5867 5868 // (1) src and dest are arrays. 5869 const Type* src_type = src->Value(&_gvn); 5870 const Type* dest_type = dest->Value(&_gvn); 5871 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5872 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5873 assert (top_src != NULL && top_src->klass() != NULL 5874 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5875 5876 // checks are the responsibility of the caller 5877 Node* src_start = src; 5878 Node* dest_start = dest; 5879 if (src_offset != NULL || dest_offset != NULL) { 5880 assert(src_offset != NULL && dest_offset != NULL, ""); 5881 src_start = array_element_address(src, src_offset, T_BYTE); 5882 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5883 } 5884 5885 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5886 // (because of the predicated logic executed earlier). 5887 // so we cast it here safely. 5888 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5889 5890 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5891 if (embeddedCipherObj == NULL) return false; 5892 5893 // cast it to what we know it will be at runtime 5894 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5895 assert(tinst != NULL, "CBC obj is null"); 5896 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5897 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5898 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5899 5900 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5901 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5902 const TypeOopPtr* xtype = aklass->as_instance_type(); 5903 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5904 aescrypt_object = _gvn.transform(aescrypt_object); 5905 5906 // we need to get the start of the aescrypt_object's expanded key array 5907 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5908 if (k_start == NULL) return false; 5909 5910 // similarly, get the start address of the r vector 5911 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5912 if (objRvec == NULL) return false; 5913 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5914 5915 Node* cbcCrypt; 5916 if (Matcher::pass_original_key_for_aes()) { 5917 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5918 // compatibility issues between Java key expansion and SPARC crypto instructions 5919 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5920 if (original_k_start == NULL) return false; 5921 5922 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5923 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5924 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5925 stubAddr, stubName, TypePtr::BOTTOM, 5926 src_start, dest_start, k_start, r_start, len, original_k_start); 5927 } else { 5928 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5929 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5930 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5931 stubAddr, stubName, TypePtr::BOTTOM, 5932 src_start, dest_start, k_start, r_start, len); 5933 } 5934 5935 // return cipher length (int) 5936 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5937 set_result(retvalue); 5938 return true; 5939 } 5940 5941 //------------------------------inline_counterMode_AESCrypt----------------------- 5942 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 5943 assert(UseAES, "need AES instruction support"); 5944 if (!UseAESCTRIntrinsics) return false; 5945 5946 address stubAddr = NULL; 5947 const char *stubName = NULL; 5948 if (id == vmIntrinsics::_counterMode_AESCrypt) { 5949 stubAddr = StubRoutines::counterMode_AESCrypt(); 5950 stubName = "counterMode_AESCrypt"; 5951 } 5952 if (stubAddr == NULL) return false; 5953 5954 Node* counterMode_object = argument(0); 5955 Node* src = argument(1); 5956 Node* src_offset = argument(2); 5957 Node* len = argument(3); 5958 Node* dest = argument(4); 5959 Node* dest_offset = argument(5); 5960 5961 // (1) src and dest are arrays. 5962 const Type* src_type = src->Value(&_gvn); 5963 const Type* dest_type = dest->Value(&_gvn); 5964 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5965 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5966 assert(top_src != NULL && top_src->klass() != NULL && 5967 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5968 5969 // checks are the responsibility of the caller 5970 Node* src_start = src; 5971 Node* dest_start = dest; 5972 if (src_offset != NULL || dest_offset != NULL) { 5973 assert(src_offset != NULL && dest_offset != NULL, ""); 5974 src_start = array_element_address(src, src_offset, T_BYTE); 5975 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5976 } 5977 5978 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5979 // (because of the predicated logic executed earlier). 5980 // so we cast it here safely. 5981 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5982 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5983 if (embeddedCipherObj == NULL) return false; 5984 // cast it to what we know it will be at runtime 5985 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 5986 assert(tinst != NULL, "CTR obj is null"); 5987 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 5988 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5989 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5990 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5991 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5992 const TypeOopPtr* xtype = aklass->as_instance_type(); 5993 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5994 aescrypt_object = _gvn.transform(aescrypt_object); 5995 // we need to get the start of the aescrypt_object's expanded key array 5996 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5997 if (k_start == NULL) return false; 5998 // similarly, get the start address of the r vector 5999 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6000 if (obj_counter == NULL) return false; 6001 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6002 6003 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6004 if (saved_encCounter == NULL) return false; 6005 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6006 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6007 6008 Node* ctrCrypt; 6009 if (Matcher::pass_original_key_for_aes()) { 6010 // no SPARC version for AES/CTR intrinsics now. 6011 return false; 6012 } 6013 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6014 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6015 OptoRuntime::counterMode_aescrypt_Type(), 6016 stubAddr, stubName, TypePtr::BOTTOM, 6017 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6018 6019 // return cipher length (int) 6020 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6021 set_result(retvalue); 6022 return true; 6023 } 6024 6025 //------------------------------get_key_start_from_aescrypt_object----------------------- 6026 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6027 #if defined(PPC64) || defined(S390) 6028 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6029 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6030 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6031 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6032 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6033 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6034 if (objSessionK == NULL) { 6035 return (Node *) NULL; 6036 } 6037 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6038 #else 6039 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6040 #endif // PPC64 6041 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6042 if (objAESCryptKey == NULL) return (Node *) NULL; 6043 6044 // now have the array, need to get the start address of the K array 6045 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6046 return k_start; 6047 } 6048 6049 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6050 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6051 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6052 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6053 if (objAESCryptKey == NULL) return (Node *) NULL; 6054 6055 // now have the array, need to get the start address of the lastKey array 6056 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6057 return original_k_start; 6058 } 6059 6060 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6061 // Return node representing slow path of predicate check. 6062 // the pseudo code we want to emulate with this predicate is: 6063 // for encryption: 6064 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6065 // for decryption: 6066 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6067 // note cipher==plain is more conservative than the original java code but that's OK 6068 // 6069 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6070 // The receiver was checked for NULL already. 6071 Node* objCBC = argument(0); 6072 6073 Node* src = argument(1); 6074 Node* dest = argument(4); 6075 6076 // Load embeddedCipher field of CipherBlockChaining object. 6077 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6078 6079 // get AESCrypt klass for instanceOf check 6080 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6081 // will have same classloader as CipherBlockChaining object 6082 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6083 assert(tinst != NULL, "CBCobj is null"); 6084 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6085 6086 // we want to do an instanceof comparison against the AESCrypt class 6087 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6088 if (!klass_AESCrypt->is_loaded()) { 6089 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6090 Node* ctrl = control(); 6091 set_control(top()); // no regular fast path 6092 return ctrl; 6093 } 6094 6095 src = must_be_not_null(src, true); 6096 dest = must_be_not_null(dest, true); 6097 6098 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6099 6100 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6101 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6102 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6103 6104 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6105 6106 // for encryption, we are done 6107 if (!decrypting) 6108 return instof_false; // even if it is NULL 6109 6110 // for decryption, we need to add a further check to avoid 6111 // taking the intrinsic path when cipher and plain are the same 6112 // see the original java code for why. 6113 RegionNode* region = new RegionNode(3); 6114 region->init_req(1, instof_false); 6115 6116 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6117 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6118 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6119 region->init_req(2, src_dest_conjoint); 6120 6121 record_for_igvn(region); 6122 return _gvn.transform(region); 6123 } 6124 6125 //----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6126 // Return node representing slow path of predicate check. 6127 // the pseudo code we want to emulate with this predicate is: 6128 // for encryption: 6129 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6130 // for decryption: 6131 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6132 // note cipher==plain is more conservative than the original java code but that's OK 6133 // 6134 6135 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6136 // The receiver was checked for NULL already. 6137 Node* objCTR = argument(0); 6138 6139 // Load embeddedCipher field of CipherBlockChaining object. 6140 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6141 6142 // get AESCrypt klass for instanceOf check 6143 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6144 // will have same classloader as CipherBlockChaining object 6145 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6146 assert(tinst != NULL, "CTRobj is null"); 6147 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6148 6149 // we want to do an instanceof comparison against the AESCrypt class 6150 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6151 if (!klass_AESCrypt->is_loaded()) { 6152 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6153 Node* ctrl = control(); 6154 set_control(top()); // no regular fast path 6155 return ctrl; 6156 } 6157 6158 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6159 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6160 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6161 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6162 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6163 6164 return instof_false; // even if it is NULL 6165 } 6166 6167 //------------------------------inline_ghash_processBlocks 6168 bool LibraryCallKit::inline_ghash_processBlocks() { 6169 address stubAddr; 6170 const char *stubName; 6171 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6172 6173 stubAddr = StubRoutines::ghash_processBlocks(); 6174 stubName = "ghash_processBlocks"; 6175 6176 Node* data = argument(0); 6177 Node* offset = argument(1); 6178 Node* len = argument(2); 6179 Node* state = argument(3); 6180 Node* subkeyH = argument(4); 6181 6182 state = must_be_not_null(state, true); 6183 subkeyH = must_be_not_null(subkeyH, true); 6184 data = must_be_not_null(data, true); 6185 6186 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6187 assert(state_start, "state is NULL"); 6188 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6189 assert(subkeyH_start, "subkeyH is NULL"); 6190 Node* data_start = array_element_address(data, offset, T_BYTE); 6191 assert(data_start, "data is NULL"); 6192 6193 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6194 OptoRuntime::ghash_processBlocks_Type(), 6195 stubAddr, stubName, TypePtr::BOTTOM, 6196 state_start, subkeyH_start, data_start, len); 6197 return true; 6198 } 6199 6200 bool LibraryCallKit::inline_base64_encodeBlock() { 6201 address stubAddr; 6202 const char *stubName; 6203 assert(UseBASE64Intrinsics, "need Base64 intrinsics support"); 6204 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters"); 6205 stubAddr = StubRoutines::base64_encodeBlock(); 6206 stubName = "encodeBlock"; 6207 6208 if (!stubAddr) return false; 6209 Node* base64obj = argument(0); 6210 Node* src = argument(1); 6211 Node* offset = argument(2); 6212 Node* len = argument(3); 6213 Node* dest = argument(4); 6214 Node* dp = argument(5); 6215 Node* isURL = argument(6); 6216 6217 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 6218 assert(src_start, "source array is NULL"); 6219 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE); 6220 assert(dest_start, "destination array is NULL"); 6221 6222 Node* base64 = make_runtime_call(RC_LEAF, 6223 OptoRuntime::base64_encodeBlock_Type(), 6224 stubAddr, stubName, TypePtr::BOTTOM, 6225 src_start, offset, len, dest_start, dp, isURL); 6226 return true; 6227 } 6228 6229 //------------------------------inline_sha_implCompress----------------------- 6230 // 6231 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6232 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6233 // 6234 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6235 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6236 // 6237 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6238 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6239 // 6240 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6241 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6242 6243 Node* sha_obj = argument(0); 6244 Node* src = argument(1); // type oop 6245 Node* ofs = argument(2); // type int 6246 6247 const Type* src_type = src->Value(&_gvn); 6248 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6249 if (top_src == NULL || top_src->klass() == NULL) { 6250 // failed array check 6251 return false; 6252 } 6253 // Figure out the size and type of the elements we will be copying. 6254 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6255 if (src_elem != T_BYTE) { 6256 return false; 6257 } 6258 // 'src_start' points to src array + offset 6259 src = must_be_not_null(src, true); 6260 Node* src_start = array_element_address(src, ofs, src_elem); 6261 Node* state = NULL; 6262 address stubAddr; 6263 const char *stubName; 6264 6265 switch(id) { 6266 case vmIntrinsics::_sha_implCompress: 6267 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6268 state = get_state_from_sha_object(sha_obj); 6269 stubAddr = StubRoutines::sha1_implCompress(); 6270 stubName = "sha1_implCompress"; 6271 break; 6272 case vmIntrinsics::_sha2_implCompress: 6273 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6274 state = get_state_from_sha_object(sha_obj); 6275 stubAddr = StubRoutines::sha256_implCompress(); 6276 stubName = "sha256_implCompress"; 6277 break; 6278 case vmIntrinsics::_sha5_implCompress: 6279 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6280 state = get_state_from_sha5_object(sha_obj); 6281 stubAddr = StubRoutines::sha512_implCompress(); 6282 stubName = "sha512_implCompress"; 6283 break; 6284 default: 6285 fatal_unexpected_iid(id); 6286 return false; 6287 } 6288 if (state == NULL) return false; 6289 6290 // Call the stub. 6291 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6292 stubAddr, stubName, TypePtr::BOTTOM, 6293 src_start, state); 6294 6295 return true; 6296 } 6297 6298 //------------------------------inline_digestBase_implCompressMB----------------------- 6299 // 6300 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6301 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6302 // 6303 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6304 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6305 "need SHA1/SHA256/SHA512 instruction support"); 6306 assert((uint)predicate < 3, "sanity"); 6307 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6308 6309 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6310 Node* src = argument(1); // byte[] array 6311 Node* ofs = argument(2); // type int 6312 Node* limit = argument(3); // type int 6313 6314 const Type* src_type = src->Value(&_gvn); 6315 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6316 if (top_src == NULL || top_src->klass() == NULL) { 6317 // failed array check 6318 return false; 6319 } 6320 // Figure out the size and type of the elements we will be copying. 6321 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6322 if (src_elem != T_BYTE) { 6323 return false; 6324 } 6325 // 'src_start' points to src array + offset 6326 src = must_be_not_null(src, false); 6327 Node* src_start = array_element_address(src, ofs, src_elem); 6328 6329 const char* klass_SHA_name = NULL; 6330 const char* stub_name = NULL; 6331 address stub_addr = NULL; 6332 bool long_state = false; 6333 6334 switch (predicate) { 6335 case 0: 6336 if (UseSHA1Intrinsics) { 6337 klass_SHA_name = "sun/security/provider/SHA"; 6338 stub_name = "sha1_implCompressMB"; 6339 stub_addr = StubRoutines::sha1_implCompressMB(); 6340 } 6341 break; 6342 case 1: 6343 if (UseSHA256Intrinsics) { 6344 klass_SHA_name = "sun/security/provider/SHA2"; 6345 stub_name = "sha256_implCompressMB"; 6346 stub_addr = StubRoutines::sha256_implCompressMB(); 6347 } 6348 break; 6349 case 2: 6350 if (UseSHA512Intrinsics) { 6351 klass_SHA_name = "sun/security/provider/SHA5"; 6352 stub_name = "sha512_implCompressMB"; 6353 stub_addr = StubRoutines::sha512_implCompressMB(); 6354 long_state = true; 6355 } 6356 break; 6357 default: 6358 fatal("unknown SHA intrinsic predicate: %d", predicate); 6359 } 6360 if (klass_SHA_name != NULL) { 6361 // get DigestBase klass to lookup for SHA klass 6362 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6363 assert(tinst != NULL, "digestBase_obj is not instance???"); 6364 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6365 6366 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6367 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6368 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6369 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6370 } 6371 return false; 6372 } 6373 //------------------------------inline_sha_implCompressMB----------------------- 6374 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6375 bool long_state, address stubAddr, const char *stubName, 6376 Node* src_start, Node* ofs, Node* limit) { 6377 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6378 const TypeOopPtr* xtype = aklass->as_instance_type(); 6379 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6380 sha_obj = _gvn.transform(sha_obj); 6381 6382 Node* state; 6383 if (long_state) { 6384 state = get_state_from_sha5_object(sha_obj); 6385 } else { 6386 state = get_state_from_sha_object(sha_obj); 6387 } 6388 if (state == NULL) return false; 6389 6390 // Call the stub. 6391 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6392 OptoRuntime::digestBase_implCompressMB_Type(), 6393 stubAddr, stubName, TypePtr::BOTTOM, 6394 src_start, state, ofs, limit); 6395 // return ofs (int) 6396 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6397 set_result(result); 6398 6399 return true; 6400 } 6401 6402 enum VectorApiObjectType { 6403 VECAPI_VECTOR, 6404 VECAPI_MASK, 6405 VECAPI_SPECIES, 6406 VECAPI_SHUFFLE, 6407 }; 6408 6409 static ciKlass* get_exact_klass_for_vector_box(ciKlass* ctx, BasicType bt, int num_elem, VectorApiObjectType obj_type) { 6410 // FIXME: use VM symbol table instead 6411 ResourceMark rm; 6412 stringStream ss; 6413 ss.print_raw("jdk/incubator/vector/"); 6414 switch (bt) { 6415 case T_BYTE: ss.print_raw("Byte"); break; 6416 case T_SHORT: ss.print_raw("Short"); break; 6417 case T_INT: ss.print_raw("Int"); break; 6418 case T_LONG: ss.print_raw("Long"); break; 6419 case T_FLOAT: ss.print_raw("Float"); break; 6420 case T_DOUBLE: ss.print_raw("Double"); break; 6421 default: fatal("unknown element type: %d", bt); 6422 } 6423 int bits = num_elem * BitsPerByte * type2aelembytes(bt); 6424 switch (bits) { 6425 case 64: // fall through 6426 case 128: // fall through 6427 case 256: // fall through 6428 case 512: ss.print("%d", bits); break; 6429 default: fatal("unknown vector size: %d", bits); 6430 } 6431 ss.print_raw("Vector"); 6432 if (obj_type != VECAPI_VECTOR) { 6433 ss.print_raw("$"); 6434 switch (bt) { 6435 case T_BYTE: ss.print_raw("Byte"); break; 6436 case T_SHORT: ss.print_raw("Short"); break; 6437 case T_INT: ss.print_raw("Int"); break; 6438 case T_LONG: ss.print_raw("Long"); break; 6439 case T_FLOAT: ss.print_raw("Float"); break; 6440 case T_DOUBLE: ss.print_raw("Double"); break; 6441 default: fatal("unknown element type: %d", bt); 6442 } 6443 switch (bits) { 6444 case 64: // fall through 6445 case 128: // fall through 6446 case 256: // fall through 6447 case 512: ss.print("%d", bits); break; 6448 default: fatal("unknown vector size: %d", bits); 6449 } 6450 switch (obj_type) { 6451 case VECAPI_MASK: ss.print_raw("Mask"); break; 6452 case VECAPI_SPECIES: ss.print_raw("Species"); break; 6453 case VECAPI_SHUFFLE: ss.print_raw("Shuffle"); break; 6454 default: fatal("unknown vector object type"); 6455 } 6456 } 6457 6458 ciSymbol* vector_klass_name = ciSymbol::make(ss.as_string()); 6459 ciKlass* vector_klass = ctx->find_klass(vector_klass_name); 6460 return vector_klass; 6461 } 6462 6463 enum VectorMaskUseType { 6464 VecMaskUseLoad, 6465 VecMaskUseStore, 6466 VecMaskUseAll, 6467 VecMaskNotUsed 6468 }; 6469 6470 static bool arch_supports_vector(int op, int num_elem, BasicType type, VectorMaskUseType mask_use_type, int op_arity = 0) { 6471 // Check that the operation is valid. 6472 if (op <= 0) { 6473 #ifndef PRODUCT 6474 if (DebugVectorApi) { 6475 tty->print_cr("Rejected intrinsification because no valid vector op could be extracted"); 6476 } 6477 #endif 6478 return false; 6479 } 6480 6481 // Check that architecture supports this op-size-type combination. 6482 if (!Matcher::match_rule_supported_vector(op, num_elem, type, op_arity)) { 6483 #ifndef PRODUCT 6484 if (DebugVectorApi) { 6485 tty->print_cr("Rejected vector op (%s,%s,%d) because architecture does not support it", 6486 NodeClassNames[op], type2name(type), num_elem); 6487 } 6488 #endif 6489 return false; 6490 } else { 6491 assert(Matcher::match_rule_supported(op), "must be supported"); 6492 } 6493 6494 // Check whether mask unboxing is supported. 6495 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseLoad) { 6496 if (!Matcher::match_rule_supported_vector(Op_VectorLoadMask, num_elem, type)) { 6497 #ifndef PRODUCT 6498 if (DebugVectorApi) { 6499 tty->print_cr("Rejected vector mask loading (%s,%s,%d) because architecture does not support it", 6500 NodeClassNames[Op_VectorLoadMask], type2name(type), num_elem); 6501 } 6502 #endif 6503 return false; 6504 } 6505 } 6506 6507 // Check whether mask boxing is supported. 6508 if (mask_use_type == VecMaskUseAll || mask_use_type == VecMaskUseStore) { 6509 if (!Matcher::match_rule_supported_vector(Op_VectorStoreMask, num_elem, type)) { 6510 #ifndef PRODUCT 6511 if (DebugVectorApi) { 6512 tty->print_cr("Rejected vector mask storing (%s,%s,%d) because architecture does not support it", 6513 NodeClassNames[Op_VectorStoreMask], type2name(type), num_elem); 6514 } 6515 #endif 6516 return false; 6517 } 6518 } 6519 6520 return true; 6521 } 6522 6523 // Should be aligned with constants in jdk.incubator.vector.VectorIntrinsics. 6524 enum { 6525 // Unary 6526 OP_ABS = 0, 6527 OP_NEG = 1, 6528 OP_SQRT = 2, 6529 OP_NOT = 3, 6530 // Binary 6531 OP_ADD = 4, 6532 OP_SUB = 5, 6533 OP_MUL = 6, 6534 OP_DIV = 7, 6535 OP_MIN = 8, 6536 OP_MAX = 9, 6537 OP_AND = 10, 6538 OP_OR = 11, 6539 OP_XOR = 12, 6540 // Ternary 6541 OP_FMA = 13, 6542 // Broadcast int 6543 OP_LSHIFT = 14, 6544 OP_RSHIFT = 15, 6545 OP_URSHIFT = 16, 6546 // Vector Math Library 6547 OP_TAN = 101, 6548 OP_SVML_START = OP_TAN, 6549 OP_TANH = 102, 6550 OP_SIN = 103, 6551 OP_SINH = 104, 6552 OP_COS = 105, 6553 OP_COSH = 106, 6554 OP_ASIN = 107, 6555 OP_ACOS = 108, 6556 OP_ATAN = 109, 6557 OP_ATAN2 = 110, 6558 OP_CBRT = 111, 6559 OP_LOG = 112, 6560 OP_LOG10 = 113, 6561 OP_LOG1P = 114, 6562 OP_POW = 115, 6563 OP_EXP = 116, 6564 OP_EXPM1 = 117, 6565 OP_HYPOT = 118, 6566 OP_SVML_END = OP_HYPOT, 6567 }; 6568 6569 static int get_opc(jint op, BasicType bt) { 6570 switch (op) { 6571 case OP_ADD: { 6572 switch (bt) { 6573 case T_BYTE: // fall-through 6574 case T_SHORT: // fall-through 6575 case T_INT: return Op_AddI; 6576 case T_LONG: return Op_AddL; 6577 case T_FLOAT: return Op_AddF; 6578 case T_DOUBLE: return Op_AddD; 6579 default: fatal("ADD: %s", type2name(bt)); 6580 } 6581 break; 6582 } 6583 case OP_SUB: { 6584 switch (bt) { 6585 case T_BYTE: // fall-through 6586 case T_SHORT: // fall-through 6587 case T_INT: return Op_SubI; 6588 case T_LONG: return Op_SubL; 6589 case T_FLOAT: return Op_SubF; 6590 case T_DOUBLE: return Op_SubD; 6591 default: fatal("SUB: %s", type2name(bt)); 6592 } 6593 break; 6594 } 6595 case OP_MUL: { 6596 switch (bt) { 6597 case T_BYTE: // fall-through 6598 case T_SHORT: // fall-through 6599 case T_INT: return Op_MulI; 6600 case T_LONG: return Op_MulL; 6601 case T_FLOAT: return Op_MulF; 6602 case T_DOUBLE: return Op_MulD; 6603 default: fatal("MUL: %s", type2name(bt)); 6604 } 6605 break; 6606 } 6607 case OP_DIV: { 6608 switch (bt) { 6609 case T_BYTE: // fall-through 6610 case T_SHORT: // fall-through 6611 case T_INT: return Op_DivI; 6612 case T_LONG: return Op_DivL; 6613 case T_FLOAT: return Op_DivF; 6614 case T_DOUBLE: return Op_DivD; 6615 default: fatal("DIV: %s", type2name(bt)); 6616 } 6617 break; 6618 } 6619 case OP_MIN: { 6620 switch (bt) { 6621 case T_BYTE: 6622 case T_SHORT: 6623 case T_INT: return Op_MinI; 6624 case T_LONG: return Op_MinL; 6625 case T_FLOAT: return Op_MinF; 6626 case T_DOUBLE: return Op_MinD; 6627 default: fatal("MIN: %s", type2name(bt)); 6628 } 6629 break; 6630 } 6631 case OP_MAX: { 6632 switch (bt) { 6633 case T_BYTE: 6634 case T_SHORT: 6635 case T_INT: return Op_MaxI; 6636 case T_LONG: return Op_MaxL; 6637 case T_FLOAT: return Op_MaxF; 6638 case T_DOUBLE: return Op_MaxD; 6639 default: fatal("MAX: %s", type2name(bt)); 6640 } 6641 break; 6642 } 6643 case OP_ABS: { 6644 switch (bt) { 6645 case T_BYTE: // fall-through 6646 case T_SHORT: // fall-through 6647 case T_LONG: // fall-through 6648 case T_INT: return Op_AbsI; 6649 case T_FLOAT: return Op_AbsF; 6650 case T_DOUBLE: return Op_AbsD; 6651 default: fatal("ABS: %s", type2name(bt)); 6652 } 6653 break; 6654 } 6655 case OP_NEG: { 6656 switch (bt) { 6657 case T_BYTE: // fall-through 6658 case T_SHORT: // fall-through 6659 case T_INT: return Op_NegI; 6660 case T_FLOAT: return Op_NegF; 6661 case T_DOUBLE: return Op_NegD; 6662 default: fatal("NEG: %s", type2name(bt)); 6663 } 6664 break; 6665 } 6666 case OP_AND: { 6667 switch (bt) { 6668 case T_BYTE: // fall-through 6669 case T_SHORT: // fall-through 6670 case T_INT: return Op_AndI; 6671 case T_LONG: return Op_AndL; 6672 default: fatal("AND: %s", type2name(bt)); 6673 } 6674 break; 6675 } 6676 case OP_OR: { 6677 switch (bt) { 6678 case T_BYTE: // fall-through 6679 case T_SHORT: // fall-through 6680 case T_INT: return Op_OrI; 6681 case T_LONG: return Op_OrL; 6682 default: fatal("OR: %s", type2name(bt)); 6683 } 6684 break; 6685 } 6686 case OP_XOR: { 6687 switch (bt) { 6688 case T_BYTE: // fall-through 6689 case T_SHORT: // fall-through 6690 case T_INT: return Op_XorI; 6691 case T_LONG: return Op_XorL; 6692 default: fatal("XOR: %s", type2name(bt)); 6693 } 6694 break; 6695 } 6696 case OP_SQRT: { 6697 switch (bt) { 6698 case T_FLOAT: return Op_SqrtF; 6699 case T_DOUBLE: return Op_SqrtD; 6700 default: fatal("SQRT: %s", type2name(bt)); 6701 } 6702 break; 6703 } 6704 case OP_NOT: { 6705 switch (bt) { 6706 case T_BYTE: // fall-through 6707 case T_SHORT: // fall-through 6708 case T_INT: // fall-through 6709 case T_LONG: return Op_Not; 6710 default: fatal("NOT: %s", type2name(bt)); 6711 } 6712 break; 6713 } 6714 case OP_FMA: { 6715 switch (bt) { 6716 case T_FLOAT: return Op_FmaF; 6717 case T_DOUBLE: return Op_FmaD; 6718 default: fatal("FMA: %s", type2name(bt)); 6719 } 6720 break; 6721 } 6722 case OP_LSHIFT: { 6723 switch (bt) { 6724 case T_BYTE: // fall-through 6725 case T_SHORT: // fall-through 6726 case T_INT: return Op_LShiftI; 6727 case T_LONG: return Op_LShiftL; 6728 default: fatal("LSHIFT: %s", type2name(bt)); 6729 } 6730 break; 6731 } 6732 case OP_RSHIFT: { 6733 switch (bt) { 6734 case T_BYTE: // fall-through 6735 case T_SHORT: // fall-through 6736 case T_INT: return Op_RShiftI; 6737 case T_LONG: return Op_RShiftL; 6738 default: fatal("RSHIFT: %s", type2name(bt)); 6739 } 6740 break; 6741 } 6742 case OP_URSHIFT: { 6743 switch (bt) { 6744 case T_BYTE: // fall-through 6745 case T_SHORT: // fall-through 6746 case T_INT: return Op_URShiftI; 6747 case T_LONG: return Op_URShiftL; 6748 default: fatal("URSHIFT: %s", type2name(bt)); 6749 } 6750 break; 6751 } 6752 case OP_TAN: 6753 case OP_TANH: 6754 case OP_SIN: 6755 case OP_SINH: 6756 case OP_COS: 6757 case OP_COSH: 6758 case OP_ASIN: 6759 case OP_ACOS: 6760 case OP_ATAN: 6761 case OP_ATAN2: 6762 case OP_CBRT: 6763 case OP_LOG: 6764 case OP_LOG10: 6765 case OP_LOG1P: 6766 case OP_POW: 6767 case OP_EXP: 6768 case OP_EXPM1: 6769 case OP_HYPOT: 6770 return Op_CallLeafVector; 6771 default: fatal("unknown op: %d", op); 6772 } 6773 return 0; // Unimplemented 6774 } 6775 6776 Node* LibraryCallKit::box_vector(Node* vector, const TypeInstPtr* vbox_type, 6777 BasicType elem_bt, int num_elem) { 6778 6779 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem); 6780 6781 VectorBoxAllocateNode* alloc = new VectorBoxAllocateNode(C, vbox_type); 6782 set_edges_for_java_call(alloc, /*must_throw=*/false, /*separate_io_proj=*/true); 6783 make_slow_call_ex(alloc, env()->Throwable_klass(), /*separate_io_proj=*/true); 6784 set_i_o(_gvn.transform( new ProjNode(alloc, TypeFunc::I_O) )); 6785 set_all_memory(_gvn.transform( new ProjNode(alloc, TypeFunc::Memory) )); 6786 Node* ret = _gvn.transform(new ProjNode(alloc, TypeFunc::Parms)); 6787 6788 VectorBoxNode* vbox = new VectorBoxNode(C, ret, vector, vbox_type, vec_type); 6789 return _gvn.transform(vbox); 6790 } 6791 6792 Node* LibraryCallKit::unbox_vector(Node* v, const TypeInstPtr* vbox_type, BasicType elem_bt, int num_elem) { 6793 const TypeInstPtr* vbox_type_v = gvn().type(v)->is_instptr(); 6794 if (vbox_type->klass() != vbox_type_v->klass()) { 6795 return NULL; // arguments don't agree on vector shapes 6796 } 6797 if (vbox_type_v->maybe_null()) { 6798 return NULL; // no nulls are allowed 6799 } 6800 const TypeVect* vec_type = TypeVect::make(elem_bt, num_elem); 6801 Node* unbox = gvn().transform(new VectorUnboxNode(C, vec_type, v, merged_memory())); 6802 return unbox; 6803 } 6804 6805 void LibraryCallKit::set_vector_result(Node* result, bool set_res) { 6806 if (DebugVectorApi) { 6807 #ifndef PRODUCT 6808 tty->print("============ "); 6809 callee()->print(); 6810 tty->print_cr(" ============"); 6811 result->dump(5); 6812 tty->print_cr("----------------------------------------------------"); 6813 #endif 6814 } 6815 if (set_res) { 6816 set_result(result); 6817 } 6818 } 6819 6820 bool LibraryCallKit::inline_vector_nary_operation(int n) { 6821 const TypeInt* opr = gvn().type(argument(0))->is_int(); 6822 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr(); 6823 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr(); 6824 const TypeInt* vlen = gvn().type(argument(3))->is_int(); 6825 6826 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) { 6827 return false; // not enough info for intrinsification 6828 } 6829 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 6830 if (!elem_type->is_primitive_type()) { 6831 return false; // should be primitive type 6832 } 6833 BasicType elem_bt = elem_type->basic_type(); 6834 int num_elem = vlen->get_con(); 6835 int opc = get_opc(opr->get_con(), elem_bt); 6836 int sopc = opc != Op_CallLeafVector ? VectorNode::opcode(opc, elem_bt) : Op_CallLeafVector; 6837 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 6838 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 6839 6840 if (opc == Op_CallLeafVector && !Matcher::supports_vector_calling_convention()) { 6841 return false; 6842 } 6843 6844 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad. 6845 if (!arch_supports_vector(sopc, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseAll : VecMaskNotUsed, n)) { 6846 return false; // not supported 6847 } 6848 6849 Node* opd1 = NULL; Node* opd2 = NULL; Node* opd3 = NULL; 6850 switch (n) { 6851 case 3: { 6852 opd3 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem); 6853 if (opd3 == NULL) { 6854 return false; 6855 } 6856 // fall-through 6857 } 6858 case 2: { 6859 opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem); 6860 if (opd2 == NULL) { 6861 return false; 6862 } 6863 // fall-through 6864 } 6865 case 1: { 6866 opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem); 6867 if (opd1 == NULL) { 6868 return false; 6869 } 6870 break; 6871 } 6872 default: fatal("unsupported arity: %d", n); 6873 } 6874 6875 Node* operation = NULL; 6876 if (sopc == Op_CallLeafVector) { 6877 operation = gen_call_to_svml(opr->get_con(), elem_bt, num_elem, opd1, opd2); 6878 if (operation == NULL) { 6879 return false; 6880 } 6881 } else { 6882 switch (n) { 6883 case 1: 6884 case 2: { 6885 operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, num_elem, elem_bt)); 6886 break; 6887 } 6888 case 3: { 6889 operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, opd3, num_elem, elem_bt)); 6890 break; 6891 } 6892 default: fatal("unsupported arity: %d", n); 6893 } 6894 } 6895 // Wrap it up in VectorBox to keep object type information. 6896 operation = box_vector(operation, vbox_type, elem_bt, num_elem); 6897 set_vector_result(operation); 6898 6899 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt)))); 6900 return true; 6901 } 6902 6903 // <V extends Vector<?,?>> 6904 // V broadcastCoerced(Class<?> vectorClass, Class<?> elementType, int vlen, 6905 // long bits, 6906 // LongFunction<V> defaultImpl) 6907 bool LibraryCallKit::inline_vector_broadcast_coerced() { 6908 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr(); 6909 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr(); 6910 const TypeInt* vlen = gvn().type(argument(2))->is_int(); 6911 6912 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) { 6913 return false; // not enough info for intrinsification 6914 } 6915 6916 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 6917 if (!elem_type->is_primitive_type()) { 6918 return false; // should be primitive type 6919 } 6920 BasicType elem_bt = elem_type->basic_type(); 6921 int num_elem = vlen->get_con(); 6922 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 6923 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 6924 6925 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad. 6926 if (!arch_supports_vector(VectorNode::replicate_opcode(elem_bt), num_elem, elem_bt, 6927 vbox_klass->is_vectormask() ? VecMaskUseStore : VecMaskNotUsed)) { 6928 return false; // not supported 6929 } 6930 6931 Node* bits = argument(3); // long 6932 6933 Node* elem = NULL; 6934 switch (elem_bt) { 6935 case T_BOOLEAN: // fall-through 6936 case T_BYTE: // fall-through 6937 case T_SHORT: // fall-through 6938 case T_CHAR: // fall-through 6939 case T_INT: { 6940 elem = gvn().transform(new ConvL2INode(bits)); 6941 break; 6942 } 6943 case T_DOUBLE: { 6944 elem = gvn().transform(new MoveL2DNode(bits)); 6945 break; 6946 } 6947 case T_FLOAT: { 6948 bits = gvn().transform(new ConvL2INode(bits)); 6949 elem = gvn().transform(new MoveI2FNode(bits)); 6950 break; 6951 } 6952 case T_LONG: { 6953 elem = bits; // no conversion needed 6954 break; 6955 } 6956 default: fatal("%s", type2name(elem_bt)); 6957 } 6958 6959 Node* broadcast = VectorNode::scalar2vector(elem, num_elem, Type::get_const_basic_type(elem_bt)); 6960 broadcast = gvn().transform(broadcast); 6961 Node* box = box_vector(broadcast, vbox_type, elem_bt, num_elem); 6962 set_vector_result(box); 6963 6964 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt)))); 6965 return true; 6966 } 6967 6968 // <C, V extends Vector<?,?>> 6969 // V load(Class<?> vectorClass, Class<?> elementType, int vlen, 6970 // Object base, long offset, 6971 // /* Vector.Mask<E,S> m*/ 6972 // Object container, int index, 6973 // LoadVectorOperation<C, V> defaultImpl) { 6974 // 6975 // <C, V extends Vector<?,?>> 6976 // void store(Class<?> vectorClass, Class<?> elementType, int vlen, 6977 // Object base, long offset, 6978 // V v, /*Vector.Mask<E,S> m*/ 6979 // Object container, int index, 6980 // StoreVectorOperation<C, V> defaultImpl) { 6981 6982 bool LibraryCallKit::inline_vector_mem_operation(bool is_store) { 6983 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr(); 6984 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr(); 6985 const TypeInt* vlen = gvn().type(argument(2))->is_int(); 6986 6987 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) { 6988 return false; // not enough info for intrinsification 6989 } 6990 6991 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 6992 if (!elem_type->is_primitive_type()) { 6993 return false; // should be primitive type 6994 } 6995 BasicType elem_bt = elem_type->basic_type(); 6996 int num_elem = vlen->get_con(); 6997 6998 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad. 6999 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, num_elem, elem_bt, VecMaskNotUsed)) { 7000 return false; // not supported 7001 } 7002 7003 Node* base = argument(3); 7004 Node* offset = ConvL2X(argument(4)); 7005 Node* addr = make_unsafe_address(base, offset, elem_bt, true); 7006 7007 const TypePtr *addr_type = gvn().type(addr)->isa_ptr(); 7008 const TypeAryPtr* arr_type = addr_type->isa_aryptr(); 7009 7010 // Now handle special case where load/store happens from/to byte array but element type is not byte. 7011 bool using_byte_array = arr_type != NULL && arr_type->elem()->array_element_basic_type() == T_BYTE && elem_bt != T_BYTE; 7012 7013 // It must be the case that if it is not special byte array case, there is consistency between 7014 // array and vector element types. 7015 if (!using_byte_array && arr_type != NULL && elem_bt != arr_type->elem()->array_element_basic_type()) { 7016 return false; 7017 } 7018 7019 // Since we are using byte array, we need to double check that the byte operations are supported by backend. 7020 if (using_byte_array) { 7021 int byte_num_elem = num_elem * type2aelembytes(elem_bt); 7022 if (!arch_supports_vector(is_store ? Op_StoreVector : Op_LoadVector, byte_num_elem, T_BYTE, VecMaskNotUsed)) { 7023 return false; // not supported 7024 } 7025 } 7026 7027 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7028 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 7029 7030 if (is_store) { 7031 Node* val = unbox_vector(argument(6), vbox_type, elem_bt, num_elem); 7032 if (val == NULL) { 7033 return false; // operand unboxing failed 7034 } 7035 set_all_memory(reset_memory()); 7036 7037 // In case the store needs to happen to byte array, reinterpret the incoming vector to byte vector. 7038 int store_num_elem = num_elem; 7039 if (using_byte_array) { 7040 store_num_elem = num_elem * type2aelembytes(elem_bt); 7041 const TypeVect* to_vect_type = TypeVect::make(T_BYTE, store_num_elem); 7042 val = gvn().transform(new VectorReinterpretNode(val, val->bottom_type()->is_vect(), to_vect_type)); 7043 } 7044 7045 Node* vstore = gvn().transform(StoreVectorNode::make(0, control(), memory(addr), addr, addr_type, val, store_num_elem)); 7046 set_memory(vstore, addr_type); 7047 set_vector_result(vstore, false); 7048 } else { 7049 // When using byte array, we need to load as byte then reinterpret the value. Otherwise, do a simple vector load. 7050 Node* vload = NULL; 7051 if (using_byte_array) { 7052 int load_num_elem = num_elem * type2aelembytes(elem_bt); 7053 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, load_num_elem, T_BYTE)); 7054 const TypeVect* to_vect_type = TypeVect::make(elem_bt, num_elem); 7055 vload = gvn().transform(new VectorReinterpretNode(vload, vload->bottom_type()->is_vect(), to_vect_type)); 7056 } else { 7057 vload = gvn().transform(LoadVectorNode::make(0, control(), memory(addr), addr, addr_type, num_elem, elem_bt)); 7058 } 7059 7060 Node* box = box_vector(vload, vbox_type, elem_bt, num_elem); 7061 set_vector_result(box); 7062 } 7063 7064 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt)))); 7065 return true; 7066 } 7067 7068 // <V extends Vector<?,?>> 7069 // long reductionCoerced(int oprId, Class<?> vectorClass, Class<?> elementType, int vlen, 7070 // V v, 7071 // Function<V,Long> defaultImpl) 7072 7073 bool LibraryCallKit::inline_vector_reduction() { 7074 const TypeInt* opr = gvn().type(argument(0))->is_int(); 7075 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr(); 7076 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr(); 7077 const TypeInt* vlen = gvn().type(argument(3))->is_int(); 7078 7079 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) { 7080 return false; // not enough info for intrinsification 7081 } 7082 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 7083 if (!elem_type->is_primitive_type()) { 7084 return false; // should be primitive type 7085 } 7086 BasicType elem_bt = elem_type->basic_type(); 7087 int num_elem = vlen->get_con(); 7088 7089 int opc = get_opc(opr->get_con(), elem_bt); 7090 int sopc = ReductionNode::opcode(opc, elem_bt); 7091 7092 // FIXME: When encountering a SubReduction, we want to check for support of 7093 // the corresponding AddReduction node. 7094 if (sopc == Op_SubReductionV) { 7095 if (gvn().type(argument(2))->isa_int()) { 7096 sopc = Op_AddReductionVI; 7097 } else if (gvn().type(argument(2))->isa_long()) { 7098 sopc = Op_AddReductionVL; 7099 } 7100 } 7101 7102 // TODO When mask usage is supported, VecMaskNotUsed needs to be VecMaskUseLoad. 7103 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) { 7104 return false; 7105 } 7106 7107 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7108 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 7109 7110 Node* opd = unbox_vector(argument(4), vbox_type, elem_bt, num_elem); 7111 if (opd == NULL) { 7112 return false; // operand unboxing failed 7113 } 7114 7115 Node* init = ReductionNode::make_reduction_input(_gvn, opc, elem_bt); 7116 Node* rn = gvn().transform(ReductionNode::make(opc, NULL, init, opd, elem_bt)); 7117 7118 Node* bits = NULL; 7119 switch (elem_bt) { 7120 case T_BYTE: 7121 case T_SHORT: 7122 case T_INT: { 7123 bits = gvn().transform(new ConvI2LNode(rn)); 7124 break; 7125 } 7126 case T_FLOAT: { 7127 rn = gvn().transform(new MoveF2INode(rn)); 7128 bits = gvn().transform(new ConvI2LNode(rn)); 7129 break; 7130 } 7131 case T_DOUBLE: { 7132 bits = gvn().transform(new MoveD2LNode(rn)); 7133 break; 7134 } 7135 case T_LONG: { 7136 bits = rn; // no conversion needed 7137 break; 7138 } 7139 default: fatal("%s", type2name(elem_bt)); 7140 } 7141 set_vector_result(bits); 7142 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt)))); 7143 return true; 7144 } 7145 7146 // static <V> boolean test(int cond, Class<?> vectorClass, Class<?> elementType, int vlen, 7147 // V v1, V v2, 7148 // BiFunction<V, V, Boolean> defaultImpl) { 7149 7150 bool LibraryCallKit::inline_vector_test() { 7151 const TypeInt* cond = gvn().type(argument(0))->is_int(); 7152 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr(); 7153 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr(); 7154 const TypeInt* vlen = gvn().type(argument(3))->is_int(); 7155 7156 if (!cond->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) { 7157 return false; // not enough info for intrinsification 7158 } 7159 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 7160 if (!elem_type->is_primitive_type()) { 7161 return false; // should be primitive type 7162 } 7163 BasicType elem_bt = elem_type->basic_type(); 7164 int num_elem = vlen->get_con(); 7165 Assembler::Condition booltest = (Assembler::Condition)cond->get_con(); 7166 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7167 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 7168 7169 if (!arch_supports_vector(Op_VectorTest, num_elem, elem_bt, vbox_klass->is_vectormask() ? VecMaskUseLoad : VecMaskNotUsed)) { 7170 return false; 7171 } 7172 7173 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem); 7174 Node* opd2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem); 7175 if (opd1 == NULL || opd2 == NULL) { 7176 return false; // operand unboxing failed 7177 } 7178 Node* test = new VectorTestNode(opd1, opd2, booltest); 7179 test = _gvn.transform(test); 7180 set_vector_result(test); 7181 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt)))); 7182 return true; 7183 } 7184 7185 // static 7186 // <V extends Vector, M extends Mask> 7187 // V blend(Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen, 7188 // V v1, V v2, M m, 7189 // VectorBlendOp<V,M> defaultImpl) { ... 7190 // 7191 bool LibraryCallKit::inline_vector_blend() { 7192 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr(); 7193 const TypeInstPtr* mask_klass = gvn().type(argument(1))->is_instptr(); 7194 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr(); 7195 const TypeInt* vlen = gvn().type(argument(3))->is_int(); 7196 7197 if (mask_klass->const_oop() == NULL || vector_klass->const_oop() == NULL || 7198 elem_klass->const_oop() == NULL || !vlen->is_con()) { 7199 return false; // not enough info for intrinsification 7200 } 7201 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 7202 if (!elem_type->is_primitive_type()) { 7203 return false; // should be primitive type 7204 } 7205 BasicType elem_bt = elem_type->basic_type(); 7206 BasicType mask_bt = elem_bt; 7207 int num_elem = vlen->get_con(); 7208 7209 if (!arch_supports_vector(Op_VectorBlend, num_elem, elem_bt, VecMaskUseLoad)) { 7210 return false; // not supported 7211 } 7212 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7213 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 7214 7215 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7216 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass); 7217 7218 Node* v1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem); 7219 Node* v2 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem); 7220 Node* mask = unbox_vector(argument(6), mbox_type, mask_bt, num_elem); 7221 7222 if (v1 == NULL || v2 == NULL || mask == NULL) { 7223 return false; // operand unboxing failed 7224 } 7225 7226 Node* blend = _gvn.transform(new VectorBlendNode(v1, v2, mask)); 7227 Node* box = box_vector(blend, vbox_type, elem_bt, num_elem); 7228 set_vector_result(box); 7229 7230 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt)))); 7231 return true; 7232 } 7233 7234 // static <V extends Vector<E,S>, 7235 // M extends Vector.Mask<E,S>, 7236 // S extends Vector.Shape, E> 7237 // M compare(int cond, Class<V> vectorClass, Class<M> maskClass, Class<?> elementType, int vlen, 7238 // V v1, V v2, 7239 // VectorCompareOp<V,M> defaultImpl) { ... 7240 // 7241 bool LibraryCallKit::inline_vector_compare() { 7242 const TypeInt* cond = gvn().type(argument(0))->is_int(); 7243 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr(); 7244 const TypeInstPtr* mask_klass = gvn().type(argument(2))->is_instptr(); 7245 const TypeInstPtr* elem_klass = gvn().type(argument(3))->is_instptr(); 7246 const TypeInt* vlen = gvn().type(argument(4))->is_int(); 7247 7248 if (!cond->is_con() || vector_klass->const_oop() == NULL || mask_klass->const_oop() == NULL || 7249 elem_klass->const_oop() == NULL || !vlen->is_con()) { 7250 return false; // not enough info for intrinsification 7251 } 7252 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 7253 if (!elem_type->is_primitive_type()) { 7254 return false; // should be primitive type 7255 } 7256 7257 int num_elem = vlen->get_con(); 7258 BasicType elem_bt = elem_type->basic_type(); 7259 BasicType mask_bt = elem_bt; 7260 7261 if (!arch_supports_vector(Op_VectorMaskCmp, num_elem, elem_bt, VecMaskUseStore)) { 7262 return false; 7263 } 7264 7265 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7266 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 7267 7268 ciKlass* mbox_klass = mask_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7269 const TypeInstPtr* mbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, mbox_klass); 7270 7271 Node* v1 = unbox_vector(argument(5), vbox_type, elem_bt, num_elem); 7272 Node* v2 = unbox_vector(argument(6), vbox_type, elem_bt, num_elem); 7273 7274 if (v1 == NULL || v2 == NULL) { 7275 return false; // operand unboxing failed 7276 } 7277 BoolTest::mask pred = (BoolTest::mask)cond->get_con(); 7278 const TypeVect* vt = TypeVect::make(mask_bt, num_elem); 7279 Node* operation = _gvn.transform(new VectorMaskCmpNode(pred, v1, v2, vt)); 7280 7281 Node* box = box_vector(operation, mbox_type, mask_bt, num_elem); 7282 set_vector_result(box); 7283 7284 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt)))); 7285 return true; 7286 } 7287 7288 Node* LibraryCallKit::shift_count(Node* cnt, int shift_op, BasicType bt, int num_elem) { 7289 assert(bt == T_INT || bt == T_LONG, "Only long and int are supported"); 7290 juint mask = (bt == T_INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 7291 const TypeInt* t = cnt->find_int_type(); 7292 if (t != NULL && t->is_con()) { 7293 juint shift = t->get_con(); 7294 if (shift > mask) { 7295 return _gvn.transform(ConNode::make(TypeInt::make(shift & mask))); 7296 } else { 7297 return cnt; 7298 } 7299 } else { 7300 Node* maskedcnt = cnt; 7301 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 7302 Node* nmask = _gvn.transform(ConNode::make(TypeInt::make(mask))); 7303 maskedcnt = _gvn.transform(new AndINode(cnt, nmask)); 7304 } 7305 return VectorNode::shift_count(shift_op, maskedcnt, num_elem, bt); 7306 } 7307 } 7308 7309 static void get_svml_address(int op, int bits, BasicType bt, const char** name_ptr, address* addr_ptr) { 7310 assert(name_ptr != NULL, "unexpected"); 7311 assert(addr_ptr != NULL, "unexpected"); 7312 7313 // Since the addresses are resolved at runtime, using switch instead of table - otherwise might get NULL addresses. 7314 if (bt == T_FLOAT) { 7315 switch(op) { 7316 case OP_EXP: { 7317 switch(bits) { 7318 case 64: *name_ptr = "vector_float64_exp"; *addr_ptr = StubRoutines::vector_float64_exp(); break; 7319 case 128: *name_ptr = "vector_float128_exp"; *addr_ptr = StubRoutines::vector_float128_exp(); break; 7320 case 256: *name_ptr = "vector_float256_exp"; *addr_ptr = StubRoutines::vector_float256_exp(); break; 7321 case 512: *name_ptr = "vector_float512_exp"; *addr_ptr = StubRoutines::vector_float512_exp(); break; 7322 default: Unimplemented(); break; 7323 } 7324 } 7325 break; 7326 case OP_LOG1P: { 7327 switch(bits) { 7328 case 64: *name_ptr = "vector_float64_log1p"; *addr_ptr = StubRoutines::vector_float64_log1p(); break; 7329 case 128: *name_ptr = "vector_float128_log1p"; *addr_ptr = StubRoutines::vector_float128_log1p(); break; 7330 case 256: *name_ptr = "vector_float256_log1p"; *addr_ptr = StubRoutines::vector_float256_log1p(); break; 7331 case 512: *name_ptr = "vector_float512_log1p"; *addr_ptr = StubRoutines::vector_float512_log1p(); break; 7332 default: Unimplemented(); break; 7333 } 7334 } 7335 break; 7336 case OP_LOG: { 7337 switch(bits) { 7338 case 64: *name_ptr = "vector_float64_log"; *addr_ptr = StubRoutines::vector_float64_log(); break; 7339 case 128: *name_ptr = "vector_float128_log"; *addr_ptr = StubRoutines::vector_float128_log(); break; 7340 case 256: *name_ptr = "vector_float256_log"; *addr_ptr = StubRoutines::vector_float256_log(); break; 7341 case 512: *name_ptr = "vector_float512_log"; *addr_ptr = StubRoutines::vector_float512_log(); break; 7342 default: Unimplemented(); break; 7343 } 7344 } 7345 break; 7346 case OP_LOG10: { 7347 switch(bits) { 7348 case 64: *name_ptr = "vector_float64_log10"; *addr_ptr = StubRoutines::vector_float64_log10(); break; 7349 case 128: *name_ptr = "vector_float128_log10"; *addr_ptr = StubRoutines::vector_float128_log10(); break; 7350 case 256: *name_ptr = "vector_float256_log10"; *addr_ptr = StubRoutines::vector_float256_log10(); break; 7351 case 512: *name_ptr = "vector_float512_log10"; *addr_ptr = StubRoutines::vector_float512_log10(); break; 7352 default: Unimplemented(); break; 7353 } 7354 } 7355 break; 7356 case OP_EXPM1: { 7357 switch(bits) { 7358 case 64: *name_ptr = "vector_float64_expm1"; *addr_ptr = StubRoutines::vector_float64_expm1(); break; 7359 case 128: *name_ptr = "vector_float128_expm1"; *addr_ptr = StubRoutines::vector_float128_expm1(); break; 7360 case 256: *name_ptr = "vector_float256_expm1"; *addr_ptr = StubRoutines::vector_float256_expm1(); break; 7361 case 512: *name_ptr = "vector_float512_expm1"; *addr_ptr = StubRoutines::vector_float512_expm1(); break; 7362 default: Unimplemented(); break; 7363 } 7364 } 7365 break; 7366 case OP_SIN: { 7367 switch(bits) { 7368 case 64: *name_ptr = "vector_float64_sin"; *addr_ptr = StubRoutines::vector_float64_sin(); break; 7369 case 128: *name_ptr = "vector_float128_sin"; *addr_ptr = StubRoutines::vector_float128_sin(); break; 7370 case 256: *name_ptr = "vector_float256_sin"; *addr_ptr = StubRoutines::vector_float256_sin(); break; 7371 case 512: *name_ptr = "vector_float512_sin"; *addr_ptr = StubRoutines::vector_float512_sin(); break; 7372 default: Unimplemented(); break; 7373 } 7374 } 7375 break; 7376 case OP_COS: { 7377 switch(bits) { 7378 case 64: *name_ptr = "vector_float64_cos"; *addr_ptr = StubRoutines::vector_float64_cos(); break; 7379 case 128: *name_ptr = "vector_float128_cos"; *addr_ptr = StubRoutines::vector_float128_cos(); break; 7380 case 256: *name_ptr = "vector_float256_cos"; *addr_ptr = StubRoutines::vector_float256_cos(); break; 7381 case 512: *name_ptr = "vector_float512_cos"; *addr_ptr = StubRoutines::vector_float512_cos(); break; 7382 default: Unimplemented(); break; 7383 } 7384 } 7385 break; 7386 case OP_TAN: { 7387 switch(bits) { 7388 case 64: *name_ptr = "vector_float64_tan"; *addr_ptr = StubRoutines::vector_float64_tan(); break; 7389 case 128: *name_ptr = "vector_float128_tan"; *addr_ptr = StubRoutines::vector_float128_tan(); break; 7390 case 256: *name_ptr = "vector_float256_tan"; *addr_ptr = StubRoutines::vector_float256_tan(); break; 7391 case 512: *name_ptr = "vector_float512_tan"; *addr_ptr = StubRoutines::vector_float512_tan(); break; 7392 default: Unimplemented(); break; 7393 } 7394 } 7395 break; 7396 case OP_SINH: { 7397 switch(bits) { 7398 case 64: *name_ptr = "vector_float64_sinh"; *addr_ptr = StubRoutines::vector_float64_sinh(); break; 7399 case 128: *name_ptr = "vector_float128_sinh"; *addr_ptr = StubRoutines::vector_float128_sinh(); break; 7400 case 256: *name_ptr = "vector_float256_sinh"; *addr_ptr = StubRoutines::vector_float256_sinh(); break; 7401 case 512: *name_ptr = "vector_float512_sinh"; *addr_ptr = StubRoutines::vector_float512_sinh(); break; 7402 default: Unimplemented(); break; 7403 } 7404 } 7405 break; 7406 case OP_COSH: { 7407 switch(bits) { 7408 case 64: *name_ptr = "vector_float64_cosh"; *addr_ptr = StubRoutines::vector_float64_cosh(); break; 7409 case 128: *name_ptr = "vector_float128_cosh"; *addr_ptr = StubRoutines::vector_float128_cosh(); break; 7410 case 256: *name_ptr = "vector_float256_cosh"; *addr_ptr = StubRoutines::vector_float256_cosh(); break; 7411 case 512: *name_ptr = "vector_float512_cosh"; *addr_ptr = StubRoutines::vector_float512_cosh(); break; 7412 default: Unimplemented(); break; 7413 } 7414 } 7415 break; 7416 case OP_TANH: { 7417 switch(bits) { 7418 case 64: *name_ptr = "vector_float64_tanh"; *addr_ptr = StubRoutines::vector_float64_tanh(); break; 7419 case 128: *name_ptr = "vector_float128_tanh"; *addr_ptr = StubRoutines::vector_float128_tanh(); break; 7420 case 256: *name_ptr = "vector_float256_tanh"; *addr_ptr = StubRoutines::vector_float256_tanh(); break; 7421 case 512: *name_ptr = "vector_float512_tanh"; *addr_ptr = StubRoutines::vector_float512_tanh(); break; 7422 default: Unimplemented(); break; 7423 } 7424 } 7425 break; 7426 case OP_ASIN: { 7427 switch(bits) { 7428 case 64: *name_ptr = "vector_float64_asin"; *addr_ptr = StubRoutines::vector_float64_asin(); break; 7429 case 128: *name_ptr = "vector_float128_asin"; *addr_ptr = StubRoutines::vector_float128_asin(); break; 7430 case 256: *name_ptr = "vector_float256_asin"; *addr_ptr = StubRoutines::vector_float256_asin(); break; 7431 case 512: *name_ptr = "vector_float512_asin"; *addr_ptr = StubRoutines::vector_float512_asin(); break; 7432 default: Unimplemented(); break; 7433 } 7434 } 7435 break; 7436 case OP_ACOS: { 7437 switch(bits) { 7438 case 64: *name_ptr = "vector_float64_acos"; *addr_ptr = StubRoutines::vector_float64_acos(); break; 7439 case 128: *name_ptr = "vector_float128_acos"; *addr_ptr = StubRoutines::vector_float128_acos(); break; 7440 case 256: *name_ptr = "vector_float256_acos"; *addr_ptr = StubRoutines::vector_float256_acos(); break; 7441 case 512: *name_ptr = "vector_float512_acos"; *addr_ptr = StubRoutines::vector_float512_acos(); break; 7442 default: Unimplemented(); break; 7443 } 7444 } 7445 break; 7446 case OP_ATAN: { 7447 switch(bits) { 7448 case 64: *name_ptr = "vector_float64_atan"; *addr_ptr = StubRoutines::vector_float64_atan(); break; 7449 case 128: *name_ptr = "vector_float128_atan"; *addr_ptr = StubRoutines::vector_float128_atan(); break; 7450 case 256: *name_ptr = "vector_float256_atan"; *addr_ptr = StubRoutines::vector_float256_atan(); break; 7451 case 512: *name_ptr = "vector_float512_atan"; *addr_ptr = StubRoutines::vector_float512_atan(); break; 7452 default: Unimplemented(); break; 7453 } 7454 } 7455 break; 7456 case OP_CBRT: { 7457 switch(bits) { 7458 case 64: *name_ptr = "vector_float64_cbrt"; *addr_ptr = StubRoutines::vector_float64_cbrt(); break; 7459 case 128: *name_ptr = "vector_float128_cbrt"; *addr_ptr = StubRoutines::vector_float128_cbrt(); break; 7460 case 256: *name_ptr = "vector_float256_cbrt"; *addr_ptr = StubRoutines::vector_float256_cbrt(); break; 7461 case 512: *name_ptr = "vector_float512_cbrt"; *addr_ptr = StubRoutines::vector_float512_cbrt(); break; 7462 default: Unimplemented(); break; 7463 } 7464 } 7465 break; 7466 case OP_HYPOT: { 7467 switch(bits) { 7468 case 64: *name_ptr = "vector_float64_hypot"; *addr_ptr = StubRoutines::vector_float64_hypot(); break; 7469 case 128: *name_ptr = "vector_float128_hypot"; *addr_ptr = StubRoutines::vector_float128_hypot(); break; 7470 case 256: *name_ptr = "vector_float256_hypot"; *addr_ptr = StubRoutines::vector_float256_hypot(); break; 7471 case 512: *name_ptr = "vector_float512_hypot"; *addr_ptr = StubRoutines::vector_float512_hypot(); break; 7472 default: Unimplemented(); break; 7473 } 7474 } 7475 break; 7476 case OP_POW: { 7477 switch(bits) { 7478 case 64: *name_ptr = "vector_float64_pow"; *addr_ptr = StubRoutines::vector_float64_pow(); break; 7479 case 128: *name_ptr = "vector_float128_pow"; *addr_ptr = StubRoutines::vector_float128_pow(); break; 7480 case 256: *name_ptr = "vector_float256_pow"; *addr_ptr = StubRoutines::vector_float256_pow(); break; 7481 case 512: *name_ptr = "vector_float512_pow"; *addr_ptr = StubRoutines::vector_float512_pow(); break; 7482 default: Unimplemented(); break; 7483 } 7484 } 7485 break; 7486 case OP_ATAN2: { 7487 switch(bits) { 7488 case 64: *name_ptr = "vector_float64_atan2"; *addr_ptr = StubRoutines::vector_float64_atan2(); break; 7489 case 128: *name_ptr = "vector_float128_atan2"; *addr_ptr = StubRoutines::vector_float128_atan2(); break; 7490 case 256: *name_ptr = "vector_float256_atan2"; *addr_ptr = StubRoutines::vector_float256_atan2(); break; 7491 case 512: *name_ptr = "vector_float512_atan2"; *addr_ptr = StubRoutines::vector_float512_atan2(); break; 7492 default: Unimplemented(); break; 7493 } 7494 } 7495 break; 7496 default: 7497 *name_ptr = "invalid"; 7498 *addr_ptr = NULL; 7499 break; 7500 } 7501 } else { 7502 assert(bt == T_DOUBLE, "must be FP type only"); 7503 switch(op) { 7504 case OP_EXP: { 7505 switch(bits) { 7506 case 64: *name_ptr = "vector_double64_exp"; *addr_ptr = StubRoutines::vector_double64_exp(); break; 7507 case 128: *name_ptr = "vector_double128_exp"; *addr_ptr = StubRoutines::vector_double128_exp(); break; 7508 case 256: *name_ptr = "vector_double256_exp"; *addr_ptr = StubRoutines::vector_double256_exp(); break; 7509 case 512: *name_ptr = "vector_double512_exp"; *addr_ptr = StubRoutines::vector_double512_exp(); break; 7510 default: Unimplemented(); break; 7511 } 7512 } 7513 break; 7514 case OP_LOG1P: { 7515 switch(bits) { 7516 case 64: *name_ptr = "vector_double64_log1p"; *addr_ptr = StubRoutines::vector_double64_log1p(); break; 7517 case 128: *name_ptr = "vector_double128_log1p"; *addr_ptr = StubRoutines::vector_double128_log1p(); break; 7518 case 256: *name_ptr = "vector_double256_log1p"; *addr_ptr = StubRoutines::vector_double256_log1p(); break; 7519 case 512: *name_ptr = "vector_double512_log1p"; *addr_ptr = StubRoutines::vector_double512_log1p(); break; 7520 default: Unimplemented(); break; 7521 } 7522 } 7523 break; 7524 case OP_LOG: { 7525 switch(bits) { 7526 case 64: *name_ptr = "vector_double64_log"; *addr_ptr = StubRoutines::vector_double64_log(); break; 7527 case 128: *name_ptr = "vector_double128_log"; *addr_ptr = StubRoutines::vector_double128_log(); break; 7528 case 256: *name_ptr = "vector_double256_log"; *addr_ptr = StubRoutines::vector_double256_log(); break; 7529 case 512: *name_ptr = "vector_double512_log"; *addr_ptr = StubRoutines::vector_double512_log(); break; 7530 default: Unimplemented(); break; 7531 } 7532 } 7533 break; 7534 case OP_LOG10: { 7535 switch(bits) { 7536 case 64: *name_ptr = "vector_double64_log10"; *addr_ptr = StubRoutines::vector_double64_log10(); break; 7537 case 128: *name_ptr = "vector_double128_log10"; *addr_ptr = StubRoutines::vector_double128_log10(); break; 7538 case 256: *name_ptr = "vector_double256_log10"; *addr_ptr = StubRoutines::vector_double256_log10(); break; 7539 case 512: *name_ptr = "vector_double512_log10"; *addr_ptr = StubRoutines::vector_double512_log10(); break; 7540 default: Unimplemented(); break; 7541 } 7542 } 7543 break; 7544 case OP_EXPM1: { 7545 switch(bits) { 7546 case 64: *name_ptr = "vector_double64_expm1"; *addr_ptr = StubRoutines::vector_double64_expm1(); break; 7547 case 128: *name_ptr = "vector_double128_expm1"; *addr_ptr = StubRoutines::vector_double128_expm1(); break; 7548 case 256: *name_ptr = "vector_double256_expm1"; *addr_ptr = StubRoutines::vector_double256_expm1(); break; 7549 case 512: *name_ptr = "vector_double512_expm1"; *addr_ptr = StubRoutines::vector_double512_expm1(); break; 7550 default: Unimplemented(); break; 7551 } 7552 } 7553 break; 7554 case OP_SIN: { 7555 switch(bits) { 7556 case 64: *name_ptr = "vector_double64_sin"; *addr_ptr = StubRoutines::vector_double64_sin(); break; 7557 case 128: *name_ptr = "vector_double128_sin"; *addr_ptr = StubRoutines::vector_double128_sin(); break; 7558 case 256: *name_ptr = "vector_double256_sin"; *addr_ptr = StubRoutines::vector_double256_sin(); break; 7559 case 512: *name_ptr = "vector_double512_sin"; *addr_ptr = StubRoutines::vector_double512_sin(); break; 7560 default: Unimplemented(); break; 7561 } 7562 } 7563 break; 7564 case OP_COS: { 7565 switch(bits) { 7566 case 64: *name_ptr = "vector_double64_cos"; *addr_ptr = StubRoutines::vector_double64_cos(); break; 7567 case 128: *name_ptr = "vector_double128_cos"; *addr_ptr = StubRoutines::vector_double128_cos(); break; 7568 case 256: *name_ptr = "vector_double256_cos"; *addr_ptr = StubRoutines::vector_double256_cos(); break; 7569 case 512: *name_ptr = "vector_double512_cos"; *addr_ptr = StubRoutines::vector_double512_cos(); break; 7570 default: Unimplemented(); break; 7571 } 7572 } 7573 break; 7574 case OP_TAN: { 7575 switch(bits) { 7576 case 64: *name_ptr = "vector_double64_tan"; *addr_ptr = StubRoutines::vector_double64_tan(); break; 7577 case 128: *name_ptr = "vector_double128_tan"; *addr_ptr = StubRoutines::vector_double128_tan(); break; 7578 case 256: *name_ptr = "vector_double256_tan"; *addr_ptr = StubRoutines::vector_double256_tan(); break; 7579 case 512: *name_ptr = "vector_double512_tan"; *addr_ptr = StubRoutines::vector_double512_tan(); break; 7580 default: Unimplemented(); break; 7581 } 7582 } 7583 break; 7584 case OP_SINH: { 7585 switch(bits) { 7586 case 64: *name_ptr = "vector_double64_sinh"; *addr_ptr = StubRoutines::vector_double64_sinh(); break; 7587 case 128: *name_ptr = "vector_double128_sinh"; *addr_ptr = StubRoutines::vector_double128_sinh(); break; 7588 case 256: *name_ptr = "vector_double256_sinh"; *addr_ptr = StubRoutines::vector_double256_sinh(); break; 7589 case 512: *name_ptr = "vector_double512_sinh"; *addr_ptr = StubRoutines::vector_double512_sinh(); break; 7590 default: Unimplemented(); break; 7591 } 7592 } 7593 break; 7594 case OP_COSH: { 7595 switch(bits) { 7596 case 64: *name_ptr = "vector_double64_cosh"; *addr_ptr = StubRoutines::vector_double64_cosh(); break; 7597 case 128: *name_ptr = "vector_double128_cosh"; *addr_ptr = StubRoutines::vector_double128_cosh(); break; 7598 case 256: *name_ptr = "vector_double256_cosh"; *addr_ptr = StubRoutines::vector_double256_cosh(); break; 7599 case 512: *name_ptr = "vector_double512_cosh"; *addr_ptr = StubRoutines::vector_double512_cosh(); break; 7600 default: Unimplemented(); break; 7601 } 7602 } 7603 break; 7604 case OP_TANH: { 7605 switch(bits) { 7606 case 64: *name_ptr = "vector_double64_tanh"; *addr_ptr = StubRoutines::vector_double64_tanh(); break; 7607 case 128: *name_ptr = "vector_double128_tanh"; *addr_ptr = StubRoutines::vector_double128_tanh(); break; 7608 case 256: *name_ptr = "vector_double256_tanh"; *addr_ptr = StubRoutines::vector_double256_tanh(); break; 7609 case 512: *name_ptr = "vector_double512_tanh"; *addr_ptr = StubRoutines::vector_double512_tanh(); break; 7610 default: Unimplemented(); break; 7611 } 7612 } 7613 break; 7614 case OP_ASIN: { 7615 switch(bits) { 7616 case 64: *name_ptr = "vector_double64_asin"; *addr_ptr = StubRoutines::vector_double64_asin(); break; 7617 case 128: *name_ptr = "vector_double128_asin"; *addr_ptr = StubRoutines::vector_double128_asin(); break; 7618 case 256: *name_ptr = "vector_double256_asin"; *addr_ptr = StubRoutines::vector_double256_asin(); break; 7619 case 512: *name_ptr = "vector_double512_asin"; *addr_ptr = StubRoutines::vector_double512_asin(); break; 7620 default: Unimplemented(); break; 7621 } 7622 } 7623 break; 7624 case OP_ACOS: { 7625 switch(bits) { 7626 case 64: *name_ptr = "vector_double64_acos"; *addr_ptr = StubRoutines::vector_double64_acos(); break; 7627 case 128: *name_ptr = "vector_double128_acos"; *addr_ptr = StubRoutines::vector_double128_acos(); break; 7628 case 256: *name_ptr = "vector_double256_acos"; *addr_ptr = StubRoutines::vector_double256_acos(); break; 7629 case 512: *name_ptr = "vector_double512_acos"; *addr_ptr = StubRoutines::vector_double512_acos(); break; 7630 default: Unimplemented(); break; 7631 } 7632 } 7633 break; 7634 case OP_ATAN: { 7635 switch(bits) { 7636 case 64: *name_ptr = "vector_double64_atan"; *addr_ptr = StubRoutines::vector_double64_atan(); break; 7637 case 128: *name_ptr = "vector_double128_atan"; *addr_ptr = StubRoutines::vector_double128_atan(); break; 7638 case 256: *name_ptr = "vector_double256_atan"; *addr_ptr = StubRoutines::vector_double256_atan(); break; 7639 case 512: *name_ptr = "vector_double512_atan"; *addr_ptr = StubRoutines::vector_double512_atan(); break; 7640 default: Unimplemented(); break; 7641 } 7642 } 7643 break; 7644 case OP_CBRT: { 7645 switch(bits) { 7646 case 64: *name_ptr = "vector_double64_cbrt"; *addr_ptr = StubRoutines::vector_double64_cbrt(); break; 7647 case 128: *name_ptr = "vector_double128_cbrt"; *addr_ptr = StubRoutines::vector_double128_cbrt(); break; 7648 case 256: *name_ptr = "vector_double256_cbrt"; *addr_ptr = StubRoutines::vector_double256_cbrt(); break; 7649 case 512: *name_ptr = "vector_double512_cbrt"; *addr_ptr = StubRoutines::vector_double512_cbrt(); break; 7650 default: Unimplemented(); break; 7651 } 7652 } 7653 break; 7654 case OP_HYPOT: { 7655 switch(bits) { 7656 case 64: *name_ptr = "vector_double64_hypot"; *addr_ptr = StubRoutines::vector_double64_hypot(); break; 7657 case 128: *name_ptr = "vector_double128_hypot"; *addr_ptr = StubRoutines::vector_double128_hypot(); break; 7658 case 256: *name_ptr = "vector_double256_hypot"; *addr_ptr = StubRoutines::vector_double256_hypot(); break; 7659 case 512: *name_ptr = "vector_double512_hypot"; *addr_ptr = StubRoutines::vector_double512_hypot(); break; 7660 default: Unimplemented(); break; 7661 } 7662 } 7663 break; 7664 case OP_POW: { 7665 switch(bits) { 7666 case 64: *name_ptr = "vector_double64_pow"; *addr_ptr = StubRoutines::vector_double64_pow(); break; 7667 case 128: *name_ptr = "vector_double128_pow"; *addr_ptr = StubRoutines::vector_double128_pow(); break; 7668 case 256: *name_ptr = "vector_double256_pow"; *addr_ptr = StubRoutines::vector_double256_pow(); break; 7669 case 512: *name_ptr = "vector_double512_pow"; *addr_ptr = StubRoutines::vector_double512_pow(); break; 7670 default: Unimplemented(); break; 7671 } 7672 } 7673 break; 7674 case OP_ATAN2: { 7675 switch(bits) { 7676 case 64: *name_ptr = "vector_double64_atan2"; *addr_ptr = StubRoutines::vector_double64_atan2(); break; 7677 case 128: *name_ptr = "vector_double128_atan2"; *addr_ptr = StubRoutines::vector_double128_atan2(); break; 7678 case 256: *name_ptr = "vector_double256_atan2"; *addr_ptr = StubRoutines::vector_double256_atan2(); break; 7679 case 512: *name_ptr = "vector_double512_atan2"; *addr_ptr = StubRoutines::vector_double512_atan2(); break; 7680 default: Unimplemented(); break; 7681 } 7682 } 7683 break; 7684 7685 default: 7686 *name_ptr = "invalid"; 7687 *addr_ptr = NULL; 7688 break; 7689 } 7690 } 7691 } 7692 7693 Node* LibraryCallKit::gen_call_to_svml(int vector_api_op_id, BasicType bt, int num_elem, Node* opd1, Node* opd2) { 7694 assert(vector_api_op_id >= OP_SVML_START && vector_api_op_id <= OP_SVML_END, "need valid op id"); 7695 assert(opd1 != NULL, "must not be null"); 7696 const TypeVect* vt = TypeVect::make(bt, num_elem); 7697 const TypeFunc* call_type = OptoRuntime::Math_Vector_Vector_Type(opd2 != NULL ? 2 : 1, vt, vt); 7698 const char* name = NULL; 7699 address addr = NULL; 7700 7701 // Get address for svml method. 7702 get_svml_address(vector_api_op_id, vt->length_in_bytes() * BitsPerByte, bt, &name, &addr); 7703 7704 if (addr == NULL) { 7705 return NULL; 7706 } 7707 7708 assert(name != NULL, "name must not be null"); 7709 Node* operation = make_runtime_call(RC_VECTOR, 7710 call_type, 7711 addr, 7712 name, 7713 TypePtr::BOTTOM, 7714 opd1, 7715 opd2); 7716 return _gvn.transform(new ProjNode(_gvn.transform(operation), TypeFunc::Parms)); 7717 } 7718 7719 // static 7720 // <V extends Vector<?,?>> 7721 // V broadcastInt(int opr, Class<V> vectorClass, Class<?> elementType, int vlen, 7722 // V v, int i, 7723 // VectorBroadcastIntOp<V> defaultImpl) { 7724 // 7725 bool LibraryCallKit::inline_vector_broadcast_int() { 7726 const TypeInt* opr = gvn().type(argument(0))->is_int(); 7727 const TypeInstPtr* vector_klass = gvn().type(argument(1))->is_instptr(); 7728 const TypeInstPtr* elem_klass = gvn().type(argument(2))->is_instptr(); 7729 const TypeInt* vlen = gvn().type(argument(3))->is_int(); 7730 7731 if (!opr->is_con() || vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con()) { 7732 return false; // not enough info for intrinsification 7733 } 7734 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 7735 if (!elem_type->is_primitive_type()) { 7736 return false; // should be primitive type 7737 } 7738 BasicType elem_bt = elem_type->basic_type(); 7739 int num_elem = vlen->get_con(); 7740 int opc = get_opc(opr->get_con(), elem_bt); 7741 int sopc = VectorNode::opcode(opc, elem_bt); // get_node_id(opr->get_con(), elem_bt); 7742 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7743 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 7744 7745 if (!arch_supports_vector(sopc, num_elem, elem_bt, VecMaskNotUsed)) { 7746 return false; // not supported 7747 } 7748 Node* opd1 = unbox_vector(argument(4), vbox_type, elem_bt, num_elem); 7749 Node* opd2 = shift_count(argument(5), opc, elem_bt, num_elem); 7750 if (opd1 == NULL || opd2 == NULL) { 7751 return false; 7752 } 7753 Node* operation = _gvn.transform(VectorNode::make(opc, opd1, opd2, num_elem, elem_bt)); 7754 7755 Node* vbox = box_vector(operation, vbox_type, elem_bt, num_elem); 7756 set_vector_result(vbox); 7757 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt)))); 7758 return true; 7759 } 7760 7761 bool LibraryCallKit::inline_vector_cast_reinterpret(bool is_cast) { 7762 const TypeInstPtr* vector_klass_from = gvn().type(argument(0))->is_instptr(); 7763 const TypeInstPtr* elem_klass_from = gvn().type(argument(1))->is_instptr(); 7764 const TypeInt* vlen_from = gvn().type(argument(2))->is_int(); 7765 7766 const TypeInstPtr* elem_klass_to = gvn().type(argument(3))->is_instptr(); 7767 const TypeInt* vlen_to = gvn().type(argument(4))->is_int(); 7768 7769 if (vector_klass_from->const_oop() == NULL || elem_klass_from->const_oop() == NULL || !vlen_from->is_con() || 7770 elem_klass_to->const_oop() == NULL || !vlen_to->is_con()) { 7771 return false; // not enough info for intrinsification 7772 } 7773 7774 ciKlass* vbox_klass_from = vector_klass_from->const_oop()->as_instance()->java_lang_Class_klass(); 7775 if (!vbox_klass_from->is_vectorapi_vector()) { 7776 return false; // only vector & mask are supported 7777 } 7778 bool is_mask = vbox_klass_from->is_vectormask(); 7779 7780 ciType* elem_type_from = elem_klass_from->const_oop()->as_instance()->java_mirror_type(); 7781 if (!elem_type_from->is_primitive_type()) { 7782 return false; // should be primitive type 7783 } 7784 BasicType elem_bt_from = elem_type_from->basic_type(); 7785 ciType* elem_type_to = elem_klass_to->const_oop()->as_instance()->java_mirror_type(); 7786 if (!elem_type_to->is_primitive_type()) { 7787 return false; // should be primitive type 7788 } 7789 BasicType elem_bt_to = elem_type_to->basic_type(); 7790 if (is_mask && elem_bt_from != elem_bt_to) { 7791 return false; // type mismatch 7792 } 7793 int num_elem_from = vlen_from->get_con(); 7794 int num_elem_to = vlen_to->get_con(); 7795 7796 // Check whether we can unbox to appropriate size. Even with casting, checking for reinterpret is needed 7797 // since we may need to change size. 7798 if (!arch_supports_vector(Op_VectorReinterpret, 7799 num_elem_from, 7800 elem_bt_from, 7801 is_mask ? VecMaskUseAll : VecMaskNotUsed)) { 7802 return false; 7803 } 7804 7805 // Check whether we can support resizing/reinterpreting to the new size. 7806 if (!arch_supports_vector(Op_VectorReinterpret, 7807 num_elem_to, 7808 elem_bt_to, 7809 is_mask ? VecMaskUseAll : VecMaskNotUsed)) { 7810 return false; 7811 } 7812 7813 // At this point, we know that both input and output vector registers are supported 7814 // by the architecture. Next check if the casted type is simply to same type - which means 7815 // that it is actually a resize and not a cast. 7816 if (is_cast && elem_bt_from == elem_bt_to) { 7817 is_cast = false; 7818 } 7819 7820 const TypeInstPtr* vbox_type_from = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_from); 7821 7822 Node* opd1 = unbox_vector(argument(5), vbox_type_from, elem_bt_from, num_elem_from); 7823 if (opd1 == NULL) { 7824 return false; 7825 } 7826 7827 const TypeVect* src_type = TypeVect::make(elem_bt_from, num_elem_from); 7828 const TypeVect* dst_type = TypeVect::make(elem_bt_to, num_elem_to); 7829 7830 Node* op = opd1; 7831 if (is_cast) { 7832 assert(!is_mask, "masks cannot be casted"); 7833 int cast_vopc = VectorCastNode::opcode(elem_bt_from); 7834 // Make sure that cast is implemented to particular type/size combination. 7835 if (!arch_supports_vector(cast_vopc, num_elem_to, elem_bt_to, VecMaskNotUsed)) { 7836 return false; 7837 } 7838 7839 if (num_elem_from < num_elem_to) { 7840 // Since input and output number of elements are not consistent, we need to make sure we 7841 // properly size. Thus, first make a cast that retains the number of elements from source. 7842 // In case the size exceeds the arch size, we do the minimum. 7843 int num_elem_for_cast = MIN2(num_elem_from, Matcher::max_vector_size(elem_bt_to)); 7844 7845 // It is possible that arch does not support this intermediate vector size 7846 // TODO More complex logic required here to handle this corner case for the sizes. 7847 if (!arch_supports_vector(cast_vopc, num_elem_for_cast, elem_bt_to, VecMaskNotUsed)) { 7848 return false; 7849 } 7850 7851 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_for_cast)); 7852 // Now ensure that the destination gets properly resized to needed size. 7853 op = _gvn.transform(new VectorReinterpretNode(op, op->bottom_type()->is_vect(), dst_type)); 7854 } else if (num_elem_from > num_elem_to) { 7855 // Since number elements from input is larger than output, simply reduce size of input (we are supposed to 7856 // drop top elements anyway). 7857 int num_elem_for_resize = MAX2(num_elem_to, Matcher::min_vector_size(elem_bt_to)); 7858 7859 // It is possible that arch does not support this intermediate vector size 7860 // TODO More complex logic required here to handle this corner case for the sizes. 7861 if (!arch_supports_vector(Op_VectorReinterpret, 7862 num_elem_for_resize, 7863 elem_bt_from, 7864 VecMaskNotUsed)) { 7865 return false; 7866 } 7867 7868 op = _gvn.transform(new VectorReinterpretNode(op, 7869 src_type, 7870 TypeVect::make(elem_bt_from, 7871 num_elem_for_resize))); 7872 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to)); 7873 } else { 7874 // Since input and output number of elements match, and since we know this vector size is 7875 // supported, simply do a cast with no resize needed. 7876 op = _gvn.transform(VectorCastNode::make(cast_vopc, op, elem_bt_to, num_elem_to)); 7877 } 7878 } else if (Type::cmp(src_type, dst_type) != 0) { 7879 assert(!is_cast, "must be reinterpret"); 7880 op = _gvn.transform(new VectorReinterpretNode(op, src_type, dst_type)); 7881 } 7882 7883 ciKlass* vbox_klass_to = get_exact_klass_for_vector_box(vbox_klass_from, elem_type_to->basic_type(), 7884 num_elem_to, is_mask ? VECAPI_MASK : VECAPI_VECTOR); 7885 const TypeInstPtr* vbox_type_to = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass_to); 7886 Node* vbox = box_vector(op, vbox_type_to, elem_bt_to, num_elem_to); 7887 set_vector_result(vbox); 7888 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem_to * type2aelembytes(elem_bt_to)))); 7889 return true; 7890 } 7891 7892 bool LibraryCallKit::inline_vector_insert() { 7893 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr(); 7894 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr(); 7895 const TypeInt* vlen = gvn().type(argument(2))->is_int(); 7896 const TypeInt* idx = gvn().type(argument(4))->is_int(); 7897 7898 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con() || !idx->is_con()) { 7899 return false; // not enough info for intrinsification 7900 } 7901 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 7902 if (!elem_type->is_primitive_type()) { 7903 return false; // should be primitive type 7904 } 7905 BasicType elem_bt = elem_type->basic_type(); 7906 int num_elem = vlen->get_con(); 7907 if (!arch_supports_vector(Op_VectorInsert, num_elem, elem_bt, VecMaskNotUsed)) { 7908 return false; // not supported 7909 } 7910 7911 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7912 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 7913 7914 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem); 7915 if (opd == NULL) { 7916 return false; 7917 } 7918 7919 Node* insert_val = argument(5); 7920 assert(gvn().type(insert_val)->isa_long() != NULL, "expected to be long"); 7921 7922 // Convert insert value back to its appropriate type. 7923 switch (elem_bt) { 7924 case T_BYTE: 7925 insert_val = gvn().transform(new ConvL2INode(insert_val)); 7926 insert_val = gvn().transform(new CastIINode(insert_val, TypeInt::BYTE)); 7927 break; 7928 case T_SHORT: 7929 insert_val = gvn().transform(new ConvL2INode(insert_val)); 7930 insert_val = gvn().transform(new CastIINode(insert_val, TypeInt::SHORT)); 7931 break; 7932 case T_INT: 7933 insert_val = gvn().transform(new ConvL2INode(insert_val)); 7934 break; 7935 case T_FLOAT: 7936 insert_val = gvn().transform(new ConvL2INode(insert_val)); 7937 insert_val = gvn().transform(new MoveI2FNode(insert_val)); 7938 break; 7939 case T_DOUBLE: 7940 insert_val = gvn().transform(new MoveL2DNode(insert_val)); 7941 break; 7942 case T_LONG: 7943 // no conversion needed 7944 break; 7945 default: fatal("%s", type2name(elem_bt)); break; 7946 } 7947 7948 Node* operation = VectorInsertNode::make(opd, insert_val, idx->get_con()); 7949 operation = box_vector(operation, vbox_type, elem_bt, num_elem); 7950 set_vector_result(operation); 7951 7952 C->set_max_vector_size(MAX2(C->max_vector_size(), (uint)(num_elem * type2aelembytes(elem_bt)))); 7953 return true; 7954 } 7955 7956 bool LibraryCallKit::inline_vector_extract() { 7957 const TypeInstPtr* vector_klass = gvn().type(argument(0))->is_instptr(); 7958 const TypeInstPtr* elem_klass = gvn().type(argument(1))->is_instptr(); 7959 const TypeInt* vlen = gvn().type(argument(2))->is_int(); 7960 const TypeInt* idx = gvn().type(argument(4))->is_int(); 7961 7962 if (vector_klass->const_oop() == NULL || elem_klass->const_oop() == NULL || !vlen->is_con() || !idx->is_con()) { 7963 return false; // not enough info for intrinsification 7964 } 7965 ciType* elem_type = elem_klass->const_oop()->as_instance()->java_mirror_type(); 7966 if (!elem_type->is_primitive_type()) { 7967 return false; // should be primitive type 7968 } 7969 BasicType elem_bt = elem_type->basic_type(); 7970 int num_elem = vlen->get_con(); 7971 int vopc = ExtractNode::opcode(elem_bt); 7972 if (!arch_supports_vector(vopc, num_elem, elem_bt, VecMaskNotUsed)) { 7973 return false; // not supported 7974 } 7975 7976 ciKlass* vbox_klass = vector_klass->const_oop()->as_instance()->java_lang_Class_klass(); 7977 const TypeInstPtr* vbox_type = TypeInstPtr::make_exact(TypePtr::NotNull, vbox_klass); 7978 7979 Node* opd = unbox_vector(argument(3), vbox_type, elem_bt, num_elem); 7980 if (opd == NULL) { 7981 return false; 7982 } 7983 7984 Node* operation = gvn().transform(ExtractNode::make(opd, idx->get_con(), elem_bt)); 7985 7986 Node* bits = NULL; 7987 switch (elem_bt) { 7988 case T_BYTE: 7989 case T_SHORT: 7990 case T_INT: { 7991 bits = gvn().transform(new ConvI2LNode(operation)); 7992 break; 7993 } 7994 case T_FLOAT: { 7995 bits = gvn().transform(new MoveF2INode(operation)); 7996 bits = gvn().transform(new ConvI2LNode(bits)); 7997 break; 7998 } 7999 case T_DOUBLE: { 8000 bits = gvn().transform(new MoveD2LNode(operation)); 8001 break; 8002 } 8003 case T_LONG: { 8004 bits = operation; // no conversion needed 8005 break; 8006 } 8007 default: fatal("%s", type2name(elem_bt)); 8008 } 8009 8010 set_vector_result(bits); 8011 return true; 8012 } 8013 8014 //------------------------------get_state_from_sha_object----------------------- 8015 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 8016 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 8017 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 8018 if (sha_state == NULL) return (Node *) NULL; 8019 8020 // now have the array, need to get the start address of the state array 8021 Node* state = array_element_address(sha_state, intcon(0), T_INT); 8022 return state; 8023 } 8024 8025 //------------------------------get_state_from_sha5_object----------------------- 8026 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 8027 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 8028 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 8029 if (sha_state == NULL) return (Node *) NULL; 8030 8031 // now have the array, need to get the start address of the state array 8032 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 8033 return state; 8034 } 8035 8036 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 8037 // Return node representing slow path of predicate check. 8038 // the pseudo code we want to emulate with this predicate is: 8039 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 8040 // 8041 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 8042 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 8043 "need SHA1/SHA256/SHA512 instruction support"); 8044 assert((uint)predicate < 3, "sanity"); 8045 8046 // The receiver was checked for NULL already. 8047 Node* digestBaseObj = argument(0); 8048 8049 // get DigestBase klass for instanceOf check 8050 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 8051 assert(tinst != NULL, "digestBaseObj is null"); 8052 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 8053 8054 const char* klass_SHA_name = NULL; 8055 switch (predicate) { 8056 case 0: 8057 if (UseSHA1Intrinsics) { 8058 // we want to do an instanceof comparison against the SHA class 8059 klass_SHA_name = "sun/security/provider/SHA"; 8060 } 8061 break; 8062 case 1: 8063 if (UseSHA256Intrinsics) { 8064 // we want to do an instanceof comparison against the SHA2 class 8065 klass_SHA_name = "sun/security/provider/SHA2"; 8066 } 8067 break; 8068 case 2: 8069 if (UseSHA512Intrinsics) { 8070 // we want to do an instanceof comparison against the SHA5 class 8071 klass_SHA_name = "sun/security/provider/SHA5"; 8072 } 8073 break; 8074 default: 8075 fatal("unknown SHA intrinsic predicate: %d", predicate); 8076 } 8077 8078 ciKlass* klass_SHA = NULL; 8079 if (klass_SHA_name != NULL) { 8080 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 8081 } 8082 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 8083 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 8084 Node* ctrl = control(); 8085 set_control(top()); // no intrinsic path 8086 return ctrl; 8087 } 8088 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 8089 8090 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 8091 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 8092 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 8093 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 8094 8095 return instof_false; // even if it is NULL 8096 } 8097 8098 //-------------inline_fma----------------------------------- 8099 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) { 8100 Node *a = NULL; 8101 Node *b = NULL; 8102 Node *c = NULL; 8103 Node* result = NULL; 8104 switch (id) { 8105 case vmIntrinsics::_fmaD: 8106 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each."); 8107 // no receiver since it is static method 8108 a = round_double_node(argument(0)); 8109 b = round_double_node(argument(2)); 8110 c = round_double_node(argument(4)); 8111 result = _gvn.transform(new FmaDNode(control(), a, b, c)); 8112 break; 8113 case vmIntrinsics::_fmaF: 8114 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each."); 8115 a = argument(0); 8116 b = argument(1); 8117 c = argument(2); 8118 result = _gvn.transform(new FmaFNode(control(), a, b, c)); 8119 break; 8120 default: 8121 fatal_unexpected_iid(id); break; 8122 } 8123 set_result(result); 8124 return true; 8125 } 8126 8127 bool LibraryCallKit::inline_profileBoolean() { 8128 Node* counts = argument(1); 8129 const TypeAryPtr* ary = NULL; 8130 ciArray* aobj = NULL; 8131 if (counts->is_Con() 8132 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 8133 && (aobj = ary->const_oop()->as_array()) != NULL 8134 && (aobj->length() == 2)) { 8135 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 8136 jint false_cnt = aobj->element_value(0).as_int(); 8137 jint true_cnt = aobj->element_value(1).as_int(); 8138 8139 if (C->log() != NULL) { 8140 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 8141 false_cnt, true_cnt); 8142 } 8143 8144 if (false_cnt + true_cnt == 0) { 8145 // According to profile, never executed. 8146 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 8147 Deoptimization::Action_reinterpret); 8148 return true; 8149 } 8150 8151 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 8152 // is a number of each value occurrences. 8153 Node* result = argument(0); 8154 if (false_cnt == 0 || true_cnt == 0) { 8155 // According to profile, one value has been never seen. 8156 int expected_val = (false_cnt == 0) ? 1 : 0; 8157 8158 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 8159 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 8160 8161 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 8162 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 8163 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 8164 8165 { // Slow path: uncommon trap for never seen value and then reexecute 8166 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 8167 // the value has been seen at least once. 8168 PreserveJVMState pjvms(this); 8169 PreserveReexecuteState preexecs(this); 8170 jvms()->set_should_reexecute(true); 8171 8172 set_control(slow_path); 8173 set_i_o(i_o()); 8174 8175 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 8176 Deoptimization::Action_reinterpret); 8177 } 8178 // The guard for never seen value enables sharpening of the result and 8179 // returning a constant. It allows to eliminate branches on the same value 8180 // later on. 8181 set_control(fast_path); 8182 result = intcon(expected_val); 8183 } 8184 // Stop profiling. 8185 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 8186 // By replacing method body with profile data (represented as ProfileBooleanNode 8187 // on IR level) we effectively disable profiling. 8188 // It enables full speed execution once optimized code is generated. 8189 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 8190 C->record_for_igvn(profile); 8191 set_result(profile); 8192 return true; 8193 } else { 8194 // Continue profiling. 8195 // Profile data isn't available at the moment. So, execute method's bytecode version. 8196 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 8197 // is compiled and counters aren't available since corresponding MethodHandle 8198 // isn't a compile-time constant. 8199 return false; 8200 } 8201 } 8202 8203 bool LibraryCallKit::inline_isCompileConstant() { 8204 Node* n = argument(0); 8205 set_result(n->is_Con() ? intcon(1) : intcon(0)); 8206 return true; 8207 }