library_call.cpp revision 11693:b82f3570b774
1/* 2 * Copyright (c) 1999, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25#include "precompiled.hpp" 26#include "asm/macroAssembler.hpp" 27#include "classfile/systemDictionary.hpp" 28#include "classfile/vmSymbols.hpp" 29#include "compiler/compileBroker.hpp" 30#include "compiler/compileLog.hpp" 31#include "memory/resourceArea.hpp" 32#include "oops/objArrayKlass.hpp" 33#include "opto/addnode.hpp" 34#include "opto/arraycopynode.hpp" 35#include "opto/c2compiler.hpp" 36#include "opto/callGenerator.hpp" 37#include "opto/castnode.hpp" 38#include "opto/cfgnode.hpp" 39#include "opto/convertnode.hpp" 40#include "opto/countbitsnode.hpp" 41#include "opto/intrinsicnode.hpp" 42#include "opto/idealKit.hpp" 43#include "opto/mathexactnode.hpp" 44#include "opto/movenode.hpp" 45#include "opto/mulnode.hpp" 46#include "opto/narrowptrnode.hpp" 47#include "opto/opaquenode.hpp" 48#include "opto/parse.hpp" 49#include "opto/runtime.hpp" 50#include "opto/subnode.hpp" 51#include "prims/nativeLookup.hpp" 52#include "prims/unsafe.hpp" 53#include "runtime/sharedRuntime.hpp" 54#ifdef TRACE_HAVE_INTRINSICS 55#include "trace/traceMacros.hpp" 56#endif 57 58class LibraryIntrinsic : public InlineCallGenerator { 59 // Extend the set of intrinsics known to the runtime: 60 public: 61 private: 62 bool _is_virtual; 63 bool _does_virtual_dispatch; 64 int8_t _predicates_count; // Intrinsic is predicated by several conditions 65 int8_t _last_predicate; // Last generated predicate 66 vmIntrinsics::ID _intrinsic_id; 67 68 public: 69 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 70 : InlineCallGenerator(m), 71 _is_virtual(is_virtual), 72 _does_virtual_dispatch(does_virtual_dispatch), 73 _predicates_count((int8_t)predicates_count), 74 _last_predicate((int8_t)-1), 75 _intrinsic_id(id) 76 { 77 } 78 virtual bool is_intrinsic() const { return true; } 79 virtual bool is_virtual() const { return _is_virtual; } 80 virtual bool is_predicated() const { return _predicates_count > 0; } 81 virtual int predicates_count() const { return _predicates_count; } 82 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 83 virtual JVMState* generate(JVMState* jvms); 84 virtual Node* generate_predicate(JVMState* jvms, int predicate); 85 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 86}; 87 88 89// Local helper class for LibraryIntrinsic: 90class LibraryCallKit : public GraphKit { 91 private: 92 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 93 Node* _result; // the result node, if any 94 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 95 96 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type); 97 98 public: 99 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 100 : GraphKit(jvms), 101 _intrinsic(intrinsic), 102 _result(NULL) 103 { 104 // Check if this is a root compile. In that case we don't have a caller. 105 if (!jvms->has_method()) { 106 _reexecute_sp = sp(); 107 } else { 108 // Find out how many arguments the interpreter needs when deoptimizing 109 // and save the stack pointer value so it can used by uncommon_trap. 110 // We find the argument count by looking at the declared signature. 111 bool ignored_will_link; 112 ciSignature* declared_signature = NULL; 113 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 114 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 115 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 116 } 117 } 118 119 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 120 121 ciMethod* caller() const { return jvms()->method(); } 122 int bci() const { return jvms()->bci(); } 123 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 124 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 125 ciMethod* callee() const { return _intrinsic->method(); } 126 127 bool try_to_inline(int predicate); 128 Node* try_to_predicate(int predicate); 129 130 void push_result() { 131 // Push the result onto the stack. 132 if (!stopped() && result() != NULL) { 133 BasicType bt = result()->bottom_type()->basic_type(); 134 push_node(bt, result()); 135 } 136 } 137 138 private: 139 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 140 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)); 141 } 142 143 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 144 void set_result(RegionNode* region, PhiNode* value); 145 Node* result() { return _result; } 146 147 virtual int reexecute_sp() { return _reexecute_sp; } 148 149 // Helper functions to inline natives 150 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 151 Node* generate_slow_guard(Node* test, RegionNode* region); 152 Node* generate_fair_guard(Node* test, RegionNode* region); 153 Node* generate_negative_guard(Node* index, RegionNode* region, 154 // resulting CastII of index: 155 Node* *pos_index = NULL); 156 Node* generate_limit_guard(Node* offset, Node* subseq_length, 157 Node* array_length, 158 RegionNode* region); 159 void generate_string_range_check(Node* array, Node* offset, 160 Node* length, bool char_count); 161 Node* generate_current_thread(Node* &tls_output); 162 Node* load_mirror_from_klass(Node* klass); 163 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 164 RegionNode* region, int null_path, 165 int offset); 166 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 167 RegionNode* region, int null_path) { 168 int offset = java_lang_Class::klass_offset_in_bytes(); 169 return load_klass_from_mirror_common(mirror, never_see_null, 170 region, null_path, 171 offset); 172 } 173 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 174 RegionNode* region, int null_path) { 175 int offset = java_lang_Class::array_klass_offset_in_bytes(); 176 return load_klass_from_mirror_common(mirror, never_see_null, 177 region, null_path, 178 offset); 179 } 180 Node* generate_access_flags_guard(Node* kls, 181 int modifier_mask, int modifier_bits, 182 RegionNode* region); 183 Node* generate_interface_guard(Node* kls, RegionNode* region); 184 Node* generate_array_guard(Node* kls, RegionNode* region) { 185 return generate_array_guard_common(kls, region, false, false); 186 } 187 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 188 return generate_array_guard_common(kls, region, false, true); 189 } 190 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 191 return generate_array_guard_common(kls, region, true, false); 192 } 193 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 194 return generate_array_guard_common(kls, region, true, true); 195 } 196 Node* generate_array_guard_common(Node* kls, RegionNode* region, 197 bool obj_array, bool not_array); 198 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 199 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 200 bool is_virtual = false, bool is_static = false); 201 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 202 return generate_method_call(method_id, false, true); 203 } 204 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 205 return generate_method_call(method_id, true, false); 206 } 207 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 208 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 209 210 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae); 211 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae); 212 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae); 213 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae); 214 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 215 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae); 216 bool inline_string_indexOfChar(); 217 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae); 218 bool inline_string_toBytesU(); 219 bool inline_string_getCharsU(); 220 bool inline_string_copy(bool compress); 221 bool inline_string_char_access(bool is_store); 222 Node* round_double_node(Node* n); 223 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 224 bool inline_math_native(vmIntrinsics::ID id); 225 bool inline_math(vmIntrinsics::ID id); 226 template <typename OverflowOp> 227 bool inline_math_overflow(Node* arg1, Node* arg2); 228 void inline_math_mathExact(Node* math, Node* test); 229 bool inline_math_addExactI(bool is_increment); 230 bool inline_math_addExactL(bool is_increment); 231 bool inline_math_multiplyExactI(); 232 bool inline_math_multiplyExactL(); 233 bool inline_math_negateExactI(); 234 bool inline_math_negateExactL(); 235 bool inline_math_subtractExactI(bool is_decrement); 236 bool inline_math_subtractExactL(bool is_decrement); 237 bool inline_min_max(vmIntrinsics::ID id); 238 bool inline_notify(vmIntrinsics::ID id); 239 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 240 // This returns Type::AnyPtr, RawPtr, or OopPtr. 241 int classify_unsafe_addr(Node* &base, Node* &offset); 242 Node* make_unsafe_address(Node* base, Node* offset); 243 // Helper for inline_unsafe_access. 244 // Generates the guards that check whether the result of 245 // Unsafe.getObject should be recorded in an SATB log buffer. 246 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 247 248 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind; 249 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned); 250 static bool klass_needs_init_guard(Node* kls); 251 bool inline_unsafe_allocate(); 252 bool inline_unsafe_newArray(bool uninitialized); 253 bool inline_unsafe_copyMemory(); 254 bool inline_native_currentThread(); 255 256 bool inline_native_time_funcs(address method, const char* funcName); 257 bool inline_native_isInterrupted(); 258 bool inline_native_Class_query(vmIntrinsics::ID id); 259 bool inline_native_subtype_check(); 260 bool inline_native_getLength(); 261 bool inline_array_copyOf(bool is_copyOfRange); 262 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae); 263 bool inline_preconditions_checkIndex(); 264 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 265 bool inline_native_clone(bool is_virtual); 266 bool inline_native_Reflection_getCallerClass(); 267 // Helper function for inlining native object hash method 268 bool inline_native_hashcode(bool is_virtual, bool is_static); 269 bool inline_native_getClass(); 270 271 // Helper functions for inlining arraycopy 272 bool inline_arraycopy(); 273 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 274 RegionNode* slow_region); 275 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 276 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 277 278 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind; 279 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store); 280 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind); 281 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind); 282 bool inline_unsafe_fence(vmIntrinsics::ID id); 283 bool inline_onspinwait(); 284 bool inline_fp_conversions(vmIntrinsics::ID id); 285 bool inline_number_methods(vmIntrinsics::ID id); 286 bool inline_reference_get(); 287 bool inline_Class_cast(); 288 bool inline_aescrypt_Block(vmIntrinsics::ID id); 289 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 290 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id); 291 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 292 Node* inline_counterMode_AESCrypt_predicate(); 293 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 294 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 295 bool inline_ghash_processBlocks(); 296 bool inline_sha_implCompress(vmIntrinsics::ID id); 297 bool inline_digestBase_implCompressMB(int predicate); 298 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 299 bool long_state, address stubAddr, const char *stubName, 300 Node* src_start, Node* ofs, Node* limit); 301 Node* get_state_from_sha_object(Node *sha_object); 302 Node* get_state_from_sha5_object(Node *sha_object); 303 Node* inline_digestBase_implCompressMB_predicate(int predicate); 304 bool inline_encodeISOArray(); 305 bool inline_updateCRC32(); 306 bool inline_updateBytesCRC32(); 307 bool inline_updateByteBufferCRC32(); 308 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 309 bool inline_updateBytesCRC32C(); 310 bool inline_updateDirectByteBufferCRC32C(); 311 bool inline_updateBytesAdler32(); 312 bool inline_updateByteBufferAdler32(); 313 bool inline_multiplyToLen(); 314 bool inline_hasNegatives(); 315 bool inline_squareToLen(); 316 bool inline_mulAdd(); 317 bool inline_montgomeryMultiply(); 318 bool inline_montgomerySquare(); 319 bool inline_vectorizedMismatch(); 320 321 bool inline_profileBoolean(); 322 bool inline_isCompileConstant(); 323}; 324 325//---------------------------make_vm_intrinsic---------------------------- 326CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 327 vmIntrinsics::ID id = m->intrinsic_id(); 328 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 329 330 if (!m->is_loaded()) { 331 // Do not attempt to inline unloaded methods. 332 return NULL; 333 } 334 335 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization); 336 bool is_available = false; 337 338 { 339 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag 340 // the compiler must transition to '_thread_in_vm' state because both 341 // methods access VM-internal data. 342 VM_ENTRY_MARK; 343 methodHandle mh(THREAD, m->get_Method()); 344 is_available = compiler->is_intrinsic_supported(mh, is_virtual) && 345 !C->directive()->is_intrinsic_disabled(mh) && 346 !vmIntrinsics::is_disabled_by_flags(mh); 347 348 } 349 350 if (is_available) { 351 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 352 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 353 return new LibraryIntrinsic(m, is_virtual, 354 vmIntrinsics::predicates_needed(id), 355 vmIntrinsics::does_virtual_dispatch(id), 356 (vmIntrinsics::ID) id); 357 } else { 358 return NULL; 359 } 360} 361 362//----------------------register_library_intrinsics----------------------- 363// Initialize this file's data structures, for each Compile instance. 364void Compile::register_library_intrinsics() { 365 // Nothing to do here. 366} 367 368JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 369 LibraryCallKit kit(jvms, this); 370 Compile* C = kit.C; 371 int nodes = C->unique(); 372#ifndef PRODUCT 373 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 374 char buf[1000]; 375 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 376 tty->print_cr("Intrinsic %s", str); 377 } 378#endif 379 ciMethod* callee = kit.callee(); 380 const int bci = kit.bci(); 381 382 // Try to inline the intrinsic. 383 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) && 384 kit.try_to_inline(_last_predicate)) { 385 if (C->print_intrinsics() || C->print_inlining()) { 386 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 387 } 388 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 389 if (C->log()) { 390 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 391 vmIntrinsics::name_at(intrinsic_id()), 392 (is_virtual() ? " virtual='1'" : ""), 393 C->unique() - nodes); 394 } 395 // Push the result from the inlined method onto the stack. 396 kit.push_result(); 397 C->print_inlining_update(this); 398 return kit.transfer_exceptions_into_jvms(); 399 } 400 401 // The intrinsic bailed out 402 if (C->print_intrinsics() || C->print_inlining()) { 403 if (jvms->has_method()) { 404 // Not a root compile. 405 const char* msg; 406 if (callee->intrinsic_candidate()) { 407 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 408 } else { 409 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated" 410 : "failed to inline (intrinsic), method not annotated"; 411 } 412 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 413 } else { 414 // Root compile 415 tty->print("Did not generate intrinsic %s%s at bci:%d in", 416 vmIntrinsics::name_at(intrinsic_id()), 417 (is_virtual() ? " (virtual)" : ""), bci); 418 } 419 } 420 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 421 C->print_inlining_update(this); 422 return NULL; 423} 424 425Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 426 LibraryCallKit kit(jvms, this); 427 Compile* C = kit.C; 428 int nodes = C->unique(); 429 _last_predicate = predicate; 430#ifndef PRODUCT 431 assert(is_predicated() && predicate < predicates_count(), "sanity"); 432 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 433 char buf[1000]; 434 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 435 tty->print_cr("Predicate for intrinsic %s", str); 436 } 437#endif 438 ciMethod* callee = kit.callee(); 439 const int bci = kit.bci(); 440 441 Node* slow_ctl = kit.try_to_predicate(predicate); 442 if (!kit.failing()) { 443 if (C->print_intrinsics() || C->print_inlining()) { 444 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 445 } 446 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 447 if (C->log()) { 448 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 449 vmIntrinsics::name_at(intrinsic_id()), 450 (is_virtual() ? " virtual='1'" : ""), 451 C->unique() - nodes); 452 } 453 return slow_ctl; // Could be NULL if the check folds. 454 } 455 456 // The intrinsic bailed out 457 if (C->print_intrinsics() || C->print_inlining()) { 458 if (jvms->has_method()) { 459 // Not a root compile. 460 const char* msg = "failed to generate predicate for intrinsic"; 461 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 462 } else { 463 // Root compile 464 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 465 vmIntrinsics::name_at(intrinsic_id()), 466 (is_virtual() ? " (virtual)" : ""), bci); 467 } 468 } 469 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 470 return NULL; 471} 472 473bool LibraryCallKit::try_to_inline(int predicate) { 474 // Handle symbolic names for otherwise undistinguished boolean switches: 475 const bool is_store = true; 476 const bool is_compress = true; 477 const bool is_static = true; 478 const bool is_volatile = true; 479 480 if (!jvms()->has_method()) { 481 // Root JVMState has a null method. 482 assert(map()->memory()->Opcode() == Op_Parm, ""); 483 // Insert the memory aliasing node 484 set_all_memory(reset_memory()); 485 } 486 assert(merged_memory(), ""); 487 488 489 switch (intrinsic_id()) { 490 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 491 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 492 case vmIntrinsics::_getClass: return inline_native_getClass(); 493 494 case vmIntrinsics::_dsin: 495 case vmIntrinsics::_dcos: 496 case vmIntrinsics::_dtan: 497 case vmIntrinsics::_dabs: 498 case vmIntrinsics::_datan2: 499 case vmIntrinsics::_dsqrt: 500 case vmIntrinsics::_dexp: 501 case vmIntrinsics::_dlog: 502 case vmIntrinsics::_dlog10: 503 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 504 505 case vmIntrinsics::_min: 506 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 507 508 case vmIntrinsics::_notify: 509 case vmIntrinsics::_notifyAll: 510 if (InlineNotify) { 511 return inline_notify(intrinsic_id()); 512 } 513 return false; 514 515 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 516 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 517 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 518 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 519 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 520 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 521 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 522 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 523 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 524 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 525 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 526 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 527 528 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 529 530 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL); 531 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU); 532 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU); 533 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL); 534 535 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL); 536 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU); 537 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL); 538 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL); 539 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU); 540 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL); 541 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar(); 542 543 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL); 544 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU); 545 546 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU(); 547 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU(); 548 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store); 549 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store); 550 551 case vmIntrinsics::_compressStringC: 552 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress); 553 case vmIntrinsics::_inflateStringC: 554 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress); 555 556 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false); 557 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false); 558 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false); 559 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false); 560 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false); 561 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false); 562 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false); 563 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false); 564 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false); 565 566 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false); 567 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false); 568 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false); 569 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false); 570 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false); 571 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false); 572 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false); 573 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false); 574 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false); 575 576 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false); 577 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false); 578 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false); 579 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false); 580 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false); 581 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false); 582 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false); 583 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false); 584 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false); 585 586 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false); 587 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false); 588 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false); 589 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false); 590 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false); 591 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false); 592 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false); 593 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false); 594 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false); 595 596 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true); 597 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true); 598 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true); 599 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true); 600 601 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true); 602 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true); 603 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true); 604 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true); 605 606 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false); 607 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false); 608 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false); 609 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false); 610 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false); 611 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false); 612 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false); 613 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false); 614 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false); 615 616 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false); 617 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false); 618 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false); 619 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false); 620 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false); 621 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false); 622 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false); 623 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false); 624 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false); 625 626 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false); 627 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false); 628 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false); 629 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false); 630 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false); 631 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false); 632 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false); 633 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false); 634 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false); 635 636 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false); 637 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false); 638 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false); 639 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false); 640 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false); 641 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false); 642 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false); 643 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false); 644 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false); 645 646 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile); 647 case vmIntrinsics::_compareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile); 648 case vmIntrinsics::_compareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile); 649 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile); 650 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile); 651 652 case vmIntrinsics::_weakCompareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed); 653 case vmIntrinsics::_weakCompareAndSwapObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire); 654 case vmIntrinsics::_weakCompareAndSwapObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release); 655 case vmIntrinsics::_weakCompareAndSwapObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile); 656 case vmIntrinsics::_weakCompareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed); 657 case vmIntrinsics::_weakCompareAndSwapByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire); 658 case vmIntrinsics::_weakCompareAndSwapByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release); 659 case vmIntrinsics::_weakCompareAndSwapByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile); 660 case vmIntrinsics::_weakCompareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed); 661 case vmIntrinsics::_weakCompareAndSwapShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire); 662 case vmIntrinsics::_weakCompareAndSwapShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release); 663 case vmIntrinsics::_weakCompareAndSwapShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile); 664 case vmIntrinsics::_weakCompareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed); 665 case vmIntrinsics::_weakCompareAndSwapIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire); 666 case vmIntrinsics::_weakCompareAndSwapIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release); 667 case vmIntrinsics::_weakCompareAndSwapIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile); 668 case vmIntrinsics::_weakCompareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed); 669 case vmIntrinsics::_weakCompareAndSwapLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire); 670 case vmIntrinsics::_weakCompareAndSwapLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release); 671 case vmIntrinsics::_weakCompareAndSwapLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile); 672 673 case vmIntrinsics::_compareAndExchangeObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile); 674 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire); 675 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release); 676 case vmIntrinsics::_compareAndExchangeByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile); 677 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire); 678 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release); 679 case vmIntrinsics::_compareAndExchangeShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile); 680 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire); 681 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release); 682 case vmIntrinsics::_compareAndExchangeIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile); 683 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire); 684 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release); 685 case vmIntrinsics::_compareAndExchangeLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile); 686 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire); 687 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release); 688 689 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile); 690 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile); 691 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile); 692 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile); 693 694 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile); 695 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile); 696 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile); 697 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile); 698 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile); 699 700 case vmIntrinsics::_loadFence: 701 case vmIntrinsics::_storeFence: 702 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 703 704 case vmIntrinsics::_onSpinWait: return inline_onspinwait(); 705 706 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 707 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 708 709#ifdef TRACE_HAVE_INTRINSICS 710 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 711#endif 712 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 713 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 714 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 715 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 716 case vmIntrinsics::_getLength: return inline_native_getLength(); 717 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 718 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 719 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL); 720 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU); 721 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex(); 722 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 723 724 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true); 725 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false); 726 727 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 728 729 case vmIntrinsics::_isInstance: 730 case vmIntrinsics::_getModifiers: 731 case vmIntrinsics::_isInterface: 732 case vmIntrinsics::_isArray: 733 case vmIntrinsics::_isPrimitive: 734 case vmIntrinsics::_getSuperclass: 735 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 736 737 case vmIntrinsics::_floatToRawIntBits: 738 case vmIntrinsics::_floatToIntBits: 739 case vmIntrinsics::_intBitsToFloat: 740 case vmIntrinsics::_doubleToRawLongBits: 741 case vmIntrinsics::_doubleToLongBits: 742 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 743 744 case vmIntrinsics::_numberOfLeadingZeros_i: 745 case vmIntrinsics::_numberOfLeadingZeros_l: 746 case vmIntrinsics::_numberOfTrailingZeros_i: 747 case vmIntrinsics::_numberOfTrailingZeros_l: 748 case vmIntrinsics::_bitCount_i: 749 case vmIntrinsics::_bitCount_l: 750 case vmIntrinsics::_reverseBytes_i: 751 case vmIntrinsics::_reverseBytes_l: 752 case vmIntrinsics::_reverseBytes_s: 753 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 754 755 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 756 757 case vmIntrinsics::_Reference_get: return inline_reference_get(); 758 759 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 760 761 case vmIntrinsics::_aescrypt_encryptBlock: 762 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 763 764 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 765 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 766 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 767 768 case vmIntrinsics::_counterMode_AESCrypt: 769 return inline_counterMode_AESCrypt(intrinsic_id()); 770 771 case vmIntrinsics::_sha_implCompress: 772 case vmIntrinsics::_sha2_implCompress: 773 case vmIntrinsics::_sha5_implCompress: 774 return inline_sha_implCompress(intrinsic_id()); 775 776 case vmIntrinsics::_digestBase_implCompressMB: 777 return inline_digestBase_implCompressMB(predicate); 778 779 case vmIntrinsics::_multiplyToLen: 780 return inline_multiplyToLen(); 781 782 case vmIntrinsics::_squareToLen: 783 return inline_squareToLen(); 784 785 case vmIntrinsics::_mulAdd: 786 return inline_mulAdd(); 787 788 case vmIntrinsics::_montgomeryMultiply: 789 return inline_montgomeryMultiply(); 790 case vmIntrinsics::_montgomerySquare: 791 return inline_montgomerySquare(); 792 793 case vmIntrinsics::_vectorizedMismatch: 794 return inline_vectorizedMismatch(); 795 796 case vmIntrinsics::_ghash_processBlocks: 797 return inline_ghash_processBlocks(); 798 799 case vmIntrinsics::_encodeISOArray: 800 case vmIntrinsics::_encodeByteISOArray: 801 return inline_encodeISOArray(); 802 803 case vmIntrinsics::_updateCRC32: 804 return inline_updateCRC32(); 805 case vmIntrinsics::_updateBytesCRC32: 806 return inline_updateBytesCRC32(); 807 case vmIntrinsics::_updateByteBufferCRC32: 808 return inline_updateByteBufferCRC32(); 809 810 case vmIntrinsics::_updateBytesCRC32C: 811 return inline_updateBytesCRC32C(); 812 case vmIntrinsics::_updateDirectByteBufferCRC32C: 813 return inline_updateDirectByteBufferCRC32C(); 814 815 case vmIntrinsics::_updateBytesAdler32: 816 return inline_updateBytesAdler32(); 817 case vmIntrinsics::_updateByteBufferAdler32: 818 return inline_updateByteBufferAdler32(); 819 820 case vmIntrinsics::_profileBoolean: 821 return inline_profileBoolean(); 822 case vmIntrinsics::_isCompileConstant: 823 return inline_isCompileConstant(); 824 825 case vmIntrinsics::_hasNegatives: 826 return inline_hasNegatives(); 827 828 default: 829 // If you get here, it may be that someone has added a new intrinsic 830 // to the list in vmSymbols.hpp without implementing it here. 831#ifndef PRODUCT 832 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 833 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 834 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 835 } 836#endif 837 return false; 838 } 839} 840 841Node* LibraryCallKit::try_to_predicate(int predicate) { 842 if (!jvms()->has_method()) { 843 // Root JVMState has a null method. 844 assert(map()->memory()->Opcode() == Op_Parm, ""); 845 // Insert the memory aliasing node 846 set_all_memory(reset_memory()); 847 } 848 assert(merged_memory(), ""); 849 850 switch (intrinsic_id()) { 851 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 852 return inline_cipherBlockChaining_AESCrypt_predicate(false); 853 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 854 return inline_cipherBlockChaining_AESCrypt_predicate(true); 855 case vmIntrinsics::_counterMode_AESCrypt: 856 return inline_counterMode_AESCrypt_predicate(); 857 case vmIntrinsics::_digestBase_implCompressMB: 858 return inline_digestBase_implCompressMB_predicate(predicate); 859 860 default: 861 // If you get here, it may be that someone has added a new intrinsic 862 // to the list in vmSymbols.hpp without implementing it here. 863#ifndef PRODUCT 864 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 865 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 866 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 867 } 868#endif 869 Node* slow_ctl = control(); 870 set_control(top()); // No fast path instrinsic 871 return slow_ctl; 872 } 873} 874 875//------------------------------set_result------------------------------- 876// Helper function for finishing intrinsics. 877void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 878 record_for_igvn(region); 879 set_control(_gvn.transform(region)); 880 set_result( _gvn.transform(value)); 881 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 882} 883 884//------------------------------generate_guard--------------------------- 885// Helper function for generating guarded fast-slow graph structures. 886// The given 'test', if true, guards a slow path. If the test fails 887// then a fast path can be taken. (We generally hope it fails.) 888// In all cases, GraphKit::control() is updated to the fast path. 889// The returned value represents the control for the slow path. 890// The return value is never 'top'; it is either a valid control 891// or NULL if it is obvious that the slow path can never be taken. 892// Also, if region and the slow control are not NULL, the slow edge 893// is appended to the region. 894Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 895 if (stopped()) { 896 // Already short circuited. 897 return NULL; 898 } 899 900 // Build an if node and its projections. 901 // If test is true we take the slow path, which we assume is uncommon. 902 if (_gvn.type(test) == TypeInt::ZERO) { 903 // The slow branch is never taken. No need to build this guard. 904 return NULL; 905 } 906 907 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 908 909 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 910 if (if_slow == top()) { 911 // The slow branch is never taken. No need to build this guard. 912 return NULL; 913 } 914 915 if (region != NULL) 916 region->add_req(if_slow); 917 918 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 919 set_control(if_fast); 920 921 return if_slow; 922} 923 924inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 925 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 926} 927inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 928 return generate_guard(test, region, PROB_FAIR); 929} 930 931inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 932 Node* *pos_index) { 933 if (stopped()) 934 return NULL; // already stopped 935 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 936 return NULL; // index is already adequately typed 937 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 938 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 939 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 940 if (is_neg != NULL && pos_index != NULL) { 941 // Emulate effect of Parse::adjust_map_after_if. 942 Node* ccast = new CastIINode(index, TypeInt::POS); 943 ccast->set_req(0, control()); 944 (*pos_index) = _gvn.transform(ccast); 945 } 946 return is_neg; 947} 948 949// Make sure that 'position' is a valid limit index, in [0..length]. 950// There are two equivalent plans for checking this: 951// A. (offset + copyLength) unsigned<= arrayLength 952// B. offset <= (arrayLength - copyLength) 953// We require that all of the values above, except for the sum and 954// difference, are already known to be non-negative. 955// Plan A is robust in the face of overflow, if offset and copyLength 956// are both hugely positive. 957// 958// Plan B is less direct and intuitive, but it does not overflow at 959// all, since the difference of two non-negatives is always 960// representable. Whenever Java methods must perform the equivalent 961// check they generally use Plan B instead of Plan A. 962// For the moment we use Plan A. 963inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 964 Node* subseq_length, 965 Node* array_length, 966 RegionNode* region) { 967 if (stopped()) 968 return NULL; // already stopped 969 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 970 if (zero_offset && subseq_length->eqv_uncast(array_length)) 971 return NULL; // common case of whole-array copy 972 Node* last = subseq_length; 973 if (!zero_offset) // last += offset 974 last = _gvn.transform(new AddINode(last, offset)); 975 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 976 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 977 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 978 return is_over; 979} 980 981// Emit range checks for the given String.value byte array 982void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) { 983 if (stopped()) { 984 return; // already stopped 985 } 986 RegionNode* bailout = new RegionNode(1); 987 record_for_igvn(bailout); 988 if (char_count) { 989 // Convert char count to byte count 990 count = _gvn.transform(new LShiftINode(count, intcon(1))); 991 } 992 993 // Offset and count must not be negative 994 generate_negative_guard(offset, bailout); 995 generate_negative_guard(count, bailout); 996 // Offset + count must not exceed length of array 997 generate_limit_guard(offset, count, load_array_length(array), bailout); 998 999 if (bailout->req() > 1) { 1000 PreserveJVMState pjvms(this); 1001 set_control(_gvn.transform(bailout)); 1002 uncommon_trap(Deoptimization::Reason_intrinsic, 1003 Deoptimization::Action_maybe_recompile); 1004 } 1005} 1006 1007//--------------------------generate_current_thread-------------------- 1008Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1009 ciKlass* thread_klass = env()->Thread_klass(); 1010 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1011 Node* thread = _gvn.transform(new ThreadLocalNode()); 1012 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1013 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1014 tls_output = thread; 1015 return threadObj; 1016} 1017 1018 1019//------------------------------make_string_method_node------------------------ 1020// Helper method for String intrinsic functions. This version is called with 1021// str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded 1022// characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes 1023// containing the lengths of str1 and str2. 1024Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) { 1025 Node* result = NULL; 1026 switch (opcode) { 1027 case Op_StrIndexOf: 1028 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES), 1029 str1_start, cnt1, str2_start, cnt2, ae); 1030 break; 1031 case Op_StrComp: 1032 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES), 1033 str1_start, cnt1, str2_start, cnt2, ae); 1034 break; 1035 case Op_StrEquals: 1036 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals'). 1037 // Use the constant length if there is one because optimized match rule may exist. 1038 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES), 1039 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae); 1040 break; 1041 default: 1042 ShouldNotReachHere(); 1043 return NULL; 1044 } 1045 1046 // All these intrinsics have checks. 1047 C->set_has_split_ifs(true); // Has chance for split-if optimization 1048 1049 return _gvn.transform(result); 1050} 1051 1052//------------------------------inline_string_compareTo------------------------ 1053bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) { 1054 Node* arg1 = argument(0); 1055 Node* arg2 = argument(1); 1056 1057 // Get start addr and length of first argument 1058 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1059 Node* arg1_cnt = load_array_length(arg1); 1060 1061 // Get start addr and length of second argument 1062 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1063 Node* arg2_cnt = load_array_length(arg2); 1064 1065 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1066 set_result(result); 1067 return true; 1068} 1069 1070//------------------------------inline_string_equals------------------------ 1071bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) { 1072 Node* arg1 = argument(0); 1073 Node* arg2 = argument(1); 1074 1075 // paths (plus control) merge 1076 RegionNode* region = new RegionNode(3); 1077 Node* phi = new PhiNode(region, TypeInt::BOOL); 1078 1079 if (!stopped()) { 1080 // Get start addr and length of first argument 1081 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE); 1082 Node* arg1_cnt = load_array_length(arg1); 1083 1084 // Get start addr and length of second argument 1085 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE); 1086 Node* arg2_cnt = load_array_length(arg2); 1087 1088 // Check for arg1_cnt != arg2_cnt 1089 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt)); 1090 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1091 Node* if_ne = generate_slow_guard(bol, NULL); 1092 if (if_ne != NULL) { 1093 phi->init_req(2, intcon(0)); 1094 region->init_req(2, if_ne); 1095 } 1096 1097 // Check for count == 0 is done by assembler code for StrEquals. 1098 1099 if (!stopped()) { 1100 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae); 1101 phi->init_req(1, equals); 1102 region->init_req(1, control()); 1103 } 1104 } 1105 1106 // post merge 1107 set_control(_gvn.transform(region)); 1108 record_for_igvn(region); 1109 1110 set_result(_gvn.transform(phi)); 1111 return true; 1112} 1113 1114//------------------------------inline_array_equals---------------------------- 1115bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) { 1116 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types"); 1117 Node* arg1 = argument(0); 1118 Node* arg2 = argument(1); 1119 1120 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES; 1121 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae))); 1122 return true; 1123} 1124 1125//------------------------------inline_hasNegatives------------------------------ 1126bool LibraryCallKit::inline_hasNegatives() { 1127 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1128 return false; 1129 } 1130 1131 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters"); 1132 // no receiver since it is static method 1133 Node* ba = argument(0); 1134 Node* offset = argument(1); 1135 Node* len = argument(2); 1136 1137 // Range checks 1138 generate_string_range_check(ba, offset, len, false); 1139 if (stopped()) { 1140 return true; 1141 } 1142 Node* ba_start = array_element_address(ba, offset, T_BYTE); 1143 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len); 1144 set_result(_gvn.transform(result)); 1145 return true; 1146} 1147 1148bool LibraryCallKit::inline_preconditions_checkIndex() { 1149 Node* index = argument(0); 1150 Node* length = argument(1); 1151 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) { 1152 return false; 1153 } 1154 1155 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0))); 1156 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge)); 1157 1158 { 1159 BuildCutout unless(this, len_pos_bol, PROB_MAX); 1160 uncommon_trap(Deoptimization::Reason_intrinsic, 1161 Deoptimization::Action_make_not_entrant); 1162 } 1163 1164 if (stopped()) { 1165 return false; 1166 } 1167 1168 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length)); 1169 BoolTest::mask btest = BoolTest::lt; 1170 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest)); 1171 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN); 1172 _gvn.set_type(rc, rc->Value(&_gvn)); 1173 if (!rc_bool->is_Con()) { 1174 record_for_igvn(rc); 1175 } 1176 set_control(_gvn.transform(new IfTrueNode(rc))); 1177 { 1178 PreserveJVMState pjvms(this); 1179 set_control(_gvn.transform(new IfFalseNode(rc))); 1180 uncommon_trap(Deoptimization::Reason_range_check, 1181 Deoptimization::Action_make_not_entrant); 1182 } 1183 1184 if (stopped()) { 1185 return false; 1186 } 1187 1188 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax)); 1189 result->set_req(0, control()); 1190 result = _gvn.transform(result); 1191 set_result(result); 1192 replace_in_map(index, result); 1193 return true; 1194} 1195 1196//------------------------------inline_string_indexOf------------------------ 1197bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) { 1198 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1199 return false; 1200 } 1201 Node* src = argument(0); 1202 Node* tgt = argument(1); 1203 1204 // Make the merge point 1205 RegionNode* result_rgn = new RegionNode(4); 1206 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1207 1208 // Get start addr and length of source string 1209 Node* src_start = array_element_address(src, intcon(0), T_BYTE); 1210 Node* src_count = load_array_length(src); 1211 1212 // Get start addr and length of substring 1213 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1214 Node* tgt_count = load_array_length(tgt); 1215 1216 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) { 1217 // Divide src size by 2 if String is UTF16 encoded 1218 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1))); 1219 } 1220 if (ae == StrIntrinsicNode::UU) { 1221 // Divide substring size by 2 if String is UTF16 encoded 1222 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1))); 1223 } 1224 1225 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae); 1226 if (result != NULL) { 1227 result_phi->init_req(3, result); 1228 result_rgn->init_req(3, control()); 1229 } 1230 set_control(_gvn.transform(result_rgn)); 1231 record_for_igvn(result_rgn); 1232 set_result(_gvn.transform(result_phi)); 1233 1234 return true; 1235} 1236 1237//-----------------------------inline_string_indexOf----------------------- 1238bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) { 1239 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1240 return false; 1241 } 1242 if (!Matcher::match_rule_supported(Op_StrIndexOf)) { 1243 return false; 1244 } 1245 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments"); 1246 Node* src = argument(0); // byte[] 1247 Node* src_count = argument(1); // char count 1248 Node* tgt = argument(2); // byte[] 1249 Node* tgt_count = argument(3); // char count 1250 Node* from_index = argument(4); // char index 1251 1252 // Multiply byte array index by 2 if String is UTF16 encoded 1253 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1))); 1254 src_count = _gvn.transform(new SubINode(src_count, from_index)); 1255 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1256 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE); 1257 1258 // Range checks 1259 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL); 1260 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU); 1261 if (stopped()) { 1262 return true; 1263 } 1264 1265 RegionNode* region = new RegionNode(5); 1266 Node* phi = new PhiNode(region, TypeInt::INT); 1267 1268 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae); 1269 if (result != NULL) { 1270 // The result is index relative to from_index if substring was found, -1 otherwise. 1271 // Generate code which will fold into cmove. 1272 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1273 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1274 1275 Node* if_lt = generate_slow_guard(bol, NULL); 1276 if (if_lt != NULL) { 1277 // result == -1 1278 phi->init_req(3, result); 1279 region->init_req(3, if_lt); 1280 } 1281 if (!stopped()) { 1282 result = _gvn.transform(new AddINode(result, from_index)); 1283 phi->init_req(4, result); 1284 region->init_req(4, control()); 1285 } 1286 } 1287 1288 set_control(_gvn.transform(region)); 1289 record_for_igvn(region); 1290 set_result(_gvn.transform(phi)); 1291 1292 return true; 1293} 1294 1295// Create StrIndexOfNode with fast path checks 1296Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count, 1297 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) { 1298 // Check for substr count > string count 1299 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count)); 1300 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1301 Node* if_gt = generate_slow_guard(bol, NULL); 1302 if (if_gt != NULL) { 1303 phi->init_req(1, intcon(-1)); 1304 region->init_req(1, if_gt); 1305 } 1306 if (!stopped()) { 1307 // Check for substr count == 0 1308 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0))); 1309 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1310 Node* if_zero = generate_slow_guard(bol, NULL); 1311 if (if_zero != NULL) { 1312 phi->init_req(2, intcon(0)); 1313 region->init_req(2, if_zero); 1314 } 1315 } 1316 if (!stopped()) { 1317 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae); 1318 } 1319 return NULL; 1320} 1321 1322//-----------------------------inline_string_indexOfChar----------------------- 1323bool LibraryCallKit::inline_string_indexOfChar() { 1324 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1325 return false; 1326 } 1327 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) { 1328 return false; 1329 } 1330 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments"); 1331 Node* src = argument(0); // byte[] 1332 Node* tgt = argument(1); // tgt is int ch 1333 Node* from_index = argument(2); 1334 Node* max = argument(3); 1335 1336 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1))); 1337 Node* src_start = array_element_address(src, src_offset, T_BYTE); 1338 Node* src_count = _gvn.transform(new SubINode(max, from_index)); 1339 1340 // Range checks 1341 generate_string_range_check(src, src_offset, src_count, true); 1342 if (stopped()) { 1343 return true; 1344 } 1345 1346 RegionNode* region = new RegionNode(3); 1347 Node* phi = new PhiNode(region, TypeInt::INT); 1348 1349 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none); 1350 C->set_has_split_ifs(true); // Has chance for split-if optimization 1351 _gvn.transform(result); 1352 1353 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0))); 1354 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt)); 1355 1356 Node* if_lt = generate_slow_guard(bol, NULL); 1357 if (if_lt != NULL) { 1358 // result == -1 1359 phi->init_req(2, result); 1360 region->init_req(2, if_lt); 1361 } 1362 if (!stopped()) { 1363 result = _gvn.transform(new AddINode(result, from_index)); 1364 phi->init_req(1, result); 1365 region->init_req(1, control()); 1366 } 1367 set_control(_gvn.transform(region)); 1368 record_for_igvn(region); 1369 set_result(_gvn.transform(phi)); 1370 1371 return true; 1372} 1373//---------------------------inline_string_copy--------------------- 1374// compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[]) 1375// int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len) 1376// int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1377// compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[]) 1378// void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len) 1379// void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len) 1380bool LibraryCallKit::inline_string_copy(bool compress) { 1381 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1382 return false; 1383 } 1384 int nargs = 5; // 2 oops, 3 ints 1385 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments"); 1386 1387 Node* src = argument(0); 1388 Node* src_offset = argument(1); 1389 Node* dst = argument(2); 1390 Node* dst_offset = argument(3); 1391 Node* length = argument(4); 1392 1393 // Check for allocation before we add nodes that would confuse 1394 // tightly_coupled_allocation() 1395 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1396 1397 // Figure out the size and type of the elements we will be copying. 1398 const Type* src_type = src->Value(&_gvn); 1399 const Type* dst_type = dst->Value(&_gvn); 1400 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1401 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 1402 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) || 1403 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)), 1404 "Unsupported array types for inline_string_copy"); 1405 1406 // Convert char[] offsets to byte[] offsets 1407 bool convert_src = (compress && src_elem == T_BYTE); 1408 bool convert_dst = (!compress && dst_elem == T_BYTE); 1409 if (convert_src) { 1410 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1))); 1411 } else if (convert_dst) { 1412 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1))); 1413 } 1414 1415 // Range checks 1416 generate_string_range_check(src, src_offset, length, convert_src); 1417 generate_string_range_check(dst, dst_offset, length, convert_dst); 1418 if (stopped()) { 1419 return true; 1420 } 1421 1422 Node* src_start = array_element_address(src, src_offset, src_elem); 1423 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 1424 // 'src_start' points to src array + scaled offset 1425 // 'dst_start' points to dst array + scaled offset 1426 Node* count = NULL; 1427 if (compress) { 1428 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length); 1429 } else { 1430 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length); 1431 } 1432 1433 if (alloc != NULL) { 1434 if (alloc->maybe_set_complete(&_gvn)) { 1435 // "You break it, you buy it." 1436 InitializeNode* init = alloc->initialization(); 1437 assert(init->is_complete(), "we just did this"); 1438 init->set_complete_with_arraycopy(); 1439 assert(dst->is_CheckCastPP(), "sanity"); 1440 assert(dst->in(0)->in(0) == init, "dest pinned"); 1441 } 1442 // Do not let stores that initialize this object be reordered with 1443 // a subsequent store that would make this object accessible by 1444 // other threads. 1445 // Record what AllocateNode this StoreStore protects so that 1446 // escape analysis can go from the MemBarStoreStoreNode to the 1447 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1448 // based on the escape status of the AllocateNode. 1449 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1450 } 1451 if (compress) { 1452 set_result(_gvn.transform(count)); 1453 } 1454 return true; 1455} 1456 1457#ifdef _LP64 1458#define XTOP ,top() /*additional argument*/ 1459#else //_LP64 1460#define XTOP /*no additional argument*/ 1461#endif //_LP64 1462 1463//------------------------inline_string_toBytesU-------------------------- 1464// public static byte[] StringUTF16.toBytes(char[] value, int off, int len) 1465bool LibraryCallKit::inline_string_toBytesU() { 1466 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1467 return false; 1468 } 1469 // Get the arguments. 1470 Node* value = argument(0); 1471 Node* offset = argument(1); 1472 Node* length = argument(2); 1473 1474 Node* newcopy = NULL; 1475 1476 // Set the original stack and the reexecute bit for the interpreter to reexecute 1477 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens. 1478 { PreserveReexecuteState preexecs(this); 1479 jvms()->set_should_reexecute(true); 1480 1481 // Check if a null path was taken unconditionally. 1482 value = null_check(value); 1483 1484 RegionNode* bailout = new RegionNode(1); 1485 record_for_igvn(bailout); 1486 1487 // Range checks 1488 generate_negative_guard(offset, bailout); 1489 generate_negative_guard(length, bailout); 1490 generate_limit_guard(offset, length, load_array_length(value), bailout); 1491 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE 1492 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout); 1493 1494 if (bailout->req() > 1) { 1495 PreserveJVMState pjvms(this); 1496 set_control(_gvn.transform(bailout)); 1497 uncommon_trap(Deoptimization::Reason_intrinsic, 1498 Deoptimization::Action_maybe_recompile); 1499 } 1500 if (stopped()) { 1501 return true; 1502 } 1503 1504 Node* size = _gvn.transform(new LShiftINode(length, intcon(1))); 1505 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE))); 1506 newcopy = new_array(klass_node, size, 0); // no arguments to push 1507 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL); 1508 1509 // Calculate starting addresses. 1510 Node* src_start = array_element_address(value, offset, T_CHAR); 1511 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE)); 1512 1513 // Check if src array address is aligned to HeapWordSize (dst is always aligned) 1514 const TypeInt* toffset = gvn().type(offset)->is_int(); 1515 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1516 1517 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1518 const char* copyfunc_name = "arraycopy"; 1519 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1520 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1521 OptoRuntime::fast_arraycopy_Type(), 1522 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1523 src_start, dst_start, ConvI2X(length) XTOP); 1524 // Do not let reads from the cloned object float above the arraycopy. 1525 if (alloc != NULL) { 1526 if (alloc->maybe_set_complete(&_gvn)) { 1527 // "You break it, you buy it." 1528 InitializeNode* init = alloc->initialization(); 1529 assert(init->is_complete(), "we just did this"); 1530 init->set_complete_with_arraycopy(); 1531 assert(newcopy->is_CheckCastPP(), "sanity"); 1532 assert(newcopy->in(0)->in(0) == init, "dest pinned"); 1533 } 1534 // Do not let stores that initialize this object be reordered with 1535 // a subsequent store that would make this object accessible by 1536 // other threads. 1537 // Record what AllocateNode this StoreStore protects so that 1538 // escape analysis can go from the MemBarStoreStoreNode to the 1539 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1540 // based on the escape status of the AllocateNode. 1541 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1542 } else { 1543 insert_mem_bar(Op_MemBarCPUOrder); 1544 } 1545 } // original reexecute is set back here 1546 1547 C->set_has_split_ifs(true); // Has chance for split-if optimization 1548 if (!stopped()) { 1549 set_result(newcopy); 1550 } 1551 return true; 1552} 1553 1554//------------------------inline_string_getCharsU-------------------------- 1555// public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin) 1556bool LibraryCallKit::inline_string_getCharsU() { 1557 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 1558 return false; 1559 } 1560 1561 // Get the arguments. 1562 Node* src = argument(0); 1563 Node* src_begin = argument(1); 1564 Node* src_end = argument(2); // exclusive offset (i < src_end) 1565 Node* dst = argument(3); 1566 Node* dst_begin = argument(4); 1567 1568 // Check for allocation before we add nodes that would confuse 1569 // tightly_coupled_allocation() 1570 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL); 1571 1572 // Check if a null path was taken unconditionally. 1573 src = null_check(src); 1574 dst = null_check(dst); 1575 if (stopped()) { 1576 return true; 1577 } 1578 1579 // Get length and convert char[] offset to byte[] offset 1580 Node* length = _gvn.transform(new SubINode(src_end, src_begin)); 1581 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1))); 1582 1583 // Range checks 1584 generate_string_range_check(src, src_begin, length, true); 1585 generate_string_range_check(dst, dst_begin, length, false); 1586 if (stopped()) { 1587 return true; 1588 } 1589 1590 if (!stopped()) { 1591 // Calculate starting addresses. 1592 Node* src_start = array_element_address(src, src_begin, T_BYTE); 1593 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR); 1594 1595 // Check if array addresses are aligned to HeapWordSize 1596 const TypeInt* tsrc = gvn().type(src_begin)->is_int(); 1597 const TypeInt* tdst = gvn().type(dst_begin)->is_int(); 1598 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) && 1599 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0); 1600 1601 // Figure out which arraycopy runtime method to call (disjoint, uninitialized). 1602 const char* copyfunc_name = "arraycopy"; 1603 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true); 1604 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 1605 OptoRuntime::fast_arraycopy_Type(), 1606 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM, 1607 src_start, dst_start, ConvI2X(length) XTOP); 1608 // Do not let reads from the cloned object float above the arraycopy. 1609 if (alloc != NULL) { 1610 if (alloc->maybe_set_complete(&_gvn)) { 1611 // "You break it, you buy it." 1612 InitializeNode* init = alloc->initialization(); 1613 assert(init->is_complete(), "we just did this"); 1614 init->set_complete_with_arraycopy(); 1615 assert(dst->is_CheckCastPP(), "sanity"); 1616 assert(dst->in(0)->in(0) == init, "dest pinned"); 1617 } 1618 // Do not let stores that initialize this object be reordered with 1619 // a subsequent store that would make this object accessible by 1620 // other threads. 1621 // Record what AllocateNode this StoreStore protects so that 1622 // escape analysis can go from the MemBarStoreStoreNode to the 1623 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 1624 // based on the escape status of the AllocateNode. 1625 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 1626 } else { 1627 insert_mem_bar(Op_MemBarCPUOrder); 1628 } 1629 } 1630 1631 C->set_has_split_ifs(true); // Has chance for split-if optimization 1632 return true; 1633} 1634 1635//----------------------inline_string_char_access---------------------------- 1636// Store/Load char to/from byte[] array. 1637// static void StringUTF16.putChar(byte[] val, int index, int c) 1638// static char StringUTF16.getChar(byte[] val, int index) 1639bool LibraryCallKit::inline_string_char_access(bool is_store) { 1640 Node* value = argument(0); 1641 Node* index = argument(1); 1642 Node* ch = is_store ? argument(2) : NULL; 1643 1644 // This intrinsic accesses byte[] array as char[] array. Computing the offsets 1645 // correctly requires matched array shapes. 1646 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE), 1647 "sanity: byte[] and char[] bases agree"); 1648 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2, 1649 "sanity: byte[] and char[] scales agree"); 1650 1651 // Bail when getChar over constants is requested: constant folding would 1652 // reject folding mismatched char access over byte[]. A normal inlining for getChar 1653 // Java method would constant fold nicely instead. 1654 if (!is_store && value->is_Con() && index->is_Con()) { 1655 return false; 1656 } 1657 1658 Node* adr = array_element_address(value, index, T_CHAR); 1659 if (is_store) { 1660 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1661 false, false, true /* mismatched */); 1662 } else { 1663 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 1664 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */); 1665 set_result(ch); 1666 } 1667 return true; 1668} 1669 1670//--------------------------round_double_node-------------------------------- 1671// Round a double node if necessary. 1672Node* LibraryCallKit::round_double_node(Node* n) { 1673 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1674 n = _gvn.transform(new RoundDoubleNode(0, n)); 1675 return n; 1676} 1677 1678//------------------------------inline_math----------------------------------- 1679// public static double Math.abs(double) 1680// public static double Math.sqrt(double) 1681// public static double Math.log(double) 1682// public static double Math.log10(double) 1683bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1684 Node* arg = round_double_node(argument(0)); 1685 Node* n = NULL; 1686 switch (id) { 1687 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1688 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1689 default: fatal_unexpected_iid(id); break; 1690 } 1691 set_result(_gvn.transform(n)); 1692 return true; 1693} 1694 1695//------------------------------runtime_math----------------------------- 1696bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1697 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1698 "must be (DD)D or (D)D type"); 1699 1700 // Inputs 1701 Node* a = round_double_node(argument(0)); 1702 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1703 1704 const TypePtr* no_memory_effects = NULL; 1705 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1706 no_memory_effects, 1707 a, top(), b, b ? top() : NULL); 1708 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1709#ifdef ASSERT 1710 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1711 assert(value_top == top(), "second value must be top"); 1712#endif 1713 1714 set_result(value); 1715 return true; 1716} 1717 1718//------------------------------inline_math_native----------------------------- 1719bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1720#define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1721 switch (id) { 1722 // These intrinsics are not properly supported on all hardware 1723 case vmIntrinsics::_dsin: 1724 return StubRoutines::dsin() != NULL ? 1725 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") : 1726 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1727 case vmIntrinsics::_dcos: 1728 return StubRoutines::dcos() != NULL ? 1729 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") : 1730 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1731 case vmIntrinsics::_dtan: 1732 return StubRoutines::dtan() != NULL ? 1733 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") : 1734 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1735 case vmIntrinsics::_dlog: 1736 return StubRoutines::dlog() != NULL ? 1737 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") : 1738 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1739 case vmIntrinsics::_dlog10: 1740 return StubRoutines::dlog10() != NULL ? 1741 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") : 1742 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1743 1744 // These intrinsics are supported on all hardware 1745 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1746 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1747 1748 case vmIntrinsics::_dexp: 1749 return StubRoutines::dexp() != NULL ? 1750 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") : 1751 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1752 case vmIntrinsics::_dpow: 1753 return StubRoutines::dpow() != NULL ? 1754 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") : 1755 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1756#undef FN_PTR 1757 1758 // These intrinsics are not yet correctly implemented 1759 case vmIntrinsics::_datan2: 1760 return false; 1761 1762 default: 1763 fatal_unexpected_iid(id); 1764 return false; 1765 } 1766} 1767 1768static bool is_simple_name(Node* n) { 1769 return (n->req() == 1 // constant 1770 || (n->is_Type() && n->as_Type()->type()->singleton()) 1771 || n->is_Proj() // parameter or return value 1772 || n->is_Phi() // local of some sort 1773 ); 1774} 1775 1776//----------------------------inline_notify-----------------------------------* 1777bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 1778 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 1779 address func; 1780 if (id == vmIntrinsics::_notify) { 1781 func = OptoRuntime::monitor_notify_Java(); 1782 } else { 1783 func = OptoRuntime::monitor_notifyAll_Java(); 1784 } 1785 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 1786 make_slow_call_ex(call, env()->Throwable_klass(), false); 1787 return true; 1788} 1789 1790 1791//----------------------------inline_min_max----------------------------------- 1792bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1793 set_result(generate_min_max(id, argument(0), argument(1))); 1794 return true; 1795} 1796 1797void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 1798 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 1799 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 1800 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 1801 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 1802 1803 { 1804 PreserveJVMState pjvms(this); 1805 PreserveReexecuteState preexecs(this); 1806 jvms()->set_should_reexecute(true); 1807 1808 set_control(slow_path); 1809 set_i_o(i_o()); 1810 1811 uncommon_trap(Deoptimization::Reason_intrinsic, 1812 Deoptimization::Action_none); 1813 } 1814 1815 set_control(fast_path); 1816 set_result(math); 1817} 1818 1819template <typename OverflowOp> 1820bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 1821 typedef typename OverflowOp::MathOp MathOp; 1822 1823 MathOp* mathOp = new MathOp(arg1, arg2); 1824 Node* operation = _gvn.transform( mathOp ); 1825 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 1826 inline_math_mathExact(operation, ofcheck); 1827 return true; 1828} 1829 1830bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 1831 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 1832} 1833 1834bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 1835 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 1836} 1837 1838bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 1839 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 1840} 1841 1842bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 1843 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 1844} 1845 1846bool LibraryCallKit::inline_math_negateExactI() { 1847 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 1848} 1849 1850bool LibraryCallKit::inline_math_negateExactL() { 1851 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 1852} 1853 1854bool LibraryCallKit::inline_math_multiplyExactI() { 1855 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 1856} 1857 1858bool LibraryCallKit::inline_math_multiplyExactL() { 1859 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 1860} 1861 1862Node* 1863LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1864 // These are the candidate return value: 1865 Node* xvalue = x0; 1866 Node* yvalue = y0; 1867 1868 if (xvalue == yvalue) { 1869 return xvalue; 1870 } 1871 1872 bool want_max = (id == vmIntrinsics::_max); 1873 1874 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1875 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1876 if (txvalue == NULL || tyvalue == NULL) return top(); 1877 // This is not really necessary, but it is consistent with a 1878 // hypothetical MaxINode::Value method: 1879 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1880 1881 // %%% This folding logic should (ideally) be in a different place. 1882 // Some should be inside IfNode, and there to be a more reliable 1883 // transformation of ?: style patterns into cmoves. We also want 1884 // more powerful optimizations around cmove and min/max. 1885 1886 // Try to find a dominating comparison of these guys. 1887 // It can simplify the index computation for Arrays.copyOf 1888 // and similar uses of System.arraycopy. 1889 // First, compute the normalized version of CmpI(x, y). 1890 int cmp_op = Op_CmpI; 1891 Node* xkey = xvalue; 1892 Node* ykey = yvalue; 1893 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 1894 if (ideal_cmpxy->is_Cmp()) { 1895 // E.g., if we have CmpI(length - offset, count), 1896 // it might idealize to CmpI(length, count + offset) 1897 cmp_op = ideal_cmpxy->Opcode(); 1898 xkey = ideal_cmpxy->in(1); 1899 ykey = ideal_cmpxy->in(2); 1900 } 1901 1902 // Start by locating any relevant comparisons. 1903 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1904 Node* cmpxy = NULL; 1905 Node* cmpyx = NULL; 1906 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1907 Node* cmp = start_from->fast_out(k); 1908 if (cmp->outcnt() > 0 && // must have prior uses 1909 cmp->in(0) == NULL && // must be context-independent 1910 cmp->Opcode() == cmp_op) { // right kind of compare 1911 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1912 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1913 } 1914 } 1915 1916 const int NCMPS = 2; 1917 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1918 int cmpn; 1919 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1920 if (cmps[cmpn] != NULL) break; // find a result 1921 } 1922 if (cmpn < NCMPS) { 1923 // Look for a dominating test that tells us the min and max. 1924 int depth = 0; // Limit search depth for speed 1925 Node* dom = control(); 1926 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1927 if (++depth >= 100) break; 1928 Node* ifproj = dom; 1929 if (!ifproj->is_Proj()) continue; 1930 Node* iff = ifproj->in(0); 1931 if (!iff->is_If()) continue; 1932 Node* bol = iff->in(1); 1933 if (!bol->is_Bool()) continue; 1934 Node* cmp = bol->in(1); 1935 if (cmp == NULL) continue; 1936 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1937 if (cmps[cmpn] == cmp) break; 1938 if (cmpn == NCMPS) continue; 1939 BoolTest::mask btest = bol->as_Bool()->_test._test; 1940 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1941 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1942 // At this point, we know that 'x btest y' is true. 1943 switch (btest) { 1944 case BoolTest::eq: 1945 // They are proven equal, so we can collapse the min/max. 1946 // Either value is the answer. Choose the simpler. 1947 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1948 return yvalue; 1949 return xvalue; 1950 case BoolTest::lt: // x < y 1951 case BoolTest::le: // x <= y 1952 return (want_max ? yvalue : xvalue); 1953 case BoolTest::gt: // x > y 1954 case BoolTest::ge: // x >= y 1955 return (want_max ? xvalue : yvalue); 1956 } 1957 } 1958 } 1959 1960 // We failed to find a dominating test. 1961 // Let's pick a test that might GVN with prior tests. 1962 Node* best_bol = NULL; 1963 BoolTest::mask best_btest = BoolTest::illegal; 1964 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1965 Node* cmp = cmps[cmpn]; 1966 if (cmp == NULL) continue; 1967 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1968 Node* bol = cmp->fast_out(j); 1969 if (!bol->is_Bool()) continue; 1970 BoolTest::mask btest = bol->as_Bool()->_test._test; 1971 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1972 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1973 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1974 best_bol = bol->as_Bool(); 1975 best_btest = btest; 1976 } 1977 } 1978 } 1979 1980 Node* answer_if_true = NULL; 1981 Node* answer_if_false = NULL; 1982 switch (best_btest) { 1983 default: 1984 if (cmpxy == NULL) 1985 cmpxy = ideal_cmpxy; 1986 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 1987 // and fall through: 1988 case BoolTest::lt: // x < y 1989 case BoolTest::le: // x <= y 1990 answer_if_true = (want_max ? yvalue : xvalue); 1991 answer_if_false = (want_max ? xvalue : yvalue); 1992 break; 1993 case BoolTest::gt: // x > y 1994 case BoolTest::ge: // x >= y 1995 answer_if_true = (want_max ? xvalue : yvalue); 1996 answer_if_false = (want_max ? yvalue : xvalue); 1997 break; 1998 } 1999 2000 jint hi, lo; 2001 if (want_max) { 2002 // We can sharpen the minimum. 2003 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2004 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2005 } else { 2006 // We can sharpen the maximum. 2007 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2008 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2009 } 2010 2011 // Use a flow-free graph structure, to avoid creating excess control edges 2012 // which could hinder other optimizations. 2013 // Since Math.min/max is often used with arraycopy, we want 2014 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2015 Node* cmov = CMoveNode::make(NULL, best_bol, 2016 answer_if_false, answer_if_true, 2017 TypeInt::make(lo, hi, widen)); 2018 2019 return _gvn.transform(cmov); 2020 2021 /* 2022 // This is not as desirable as it may seem, since Min and Max 2023 // nodes do not have a full set of optimizations. 2024 // And they would interfere, anyway, with 'if' optimizations 2025 // and with CMoveI canonical forms. 2026 switch (id) { 2027 case vmIntrinsics::_min: 2028 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2029 case vmIntrinsics::_max: 2030 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2031 default: 2032 ShouldNotReachHere(); 2033 } 2034 */ 2035} 2036 2037inline int 2038LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2039 const TypePtr* base_type = TypePtr::NULL_PTR; 2040 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2041 if (base_type == NULL) { 2042 // Unknown type. 2043 return Type::AnyPtr; 2044 } else if (base_type == TypePtr::NULL_PTR) { 2045 // Since this is a NULL+long form, we have to switch to a rawptr. 2046 base = _gvn.transform(new CastX2PNode(offset)); 2047 offset = MakeConX(0); 2048 return Type::RawPtr; 2049 } else if (base_type->base() == Type::RawPtr) { 2050 return Type::RawPtr; 2051 } else if (base_type->isa_oopptr()) { 2052 // Base is never null => always a heap address. 2053 if (base_type->ptr() == TypePtr::NotNull) { 2054 return Type::OopPtr; 2055 } 2056 // Offset is small => always a heap address. 2057 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2058 if (offset_type != NULL && 2059 base_type->offset() == 0 && // (should always be?) 2060 offset_type->_lo >= 0 && 2061 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2062 return Type::OopPtr; 2063 } 2064 // Otherwise, it might either be oop+off or NULL+addr. 2065 return Type::AnyPtr; 2066 } else { 2067 // No information: 2068 return Type::AnyPtr; 2069 } 2070} 2071 2072inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2073 int kind = classify_unsafe_addr(base, offset); 2074 if (kind == Type::RawPtr) { 2075 return basic_plus_adr(top(), base, offset); 2076 } else { 2077 return basic_plus_adr(base, offset); 2078 } 2079} 2080 2081//--------------------------inline_number_methods----------------------------- 2082// inline int Integer.numberOfLeadingZeros(int) 2083// inline int Long.numberOfLeadingZeros(long) 2084// 2085// inline int Integer.numberOfTrailingZeros(int) 2086// inline int Long.numberOfTrailingZeros(long) 2087// 2088// inline int Integer.bitCount(int) 2089// inline int Long.bitCount(long) 2090// 2091// inline char Character.reverseBytes(char) 2092// inline short Short.reverseBytes(short) 2093// inline int Integer.reverseBytes(int) 2094// inline long Long.reverseBytes(long) 2095bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2096 Node* arg = argument(0); 2097 Node* n = NULL; 2098 switch (id) { 2099 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2100 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2101 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2102 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2103 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2104 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2105 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2106 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2107 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2108 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2109 default: fatal_unexpected_iid(id); break; 2110 } 2111 set_result(_gvn.transform(n)); 2112 return true; 2113} 2114 2115//----------------------------inline_unsafe_access---------------------------- 2116 2117// Helper that guards and inserts a pre-barrier. 2118void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2119 Node* pre_val, bool need_mem_bar) { 2120 // We could be accessing the referent field of a reference object. If so, when G1 2121 // is enabled, we need to log the value in the referent field in an SATB buffer. 2122 // This routine performs some compile time filters and generates suitable 2123 // runtime filters that guard the pre-barrier code. 2124 // Also add memory barrier for non volatile load from the referent field 2125 // to prevent commoning of loads across safepoint. 2126 if (!UseG1GC && !need_mem_bar) 2127 return; 2128 2129 // Some compile time checks. 2130 2131 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2132 const TypeX* otype = offset->find_intptr_t_type(); 2133 if (otype != NULL && otype->is_con() && 2134 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2135 // Constant offset but not the reference_offset so just return 2136 return; 2137 } 2138 2139 // We only need to generate the runtime guards for instances. 2140 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2141 if (btype != NULL) { 2142 if (btype->isa_aryptr()) { 2143 // Array type so nothing to do 2144 return; 2145 } 2146 2147 const TypeInstPtr* itype = btype->isa_instptr(); 2148 if (itype != NULL) { 2149 // Can the klass of base_oop be statically determined to be 2150 // _not_ a sub-class of Reference and _not_ Object? 2151 ciKlass* klass = itype->klass(); 2152 if ( klass->is_loaded() && 2153 !klass->is_subtype_of(env()->Reference_klass()) && 2154 !env()->Object_klass()->is_subtype_of(klass)) { 2155 return; 2156 } 2157 } 2158 } 2159 2160 // The compile time filters did not reject base_oop/offset so 2161 // we need to generate the following runtime filters 2162 // 2163 // if (offset == java_lang_ref_Reference::_reference_offset) { 2164 // if (instance_of(base, java.lang.ref.Reference)) { 2165 // pre_barrier(_, pre_val, ...); 2166 // } 2167 // } 2168 2169 float likely = PROB_LIKELY( 0.999); 2170 float unlikely = PROB_UNLIKELY(0.999); 2171 2172 IdealKit ideal(this); 2173#define __ ideal. 2174 2175 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2176 2177 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2178 // Update graphKit memory and control from IdealKit. 2179 sync_kit(ideal); 2180 2181 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2182 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2183 2184 // Update IdealKit memory and control from graphKit. 2185 __ sync_kit(this); 2186 2187 Node* one = __ ConI(1); 2188 // is_instof == 0 if base_oop == NULL 2189 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2190 2191 // Update graphKit from IdeakKit. 2192 sync_kit(ideal); 2193 2194 // Use the pre-barrier to record the value in the referent field 2195 pre_barrier(false /* do_load */, 2196 __ ctrl(), 2197 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2198 pre_val /* pre_val */, 2199 T_OBJECT); 2200 if (need_mem_bar) { 2201 // Add memory barrier to prevent commoning reads from this field 2202 // across safepoint since GC can change its value. 2203 insert_mem_bar(Op_MemBarCPUOrder); 2204 } 2205 // Update IdealKit from graphKit. 2206 __ sync_kit(this); 2207 2208 } __ end_if(); // _ref_type != ref_none 2209 } __ end_if(); // offset == referent_offset 2210 2211 // Final sync IdealKit and GraphKit. 2212 final_sync(ideal); 2213#undef __ 2214} 2215 2216 2217const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) { 2218 // Attempt to infer a sharper value type from the offset and base type. 2219 ciKlass* sharpened_klass = NULL; 2220 2221 // See if it is an instance field, with an object type. 2222 if (alias_type->field() != NULL) { 2223 if (alias_type->field()->type()->is_klass()) { 2224 sharpened_klass = alias_type->field()->type()->as_klass(); 2225 } 2226 } 2227 2228 // See if it is a narrow oop array. 2229 if (adr_type->isa_aryptr()) { 2230 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2231 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2232 if (elem_type != NULL) { 2233 sharpened_klass = elem_type->klass(); 2234 } 2235 } 2236 } 2237 2238 // The sharpened class might be unloaded if there is no class loader 2239 // contraint in place. 2240 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2241 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2242 2243#ifndef PRODUCT 2244 if (C->print_intrinsics() || C->print_inlining()) { 2245 tty->print(" from base type: "); adr_type->dump(); 2246 tty->print(" sharpened value: "); tjp->dump(); 2247 } 2248#endif 2249 // Sharpen the value type. 2250 return tjp; 2251 } 2252 return NULL; 2253} 2254 2255bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) { 2256 if (callee()->is_static()) return false; // caller must have the capability! 2257 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads"); 2258 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores"); 2259 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2260 2261#ifndef PRODUCT 2262 { 2263 ResourceMark rm; 2264 // Check the signatures. 2265 ciSignature* sig = callee()->signature(); 2266#ifdef ASSERT 2267 if (!is_store) { 2268 // Object getObject(Object base, int/long offset), etc. 2269 BasicType rtype = sig->return_type()->basic_type(); 2270 assert(rtype == type, "getter must return the expected value"); 2271 assert(sig->count() == 2, "oop getter has 2 arguments"); 2272 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2273 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2274 } else { 2275 // void putObject(Object base, int/long offset, Object x), etc. 2276 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2277 assert(sig->count() == 3, "oop putter has 3 arguments"); 2278 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2279 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2280 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2281 assert(vtype == type, "putter must accept the expected value"); 2282 } 2283#endif // ASSERT 2284 } 2285#endif //PRODUCT 2286 2287 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2288 2289 Node* receiver = argument(0); // type: oop 2290 2291 // Build address expression. 2292 Node* adr; 2293 Node* heap_base_oop = top(); 2294 Node* offset = top(); 2295 Node* val; 2296 2297 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2298 Node* base = argument(1); // type: oop 2299 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2300 offset = argument(2); // type: long 2301 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2302 // to be plain byte offsets, which are also the same as those accepted 2303 // by oopDesc::field_base. 2304 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2305 "fieldOffset must be byte-scaled"); 2306 // 32-bit machines ignore the high half! 2307 offset = ConvL2X(offset); 2308 adr = make_unsafe_address(base, offset); 2309 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) { 2310 heap_base_oop = base; 2311 } 2312 val = is_store ? argument(4) : NULL; 2313 2314 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2315 2316 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2317 // there was not enough information to nail it down. 2318 Compile::AliasType* alias_type = C->alias_type(adr_type); 2319 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2320 2321 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM || 2322 alias_type->basic_type() != T_ILLEGAL, "field, array element or unknown"); 2323 bool mismatched = false; 2324 BasicType bt = alias_type->basic_type(); 2325 if (bt != T_ILLEGAL) { 2326 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2327 // Alias type doesn't differentiate between byte[] and boolean[]). 2328 // Use address type to get the element type. 2329 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2330 } 2331 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2332 // accessing an array field with getObject is not a mismatch 2333 bt = T_OBJECT; 2334 } 2335 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2336 // Don't intrinsify mismatched object accesses 2337 return false; 2338 } 2339 mismatched = (bt != type); 2340 } 2341 2342 // First guess at the value type. 2343 const Type *value_type = Type::get_const_basic_type(type); 2344 2345 // We will need memory barriers unless we can determine a unique 2346 // alias category for this reference. (Note: If for some reason 2347 // the barriers get omitted and the unsafe reference begins to "pollute" 2348 // the alias analysis of the rest of the graph, either Compile::can_alias 2349 // or Compile::must_alias will throw a diagnostic assert.) 2350 bool need_mem_bar; 2351 switch (kind) { 2352 case Relaxed: 2353 need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2354 break; 2355 case Opaque: 2356 // Opaque uses CPUOrder membars for protection against code movement. 2357 case Acquire: 2358 case Release: 2359 case Volatile: 2360 need_mem_bar = true; 2361 break; 2362 default: 2363 ShouldNotReachHere(); 2364 } 2365 2366 // Some accesses require access atomicity for all types, notably longs and doubles. 2367 // When AlwaysAtomicAccesses is enabled, all accesses are atomic. 2368 bool requires_atomic_access = false; 2369 switch (kind) { 2370 case Relaxed: 2371 requires_atomic_access = AlwaysAtomicAccesses; 2372 break; 2373 case Opaque: 2374 // Opaque accesses are atomic. 2375 case Acquire: 2376 case Release: 2377 case Volatile: 2378 requires_atomic_access = true; 2379 break; 2380 default: 2381 ShouldNotReachHere(); 2382 } 2383 2384 // Figure out the memory ordering. 2385 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd 2386 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store); 2387 2388 // If we are reading the value of the referent field of a Reference 2389 // object (either by using Unsafe directly or through reflection) 2390 // then, if G1 is enabled, we need to record the referent in an 2391 // SATB log buffer using the pre-barrier mechanism. 2392 // Also we need to add memory barrier to prevent commoning reads 2393 // from this field across safepoint since GC can change its value. 2394 bool need_read_barrier = !is_store && 2395 offset != top() && heap_base_oop != top(); 2396 2397 if (!is_store && type == T_OBJECT) { 2398 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2399 if (tjp != NULL) { 2400 value_type = tjp; 2401 } 2402 } 2403 2404 receiver = null_check(receiver); 2405 if (stopped()) { 2406 return true; 2407 } 2408 // Heap pointers get a null-check from the interpreter, 2409 // as a courtesy. However, this is not guaranteed by Unsafe, 2410 // and it is not possible to fully distinguish unintended nulls 2411 // from intended ones in this API. 2412 2413 // We need to emit leading and trailing CPU membars (see below) in 2414 // addition to memory membars for special access modes. This is a little 2415 // too strong, but avoids the need to insert per-alias-type 2416 // volatile membars (for stores; compare Parse::do_put_xxx), which 2417 // we cannot do effectively here because we probably only have a 2418 // rough approximation of type. 2419 2420 switch(kind) { 2421 case Relaxed: 2422 case Opaque: 2423 case Acquire: 2424 break; 2425 case Release: 2426 case Volatile: 2427 if (is_store) { 2428 insert_mem_bar(Op_MemBarRelease); 2429 } else { 2430 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2431 insert_mem_bar(Op_MemBarVolatile); 2432 } 2433 } 2434 break; 2435 default: 2436 ShouldNotReachHere(); 2437 } 2438 2439 // Memory barrier to prevent normal and 'unsafe' accesses from 2440 // bypassing each other. Happens after null checks, so the 2441 // exception paths do not take memory state from the memory barrier, 2442 // so there's no problems making a strong assert about mixing users 2443 // of safe & unsafe memory. 2444 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2445 2446 if (!is_store) { 2447 Node* p = NULL; 2448 // Try to constant fold a load from a constant field 2449 ciField* field = alias_type->field(); 2450 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) { 2451 // final or stable field 2452 p = make_constant_from_field(field, heap_base_oop); 2453 } 2454 if (p == NULL) { 2455 // To be valid, unsafe loads may depend on other conditions than 2456 // the one that guards them: pin the Load node 2457 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, requires_atomic_access, unaligned, mismatched); 2458 // load value 2459 switch (type) { 2460 case T_BOOLEAN: 2461 case T_CHAR: 2462 case T_BYTE: 2463 case T_SHORT: 2464 case T_INT: 2465 case T_LONG: 2466 case T_FLOAT: 2467 case T_DOUBLE: 2468 break; 2469 case T_OBJECT: 2470 if (need_read_barrier) { 2471 // We do not require a mem bar inside pre_barrier if need_mem_bar 2472 // is set: the barriers would be emitted by us. 2473 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar); 2474 } 2475 break; 2476 case T_ADDRESS: 2477 // Cast to an int type. 2478 p = _gvn.transform(new CastP2XNode(NULL, p)); 2479 p = ConvX2UL(p); 2480 break; 2481 default: 2482 fatal("unexpected type %d: %s", type, type2name(type)); 2483 break; 2484 } 2485 } 2486 // The load node has the control of the preceding MemBarCPUOrder. All 2487 // following nodes will have the control of the MemBarCPUOrder inserted at 2488 // the end of this method. So, pushing the load onto the stack at a later 2489 // point is fine. 2490 set_result(p); 2491 } else { 2492 // place effect of store into memory 2493 switch (type) { 2494 case T_DOUBLE: 2495 val = dstore_rounding(val); 2496 break; 2497 case T_ADDRESS: 2498 // Repackage the long as a pointer. 2499 val = ConvL2X(val); 2500 val = _gvn.transform(new CastX2PNode(val)); 2501 break; 2502 } 2503 2504 if (type != T_OBJECT) { 2505 (void) store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched); 2506 } else { 2507 // Possibly an oop being stored to Java heap or native memory 2508 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2509 // oop to Java heap. 2510 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2511 } else { 2512 // We can't tell at compile time if we are storing in the Java heap or outside 2513 // of it. So we need to emit code to conditionally do the proper type of 2514 // store. 2515 2516 IdealKit ideal(this); 2517#define __ ideal. 2518 // QQQ who knows what probability is here?? 2519 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2520 // Sync IdealKit and graphKit. 2521 sync_kit(ideal); 2522 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2523 // Update IdealKit memory. 2524 __ sync_kit(this); 2525 } __ else_(); { 2526 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, requires_atomic_access, mismatched); 2527 } __ end_if(); 2528 // Final sync IdealKit and GraphKit. 2529 final_sync(ideal); 2530#undef __ 2531 } 2532 } 2533 } 2534 2535 switch(kind) { 2536 case Relaxed: 2537 case Opaque: 2538 case Release: 2539 break; 2540 case Acquire: 2541 case Volatile: 2542 if (!is_store) { 2543 insert_mem_bar(Op_MemBarAcquire); 2544 } else { 2545 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2546 insert_mem_bar(Op_MemBarVolatile); 2547 } 2548 } 2549 break; 2550 default: 2551 ShouldNotReachHere(); 2552 } 2553 2554 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2555 2556 return true; 2557} 2558 2559//----------------------------inline_unsafe_load_store---------------------------- 2560// This method serves a couple of different customers (depending on LoadStoreKind): 2561// 2562// LS_cmp_swap: 2563// 2564// boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2565// boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2566// boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2567// 2568// LS_cmp_swap_weak: 2569// 2570// boolean weakCompareAndSwapObject( Object o, long offset, Object expected, Object x); 2571// boolean weakCompareAndSwapObjectAcquire(Object o, long offset, Object expected, Object x); 2572// boolean weakCompareAndSwapObjectRelease(Object o, long offset, Object expected, Object x); 2573// 2574// boolean weakCompareAndSwapInt( Object o, long offset, int expected, int x); 2575// boolean weakCompareAndSwapIntAcquire( Object o, long offset, int expected, int x); 2576// boolean weakCompareAndSwapIntRelease( Object o, long offset, int expected, int x); 2577// 2578// boolean weakCompareAndSwapLong( Object o, long offset, long expected, long x); 2579// boolean weakCompareAndSwapLongAcquire( Object o, long offset, long expected, long x); 2580// boolean weakCompareAndSwapLongRelease( Object o, long offset, long expected, long x); 2581// 2582// LS_cmp_exchange: 2583// 2584// Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x); 2585// Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x); 2586// Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x); 2587// 2588// Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x); 2589// Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x); 2590// Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x); 2591// 2592// Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x); 2593// Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x); 2594// Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x); 2595// 2596// LS_get_add: 2597// 2598// int getAndAddInt( Object o, long offset, int delta) 2599// long getAndAddLong(Object o, long offset, long delta) 2600// 2601// LS_get_set: 2602// 2603// int getAndSet(Object o, long offset, int newValue) 2604// long getAndSet(Object o, long offset, long newValue) 2605// Object getAndSet(Object o, long offset, Object newValue) 2606// 2607bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) { 2608 // This basic scheme here is the same as inline_unsafe_access, but 2609 // differs in enough details that combining them would make the code 2610 // overly confusing. (This is a true fact! I originally combined 2611 // them, but even I was confused by it!) As much code/comments as 2612 // possible are retained from inline_unsafe_access though to make 2613 // the correspondences clearer. - dl 2614 2615 if (callee()->is_static()) return false; // caller must have the capability! 2616 2617#ifndef PRODUCT 2618 BasicType rtype; 2619 { 2620 ResourceMark rm; 2621 // Check the signatures. 2622 ciSignature* sig = callee()->signature(); 2623 rtype = sig->return_type()->basic_type(); 2624 switch(kind) { 2625 case LS_get_add: 2626 case LS_get_set: { 2627 // Check the signatures. 2628#ifdef ASSERT 2629 assert(rtype == type, "get and set must return the expected type"); 2630 assert(sig->count() == 3, "get and set has 3 arguments"); 2631 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2632 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2633 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2634 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation"); 2635#endif // ASSERT 2636 break; 2637 } 2638 case LS_cmp_swap: 2639 case LS_cmp_swap_weak: { 2640 // Check the signatures. 2641#ifdef ASSERT 2642 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2643 assert(sig->count() == 4, "CAS has 4 arguments"); 2644 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2645 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2646#endif // ASSERT 2647 break; 2648 } 2649 case LS_cmp_exchange: { 2650 // Check the signatures. 2651#ifdef ASSERT 2652 assert(rtype == type, "CAS must return the expected type"); 2653 assert(sig->count() == 4, "CAS has 4 arguments"); 2654 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2655 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2656#endif // ASSERT 2657 break; 2658 } 2659 default: 2660 ShouldNotReachHere(); 2661 } 2662 } 2663#endif //PRODUCT 2664 2665 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2666 2667 // Get arguments: 2668 Node* receiver = NULL; 2669 Node* base = NULL; 2670 Node* offset = NULL; 2671 Node* oldval = NULL; 2672 Node* newval = NULL; 2673 switch(kind) { 2674 case LS_cmp_swap: 2675 case LS_cmp_swap_weak: 2676 case LS_cmp_exchange: { 2677 const bool two_slot_type = type2size[type] == 2; 2678 receiver = argument(0); // type: oop 2679 base = argument(1); // type: oop 2680 offset = argument(2); // type: long 2681 oldval = argument(4); // type: oop, int, or long 2682 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2683 break; 2684 } 2685 case LS_get_add: 2686 case LS_get_set: { 2687 receiver = argument(0); // type: oop 2688 base = argument(1); // type: oop 2689 offset = argument(2); // type: long 2690 oldval = NULL; 2691 newval = argument(4); // type: oop, int, or long 2692 break; 2693 } 2694 default: 2695 ShouldNotReachHere(); 2696 } 2697 2698 // Null check receiver. 2699 receiver = null_check(receiver); 2700 if (stopped()) { 2701 return true; 2702 } 2703 2704 // Build field offset expression. 2705 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2706 // to be plain byte offsets, which are also the same as those accepted 2707 // by oopDesc::field_base. 2708 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2709 // 32-bit machines ignore the high half of long offsets 2710 offset = ConvL2X(offset); 2711 Node* adr = make_unsafe_address(base, offset); 2712 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2713 2714 Compile::AliasType* alias_type = C->alias_type(adr_type); 2715 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM || 2716 alias_type->basic_type() != T_ILLEGAL, "field, array element or unknown"); 2717 BasicType bt = alias_type->basic_type(); 2718 if (bt != T_ILLEGAL && 2719 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2720 // Don't intrinsify mismatched object accesses. 2721 return false; 2722 } 2723 2724 // For CAS, unlike inline_unsafe_access, there seems no point in 2725 // trying to refine types. Just use the coarse types here. 2726 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2727 const Type *value_type = Type::get_const_basic_type(type); 2728 2729 switch (kind) { 2730 case LS_get_set: 2731 case LS_cmp_exchange: { 2732 if (type == T_OBJECT) { 2733 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2734 if (tjp != NULL) { 2735 value_type = tjp; 2736 } 2737 } 2738 break; 2739 } 2740 case LS_cmp_swap: 2741 case LS_cmp_swap_weak: 2742 case LS_get_add: 2743 break; 2744 default: 2745 ShouldNotReachHere(); 2746 } 2747 2748 int alias_idx = C->get_alias_index(adr_type); 2749 2750 // Memory-model-wise, a LoadStore acts like a little synchronized 2751 // block, so needs barriers on each side. These don't translate 2752 // into actual barriers on most machines, but we still need rest of 2753 // compiler to respect ordering. 2754 2755 switch (access_kind) { 2756 case Relaxed: 2757 case Acquire: 2758 break; 2759 case Release: 2760 insert_mem_bar(Op_MemBarRelease); 2761 break; 2762 case Volatile: 2763 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2764 insert_mem_bar(Op_MemBarVolatile); 2765 } else { 2766 insert_mem_bar(Op_MemBarRelease); 2767 } 2768 break; 2769 default: 2770 ShouldNotReachHere(); 2771 } 2772 insert_mem_bar(Op_MemBarCPUOrder); 2773 2774 // Figure out the memory ordering. 2775 MemNode::MemOrd mo = access_kind_to_memord(access_kind); 2776 2777 // 4984716: MemBars must be inserted before this 2778 // memory node in order to avoid a false 2779 // dependency which will confuse the scheduler. 2780 Node *mem = memory(alias_idx); 2781 2782 // For now, we handle only those cases that actually exist: ints, 2783 // longs, and Object. Adding others should be straightforward. 2784 Node* load_store = NULL; 2785 switch(type) { 2786 case T_BYTE: 2787 switch(kind) { 2788 case LS_get_add: 2789 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type)); 2790 break; 2791 case LS_get_set: 2792 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type)); 2793 break; 2794 case LS_cmp_swap_weak: 2795 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo)); 2796 break; 2797 case LS_cmp_swap: 2798 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo)); 2799 break; 2800 case LS_cmp_exchange: 2801 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo)); 2802 break; 2803 default: 2804 ShouldNotReachHere(); 2805 } 2806 break; 2807 case T_SHORT: 2808 switch(kind) { 2809 case LS_get_add: 2810 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type)); 2811 break; 2812 case LS_get_set: 2813 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type)); 2814 break; 2815 case LS_cmp_swap_weak: 2816 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo)); 2817 break; 2818 case LS_cmp_swap: 2819 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo)); 2820 break; 2821 case LS_cmp_exchange: 2822 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo)); 2823 break; 2824 default: 2825 ShouldNotReachHere(); 2826 } 2827 break; 2828 case T_INT: 2829 switch(kind) { 2830 case LS_get_add: 2831 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2832 break; 2833 case LS_get_set: 2834 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2835 break; 2836 case LS_cmp_swap_weak: 2837 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo)); 2838 break; 2839 case LS_cmp_swap: 2840 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo)); 2841 break; 2842 case LS_cmp_exchange: 2843 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo)); 2844 break; 2845 default: 2846 ShouldNotReachHere(); 2847 } 2848 break; 2849 case T_LONG: 2850 switch(kind) { 2851 case LS_get_add: 2852 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2853 break; 2854 case LS_get_set: 2855 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2856 break; 2857 case LS_cmp_swap_weak: 2858 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo)); 2859 break; 2860 case LS_cmp_swap: 2861 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo)); 2862 break; 2863 case LS_cmp_exchange: 2864 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo)); 2865 break; 2866 default: 2867 ShouldNotReachHere(); 2868 } 2869 break; 2870 case T_OBJECT: 2871 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2872 // could be delayed during Parse (for example, in adjust_map_after_if()). 2873 // Execute transformation here to avoid barrier generation in such case. 2874 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2875 newval = _gvn.makecon(TypePtr::NULL_PTR); 2876 2877 // Reference stores need a store barrier. 2878 switch(kind) { 2879 case LS_get_set: { 2880 // If pre-barrier must execute before the oop store, old value will require do_load here. 2881 if (!can_move_pre_barrier()) { 2882 pre_barrier(true /* do_load*/, 2883 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2884 NULL /* pre_val*/, 2885 T_OBJECT); 2886 } // Else move pre_barrier to use load_store value, see below. 2887 break; 2888 } 2889 case LS_cmp_swap_weak: 2890 case LS_cmp_swap: 2891 case LS_cmp_exchange: { 2892 // Same as for newval above: 2893 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2894 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2895 } 2896 // The only known value which might get overwritten is oldval. 2897 pre_barrier(false /* do_load */, 2898 control(), NULL, NULL, max_juint, NULL, NULL, 2899 oldval /* pre_val */, 2900 T_OBJECT); 2901 break; 2902 } 2903 default: 2904 ShouldNotReachHere(); 2905 } 2906 2907#ifdef _LP64 2908 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2909 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2910 2911 switch(kind) { 2912 case LS_get_set: 2913 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop())); 2914 break; 2915 case LS_cmp_swap_weak: { 2916 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2917 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo)); 2918 break; 2919 } 2920 case LS_cmp_swap: { 2921 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2922 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo)); 2923 break; 2924 } 2925 case LS_cmp_exchange: { 2926 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2927 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo)); 2928 break; 2929 } 2930 default: 2931 ShouldNotReachHere(); 2932 } 2933 } else 2934#endif 2935 switch (kind) { 2936 case LS_get_set: 2937 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2938 break; 2939 case LS_cmp_swap_weak: 2940 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo)); 2941 break; 2942 case LS_cmp_swap: 2943 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo)); 2944 break; 2945 case LS_cmp_exchange: 2946 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo)); 2947 break; 2948 default: 2949 ShouldNotReachHere(); 2950 } 2951 2952 // Emit the post barrier only when the actual store happened. This makes sense 2953 // to check only for LS_cmp_* that can fail to set the value. 2954 // LS_cmp_exchange does not produce any branches by default, so there is no 2955 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with 2956 // CompareAndExchange and move branches here, it would make sense to conditionalize 2957 // post_barriers for LS_cmp_exchange as well. 2958 // 2959 // CAS success path is marked more likely since we anticipate this is a performance 2960 // critical path, while CAS failure path can use the penalty for going through unlikely 2961 // path as backoff. Which is still better than doing a store barrier there. 2962 switch (kind) { 2963 case LS_get_set: 2964 case LS_cmp_exchange: { 2965 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2966 break; 2967 } 2968 case LS_cmp_swap_weak: 2969 case LS_cmp_swap: { 2970 IdealKit ideal(this); 2971 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); { 2972 sync_kit(ideal); 2973 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2974 ideal.sync_kit(this); 2975 } ideal.end_if(); 2976 final_sync(ideal); 2977 break; 2978 } 2979 default: 2980 ShouldNotReachHere(); 2981 } 2982 break; 2983 default: 2984 fatal("unexpected type %d: %s", type, type2name(type)); 2985 break; 2986 } 2987 2988 // SCMemProjNodes represent the memory state of a LoadStore. Their 2989 // main role is to prevent LoadStore nodes from being optimized away 2990 // when their results aren't used. 2991 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2992 set_memory(proj, alias_idx); 2993 2994 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) { 2995#ifdef _LP64 2996 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2997 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2998 } 2999#endif 3000 if (can_move_pre_barrier()) { 3001 // Don't need to load pre_val. The old value is returned by load_store. 3002 // The pre_barrier can execute after the xchg as long as no safepoint 3003 // gets inserted between them. 3004 pre_barrier(false /* do_load */, 3005 control(), NULL, NULL, max_juint, NULL, NULL, 3006 load_store /* pre_val */, 3007 T_OBJECT); 3008 } 3009 } 3010 3011 // Add the trailing membar surrounding the access 3012 insert_mem_bar(Op_MemBarCPUOrder); 3013 3014 switch (access_kind) { 3015 case Relaxed: 3016 case Release: 3017 break; // do nothing 3018 case Acquire: 3019 case Volatile: 3020 insert_mem_bar(Op_MemBarAcquire); 3021 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code 3022 break; 3023 default: 3024 ShouldNotReachHere(); 3025 } 3026 3027 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3028 set_result(load_store); 3029 return true; 3030} 3031 3032MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) { 3033 MemNode::MemOrd mo = MemNode::unset; 3034 switch(kind) { 3035 case Opaque: 3036 case Relaxed: mo = MemNode::unordered; break; 3037 case Acquire: mo = MemNode::acquire; break; 3038 case Release: mo = MemNode::release; break; 3039 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break; 3040 default: 3041 ShouldNotReachHere(); 3042 } 3043 guarantee(mo != MemNode::unset, "Should select memory ordering"); 3044 return mo; 3045} 3046 3047MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) { 3048 MemNode::MemOrd mo = MemNode::unset; 3049 switch(kind) { 3050 case Opaque: 3051 case Relaxed: mo = MemNode::unordered; break; 3052 case Acquire: mo = MemNode::acquire; break; 3053 case Release: mo = MemNode::release; break; 3054 case Volatile: mo = MemNode::seqcst; break; 3055 default: 3056 ShouldNotReachHere(); 3057 } 3058 guarantee(mo != MemNode::unset, "Should select memory ordering"); 3059 return mo; 3060} 3061 3062bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3063 // Regardless of form, don't allow previous ld/st to move down, 3064 // then issue acquire, release, or volatile mem_bar. 3065 insert_mem_bar(Op_MemBarCPUOrder); 3066 switch(id) { 3067 case vmIntrinsics::_loadFence: 3068 insert_mem_bar(Op_LoadFence); 3069 return true; 3070 case vmIntrinsics::_storeFence: 3071 insert_mem_bar(Op_StoreFence); 3072 return true; 3073 case vmIntrinsics::_fullFence: 3074 insert_mem_bar(Op_MemBarVolatile); 3075 return true; 3076 default: 3077 fatal_unexpected_iid(id); 3078 return false; 3079 } 3080} 3081 3082bool LibraryCallKit::inline_onspinwait() { 3083 insert_mem_bar(Op_OnSpinWait); 3084 return true; 3085} 3086 3087bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3088 if (!kls->is_Con()) { 3089 return true; 3090 } 3091 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3092 if (klsptr == NULL) { 3093 return true; 3094 } 3095 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3096 // don't need a guard for a klass that is already initialized 3097 return !ik->is_initialized(); 3098} 3099 3100//----------------------------inline_unsafe_allocate--------------------------- 3101// public native Object Unsafe.allocateInstance(Class<?> cls); 3102bool LibraryCallKit::inline_unsafe_allocate() { 3103 if (callee()->is_static()) return false; // caller must have the capability! 3104 3105 null_check_receiver(); // null-check, then ignore 3106 Node* cls = null_check(argument(1)); 3107 if (stopped()) return true; 3108 3109 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3110 kls = null_check(kls); 3111 if (stopped()) return true; // argument was like int.class 3112 3113 Node* test = NULL; 3114 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3115 // Note: The argument might still be an illegal value like 3116 // Serializable.class or Object[].class. The runtime will handle it. 3117 // But we must make an explicit check for initialization. 3118 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3119 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3120 // can generate code to load it as unsigned byte. 3121 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3122 Node* bits = intcon(InstanceKlass::fully_initialized); 3123 test = _gvn.transform(new SubINode(inst, bits)); 3124 // The 'test' is non-zero if we need to take a slow path. 3125 } 3126 3127 Node* obj = new_instance(kls, test); 3128 set_result(obj); 3129 return true; 3130} 3131 3132//------------------------inline_native_time_funcs-------------- 3133// inline code for System.currentTimeMillis() and System.nanoTime() 3134// these have the same type and signature 3135bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3136 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3137 const TypePtr* no_memory_effects = NULL; 3138 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3139 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3140#ifdef ASSERT 3141 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3142 assert(value_top == top(), "second value must be top"); 3143#endif 3144 set_result(value); 3145 return true; 3146} 3147 3148//------------------------inline_native_currentThread------------------ 3149bool LibraryCallKit::inline_native_currentThread() { 3150 Node* junk = NULL; 3151 set_result(generate_current_thread(junk)); 3152 return true; 3153} 3154 3155//------------------------inline_native_isInterrupted------------------ 3156// private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3157bool LibraryCallKit::inline_native_isInterrupted() { 3158 // Add a fast path to t.isInterrupted(clear_int): 3159 // (t == Thread.current() && 3160 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3161 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3162 // So, in the common case that the interrupt bit is false, 3163 // we avoid making a call into the VM. Even if the interrupt bit 3164 // is true, if the clear_int argument is false, we avoid the VM call. 3165 // However, if the receiver is not currentThread, we must call the VM, 3166 // because there must be some locking done around the operation. 3167 3168 // We only go to the fast case code if we pass two guards. 3169 // Paths which do not pass are accumulated in the slow_region. 3170 3171 enum { 3172 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3173 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3174 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3175 PATH_LIMIT 3176 }; 3177 3178 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3179 // out of the function. 3180 insert_mem_bar(Op_MemBarCPUOrder); 3181 3182 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3183 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3184 3185 RegionNode* slow_region = new RegionNode(1); 3186 record_for_igvn(slow_region); 3187 3188 // (a) Receiving thread must be the current thread. 3189 Node* rec_thr = argument(0); 3190 Node* tls_ptr = NULL; 3191 Node* cur_thr = generate_current_thread(tls_ptr); 3192 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3193 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3194 3195 generate_slow_guard(bol_thr, slow_region); 3196 3197 // (b) Interrupt bit on TLS must be false. 3198 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3199 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3200 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3201 3202 // Set the control input on the field _interrupted read to prevent it floating up. 3203 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3204 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3205 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3206 3207 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3208 3209 // First fast path: if (!TLS._interrupted) return false; 3210 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3211 result_rgn->init_req(no_int_result_path, false_bit); 3212 result_val->init_req(no_int_result_path, intcon(0)); 3213 3214 // drop through to next case 3215 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3216 3217#ifndef TARGET_OS_FAMILY_windows 3218 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3219 Node* clr_arg = argument(1); 3220 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3221 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3222 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3223 3224 // Second fast path: ... else if (!clear_int) return true; 3225 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3226 result_rgn->init_req(no_clear_result_path, false_arg); 3227 result_val->init_req(no_clear_result_path, intcon(1)); 3228 3229 // drop through to next case 3230 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3231#else 3232 // To return true on Windows you must read the _interrupted field 3233 // and check the event state i.e. take the slow path. 3234#endif // TARGET_OS_FAMILY_windows 3235 3236 // (d) Otherwise, go to the slow path. 3237 slow_region->add_req(control()); 3238 set_control( _gvn.transform(slow_region)); 3239 3240 if (stopped()) { 3241 // There is no slow path. 3242 result_rgn->init_req(slow_result_path, top()); 3243 result_val->init_req(slow_result_path, top()); 3244 } else { 3245 // non-virtual because it is a private non-static 3246 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3247 3248 Node* slow_val = set_results_for_java_call(slow_call); 3249 // this->control() comes from set_results_for_java_call 3250 3251 Node* fast_io = slow_call->in(TypeFunc::I_O); 3252 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3253 3254 // These two phis are pre-filled with copies of of the fast IO and Memory 3255 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3256 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3257 3258 result_rgn->init_req(slow_result_path, control()); 3259 result_io ->init_req(slow_result_path, i_o()); 3260 result_mem->init_req(slow_result_path, reset_memory()); 3261 result_val->init_req(slow_result_path, slow_val); 3262 3263 set_all_memory(_gvn.transform(result_mem)); 3264 set_i_o( _gvn.transform(result_io)); 3265 } 3266 3267 C->set_has_split_ifs(true); // Has chance for split-if optimization 3268 set_result(result_rgn, result_val); 3269 return true; 3270} 3271 3272//---------------------------load_mirror_from_klass---------------------------- 3273// Given a klass oop, load its java mirror (a java.lang.Class oop). 3274Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3275 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3276 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3277} 3278 3279//-----------------------load_klass_from_mirror_common------------------------- 3280// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3281// Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3282// and branch to the given path on the region. 3283// If never_see_null, take an uncommon trap on null, so we can optimistically 3284// compile for the non-null case. 3285// If the region is NULL, force never_see_null = true. 3286Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3287 bool never_see_null, 3288 RegionNode* region, 3289 int null_path, 3290 int offset) { 3291 if (region == NULL) never_see_null = true; 3292 Node* p = basic_plus_adr(mirror, offset); 3293 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3294 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3295 Node* null_ctl = top(); 3296 kls = null_check_oop(kls, &null_ctl, never_see_null); 3297 if (region != NULL) { 3298 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3299 region->init_req(null_path, null_ctl); 3300 } else { 3301 assert(null_ctl == top(), "no loose ends"); 3302 } 3303 return kls; 3304} 3305 3306//--------------------(inline_native_Class_query helpers)--------------------- 3307// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER. 3308// Fall through if (mods & mask) == bits, take the guard otherwise. 3309Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3310 // Branch around if the given klass has the given modifier bit set. 3311 // Like generate_guard, adds a new path onto the region. 3312 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3313 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3314 Node* mask = intcon(modifier_mask); 3315 Node* bits = intcon(modifier_bits); 3316 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3317 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3318 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3319 return generate_fair_guard(bol, region); 3320} 3321Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3322 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3323} 3324 3325//-------------------------inline_native_Class_query------------------- 3326bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3327 const Type* return_type = TypeInt::BOOL; 3328 Node* prim_return_value = top(); // what happens if it's a primitive class? 3329 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3330 bool expect_prim = false; // most of these guys expect to work on refs 3331 3332 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3333 3334 Node* mirror = argument(0); 3335 Node* obj = top(); 3336 3337 switch (id) { 3338 case vmIntrinsics::_isInstance: 3339 // nothing is an instance of a primitive type 3340 prim_return_value = intcon(0); 3341 obj = argument(1); 3342 break; 3343 case vmIntrinsics::_getModifiers: 3344 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3345 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3346 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3347 break; 3348 case vmIntrinsics::_isInterface: 3349 prim_return_value = intcon(0); 3350 break; 3351 case vmIntrinsics::_isArray: 3352 prim_return_value = intcon(0); 3353 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3354 break; 3355 case vmIntrinsics::_isPrimitive: 3356 prim_return_value = intcon(1); 3357 expect_prim = true; // obviously 3358 break; 3359 case vmIntrinsics::_getSuperclass: 3360 prim_return_value = null(); 3361 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3362 break; 3363 case vmIntrinsics::_getClassAccessFlags: 3364 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3365 return_type = TypeInt::INT; // not bool! 6297094 3366 break; 3367 default: 3368 fatal_unexpected_iid(id); 3369 break; 3370 } 3371 3372 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3373 if (mirror_con == NULL) return false; // cannot happen? 3374 3375#ifndef PRODUCT 3376 if (C->print_intrinsics() || C->print_inlining()) { 3377 ciType* k = mirror_con->java_mirror_type(); 3378 if (k) { 3379 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3380 k->print_name(); 3381 tty->cr(); 3382 } 3383 } 3384#endif 3385 3386 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3387 RegionNode* region = new RegionNode(PATH_LIMIT); 3388 record_for_igvn(region); 3389 PhiNode* phi = new PhiNode(region, return_type); 3390 3391 // The mirror will never be null of Reflection.getClassAccessFlags, however 3392 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3393 // if it is. See bug 4774291. 3394 3395 // For Reflection.getClassAccessFlags(), the null check occurs in 3396 // the wrong place; see inline_unsafe_access(), above, for a similar 3397 // situation. 3398 mirror = null_check(mirror); 3399 // If mirror or obj is dead, only null-path is taken. 3400 if (stopped()) return true; 3401 3402 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3403 3404 // Now load the mirror's klass metaobject, and null-check it. 3405 // Side-effects region with the control path if the klass is null. 3406 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3407 // If kls is null, we have a primitive mirror. 3408 phi->init_req(_prim_path, prim_return_value); 3409 if (stopped()) { set_result(region, phi); return true; } 3410 bool safe_for_replace = (region->in(_prim_path) == top()); 3411 3412 Node* p; // handy temp 3413 Node* null_ctl; 3414 3415 // Now that we have the non-null klass, we can perform the real query. 3416 // For constant classes, the query will constant-fold in LoadNode::Value. 3417 Node* query_value = top(); 3418 switch (id) { 3419 case vmIntrinsics::_isInstance: 3420 // nothing is an instance of a primitive type 3421 query_value = gen_instanceof(obj, kls, safe_for_replace); 3422 break; 3423 3424 case vmIntrinsics::_getModifiers: 3425 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3426 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3427 break; 3428 3429 case vmIntrinsics::_isInterface: 3430 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3431 if (generate_interface_guard(kls, region) != NULL) 3432 // A guard was added. If the guard is taken, it was an interface. 3433 phi->add_req(intcon(1)); 3434 // If we fall through, it's a plain class. 3435 query_value = intcon(0); 3436 break; 3437 3438 case vmIntrinsics::_isArray: 3439 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3440 if (generate_array_guard(kls, region) != NULL) 3441 // A guard was added. If the guard is taken, it was an array. 3442 phi->add_req(intcon(1)); 3443 // If we fall through, it's a plain class. 3444 query_value = intcon(0); 3445 break; 3446 3447 case vmIntrinsics::_isPrimitive: 3448 query_value = intcon(0); // "normal" path produces false 3449 break; 3450 3451 case vmIntrinsics::_getSuperclass: 3452 // The rules here are somewhat unfortunate, but we can still do better 3453 // with random logic than with a JNI call. 3454 // Interfaces store null or Object as _super, but must report null. 3455 // Arrays store an intermediate super as _super, but must report Object. 3456 // Other types can report the actual _super. 3457 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3458 if (generate_interface_guard(kls, region) != NULL) 3459 // A guard was added. If the guard is taken, it was an interface. 3460 phi->add_req(null()); 3461 if (generate_array_guard(kls, region) != NULL) 3462 // A guard was added. If the guard is taken, it was an array. 3463 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3464 // If we fall through, it's a plain class. Get its _super. 3465 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3466 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3467 null_ctl = top(); 3468 kls = null_check_oop(kls, &null_ctl); 3469 if (null_ctl != top()) { 3470 // If the guard is taken, Object.superClass is null (both klass and mirror). 3471 region->add_req(null_ctl); 3472 phi ->add_req(null()); 3473 } 3474 if (!stopped()) { 3475 query_value = load_mirror_from_klass(kls); 3476 } 3477 break; 3478 3479 case vmIntrinsics::_getClassAccessFlags: 3480 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3481 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3482 break; 3483 3484 default: 3485 fatal_unexpected_iid(id); 3486 break; 3487 } 3488 3489 // Fall-through is the normal case of a query to a real class. 3490 phi->init_req(1, query_value); 3491 region->init_req(1, control()); 3492 3493 C->set_has_split_ifs(true); // Has chance for split-if optimization 3494 set_result(region, phi); 3495 return true; 3496} 3497 3498//-------------------------inline_Class_cast------------------- 3499bool LibraryCallKit::inline_Class_cast() { 3500 Node* mirror = argument(0); // Class 3501 Node* obj = argument(1); 3502 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3503 if (mirror_con == NULL) { 3504 return false; // dead path (mirror->is_top()). 3505 } 3506 if (obj == NULL || obj->is_top()) { 3507 return false; // dead path 3508 } 3509 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3510 3511 // First, see if Class.cast() can be folded statically. 3512 // java_mirror_type() returns non-null for compile-time Class constants. 3513 ciType* tm = mirror_con->java_mirror_type(); 3514 if (tm != NULL && tm->is_klass() && 3515 tp != NULL && tp->klass() != NULL) { 3516 if (!tp->klass()->is_loaded()) { 3517 // Don't use intrinsic when class is not loaded. 3518 return false; 3519 } else { 3520 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3521 if (static_res == Compile::SSC_always_true) { 3522 // isInstance() is true - fold the code. 3523 set_result(obj); 3524 return true; 3525 } else if (static_res == Compile::SSC_always_false) { 3526 // Don't use intrinsic, have to throw ClassCastException. 3527 // If the reference is null, the non-intrinsic bytecode will 3528 // be optimized appropriately. 3529 return false; 3530 } 3531 } 3532 } 3533 3534 // Bailout intrinsic and do normal inlining if exception path is frequent. 3535 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3536 return false; 3537 } 3538 3539 // Generate dynamic checks. 3540 // Class.cast() is java implementation of _checkcast bytecode. 3541 // Do checkcast (Parse::do_checkcast()) optimizations here. 3542 3543 mirror = null_check(mirror); 3544 // If mirror is dead, only null-path is taken. 3545 if (stopped()) { 3546 return true; 3547 } 3548 3549 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3550 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3551 RegionNode* region = new RegionNode(PATH_LIMIT); 3552 record_for_igvn(region); 3553 3554 // Now load the mirror's klass metaobject, and null-check it. 3555 // If kls is null, we have a primitive mirror and 3556 // nothing is an instance of a primitive type. 3557 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3558 3559 Node* res = top(); 3560 if (!stopped()) { 3561 Node* bad_type_ctrl = top(); 3562 // Do checkcast optimizations. 3563 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3564 region->init_req(_bad_type_path, bad_type_ctrl); 3565 } 3566 if (region->in(_prim_path) != top() || 3567 region->in(_bad_type_path) != top()) { 3568 // Let Interpreter throw ClassCastException. 3569 PreserveJVMState pjvms(this); 3570 set_control(_gvn.transform(region)); 3571 uncommon_trap(Deoptimization::Reason_intrinsic, 3572 Deoptimization::Action_maybe_recompile); 3573 } 3574 if (!stopped()) { 3575 set_result(res); 3576 } 3577 return true; 3578} 3579 3580 3581//--------------------------inline_native_subtype_check------------------------ 3582// This intrinsic takes the JNI calls out of the heart of 3583// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3584bool LibraryCallKit::inline_native_subtype_check() { 3585 // Pull both arguments off the stack. 3586 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3587 args[0] = argument(0); 3588 args[1] = argument(1); 3589 Node* klasses[2]; // corresponding Klasses: superk, subk 3590 klasses[0] = klasses[1] = top(); 3591 3592 enum { 3593 // A full decision tree on {superc is prim, subc is prim}: 3594 _prim_0_path = 1, // {P,N} => false 3595 // {P,P} & superc!=subc => false 3596 _prim_same_path, // {P,P} & superc==subc => true 3597 _prim_1_path, // {N,P} => false 3598 _ref_subtype_path, // {N,N} & subtype check wins => true 3599 _both_ref_path, // {N,N} & subtype check loses => false 3600 PATH_LIMIT 3601 }; 3602 3603 RegionNode* region = new RegionNode(PATH_LIMIT); 3604 Node* phi = new PhiNode(region, TypeInt::BOOL); 3605 record_for_igvn(region); 3606 3607 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3608 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3609 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3610 3611 // First null-check both mirrors and load each mirror's klass metaobject. 3612 int which_arg; 3613 for (which_arg = 0; which_arg <= 1; which_arg++) { 3614 Node* arg = args[which_arg]; 3615 arg = null_check(arg); 3616 if (stopped()) break; 3617 args[which_arg] = arg; 3618 3619 Node* p = basic_plus_adr(arg, class_klass_offset); 3620 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3621 klasses[which_arg] = _gvn.transform(kls); 3622 } 3623 3624 // Having loaded both klasses, test each for null. 3625 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3626 for (which_arg = 0; which_arg <= 1; which_arg++) { 3627 Node* kls = klasses[which_arg]; 3628 Node* null_ctl = top(); 3629 kls = null_check_oop(kls, &null_ctl, never_see_null); 3630 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3631 region->init_req(prim_path, null_ctl); 3632 if (stopped()) break; 3633 klasses[which_arg] = kls; 3634 } 3635 3636 if (!stopped()) { 3637 // now we have two reference types, in klasses[0..1] 3638 Node* subk = klasses[1]; // the argument to isAssignableFrom 3639 Node* superk = klasses[0]; // the receiver 3640 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3641 // now we have a successful reference subtype check 3642 region->set_req(_ref_subtype_path, control()); 3643 } 3644 3645 // If both operands are primitive (both klasses null), then 3646 // we must return true when they are identical primitives. 3647 // It is convenient to test this after the first null klass check. 3648 set_control(region->in(_prim_0_path)); // go back to first null check 3649 if (!stopped()) { 3650 // Since superc is primitive, make a guard for the superc==subc case. 3651 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3652 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3653 generate_guard(bol_eq, region, PROB_FAIR); 3654 if (region->req() == PATH_LIMIT+1) { 3655 // A guard was added. If the added guard is taken, superc==subc. 3656 region->swap_edges(PATH_LIMIT, _prim_same_path); 3657 region->del_req(PATH_LIMIT); 3658 } 3659 region->set_req(_prim_0_path, control()); // Not equal after all. 3660 } 3661 3662 // these are the only paths that produce 'true': 3663 phi->set_req(_prim_same_path, intcon(1)); 3664 phi->set_req(_ref_subtype_path, intcon(1)); 3665 3666 // pull together the cases: 3667 assert(region->req() == PATH_LIMIT, "sane region"); 3668 for (uint i = 1; i < region->req(); i++) { 3669 Node* ctl = region->in(i); 3670 if (ctl == NULL || ctl == top()) { 3671 region->set_req(i, top()); 3672 phi ->set_req(i, top()); 3673 } else if (phi->in(i) == NULL) { 3674 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3675 } 3676 } 3677 3678 set_control(_gvn.transform(region)); 3679 set_result(_gvn.transform(phi)); 3680 return true; 3681} 3682 3683//---------------------generate_array_guard_common------------------------ 3684Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3685 bool obj_array, bool not_array) { 3686 3687 if (stopped()) { 3688 return NULL; 3689 } 3690 3691 // If obj_array/non_array==false/false: 3692 // Branch around if the given klass is in fact an array (either obj or prim). 3693 // If obj_array/non_array==false/true: 3694 // Branch around if the given klass is not an array klass of any kind. 3695 // If obj_array/non_array==true/true: 3696 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3697 // If obj_array/non_array==true/false: 3698 // Branch around if the kls is an oop array (Object[] or subtype) 3699 // 3700 // Like generate_guard, adds a new path onto the region. 3701 jint layout_con = 0; 3702 Node* layout_val = get_layout_helper(kls, layout_con); 3703 if (layout_val == NULL) { 3704 bool query = (obj_array 3705 ? Klass::layout_helper_is_objArray(layout_con) 3706 : Klass::layout_helper_is_array(layout_con)); 3707 if (query == not_array) { 3708 return NULL; // never a branch 3709 } else { // always a branch 3710 Node* always_branch = control(); 3711 if (region != NULL) 3712 region->add_req(always_branch); 3713 set_control(top()); 3714 return always_branch; 3715 } 3716 } 3717 // Now test the correct condition. 3718 jint nval = (obj_array 3719 ? (jint)(Klass::_lh_array_tag_type_value 3720 << Klass::_lh_array_tag_shift) 3721 : Klass::_lh_neutral_value); 3722 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3723 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3724 // invert the test if we are looking for a non-array 3725 if (not_array) btest = BoolTest(btest).negate(); 3726 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3727 return generate_fair_guard(bol, region); 3728} 3729 3730 3731//-----------------------inline_native_newArray-------------------------- 3732// private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3733// private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size); 3734bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) { 3735 Node* mirror; 3736 Node* count_val; 3737 if (uninitialized) { 3738 mirror = argument(1); 3739 count_val = argument(2); 3740 } else { 3741 mirror = argument(0); 3742 count_val = argument(1); 3743 } 3744 3745 mirror = null_check(mirror); 3746 // If mirror or obj is dead, only null-path is taken. 3747 if (stopped()) return true; 3748 3749 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3750 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3751 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3752 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3753 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3754 3755 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3756 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3757 result_reg, _slow_path); 3758 Node* normal_ctl = control(); 3759 Node* no_array_ctl = result_reg->in(_slow_path); 3760 3761 // Generate code for the slow case. We make a call to newArray(). 3762 set_control(no_array_ctl); 3763 if (!stopped()) { 3764 // Either the input type is void.class, or else the 3765 // array klass has not yet been cached. Either the 3766 // ensuing call will throw an exception, or else it 3767 // will cache the array klass for next time. 3768 PreserveJVMState pjvms(this); 3769 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3770 Node* slow_result = set_results_for_java_call(slow_call); 3771 // this->control() comes from set_results_for_java_call 3772 result_reg->set_req(_slow_path, control()); 3773 result_val->set_req(_slow_path, slow_result); 3774 result_io ->set_req(_slow_path, i_o()); 3775 result_mem->set_req(_slow_path, reset_memory()); 3776 } 3777 3778 set_control(normal_ctl); 3779 if (!stopped()) { 3780 // Normal case: The array type has been cached in the java.lang.Class. 3781 // The following call works fine even if the array type is polymorphic. 3782 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3783 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3784 result_reg->init_req(_normal_path, control()); 3785 result_val->init_req(_normal_path, obj); 3786 result_io ->init_req(_normal_path, i_o()); 3787 result_mem->init_req(_normal_path, reset_memory()); 3788 3789 if (uninitialized) { 3790 // Mark the allocation so that zeroing is skipped 3791 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn); 3792 alloc->maybe_set_complete(&_gvn); 3793 } 3794 } 3795 3796 // Return the combined state. 3797 set_i_o( _gvn.transform(result_io) ); 3798 set_all_memory( _gvn.transform(result_mem)); 3799 3800 C->set_has_split_ifs(true); // Has chance for split-if optimization 3801 set_result(result_reg, result_val); 3802 return true; 3803} 3804 3805//----------------------inline_native_getLength-------------------------- 3806// public static native int java.lang.reflect.Array.getLength(Object array); 3807bool LibraryCallKit::inline_native_getLength() { 3808 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3809 3810 Node* array = null_check(argument(0)); 3811 // If array is dead, only null-path is taken. 3812 if (stopped()) return true; 3813 3814 // Deoptimize if it is a non-array. 3815 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3816 3817 if (non_array != NULL) { 3818 PreserveJVMState pjvms(this); 3819 set_control(non_array); 3820 uncommon_trap(Deoptimization::Reason_intrinsic, 3821 Deoptimization::Action_maybe_recompile); 3822 } 3823 3824 // If control is dead, only non-array-path is taken. 3825 if (stopped()) return true; 3826 3827 // The works fine even if the array type is polymorphic. 3828 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3829 Node* result = load_array_length(array); 3830 3831 C->set_has_split_ifs(true); // Has chance for split-if optimization 3832 set_result(result); 3833 return true; 3834} 3835 3836//------------------------inline_array_copyOf---------------------------- 3837// public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3838// public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3839bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3840 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3841 3842 // Get the arguments. 3843 Node* original = argument(0); 3844 Node* start = is_copyOfRange? argument(1): intcon(0); 3845 Node* end = is_copyOfRange? argument(2): argument(1); 3846 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3847 3848 Node* newcopy = NULL; 3849 3850 // Set the original stack and the reexecute bit for the interpreter to reexecute 3851 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3852 { PreserveReexecuteState preexecs(this); 3853 jvms()->set_should_reexecute(true); 3854 3855 array_type_mirror = null_check(array_type_mirror); 3856 original = null_check(original); 3857 3858 // Check if a null path was taken unconditionally. 3859 if (stopped()) return true; 3860 3861 Node* orig_length = load_array_length(original); 3862 3863 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3864 klass_node = null_check(klass_node); 3865 3866 RegionNode* bailout = new RegionNode(1); 3867 record_for_igvn(bailout); 3868 3869 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3870 // Bail out if that is so. 3871 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3872 if (not_objArray != NULL) { 3873 // Improve the klass node's type from the new optimistic assumption: 3874 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3875 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3876 Node* cast = new CastPPNode(klass_node, akls); 3877 cast->init_req(0, control()); 3878 klass_node = _gvn.transform(cast); 3879 } 3880 3881 // Bail out if either start or end is negative. 3882 generate_negative_guard(start, bailout, &start); 3883 generate_negative_guard(end, bailout, &end); 3884 3885 Node* length = end; 3886 if (_gvn.type(start) != TypeInt::ZERO) { 3887 length = _gvn.transform(new SubINode(end, start)); 3888 } 3889 3890 // Bail out if length is negative. 3891 // Without this the new_array would throw 3892 // NegativeArraySizeException but IllegalArgumentException is what 3893 // should be thrown 3894 generate_negative_guard(length, bailout, &length); 3895 3896 if (bailout->req() > 1) { 3897 PreserveJVMState pjvms(this); 3898 set_control(_gvn.transform(bailout)); 3899 uncommon_trap(Deoptimization::Reason_intrinsic, 3900 Deoptimization::Action_maybe_recompile); 3901 } 3902 3903 if (!stopped()) { 3904 // How many elements will we copy from the original? 3905 // The answer is MinI(orig_length - start, length). 3906 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3907 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3908 3909 // Generate a direct call to the right arraycopy function(s). 3910 // We know the copy is disjoint but we might not know if the 3911 // oop stores need checking. 3912 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3913 // This will fail a store-check if x contains any non-nulls. 3914 3915 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3916 // loads/stores but it is legal only if we're sure the 3917 // Arrays.copyOf would succeed. So we need all input arguments 3918 // to the copyOf to be validated, including that the copy to the 3919 // new array won't trigger an ArrayStoreException. That subtype 3920 // check can be optimized if we know something on the type of 3921 // the input array from type speculation. 3922 if (_gvn.type(klass_node)->singleton()) { 3923 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3924 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3925 3926 int test = C->static_subtype_check(superk, subk); 3927 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3928 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3929 if (t_original->speculative_type() != NULL) { 3930 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3931 } 3932 } 3933 } 3934 3935 bool validated = false; 3936 // Reason_class_check rather than Reason_intrinsic because we 3937 // want to intrinsify even if this traps. 3938 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3939 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3940 klass_node); 3941 3942 if (not_subtype_ctrl != top()) { 3943 PreserveJVMState pjvms(this); 3944 set_control(not_subtype_ctrl); 3945 uncommon_trap(Deoptimization::Reason_class_check, 3946 Deoptimization::Action_make_not_entrant); 3947 assert(stopped(), "Should be stopped"); 3948 } 3949 validated = true; 3950 } 3951 3952 if (!stopped()) { 3953 newcopy = new_array(klass_node, length, 0); // no arguments to push 3954 3955 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3956 load_object_klass(original), klass_node); 3957 if (!is_copyOfRange) { 3958 ac->set_copyof(validated); 3959 } else { 3960 ac->set_copyofrange(validated); 3961 } 3962 Node* n = _gvn.transform(ac); 3963 if (n == ac) { 3964 ac->connect_outputs(this); 3965 } else { 3966 assert(validated, "shouldn't transform if all arguments not validated"); 3967 set_all_memory(n); 3968 } 3969 } 3970 } 3971 } // original reexecute is set back here 3972 3973 C->set_has_split_ifs(true); // Has chance for split-if optimization 3974 if (!stopped()) { 3975 set_result(newcopy); 3976 } 3977 return true; 3978} 3979 3980 3981//----------------------generate_virtual_guard--------------------------- 3982// Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3983Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3984 RegionNode* slow_region) { 3985 ciMethod* method = callee(); 3986 int vtable_index = method->vtable_index(); 3987 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3988 "bad index %d", vtable_index); 3989 // Get the Method* out of the appropriate vtable entry. 3990 int entry_offset = in_bytes(Klass::vtable_start_offset()) + 3991 vtable_index*vtableEntry::size_in_bytes() + 3992 vtableEntry::method_offset_in_bytes(); 3993 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3994 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3995 3996 // Compare the target method with the expected method (e.g., Object.hashCode). 3997 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3998 3999 Node* native_call = makecon(native_call_addr); 4000 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4001 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4002 4003 return generate_slow_guard(test_native, slow_region); 4004} 4005 4006//-----------------------generate_method_call---------------------------- 4007// Use generate_method_call to make a slow-call to the real 4008// method if the fast path fails. An alternative would be to 4009// use a stub like OptoRuntime::slow_arraycopy_Java. 4010// This only works for expanding the current library call, 4011// not another intrinsic. (E.g., don't use this for making an 4012// arraycopy call inside of the copyOf intrinsic.) 4013CallJavaNode* 4014LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4015 // When compiling the intrinsic method itself, do not use this technique. 4016 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4017 4018 ciMethod* method = callee(); 4019 // ensure the JVMS we have will be correct for this call 4020 guarantee(method_id == method->intrinsic_id(), "must match"); 4021 4022 const TypeFunc* tf = TypeFunc::make(method); 4023 CallJavaNode* slow_call; 4024 if (is_static) { 4025 assert(!is_virtual, ""); 4026 slow_call = new CallStaticJavaNode(C, tf, 4027 SharedRuntime::get_resolve_static_call_stub(), 4028 method, bci()); 4029 } else if (is_virtual) { 4030 null_check_receiver(); 4031 int vtable_index = Method::invalid_vtable_index; 4032 if (UseInlineCaches) { 4033 // Suppress the vtable call 4034 } else { 4035 // hashCode and clone are not a miranda methods, 4036 // so the vtable index is fixed. 4037 // No need to use the linkResolver to get it. 4038 vtable_index = method->vtable_index(); 4039 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4040 "bad index %d", vtable_index); 4041 } 4042 slow_call = new CallDynamicJavaNode(tf, 4043 SharedRuntime::get_resolve_virtual_call_stub(), 4044 method, vtable_index, bci()); 4045 } else { // neither virtual nor static: opt_virtual 4046 null_check_receiver(); 4047 slow_call = new CallStaticJavaNode(C, tf, 4048 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4049 method, bci()); 4050 slow_call->set_optimized_virtual(true); 4051 } 4052 set_arguments_for_java_call(slow_call); 4053 set_edges_for_java_call(slow_call); 4054 return slow_call; 4055} 4056 4057 4058/** 4059 * Build special case code for calls to hashCode on an object. This call may 4060 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4061 * slightly different code. 4062 */ 4063bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4064 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4065 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4066 4067 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4068 4069 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4070 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4071 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4072 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4073 Node* obj = NULL; 4074 if (!is_static) { 4075 // Check for hashing null object 4076 obj = null_check_receiver(); 4077 if (stopped()) return true; // unconditionally null 4078 result_reg->init_req(_null_path, top()); 4079 result_val->init_req(_null_path, top()); 4080 } else { 4081 // Do a null check, and return zero if null. 4082 // System.identityHashCode(null) == 0 4083 obj = argument(0); 4084 Node* null_ctl = top(); 4085 obj = null_check_oop(obj, &null_ctl); 4086 result_reg->init_req(_null_path, null_ctl); 4087 result_val->init_req(_null_path, _gvn.intcon(0)); 4088 } 4089 4090 // Unconditionally null? Then return right away. 4091 if (stopped()) { 4092 set_control( result_reg->in(_null_path)); 4093 if (!stopped()) 4094 set_result(result_val->in(_null_path)); 4095 return true; 4096 } 4097 4098 // We only go to the fast case code if we pass a number of guards. The 4099 // paths which do not pass are accumulated in the slow_region. 4100 RegionNode* slow_region = new RegionNode(1); 4101 record_for_igvn(slow_region); 4102 4103 // If this is a virtual call, we generate a funny guard. We pull out 4104 // the vtable entry corresponding to hashCode() from the target object. 4105 // If the target method which we are calling happens to be the native 4106 // Object hashCode() method, we pass the guard. We do not need this 4107 // guard for non-virtual calls -- the caller is known to be the native 4108 // Object hashCode(). 4109 if (is_virtual) { 4110 // After null check, get the object's klass. 4111 Node* obj_klass = load_object_klass(obj); 4112 generate_virtual_guard(obj_klass, slow_region); 4113 } 4114 4115 // Get the header out of the object, use LoadMarkNode when available 4116 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4117 // The control of the load must be NULL. Otherwise, the load can move before 4118 // the null check after castPP removal. 4119 Node* no_ctrl = NULL; 4120 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4121 4122 // Test the header to see if it is unlocked. 4123 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4124 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4125 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4126 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4127 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4128 4129 generate_slow_guard(test_unlocked, slow_region); 4130 4131 // Get the hash value and check to see that it has been properly assigned. 4132 // We depend on hash_mask being at most 32 bits and avoid the use of 4133 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4134 // vm: see markOop.hpp. 4135 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4136 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4137 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4138 // This hack lets the hash bits live anywhere in the mark object now, as long 4139 // as the shift drops the relevant bits into the low 32 bits. Note that 4140 // Java spec says that HashCode is an int so there's no point in capturing 4141 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4142 hshifted_header = ConvX2I(hshifted_header); 4143 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4144 4145 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4146 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4147 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4148 4149 generate_slow_guard(test_assigned, slow_region); 4150 4151 Node* init_mem = reset_memory(); 4152 // fill in the rest of the null path: 4153 result_io ->init_req(_null_path, i_o()); 4154 result_mem->init_req(_null_path, init_mem); 4155 4156 result_val->init_req(_fast_path, hash_val); 4157 result_reg->init_req(_fast_path, control()); 4158 result_io ->init_req(_fast_path, i_o()); 4159 result_mem->init_req(_fast_path, init_mem); 4160 4161 // Generate code for the slow case. We make a call to hashCode(). 4162 set_control(_gvn.transform(slow_region)); 4163 if (!stopped()) { 4164 // No need for PreserveJVMState, because we're using up the present state. 4165 set_all_memory(init_mem); 4166 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4167 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4168 Node* slow_result = set_results_for_java_call(slow_call); 4169 // this->control() comes from set_results_for_java_call 4170 result_reg->init_req(_slow_path, control()); 4171 result_val->init_req(_slow_path, slow_result); 4172 result_io ->set_req(_slow_path, i_o()); 4173 result_mem ->set_req(_slow_path, reset_memory()); 4174 } 4175 4176 // Return the combined state. 4177 set_i_o( _gvn.transform(result_io) ); 4178 set_all_memory( _gvn.transform(result_mem)); 4179 4180 set_result(result_reg, result_val); 4181 return true; 4182} 4183 4184//---------------------------inline_native_getClass---------------------------- 4185// public final native Class<?> java.lang.Object.getClass(); 4186// 4187// Build special case code for calls to getClass on an object. 4188bool LibraryCallKit::inline_native_getClass() { 4189 Node* obj = null_check_receiver(); 4190 if (stopped()) return true; 4191 set_result(load_mirror_from_klass(load_object_klass(obj))); 4192 return true; 4193} 4194 4195//-----------------inline_native_Reflection_getCallerClass--------------------- 4196// public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4197// 4198// In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4199// 4200// NOTE: This code must perform the same logic as JVM_GetCallerClass 4201// in that it must skip particular security frames and checks for 4202// caller sensitive methods. 4203bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4204#ifndef PRODUCT 4205 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4206 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4207 } 4208#endif 4209 4210 if (!jvms()->has_method()) { 4211#ifndef PRODUCT 4212 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4213 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4214 } 4215#endif 4216 return false; 4217 } 4218 4219 // Walk back up the JVM state to find the caller at the required 4220 // depth. 4221 JVMState* caller_jvms = jvms(); 4222 4223 // Cf. JVM_GetCallerClass 4224 // NOTE: Start the loop at depth 1 because the current JVM state does 4225 // not include the Reflection.getCallerClass() frame. 4226 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4227 ciMethod* m = caller_jvms->method(); 4228 switch (n) { 4229 case 0: 4230 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4231 break; 4232 case 1: 4233 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4234 if (!m->caller_sensitive()) { 4235#ifndef PRODUCT 4236 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4237 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4238 } 4239#endif 4240 return false; // bail-out; let JVM_GetCallerClass do the work 4241 } 4242 break; 4243 default: 4244 if (!m->is_ignored_by_security_stack_walk()) { 4245 // We have reached the desired frame; return the holder class. 4246 // Acquire method holder as java.lang.Class and push as constant. 4247 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4248 ciInstance* caller_mirror = caller_klass->java_mirror(); 4249 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4250 4251#ifndef PRODUCT 4252 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4253 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()); 4254 tty->print_cr(" JVM state at this point:"); 4255 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4256 ciMethod* m = jvms()->of_depth(i)->method(); 4257 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4258 } 4259 } 4260#endif 4261 return true; 4262 } 4263 break; 4264 } 4265 } 4266 4267#ifndef PRODUCT 4268 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4269 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4270 tty->print_cr(" JVM state at this point:"); 4271 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4272 ciMethod* m = jvms()->of_depth(i)->method(); 4273 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4274 } 4275 } 4276#endif 4277 4278 return false; // bail-out; let JVM_GetCallerClass do the work 4279} 4280 4281bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4282 Node* arg = argument(0); 4283 Node* result = NULL; 4284 4285 switch (id) { 4286 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4287 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4288 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4289 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4290 4291 case vmIntrinsics::_doubleToLongBits: { 4292 // two paths (plus control) merge in a wood 4293 RegionNode *r = new RegionNode(3); 4294 Node *phi = new PhiNode(r, TypeLong::LONG); 4295 4296 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4297 // Build the boolean node 4298 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4299 4300 // Branch either way. 4301 // NaN case is less traveled, which makes all the difference. 4302 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4303 Node *opt_isnan = _gvn.transform(ifisnan); 4304 assert( opt_isnan->is_If(), "Expect an IfNode"); 4305 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4306 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4307 4308 set_control(iftrue); 4309 4310 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4311 Node *slow_result = longcon(nan_bits); // return NaN 4312 phi->init_req(1, _gvn.transform( slow_result )); 4313 r->init_req(1, iftrue); 4314 4315 // Else fall through 4316 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4317 set_control(iffalse); 4318 4319 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4320 r->init_req(2, iffalse); 4321 4322 // Post merge 4323 set_control(_gvn.transform(r)); 4324 record_for_igvn(r); 4325 4326 C->set_has_split_ifs(true); // Has chance for split-if optimization 4327 result = phi; 4328 assert(result->bottom_type()->isa_long(), "must be"); 4329 break; 4330 } 4331 4332 case vmIntrinsics::_floatToIntBits: { 4333 // two paths (plus control) merge in a wood 4334 RegionNode *r = new RegionNode(3); 4335 Node *phi = new PhiNode(r, TypeInt::INT); 4336 4337 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4338 // Build the boolean node 4339 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4340 4341 // Branch either way. 4342 // NaN case is less traveled, which makes all the difference. 4343 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4344 Node *opt_isnan = _gvn.transform(ifisnan); 4345 assert( opt_isnan->is_If(), "Expect an IfNode"); 4346 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4347 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4348 4349 set_control(iftrue); 4350 4351 static const jint nan_bits = 0x7fc00000; 4352 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4353 phi->init_req(1, _gvn.transform( slow_result )); 4354 r->init_req(1, iftrue); 4355 4356 // Else fall through 4357 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4358 set_control(iffalse); 4359 4360 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4361 r->init_req(2, iffalse); 4362 4363 // Post merge 4364 set_control(_gvn.transform(r)); 4365 record_for_igvn(r); 4366 4367 C->set_has_split_ifs(true); // Has chance for split-if optimization 4368 result = phi; 4369 assert(result->bottom_type()->isa_int(), "must be"); 4370 break; 4371 } 4372 4373 default: 4374 fatal_unexpected_iid(id); 4375 break; 4376 } 4377 set_result(_gvn.transform(result)); 4378 return true; 4379} 4380 4381//----------------------inline_unsafe_copyMemory------------------------- 4382// public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4383bool LibraryCallKit::inline_unsafe_copyMemory() { 4384 if (callee()->is_static()) return false; // caller must have the capability! 4385 null_check_receiver(); // null-check receiver 4386 if (stopped()) return true; 4387 4388 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4389 4390 Node* src_ptr = argument(1); // type: oop 4391 Node* src_off = ConvL2X(argument(2)); // type: long 4392 Node* dst_ptr = argument(4); // type: oop 4393 Node* dst_off = ConvL2X(argument(5)); // type: long 4394 Node* size = ConvL2X(argument(7)); // type: long 4395 4396 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4397 "fieldOffset must be byte-scaled"); 4398 4399 Node* src = make_unsafe_address(src_ptr, src_off); 4400 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4401 4402 // Conservatively insert a memory barrier on all memory slices. 4403 // Do not let writes of the copy source or destination float below the copy. 4404 insert_mem_bar(Op_MemBarCPUOrder); 4405 4406 // Call it. Note that the length argument is not scaled. 4407 make_runtime_call(RC_LEAF|RC_NO_FP, 4408 OptoRuntime::fast_arraycopy_Type(), 4409 StubRoutines::unsafe_arraycopy(), 4410 "unsafe_arraycopy", 4411 TypeRawPtr::BOTTOM, 4412 src, dst, size XTOP); 4413 4414 // Do not let reads of the copy destination float above the copy. 4415 insert_mem_bar(Op_MemBarCPUOrder); 4416 4417 return true; 4418} 4419 4420//------------------------clone_coping----------------------------------- 4421// Helper function for inline_native_clone. 4422void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4423 assert(obj_size != NULL, ""); 4424 Node* raw_obj = alloc_obj->in(1); 4425 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4426 4427 AllocateNode* alloc = NULL; 4428 if (ReduceBulkZeroing) { 4429 // We will be completely responsible for initializing this object - 4430 // mark Initialize node as complete. 4431 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4432 // The object was just allocated - there should be no any stores! 4433 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4434 // Mark as complete_with_arraycopy so that on AllocateNode 4435 // expansion, we know this AllocateNode is initialized by an array 4436 // copy and a StoreStore barrier exists after the array copy. 4437 alloc->initialization()->set_complete_with_arraycopy(); 4438 } 4439 4440 // Copy the fastest available way. 4441 // TODO: generate fields copies for small objects instead. 4442 Node* src = obj; 4443 Node* dest = alloc_obj; 4444 Node* size = _gvn.transform(obj_size); 4445 4446 // Exclude the header but include array length to copy by 8 bytes words. 4447 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4448 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4449 instanceOopDesc::base_offset_in_bytes(); 4450 // base_off: 4451 // 8 - 32-bit VM 4452 // 12 - 64-bit VM, compressed klass 4453 // 16 - 64-bit VM, normal klass 4454 if (base_off % BytesPerLong != 0) { 4455 assert(UseCompressedClassPointers, ""); 4456 if (is_array) { 4457 // Exclude length to copy by 8 bytes words. 4458 base_off += sizeof(int); 4459 } else { 4460 // Include klass to copy by 8 bytes words. 4461 base_off = instanceOopDesc::klass_offset_in_bytes(); 4462 } 4463 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4464 } 4465 src = basic_plus_adr(src, base_off); 4466 dest = basic_plus_adr(dest, base_off); 4467 4468 // Compute the length also, if needed: 4469 Node* countx = size; 4470 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4471 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4472 4473 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4474 4475 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4476 ac->set_clonebasic(); 4477 Node* n = _gvn.transform(ac); 4478 if (n == ac) { 4479 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4480 } else { 4481 set_all_memory(n); 4482 } 4483 4484 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4485 if (card_mark) { 4486 assert(!is_array, ""); 4487 // Put in store barrier for any and all oops we are sticking 4488 // into this object. (We could avoid this if we could prove 4489 // that the object type contains no oop fields at all.) 4490 Node* no_particular_value = NULL; 4491 Node* no_particular_field = NULL; 4492 int raw_adr_idx = Compile::AliasIdxRaw; 4493 post_barrier(control(), 4494 memory(raw_adr_type), 4495 alloc_obj, 4496 no_particular_field, 4497 raw_adr_idx, 4498 no_particular_value, 4499 T_OBJECT, 4500 false); 4501 } 4502 4503 // Do not let reads from the cloned object float above the arraycopy. 4504 if (alloc != NULL) { 4505 // Do not let stores that initialize this object be reordered with 4506 // a subsequent store that would make this object accessible by 4507 // other threads. 4508 // Record what AllocateNode this StoreStore protects so that 4509 // escape analysis can go from the MemBarStoreStoreNode to the 4510 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4511 // based on the escape status of the AllocateNode. 4512 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4513 } else { 4514 insert_mem_bar(Op_MemBarCPUOrder); 4515 } 4516} 4517 4518//------------------------inline_native_clone---------------------------- 4519// protected native Object java.lang.Object.clone(); 4520// 4521// Here are the simple edge cases: 4522// null receiver => normal trap 4523// virtual and clone was overridden => slow path to out-of-line clone 4524// not cloneable or finalizer => slow path to out-of-line Object.clone 4525// 4526// The general case has two steps, allocation and copying. 4527// Allocation has two cases, and uses GraphKit::new_instance or new_array. 4528// 4529// Copying also has two cases, oop arrays and everything else. 4530// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4531// Everything else uses the tight inline loop supplied by CopyArrayNode. 4532// 4533// These steps fold up nicely if and when the cloned object's klass 4534// can be sharply typed as an object array, a type array, or an instance. 4535// 4536bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4537 PhiNode* result_val; 4538 4539 // Set the reexecute bit for the interpreter to reexecute 4540 // the bytecode that invokes Object.clone if deoptimization happens. 4541 { PreserveReexecuteState preexecs(this); 4542 jvms()->set_should_reexecute(true); 4543 4544 Node* obj = null_check_receiver(); 4545 if (stopped()) return true; 4546 4547 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4548 4549 // If we are going to clone an instance, we need its exact type to 4550 // know the number and types of fields to convert the clone to 4551 // loads/stores. Maybe a speculative type can help us. 4552 if (!obj_type->klass_is_exact() && 4553 obj_type->speculative_type() != NULL && 4554 obj_type->speculative_type()->is_instance_klass()) { 4555 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4556 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4557 !spec_ik->has_injected_fields()) { 4558 ciKlass* k = obj_type->klass(); 4559 if (!k->is_instance_klass() || 4560 k->as_instance_klass()->is_interface() || 4561 k->as_instance_klass()->has_subklass()) { 4562 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4563 } 4564 } 4565 } 4566 4567 Node* obj_klass = load_object_klass(obj); 4568 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4569 const TypeOopPtr* toop = ((tklass != NULL) 4570 ? tklass->as_instance_type() 4571 : TypeInstPtr::NOTNULL); 4572 4573 // Conservatively insert a memory barrier on all memory slices. 4574 // Do not let writes into the original float below the clone. 4575 insert_mem_bar(Op_MemBarCPUOrder); 4576 4577 // paths into result_reg: 4578 enum { 4579 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4580 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4581 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4582 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4583 PATH_LIMIT 4584 }; 4585 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4586 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4587 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4588 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4589 record_for_igvn(result_reg); 4590 4591 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4592 int raw_adr_idx = Compile::AliasIdxRaw; 4593 4594 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4595 if (array_ctl != NULL) { 4596 // It's an array. 4597 PreserveJVMState pjvms(this); 4598 set_control(array_ctl); 4599 Node* obj_length = load_array_length(obj); 4600 Node* obj_size = NULL; 4601 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4602 4603 if (!use_ReduceInitialCardMarks()) { 4604 // If it is an oop array, it requires very special treatment, 4605 // because card marking is required on each card of the array. 4606 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4607 if (is_obja != NULL) { 4608 PreserveJVMState pjvms2(this); 4609 set_control(is_obja); 4610 // Generate a direct call to the right arraycopy function(s). 4611 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4612 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4613 ac->set_cloneoop(); 4614 Node* n = _gvn.transform(ac); 4615 assert(n == ac, "cannot disappear"); 4616 ac->connect_outputs(this); 4617 4618 result_reg->init_req(_objArray_path, control()); 4619 result_val->init_req(_objArray_path, alloc_obj); 4620 result_i_o ->set_req(_objArray_path, i_o()); 4621 result_mem ->set_req(_objArray_path, reset_memory()); 4622 } 4623 } 4624 // Otherwise, there are no card marks to worry about. 4625 // (We can dispense with card marks if we know the allocation 4626 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4627 // causes the non-eden paths to take compensating steps to 4628 // simulate a fresh allocation, so that no further 4629 // card marks are required in compiled code to initialize 4630 // the object.) 4631 4632 if (!stopped()) { 4633 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4634 4635 // Present the results of the copy. 4636 result_reg->init_req(_array_path, control()); 4637 result_val->init_req(_array_path, alloc_obj); 4638 result_i_o ->set_req(_array_path, i_o()); 4639 result_mem ->set_req(_array_path, reset_memory()); 4640 } 4641 } 4642 4643 // We only go to the instance fast case code if we pass a number of guards. 4644 // The paths which do not pass are accumulated in the slow_region. 4645 RegionNode* slow_region = new RegionNode(1); 4646 record_for_igvn(slow_region); 4647 if (!stopped()) { 4648 // It's an instance (we did array above). Make the slow-path tests. 4649 // If this is a virtual call, we generate a funny guard. We grab 4650 // the vtable entry corresponding to clone() from the target object. 4651 // If the target method which we are calling happens to be the 4652 // Object clone() method, we pass the guard. We do not need this 4653 // guard for non-virtual calls; the caller is known to be the native 4654 // Object clone(). 4655 if (is_virtual) { 4656 generate_virtual_guard(obj_klass, slow_region); 4657 } 4658 4659 // The object must be easily cloneable and must not have a finalizer. 4660 // Both of these conditions may be checked in a single test. 4661 // We could optimize the test further, but we don't care. 4662 generate_access_flags_guard(obj_klass, 4663 // Test both conditions: 4664 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER, 4665 // Must be cloneable but not finalizer: 4666 JVM_ACC_IS_CLONEABLE_FAST, 4667 slow_region); 4668 } 4669 4670 if (!stopped()) { 4671 // It's an instance, and it passed the slow-path tests. 4672 PreserveJVMState pjvms(this); 4673 Node* obj_size = NULL; 4674 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4675 // is reexecuted if deoptimization occurs and there could be problems when merging 4676 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4677 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4678 4679 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4680 4681 // Present the results of the slow call. 4682 result_reg->init_req(_instance_path, control()); 4683 result_val->init_req(_instance_path, alloc_obj); 4684 result_i_o ->set_req(_instance_path, i_o()); 4685 result_mem ->set_req(_instance_path, reset_memory()); 4686 } 4687 4688 // Generate code for the slow case. We make a call to clone(). 4689 set_control(_gvn.transform(slow_region)); 4690 if (!stopped()) { 4691 PreserveJVMState pjvms(this); 4692 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4693 Node* slow_result = set_results_for_java_call(slow_call); 4694 // this->control() comes from set_results_for_java_call 4695 result_reg->init_req(_slow_path, control()); 4696 result_val->init_req(_slow_path, slow_result); 4697 result_i_o ->set_req(_slow_path, i_o()); 4698 result_mem ->set_req(_slow_path, reset_memory()); 4699 } 4700 4701 // Return the combined state. 4702 set_control( _gvn.transform(result_reg)); 4703 set_i_o( _gvn.transform(result_i_o)); 4704 set_all_memory( _gvn.transform(result_mem)); 4705 } // original reexecute is set back here 4706 4707 set_result(_gvn.transform(result_val)); 4708 return true; 4709} 4710 4711// If we have a tighly coupled allocation, the arraycopy may take care 4712// of the array initialization. If one of the guards we insert between 4713// the allocation and the arraycopy causes a deoptimization, an 4714// unitialized array will escape the compiled method. To prevent that 4715// we set the JVM state for uncommon traps between the allocation and 4716// the arraycopy to the state before the allocation so, in case of 4717// deoptimization, we'll reexecute the allocation and the 4718// initialization. 4719JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4720 if (alloc != NULL) { 4721 ciMethod* trap_method = alloc->jvms()->method(); 4722 int trap_bci = alloc->jvms()->bci(); 4723 4724 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4725 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4726 // Make sure there's no store between the allocation and the 4727 // arraycopy otherwise visible side effects could be rexecuted 4728 // in case of deoptimization and cause incorrect execution. 4729 bool no_interfering_store = true; 4730 Node* mem = alloc->in(TypeFunc::Memory); 4731 if (mem->is_MergeMem()) { 4732 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4733 Node* n = mms.memory(); 4734 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4735 assert(n->is_Store(), "what else?"); 4736 no_interfering_store = false; 4737 break; 4738 } 4739 } 4740 } else { 4741 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4742 Node* n = mms.memory(); 4743 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4744 assert(n->is_Store(), "what else?"); 4745 no_interfering_store = false; 4746 break; 4747 } 4748 } 4749 } 4750 4751 if (no_interfering_store) { 4752 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4753 uint size = alloc->req(); 4754 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4755 old_jvms->set_map(sfpt); 4756 for (uint i = 0; i < size; i++) { 4757 sfpt->init_req(i, alloc->in(i)); 4758 } 4759 // re-push array length for deoptimization 4760 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4761 old_jvms->set_sp(old_jvms->sp()+1); 4762 old_jvms->set_monoff(old_jvms->monoff()+1); 4763 old_jvms->set_scloff(old_jvms->scloff()+1); 4764 old_jvms->set_endoff(old_jvms->endoff()+1); 4765 old_jvms->set_should_reexecute(true); 4766 4767 sfpt->set_i_o(map()->i_o()); 4768 sfpt->set_memory(map()->memory()); 4769 sfpt->set_control(map()->control()); 4770 4771 JVMState* saved_jvms = jvms(); 4772 saved_reexecute_sp = _reexecute_sp; 4773 4774 set_jvms(sfpt->jvms()); 4775 _reexecute_sp = jvms()->sp(); 4776 4777 return saved_jvms; 4778 } 4779 } 4780 } 4781 return NULL; 4782} 4783 4784// In case of a deoptimization, we restart execution at the 4785// allocation, allocating a new array. We would leave an uninitialized 4786// array in the heap that GCs wouldn't expect. Move the allocation 4787// after the traps so we don't allocate the array if we 4788// deoptimize. This is possible because tightly_coupled_allocation() 4789// guarantees there's no observer of the allocated array at this point 4790// and the control flow is simple enough. 4791void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4792 if (saved_jvms != NULL && !stopped()) { 4793 assert(alloc != NULL, "only with a tightly coupled allocation"); 4794 // restore JVM state to the state at the arraycopy 4795 saved_jvms->map()->set_control(map()->control()); 4796 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4797 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4798 // If we've improved the types of some nodes (null check) while 4799 // emitting the guards, propagate them to the current state 4800 map()->replaced_nodes().apply(saved_jvms->map()); 4801 set_jvms(saved_jvms); 4802 _reexecute_sp = saved_reexecute_sp; 4803 4804 // Remove the allocation from above the guards 4805 CallProjections callprojs; 4806 alloc->extract_projections(&callprojs, true); 4807 InitializeNode* init = alloc->initialization(); 4808 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4809 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4810 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4811 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4812 4813 // move the allocation here (after the guards) 4814 _gvn.hash_delete(alloc); 4815 alloc->set_req(TypeFunc::Control, control()); 4816 alloc->set_req(TypeFunc::I_O, i_o()); 4817 Node *mem = reset_memory(); 4818 set_all_memory(mem); 4819 alloc->set_req(TypeFunc::Memory, mem); 4820 set_control(init->proj_out(TypeFunc::Control)); 4821 set_i_o(callprojs.fallthrough_ioproj); 4822 4823 // Update memory as done in GraphKit::set_output_for_allocation() 4824 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4825 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4826 if (ary_type->isa_aryptr() && length_type != NULL) { 4827 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4828 } 4829 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4830 int elemidx = C->get_alias_index(telemref); 4831 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4832 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4833 4834 Node* allocx = _gvn.transform(alloc); 4835 assert(allocx == alloc, "where has the allocation gone?"); 4836 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4837 4838 _gvn.hash_delete(dest); 4839 dest->set_req(0, control()); 4840 Node* destx = _gvn.transform(dest); 4841 assert(destx == dest, "where has the allocation result gone?"); 4842 } 4843} 4844 4845 4846//------------------------------inline_arraycopy----------------------- 4847// public static native void java.lang.System.arraycopy(Object src, int srcPos, 4848// Object dest, int destPos, 4849// int length); 4850bool LibraryCallKit::inline_arraycopy() { 4851 // Get the arguments. 4852 Node* src = argument(0); // type: oop 4853 Node* src_offset = argument(1); // type: int 4854 Node* dest = argument(2); // type: oop 4855 Node* dest_offset = argument(3); // type: int 4856 Node* length = argument(4); // type: int 4857 4858 4859 // Check for allocation before we add nodes that would confuse 4860 // tightly_coupled_allocation() 4861 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4862 4863 int saved_reexecute_sp = -1; 4864 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4865 // See arraycopy_restore_alloc_state() comment 4866 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4867 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4868 // if saved_jvms == NULL and alloc != NULL, we can���t emit any guards 4869 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4870 4871 // The following tests must be performed 4872 // (1) src and dest are arrays. 4873 // (2) src and dest arrays must have elements of the same BasicType 4874 // (3) src and dest must not be null. 4875 // (4) src_offset must not be negative. 4876 // (5) dest_offset must not be negative. 4877 // (6) length must not be negative. 4878 // (7) src_offset + length must not exceed length of src. 4879 // (8) dest_offset + length must not exceed length of dest. 4880 // (9) each element of an oop array must be assignable 4881 4882 // (3) src and dest must not be null. 4883 // always do this here because we need the JVM state for uncommon traps 4884 Node* null_ctl = top(); 4885 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4886 assert(null_ctl->is_top(), "no null control here"); 4887 dest = null_check(dest, T_ARRAY); 4888 4889 if (!can_emit_guards) { 4890 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4891 // guards but the arraycopy node could still take advantage of a 4892 // tightly allocated allocation. tightly_coupled_allocation() is 4893 // called again to make sure it takes the null check above into 4894 // account: the null check is mandatory and if it caused an 4895 // uncommon trap to be emitted then the allocation can't be 4896 // considered tightly coupled in this context. 4897 alloc = tightly_coupled_allocation(dest, NULL); 4898 } 4899 4900 bool validated = false; 4901 4902 const Type* src_type = _gvn.type(src); 4903 const Type* dest_type = _gvn.type(dest); 4904 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4905 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4906 4907 // Do we have the type of src? 4908 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4909 // Do we have the type of dest? 4910 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4911 // Is the type for src from speculation? 4912 bool src_spec = false; 4913 // Is the type for dest from speculation? 4914 bool dest_spec = false; 4915 4916 if ((!has_src || !has_dest) && can_emit_guards) { 4917 // We don't have sufficient type information, let's see if 4918 // speculative types can help. We need to have types for both src 4919 // and dest so that it pays off. 4920 4921 // Do we already have or could we have type information for src 4922 bool could_have_src = has_src; 4923 // Do we already have or could we have type information for dest 4924 bool could_have_dest = has_dest; 4925 4926 ciKlass* src_k = NULL; 4927 if (!has_src) { 4928 src_k = src_type->speculative_type_not_null(); 4929 if (src_k != NULL && src_k->is_array_klass()) { 4930 could_have_src = true; 4931 } 4932 } 4933 4934 ciKlass* dest_k = NULL; 4935 if (!has_dest) { 4936 dest_k = dest_type->speculative_type_not_null(); 4937 if (dest_k != NULL && dest_k->is_array_klass()) { 4938 could_have_dest = true; 4939 } 4940 } 4941 4942 if (could_have_src && could_have_dest) { 4943 // This is going to pay off so emit the required guards 4944 if (!has_src) { 4945 src = maybe_cast_profiled_obj(src, src_k, true); 4946 src_type = _gvn.type(src); 4947 top_src = src_type->isa_aryptr(); 4948 has_src = (top_src != NULL && top_src->klass() != NULL); 4949 src_spec = true; 4950 } 4951 if (!has_dest) { 4952 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4953 dest_type = _gvn.type(dest); 4954 top_dest = dest_type->isa_aryptr(); 4955 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4956 dest_spec = true; 4957 } 4958 } 4959 } 4960 4961 if (has_src && has_dest && can_emit_guards) { 4962 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4963 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4964 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4965 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4966 4967 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4968 // If both arrays are object arrays then having the exact types 4969 // for both will remove the need for a subtype check at runtime 4970 // before the call and may make it possible to pick a faster copy 4971 // routine (without a subtype check on every element) 4972 // Do we have the exact type of src? 4973 bool could_have_src = src_spec; 4974 // Do we have the exact type of dest? 4975 bool could_have_dest = dest_spec; 4976 ciKlass* src_k = top_src->klass(); 4977 ciKlass* dest_k = top_dest->klass(); 4978 if (!src_spec) { 4979 src_k = src_type->speculative_type_not_null(); 4980 if (src_k != NULL && src_k->is_array_klass()) { 4981 could_have_src = true; 4982 } 4983 } 4984 if (!dest_spec) { 4985 dest_k = dest_type->speculative_type_not_null(); 4986 if (dest_k != NULL && dest_k->is_array_klass()) { 4987 could_have_dest = true; 4988 } 4989 } 4990 if (could_have_src && could_have_dest) { 4991 // If we can have both exact types, emit the missing guards 4992 if (could_have_src && !src_spec) { 4993 src = maybe_cast_profiled_obj(src, src_k, true); 4994 } 4995 if (could_have_dest && !dest_spec) { 4996 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4997 } 4998 } 4999 } 5000 } 5001 5002 ciMethod* trap_method = method(); 5003 int trap_bci = bci(); 5004 if (saved_jvms != NULL) { 5005 trap_method = alloc->jvms()->method(); 5006 trap_bci = alloc->jvms()->bci(); 5007 } 5008 5009 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5010 can_emit_guards && 5011 !src->is_top() && !dest->is_top()) { 5012 // validate arguments: enables transformation the ArrayCopyNode 5013 validated = true; 5014 5015 RegionNode* slow_region = new RegionNode(1); 5016 record_for_igvn(slow_region); 5017 5018 // (1) src and dest are arrays. 5019 generate_non_array_guard(load_object_klass(src), slow_region); 5020 generate_non_array_guard(load_object_klass(dest), slow_region); 5021 5022 // (2) src and dest arrays must have elements of the same BasicType 5023 // done at macro expansion or at Ideal transformation time 5024 5025 // (4) src_offset must not be negative. 5026 generate_negative_guard(src_offset, slow_region); 5027 5028 // (5) dest_offset must not be negative. 5029 generate_negative_guard(dest_offset, slow_region); 5030 5031 // (7) src_offset + length must not exceed length of src. 5032 generate_limit_guard(src_offset, length, 5033 load_array_length(src), 5034 slow_region); 5035 5036 // (8) dest_offset + length must not exceed length of dest. 5037 generate_limit_guard(dest_offset, length, 5038 load_array_length(dest), 5039 slow_region); 5040 5041 // (9) each element of an oop array must be assignable 5042 Node* src_klass = load_object_klass(src); 5043 Node* dest_klass = load_object_klass(dest); 5044 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5045 5046 if (not_subtype_ctrl != top()) { 5047 PreserveJVMState pjvms(this); 5048 set_control(not_subtype_ctrl); 5049 uncommon_trap(Deoptimization::Reason_intrinsic, 5050 Deoptimization::Action_make_not_entrant); 5051 assert(stopped(), "Should be stopped"); 5052 } 5053 { 5054 PreserveJVMState pjvms(this); 5055 set_control(_gvn.transform(slow_region)); 5056 uncommon_trap(Deoptimization::Reason_intrinsic, 5057 Deoptimization::Action_make_not_entrant); 5058 assert(stopped(), "Should be stopped"); 5059 } 5060 } 5061 5062 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5063 5064 if (stopped()) { 5065 return true; 5066 } 5067 5068 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5069 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5070 // so the compiler has a chance to eliminate them: during macro expansion, 5071 // we have to set their control (CastPP nodes are eliminated). 5072 load_object_klass(src), load_object_klass(dest), 5073 load_array_length(src), load_array_length(dest)); 5074 5075 ac->set_arraycopy(validated); 5076 5077 Node* n = _gvn.transform(ac); 5078 if (n == ac) { 5079 ac->connect_outputs(this); 5080 } else { 5081 assert(validated, "shouldn't transform if all arguments not validated"); 5082 set_all_memory(n); 5083 } 5084 5085 return true; 5086} 5087 5088 5089// Helper function which determines if an arraycopy immediately follows 5090// an allocation, with no intervening tests or other escapes for the object. 5091AllocateArrayNode* 5092LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5093 RegionNode* slow_region) { 5094 if (stopped()) return NULL; // no fast path 5095 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5096 5097 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5098 if (alloc == NULL) return NULL; 5099 5100 Node* rawmem = memory(Compile::AliasIdxRaw); 5101 // Is the allocation's memory state untouched? 5102 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5103 // Bail out if there have been raw-memory effects since the allocation. 5104 // (Example: There might have been a call or safepoint.) 5105 return NULL; 5106 } 5107 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5108 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5109 return NULL; 5110 } 5111 5112 // There must be no unexpected observers of this allocation. 5113 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5114 Node* obs = ptr->fast_out(i); 5115 if (obs != this->map()) { 5116 return NULL; 5117 } 5118 } 5119 5120 // This arraycopy must unconditionally follow the allocation of the ptr. 5121 Node* alloc_ctl = ptr->in(0); 5122 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5123 5124 Node* ctl = control(); 5125 while (ctl != alloc_ctl) { 5126 // There may be guards which feed into the slow_region. 5127 // Any other control flow means that we might not get a chance 5128 // to finish initializing the allocated object. 5129 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5130 IfNode* iff = ctl->in(0)->as_If(); 5131 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5132 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5133 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5134 ctl = iff->in(0); // This test feeds the known slow_region. 5135 continue; 5136 } 5137 // One more try: Various low-level checks bottom out in 5138 // uncommon traps. If the debug-info of the trap omits 5139 // any reference to the allocation, as we've already 5140 // observed, then there can be no objection to the trap. 5141 bool found_trap = false; 5142 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5143 Node* obs = not_ctl->fast_out(j); 5144 if (obs->in(0) == not_ctl && obs->is_Call() && 5145 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5146 found_trap = true; break; 5147 } 5148 } 5149 if (found_trap) { 5150 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5151 continue; 5152 } 5153 } 5154 return NULL; 5155 } 5156 5157 // If we get this far, we have an allocation which immediately 5158 // precedes the arraycopy, and we can take over zeroing the new object. 5159 // The arraycopy will finish the initialization, and provide 5160 // a new control state to which we will anchor the destination pointer. 5161 5162 return alloc; 5163} 5164 5165//-------------inline_encodeISOArray----------------------------------- 5166// encode char[] to byte[] in ISO_8859_1 5167bool LibraryCallKit::inline_encodeISOArray() { 5168 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5169 // no receiver since it is static method 5170 Node *src = argument(0); 5171 Node *src_offset = argument(1); 5172 Node *dst = argument(2); 5173 Node *dst_offset = argument(3); 5174 Node *length = argument(4); 5175 5176 const Type* src_type = src->Value(&_gvn); 5177 const Type* dst_type = dst->Value(&_gvn); 5178 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5179 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5180 if (top_src == NULL || top_src->klass() == NULL || 5181 top_dest == NULL || top_dest->klass() == NULL) { 5182 // failed array check 5183 return false; 5184 } 5185 5186 // Figure out the size and type of the elements we will be copying. 5187 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5188 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5189 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) { 5190 return false; 5191 } 5192 5193 Node* src_start = array_element_address(src, src_offset, T_CHAR); 5194 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5195 // 'src_start' points to src array + scaled offset 5196 // 'dst_start' points to dst array + scaled offset 5197 5198 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5199 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5200 enc = _gvn.transform(enc); 5201 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5202 set_memory(res_mem, mtype); 5203 set_result(enc); 5204 return true; 5205} 5206 5207//-------------inline_multiplyToLen----------------------------------- 5208bool LibraryCallKit::inline_multiplyToLen() { 5209 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform"); 5210 5211 address stubAddr = StubRoutines::multiplyToLen(); 5212 if (stubAddr == NULL) { 5213 return false; // Intrinsic's stub is not implemented on this platform 5214 } 5215 const char* stubName = "multiplyToLen"; 5216 5217 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5218 5219 // no receiver because it is a static method 5220 Node* x = argument(0); 5221 Node* xlen = argument(1); 5222 Node* y = argument(2); 5223 Node* ylen = argument(3); 5224 Node* z = argument(4); 5225 5226 const Type* x_type = x->Value(&_gvn); 5227 const Type* y_type = y->Value(&_gvn); 5228 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5229 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5230 if (top_x == NULL || top_x->klass() == NULL || 5231 top_y == NULL || top_y->klass() == NULL) { 5232 // failed array check 5233 return false; 5234 } 5235 5236 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5237 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5238 if (x_elem != T_INT || y_elem != T_INT) { 5239 return false; 5240 } 5241 5242 // Set the original stack and the reexecute bit for the interpreter to reexecute 5243 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5244 // on the return from z array allocation in runtime. 5245 { PreserveReexecuteState preexecs(this); 5246 jvms()->set_should_reexecute(true); 5247 5248 Node* x_start = array_element_address(x, intcon(0), x_elem); 5249 Node* y_start = array_element_address(y, intcon(0), y_elem); 5250 // 'x_start' points to x array + scaled xlen 5251 // 'y_start' points to y array + scaled ylen 5252 5253 // Allocate the result array 5254 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5255 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5256 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5257 5258 IdealKit ideal(this); 5259 5260#define __ ideal. 5261 Node* one = __ ConI(1); 5262 Node* zero = __ ConI(0); 5263 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5264 __ set(need_alloc, zero); 5265 __ set(z_alloc, z); 5266 __ if_then(z, BoolTest::eq, null()); { 5267 __ increment (need_alloc, one); 5268 } __ else_(); { 5269 // Update graphKit memory and control from IdealKit. 5270 sync_kit(ideal); 5271 Node* zlen_arg = load_array_length(z); 5272 // Update IdealKit memory and control from graphKit. 5273 __ sync_kit(this); 5274 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5275 __ increment (need_alloc, one); 5276 } __ end_if(); 5277 } __ end_if(); 5278 5279 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5280 // Update graphKit memory and control from IdealKit. 5281 sync_kit(ideal); 5282 Node * narr = new_array(klass_node, zlen, 1); 5283 // Update IdealKit memory and control from graphKit. 5284 __ sync_kit(this); 5285 __ set(z_alloc, narr); 5286 } __ end_if(); 5287 5288 sync_kit(ideal); 5289 z = __ value(z_alloc); 5290 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5291 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5292 // Final sync IdealKit and GraphKit. 5293 final_sync(ideal); 5294#undef __ 5295 5296 Node* z_start = array_element_address(z, intcon(0), T_INT); 5297 5298 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5299 OptoRuntime::multiplyToLen_Type(), 5300 stubAddr, stubName, TypePtr::BOTTOM, 5301 x_start, xlen, y_start, ylen, z_start, zlen); 5302 } // original reexecute is set back here 5303 5304 C->set_has_split_ifs(true); // Has chance for split-if optimization 5305 set_result(z); 5306 return true; 5307} 5308 5309//-------------inline_squareToLen------------------------------------ 5310bool LibraryCallKit::inline_squareToLen() { 5311 assert(UseSquareToLenIntrinsic, "not implemented on this platform"); 5312 5313 address stubAddr = StubRoutines::squareToLen(); 5314 if (stubAddr == NULL) { 5315 return false; // Intrinsic's stub is not implemented on this platform 5316 } 5317 const char* stubName = "squareToLen"; 5318 5319 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5320 5321 Node* x = argument(0); 5322 Node* len = argument(1); 5323 Node* z = argument(2); 5324 Node* zlen = argument(3); 5325 5326 const Type* x_type = x->Value(&_gvn); 5327 const Type* z_type = z->Value(&_gvn); 5328 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5329 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5330 if (top_x == NULL || top_x->klass() == NULL || 5331 top_z == NULL || top_z->klass() == NULL) { 5332 // failed array check 5333 return false; 5334 } 5335 5336 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5337 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5338 if (x_elem != T_INT || z_elem != T_INT) { 5339 return false; 5340 } 5341 5342 5343 Node* x_start = array_element_address(x, intcon(0), x_elem); 5344 Node* z_start = array_element_address(z, intcon(0), z_elem); 5345 5346 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5347 OptoRuntime::squareToLen_Type(), 5348 stubAddr, stubName, TypePtr::BOTTOM, 5349 x_start, len, z_start, zlen); 5350 5351 set_result(z); 5352 return true; 5353} 5354 5355//-------------inline_mulAdd------------------------------------------ 5356bool LibraryCallKit::inline_mulAdd() { 5357 assert(UseMulAddIntrinsic, "not implemented on this platform"); 5358 5359 address stubAddr = StubRoutines::mulAdd(); 5360 if (stubAddr == NULL) { 5361 return false; // Intrinsic's stub is not implemented on this platform 5362 } 5363 const char* stubName = "mulAdd"; 5364 5365 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5366 5367 Node* out = argument(0); 5368 Node* in = argument(1); 5369 Node* offset = argument(2); 5370 Node* len = argument(3); 5371 Node* k = argument(4); 5372 5373 const Type* out_type = out->Value(&_gvn); 5374 const Type* in_type = in->Value(&_gvn); 5375 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5376 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5377 if (top_out == NULL || top_out->klass() == NULL || 5378 top_in == NULL || top_in->klass() == NULL) { 5379 // failed array check 5380 return false; 5381 } 5382 5383 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5384 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5385 if (out_elem != T_INT || in_elem != T_INT) { 5386 return false; 5387 } 5388 5389 Node* outlen = load_array_length(out); 5390 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5391 Node* out_start = array_element_address(out, intcon(0), out_elem); 5392 Node* in_start = array_element_address(in, intcon(0), in_elem); 5393 5394 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5395 OptoRuntime::mulAdd_Type(), 5396 stubAddr, stubName, TypePtr::BOTTOM, 5397 out_start,in_start, new_offset, len, k); 5398 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5399 set_result(result); 5400 return true; 5401} 5402 5403//-------------inline_montgomeryMultiply----------------------------------- 5404bool LibraryCallKit::inline_montgomeryMultiply() { 5405 address stubAddr = StubRoutines::montgomeryMultiply(); 5406 if (stubAddr == NULL) { 5407 return false; // Intrinsic's stub is not implemented on this platform 5408 } 5409 5410 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5411 const char* stubName = "montgomery_square"; 5412 5413 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5414 5415 Node* a = argument(0); 5416 Node* b = argument(1); 5417 Node* n = argument(2); 5418 Node* len = argument(3); 5419 Node* inv = argument(4); 5420 Node* m = argument(6); 5421 5422 const Type* a_type = a->Value(&_gvn); 5423 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5424 const Type* b_type = b->Value(&_gvn); 5425 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5426 const Type* n_type = a->Value(&_gvn); 5427 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5428 const Type* m_type = a->Value(&_gvn); 5429 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5430 if (top_a == NULL || top_a->klass() == NULL || 5431 top_b == NULL || top_b->klass() == NULL || 5432 top_n == NULL || top_n->klass() == NULL || 5433 top_m == NULL || top_m->klass() == NULL) { 5434 // failed array check 5435 return false; 5436 } 5437 5438 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5439 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5440 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5441 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5442 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5443 return false; 5444 } 5445 5446 // Make the call 5447 { 5448 Node* a_start = array_element_address(a, intcon(0), a_elem); 5449 Node* b_start = array_element_address(b, intcon(0), b_elem); 5450 Node* n_start = array_element_address(n, intcon(0), n_elem); 5451 Node* m_start = array_element_address(m, intcon(0), m_elem); 5452 5453 Node* call = make_runtime_call(RC_LEAF, 5454 OptoRuntime::montgomeryMultiply_Type(), 5455 stubAddr, stubName, TypePtr::BOTTOM, 5456 a_start, b_start, n_start, len, inv, top(), 5457 m_start); 5458 set_result(m); 5459 } 5460 5461 return true; 5462} 5463 5464bool LibraryCallKit::inline_montgomerySquare() { 5465 address stubAddr = StubRoutines::montgomerySquare(); 5466 if (stubAddr == NULL) { 5467 return false; // Intrinsic's stub is not implemented on this platform 5468 } 5469 5470 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5471 const char* stubName = "montgomery_square"; 5472 5473 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5474 5475 Node* a = argument(0); 5476 Node* n = argument(1); 5477 Node* len = argument(2); 5478 Node* inv = argument(3); 5479 Node* m = argument(5); 5480 5481 const Type* a_type = a->Value(&_gvn); 5482 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5483 const Type* n_type = a->Value(&_gvn); 5484 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5485 const Type* m_type = a->Value(&_gvn); 5486 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5487 if (top_a == NULL || top_a->klass() == NULL || 5488 top_n == NULL || top_n->klass() == NULL || 5489 top_m == NULL || top_m->klass() == NULL) { 5490 // failed array check 5491 return false; 5492 } 5493 5494 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5495 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5496 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5497 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5498 return false; 5499 } 5500 5501 // Make the call 5502 { 5503 Node* a_start = array_element_address(a, intcon(0), a_elem); 5504 Node* n_start = array_element_address(n, intcon(0), n_elem); 5505 Node* m_start = array_element_address(m, intcon(0), m_elem); 5506 5507 Node* call = make_runtime_call(RC_LEAF, 5508 OptoRuntime::montgomerySquare_Type(), 5509 stubAddr, stubName, TypePtr::BOTTOM, 5510 a_start, n_start, len, inv, top(), 5511 m_start); 5512 set_result(m); 5513 } 5514 5515 return true; 5516} 5517 5518//-------------inline_vectorizedMismatch------------------------------ 5519bool LibraryCallKit::inline_vectorizedMismatch() { 5520 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform"); 5521 5522 address stubAddr = StubRoutines::vectorizedMismatch(); 5523 if (stubAddr == NULL) { 5524 return false; // Intrinsic's stub is not implemented on this platform 5525 } 5526 const char* stubName = "vectorizedMismatch"; 5527 int size_l = callee()->signature()->size(); 5528 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters"); 5529 5530 Node* obja = argument(0); 5531 Node* aoffset = argument(1); 5532 Node* objb = argument(3); 5533 Node* boffset = argument(4); 5534 Node* length = argument(6); 5535 Node* scale = argument(7); 5536 5537 const Type* a_type = obja->Value(&_gvn); 5538 const Type* b_type = objb->Value(&_gvn); 5539 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5540 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5541 if (top_a == NULL || top_a->klass() == NULL || 5542 top_b == NULL || top_b->klass() == NULL) { 5543 // failed array check 5544 return false; 5545 } 5546 5547 Node* call; 5548 jvms()->set_should_reexecute(true); 5549 5550 Node* obja_adr = make_unsafe_address(obja, aoffset); 5551 Node* objb_adr = make_unsafe_address(objb, boffset); 5552 5553 call = make_runtime_call(RC_LEAF, 5554 OptoRuntime::vectorizedMismatch_Type(), 5555 stubAddr, stubName, TypePtr::BOTTOM, 5556 obja_adr, objb_adr, length, scale); 5557 5558 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5559 set_result(result); 5560 return true; 5561} 5562 5563/** 5564 * Calculate CRC32 for byte. 5565 * int java.util.zip.CRC32.update(int crc, int b) 5566 */ 5567bool LibraryCallKit::inline_updateCRC32() { 5568 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5569 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5570 // no receiver since it is static method 5571 Node* crc = argument(0); // type: int 5572 Node* b = argument(1); // type: int 5573 5574 /* 5575 * int c = ~ crc; 5576 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5577 * b = b ^ (c >>> 8); 5578 * crc = ~b; 5579 */ 5580 5581 Node* M1 = intcon(-1); 5582 crc = _gvn.transform(new XorINode(crc, M1)); 5583 Node* result = _gvn.transform(new XorINode(crc, b)); 5584 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5585 5586 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5587 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5588 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5589 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5590 5591 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5592 result = _gvn.transform(new XorINode(crc, result)); 5593 result = _gvn.transform(new XorINode(result, M1)); 5594 set_result(result); 5595 return true; 5596} 5597 5598/** 5599 * Calculate CRC32 for byte[] array. 5600 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5601 */ 5602bool LibraryCallKit::inline_updateBytesCRC32() { 5603 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5604 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5605 // no receiver since it is static method 5606 Node* crc = argument(0); // type: int 5607 Node* src = argument(1); // type: oop 5608 Node* offset = argument(2); // type: int 5609 Node* length = argument(3); // type: int 5610 5611 const Type* src_type = src->Value(&_gvn); 5612 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5613 if (top_src == NULL || top_src->klass() == NULL) { 5614 // failed array check 5615 return false; 5616 } 5617 5618 // Figure out the size and type of the elements we will be copying. 5619 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5620 if (src_elem != T_BYTE) { 5621 return false; 5622 } 5623 5624 // 'src_start' points to src array + scaled offset 5625 Node* src_start = array_element_address(src, offset, src_elem); 5626 5627 // We assume that range check is done by caller. 5628 // TODO: generate range check (offset+length < src.length) in debug VM. 5629 5630 // Call the stub. 5631 address stubAddr = StubRoutines::updateBytesCRC32(); 5632 const char *stubName = "updateBytesCRC32"; 5633 5634 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5635 stubAddr, stubName, TypePtr::BOTTOM, 5636 crc, src_start, length); 5637 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5638 set_result(result); 5639 return true; 5640} 5641 5642/** 5643 * Calculate CRC32 for ByteBuffer. 5644 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5645 */ 5646bool LibraryCallKit::inline_updateByteBufferCRC32() { 5647 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5648 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5649 // no receiver since it is static method 5650 Node* crc = argument(0); // type: int 5651 Node* src = argument(1); // type: long 5652 Node* offset = argument(3); // type: int 5653 Node* length = argument(4); // type: int 5654 5655 src = ConvL2X(src); // adjust Java long to machine word 5656 Node* base = _gvn.transform(new CastX2PNode(src)); 5657 offset = ConvI2X(offset); 5658 5659 // 'src_start' points to src array + scaled offset 5660 Node* src_start = basic_plus_adr(top(), base, offset); 5661 5662 // Call the stub. 5663 address stubAddr = StubRoutines::updateBytesCRC32(); 5664 const char *stubName = "updateBytesCRC32"; 5665 5666 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5667 stubAddr, stubName, TypePtr::BOTTOM, 5668 crc, src_start, length); 5669 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5670 set_result(result); 5671 return true; 5672} 5673 5674//------------------------------get_table_from_crc32c_class----------------------- 5675Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5676 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5677 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5678 5679 return table; 5680} 5681 5682//------------------------------inline_updateBytesCRC32C----------------------- 5683// 5684// Calculate CRC32C for byte[] array. 5685// int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5686// 5687bool LibraryCallKit::inline_updateBytesCRC32C() { 5688 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5689 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5690 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5691 // no receiver since it is a static method 5692 Node* crc = argument(0); // type: int 5693 Node* src = argument(1); // type: oop 5694 Node* offset = argument(2); // type: int 5695 Node* end = argument(3); // type: int 5696 5697 Node* length = _gvn.transform(new SubINode(end, offset)); 5698 5699 const Type* src_type = src->Value(&_gvn); 5700 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5701 if (top_src == NULL || top_src->klass() == NULL) { 5702 // failed array check 5703 return false; 5704 } 5705 5706 // Figure out the size and type of the elements we will be copying. 5707 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5708 if (src_elem != T_BYTE) { 5709 return false; 5710 } 5711 5712 // 'src_start' points to src array + scaled offset 5713 Node* src_start = array_element_address(src, offset, src_elem); 5714 5715 // static final int[] byteTable in class CRC32C 5716 Node* table = get_table_from_crc32c_class(callee()->holder()); 5717 Node* table_start = array_element_address(table, intcon(0), T_INT); 5718 5719 // We assume that range check is done by caller. 5720 // TODO: generate range check (offset+length < src.length) in debug VM. 5721 5722 // Call the stub. 5723 address stubAddr = StubRoutines::updateBytesCRC32C(); 5724 const char *stubName = "updateBytesCRC32C"; 5725 5726 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5727 stubAddr, stubName, TypePtr::BOTTOM, 5728 crc, src_start, length, table_start); 5729 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5730 set_result(result); 5731 return true; 5732} 5733 5734//------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5735// 5736// Calculate CRC32C for DirectByteBuffer. 5737// int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5738// 5739bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5740 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5741 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5742 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5743 // no receiver since it is a static method 5744 Node* crc = argument(0); // type: int 5745 Node* src = argument(1); // type: long 5746 Node* offset = argument(3); // type: int 5747 Node* end = argument(4); // type: int 5748 5749 Node* length = _gvn.transform(new SubINode(end, offset)); 5750 5751 src = ConvL2X(src); // adjust Java long to machine word 5752 Node* base = _gvn.transform(new CastX2PNode(src)); 5753 offset = ConvI2X(offset); 5754 5755 // 'src_start' points to src array + scaled offset 5756 Node* src_start = basic_plus_adr(top(), base, offset); 5757 5758 // static final int[] byteTable in class CRC32C 5759 Node* table = get_table_from_crc32c_class(callee()->holder()); 5760 Node* table_start = array_element_address(table, intcon(0), T_INT); 5761 5762 // Call the stub. 5763 address stubAddr = StubRoutines::updateBytesCRC32C(); 5764 const char *stubName = "updateBytesCRC32C"; 5765 5766 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5767 stubAddr, stubName, TypePtr::BOTTOM, 5768 crc, src_start, length, table_start); 5769 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5770 set_result(result); 5771 return true; 5772} 5773 5774//------------------------------inline_updateBytesAdler32---------------------- 5775// 5776// Calculate Adler32 checksum for byte[] array. 5777// int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len) 5778// 5779bool LibraryCallKit::inline_updateBytesAdler32() { 5780 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5781 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5782 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5783 // no receiver since it is static method 5784 Node* crc = argument(0); // type: int 5785 Node* src = argument(1); // type: oop 5786 Node* offset = argument(2); // type: int 5787 Node* length = argument(3); // type: int 5788 5789 const Type* src_type = src->Value(&_gvn); 5790 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5791 if (top_src == NULL || top_src->klass() == NULL) { 5792 // failed array check 5793 return false; 5794 } 5795 5796 // Figure out the size and type of the elements we will be copying. 5797 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5798 if (src_elem != T_BYTE) { 5799 return false; 5800 } 5801 5802 // 'src_start' points to src array + scaled offset 5803 Node* src_start = array_element_address(src, offset, src_elem); 5804 5805 // We assume that range check is done by caller. 5806 // TODO: generate range check (offset+length < src.length) in debug VM. 5807 5808 // Call the stub. 5809 address stubAddr = StubRoutines::updateBytesAdler32(); 5810 const char *stubName = "updateBytesAdler32"; 5811 5812 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5813 stubAddr, stubName, TypePtr::BOTTOM, 5814 crc, src_start, length); 5815 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5816 set_result(result); 5817 return true; 5818} 5819 5820//------------------------------inline_updateByteBufferAdler32--------------- 5821// 5822// Calculate Adler32 checksum for DirectByteBuffer. 5823// int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len) 5824// 5825bool LibraryCallKit::inline_updateByteBufferAdler32() { 5826 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one 5827 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5828 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded"); 5829 // no receiver since it is static method 5830 Node* crc = argument(0); // type: int 5831 Node* src = argument(1); // type: long 5832 Node* offset = argument(3); // type: int 5833 Node* length = argument(4); // type: int 5834 5835 src = ConvL2X(src); // adjust Java long to machine word 5836 Node* base = _gvn.transform(new CastX2PNode(src)); 5837 offset = ConvI2X(offset); 5838 5839 // 'src_start' points to src array + scaled offset 5840 Node* src_start = basic_plus_adr(top(), base, offset); 5841 5842 // Call the stub. 5843 address stubAddr = StubRoutines::updateBytesAdler32(); 5844 const char *stubName = "updateBytesAdler32"; 5845 5846 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(), 5847 stubAddr, stubName, TypePtr::BOTTOM, 5848 crc, src_start, length); 5849 5850 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5851 set_result(result); 5852 return true; 5853} 5854 5855//----------------------------inline_reference_get---------------------------- 5856// public T java.lang.ref.Reference.get(); 5857bool LibraryCallKit::inline_reference_get() { 5858 const int referent_offset = java_lang_ref_Reference::referent_offset; 5859 guarantee(referent_offset > 0, "should have already been set"); 5860 5861 // Get the argument: 5862 Node* reference_obj = null_check_receiver(); 5863 if (stopped()) return true; 5864 5865 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5866 5867 ciInstanceKlass* klass = env()->Object_klass(); 5868 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5869 5870 Node* no_ctrl = NULL; 5871 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5872 5873 // Use the pre-barrier to record the value in the referent field 5874 pre_barrier(false /* do_load */, 5875 control(), 5876 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5877 result /* pre_val */, 5878 T_OBJECT); 5879 5880 // Add memory barrier to prevent commoning reads from this field 5881 // across safepoint since GC can change its value. 5882 insert_mem_bar(Op_MemBarCPUOrder); 5883 5884 set_result(result); 5885 return true; 5886} 5887 5888 5889Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5890 bool is_exact=true, bool is_static=false, 5891 ciInstanceKlass * fromKls=NULL) { 5892 if (fromKls == NULL) { 5893 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5894 assert(tinst != NULL, "obj is null"); 5895 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5896 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5897 fromKls = tinst->klass()->as_instance_klass(); 5898 } else { 5899 assert(is_static, "only for static field access"); 5900 } 5901 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5902 ciSymbol::make(fieldTypeString), 5903 is_static); 5904 5905 assert (field != NULL, "undefined field"); 5906 if (field == NULL) return (Node *) NULL; 5907 5908 if (is_static) { 5909 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5910 fromObj = makecon(tip); 5911 } 5912 5913 // Next code copied from Parse::do_get_xxx(): 5914 5915 // Compute address and memory type. 5916 int offset = field->offset_in_bytes(); 5917 bool is_vol = field->is_volatile(); 5918 ciType* field_klass = field->type(); 5919 assert(field_klass->is_loaded(), "should be loaded"); 5920 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5921 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5922 BasicType bt = field->layout_type(); 5923 5924 // Build the resultant type of the load 5925 const Type *type; 5926 if (bt == T_OBJECT) { 5927 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5928 } else { 5929 type = Type::get_const_basic_type(bt); 5930 } 5931 5932 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5933 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5934 } 5935 // Build the load. 5936 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5937 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5938 // If reference is volatile, prevent following memory ops from 5939 // floating up past the volatile read. Also prevents commoning 5940 // another volatile read. 5941 if (is_vol) { 5942 // Memory barrier includes bogus read of value to force load BEFORE membar 5943 insert_mem_bar(Op_MemBarAcquire, loadedField); 5944 } 5945 return loadedField; 5946} 5947 5948Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5949 bool is_exact = true, bool is_static = false, 5950 ciInstanceKlass * fromKls = NULL) { 5951 if (fromKls == NULL) { 5952 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5953 assert(tinst != NULL, "obj is null"); 5954 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5955 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5956 fromKls = tinst->klass()->as_instance_klass(); 5957 } 5958 else { 5959 assert(is_static, "only for static field access"); 5960 } 5961 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5962 ciSymbol::make(fieldTypeString), 5963 is_static); 5964 5965 assert(field != NULL, "undefined field"); 5966 assert(!field->is_volatile(), "not defined for volatile fields"); 5967 5968 if (is_static) { 5969 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5970 fromObj = makecon(tip); 5971 } 5972 5973 // Next code copied from Parse::do_get_xxx(): 5974 5975 // Compute address and memory type. 5976 int offset = field->offset_in_bytes(); 5977 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5978 5979 return adr; 5980} 5981 5982//------------------------------inline_aescrypt_Block----------------------- 5983bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5984 address stubAddr = NULL; 5985 const char *stubName; 5986 assert(UseAES, "need AES instruction support"); 5987 5988 switch(id) { 5989 case vmIntrinsics::_aescrypt_encryptBlock: 5990 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5991 stubName = "aescrypt_encryptBlock"; 5992 break; 5993 case vmIntrinsics::_aescrypt_decryptBlock: 5994 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5995 stubName = "aescrypt_decryptBlock"; 5996 break; 5997 } 5998 if (stubAddr == NULL) return false; 5999 6000 Node* aescrypt_object = argument(0); 6001 Node* src = argument(1); 6002 Node* src_offset = argument(2); 6003 Node* dest = argument(3); 6004 Node* dest_offset = argument(4); 6005 6006 // (1) src and dest are arrays. 6007 const Type* src_type = src->Value(&_gvn); 6008 const Type* dest_type = dest->Value(&_gvn); 6009 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6010 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6011 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6012 6013 // for the quick and dirty code we will skip all the checks. 6014 // we are just trying to get the call to be generated. 6015 Node* src_start = src; 6016 Node* dest_start = dest; 6017 if (src_offset != NULL || dest_offset != NULL) { 6018 assert(src_offset != NULL && dest_offset != NULL, ""); 6019 src_start = array_element_address(src, src_offset, T_BYTE); 6020 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6021 } 6022 6023 // now need to get the start of its expanded key array 6024 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6025 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6026 if (k_start == NULL) return false; 6027 6028 if (Matcher::pass_original_key_for_aes()) { 6029 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6030 // compatibility issues between Java key expansion and SPARC crypto instructions 6031 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6032 if (original_k_start == NULL) return false; 6033 6034 // Call the stub. 6035 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6036 stubAddr, stubName, TypePtr::BOTTOM, 6037 src_start, dest_start, k_start, original_k_start); 6038 } else { 6039 // Call the stub. 6040 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6041 stubAddr, stubName, TypePtr::BOTTOM, 6042 src_start, dest_start, k_start); 6043 } 6044 6045 return true; 6046} 6047 6048//------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 6049bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 6050 address stubAddr = NULL; 6051 const char *stubName = NULL; 6052 6053 assert(UseAES, "need AES instruction support"); 6054 6055 switch(id) { 6056 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6057 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6058 stubName = "cipherBlockChaining_encryptAESCrypt"; 6059 break; 6060 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6061 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6062 stubName = "cipherBlockChaining_decryptAESCrypt"; 6063 break; 6064 } 6065 if (stubAddr == NULL) return false; 6066 6067 Node* cipherBlockChaining_object = argument(0); 6068 Node* src = argument(1); 6069 Node* src_offset = argument(2); 6070 Node* len = argument(3); 6071 Node* dest = argument(4); 6072 Node* dest_offset = argument(5); 6073 6074 // (1) src and dest are arrays. 6075 const Type* src_type = src->Value(&_gvn); 6076 const Type* dest_type = dest->Value(&_gvn); 6077 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6078 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6079 assert (top_src != NULL && top_src->klass() != NULL 6080 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6081 6082 // checks are the responsibility of the caller 6083 Node* src_start = src; 6084 Node* dest_start = dest; 6085 if (src_offset != NULL || dest_offset != NULL) { 6086 assert(src_offset != NULL && dest_offset != NULL, ""); 6087 src_start = array_element_address(src, src_offset, T_BYTE); 6088 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6089 } 6090 6091 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6092 // (because of the predicated logic executed earlier). 6093 // so we cast it here safely. 6094 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6095 6096 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6097 if (embeddedCipherObj == NULL) return false; 6098 6099 // cast it to what we know it will be at runtime 6100 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6101 assert(tinst != NULL, "CBC obj is null"); 6102 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6103 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6104 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6105 6106 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6107 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6108 const TypeOopPtr* xtype = aklass->as_instance_type(); 6109 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6110 aescrypt_object = _gvn.transform(aescrypt_object); 6111 6112 // we need to get the start of the aescrypt_object's expanded key array 6113 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6114 if (k_start == NULL) return false; 6115 6116 // similarly, get the start address of the r vector 6117 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6118 if (objRvec == NULL) return false; 6119 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6120 6121 Node* cbcCrypt; 6122 if (Matcher::pass_original_key_for_aes()) { 6123 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6124 // compatibility issues between Java key expansion and SPARC crypto instructions 6125 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6126 if (original_k_start == NULL) return false; 6127 6128 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6129 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6130 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6131 stubAddr, stubName, TypePtr::BOTTOM, 6132 src_start, dest_start, k_start, r_start, len, original_k_start); 6133 } else { 6134 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6135 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6136 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6137 stubAddr, stubName, TypePtr::BOTTOM, 6138 src_start, dest_start, k_start, r_start, len); 6139 } 6140 6141 // return cipher length (int) 6142 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6143 set_result(retvalue); 6144 return true; 6145} 6146 6147//------------------------------inline_counterMode_AESCrypt----------------------- 6148bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) { 6149 assert(UseAES, "need AES instruction support"); 6150 if (!UseAESCTRIntrinsics) return false; 6151 6152 address stubAddr = NULL; 6153 const char *stubName = NULL; 6154 if (id == vmIntrinsics::_counterMode_AESCrypt) { 6155 stubAddr = StubRoutines::counterMode_AESCrypt(); 6156 stubName = "counterMode_AESCrypt"; 6157 } 6158 if (stubAddr == NULL) return false; 6159 6160 Node* counterMode_object = argument(0); 6161 Node* src = argument(1); 6162 Node* src_offset = argument(2); 6163 Node* len = argument(3); 6164 Node* dest = argument(4); 6165 Node* dest_offset = argument(5); 6166 6167 // (1) src and dest are arrays. 6168 const Type* src_type = src->Value(&_gvn); 6169 const Type* dest_type = dest->Value(&_gvn); 6170 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6171 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6172 assert(top_src != NULL && top_src->klass() != NULL && 6173 top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6174 6175 // checks are the responsibility of the caller 6176 Node* src_start = src; 6177 Node* dest_start = dest; 6178 if (src_offset != NULL || dest_offset != NULL) { 6179 assert(src_offset != NULL && dest_offset != NULL, ""); 6180 src_start = array_element_address(src, src_offset, T_BYTE); 6181 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6182 } 6183 6184 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6185 // (because of the predicated logic executed earlier). 6186 // so we cast it here safely. 6187 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6188 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6189 if (embeddedCipherObj == NULL) return false; 6190 // cast it to what we know it will be at runtime 6191 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr(); 6192 assert(tinst != NULL, "CTR obj is null"); 6193 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded"); 6194 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6195 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6196 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6197 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6198 const TypeOopPtr* xtype = aklass->as_instance_type(); 6199 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6200 aescrypt_object = _gvn.transform(aescrypt_object); 6201 // we need to get the start of the aescrypt_object's expanded key array 6202 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6203 if (k_start == NULL) return false; 6204 // similarly, get the start address of the r vector 6205 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false); 6206 if (obj_counter == NULL) return false; 6207 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE); 6208 6209 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false); 6210 if (saved_encCounter == NULL) return false; 6211 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE); 6212 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false); 6213 6214 Node* ctrCrypt; 6215 if (Matcher::pass_original_key_for_aes()) { 6216 // no SPARC version for AES/CTR intrinsics now. 6217 return false; 6218 } 6219 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6220 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6221 OptoRuntime::counterMode_aescrypt_Type(), 6222 stubAddr, stubName, TypePtr::BOTTOM, 6223 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used); 6224 6225 // return cipher length (int) 6226 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms)); 6227 set_result(retvalue); 6228 return true; 6229} 6230 6231//------------------------------get_key_start_from_aescrypt_object----------------------- 6232Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6233#ifdef PPC64 6234 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6235 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6236 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6237 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6238 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6239 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6240 if (objSessionK == NULL) { 6241 return (Node *) NULL; 6242 } 6243 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6244#else 6245 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6246#endif // PPC64 6247 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6248 if (objAESCryptKey == NULL) return (Node *) NULL; 6249 6250 // now have the array, need to get the start address of the K array 6251 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6252 return k_start; 6253} 6254 6255//------------------------------get_original_key_start_from_aescrypt_object----------------------- 6256Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6257 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6258 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6259 if (objAESCryptKey == NULL) return (Node *) NULL; 6260 6261 // now have the array, need to get the start address of the lastKey array 6262 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6263 return original_k_start; 6264} 6265 6266//----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6267// Return node representing slow path of predicate check. 6268// the pseudo code we want to emulate with this predicate is: 6269// for encryption: 6270// if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6271// for decryption: 6272// if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6273// note cipher==plain is more conservative than the original java code but that's OK 6274// 6275Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6276 // The receiver was checked for NULL already. 6277 Node* objCBC = argument(0); 6278 6279 // Load embeddedCipher field of CipherBlockChaining object. 6280 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6281 6282 // get AESCrypt klass for instanceOf check 6283 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6284 // will have same classloader as CipherBlockChaining object 6285 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6286 assert(tinst != NULL, "CBCobj is null"); 6287 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6288 6289 // we want to do an instanceof comparison against the AESCrypt class 6290 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6291 if (!klass_AESCrypt->is_loaded()) { 6292 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6293 Node* ctrl = control(); 6294 set_control(top()); // no regular fast path 6295 return ctrl; 6296 } 6297 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6298 6299 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6300 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6301 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6302 6303 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6304 6305 // for encryption, we are done 6306 if (!decrypting) 6307 return instof_false; // even if it is NULL 6308 6309 // for decryption, we need to add a further check to avoid 6310 // taking the intrinsic path when cipher and plain are the same 6311 // see the original java code for why. 6312 RegionNode* region = new RegionNode(3); 6313 region->init_req(1, instof_false); 6314 Node* src = argument(1); 6315 Node* dest = argument(4); 6316 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6317 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6318 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6319 region->init_req(2, src_dest_conjoint); 6320 6321 record_for_igvn(region); 6322 return _gvn.transform(region); 6323} 6324 6325//----------------------------inline_counterMode_AESCrypt_predicate---------------------------- 6326// Return node representing slow path of predicate check. 6327// the pseudo code we want to emulate with this predicate is: 6328// for encryption: 6329// if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6330// for decryption: 6331// if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6332// note cipher==plain is more conservative than the original java code but that's OK 6333// 6334 6335Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() { 6336 // The receiver was checked for NULL already. 6337 Node* objCTR = argument(0); 6338 6339 // Load embeddedCipher field of CipherBlockChaining object. 6340 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6341 6342 // get AESCrypt klass for instanceOf check 6343 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6344 // will have same classloader as CipherBlockChaining object 6345 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr(); 6346 assert(tinst != NULL, "CTRobj is null"); 6347 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded"); 6348 6349 // we want to do an instanceof comparison against the AESCrypt class 6350 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6351 if (!klass_AESCrypt->is_loaded()) { 6352 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6353 Node* ctrl = control(); 6354 set_control(top()); // no regular fast path 6355 return ctrl; 6356 } 6357 6358 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6359 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6360 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6361 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6362 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6363 6364 return instof_false; // even if it is NULL 6365} 6366 6367//------------------------------inline_ghash_processBlocks 6368bool LibraryCallKit::inline_ghash_processBlocks() { 6369 address stubAddr; 6370 const char *stubName; 6371 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6372 6373 stubAddr = StubRoutines::ghash_processBlocks(); 6374 stubName = "ghash_processBlocks"; 6375 6376 Node* data = argument(0); 6377 Node* offset = argument(1); 6378 Node* len = argument(2); 6379 Node* state = argument(3); 6380 Node* subkeyH = argument(4); 6381 6382 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6383 assert(state_start, "state is NULL"); 6384 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6385 assert(subkeyH_start, "subkeyH is NULL"); 6386 Node* data_start = array_element_address(data, offset, T_BYTE); 6387 assert(data_start, "data is NULL"); 6388 6389 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6390 OptoRuntime::ghash_processBlocks_Type(), 6391 stubAddr, stubName, TypePtr::BOTTOM, 6392 state_start, subkeyH_start, data_start, len); 6393 return true; 6394} 6395 6396//------------------------------inline_sha_implCompress----------------------- 6397// 6398// Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6399// void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6400// 6401// Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6402// void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6403// 6404// Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6405// void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6406// 6407bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6408 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6409 6410 Node* sha_obj = argument(0); 6411 Node* src = argument(1); // type oop 6412 Node* ofs = argument(2); // type int 6413 6414 const Type* src_type = src->Value(&_gvn); 6415 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6416 if (top_src == NULL || top_src->klass() == NULL) { 6417 // failed array check 6418 return false; 6419 } 6420 // Figure out the size and type of the elements we will be copying. 6421 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6422 if (src_elem != T_BYTE) { 6423 return false; 6424 } 6425 // 'src_start' points to src array + offset 6426 Node* src_start = array_element_address(src, ofs, src_elem); 6427 Node* state = NULL; 6428 address stubAddr; 6429 const char *stubName; 6430 6431 switch(id) { 6432 case vmIntrinsics::_sha_implCompress: 6433 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6434 state = get_state_from_sha_object(sha_obj); 6435 stubAddr = StubRoutines::sha1_implCompress(); 6436 stubName = "sha1_implCompress"; 6437 break; 6438 case vmIntrinsics::_sha2_implCompress: 6439 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6440 state = get_state_from_sha_object(sha_obj); 6441 stubAddr = StubRoutines::sha256_implCompress(); 6442 stubName = "sha256_implCompress"; 6443 break; 6444 case vmIntrinsics::_sha5_implCompress: 6445 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6446 state = get_state_from_sha5_object(sha_obj); 6447 stubAddr = StubRoutines::sha512_implCompress(); 6448 stubName = "sha512_implCompress"; 6449 break; 6450 default: 6451 fatal_unexpected_iid(id); 6452 return false; 6453 } 6454 if (state == NULL) return false; 6455 6456 // Call the stub. 6457 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6458 stubAddr, stubName, TypePtr::BOTTOM, 6459 src_start, state); 6460 6461 return true; 6462} 6463 6464//------------------------------inline_digestBase_implCompressMB----------------------- 6465// 6466// Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6467// int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6468// 6469bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6470 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6471 "need SHA1/SHA256/SHA512 instruction support"); 6472 assert((uint)predicate < 3, "sanity"); 6473 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6474 6475 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6476 Node* src = argument(1); // byte[] array 6477 Node* ofs = argument(2); // type int 6478 Node* limit = argument(3); // type int 6479 6480 const Type* src_type = src->Value(&_gvn); 6481 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6482 if (top_src == NULL || top_src->klass() == NULL) { 6483 // failed array check 6484 return false; 6485 } 6486 // Figure out the size and type of the elements we will be copying. 6487 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6488 if (src_elem != T_BYTE) { 6489 return false; 6490 } 6491 // 'src_start' points to src array + offset 6492 Node* src_start = array_element_address(src, ofs, src_elem); 6493 6494 const char* klass_SHA_name = NULL; 6495 const char* stub_name = NULL; 6496 address stub_addr = NULL; 6497 bool long_state = false; 6498 6499 switch (predicate) { 6500 case 0: 6501 if (UseSHA1Intrinsics) { 6502 klass_SHA_name = "sun/security/provider/SHA"; 6503 stub_name = "sha1_implCompressMB"; 6504 stub_addr = StubRoutines::sha1_implCompressMB(); 6505 } 6506 break; 6507 case 1: 6508 if (UseSHA256Intrinsics) { 6509 klass_SHA_name = "sun/security/provider/SHA2"; 6510 stub_name = "sha256_implCompressMB"; 6511 stub_addr = StubRoutines::sha256_implCompressMB(); 6512 } 6513 break; 6514 case 2: 6515 if (UseSHA512Intrinsics) { 6516 klass_SHA_name = "sun/security/provider/SHA5"; 6517 stub_name = "sha512_implCompressMB"; 6518 stub_addr = StubRoutines::sha512_implCompressMB(); 6519 long_state = true; 6520 } 6521 break; 6522 default: 6523 fatal("unknown SHA intrinsic predicate: %d", predicate); 6524 } 6525 if (klass_SHA_name != NULL) { 6526 // get DigestBase klass to lookup for SHA klass 6527 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6528 assert(tinst != NULL, "digestBase_obj is not instance???"); 6529 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6530 6531 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6532 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6533 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6534 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6535 } 6536 return false; 6537} 6538//------------------------------inline_sha_implCompressMB----------------------- 6539bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6540 bool long_state, address stubAddr, const char *stubName, 6541 Node* src_start, Node* ofs, Node* limit) { 6542 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6543 const TypeOopPtr* xtype = aklass->as_instance_type(); 6544 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6545 sha_obj = _gvn.transform(sha_obj); 6546 6547 Node* state; 6548 if (long_state) { 6549 state = get_state_from_sha5_object(sha_obj); 6550 } else { 6551 state = get_state_from_sha_object(sha_obj); 6552 } 6553 if (state == NULL) return false; 6554 6555 // Call the stub. 6556 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6557 OptoRuntime::digestBase_implCompressMB_Type(), 6558 stubAddr, stubName, TypePtr::BOTTOM, 6559 src_start, state, ofs, limit); 6560 // return ofs (int) 6561 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6562 set_result(result); 6563 6564 return true; 6565} 6566 6567//------------------------------get_state_from_sha_object----------------------- 6568Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6569 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6570 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6571 if (sha_state == NULL) return (Node *) NULL; 6572 6573 // now have the array, need to get the start address of the state array 6574 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6575 return state; 6576} 6577 6578//------------------------------get_state_from_sha5_object----------------------- 6579Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6580 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6581 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6582 if (sha_state == NULL) return (Node *) NULL; 6583 6584 // now have the array, need to get the start address of the state array 6585 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6586 return state; 6587} 6588 6589//----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6590// Return node representing slow path of predicate check. 6591// the pseudo code we want to emulate with this predicate is: 6592// if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6593// 6594Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6595 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6596 "need SHA1/SHA256/SHA512 instruction support"); 6597 assert((uint)predicate < 3, "sanity"); 6598 6599 // The receiver was checked for NULL already. 6600 Node* digestBaseObj = argument(0); 6601 6602 // get DigestBase klass for instanceOf check 6603 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6604 assert(tinst != NULL, "digestBaseObj is null"); 6605 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6606 6607 const char* klass_SHA_name = NULL; 6608 switch (predicate) { 6609 case 0: 6610 if (UseSHA1Intrinsics) { 6611 // we want to do an instanceof comparison against the SHA class 6612 klass_SHA_name = "sun/security/provider/SHA"; 6613 } 6614 break; 6615 case 1: 6616 if (UseSHA256Intrinsics) { 6617 // we want to do an instanceof comparison against the SHA2 class 6618 klass_SHA_name = "sun/security/provider/SHA2"; 6619 } 6620 break; 6621 case 2: 6622 if (UseSHA512Intrinsics) { 6623 // we want to do an instanceof comparison against the SHA5 class 6624 klass_SHA_name = "sun/security/provider/SHA5"; 6625 } 6626 break; 6627 default: 6628 fatal("unknown SHA intrinsic predicate: %d", predicate); 6629 } 6630 6631 ciKlass* klass_SHA = NULL; 6632 if (klass_SHA_name != NULL) { 6633 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6634 } 6635 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6636 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6637 Node* ctrl = control(); 6638 set_control(top()); // no intrinsic path 6639 return ctrl; 6640 } 6641 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6642 6643 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6644 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6645 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6646 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6647 6648 return instof_false; // even if it is NULL 6649} 6650 6651bool LibraryCallKit::inline_profileBoolean() { 6652 Node* counts = argument(1); 6653 const TypeAryPtr* ary = NULL; 6654 ciArray* aobj = NULL; 6655 if (counts->is_Con() 6656 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6657 && (aobj = ary->const_oop()->as_array()) != NULL 6658 && (aobj->length() == 2)) { 6659 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6660 jint false_cnt = aobj->element_value(0).as_int(); 6661 jint true_cnt = aobj->element_value(1).as_int(); 6662 6663 if (C->log() != NULL) { 6664 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6665 false_cnt, true_cnt); 6666 } 6667 6668 if (false_cnt + true_cnt == 0) { 6669 // According to profile, never executed. 6670 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6671 Deoptimization::Action_reinterpret); 6672 return true; 6673 } 6674 6675 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6676 // is a number of each value occurrences. 6677 Node* result = argument(0); 6678 if (false_cnt == 0 || true_cnt == 0) { 6679 // According to profile, one value has been never seen. 6680 int expected_val = (false_cnt == 0) ? 1 : 0; 6681 6682 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6683 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6684 6685 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6686 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6687 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6688 6689 { // Slow path: uncommon trap for never seen value and then reexecute 6690 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6691 // the value has been seen at least once. 6692 PreserveJVMState pjvms(this); 6693 PreserveReexecuteState preexecs(this); 6694 jvms()->set_should_reexecute(true); 6695 6696 set_control(slow_path); 6697 set_i_o(i_o()); 6698 6699 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6700 Deoptimization::Action_reinterpret); 6701 } 6702 // The guard for never seen value enables sharpening of the result and 6703 // returning a constant. It allows to eliminate branches on the same value 6704 // later on. 6705 set_control(fast_path); 6706 result = intcon(expected_val); 6707 } 6708 // Stop profiling. 6709 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6710 // By replacing method body with profile data (represented as ProfileBooleanNode 6711 // on IR level) we effectively disable profiling. 6712 // It enables full speed execution once optimized code is generated. 6713 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6714 C->record_for_igvn(profile); 6715 set_result(profile); 6716 return true; 6717 } else { 6718 // Continue profiling. 6719 // Profile data isn't available at the moment. So, execute method's bytecode version. 6720 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6721 // is compiled and counters aren't available since corresponding MethodHandle 6722 // isn't a compile-time constant. 6723 return false; 6724 } 6725} 6726 6727bool LibraryCallKit::inline_isCompileConstant() { 6728 Node* n = argument(0); 6729 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6730 return true; 6731} 6732