1 /* 2 * Copyright (c) 1999, 2012, 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 "classfile/systemDictionary.hpp" 27 #include "classfile/vmSymbols.hpp" 28 #include "compiler/compileBroker.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "oops/objArrayKlass.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/callGenerator.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/idealKit.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/parse.hpp" 37 #include "opto/runtime.hpp" 38 #include "opto/subnode.hpp" 39 #include "prims/nativeLookup.hpp" 40 #include "runtime/sharedRuntime.hpp" 41 42 class LibraryIntrinsic : public InlineCallGenerator { 43 // Extend the set of intrinsics known to the runtime: 44 public: 45 private: 46 bool _is_virtual; 47 vmIntrinsics::ID _intrinsic_id; 48 49 public: 50 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id) 51 : InlineCallGenerator(m), 52 _is_virtual(is_virtual), 53 _intrinsic_id(id) 54 { 55 } 56 virtual bool is_intrinsic() const { return true; } 57 virtual bool is_virtual() const { return _is_virtual; } 58 virtual JVMState* generate(JVMState* jvms); 59 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 60 }; 61 62 63 // Local helper class for LibraryIntrinsic: 64 class LibraryCallKit : public GraphKit { 65 private: 66 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 67 68 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 69 70 public: 71 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic) 72 : GraphKit(caller), 73 _intrinsic(intrinsic) 74 { 75 } 76 77 ciMethod* caller() const { return jvms()->method(); } 78 int bci() const { return jvms()->bci(); } 79 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 80 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 81 ciMethod* callee() const { return _intrinsic->method(); } 82 ciSignature* signature() const { return callee()->signature(); } 83 int arg_size() const { return callee()->arg_size(); } 84 85 bool try_to_inline(); 86 87 // Helper functions to inline natives 88 void push_result(RegionNode* region, PhiNode* value); 89 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 90 Node* generate_slow_guard(Node* test, RegionNode* region); 91 Node* generate_fair_guard(Node* test, RegionNode* region); 92 Node* generate_negative_guard(Node* index, RegionNode* region, 93 // resulting CastII of index: 94 Node* *pos_index = NULL); 95 Node* generate_nonpositive_guard(Node* index, bool never_negative, 96 // resulting CastII of index: 97 Node* *pos_index = NULL); 98 Node* generate_limit_guard(Node* offset, Node* subseq_length, 99 Node* array_length, 100 RegionNode* region); 101 Node* generate_current_thread(Node* &tls_output); 102 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 103 bool disjoint_bases, const char* &name, bool dest_uninitialized); 104 Node* load_mirror_from_klass(Node* klass); 105 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 106 int nargs, 107 RegionNode* region, int null_path, 108 int offset); 109 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs, 110 RegionNode* region, int null_path) { 111 int offset = java_lang_Class::klass_offset_in_bytes(); 112 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 113 region, null_path, 114 offset); 115 } 116 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 117 int nargs, 118 RegionNode* region, int null_path) { 119 int offset = java_lang_Class::array_klass_offset_in_bytes(); 120 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 121 region, null_path, 122 offset); 123 } 124 Node* generate_access_flags_guard(Node* kls, 125 int modifier_mask, int modifier_bits, 126 RegionNode* region); 127 Node* generate_interface_guard(Node* kls, RegionNode* region); 128 Node* generate_array_guard(Node* kls, RegionNode* region) { 129 return generate_array_guard_common(kls, region, false, false); 130 } 131 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 132 return generate_array_guard_common(kls, region, false, true); 133 } 134 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 135 return generate_array_guard_common(kls, region, true, false); 136 } 137 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 138 return generate_array_guard_common(kls, region, true, true); 139 } 140 Node* generate_array_guard_common(Node* kls, RegionNode* region, 141 bool obj_array, bool not_array); 142 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 143 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 144 bool is_virtual = false, bool is_static = false); 145 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 146 return generate_method_call(method_id, false, true); 147 } 148 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 149 return generate_method_call(method_id, true, false); 150 } 151 152 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 153 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 154 bool inline_string_compareTo(); 155 bool inline_string_indexOf(); 156 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 157 bool inline_string_equals(); 158 Node* pop_math_arg(); 159 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 160 bool inline_math_native(vmIntrinsics::ID id); 161 bool inline_trig(vmIntrinsics::ID id); 162 bool inline_trans(vmIntrinsics::ID id); 163 bool inline_abs(vmIntrinsics::ID id); 164 bool inline_sqrt(vmIntrinsics::ID id); 165 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 166 bool inline_pow(vmIntrinsics::ID id); 167 bool inline_exp(vmIntrinsics::ID id); 168 bool inline_min_max(vmIntrinsics::ID id); 169 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 170 // This returns Type::AnyPtr, RawPtr, or OopPtr. 171 int classify_unsafe_addr(Node* &base, Node* &offset); 172 Node* make_unsafe_address(Node* base, Node* offset); 173 // Helper for inline_unsafe_access. 174 // Generates the guards that check whether the result of 175 // Unsafe.getObject should be recorded in an SATB log buffer. 176 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, int nargs, bool need_mem_bar); 177 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 178 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 179 bool inline_unsafe_allocate(); 180 bool inline_unsafe_copyMemory(); 181 bool inline_native_currentThread(); 182 #ifdef TRACE_HAVE_INTRINSICS 183 bool inline_native_classID(); 184 bool inline_native_threadID(); 185 #endif 186 bool inline_native_time_funcs(address method, const char* funcName); 187 bool inline_native_isInterrupted(); 188 bool inline_native_Class_query(vmIntrinsics::ID id); 189 bool inline_native_subtype_check(); 190 191 bool inline_native_newArray(); 192 bool inline_native_getLength(); 193 bool inline_array_copyOf(bool is_copyOfRange); 194 bool inline_array_equals(); 195 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 196 bool inline_native_clone(bool is_virtual); 197 bool inline_native_Reflection_getCallerClass(); 198 bool is_method_invoke_or_aux_frame(JVMState* jvms); 199 // Helper function for inlining native object hash method 200 bool inline_native_hashcode(bool is_virtual, bool is_static); 201 bool inline_native_getClass(); 202 203 // Helper functions for inlining arraycopy 204 bool inline_arraycopy(); 205 void generate_arraycopy(const TypePtr* adr_type, 206 BasicType basic_elem_type, 207 Node* src, Node* src_offset, 208 Node* dest, Node* dest_offset, 209 Node* copy_length, 210 bool disjoint_bases = false, 211 bool length_never_negative = false, 212 RegionNode* slow_region = NULL); 213 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 214 RegionNode* slow_region); 215 void generate_clear_array(const TypePtr* adr_type, 216 Node* dest, 217 BasicType basic_elem_type, 218 Node* slice_off, 219 Node* slice_len, 220 Node* slice_end); 221 bool generate_block_arraycopy(const TypePtr* adr_type, 222 BasicType basic_elem_type, 223 AllocateNode* alloc, 224 Node* src, Node* src_offset, 225 Node* dest, Node* dest_offset, 226 Node* dest_size, bool dest_uninitialized); 227 void generate_slow_arraycopy(const TypePtr* adr_type, 228 Node* src, Node* src_offset, 229 Node* dest, Node* dest_offset, 230 Node* copy_length, bool dest_uninitialized); 231 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 232 Node* dest_elem_klass, 233 Node* src, Node* src_offset, 234 Node* dest, Node* dest_offset, 235 Node* copy_length, bool dest_uninitialized); 236 Node* generate_generic_arraycopy(const TypePtr* adr_type, 237 Node* src, Node* src_offset, 238 Node* dest, Node* dest_offset, 239 Node* copy_length, bool dest_uninitialized); 240 void generate_unchecked_arraycopy(const TypePtr* adr_type, 241 BasicType basic_elem_type, 242 bool disjoint_bases, 243 Node* src, Node* src_offset, 244 Node* dest, Node* dest_offset, 245 Node* copy_length, bool dest_uninitialized); 246 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 247 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 248 bool inline_unsafe_ordered_store(BasicType type); 249 bool inline_fp_conversions(vmIntrinsics::ID id); 250 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id); 251 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id); 252 bool inline_bitCount(vmIntrinsics::ID id); 253 bool inline_reverseBytes(vmIntrinsics::ID id); 254 255 bool inline_reference_get(); 256 }; 257 258 259 //---------------------------make_vm_intrinsic---------------------------- 260 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 261 vmIntrinsics::ID id = m->intrinsic_id(); 262 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 263 264 if (DisableIntrinsic[0] != '\0' 265 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 266 // disabled by a user request on the command line: 267 // example: -XX:DisableIntrinsic=_hashCode,_getClass 268 return NULL; 269 } 270 271 if (!m->is_loaded()) { 272 // do not attempt to inline unloaded methods 273 return NULL; 274 } 275 276 // Only a few intrinsics implement a virtual dispatch. 277 // They are expensive calls which are also frequently overridden. 278 if (is_virtual) { 279 switch (id) { 280 case vmIntrinsics::_hashCode: 281 case vmIntrinsics::_clone: 282 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 283 break; 284 default: 285 return NULL; 286 } 287 } 288 289 // -XX:-InlineNatives disables nearly all intrinsics: 290 if (!InlineNatives) { 291 switch (id) { 292 case vmIntrinsics::_indexOf: 293 case vmIntrinsics::_compareTo: 294 case vmIntrinsics::_equals: 295 case vmIntrinsics::_equalsC: 296 case vmIntrinsics::_getAndAddInt: 297 case vmIntrinsics::_getAndAddLong: 298 case vmIntrinsics::_getAndSetInt: 299 case vmIntrinsics::_getAndSetLong: 300 case vmIntrinsics::_getAndSetObject: 301 break; // InlineNatives does not control String.compareTo 302 case vmIntrinsics::_Reference_get: 303 break; // InlineNatives does not control Reference.get 304 default: 305 return NULL; 306 } 307 } 308 309 switch (id) { 310 case vmIntrinsics::_compareTo: 311 if (!SpecialStringCompareTo) return NULL; 312 break; 313 case vmIntrinsics::_indexOf: 314 if (!SpecialStringIndexOf) return NULL; 315 break; 316 case vmIntrinsics::_equals: 317 if (!SpecialStringEquals) return NULL; 318 break; 319 case vmIntrinsics::_equalsC: 320 if (!SpecialArraysEquals) return NULL; 321 break; 322 case vmIntrinsics::_arraycopy: 323 if (!InlineArrayCopy) return NULL; 324 break; 325 case vmIntrinsics::_copyMemory: 326 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 327 if (!InlineArrayCopy) return NULL; 328 break; 329 case vmIntrinsics::_hashCode: 330 if (!InlineObjectHash) return NULL; 331 break; 332 case vmIntrinsics::_clone: 333 case vmIntrinsics::_copyOf: 334 case vmIntrinsics::_copyOfRange: 335 if (!InlineObjectCopy) return NULL; 336 // These also use the arraycopy intrinsic mechanism: 337 if (!InlineArrayCopy) return NULL; 338 break; 339 case vmIntrinsics::_checkIndex: 340 // We do not intrinsify this. The optimizer does fine with it. 341 return NULL; 342 343 case vmIntrinsics::_getCallerClass: 344 if (!UseNewReflection) return NULL; 345 if (!InlineReflectionGetCallerClass) return NULL; 346 if (!JDK_Version::is_gte_jdk14x_version()) return NULL; 347 break; 348 349 case vmIntrinsics::_bitCount_i: 350 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 351 break; 352 353 case vmIntrinsics::_bitCount_l: 354 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 355 break; 356 357 case vmIntrinsics::_numberOfLeadingZeros_i: 358 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 359 break; 360 361 case vmIntrinsics::_numberOfLeadingZeros_l: 362 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 363 break; 364 365 case vmIntrinsics::_numberOfTrailingZeros_i: 366 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 367 break; 368 369 case vmIntrinsics::_numberOfTrailingZeros_l: 370 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 371 break; 372 373 case vmIntrinsics::_Reference_get: 374 // Use the intrinsic version of Reference.get() so that the value in 375 // the referent field can be registered by the G1 pre-barrier code. 376 // Also add memory barrier to prevent commoning reads from this field 377 // across safepoint since GC can change it value. 378 break; 379 380 case vmIntrinsics::_compareAndSwapObject: 381 #ifdef _LP64 382 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 383 #endif 384 break; 385 386 case vmIntrinsics::_compareAndSwapLong: 387 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 388 break; 389 390 case vmIntrinsics::_getAndAddInt: 391 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 392 break; 393 394 case vmIntrinsics::_getAndAddLong: 395 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 396 break; 397 398 case vmIntrinsics::_getAndSetInt: 399 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 400 break; 401 402 case vmIntrinsics::_getAndSetLong: 403 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 404 break; 405 406 case vmIntrinsics::_getAndSetObject: 407 #ifdef _LP64 408 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 409 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 410 break; 411 #else 412 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 413 break; 414 #endif 415 416 default: 417 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 418 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 419 break; 420 } 421 422 // -XX:-InlineClassNatives disables natives from the Class class. 423 // The flag applies to all reflective calls, notably Array.newArray 424 // (visible to Java programmers as Array.newInstance). 425 if (m->holder()->name() == ciSymbol::java_lang_Class() || 426 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 427 if (!InlineClassNatives) return NULL; 428 } 429 430 // -XX:-InlineThreadNatives disables natives from the Thread class. 431 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 432 if (!InlineThreadNatives) return NULL; 433 } 434 435 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 436 if (m->holder()->name() == ciSymbol::java_lang_Math() || 437 m->holder()->name() == ciSymbol::java_lang_Float() || 438 m->holder()->name() == ciSymbol::java_lang_Double()) { 439 if (!InlineMathNatives) return NULL; 440 } 441 442 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 443 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 444 if (!InlineUnsafeOps) return NULL; 445 } 446 447 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); 448 } 449 450 //----------------------register_library_intrinsics----------------------- 451 // Initialize this file's data structures, for each Compile instance. 452 void Compile::register_library_intrinsics() { 453 // Nothing to do here. 454 } 455 456 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 457 LibraryCallKit kit(jvms, this); 458 Compile* C = kit.C; 459 int nodes = C->unique(); 460 #ifndef PRODUCT 461 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 462 char buf[1000]; 463 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 464 tty->print_cr("Intrinsic %s", str); 465 } 466 #endif 467 468 if (kit.try_to_inline()) { 469 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 470 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 471 } 472 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 473 if (C->log()) { 474 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 475 vmIntrinsics::name_at(intrinsic_id()), 476 (is_virtual() ? " virtual='1'" : ""), 477 C->unique() - nodes); 478 } 479 return kit.transfer_exceptions_into_jvms(); 480 } 481 482 // The intrinsic bailed out 483 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 484 if (jvms->has_method()) { 485 // Not a root compile. 486 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 487 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg); 488 } else { 489 // Root compile 490 tty->print("Did not generate intrinsic %s%s at bci:%d in", 491 vmIntrinsics::name_at(intrinsic_id()), 492 (is_virtual() ? " (virtual)" : ""), kit.bci()); 493 } 494 } 495 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 496 return NULL; 497 } 498 499 bool LibraryCallKit::try_to_inline() { 500 // Handle symbolic names for otherwise undistinguished boolean switches: 501 const bool is_store = true; 502 const bool is_native_ptr = true; 503 const bool is_static = true; 504 505 if (!jvms()->has_method()) { 506 // Root JVMState has a null method. 507 assert(map()->memory()->Opcode() == Op_Parm, ""); 508 // Insert the memory aliasing node 509 set_all_memory(reset_memory()); 510 } 511 assert(merged_memory(), ""); 512 513 switch (intrinsic_id()) { 514 case vmIntrinsics::_hashCode: 515 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 516 case vmIntrinsics::_identityHashCode: 517 return inline_native_hashcode(/*!virtual*/ false, is_static); 518 case vmIntrinsics::_getClass: 519 return inline_native_getClass(); 520 521 case vmIntrinsics::_dsin: 522 case vmIntrinsics::_dcos: 523 case vmIntrinsics::_dtan: 524 case vmIntrinsics::_dabs: 525 case vmIntrinsics::_datan2: 526 case vmIntrinsics::_dsqrt: 527 case vmIntrinsics::_dexp: 528 case vmIntrinsics::_dlog: 529 case vmIntrinsics::_dlog10: 530 case vmIntrinsics::_dpow: 531 return inline_math_native(intrinsic_id()); 532 533 case vmIntrinsics::_min: 534 case vmIntrinsics::_max: 535 return inline_min_max(intrinsic_id()); 536 537 case vmIntrinsics::_arraycopy: 538 return inline_arraycopy(); 539 540 case vmIntrinsics::_compareTo: 541 return inline_string_compareTo(); 542 case vmIntrinsics::_indexOf: 543 return inline_string_indexOf(); 544 case vmIntrinsics::_equals: 545 return inline_string_equals(); 546 547 case vmIntrinsics::_getObject: 548 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); 549 case vmIntrinsics::_getBoolean: 550 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); 551 case vmIntrinsics::_getByte: 552 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); 553 case vmIntrinsics::_getShort: 554 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); 555 case vmIntrinsics::_getChar: 556 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); 557 case vmIntrinsics::_getInt: 558 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); 559 case vmIntrinsics::_getLong: 560 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); 561 case vmIntrinsics::_getFloat: 562 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); 563 case vmIntrinsics::_getDouble: 564 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); 565 566 case vmIntrinsics::_putObject: 567 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); 568 case vmIntrinsics::_putBoolean: 569 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); 570 case vmIntrinsics::_putByte: 571 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); 572 case vmIntrinsics::_putShort: 573 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); 574 case vmIntrinsics::_putChar: 575 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); 576 case vmIntrinsics::_putInt: 577 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); 578 case vmIntrinsics::_putLong: 579 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); 580 case vmIntrinsics::_putFloat: 581 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); 582 case vmIntrinsics::_putDouble: 583 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); 584 585 case vmIntrinsics::_getByte_raw: 586 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); 587 case vmIntrinsics::_getShort_raw: 588 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); 589 case vmIntrinsics::_getChar_raw: 590 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); 591 case vmIntrinsics::_getInt_raw: 592 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); 593 case vmIntrinsics::_getLong_raw: 594 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); 595 case vmIntrinsics::_getFloat_raw: 596 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); 597 case vmIntrinsics::_getDouble_raw: 598 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); 599 case vmIntrinsics::_getAddress_raw: 600 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); 601 602 case vmIntrinsics::_putByte_raw: 603 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); 604 case vmIntrinsics::_putShort_raw: 605 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); 606 case vmIntrinsics::_putChar_raw: 607 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); 608 case vmIntrinsics::_putInt_raw: 609 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); 610 case vmIntrinsics::_putLong_raw: 611 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); 612 case vmIntrinsics::_putFloat_raw: 613 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); 614 case vmIntrinsics::_putDouble_raw: 615 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); 616 case vmIntrinsics::_putAddress_raw: 617 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); 618 619 case vmIntrinsics::_getObjectVolatile: 620 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); 621 case vmIntrinsics::_getBooleanVolatile: 622 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); 623 case vmIntrinsics::_getByteVolatile: 624 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); 625 case vmIntrinsics::_getShortVolatile: 626 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); 627 case vmIntrinsics::_getCharVolatile: 628 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); 629 case vmIntrinsics::_getIntVolatile: 630 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); 631 case vmIntrinsics::_getLongVolatile: 632 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); 633 case vmIntrinsics::_getFloatVolatile: 634 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); 635 case vmIntrinsics::_getDoubleVolatile: 636 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); 637 638 case vmIntrinsics::_putObjectVolatile: 639 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); 640 case vmIntrinsics::_putBooleanVolatile: 641 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); 642 case vmIntrinsics::_putByteVolatile: 643 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); 644 case vmIntrinsics::_putShortVolatile: 645 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); 646 case vmIntrinsics::_putCharVolatile: 647 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); 648 case vmIntrinsics::_putIntVolatile: 649 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); 650 case vmIntrinsics::_putLongVolatile: 651 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); 652 case vmIntrinsics::_putFloatVolatile: 653 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); 654 case vmIntrinsics::_putDoubleVolatile: 655 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); 656 657 case vmIntrinsics::_prefetchRead: 658 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 659 case vmIntrinsics::_prefetchWrite: 660 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 661 case vmIntrinsics::_prefetchReadStatic: 662 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 663 case vmIntrinsics::_prefetchWriteStatic: 664 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 665 666 case vmIntrinsics::_compareAndSwapObject: 667 return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 668 case vmIntrinsics::_compareAndSwapInt: 669 return inline_unsafe_load_store(T_INT, LS_cmpxchg); 670 case vmIntrinsics::_compareAndSwapLong: 671 return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 672 673 case vmIntrinsics::_putOrderedObject: 674 return inline_unsafe_ordered_store(T_OBJECT); 675 case vmIntrinsics::_putOrderedInt: 676 return inline_unsafe_ordered_store(T_INT); 677 case vmIntrinsics::_putOrderedLong: 678 return inline_unsafe_ordered_store(T_LONG); 679 680 case vmIntrinsics::_getAndAddInt: 681 return inline_unsafe_load_store(T_INT, LS_xadd); 682 case vmIntrinsics::_getAndAddLong: 683 return inline_unsafe_load_store(T_LONG, LS_xadd); 684 case vmIntrinsics::_getAndSetInt: 685 return inline_unsafe_load_store(T_INT, LS_xchg); 686 case vmIntrinsics::_getAndSetLong: 687 return inline_unsafe_load_store(T_LONG, LS_xchg); 688 case vmIntrinsics::_getAndSetObject: 689 return inline_unsafe_load_store(T_OBJECT, LS_xchg); 690 691 case vmIntrinsics::_currentThread: 692 return inline_native_currentThread(); 693 case vmIntrinsics::_isInterrupted: 694 return inline_native_isInterrupted(); 695 696 #ifdef TRACE_HAVE_INTRINSICS 697 case vmIntrinsics::_classID: 698 return inline_native_classID(); 699 case vmIntrinsics::_threadID: 700 return inline_native_threadID(); 701 case vmIntrinsics::_counterTime: 702 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 703 #endif 704 case vmIntrinsics::_currentTimeMillis: 705 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 706 case vmIntrinsics::_nanoTime: 707 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 708 case vmIntrinsics::_allocateInstance: 709 return inline_unsafe_allocate(); 710 case vmIntrinsics::_copyMemory: 711 return inline_unsafe_copyMemory(); 712 case vmIntrinsics::_newArray: 713 return inline_native_newArray(); 714 case vmIntrinsics::_getLength: 715 return inline_native_getLength(); 716 case vmIntrinsics::_copyOf: 717 return inline_array_copyOf(false); 718 case vmIntrinsics::_copyOfRange: 719 return inline_array_copyOf(true); 720 case vmIntrinsics::_equalsC: 721 return inline_array_equals(); 722 case vmIntrinsics::_clone: 723 return inline_native_clone(intrinsic()->is_virtual()); 724 725 case vmIntrinsics::_isAssignableFrom: 726 return inline_native_subtype_check(); 727 728 case vmIntrinsics::_isInstance: 729 case vmIntrinsics::_getModifiers: 730 case vmIntrinsics::_isInterface: 731 case vmIntrinsics::_isArray: 732 case vmIntrinsics::_isPrimitive: 733 case vmIntrinsics::_getSuperclass: 734 case vmIntrinsics::_getComponentType: 735 case vmIntrinsics::_getClassAccessFlags: 736 return inline_native_Class_query(intrinsic_id()); 737 738 case vmIntrinsics::_floatToRawIntBits: 739 case vmIntrinsics::_floatToIntBits: 740 case vmIntrinsics::_intBitsToFloat: 741 case vmIntrinsics::_doubleToRawLongBits: 742 case vmIntrinsics::_doubleToLongBits: 743 case vmIntrinsics::_longBitsToDouble: 744 return inline_fp_conversions(intrinsic_id()); 745 746 case vmIntrinsics::_numberOfLeadingZeros_i: 747 case vmIntrinsics::_numberOfLeadingZeros_l: 748 return inline_numberOfLeadingZeros(intrinsic_id()); 749 750 case vmIntrinsics::_numberOfTrailingZeros_i: 751 case vmIntrinsics::_numberOfTrailingZeros_l: 752 return inline_numberOfTrailingZeros(intrinsic_id()); 753 754 case vmIntrinsics::_bitCount_i: 755 case vmIntrinsics::_bitCount_l: 756 return inline_bitCount(intrinsic_id()); 757 758 case vmIntrinsics::_reverseBytes_i: 759 case vmIntrinsics::_reverseBytes_l: 760 case vmIntrinsics::_reverseBytes_s: 761 case vmIntrinsics::_reverseBytes_c: 762 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); 763 764 case vmIntrinsics::_getCallerClass: 765 return inline_native_Reflection_getCallerClass(); 766 767 case vmIntrinsics::_Reference_get: 768 return inline_reference_get(); 769 770 default: 771 // If you get here, it may be that someone has added a new intrinsic 772 // to the list in vmSymbols.hpp without implementing it here. 773 #ifndef PRODUCT 774 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 775 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 776 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 777 } 778 #endif 779 return false; 780 } 781 } 782 783 //------------------------------push_result------------------------------ 784 // Helper function for finishing intrinsics. 785 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { 786 record_for_igvn(region); 787 set_control(_gvn.transform(region)); 788 BasicType value_type = value->type()->basic_type(); 789 push_node(value_type, _gvn.transform(value)); 790 } 791 792 //------------------------------generate_guard--------------------------- 793 // Helper function for generating guarded fast-slow graph structures. 794 // The given 'test', if true, guards a slow path. If the test fails 795 // then a fast path can be taken. (We generally hope it fails.) 796 // In all cases, GraphKit::control() is updated to the fast path. 797 // The returned value represents the control for the slow path. 798 // The return value is never 'top'; it is either a valid control 799 // or NULL if it is obvious that the slow path can never be taken. 800 // Also, if region and the slow control are not NULL, the slow edge 801 // is appended to the region. 802 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 803 if (stopped()) { 804 // Already short circuited. 805 return NULL; 806 } 807 808 // Build an if node and its projections. 809 // If test is true we take the slow path, which we assume is uncommon. 810 if (_gvn.type(test) == TypeInt::ZERO) { 811 // The slow branch is never taken. No need to build this guard. 812 return NULL; 813 } 814 815 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 816 817 Node* if_slow = _gvn.transform( new (C) IfTrueNode(iff) ); 818 if (if_slow == top()) { 819 // The slow branch is never taken. No need to build this guard. 820 return NULL; 821 } 822 823 if (region != NULL) 824 region->add_req(if_slow); 825 826 Node* if_fast = _gvn.transform( new (C) IfFalseNode(iff) ); 827 set_control(if_fast); 828 829 return if_slow; 830 } 831 832 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 833 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 834 } 835 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 836 return generate_guard(test, region, PROB_FAIR); 837 } 838 839 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 840 Node* *pos_index) { 841 if (stopped()) 842 return NULL; // already stopped 843 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 844 return NULL; // index is already adequately typed 845 Node* cmp_lt = _gvn.transform( new (C) CmpINode(index, intcon(0)) ); 846 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) ); 847 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 848 if (is_neg != NULL && pos_index != NULL) { 849 // Emulate effect of Parse::adjust_map_after_if. 850 Node* ccast = new (C) CastIINode(index, TypeInt::POS); 851 ccast->set_req(0, control()); 852 (*pos_index) = _gvn.transform(ccast); 853 } 854 return is_neg; 855 } 856 857 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 858 Node* *pos_index) { 859 if (stopped()) 860 return NULL; // already stopped 861 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 862 return NULL; // index is already adequately typed 863 Node* cmp_le = _gvn.transform( new (C) CmpINode(index, intcon(0)) ); 864 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 865 Node* bol_le = _gvn.transform( new (C) BoolNode(cmp_le, le_or_eq) ); 866 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 867 if (is_notp != NULL && pos_index != NULL) { 868 // Emulate effect of Parse::adjust_map_after_if. 869 Node* ccast = new (C) CastIINode(index, TypeInt::POS1); 870 ccast->set_req(0, control()); 871 (*pos_index) = _gvn.transform(ccast); 872 } 873 return is_notp; 874 } 875 876 // Make sure that 'position' is a valid limit index, in [0..length]. 877 // There are two equivalent plans for checking this: 878 // A. (offset + copyLength) unsigned<= arrayLength 879 // B. offset <= (arrayLength - copyLength) 880 // We require that all of the values above, except for the sum and 881 // difference, are already known to be non-negative. 882 // Plan A is robust in the face of overflow, if offset and copyLength 883 // are both hugely positive. 884 // 885 // Plan B is less direct and intuitive, but it does not overflow at 886 // all, since the difference of two non-negatives is always 887 // representable. Whenever Java methods must perform the equivalent 888 // check they generally use Plan B instead of Plan A. 889 // For the moment we use Plan A. 890 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 891 Node* subseq_length, 892 Node* array_length, 893 RegionNode* region) { 894 if (stopped()) 895 return NULL; // already stopped 896 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 897 if (zero_offset && subseq_length->eqv_uncast(array_length)) 898 return NULL; // common case of whole-array copy 899 Node* last = subseq_length; 900 if (!zero_offset) // last += offset 901 last = _gvn.transform( new (C) AddINode(last, offset)); 902 Node* cmp_lt = _gvn.transform( new (C) CmpUNode(array_length, last) ); 903 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) ); 904 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 905 return is_over; 906 } 907 908 909 //--------------------------generate_current_thread-------------------- 910 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 911 ciKlass* thread_klass = env()->Thread_klass(); 912 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 913 Node* thread = _gvn.transform(new (C) ThreadLocalNode()); 914 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 915 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 916 tls_output = thread; 917 return threadObj; 918 } 919 920 921 //------------------------------make_string_method_node------------------------ 922 // Helper method for String intrinsic functions. This version is called 923 // with str1 and str2 pointing to String object nodes. 924 // 925 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 926 Node* no_ctrl = NULL; 927 928 // Get start addr of string 929 Node* str1_value = load_String_value(no_ctrl, str1); 930 Node* str1_offset = load_String_offset(no_ctrl, str1); 931 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 932 933 // Get length of string 1 934 Node* str1_len = load_String_length(no_ctrl, str1); 935 936 Node* str2_value = load_String_value(no_ctrl, str2); 937 Node* str2_offset = load_String_offset(no_ctrl, str2); 938 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 939 940 Node* str2_len = NULL; 941 Node* result = NULL; 942 943 switch (opcode) { 944 case Op_StrIndexOf: 945 // Get length of string 2 946 str2_len = load_String_length(no_ctrl, str2); 947 948 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 949 str1_start, str1_len, str2_start, str2_len); 950 break; 951 case Op_StrComp: 952 // Get length of string 2 953 str2_len = load_String_length(no_ctrl, str2); 954 955 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 956 str1_start, str1_len, str2_start, str2_len); 957 break; 958 case Op_StrEquals: 959 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 960 str1_start, str2_start, str1_len); 961 break; 962 default: 963 ShouldNotReachHere(); 964 return NULL; 965 } 966 967 // All these intrinsics have checks. 968 C->set_has_split_ifs(true); // Has chance for split-if optimization 969 970 return _gvn.transform(result); 971 } 972 973 // Helper method for String intrinsic functions. This version is called 974 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 975 // to Int nodes containing the lenghts of str1 and str2. 976 // 977 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 978 979 Node* result = NULL; 980 switch (opcode) { 981 case Op_StrIndexOf: 982 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 983 str1_start, cnt1, str2_start, cnt2); 984 break; 985 case Op_StrComp: 986 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 987 str1_start, cnt1, str2_start, cnt2); 988 break; 989 case Op_StrEquals: 990 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 991 str1_start, str2_start, cnt1); 992 break; 993 default: 994 ShouldNotReachHere(); 995 return NULL; 996 } 997 998 // All these intrinsics have checks. 999 C->set_has_split_ifs(true); // Has chance for split-if optimization 1000 1001 return _gvn.transform(result); 1002 } 1003 1004 //------------------------------inline_string_compareTo------------------------ 1005 bool LibraryCallKit::inline_string_compareTo() { 1006 1007 if (!Matcher::has_match_rule(Op_StrComp)) return false; 1008 1009 _sp += 2; 1010 Node *argument = pop(); // pop non-receiver first: it was pushed second 1011 Node *receiver = pop(); 1012 1013 // Null check on self without removing any arguments. The argument 1014 // null check technically happens in the wrong place, which can lead to 1015 // invalid stack traces when string compare is inlined into a method 1016 // which handles NullPointerExceptions. 1017 _sp += 2; 1018 receiver = do_null_check(receiver, T_OBJECT); 1019 argument = do_null_check(argument, T_OBJECT); 1020 _sp -= 2; 1021 if (stopped()) { 1022 return true; 1023 } 1024 1025 Node* compare = make_string_method_node(Op_StrComp, receiver, argument); 1026 push(compare); 1027 return true; 1028 } 1029 1030 //------------------------------inline_string_equals------------------------ 1031 bool LibraryCallKit::inline_string_equals() { 1032 1033 if (!Matcher::has_match_rule(Op_StrEquals)) return false; 1034 1035 int nargs = 2; 1036 _sp += nargs; 1037 Node* argument = pop(); // pop non-receiver first: it was pushed second 1038 Node* receiver = pop(); 1039 1040 // Null check on self without removing any arguments. The argument 1041 // null check technically happens in the wrong place, which can lead to 1042 // invalid stack traces when string compare is inlined into a method 1043 // which handles NullPointerExceptions. 1044 _sp += nargs; 1045 receiver = do_null_check(receiver, T_OBJECT); 1046 //should not do null check for argument for String.equals(), because spec 1047 //allows to specify NULL as argument. 1048 _sp -= nargs; 1049 1050 if (stopped()) { 1051 return true; 1052 } 1053 1054 // paths (plus control) merge 1055 RegionNode* region = new (C) RegionNode(5); 1056 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 1057 1058 // does source == target string? 1059 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument)); 1060 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 1061 1062 Node* if_eq = generate_slow_guard(bol, NULL); 1063 if (if_eq != NULL) { 1064 // receiver == argument 1065 phi->init_req(2, intcon(1)); 1066 region->init_req(2, if_eq); 1067 } 1068 1069 // get String klass for instanceOf 1070 ciInstanceKlass* klass = env()->String_klass(); 1071 1072 if (!stopped()) { 1073 _sp += nargs; // gen_instanceof might do an uncommon trap 1074 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1075 _sp -= nargs; 1076 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1))); 1077 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 1078 1079 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1080 //instanceOf == true, fallthrough 1081 1082 if (inst_false != NULL) { 1083 phi->init_req(3, intcon(0)); 1084 region->init_req(3, inst_false); 1085 } 1086 } 1087 1088 if (!stopped()) { 1089 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1090 1091 // Properly cast the argument to String 1092 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type)); 1093 // This path is taken only when argument's type is String:NotNull. 1094 argument = cast_not_null(argument, false); 1095 1096 Node* no_ctrl = NULL; 1097 1098 // Get start addr of receiver 1099 Node* receiver_val = load_String_value(no_ctrl, receiver); 1100 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1101 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1102 1103 // Get length of receiver 1104 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1105 1106 // Get start addr of argument 1107 Node* argument_val = load_String_value(no_ctrl, argument); 1108 Node* argument_offset = load_String_offset(no_ctrl, argument); 1109 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1110 1111 // Get length of argument 1112 Node* argument_cnt = load_String_length(no_ctrl, argument); 1113 1114 // Check for receiver count != argument count 1115 Node* cmp = _gvn.transform( new(C) CmpINode(receiver_cnt, argument_cnt) ); 1116 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::ne) ); 1117 Node* if_ne = generate_slow_guard(bol, NULL); 1118 if (if_ne != NULL) { 1119 phi->init_req(4, intcon(0)); 1120 region->init_req(4, if_ne); 1121 } 1122 1123 // Check for count == 0 is done by assembler code for StrEquals. 1124 1125 if (!stopped()) { 1126 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1127 phi->init_req(1, equals); 1128 region->init_req(1, control()); 1129 } 1130 } 1131 1132 // post merge 1133 set_control(_gvn.transform(region)); 1134 record_for_igvn(region); 1135 1136 push(_gvn.transform(phi)); 1137 1138 return true; 1139 } 1140 1141 //------------------------------inline_array_equals---------------------------- 1142 bool LibraryCallKit::inline_array_equals() { 1143 1144 if (!Matcher::has_match_rule(Op_AryEq)) return false; 1145 1146 _sp += 2; 1147 Node *argument2 = pop(); 1148 Node *argument1 = pop(); 1149 1150 Node* equals = 1151 _gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), 1152 argument1, argument2) ); 1153 push(equals); 1154 return true; 1155 } 1156 1157 // Java version of String.indexOf(constant string) 1158 // class StringDecl { 1159 // StringDecl(char[] ca) { 1160 // offset = 0; 1161 // count = ca.length; 1162 // value = ca; 1163 // } 1164 // int offset; 1165 // int count; 1166 // char[] value; 1167 // } 1168 // 1169 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1170 // int targetOffset, int cache_i, int md2) { 1171 // int cache = cache_i; 1172 // int sourceOffset = string_object.offset; 1173 // int sourceCount = string_object.count; 1174 // int targetCount = target_object.length; 1175 // 1176 // int targetCountLess1 = targetCount - 1; 1177 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1178 // 1179 // char[] source = string_object.value; 1180 // char[] target = target_object; 1181 // int lastChar = target[targetCountLess1]; 1182 // 1183 // outer_loop: 1184 // for (int i = sourceOffset; i < sourceEnd; ) { 1185 // int src = source[i + targetCountLess1]; 1186 // if (src == lastChar) { 1187 // // With random strings and a 4-character alphabet, 1188 // // reverse matching at this point sets up 0.8% fewer 1189 // // frames, but (paradoxically) makes 0.3% more probes. 1190 // // Since those probes are nearer the lastChar probe, 1191 // // there is may be a net D$ win with reverse matching. 1192 // // But, reversing loop inhibits unroll of inner loop 1193 // // for unknown reason. So, does running outer loop from 1194 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1195 // for (int j = 0; j < targetCountLess1; j++) { 1196 // if (target[targetOffset + j] != source[i+j]) { 1197 // if ((cache & (1 << source[i+j])) == 0) { 1198 // if (md2 < j+1) { 1199 // i += j+1; 1200 // continue outer_loop; 1201 // } 1202 // } 1203 // i += md2; 1204 // continue outer_loop; 1205 // } 1206 // } 1207 // return i - sourceOffset; 1208 // } 1209 // if ((cache & (1 << src)) == 0) { 1210 // i += targetCountLess1; 1211 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1212 // i++; 1213 // } 1214 // return -1; 1215 // } 1216 1217 //------------------------------string_indexOf------------------------ 1218 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1219 jint cache_i, jint md2_i) { 1220 1221 Node* no_ctrl = NULL; 1222 float likely = PROB_LIKELY(0.9); 1223 float unlikely = PROB_UNLIKELY(0.9); 1224 1225 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate 1226 1227 Node* source = load_String_value(no_ctrl, string_object); 1228 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1229 Node* sourceCount = load_String_length(no_ctrl, string_object); 1230 1231 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) ); 1232 jint target_length = target_array->length(); 1233 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1234 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1235 1236 IdealKit kit(this, false, true); 1237 #define __ kit. 1238 Node* zero = __ ConI(0); 1239 Node* one = __ ConI(1); 1240 Node* cache = __ ConI(cache_i); 1241 Node* md2 = __ ConI(md2_i); 1242 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1243 Node* targetCount = __ ConI(target_length); 1244 Node* targetCountLess1 = __ ConI(target_length - 1); 1245 Node* targetOffset = __ ConI(targetOffset_i); 1246 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1247 1248 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1249 Node* outer_loop = __ make_label(2 /* goto */); 1250 Node* return_ = __ make_label(1); 1251 1252 __ set(rtn,__ ConI(-1)); 1253 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1254 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1255 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1256 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1257 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1258 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1259 Node* tpj = __ AddI(targetOffset, __ value(j)); 1260 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1261 Node* ipj = __ AddI(__ value(i), __ value(j)); 1262 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1263 __ if_then(targ, BoolTest::ne, src2); { 1264 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1265 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1266 __ increment(i, __ AddI(__ value(j), one)); 1267 __ goto_(outer_loop); 1268 } __ end_if(); __ dead(j); 1269 }__ end_if(); __ dead(j); 1270 __ increment(i, md2); 1271 __ goto_(outer_loop); 1272 }__ end_if(); 1273 __ increment(j, one); 1274 }__ end_loop(); __ dead(j); 1275 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1276 __ goto_(return_); 1277 }__ end_if(); 1278 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1279 __ increment(i, targetCountLess1); 1280 }__ end_if(); 1281 __ increment(i, one); 1282 __ bind(outer_loop); 1283 }__ end_loop(); __ dead(i); 1284 __ bind(return_); 1285 1286 // Final sync IdealKit and GraphKit. 1287 final_sync(kit); 1288 Node* result = __ value(rtn); 1289 #undef __ 1290 C->set_has_loops(true); 1291 return result; 1292 } 1293 1294 //------------------------------inline_string_indexOf------------------------ 1295 bool LibraryCallKit::inline_string_indexOf() { 1296 1297 _sp += 2; 1298 Node *argument = pop(); // pop non-receiver first: it was pushed second 1299 Node *receiver = pop(); 1300 1301 Node* result; 1302 // Disable the use of pcmpestri until it can be guaranteed that 1303 // the load doesn't cross into the uncommited space. 1304 if (Matcher::has_match_rule(Op_StrIndexOf) && 1305 UseSSE42Intrinsics) { 1306 // Generate SSE4.2 version of indexOf 1307 // We currently only have match rules that use SSE4.2 1308 1309 // Null check on self without removing any arguments. The argument 1310 // null check technically happens in the wrong place, which can lead to 1311 // invalid stack traces when string compare is inlined into a method 1312 // which handles NullPointerExceptions. 1313 _sp += 2; 1314 receiver = do_null_check(receiver, T_OBJECT); 1315 argument = do_null_check(argument, T_OBJECT); 1316 _sp -= 2; 1317 1318 if (stopped()) { 1319 return true; 1320 } 1321 1322 ciInstanceKlass* str_klass = env()->String_klass(); 1323 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1324 1325 // Make the merge point 1326 RegionNode* result_rgn = new (C) RegionNode(4); 1327 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT); 1328 Node* no_ctrl = NULL; 1329 1330 // Get start addr of source string 1331 Node* source = load_String_value(no_ctrl, receiver); 1332 Node* source_offset = load_String_offset(no_ctrl, receiver); 1333 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1334 1335 // Get length of source string 1336 Node* source_cnt = load_String_length(no_ctrl, receiver); 1337 1338 // Get start addr of substring 1339 Node* substr = load_String_value(no_ctrl, argument); 1340 Node* substr_offset = load_String_offset(no_ctrl, argument); 1341 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1342 1343 // Get length of source string 1344 Node* substr_cnt = load_String_length(no_ctrl, argument); 1345 1346 // Check for substr count > string count 1347 Node* cmp = _gvn.transform( new(C) CmpINode(substr_cnt, source_cnt) ); 1348 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::gt) ); 1349 Node* if_gt = generate_slow_guard(bol, NULL); 1350 if (if_gt != NULL) { 1351 result_phi->init_req(2, intcon(-1)); 1352 result_rgn->init_req(2, if_gt); 1353 } 1354 1355 if (!stopped()) { 1356 // Check for substr count == 0 1357 cmp = _gvn.transform( new(C) CmpINode(substr_cnt, intcon(0)) ); 1358 bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 1359 Node* if_zero = generate_slow_guard(bol, NULL); 1360 if (if_zero != NULL) { 1361 result_phi->init_req(3, intcon(0)); 1362 result_rgn->init_req(3, if_zero); 1363 } 1364 } 1365 1366 if (!stopped()) { 1367 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1368 result_phi->init_req(1, result); 1369 result_rgn->init_req(1, control()); 1370 } 1371 set_control(_gvn.transform(result_rgn)); 1372 record_for_igvn(result_rgn); 1373 result = _gvn.transform(result_phi); 1374 1375 } else { // Use LibraryCallKit::string_indexOf 1376 // don't intrinsify if argument isn't a constant string. 1377 if (!argument->is_Con()) { 1378 return false; 1379 } 1380 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); 1381 if (str_type == NULL) { 1382 return false; 1383 } 1384 ciInstanceKlass* klass = env()->String_klass(); 1385 ciObject* str_const = str_type->const_oop(); 1386 if (str_const == NULL || str_const->klass() != klass) { 1387 return false; 1388 } 1389 ciInstance* str = str_const->as_instance(); 1390 assert(str != NULL, "must be instance"); 1391 1392 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1393 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1394 1395 int o; 1396 int c; 1397 if (java_lang_String::has_offset_field()) { 1398 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1399 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1400 } else { 1401 o = 0; 1402 c = pat->length(); 1403 } 1404 1405 // constant strings have no offset and count == length which 1406 // simplifies the resulting code somewhat so lets optimize for that. 1407 if (o != 0 || c != pat->length()) { 1408 return false; 1409 } 1410 1411 // Null check on self without removing any arguments. The argument 1412 // null check technically happens in the wrong place, which can lead to 1413 // invalid stack traces when string compare is inlined into a method 1414 // which handles NullPointerExceptions. 1415 _sp += 2; 1416 receiver = do_null_check(receiver, T_OBJECT); 1417 // No null check on the argument is needed since it's a constant String oop. 1418 _sp -= 2; 1419 if (stopped()) { 1420 return true; 1421 } 1422 1423 // The null string as a pattern always returns 0 (match at beginning of string) 1424 if (c == 0) { 1425 push(intcon(0)); 1426 return true; 1427 } 1428 1429 // Generate default indexOf 1430 jchar lastChar = pat->char_at(o + (c - 1)); 1431 int cache = 0; 1432 int i; 1433 for (i = 0; i < c - 1; i++) { 1434 assert(i < pat->length(), "out of range"); 1435 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1436 } 1437 1438 int md2 = c; 1439 for (i = 0; i < c - 1; i++) { 1440 assert(i < pat->length(), "out of range"); 1441 if (pat->char_at(o + i) == lastChar) { 1442 md2 = (c - 1) - i; 1443 } 1444 } 1445 1446 result = string_indexOf(receiver, pat, o, cache, md2); 1447 } 1448 1449 push(result); 1450 return true; 1451 } 1452 1453 //--------------------------pop_math_arg-------------------------------- 1454 // Pop a double argument to a math function from the stack 1455 // rounding it if necessary. 1456 Node * LibraryCallKit::pop_math_arg() { 1457 Node *arg = pop_pair(); 1458 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) 1459 arg = _gvn.transform( new (C) RoundDoubleNode(0, arg) ); 1460 return arg; 1461 } 1462 1463 //------------------------------inline_trig---------------------------------- 1464 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1465 // argument reduction which will turn into a fast/slow diamond. 1466 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1467 _sp += arg_size(); // restore stack pointer 1468 Node* arg = pop_math_arg(); 1469 Node* trig = NULL; 1470 1471 switch (id) { 1472 case vmIntrinsics::_dsin: 1473 trig = _gvn.transform((Node*)new (C) SinDNode(arg)); 1474 break; 1475 case vmIntrinsics::_dcos: 1476 trig = _gvn.transform((Node*)new (C) CosDNode(arg)); 1477 break; 1478 case vmIntrinsics::_dtan: 1479 trig = _gvn.transform((Node*)new (C) TanDNode(arg)); 1480 break; 1481 default: 1482 assert(false, "bad intrinsic was passed in"); 1483 return false; 1484 } 1485 1486 // Rounding required? Check for argument reduction! 1487 if( Matcher::strict_fp_requires_explicit_rounding ) { 1488 1489 static const double pi_4 = 0.7853981633974483; 1490 static const double neg_pi_4 = -0.7853981633974483; 1491 // pi/2 in 80-bit extended precision 1492 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1493 // -pi/2 in 80-bit extended precision 1494 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00}; 1495 // Cutoff value for using this argument reduction technique 1496 //static const double pi_2_minus_epsilon = 1.564660403643354; 1497 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1498 1499 // Pseudocode for sin: 1500 // if (x <= Math.PI / 4.0) { 1501 // if (x >= -Math.PI / 4.0) return fsin(x); 1502 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1503 // } else { 1504 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1505 // } 1506 // return StrictMath.sin(x); 1507 1508 // Pseudocode for cos: 1509 // if (x <= Math.PI / 4.0) { 1510 // if (x >= -Math.PI / 4.0) return fcos(x); 1511 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1512 // } else { 1513 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1514 // } 1515 // return StrictMath.cos(x); 1516 1517 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1518 // requires a special machine instruction to load it. Instead we'll try 1519 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1520 // probably do the math inside the SIN encoding. 1521 1522 // Make the merge point 1523 RegionNode *r = new (C) RegionNode(3); 1524 Node *phi = new (C) PhiNode(r,Type::DOUBLE); 1525 1526 // Flatten arg so we need only 1 test 1527 Node *abs = _gvn.transform(new (C) AbsDNode(arg)); 1528 // Node for PI/4 constant 1529 Node *pi4 = makecon(TypeD::make(pi_4)); 1530 // Check PI/4 : abs(arg) 1531 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs)); 1532 // Check: If PI/4 < abs(arg) then go slow 1533 Node *bol = _gvn.transform( new (C) BoolNode( cmp, BoolTest::lt ) ); 1534 // Branch either way 1535 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1536 set_control(opt_iff(r,iff)); 1537 1538 // Set fast path result 1539 phi->init_req(2,trig); 1540 1541 // Slow path - non-blocking leaf call 1542 Node* call = NULL; 1543 switch (id) { 1544 case vmIntrinsics::_dsin: 1545 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1546 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1547 "Sin", NULL, arg, top()); 1548 break; 1549 case vmIntrinsics::_dcos: 1550 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1551 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1552 "Cos", NULL, arg, top()); 1553 break; 1554 case vmIntrinsics::_dtan: 1555 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1556 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1557 "Tan", NULL, arg, top()); 1558 break; 1559 } 1560 assert(control()->in(0) == call, ""); 1561 Node* slow_result = _gvn.transform(new (C) ProjNode(call,TypeFunc::Parms)); 1562 r->init_req(1,control()); 1563 phi->init_req(1,slow_result); 1564 1565 // Post-merge 1566 set_control(_gvn.transform(r)); 1567 record_for_igvn(r); 1568 trig = _gvn.transform(phi); 1569 1570 C->set_has_split_ifs(true); // Has chance for split-if optimization 1571 } 1572 // Push result back on JVM stack 1573 push_pair(trig); 1574 return true; 1575 } 1576 1577 //------------------------------inline_sqrt------------------------------------- 1578 // Inline square root instruction, if possible. 1579 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { 1580 assert(id == vmIntrinsics::_dsqrt, "Not square root"); 1581 _sp += arg_size(); // restore stack pointer 1582 push_pair(_gvn.transform(new (C) SqrtDNode(0, pop_math_arg()))); 1583 return true; 1584 } 1585 1586 //------------------------------inline_abs------------------------------------- 1587 // Inline absolute value instruction, if possible. 1588 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { 1589 assert(id == vmIntrinsics::_dabs, "Not absolute value"); 1590 _sp += arg_size(); // restore stack pointer 1591 push_pair(_gvn.transform(new (C) AbsDNode(pop_math_arg()))); 1592 return true; 1593 } 1594 1595 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1596 //------------------- 1597 //result=(result.isNaN())? funcAddr():result; 1598 // Check: If isNaN() by checking result!=result? then either trap 1599 // or go to runtime 1600 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result,result)); 1601 // Build the boolean node 1602 Node* bolisnum = _gvn.transform( new (C) BoolNode(cmpisnan, BoolTest::eq) ); 1603 1604 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1605 { 1606 BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1607 // End the current control-flow path 1608 push_pair(x); 1609 if (y != NULL) { 1610 push_pair(y); 1611 } 1612 // The pow or exp intrinsic returned a NaN, which requires a call 1613 // to the runtime. Recompile with the runtime call. 1614 uncommon_trap(Deoptimization::Reason_intrinsic, 1615 Deoptimization::Action_make_not_entrant); 1616 } 1617 push_pair(result); 1618 } else { 1619 // If this inlining ever returned NaN in the past, we compile a call 1620 // to the runtime to properly handle corner cases 1621 1622 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1623 Node* if_slow = _gvn.transform( new (C) IfFalseNode(iff) ); 1624 Node* if_fast = _gvn.transform( new (C) IfTrueNode(iff) ); 1625 1626 if (!if_slow->is_top()) { 1627 RegionNode* result_region = new(C) RegionNode(3); 1628 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE); 1629 1630 result_region->init_req(1, if_fast); 1631 result_val->init_req(1, result); 1632 1633 set_control(if_slow); 1634 1635 const TypePtr* no_memory_effects = NULL; 1636 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1637 no_memory_effects, 1638 x, top(), y, y ? top() : NULL); 1639 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0)); 1640 #ifdef ASSERT 1641 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1)); 1642 assert(value_top == top(), "second value must be top"); 1643 #endif 1644 1645 result_region->init_req(2, control()); 1646 result_val->init_req(2, value); 1647 push_result(result_region, result_val); 1648 } else { 1649 push_pair(result); 1650 } 1651 } 1652 } 1653 1654 //------------------------------inline_exp------------------------------------- 1655 // Inline exp instructions, if possible. The Intel hardware only misses 1656 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1657 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { 1658 assert(id == vmIntrinsics::_dexp, "Not exp"); 1659 1660 _sp += arg_size(); // restore stack pointer 1661 Node *x = pop_math_arg(); 1662 Node *result = _gvn.transform(new (C) ExpDNode(0,x)); 1663 1664 finish_pow_exp(result, x, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1665 1666 C->set_has_split_ifs(true); // Has chance for split-if optimization 1667 1668 return true; 1669 } 1670 1671 //------------------------------inline_pow------------------------------------- 1672 // Inline power instructions, if possible. 1673 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { 1674 assert(id == vmIntrinsics::_dpow, "Not pow"); 1675 1676 // Pseudocode for pow 1677 // if (x <= 0.0) { 1678 // long longy = (long)y; 1679 // if ((double)longy == y) { // if y is long 1680 // if (y + 1 == y) longy = 0; // huge number: even 1681 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1682 // } else { 1683 // result = NaN; 1684 // } 1685 // } else { 1686 // result = DPow(x,y); 1687 // } 1688 // if (result != result)? { 1689 // result = uncommon_trap() or runtime_call(); 1690 // } 1691 // return result; 1692 1693 _sp += arg_size(); // restore stack pointer 1694 Node* y = pop_math_arg(); 1695 Node* x = pop_math_arg(); 1696 1697 Node* result = NULL; 1698 1699 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1700 // Short form: skip the fancy tests and just check for NaN result. 1701 result = _gvn.transform( new (C) PowDNode(0, x, y) ); 1702 } else { 1703 // If this inlining ever returned NaN in the past, include all 1704 // checks + call to the runtime. 1705 1706 // Set the merge point for If node with condition of (x <= 0.0) 1707 // There are four possible paths to region node and phi node 1708 RegionNode *r = new (C) RegionNode(4); 1709 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1710 1711 // Build the first if node: if (x <= 0.0) 1712 // Node for 0 constant 1713 Node *zeronode = makecon(TypeD::ZERO); 1714 // Check x:0 1715 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode)); 1716 // Check: If (x<=0) then go complex path 1717 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp, BoolTest::le ) ); 1718 // Branch either way 1719 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1720 // Fast path taken; set region slot 3 1721 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(if1) ); 1722 r->init_req(3,fast_taken); // Capture fast-control 1723 1724 // Fast path not-taken, i.e. slow path 1725 Node *complex_path = _gvn.transform( new (C) IfTrueNode(if1) ); 1726 1727 // Set fast path result 1728 Node *fast_result = _gvn.transform( new (C) PowDNode(0, x, y) ); 1729 phi->init_req(3, fast_result); 1730 1731 // Complex path 1732 // Build the second if node (if y is long) 1733 // Node for (long)y 1734 Node *longy = _gvn.transform( new (C) ConvD2LNode(y)); 1735 // Node for (double)((long) y) 1736 Node *doublelongy= _gvn.transform( new (C) ConvL2DNode(longy)); 1737 // Check (double)((long) y) : y 1738 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y)); 1739 // Check if (y isn't long) then go to slow path 1740 1741 Node *bol2 = _gvn.transform( new (C) BoolNode( cmplongy, BoolTest::ne ) ); 1742 // Branch either way 1743 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1744 Node* ylong_path = _gvn.transform( new (C) IfFalseNode(if2)); 1745 1746 Node *slow_path = _gvn.transform( new (C) IfTrueNode(if2) ); 1747 1748 // Calculate DPow(abs(x), y)*(1 & (long)y) 1749 // Node for constant 1 1750 Node *conone = longcon(1); 1751 // 1& (long)y 1752 Node *signnode= _gvn.transform( new (C) AndLNode(conone, longy) ); 1753 1754 // A huge number is always even. Detect a huge number by checking 1755 // if y + 1 == y and set integer to be tested for parity to 0. 1756 // Required for corner case: 1757 // (long)9.223372036854776E18 = max_jlong 1758 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1759 // max_jlong is odd but 9.223372036854776E18 is even 1760 Node* yplus1 = _gvn.transform( new (C) AddDNode(y, makecon(TypeD::make(1)))); 1761 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y)); 1762 Node *bolyplus1 = _gvn.transform( new (C) BoolNode( cmpyplus1, BoolTest::eq ) ); 1763 Node* correctedsign = NULL; 1764 if (ConditionalMoveLimit != 0) { 1765 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1766 } else { 1767 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1768 RegionNode *r = new (C) RegionNode(3); 1769 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 1770 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyplus1))); 1771 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyplus1))); 1772 phi->init_req(1, signnode); 1773 phi->init_req(2, longcon(0)); 1774 correctedsign = _gvn.transform(phi); 1775 ylong_path = _gvn.transform(r); 1776 record_for_igvn(r); 1777 } 1778 1779 // zero node 1780 Node *conzero = longcon(0); 1781 // Check (1&(long)y)==0? 1782 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero)); 1783 // Check if (1&(long)y)!=0?, if so the result is negative 1784 Node *bol3 = _gvn.transform( new (C) BoolNode( cmpeq1, BoolTest::ne ) ); 1785 // abs(x) 1786 Node *absx=_gvn.transform( new (C) AbsDNode(x)); 1787 // abs(x)^y 1788 Node *absxpowy = _gvn.transform( new (C) PowDNode(0, absx, y) ); 1789 // -abs(x)^y 1790 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy)); 1791 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1792 Node *signresult = NULL; 1793 if (ConditionalMoveLimit != 0) { 1794 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1795 } else { 1796 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1797 RegionNode *r = new (C) RegionNode(3); 1798 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1799 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyeven))); 1800 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyeven))); 1801 phi->init_req(1, absxpowy); 1802 phi->init_req(2, negabsxpowy); 1803 signresult = _gvn.transform(phi); 1804 ylong_path = _gvn.transform(r); 1805 record_for_igvn(r); 1806 } 1807 // Set complex path fast result 1808 r->init_req(2, ylong_path); 1809 phi->init_req(2, signresult); 1810 1811 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1812 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1813 r->init_req(1,slow_path); 1814 phi->init_req(1,slow_result); 1815 1816 // Post merge 1817 set_control(_gvn.transform(r)); 1818 record_for_igvn(r); 1819 result=_gvn.transform(phi); 1820 } 1821 1822 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1823 1824 C->set_has_split_ifs(true); // Has chance for split-if optimization 1825 1826 return true; 1827 } 1828 1829 //------------------------------inline_trans------------------------------------- 1830 // Inline transcendental instructions, if possible. The Intel hardware gets 1831 // these right, no funny corner cases missed. 1832 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { 1833 _sp += arg_size(); // restore stack pointer 1834 Node* arg = pop_math_arg(); 1835 Node* trans = NULL; 1836 1837 switch (id) { 1838 case vmIntrinsics::_dlog: 1839 trans = _gvn.transform((Node*)new (C) LogDNode(arg)); 1840 break; 1841 case vmIntrinsics::_dlog10: 1842 trans = _gvn.transform((Node*)new (C) Log10DNode(arg)); 1843 break; 1844 default: 1845 assert(false, "bad intrinsic was passed in"); 1846 return false; 1847 } 1848 1849 // Push result back on JVM stack 1850 push_pair(trans); 1851 return true; 1852 } 1853 1854 //------------------------------runtime_math----------------------------- 1855 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1856 Node* a = NULL; 1857 Node* b = NULL; 1858 1859 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1860 "must be (DD)D or (D)D type"); 1861 1862 // Inputs 1863 _sp += arg_size(); // restore stack pointer 1864 if (call_type == OptoRuntime::Math_DD_D_Type()) { 1865 b = pop_math_arg(); 1866 } 1867 a = pop_math_arg(); 1868 1869 const TypePtr* no_memory_effects = NULL; 1870 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1871 no_memory_effects, 1872 a, top(), b, b ? top() : NULL); 1873 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0)); 1874 #ifdef ASSERT 1875 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1)); 1876 assert(value_top == top(), "second value must be top"); 1877 #endif 1878 1879 push_pair(value); 1880 return true; 1881 } 1882 1883 //------------------------------inline_math_native----------------------------- 1884 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1885 switch (id) { 1886 // These intrinsics are not properly supported on all hardware 1887 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1888 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); 1889 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1890 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); 1891 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1892 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); 1893 1894 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : 1895 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); 1896 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : 1897 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); 1898 1899 // These intrinsics are supported on all hardware 1900 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; 1901 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; 1902 1903 case vmIntrinsics::_dexp: return 1904 Matcher::has_match_rule(Op_ExpD) ? inline_exp(id) : 1905 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1906 case vmIntrinsics::_dpow: return 1907 Matcher::has_match_rule(Op_PowD) ? inline_pow(id) : 1908 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1909 1910 // These intrinsics are not yet correctly implemented 1911 case vmIntrinsics::_datan2: 1912 return false; 1913 1914 default: 1915 ShouldNotReachHere(); 1916 return false; 1917 } 1918 } 1919 1920 static bool is_simple_name(Node* n) { 1921 return (n->req() == 1 // constant 1922 || (n->is_Type() && n->as_Type()->type()->singleton()) 1923 || n->is_Proj() // parameter or return value 1924 || n->is_Phi() // local of some sort 1925 ); 1926 } 1927 1928 //----------------------------inline_min_max----------------------------------- 1929 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1930 push(generate_min_max(id, argument(0), argument(1))); 1931 1932 return true; 1933 } 1934 1935 Node* 1936 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1937 // These are the candidate return value: 1938 Node* xvalue = x0; 1939 Node* yvalue = y0; 1940 1941 if (xvalue == yvalue) { 1942 return xvalue; 1943 } 1944 1945 bool want_max = (id == vmIntrinsics::_max); 1946 1947 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1948 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1949 if (txvalue == NULL || tyvalue == NULL) return top(); 1950 // This is not really necessary, but it is consistent with a 1951 // hypothetical MaxINode::Value method: 1952 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1953 1954 // %%% This folding logic should (ideally) be in a different place. 1955 // Some should be inside IfNode, and there to be a more reliable 1956 // transformation of ?: style patterns into cmoves. We also want 1957 // more powerful optimizations around cmove and min/max. 1958 1959 // Try to find a dominating comparison of these guys. 1960 // It can simplify the index computation for Arrays.copyOf 1961 // and similar uses of System.arraycopy. 1962 // First, compute the normalized version of CmpI(x, y). 1963 int cmp_op = Op_CmpI; 1964 Node* xkey = xvalue; 1965 Node* ykey = yvalue; 1966 Node* ideal_cmpxy = _gvn.transform( new(C) CmpINode(xkey, ykey) ); 1967 if (ideal_cmpxy->is_Cmp()) { 1968 // E.g., if we have CmpI(length - offset, count), 1969 // it might idealize to CmpI(length, count + offset) 1970 cmp_op = ideal_cmpxy->Opcode(); 1971 xkey = ideal_cmpxy->in(1); 1972 ykey = ideal_cmpxy->in(2); 1973 } 1974 1975 // Start by locating any relevant comparisons. 1976 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1977 Node* cmpxy = NULL; 1978 Node* cmpyx = NULL; 1979 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1980 Node* cmp = start_from->fast_out(k); 1981 if (cmp->outcnt() > 0 && // must have prior uses 1982 cmp->in(0) == NULL && // must be context-independent 1983 cmp->Opcode() == cmp_op) { // right kind of compare 1984 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1985 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1986 } 1987 } 1988 1989 const int NCMPS = 2; 1990 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1991 int cmpn; 1992 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1993 if (cmps[cmpn] != NULL) break; // find a result 1994 } 1995 if (cmpn < NCMPS) { 1996 // Look for a dominating test that tells us the min and max. 1997 int depth = 0; // Limit search depth for speed 1998 Node* dom = control(); 1999 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2000 if (++depth >= 100) break; 2001 Node* ifproj = dom; 2002 if (!ifproj->is_Proj()) continue; 2003 Node* iff = ifproj->in(0); 2004 if (!iff->is_If()) continue; 2005 Node* bol = iff->in(1); 2006 if (!bol->is_Bool()) continue; 2007 Node* cmp = bol->in(1); 2008 if (cmp == NULL) continue; 2009 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2010 if (cmps[cmpn] == cmp) break; 2011 if (cmpn == NCMPS) continue; 2012 BoolTest::mask btest = bol->as_Bool()->_test._test; 2013 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2014 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2015 // At this point, we know that 'x btest y' is true. 2016 switch (btest) { 2017 case BoolTest::eq: 2018 // They are proven equal, so we can collapse the min/max. 2019 // Either value is the answer. Choose the simpler. 2020 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2021 return yvalue; 2022 return xvalue; 2023 case BoolTest::lt: // x < y 2024 case BoolTest::le: // x <= y 2025 return (want_max ? yvalue : xvalue); 2026 case BoolTest::gt: // x > y 2027 case BoolTest::ge: // x >= y 2028 return (want_max ? xvalue : yvalue); 2029 } 2030 } 2031 } 2032 2033 // We failed to find a dominating test. 2034 // Let's pick a test that might GVN with prior tests. 2035 Node* best_bol = NULL; 2036 BoolTest::mask best_btest = BoolTest::illegal; 2037 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2038 Node* cmp = cmps[cmpn]; 2039 if (cmp == NULL) continue; 2040 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2041 Node* bol = cmp->fast_out(j); 2042 if (!bol->is_Bool()) continue; 2043 BoolTest::mask btest = bol->as_Bool()->_test._test; 2044 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2045 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2046 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2047 best_bol = bol->as_Bool(); 2048 best_btest = btest; 2049 } 2050 } 2051 } 2052 2053 Node* answer_if_true = NULL; 2054 Node* answer_if_false = NULL; 2055 switch (best_btest) { 2056 default: 2057 if (cmpxy == NULL) 2058 cmpxy = ideal_cmpxy; 2059 best_bol = _gvn.transform( new(C) BoolNode(cmpxy, BoolTest::lt) ); 2060 // and fall through: 2061 case BoolTest::lt: // x < y 2062 case BoolTest::le: // x <= y 2063 answer_if_true = (want_max ? yvalue : xvalue); 2064 answer_if_false = (want_max ? xvalue : yvalue); 2065 break; 2066 case BoolTest::gt: // x > y 2067 case BoolTest::ge: // x >= y 2068 answer_if_true = (want_max ? xvalue : yvalue); 2069 answer_if_false = (want_max ? yvalue : xvalue); 2070 break; 2071 } 2072 2073 jint hi, lo; 2074 if (want_max) { 2075 // We can sharpen the minimum. 2076 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2077 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2078 } else { 2079 // We can sharpen the maximum. 2080 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2081 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2082 } 2083 2084 // Use a flow-free graph structure, to avoid creating excess control edges 2085 // which could hinder other optimizations. 2086 // Since Math.min/max is often used with arraycopy, we want 2087 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2088 Node* cmov = CMoveNode::make(C, NULL, best_bol, 2089 answer_if_false, answer_if_true, 2090 TypeInt::make(lo, hi, widen)); 2091 2092 return _gvn.transform(cmov); 2093 2094 /* 2095 // This is not as desirable as it may seem, since Min and Max 2096 // nodes do not have a full set of optimizations. 2097 // And they would interfere, anyway, with 'if' optimizations 2098 // and with CMoveI canonical forms. 2099 switch (id) { 2100 case vmIntrinsics::_min: 2101 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2102 case vmIntrinsics::_max: 2103 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2104 default: 2105 ShouldNotReachHere(); 2106 } 2107 */ 2108 } 2109 2110 inline int 2111 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2112 const TypePtr* base_type = TypePtr::NULL_PTR; 2113 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2114 if (base_type == NULL) { 2115 // Unknown type. 2116 return Type::AnyPtr; 2117 } else if (base_type == TypePtr::NULL_PTR) { 2118 // Since this is a NULL+long form, we have to switch to a rawptr. 2119 base = _gvn.transform( new (C) CastX2PNode(offset) ); 2120 offset = MakeConX(0); 2121 return Type::RawPtr; 2122 } else if (base_type->base() == Type::RawPtr) { 2123 return Type::RawPtr; 2124 } else if (base_type->isa_oopptr()) { 2125 // Base is never null => always a heap address. 2126 if (base_type->ptr() == TypePtr::NotNull) { 2127 return Type::OopPtr; 2128 } 2129 // Offset is small => always a heap address. 2130 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2131 if (offset_type != NULL && 2132 base_type->offset() == 0 && // (should always be?) 2133 offset_type->_lo >= 0 && 2134 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2135 return Type::OopPtr; 2136 } 2137 // Otherwise, it might either be oop+off or NULL+addr. 2138 return Type::AnyPtr; 2139 } else { 2140 // No information: 2141 return Type::AnyPtr; 2142 } 2143 } 2144 2145 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2146 int kind = classify_unsafe_addr(base, offset); 2147 if (kind == Type::RawPtr) { 2148 return basic_plus_adr(top(), base, offset); 2149 } else { 2150 return basic_plus_adr(base, offset); 2151 } 2152 } 2153 2154 //-------------------inline_numberOfLeadingZeros_int/long----------------------- 2155 // inline int Integer.numberOfLeadingZeros(int) 2156 // inline int Long.numberOfLeadingZeros(long) 2157 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) { 2158 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros"); 2159 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false; 2160 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false; 2161 _sp += arg_size(); // restore stack pointer 2162 switch (id) { 2163 case vmIntrinsics::_numberOfLeadingZeros_i: 2164 push(_gvn.transform(new (C) CountLeadingZerosINode(pop()))); 2165 break; 2166 case vmIntrinsics::_numberOfLeadingZeros_l: 2167 push(_gvn.transform(new (C) CountLeadingZerosLNode(pop_pair()))); 2168 break; 2169 default: 2170 ShouldNotReachHere(); 2171 } 2172 return true; 2173 } 2174 2175 //-------------------inline_numberOfTrailingZeros_int/long---------------------- 2176 // inline int Integer.numberOfTrailingZeros(int) 2177 // inline int Long.numberOfTrailingZeros(long) 2178 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) { 2179 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros"); 2180 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false; 2181 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false; 2182 _sp += arg_size(); // restore stack pointer 2183 switch (id) { 2184 case vmIntrinsics::_numberOfTrailingZeros_i: 2185 push(_gvn.transform(new (C) CountTrailingZerosINode(pop()))); 2186 break; 2187 case vmIntrinsics::_numberOfTrailingZeros_l: 2188 push(_gvn.transform(new (C) CountTrailingZerosLNode(pop_pair()))); 2189 break; 2190 default: 2191 ShouldNotReachHere(); 2192 } 2193 return true; 2194 } 2195 2196 //----------------------------inline_bitCount_int/long----------------------- 2197 // inline int Integer.bitCount(int) 2198 // inline int Long.bitCount(long) 2199 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) { 2200 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount"); 2201 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false; 2202 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false; 2203 _sp += arg_size(); // restore stack pointer 2204 switch (id) { 2205 case vmIntrinsics::_bitCount_i: 2206 push(_gvn.transform(new (C) PopCountINode(pop()))); 2207 break; 2208 case vmIntrinsics::_bitCount_l: 2209 push(_gvn.transform(new (C) PopCountLNode(pop_pair()))); 2210 break; 2211 default: 2212 ShouldNotReachHere(); 2213 } 2214 return true; 2215 } 2216 2217 //----------------------------inline_reverseBytes_int/long/char/short------------------- 2218 // inline Integer.reverseBytes(int) 2219 // inline Long.reverseBytes(long) 2220 // inline Character.reverseBytes(char) 2221 // inline Short.reverseBytes(short) 2222 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { 2223 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l || 2224 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s, 2225 "not reverse Bytes"); 2226 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; 2227 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; 2228 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false; 2229 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false; 2230 _sp += arg_size(); // restore stack pointer 2231 switch (id) { 2232 case vmIntrinsics::_reverseBytes_i: 2233 push(_gvn.transform(new (C) ReverseBytesINode(0, pop()))); 2234 break; 2235 case vmIntrinsics::_reverseBytes_l: 2236 push_pair(_gvn.transform(new (C) ReverseBytesLNode(0, pop_pair()))); 2237 break; 2238 case vmIntrinsics::_reverseBytes_c: 2239 push(_gvn.transform(new (C) ReverseBytesUSNode(0, pop()))); 2240 break; 2241 case vmIntrinsics::_reverseBytes_s: 2242 push(_gvn.transform(new (C) ReverseBytesSNode(0, pop()))); 2243 break; 2244 default: 2245 ; 2246 } 2247 return true; 2248 } 2249 2250 //----------------------------inline_unsafe_access---------------------------- 2251 2252 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2253 2254 // Helper that guards and inserts a pre-barrier. 2255 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2256 Node* pre_val, int nargs, bool need_mem_bar) { 2257 // We could be accessing the referent field of a reference object. If so, when G1 2258 // is enabled, we need to log the value in the referent field in an SATB buffer. 2259 // This routine performs some compile time filters and generates suitable 2260 // runtime filters that guard the pre-barrier code. 2261 // Also add memory barrier for non volatile load from the referent field 2262 // to prevent commoning of loads across safepoint. 2263 if (!UseG1GC && !need_mem_bar) 2264 return; 2265 2266 // Some compile time checks. 2267 2268 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2269 const TypeX* otype = offset->find_intptr_t_type(); 2270 if (otype != NULL && otype->is_con() && 2271 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2272 // Constant offset but not the reference_offset so just return 2273 return; 2274 } 2275 2276 // We only need to generate the runtime guards for instances. 2277 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2278 if (btype != NULL) { 2279 if (btype->isa_aryptr()) { 2280 // Array type so nothing to do 2281 return; 2282 } 2283 2284 const TypeInstPtr* itype = btype->isa_instptr(); 2285 if (itype != NULL) { 2286 // Can the klass of base_oop be statically determined to be 2287 // _not_ a sub-class of Reference and _not_ Object? 2288 ciKlass* klass = itype->klass(); 2289 if ( klass->is_loaded() && 2290 !klass->is_subtype_of(env()->Reference_klass()) && 2291 !env()->Object_klass()->is_subtype_of(klass)) { 2292 return; 2293 } 2294 } 2295 } 2296 2297 // The compile time filters did not reject base_oop/offset so 2298 // we need to generate the following runtime filters 2299 // 2300 // if (offset == java_lang_ref_Reference::_reference_offset) { 2301 // if (instance_of(base, java.lang.ref.Reference)) { 2302 // pre_barrier(_, pre_val, ...); 2303 // } 2304 // } 2305 2306 float likely = PROB_LIKELY(0.999); 2307 float unlikely = PROB_UNLIKELY(0.999); 2308 2309 IdealKit ideal(this); 2310 #define __ ideal. 2311 2312 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2313 2314 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2315 // Update graphKit memory and control from IdealKit. 2316 sync_kit(ideal); 2317 2318 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2319 _sp += nargs; // gen_instanceof might do an uncommon trap 2320 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2321 _sp -= nargs; 2322 2323 // Update IdealKit memory and control from graphKit. 2324 __ sync_kit(this); 2325 2326 Node* one = __ ConI(1); 2327 // is_instof == 0 if base_oop == NULL 2328 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2329 2330 // Update graphKit from IdeakKit. 2331 sync_kit(ideal); 2332 2333 // Use the pre-barrier to record the value in the referent field 2334 pre_barrier(false /* do_load */, 2335 __ ctrl(), 2336 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2337 pre_val /* pre_val */, 2338 T_OBJECT); 2339 if (need_mem_bar) { 2340 // Add memory barrier to prevent commoning reads from this field 2341 // across safepoint since GC can change its value. 2342 insert_mem_bar(Op_MemBarCPUOrder); 2343 } 2344 // Update IdealKit from graphKit. 2345 __ sync_kit(this); 2346 2347 } __ end_if(); // _ref_type != ref_none 2348 } __ end_if(); // offset == referent_offset 2349 2350 // Final sync IdealKit and GraphKit. 2351 final_sync(ideal); 2352 #undef __ 2353 } 2354 2355 2356 // Interpret Unsafe.fieldOffset cookies correctly: 2357 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2358 2359 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2360 // Attempt to infer a sharper value type from the offset and base type. 2361 ciKlass* sharpened_klass = NULL; 2362 2363 // See if it is an instance field, with an object type. 2364 if (alias_type->field() != NULL) { 2365 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2366 if (alias_type->field()->type()->is_klass()) { 2367 sharpened_klass = alias_type->field()->type()->as_klass(); 2368 } 2369 } 2370 2371 // See if it is a narrow oop array. 2372 if (adr_type->isa_aryptr()) { 2373 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2374 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2375 if (elem_type != NULL) { 2376 sharpened_klass = elem_type->klass(); 2377 } 2378 } 2379 } 2380 2381 // The sharpened class might be unloaded if there is no class loader 2382 // contraint in place. 2383 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2384 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2385 2386 #ifndef PRODUCT 2387 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2388 tty->print(" from base type: "); adr_type->dump(); 2389 tty->print(" sharpened value: "); tjp->dump(); 2390 } 2391 #endif 2392 // Sharpen the value type. 2393 return tjp; 2394 } 2395 return NULL; 2396 } 2397 2398 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2399 if (callee()->is_static()) return false; // caller must have the capability! 2400 2401 #ifndef PRODUCT 2402 { 2403 ResourceMark rm; 2404 // Check the signatures. 2405 ciSignature* sig = signature(); 2406 #ifdef ASSERT 2407 if (!is_store) { 2408 // Object getObject(Object base, int/long offset), etc. 2409 BasicType rtype = sig->return_type()->basic_type(); 2410 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2411 rtype = T_ADDRESS; // it is really a C void* 2412 assert(rtype == type, "getter must return the expected value"); 2413 if (!is_native_ptr) { 2414 assert(sig->count() == 2, "oop getter has 2 arguments"); 2415 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2416 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2417 } else { 2418 assert(sig->count() == 1, "native getter has 1 argument"); 2419 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2420 } 2421 } else { 2422 // void putObject(Object base, int/long offset, Object x), etc. 2423 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2424 if (!is_native_ptr) { 2425 assert(sig->count() == 3, "oop putter has 3 arguments"); 2426 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2427 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2428 } else { 2429 assert(sig->count() == 2, "native putter has 2 arguments"); 2430 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2431 } 2432 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2433 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2434 vtype = T_ADDRESS; // it is really a C void* 2435 assert(vtype == type, "putter must accept the expected value"); 2436 } 2437 #endif // ASSERT 2438 } 2439 #endif //PRODUCT 2440 2441 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2442 2443 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; 2444 2445 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words 2446 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); 2447 assert(callee()->arg_size() == nargs, "must be"); 2448 2449 debug_only(int saved_sp = _sp); 2450 _sp += nargs; 2451 2452 Node* val; 2453 debug_only(val = (Node*)(uintptr_t)-1); 2454 2455 2456 if (is_store) { 2457 // Get the value being stored. (Pop it first; it was pushed last.) 2458 switch (type) { 2459 case T_DOUBLE: 2460 case T_LONG: 2461 case T_ADDRESS: 2462 val = pop_pair(); 2463 break; 2464 default: 2465 val = pop(); 2466 } 2467 } 2468 2469 // Build address expression. See the code in inline_unsafe_prefetch. 2470 Node *adr; 2471 Node *heap_base_oop = top(); 2472 Node* offset = top(); 2473 2474 if (!is_native_ptr) { 2475 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2476 offset = pop_pair(); 2477 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2478 Node* base = pop(); 2479 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2480 // to be plain byte offsets, which are also the same as those accepted 2481 // by oopDesc::field_base. 2482 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2483 "fieldOffset must be byte-scaled"); 2484 // 32-bit machines ignore the high half! 2485 offset = ConvL2X(offset); 2486 adr = make_unsafe_address(base, offset); 2487 heap_base_oop = base; 2488 } else { 2489 Node* ptr = pop_pair(); 2490 // Adjust Java long to machine word: 2491 ptr = ConvL2X(ptr); 2492 adr = make_unsafe_address(NULL, ptr); 2493 } 2494 2495 // Pop receiver last: it was pushed first. 2496 Node *receiver = pop(); 2497 2498 assert(saved_sp == _sp, "must have correct argument count"); 2499 2500 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2501 2502 // First guess at the value type. 2503 const Type *value_type = Type::get_const_basic_type(type); 2504 2505 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2506 // there was not enough information to nail it down. 2507 Compile::AliasType* alias_type = C->alias_type(adr_type); 2508 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2509 2510 // We will need memory barriers unless we can determine a unique 2511 // alias category for this reference. (Note: If for some reason 2512 // the barriers get omitted and the unsafe reference begins to "pollute" 2513 // the alias analysis of the rest of the graph, either Compile::can_alias 2514 // or Compile::must_alias will throw a diagnostic assert.) 2515 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2516 2517 // If we are reading the value of the referent field of a Reference 2518 // object (either by using Unsafe directly or through reflection) 2519 // then, if G1 is enabled, we need to record the referent in an 2520 // SATB log buffer using the pre-barrier mechanism. 2521 // Also we need to add memory barrier to prevent commoning reads 2522 // from this field across safepoint since GC can change its value. 2523 bool need_read_barrier = !is_native_ptr && !is_store && 2524 offset != top() && heap_base_oop != top(); 2525 2526 if (!is_store && type == T_OBJECT) { 2527 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2528 if (tjp != NULL) { 2529 value_type = tjp; 2530 } 2531 } 2532 2533 // Null check on self without removing any arguments. The argument 2534 // null check technically happens in the wrong place, which can lead to 2535 // invalid stack traces when the primitive is inlined into a method 2536 // which handles NullPointerExceptions. 2537 _sp += nargs; 2538 do_null_check(receiver, T_OBJECT); 2539 _sp -= nargs; 2540 if (stopped()) { 2541 return true; 2542 } 2543 // Heap pointers get a null-check from the interpreter, 2544 // as a courtesy. However, this is not guaranteed by Unsafe, 2545 // and it is not possible to fully distinguish unintended nulls 2546 // from intended ones in this API. 2547 2548 if (is_volatile) { 2549 // We need to emit leading and trailing CPU membars (see below) in 2550 // addition to memory membars when is_volatile. This is a little 2551 // too strong, but avoids the need to insert per-alias-type 2552 // volatile membars (for stores; compare Parse::do_put_xxx), which 2553 // we cannot do effectively here because we probably only have a 2554 // rough approximation of type. 2555 need_mem_bar = true; 2556 // For Stores, place a memory ordering barrier now. 2557 if (is_store) 2558 insert_mem_bar(Op_MemBarRelease); 2559 } 2560 2561 // Memory barrier to prevent normal and 'unsafe' accesses from 2562 // bypassing each other. Happens after null checks, so the 2563 // exception paths do not take memory state from the memory barrier, 2564 // so there's no problems making a strong assert about mixing users 2565 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2566 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2567 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2568 2569 if (!is_store) { 2570 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2571 // load value and push onto stack 2572 switch (type) { 2573 case T_BOOLEAN: 2574 case T_CHAR: 2575 case T_BYTE: 2576 case T_SHORT: 2577 case T_INT: 2578 case T_FLOAT: 2579 push(p); 2580 break; 2581 case T_OBJECT: 2582 if (need_read_barrier) { 2583 insert_pre_barrier(heap_base_oop, offset, p, nargs, !(is_volatile || need_mem_bar)); 2584 } 2585 push(p); 2586 break; 2587 case T_ADDRESS: 2588 // Cast to an int type. 2589 p = _gvn.transform( new (C) CastP2XNode(NULL,p) ); 2590 p = ConvX2L(p); 2591 push_pair(p); 2592 break; 2593 case T_DOUBLE: 2594 case T_LONG: 2595 push_pair( p ); 2596 break; 2597 default: ShouldNotReachHere(); 2598 } 2599 } else { 2600 // place effect of store into memory 2601 switch (type) { 2602 case T_DOUBLE: 2603 val = dstore_rounding(val); 2604 break; 2605 case T_ADDRESS: 2606 // Repackage the long as a pointer. 2607 val = ConvL2X(val); 2608 val = _gvn.transform( new (C) CastX2PNode(val) ); 2609 break; 2610 } 2611 2612 if (type != T_OBJECT ) { 2613 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2614 } else { 2615 // Possibly an oop being stored to Java heap or native memory 2616 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2617 // oop to Java heap. 2618 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2619 } else { 2620 // We can't tell at compile time if we are storing in the Java heap or outside 2621 // of it. So we need to emit code to conditionally do the proper type of 2622 // store. 2623 2624 IdealKit ideal(this); 2625 #define __ ideal. 2626 // QQQ who knows what probability is here?? 2627 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2628 // Sync IdealKit and graphKit. 2629 sync_kit(ideal); 2630 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2631 // Update IdealKit memory. 2632 __ sync_kit(this); 2633 } __ else_(); { 2634 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile); 2635 } __ end_if(); 2636 // Final sync IdealKit and GraphKit. 2637 final_sync(ideal); 2638 #undef __ 2639 } 2640 } 2641 } 2642 2643 if (is_volatile) { 2644 if (!is_store) 2645 insert_mem_bar(Op_MemBarAcquire); 2646 else 2647 insert_mem_bar(Op_MemBarVolatile); 2648 } 2649 2650 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2651 2652 return true; 2653 } 2654 2655 //----------------------------inline_unsafe_prefetch---------------------------- 2656 2657 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2658 #ifndef PRODUCT 2659 { 2660 ResourceMark rm; 2661 // Check the signatures. 2662 ciSignature* sig = signature(); 2663 #ifdef ASSERT 2664 // Object getObject(Object base, int/long offset), etc. 2665 BasicType rtype = sig->return_type()->basic_type(); 2666 if (!is_native_ptr) { 2667 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2668 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2669 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2670 } else { 2671 assert(sig->count() == 1, "native prefetch has 1 argument"); 2672 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2673 } 2674 #endif // ASSERT 2675 } 2676 #endif // !PRODUCT 2677 2678 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2679 2680 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args 2681 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); 2682 2683 debug_only(int saved_sp = _sp); 2684 _sp += nargs; 2685 2686 // Build address expression. See the code in inline_unsafe_access. 2687 Node *adr; 2688 if (!is_native_ptr) { 2689 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2690 Node* offset = pop_pair(); 2691 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2692 Node* base = pop(); 2693 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2694 // to be plain byte offsets, which are also the same as those accepted 2695 // by oopDesc::field_base. 2696 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2697 "fieldOffset must be byte-scaled"); 2698 // 32-bit machines ignore the high half! 2699 offset = ConvL2X(offset); 2700 adr = make_unsafe_address(base, offset); 2701 } else { 2702 Node* ptr = pop_pair(); 2703 // Adjust Java long to machine word: 2704 ptr = ConvL2X(ptr); 2705 adr = make_unsafe_address(NULL, ptr); 2706 } 2707 2708 if (is_static) { 2709 assert(saved_sp == _sp, "must have correct argument count"); 2710 } else { 2711 // Pop receiver last: it was pushed first. 2712 Node *receiver = pop(); 2713 assert(saved_sp == _sp, "must have correct argument count"); 2714 2715 // Null check on self without removing any arguments. The argument 2716 // null check technically happens in the wrong place, which can lead to 2717 // invalid stack traces when the primitive is inlined into a method 2718 // which handles NullPointerExceptions. 2719 _sp += nargs; 2720 do_null_check(receiver, T_OBJECT); 2721 _sp -= nargs; 2722 if (stopped()) { 2723 return true; 2724 } 2725 } 2726 2727 // Generate the read or write prefetch 2728 Node *prefetch; 2729 if (is_store) { 2730 prefetch = new (C) PrefetchWriteNode(i_o(), adr); 2731 } else { 2732 prefetch = new (C) PrefetchReadNode(i_o(), adr); 2733 } 2734 prefetch->init_req(0, control()); 2735 set_i_o(_gvn.transform(prefetch)); 2736 2737 return true; 2738 } 2739 2740 //----------------------------inline_unsafe_load_store---------------------------- 2741 2742 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2743 // This basic scheme here is the same as inline_unsafe_access, but 2744 // differs in enough details that combining them would make the code 2745 // overly confusing. (This is a true fact! I originally combined 2746 // them, but even I was confused by it!) As much code/comments as 2747 // possible are retained from inline_unsafe_access though to make 2748 // the correspondences clearer. - dl 2749 2750 if (callee()->is_static()) return false; // caller must have the capability! 2751 2752 #ifndef PRODUCT 2753 BasicType rtype; 2754 { 2755 ResourceMark rm; 2756 ciSignature* sig = signature(); 2757 rtype = sig->return_type()->basic_type(); 2758 if (kind == LS_xadd || kind == LS_xchg) { 2759 // Check the signatures. 2760 #ifdef ASSERT 2761 assert(rtype == type, "get and set must return the expected type"); 2762 assert(sig->count() == 3, "get and set has 3 arguments"); 2763 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2764 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2765 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2766 #endif // ASSERT 2767 } else if (kind == LS_cmpxchg) { 2768 // Check the signatures. 2769 #ifdef ASSERT 2770 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2771 assert(sig->count() == 4, "CAS has 4 arguments"); 2772 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2773 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2774 #endif // ASSERT 2775 } else { 2776 ShouldNotReachHere(); 2777 } 2778 } 2779 #endif //PRODUCT 2780 2781 // number of stack slots per value argument (1 or 2) 2782 int type_words = type2size[type]; 2783 2784 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2785 2786 // Argument words: "this" plus oop plus offset (plus oldvalue) plus newvalue/delta; 2787 int nargs = 1 + 1 + 2 + ((kind == LS_cmpxchg) ? type_words : 0) + type_words; 2788 2789 // pop arguments: newval, offset, base, and receiver 2790 debug_only(int saved_sp = _sp); 2791 _sp += nargs; 2792 Node* newval = (type_words == 1) ? pop() : pop_pair(); 2793 Node* oldval = (kind == LS_cmpxchg) ? ((type_words == 1) ? pop() : pop_pair()) : NULL; 2794 Node *offset = pop_pair(); 2795 Node *base = pop(); 2796 Node *receiver = pop(); 2797 assert(saved_sp == _sp, "must have correct argument count"); 2798 2799 // Null check receiver. 2800 _sp += nargs; 2801 do_null_check(receiver, T_OBJECT); 2802 _sp -= nargs; 2803 if (stopped()) { 2804 return true; 2805 } 2806 2807 // Build field offset expression. 2808 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2809 // to be plain byte offsets, which are also the same as those accepted 2810 // by oopDesc::field_base. 2811 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2812 // 32-bit machines ignore the high half of long offsets 2813 offset = ConvL2X(offset); 2814 Node* adr = make_unsafe_address(base, offset); 2815 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2816 2817 // For CAS, unlike inline_unsafe_access, there seems no point in 2818 // trying to refine types. Just use the coarse types here. 2819 const Type *value_type = Type::get_const_basic_type(type); 2820 Compile::AliasType* alias_type = C->alias_type(adr_type); 2821 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2822 2823 if (kind == LS_xchg && type == T_OBJECT) { 2824 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2825 if (tjp != NULL) { 2826 value_type = tjp; 2827 } 2828 } 2829 2830 int alias_idx = C->get_alias_index(adr_type); 2831 2832 // Memory-model-wise, a LoadStore acts like a little synchronized 2833 // block, so needs barriers on each side. These don't translate 2834 // into actual barriers on most machines, but we still need rest of 2835 // compiler to respect ordering. 2836 2837 insert_mem_bar(Op_MemBarRelease); 2838 insert_mem_bar(Op_MemBarCPUOrder); 2839 2840 // 4984716: MemBars must be inserted before this 2841 // memory node in order to avoid a false 2842 // dependency which will confuse the scheduler. 2843 Node *mem = memory(alias_idx); 2844 2845 // For now, we handle only those cases that actually exist: ints, 2846 // longs, and Object. Adding others should be straightforward. 2847 Node* load_store; 2848 switch(type) { 2849 case T_INT: 2850 if (kind == LS_xadd) { 2851 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type)); 2852 } else if (kind == LS_xchg) { 2853 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type)); 2854 } else if (kind == LS_cmpxchg) { 2855 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2856 } else { 2857 ShouldNotReachHere(); 2858 } 2859 break; 2860 case T_LONG: 2861 if (kind == LS_xadd) { 2862 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2863 } else if (kind == LS_xchg) { 2864 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2865 } else if (kind == LS_cmpxchg) { 2866 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2867 } else { 2868 ShouldNotReachHere(); 2869 } 2870 break; 2871 case T_OBJECT: 2872 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2873 // could be delayed during Parse (for example, in adjust_map_after_if()). 2874 // Execute transformation here to avoid barrier generation in such case. 2875 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2876 newval = _gvn.makecon(TypePtr::NULL_PTR); 2877 2878 // Reference stores need a store barrier. 2879 pre_barrier(true /* do_load*/, 2880 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2881 NULL /* pre_val*/, 2882 T_OBJECT); 2883 #ifdef _LP64 2884 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2885 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2886 if (kind == LS_xchg) { 2887 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr, 2888 newval_enc, adr_type, value_type->make_narrowoop())); 2889 } else { 2890 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2891 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2892 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr, 2893 newval_enc, oldval_enc)); 2894 } 2895 } else 2896 #endif 2897 { 2898 if (kind == LS_xchg) { 2899 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2900 } else { 2901 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2902 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2903 } 2904 } 2905 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2906 break; 2907 default: 2908 ShouldNotReachHere(); 2909 break; 2910 } 2911 2912 // SCMemProjNodes represent the memory state of a LoadStore. Their 2913 // main role is to prevent LoadStore nodes from being optimized away 2914 // when their results aren't used. 2915 Node* proj = _gvn.transform( new (C) SCMemProjNode(load_store)); 2916 set_memory(proj, alias_idx); 2917 2918 // Add the trailing membar surrounding the access 2919 insert_mem_bar(Op_MemBarCPUOrder); 2920 insert_mem_bar(Op_MemBarAcquire); 2921 2922 #ifdef _LP64 2923 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) { 2924 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->bottom_type()->make_ptr())); 2925 } 2926 #endif 2927 2928 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 2929 push_node(load_store->bottom_type()->basic_type(), load_store); 2930 return true; 2931 } 2932 2933 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2934 // This is another variant of inline_unsafe_access, differing in 2935 // that it always issues store-store ("release") barrier and ensures 2936 // store-atomicity (which only matters for "long"). 2937 2938 if (callee()->is_static()) return false; // caller must have the capability! 2939 2940 #ifndef PRODUCT 2941 { 2942 ResourceMark rm; 2943 // Check the signatures. 2944 ciSignature* sig = signature(); 2945 #ifdef ASSERT 2946 BasicType rtype = sig->return_type()->basic_type(); 2947 assert(rtype == T_VOID, "must return void"); 2948 assert(sig->count() == 3, "has 3 arguments"); 2949 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2950 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2951 #endif // ASSERT 2952 } 2953 #endif //PRODUCT 2954 2955 // number of stack slots per value argument (1 or 2) 2956 int type_words = type2size[type]; 2957 2958 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2959 2960 // Argument words: "this" plus oop plus offset plus value; 2961 int nargs = 1 + 1 + 2 + type_words; 2962 2963 // pop arguments: val, offset, base, and receiver 2964 debug_only(int saved_sp = _sp); 2965 _sp += nargs; 2966 Node* val = (type_words == 1) ? pop() : pop_pair(); 2967 Node *offset = pop_pair(); 2968 Node *base = pop(); 2969 Node *receiver = pop(); 2970 assert(saved_sp == _sp, "must have correct argument count"); 2971 2972 // Null check receiver. 2973 _sp += nargs; 2974 do_null_check(receiver, T_OBJECT); 2975 _sp -= nargs; 2976 if (stopped()) { 2977 return true; 2978 } 2979 2980 // Build field offset expression. 2981 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2982 // 32-bit machines ignore the high half of long offsets 2983 offset = ConvL2X(offset); 2984 Node* adr = make_unsafe_address(base, offset); 2985 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2986 const Type *value_type = Type::get_const_basic_type(type); 2987 Compile::AliasType* alias_type = C->alias_type(adr_type); 2988 2989 insert_mem_bar(Op_MemBarRelease); 2990 insert_mem_bar(Op_MemBarCPUOrder); 2991 // Ensure that the store is atomic for longs: 2992 bool require_atomic_access = true; 2993 Node* store; 2994 if (type == T_OBJECT) // reference stores need a store barrier. 2995 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2996 else { 2997 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2998 } 2999 insert_mem_bar(Op_MemBarCPUOrder); 3000 return true; 3001 } 3002 3003 bool LibraryCallKit::inline_unsafe_allocate() { 3004 if (callee()->is_static()) return false; // caller must have the capability! 3005 int nargs = 1 + 1; 3006 assert(signature()->size() == nargs-1, "alloc has 1 argument"); 3007 null_check_receiver(callee()); // check then ignore argument(0) 3008 _sp += nargs; // set original stack for use by uncommon_trap 3009 Node* cls = do_null_check(argument(1), T_OBJECT); 3010 _sp -= nargs; 3011 if (stopped()) return true; 3012 3013 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 3014 _sp += nargs; // set original stack for use by uncommon_trap 3015 kls = do_null_check(kls, T_OBJECT); 3016 _sp -= nargs; 3017 if (stopped()) return true; // argument was like int.class 3018 3019 // Note: The argument might still be an illegal value like 3020 // Serializable.class or Object[].class. The runtime will handle it. 3021 // But we must make an explicit check for initialization. 3022 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3023 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3024 // can generate code to load it as unsigned byte. 3025 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN); 3026 Node* bits = intcon(InstanceKlass::fully_initialized); 3027 Node* test = _gvn.transform( new (C) SubINode(inst, bits) ); 3028 // The 'test' is non-zero if we need to take a slow path. 3029 3030 Node* obj = new_instance(kls, test); 3031 push(obj); 3032 3033 return true; 3034 } 3035 3036 #ifdef TRACE_HAVE_INTRINSICS 3037 /* 3038 * oop -> myklass 3039 * myklass->trace_id |= USED 3040 * return myklass->trace_id & ~0x3 3041 */ 3042 bool LibraryCallKit::inline_native_classID() { 3043 int nargs = 1 + 1; 3044 null_check_receiver(callee()); // check then ignore argument(0) 3045 _sp += nargs; 3046 Node* cls = do_null_check(argument(1), T_OBJECT); 3047 _sp -= nargs; 3048 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 3049 _sp += nargs; 3050 kls = do_null_check(kls, T_OBJECT); 3051 _sp -= nargs; 3052 ByteSize offset = TRACE_ID_OFFSET; 3053 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3054 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG); 3055 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3056 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits)); 3057 Node* clsused = longcon(0x01l); // set the class bit 3058 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused)); 3059 3060 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3061 store_to_memory(control(), insp, orl, T_LONG, adr_type); 3062 push_pair(andl); 3063 return true; 3064 } 3065 3066 bool LibraryCallKit::inline_native_threadID() { 3067 Node* tls_ptr = NULL; 3068 Node* cur_thr = generate_current_thread(tls_ptr); 3069 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3070 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 3071 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3072 3073 Node* threadid = NULL; 3074 size_t thread_id_size = OSThread::thread_id_size(); 3075 if (thread_id_size == (size_t) BytesPerLong) { 3076 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG)); 3077 push(threadid); 3078 } else if (thread_id_size == (size_t) BytesPerInt) { 3079 threadid = make_load(control(), p, TypeInt::INT, T_INT); 3080 push(threadid); 3081 } else { 3082 ShouldNotReachHere(); 3083 } 3084 return true; 3085 } 3086 #endif 3087 3088 //------------------------inline_native_time_funcs-------------- 3089 // inline code for System.currentTimeMillis() and System.nanoTime() 3090 // these have the same type and signature 3091 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3092 const TypeFunc *tf = OptoRuntime::void_long_Type(); 3093 const TypePtr* no_memory_effects = NULL; 3094 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3095 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0)); 3096 #ifdef ASSERT 3097 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms + 1)); 3098 assert(value_top == top(), "second value must be top"); 3099 #endif 3100 push_pair(value); 3101 return true; 3102 } 3103 3104 //------------------------inline_native_currentThread------------------ 3105 bool LibraryCallKit::inline_native_currentThread() { 3106 Node* junk = NULL; 3107 push(generate_current_thread(junk)); 3108 return true; 3109 } 3110 3111 //------------------------inline_native_isInterrupted------------------ 3112 bool LibraryCallKit::inline_native_isInterrupted() { 3113 const int nargs = 1+1; // receiver + boolean 3114 assert(nargs == arg_size(), "sanity"); 3115 // Add a fast path to t.isInterrupted(clear_int): 3116 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 3117 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3118 // So, in the common case that the interrupt bit is false, 3119 // we avoid making a call into the VM. Even if the interrupt bit 3120 // is true, if the clear_int argument is false, we avoid the VM call. 3121 // However, if the receiver is not currentThread, we must call the VM, 3122 // because there must be some locking done around the operation. 3123 3124 // We only go to the fast case code if we pass two guards. 3125 // Paths which do not pass are accumulated in the slow_region. 3126 RegionNode* slow_region = new (C) RegionNode(1); 3127 record_for_igvn(slow_region); 3128 RegionNode* result_rgn = new (C) RegionNode(1+3); // fast1, fast2, slow 3129 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL); 3130 enum { no_int_result_path = 1, 3131 no_clear_result_path = 2, 3132 slow_result_path = 3 3133 }; 3134 3135 // (a) Receiving thread must be the current thread. 3136 Node* rec_thr = argument(0); 3137 Node* tls_ptr = NULL; 3138 Node* cur_thr = generate_current_thread(tls_ptr); 3139 Node* cmp_thr = _gvn.transform( new (C) CmpPNode(cur_thr, rec_thr) ); 3140 Node* bol_thr = _gvn.transform( new (C) BoolNode(cmp_thr, BoolTest::ne) ); 3141 3142 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); 3143 if (!known_current_thread) 3144 generate_slow_guard(bol_thr, slow_region); 3145 3146 // (b) Interrupt bit on TLS must be false. 3147 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3148 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 3149 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3150 // Set the control input on the field _interrupted read to prevent it floating up. 3151 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 3152 Node* cmp_bit = _gvn.transform( new (C) CmpINode(int_bit, intcon(0)) ); 3153 Node* bol_bit = _gvn.transform( new (C) BoolNode(cmp_bit, BoolTest::ne) ); 3154 3155 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3156 3157 // First fast path: if (!TLS._interrupted) return false; 3158 Node* false_bit = _gvn.transform( new (C) IfFalseNode(iff_bit) ); 3159 result_rgn->init_req(no_int_result_path, false_bit); 3160 result_val->init_req(no_int_result_path, intcon(0)); 3161 3162 // drop through to next case 3163 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)) ); 3164 3165 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3166 Node* clr_arg = argument(1); 3167 Node* cmp_arg = _gvn.transform( new (C) CmpINode(clr_arg, intcon(0)) ); 3168 Node* bol_arg = _gvn.transform( new (C) BoolNode(cmp_arg, BoolTest::ne) ); 3169 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3170 3171 // Second fast path: ... else if (!clear_int) return true; 3172 Node* false_arg = _gvn.transform( new (C) IfFalseNode(iff_arg) ); 3173 result_rgn->init_req(no_clear_result_path, false_arg); 3174 result_val->init_req(no_clear_result_path, intcon(1)); 3175 3176 // drop through to next case 3177 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)) ); 3178 3179 // (d) Otherwise, go to the slow path. 3180 slow_region->add_req(control()); 3181 set_control( _gvn.transform(slow_region) ); 3182 3183 if (stopped()) { 3184 // There is no slow path. 3185 result_rgn->init_req(slow_result_path, top()); 3186 result_val->init_req(slow_result_path, top()); 3187 } else { 3188 // non-virtual because it is a private non-static 3189 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3190 3191 Node* slow_val = set_results_for_java_call(slow_call); 3192 // this->control() comes from set_results_for_java_call 3193 3194 // If we know that the result of the slow call will be true, tell the optimizer! 3195 if (known_current_thread) slow_val = intcon(1); 3196 3197 Node* fast_io = slow_call->in(TypeFunc::I_O); 3198 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3199 // These two phis are pre-filled with copies of of the fast IO and Memory 3200 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3201 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3202 3203 result_rgn->init_req(slow_result_path, control()); 3204 io_phi ->init_req(slow_result_path, i_o()); 3205 mem_phi ->init_req(slow_result_path, reset_memory()); 3206 result_val->init_req(slow_result_path, slow_val); 3207 3208 set_all_memory( _gvn.transform(mem_phi) ); 3209 set_i_o( _gvn.transform(io_phi) ); 3210 } 3211 3212 push_result(result_rgn, result_val); 3213 C->set_has_split_ifs(true); // Has chance for split-if optimization 3214 3215 return true; 3216 } 3217 3218 //---------------------------load_mirror_from_klass---------------------------- 3219 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3220 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3221 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3222 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 3223 } 3224 3225 //-----------------------load_klass_from_mirror_common------------------------- 3226 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3227 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3228 // and branch to the given path on the region. 3229 // If never_see_null, take an uncommon trap on null, so we can optimistically 3230 // compile for the non-null case. 3231 // If the region is NULL, force never_see_null = true. 3232 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3233 bool never_see_null, 3234 int nargs, 3235 RegionNode* region, 3236 int null_path, 3237 int offset) { 3238 if (region == NULL) never_see_null = true; 3239 Node* p = basic_plus_adr(mirror, offset); 3240 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3241 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); 3242 _sp += nargs; // any deopt will start just before call to enclosing method 3243 Node* null_ctl = top(); 3244 kls = null_check_oop(kls, &null_ctl, never_see_null); 3245 if (region != NULL) { 3246 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3247 region->init_req(null_path, null_ctl); 3248 } else { 3249 assert(null_ctl == top(), "no loose ends"); 3250 } 3251 _sp -= nargs; 3252 return kls; 3253 } 3254 3255 //--------------------(inline_native_Class_query helpers)--------------------- 3256 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3257 // Fall through if (mods & mask) == bits, take the guard otherwise. 3258 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3259 // Branch around if the given klass has the given modifier bit set. 3260 // Like generate_guard, adds a new path onto the region. 3261 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3262 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 3263 Node* mask = intcon(modifier_mask); 3264 Node* bits = intcon(modifier_bits); 3265 Node* mbit = _gvn.transform( new (C) AndINode(mods, mask) ); 3266 Node* cmp = _gvn.transform( new (C) CmpINode(mbit, bits) ); 3267 Node* bol = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) ); 3268 return generate_fair_guard(bol, region); 3269 } 3270 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3271 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3272 } 3273 3274 //-------------------------inline_native_Class_query------------------- 3275 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3276 int nargs = 1+0; // just the Class mirror, in most cases 3277 const Type* return_type = TypeInt::BOOL; 3278 Node* prim_return_value = top(); // what happens if it's a primitive class? 3279 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3280 bool expect_prim = false; // most of these guys expect to work on refs 3281 3282 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3283 3284 switch (id) { 3285 case vmIntrinsics::_isInstance: 3286 nargs = 1+1; // the Class mirror, plus the object getting queried about 3287 // nothing is an instance of a primitive type 3288 prim_return_value = intcon(0); 3289 break; 3290 case vmIntrinsics::_getModifiers: 3291 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3292 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3293 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3294 break; 3295 case vmIntrinsics::_isInterface: 3296 prim_return_value = intcon(0); 3297 break; 3298 case vmIntrinsics::_isArray: 3299 prim_return_value = intcon(0); 3300 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3301 break; 3302 case vmIntrinsics::_isPrimitive: 3303 prim_return_value = intcon(1); 3304 expect_prim = true; // obviously 3305 break; 3306 case vmIntrinsics::_getSuperclass: 3307 prim_return_value = null(); 3308 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3309 break; 3310 case vmIntrinsics::_getComponentType: 3311 prim_return_value = null(); 3312 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3313 break; 3314 case vmIntrinsics::_getClassAccessFlags: 3315 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3316 return_type = TypeInt::INT; // not bool! 6297094 3317 break; 3318 default: 3319 ShouldNotReachHere(); 3320 } 3321 3322 Node* mirror = argument(0); 3323 Node* obj = (nargs <= 1)? top(): argument(1); 3324 3325 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3326 if (mirror_con == NULL) return false; // cannot happen? 3327 3328 #ifndef PRODUCT 3329 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 3330 ciType* k = mirror_con->java_mirror_type(); 3331 if (k) { 3332 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3333 k->print_name(); 3334 tty->cr(); 3335 } 3336 } 3337 #endif 3338 3339 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3340 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3341 record_for_igvn(region); 3342 PhiNode* phi = new (C) PhiNode(region, return_type); 3343 3344 // The mirror will never be null of Reflection.getClassAccessFlags, however 3345 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3346 // if it is. See bug 4774291. 3347 3348 // For Reflection.getClassAccessFlags(), the null check occurs in 3349 // the wrong place; see inline_unsafe_access(), above, for a similar 3350 // situation. 3351 _sp += nargs; // set original stack for use by uncommon_trap 3352 mirror = do_null_check(mirror, T_OBJECT); 3353 _sp -= nargs; 3354 // If mirror or obj is dead, only null-path is taken. 3355 if (stopped()) return true; 3356 3357 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3358 3359 // Now load the mirror's klass metaobject, and null-check it. 3360 // Side-effects region with the control path if the klass is null. 3361 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, 3362 region, _prim_path); 3363 // If kls is null, we have a primitive mirror. 3364 phi->init_req(_prim_path, prim_return_value); 3365 if (stopped()) { push_result(region, phi); return true; } 3366 3367 Node* p; // handy temp 3368 Node* null_ctl; 3369 3370 // Now that we have the non-null klass, we can perform the real query. 3371 // For constant classes, the query will constant-fold in LoadNode::Value. 3372 Node* query_value = top(); 3373 switch (id) { 3374 case vmIntrinsics::_isInstance: 3375 // nothing is an instance of a primitive type 3376 _sp += nargs; // gen_instanceof might do an uncommon trap 3377 query_value = gen_instanceof(obj, kls); 3378 _sp -= nargs; 3379 break; 3380 3381 case vmIntrinsics::_getModifiers: 3382 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3383 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3384 break; 3385 3386 case vmIntrinsics::_isInterface: 3387 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3388 if (generate_interface_guard(kls, region) != NULL) 3389 // A guard was added. If the guard is taken, it was an interface. 3390 phi->add_req(intcon(1)); 3391 // If we fall through, it's a plain class. 3392 query_value = intcon(0); 3393 break; 3394 3395 case vmIntrinsics::_isArray: 3396 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3397 if (generate_array_guard(kls, region) != NULL) 3398 // A guard was added. If the guard is taken, it was an array. 3399 phi->add_req(intcon(1)); 3400 // If we fall through, it's a plain class. 3401 query_value = intcon(0); 3402 break; 3403 3404 case vmIntrinsics::_isPrimitive: 3405 query_value = intcon(0); // "normal" path produces false 3406 break; 3407 3408 case vmIntrinsics::_getSuperclass: 3409 // The rules here are somewhat unfortunate, but we can still do better 3410 // with random logic than with a JNI call. 3411 // Interfaces store null or Object as _super, but must report null. 3412 // Arrays store an intermediate super as _super, but must report Object. 3413 // Other types can report the actual _super. 3414 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3415 if (generate_interface_guard(kls, region) != NULL) 3416 // A guard was added. If the guard is taken, it was an interface. 3417 phi->add_req(null()); 3418 if (generate_array_guard(kls, region) != NULL) 3419 // A guard was added. If the guard is taken, it was an array. 3420 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3421 // If we fall through, it's a plain class. Get its _super. 3422 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3423 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); 3424 null_ctl = top(); 3425 kls = null_check_oop(kls, &null_ctl); 3426 if (null_ctl != top()) { 3427 // If the guard is taken, Object.superClass is null (both klass and mirror). 3428 region->add_req(null_ctl); 3429 phi ->add_req(null()); 3430 } 3431 if (!stopped()) { 3432 query_value = load_mirror_from_klass(kls); 3433 } 3434 break; 3435 3436 case vmIntrinsics::_getComponentType: 3437 if (generate_array_guard(kls, region) != NULL) { 3438 // Be sure to pin the oop load to the guard edge just created: 3439 Node* is_array_ctrl = region->in(region->req()-1); 3440 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); 3441 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 3442 phi->add_req(cmo); 3443 } 3444 query_value = null(); // non-array case is null 3445 break; 3446 3447 case vmIntrinsics::_getClassAccessFlags: 3448 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3449 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3450 break; 3451 3452 default: 3453 ShouldNotReachHere(); 3454 } 3455 3456 // Fall-through is the normal case of a query to a real class. 3457 phi->init_req(1, query_value); 3458 region->init_req(1, control()); 3459 3460 push_result(region, phi); 3461 C->set_has_split_ifs(true); // Has chance for split-if optimization 3462 3463 return true; 3464 } 3465 3466 //--------------------------inline_native_subtype_check------------------------ 3467 // This intrinsic takes the JNI calls out of the heart of 3468 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3469 bool LibraryCallKit::inline_native_subtype_check() { 3470 int nargs = 1+1; // the Class mirror, plus the other class getting examined 3471 3472 // Pull both arguments off the stack. 3473 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3474 args[0] = argument(0); 3475 args[1] = argument(1); 3476 Node* klasses[2]; // corresponding Klasses: superk, subk 3477 klasses[0] = klasses[1] = top(); 3478 3479 enum { 3480 // A full decision tree on {superc is prim, subc is prim}: 3481 _prim_0_path = 1, // {P,N} => false 3482 // {P,P} & superc!=subc => false 3483 _prim_same_path, // {P,P} & superc==subc => true 3484 _prim_1_path, // {N,P} => false 3485 _ref_subtype_path, // {N,N} & subtype check wins => true 3486 _both_ref_path, // {N,N} & subtype check loses => false 3487 PATH_LIMIT 3488 }; 3489 3490 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3491 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 3492 record_for_igvn(region); 3493 3494 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3495 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3496 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3497 3498 // First null-check both mirrors and load each mirror's klass metaobject. 3499 int which_arg; 3500 for (which_arg = 0; which_arg <= 1; which_arg++) { 3501 Node* arg = args[which_arg]; 3502 _sp += nargs; // set original stack for use by uncommon_trap 3503 arg = do_null_check(arg, T_OBJECT); 3504 _sp -= nargs; 3505 if (stopped()) break; 3506 args[which_arg] = _gvn.transform(arg); 3507 3508 Node* p = basic_plus_adr(arg, class_klass_offset); 3509 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3510 klasses[which_arg] = _gvn.transform(kls); 3511 } 3512 3513 // Having loaded both klasses, test each for null. 3514 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3515 for (which_arg = 0; which_arg <= 1; which_arg++) { 3516 Node* kls = klasses[which_arg]; 3517 Node* null_ctl = top(); 3518 _sp += nargs; // set original stack for use by uncommon_trap 3519 kls = null_check_oop(kls, &null_ctl, never_see_null); 3520 _sp -= nargs; 3521 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3522 region->init_req(prim_path, null_ctl); 3523 if (stopped()) break; 3524 klasses[which_arg] = kls; 3525 } 3526 3527 if (!stopped()) { 3528 // now we have two reference types, in klasses[0..1] 3529 Node* subk = klasses[1]; // the argument to isAssignableFrom 3530 Node* superk = klasses[0]; // the receiver 3531 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3532 // now we have a successful reference subtype check 3533 region->set_req(_ref_subtype_path, control()); 3534 } 3535 3536 // If both operands are primitive (both klasses null), then 3537 // we must return true when they are identical primitives. 3538 // It is convenient to test this after the first null klass check. 3539 set_control(region->in(_prim_0_path)); // go back to first null check 3540 if (!stopped()) { 3541 // Since superc is primitive, make a guard for the superc==subc case. 3542 Node* cmp_eq = _gvn.transform( new (C) CmpPNode(args[0], args[1]) ); 3543 Node* bol_eq = _gvn.transform( new (C) BoolNode(cmp_eq, BoolTest::eq) ); 3544 generate_guard(bol_eq, region, PROB_FAIR); 3545 if (region->req() == PATH_LIMIT+1) { 3546 // A guard was added. If the added guard is taken, superc==subc. 3547 region->swap_edges(PATH_LIMIT, _prim_same_path); 3548 region->del_req(PATH_LIMIT); 3549 } 3550 region->set_req(_prim_0_path, control()); // Not equal after all. 3551 } 3552 3553 // these are the only paths that produce 'true': 3554 phi->set_req(_prim_same_path, intcon(1)); 3555 phi->set_req(_ref_subtype_path, intcon(1)); 3556 3557 // pull together the cases: 3558 assert(region->req() == PATH_LIMIT, "sane region"); 3559 for (uint i = 1; i < region->req(); i++) { 3560 Node* ctl = region->in(i); 3561 if (ctl == NULL || ctl == top()) { 3562 region->set_req(i, top()); 3563 phi ->set_req(i, top()); 3564 } else if (phi->in(i) == NULL) { 3565 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3566 } 3567 } 3568 3569 set_control(_gvn.transform(region)); 3570 push(_gvn.transform(phi)); 3571 3572 return true; 3573 } 3574 3575 //---------------------generate_array_guard_common------------------------ 3576 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3577 bool obj_array, bool not_array) { 3578 // If obj_array/non_array==false/false: 3579 // Branch around if the given klass is in fact an array (either obj or prim). 3580 // If obj_array/non_array==false/true: 3581 // Branch around if the given klass is not an array klass of any kind. 3582 // If obj_array/non_array==true/true: 3583 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3584 // If obj_array/non_array==true/false: 3585 // Branch around if the kls is an oop array (Object[] or subtype) 3586 // 3587 // Like generate_guard, adds a new path onto the region. 3588 jint layout_con = 0; 3589 Node* layout_val = get_layout_helper(kls, layout_con); 3590 if (layout_val == NULL) { 3591 bool query = (obj_array 3592 ? Klass::layout_helper_is_objArray(layout_con) 3593 : Klass::layout_helper_is_array(layout_con)); 3594 if (query == not_array) { 3595 return NULL; // never a branch 3596 } else { // always a branch 3597 Node* always_branch = control(); 3598 if (region != NULL) 3599 region->add_req(always_branch); 3600 set_control(top()); 3601 return always_branch; 3602 } 3603 } 3604 // Now test the correct condition. 3605 jint nval = (obj_array 3606 ? ((jint)Klass::_lh_array_tag_type_value 3607 << Klass::_lh_array_tag_shift) 3608 : Klass::_lh_neutral_value); 3609 Node* cmp = _gvn.transform( new(C) CmpINode(layout_val, intcon(nval)) ); 3610 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3611 // invert the test if we are looking for a non-array 3612 if (not_array) btest = BoolTest(btest).negate(); 3613 Node* bol = _gvn.transform( new(C) BoolNode(cmp, btest) ); 3614 return generate_fair_guard(bol, region); 3615 } 3616 3617 3618 //-----------------------inline_native_newArray-------------------------- 3619 bool LibraryCallKit::inline_native_newArray() { 3620 int nargs = 2; 3621 Node* mirror = argument(0); 3622 Node* count_val = argument(1); 3623 3624 _sp += nargs; // set original stack for use by uncommon_trap 3625 mirror = do_null_check(mirror, T_OBJECT); 3626 _sp -= nargs; 3627 // If mirror or obj is dead, only null-path is taken. 3628 if (stopped()) return true; 3629 3630 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3631 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3632 PhiNode* result_val = new(C) PhiNode(result_reg, 3633 TypeInstPtr::NOTNULL); 3634 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3635 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3636 TypePtr::BOTTOM); 3637 3638 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3639 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3640 nargs, 3641 result_reg, _slow_path); 3642 Node* normal_ctl = control(); 3643 Node* no_array_ctl = result_reg->in(_slow_path); 3644 3645 // Generate code for the slow case. We make a call to newArray(). 3646 set_control(no_array_ctl); 3647 if (!stopped()) { 3648 // Either the input type is void.class, or else the 3649 // array klass has not yet been cached. Either the 3650 // ensuing call will throw an exception, or else it 3651 // will cache the array klass for next time. 3652 PreserveJVMState pjvms(this); 3653 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3654 Node* slow_result = set_results_for_java_call(slow_call); 3655 // this->control() comes from set_results_for_java_call 3656 result_reg->set_req(_slow_path, control()); 3657 result_val->set_req(_slow_path, slow_result); 3658 result_io ->set_req(_slow_path, i_o()); 3659 result_mem->set_req(_slow_path, reset_memory()); 3660 } 3661 3662 set_control(normal_ctl); 3663 if (!stopped()) { 3664 // Normal case: The array type has been cached in the java.lang.Class. 3665 // The following call works fine even if the array type is polymorphic. 3666 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3667 Node* obj = new_array(klass_node, count_val, nargs); 3668 result_reg->init_req(_normal_path, control()); 3669 result_val->init_req(_normal_path, obj); 3670 result_io ->init_req(_normal_path, i_o()); 3671 result_mem->init_req(_normal_path, reset_memory()); 3672 } 3673 3674 // Return the combined state. 3675 set_i_o( _gvn.transform(result_io) ); 3676 set_all_memory( _gvn.transform(result_mem) ); 3677 push_result(result_reg, result_val); 3678 C->set_has_split_ifs(true); // Has chance for split-if optimization 3679 3680 return true; 3681 } 3682 3683 //----------------------inline_native_getLength-------------------------- 3684 bool LibraryCallKit::inline_native_getLength() { 3685 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3686 3687 int nargs = 1; 3688 Node* array = argument(0); 3689 3690 _sp += nargs; // set original stack for use by uncommon_trap 3691 array = do_null_check(array, T_OBJECT); 3692 _sp -= nargs; 3693 3694 // If array is dead, only null-path is taken. 3695 if (stopped()) return true; 3696 3697 // Deoptimize if it is a non-array. 3698 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3699 3700 if (non_array != NULL) { 3701 PreserveJVMState pjvms(this); 3702 set_control(non_array); 3703 _sp += nargs; // push the arguments back on the stack 3704 uncommon_trap(Deoptimization::Reason_intrinsic, 3705 Deoptimization::Action_maybe_recompile); 3706 } 3707 3708 // If control is dead, only non-array-path is taken. 3709 if (stopped()) return true; 3710 3711 // The works fine even if the array type is polymorphic. 3712 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3713 push( load_array_length(array) ); 3714 3715 C->set_has_split_ifs(true); // Has chance for split-if optimization 3716 3717 return true; 3718 } 3719 3720 //------------------------inline_array_copyOf---------------------------- 3721 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3722 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3723 3724 // Restore the stack and pop off the arguments. 3725 int nargs = 3 + (is_copyOfRange? 1: 0); 3726 Node* original = argument(0); 3727 Node* start = is_copyOfRange? argument(1): intcon(0); 3728 Node* end = is_copyOfRange? argument(2): argument(1); 3729 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3730 3731 Node* newcopy; 3732 3733 //set the original stack and the reexecute bit for the interpreter to reexecute 3734 //the bytecode that invokes Arrays.copyOf if deoptimization happens 3735 { PreserveReexecuteState preexecs(this); 3736 _sp += nargs; 3737 jvms()->set_should_reexecute(true); 3738 3739 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); 3740 original = do_null_check(original, T_OBJECT); 3741 3742 // Check if a null path was taken unconditionally. 3743 if (stopped()) return true; 3744 3745 Node* orig_length = load_array_length(original); 3746 3747 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0, 3748 NULL, 0); 3749 klass_node = do_null_check(klass_node, T_OBJECT); 3750 3751 RegionNode* bailout = new (C) RegionNode(1); 3752 record_for_igvn(bailout); 3753 3754 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3755 // Bail out if that is so. 3756 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3757 if (not_objArray != NULL) { 3758 // Improve the klass node's type from the new optimistic assumption: 3759 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3760 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3761 Node* cast = new (C) CastPPNode(klass_node, akls); 3762 cast->init_req(0, control()); 3763 klass_node = _gvn.transform(cast); 3764 } 3765 3766 // Bail out if either start or end is negative. 3767 generate_negative_guard(start, bailout, &start); 3768 generate_negative_guard(end, bailout, &end); 3769 3770 Node* length = end; 3771 if (_gvn.type(start) != TypeInt::ZERO) { 3772 length = _gvn.transform( new (C) SubINode(end, start) ); 3773 } 3774 3775 // Bail out if length is negative. 3776 // Without this the new_array would throw 3777 // NegativeArraySizeException but IllegalArgumentException is what 3778 // should be thrown 3779 generate_negative_guard(length, bailout, &length); 3780 3781 if (bailout->req() > 1) { 3782 PreserveJVMState pjvms(this); 3783 set_control( _gvn.transform(bailout) ); 3784 uncommon_trap(Deoptimization::Reason_intrinsic, 3785 Deoptimization::Action_maybe_recompile); 3786 } 3787 3788 if (!stopped()) { 3789 3790 // How many elements will we copy from the original? 3791 // The answer is MinI(orig_length - start, length). 3792 Node* orig_tail = _gvn.transform( new(C) SubINode(orig_length, start) ); 3793 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3794 3795 newcopy = new_array(klass_node, length, 0); 3796 3797 // Generate a direct call to the right arraycopy function(s). 3798 // We know the copy is disjoint but we might not know if the 3799 // oop stores need checking. 3800 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3801 // This will fail a store-check if x contains any non-nulls. 3802 bool disjoint_bases = true; 3803 // if start > orig_length then the length of the copy may be 3804 // negative. 3805 bool length_never_negative = !is_copyOfRange; 3806 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3807 original, start, newcopy, intcon(0), moved, 3808 disjoint_bases, length_never_negative); 3809 } 3810 } //original reexecute and sp are set back here 3811 3812 if(!stopped()) { 3813 push(newcopy); 3814 } 3815 3816 C->set_has_split_ifs(true); // Has chance for split-if optimization 3817 3818 return true; 3819 } 3820 3821 3822 //----------------------generate_virtual_guard--------------------------- 3823 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3824 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3825 RegionNode* slow_region) { 3826 ciMethod* method = callee(); 3827 int vtable_index = method->vtable_index(); 3828 // Get the Method* out of the appropriate vtable entry. 3829 int entry_offset = (InstanceKlass::vtable_start_offset() + 3830 vtable_index*vtableEntry::size()) * wordSize + 3831 vtableEntry::method_offset_in_bytes(); 3832 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3833 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS); 3834 3835 // Compare the target method with the expected method (e.g., Object.hashCode). 3836 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3837 3838 Node* native_call = makecon(native_call_addr); 3839 Node* chk_native = _gvn.transform( new(C) CmpPNode(target_call, native_call) ); 3840 Node* test_native = _gvn.transform( new(C) BoolNode(chk_native, BoolTest::ne) ); 3841 3842 return generate_slow_guard(test_native, slow_region); 3843 } 3844 3845 //-----------------------generate_method_call---------------------------- 3846 // Use generate_method_call to make a slow-call to the real 3847 // method if the fast path fails. An alternative would be to 3848 // use a stub like OptoRuntime::slow_arraycopy_Java. 3849 // This only works for expanding the current library call, 3850 // not another intrinsic. (E.g., don't use this for making an 3851 // arraycopy call inside of the copyOf intrinsic.) 3852 CallJavaNode* 3853 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3854 // When compiling the intrinsic method itself, do not use this technique. 3855 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3856 3857 ciMethod* method = callee(); 3858 // ensure the JVMS we have will be correct for this call 3859 guarantee(method_id == method->intrinsic_id(), "must match"); 3860 3861 const TypeFunc* tf = TypeFunc::make(method); 3862 CallJavaNode* slow_call; 3863 if (is_static) { 3864 assert(!is_virtual, ""); 3865 slow_call = new(C) CallStaticJavaNode(tf, 3866 SharedRuntime::get_resolve_static_call_stub(), 3867 method, bci()); 3868 } else if (is_virtual) { 3869 null_check_receiver(method); 3870 int vtable_index = Method::invalid_vtable_index; 3871 if (UseInlineCaches) { 3872 // Suppress the vtable call 3873 } else { 3874 // hashCode and clone are not a miranda methods, 3875 // so the vtable index is fixed. 3876 // No need to use the linkResolver to get it. 3877 vtable_index = method->vtable_index(); 3878 } 3879 slow_call = new(C) CallDynamicJavaNode(tf, 3880 SharedRuntime::get_resolve_virtual_call_stub(), 3881 method, vtable_index, bci()); 3882 } else { // neither virtual nor static: opt_virtual 3883 null_check_receiver(method); 3884 slow_call = new(C) CallStaticJavaNode(tf, 3885 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3886 method, bci()); 3887 slow_call->set_optimized_virtual(true); 3888 } 3889 set_arguments_for_java_call(slow_call); 3890 set_edges_for_java_call(slow_call); 3891 return slow_call; 3892 } 3893 3894 3895 //------------------------------inline_native_hashcode-------------------- 3896 // Build special case code for calls to hashCode on an object. 3897 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3898 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3899 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3900 3901 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3902 3903 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3904 PhiNode* result_val = new(C) PhiNode(result_reg, 3905 TypeInt::INT); 3906 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3907 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3908 TypePtr::BOTTOM); 3909 Node* obj = NULL; 3910 if (!is_static) { 3911 // Check for hashing null object 3912 obj = null_check_receiver(callee()); 3913 if (stopped()) return true; // unconditionally null 3914 result_reg->init_req(_null_path, top()); 3915 result_val->init_req(_null_path, top()); 3916 } else { 3917 // Do a null check, and return zero if null. 3918 // System.identityHashCode(null) == 0 3919 obj = argument(0); 3920 Node* null_ctl = top(); 3921 obj = null_check_oop(obj, &null_ctl); 3922 result_reg->init_req(_null_path, null_ctl); 3923 result_val->init_req(_null_path, _gvn.intcon(0)); 3924 } 3925 3926 // Unconditionally null? Then return right away. 3927 if (stopped()) { 3928 set_control( result_reg->in(_null_path) ); 3929 if (!stopped()) 3930 push( result_val ->in(_null_path) ); 3931 return true; 3932 } 3933 3934 // After null check, get the object's klass. 3935 Node* obj_klass = load_object_klass(obj); 3936 3937 // This call may be virtual (invokevirtual) or bound (invokespecial). 3938 // For each case we generate slightly different code. 3939 3940 // We only go to the fast case code if we pass a number of guards. The 3941 // paths which do not pass are accumulated in the slow_region. 3942 RegionNode* slow_region = new (C) RegionNode(1); 3943 record_for_igvn(slow_region); 3944 3945 // If this is a virtual call, we generate a funny guard. We pull out 3946 // the vtable entry corresponding to hashCode() from the target object. 3947 // If the target method which we are calling happens to be the native 3948 // Object hashCode() method, we pass the guard. We do not need this 3949 // guard for non-virtual calls -- the caller is known to be the native 3950 // Object hashCode(). 3951 if (is_virtual) { 3952 generate_virtual_guard(obj_klass, slow_region); 3953 } 3954 3955 // Get the header out of the object, use LoadMarkNode when available 3956 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3957 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type()); 3958 3959 // Test the header to see if it is unlocked. 3960 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3961 Node *lmasked_header = _gvn.transform( new (C) AndXNode(header, lock_mask) ); 3962 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3963 Node *chk_unlocked = _gvn.transform( new (C) CmpXNode( lmasked_header, unlocked_val)); 3964 Node *test_unlocked = _gvn.transform( new (C) BoolNode( chk_unlocked, BoolTest::ne) ); 3965 3966 generate_slow_guard(test_unlocked, slow_region); 3967 3968 // Get the hash value and check to see that it has been properly assigned. 3969 // We depend on hash_mask being at most 32 bits and avoid the use of 3970 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3971 // vm: see markOop.hpp. 3972 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3973 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3974 Node *hshifted_header= _gvn.transform( new (C) URShiftXNode(header, hash_shift) ); 3975 // This hack lets the hash bits live anywhere in the mark object now, as long 3976 // as the shift drops the relevant bits into the low 32 bits. Note that 3977 // Java spec says that HashCode is an int so there's no point in capturing 3978 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3979 hshifted_header = ConvX2I(hshifted_header); 3980 Node *hash_val = _gvn.transform( new (C) AndINode(hshifted_header, hash_mask) ); 3981 3982 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3983 Node *chk_assigned = _gvn.transform( new (C) CmpINode( hash_val, no_hash_val)); 3984 Node *test_assigned = _gvn.transform( new (C) BoolNode( chk_assigned, BoolTest::eq) ); 3985 3986 generate_slow_guard(test_assigned, slow_region); 3987 3988 Node* init_mem = reset_memory(); 3989 // fill in the rest of the null path: 3990 result_io ->init_req(_null_path, i_o()); 3991 result_mem->init_req(_null_path, init_mem); 3992 3993 result_val->init_req(_fast_path, hash_val); 3994 result_reg->init_req(_fast_path, control()); 3995 result_io ->init_req(_fast_path, i_o()); 3996 result_mem->init_req(_fast_path, init_mem); 3997 3998 // Generate code for the slow case. We make a call to hashCode(). 3999 set_control(_gvn.transform(slow_region)); 4000 if (!stopped()) { 4001 // No need for PreserveJVMState, because we're using up the present state. 4002 set_all_memory(init_mem); 4003 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; 4004 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; 4005 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4006 Node* slow_result = set_results_for_java_call(slow_call); 4007 // this->control() comes from set_results_for_java_call 4008 result_reg->init_req(_slow_path, control()); 4009 result_val->init_req(_slow_path, slow_result); 4010 result_io ->set_req(_slow_path, i_o()); 4011 result_mem ->set_req(_slow_path, reset_memory()); 4012 } 4013 4014 // Return the combined state. 4015 set_i_o( _gvn.transform(result_io) ); 4016 set_all_memory( _gvn.transform(result_mem) ); 4017 push_result(result_reg, result_val); 4018 4019 return true; 4020 } 4021 4022 //---------------------------inline_native_getClass---------------------------- 4023 // Build special case code for calls to getClass on an object. 4024 bool LibraryCallKit::inline_native_getClass() { 4025 Node* obj = null_check_receiver(callee()); 4026 if (stopped()) return true; 4027 push( load_mirror_from_klass(load_object_klass(obj)) ); 4028 return true; 4029 } 4030 4031 //-----------------inline_native_Reflection_getCallerClass--------------------- 4032 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4033 // 4034 // NOTE that this code must perform the same logic as 4035 // vframeStream::security_get_caller_frame in that it must skip 4036 // Method.invoke() and auxiliary frames. 4037 4038 4039 4040 4041 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4042 ciMethod* method = callee(); 4043 4044 #ifndef PRODUCT 4045 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 4046 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4047 } 4048 #endif 4049 4050 debug_only(int saved_sp = _sp); 4051 4052 // Argument words: (int depth) 4053 int nargs = 1; 4054 4055 _sp += nargs; 4056 Node* caller_depth_node = pop(); 4057 4058 assert(saved_sp == _sp, "must have correct argument count"); 4059 4060 // The depth value must be a constant in order for the runtime call 4061 // to be eliminated. 4062 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); 4063 if (caller_depth_type == NULL || !caller_depth_type->is_con()) { 4064 #ifndef PRODUCT 4065 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 4066 tty->print_cr(" Bailing out because caller depth was not a constant"); 4067 } 4068 #endif 4069 return false; 4070 } 4071 // Note that the JVM state at this point does not include the 4072 // getCallerClass() frame which we are trying to inline. The 4073 // semantics of getCallerClass(), however, are that the "first" 4074 // frame is the getCallerClass() frame, so we subtract one from the 4075 // requested depth before continuing. We don't inline requests of 4076 // getCallerClass(0). 4077 int caller_depth = caller_depth_type->get_con() - 1; 4078 if (caller_depth < 0) { 4079 #ifndef PRODUCT 4080 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 4081 tty->print_cr(" Bailing out because caller depth was %d", caller_depth); 4082 } 4083 #endif 4084 return false; 4085 } 4086 4087 if (!jvms()->has_method()) { 4088 #ifndef PRODUCT 4089 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 4090 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4091 } 4092 #endif 4093 return false; 4094 } 4095 int _depth = jvms()->depth(); // cache call chain depth 4096 4097 // Walk back up the JVM state to find the caller at the required 4098 // depth. NOTE that this code must perform the same logic as 4099 // vframeStream::security_get_caller_frame in that it must skip 4100 // Method.invoke() and auxiliary frames. Note also that depth is 4101 // 1-based (1 is the bottom of the inlining). 4102 int inlining_depth = _depth; 4103 JVMState* caller_jvms = NULL; 4104 4105 if (inlining_depth > 0) { 4106 caller_jvms = jvms(); 4107 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); 4108 do { 4109 // The following if-tests should be performed in this order 4110 if (is_method_invoke_or_aux_frame(caller_jvms)) { 4111 // Skip a Method.invoke() or auxiliary frame 4112 } else if (caller_depth > 0) { 4113 // Skip real frame 4114 --caller_depth; 4115 } else { 4116 // We're done: reached desired caller after skipping. 4117 break; 4118 } 4119 caller_jvms = caller_jvms->caller(); 4120 --inlining_depth; 4121 } while (inlining_depth > 0); 4122 } 4123 4124 if (inlining_depth == 0) { 4125 #ifndef PRODUCT 4126 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 4127 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); 4128 tty->print_cr(" JVM state at this point:"); 4129 for (int i = _depth; i >= 1; i--) { 4130 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 4131 } 4132 } 4133 #endif 4134 return false; // Reached end of inlining 4135 } 4136 4137 // Acquire method holder as java.lang.Class 4138 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4139 ciInstance* caller_mirror = caller_klass->java_mirror(); 4140 // Push this as a constant 4141 push(makecon(TypeInstPtr::make(caller_mirror))); 4142 #ifndef PRODUCT 4143 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 4144 tty->print_cr(" Succeeded: caller = %s.%s, caller depth = %d, depth = %d", caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), caller_depth_type->get_con(), _depth); 4145 tty->print_cr(" JVM state at this point:"); 4146 for (int i = _depth; i >= 1; i--) { 4147 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 4148 } 4149 } 4150 #endif 4151 return true; 4152 } 4153 4154 // Helper routine for above 4155 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { 4156 ciMethod* method = jvms->method(); 4157 4158 // Is this the Method.invoke method itself? 4159 if (method->intrinsic_id() == vmIntrinsics::_invoke) 4160 return true; 4161 4162 // Is this a helper, defined somewhere underneath MethodAccessorImpl. 4163 ciKlass* k = method->holder(); 4164 if (k->is_instance_klass()) { 4165 ciInstanceKlass* ik = k->as_instance_klass(); 4166 for (; ik != NULL; ik = ik->super()) { 4167 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && 4168 ik == env()->find_system_klass(ik->name())) { 4169 return true; 4170 } 4171 } 4172 } 4173 else if (method->is_method_handle_intrinsic() || 4174 method->is_compiled_lambda_form()) { 4175 // This is an internal adapter frame from the MethodHandleCompiler -- skip it 4176 return true; 4177 } 4178 4179 return false; 4180 } 4181 4182 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4183 // restore the arguments 4184 _sp += arg_size(); 4185 4186 switch (id) { 4187 case vmIntrinsics::_floatToRawIntBits: 4188 push(_gvn.transform( new (C) MoveF2INode(pop()))); 4189 break; 4190 4191 case vmIntrinsics::_intBitsToFloat: 4192 push(_gvn.transform( new (C) MoveI2FNode(pop()))); 4193 break; 4194 4195 case vmIntrinsics::_doubleToRawLongBits: 4196 push_pair(_gvn.transform( new (C) MoveD2LNode(pop_pair()))); 4197 break; 4198 4199 case vmIntrinsics::_longBitsToDouble: 4200 push_pair(_gvn.transform( new (C) MoveL2DNode(pop_pair()))); 4201 break; 4202 4203 case vmIntrinsics::_doubleToLongBits: { 4204 Node* value = pop_pair(); 4205 4206 // two paths (plus control) merge in a wood 4207 RegionNode *r = new (C) RegionNode(3); 4208 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 4209 4210 Node *cmpisnan = _gvn.transform( new (C) CmpDNode(value, value)); 4211 // Build the boolean node 4212 Node *bolisnan = _gvn.transform( new (C) BoolNode( cmpisnan, BoolTest::ne ) ); 4213 4214 // Branch either way. 4215 // NaN case is less traveled, which makes all the difference. 4216 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4217 Node *opt_isnan = _gvn.transform(ifisnan); 4218 assert( opt_isnan->is_If(), "Expect an IfNode"); 4219 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4220 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) ); 4221 4222 set_control(iftrue); 4223 4224 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4225 Node *slow_result = longcon(nan_bits); // return NaN 4226 phi->init_req(1, _gvn.transform( slow_result )); 4227 r->init_req(1, iftrue); 4228 4229 // Else fall through 4230 Node *iffalse = _gvn.transform( new (C) IfFalseNode(opt_ifisnan) ); 4231 set_control(iffalse); 4232 4233 phi->init_req(2, _gvn.transform( new (C) MoveD2LNode(value))); 4234 r->init_req(2, iffalse); 4235 4236 // Post merge 4237 set_control(_gvn.transform(r)); 4238 record_for_igvn(r); 4239 4240 Node* result = _gvn.transform(phi); 4241 assert(result->bottom_type()->isa_long(), "must be"); 4242 push_pair(result); 4243 4244 C->set_has_split_ifs(true); // Has chance for split-if optimization 4245 4246 break; 4247 } 4248 4249 case vmIntrinsics::_floatToIntBits: { 4250 Node* value = pop(); 4251 4252 // two paths (plus control) merge in a wood 4253 RegionNode *r = new (C) RegionNode(3); 4254 Node *phi = new (C) PhiNode(r, TypeInt::INT); 4255 4256 Node *cmpisnan = _gvn.transform( new (C) CmpFNode(value, value)); 4257 // Build the boolean node 4258 Node *bolisnan = _gvn.transform( new (C) BoolNode( cmpisnan, BoolTest::ne ) ); 4259 4260 // Branch either way. 4261 // NaN case is less traveled, which makes all the difference. 4262 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4263 Node *opt_isnan = _gvn.transform(ifisnan); 4264 assert( opt_isnan->is_If(), "Expect an IfNode"); 4265 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4266 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) ); 4267 4268 set_control(iftrue); 4269 4270 static const jint nan_bits = 0x7fc00000; 4271 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4272 phi->init_req(1, _gvn.transform( slow_result )); 4273 r->init_req(1, iftrue); 4274 4275 // Else fall through 4276 Node *iffalse = _gvn.transform( new (C) IfFalseNode(opt_ifisnan) ); 4277 set_control(iffalse); 4278 4279 phi->init_req(2, _gvn.transform( new (C) MoveF2INode(value))); 4280 r->init_req(2, iffalse); 4281 4282 // Post merge 4283 set_control(_gvn.transform(r)); 4284 record_for_igvn(r); 4285 4286 Node* result = _gvn.transform(phi); 4287 assert(result->bottom_type()->isa_int(), "must be"); 4288 push(result); 4289 4290 C->set_has_split_ifs(true); // Has chance for split-if optimization 4291 4292 break; 4293 } 4294 4295 default: 4296 ShouldNotReachHere(); 4297 } 4298 4299 return true; 4300 } 4301 4302 #ifdef _LP64 4303 #define XTOP ,top() /*additional argument*/ 4304 #else //_LP64 4305 #define XTOP /*no additional argument*/ 4306 #endif //_LP64 4307 4308 //----------------------inline_unsafe_copyMemory------------------------- 4309 bool LibraryCallKit::inline_unsafe_copyMemory() { 4310 if (callee()->is_static()) return false; // caller must have the capability! 4311 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) 4312 assert(signature()->size() == nargs-1, "copy has 5 arguments"); 4313 null_check_receiver(callee()); // check then ignore argument(0) 4314 if (stopped()) return true; 4315 4316 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4317 4318 Node* src_ptr = argument(1); 4319 Node* src_off = ConvL2X(argument(2)); 4320 assert(argument(3)->is_top(), "2nd half of long"); 4321 Node* dst_ptr = argument(4); 4322 Node* dst_off = ConvL2X(argument(5)); 4323 assert(argument(6)->is_top(), "2nd half of long"); 4324 Node* size = ConvL2X(argument(7)); 4325 assert(argument(8)->is_top(), "2nd half of long"); 4326 4327 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4328 "fieldOffset must be byte-scaled"); 4329 4330 Node* src = make_unsafe_address(src_ptr, src_off); 4331 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4332 4333 // Conservatively insert a memory barrier on all memory slices. 4334 // Do not let writes of the copy source or destination float below the copy. 4335 insert_mem_bar(Op_MemBarCPUOrder); 4336 4337 // Call it. Note that the length argument is not scaled. 4338 make_runtime_call(RC_LEAF|RC_NO_FP, 4339 OptoRuntime::fast_arraycopy_Type(), 4340 StubRoutines::unsafe_arraycopy(), 4341 "unsafe_arraycopy", 4342 TypeRawPtr::BOTTOM, 4343 src, dst, size XTOP); 4344 4345 // Do not let reads of the copy destination float above the copy. 4346 insert_mem_bar(Op_MemBarCPUOrder); 4347 4348 return true; 4349 } 4350 4351 //------------------------clone_coping----------------------------------- 4352 // Helper function for inline_native_clone. 4353 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4354 assert(obj_size != NULL, ""); 4355 Node* raw_obj = alloc_obj->in(1); 4356 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4357 4358 AllocateNode* alloc = NULL; 4359 if (ReduceBulkZeroing) { 4360 // We will be completely responsible for initializing this object - 4361 // mark Initialize node as complete. 4362 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4363 // The object was just allocated - there should be no any stores! 4364 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4365 // Mark as complete_with_arraycopy so that on AllocateNode 4366 // expansion, we know this AllocateNode is initialized by an array 4367 // copy and a StoreStore barrier exists after the array copy. 4368 alloc->initialization()->set_complete_with_arraycopy(); 4369 } 4370 4371 // Copy the fastest available way. 4372 // TODO: generate fields copies for small objects instead. 4373 Node* src = obj; 4374 Node* dest = alloc_obj; 4375 Node* size = _gvn.transform(obj_size); 4376 4377 // Exclude the header but include array length to copy by 8 bytes words. 4378 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4379 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4380 instanceOopDesc::base_offset_in_bytes(); 4381 // base_off: 4382 // 8 - 32-bit VM 4383 // 12 - 64-bit VM, compressed klass 4384 // 16 - 64-bit VM, normal klass 4385 if (base_off % BytesPerLong != 0) { 4386 assert(UseCompressedKlassPointers, ""); 4387 if (is_array) { 4388 // Exclude length to copy by 8 bytes words. 4389 base_off += sizeof(int); 4390 } else { 4391 // Include klass to copy by 8 bytes words. 4392 base_off = instanceOopDesc::klass_offset_in_bytes(); 4393 } 4394 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4395 } 4396 src = basic_plus_adr(src, base_off); 4397 dest = basic_plus_adr(dest, base_off); 4398 4399 // Compute the length also, if needed: 4400 Node* countx = size; 4401 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(base_off)) ); 4402 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4403 4404 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4405 bool disjoint_bases = true; 4406 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4407 src, NULL, dest, NULL, countx, 4408 /*dest_uninitialized*/true); 4409 4410 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4411 if (card_mark) { 4412 assert(!is_array, ""); 4413 // Put in store barrier for any and all oops we are sticking 4414 // into this object. (We could avoid this if we could prove 4415 // that the object type contains no oop fields at all.) 4416 Node* no_particular_value = NULL; 4417 Node* no_particular_field = NULL; 4418 int raw_adr_idx = Compile::AliasIdxRaw; 4419 post_barrier(control(), 4420 memory(raw_adr_type), 4421 alloc_obj, 4422 no_particular_field, 4423 raw_adr_idx, 4424 no_particular_value, 4425 T_OBJECT, 4426 false); 4427 } 4428 4429 // Do not let reads from the cloned object float above the arraycopy. 4430 if (alloc != NULL) { 4431 // Do not let stores that initialize this object be reordered with 4432 // a subsequent store that would make this object accessible by 4433 // other threads. 4434 // Record what AllocateNode this StoreStore protects so that 4435 // escape analysis can go from the MemBarStoreStoreNode to the 4436 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4437 // based on the escape status of the AllocateNode. 4438 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4439 } else { 4440 insert_mem_bar(Op_MemBarCPUOrder); 4441 } 4442 } 4443 4444 //------------------------inline_native_clone---------------------------- 4445 // Here are the simple edge cases: 4446 // null receiver => normal trap 4447 // virtual and clone was overridden => slow path to out-of-line clone 4448 // not cloneable or finalizer => slow path to out-of-line Object.clone 4449 // 4450 // The general case has two steps, allocation and copying. 4451 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4452 // 4453 // Copying also has two cases, oop arrays and everything else. 4454 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4455 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4456 // 4457 // These steps fold up nicely if and when the cloned object's klass 4458 // can be sharply typed as an object array, a type array, or an instance. 4459 // 4460 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4461 int nargs = 1; 4462 PhiNode* result_val; 4463 4464 //set the original stack and the reexecute bit for the interpreter to reexecute 4465 //the bytecode that invokes Object.clone if deoptimization happens 4466 { PreserveReexecuteState preexecs(this); 4467 jvms()->set_should_reexecute(true); 4468 4469 //null_check_receiver will adjust _sp (push and pop) 4470 Node* obj = null_check_receiver(callee()); 4471 if (stopped()) return true; 4472 4473 _sp += nargs; 4474 4475 Node* obj_klass = load_object_klass(obj); 4476 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4477 const TypeOopPtr* toop = ((tklass != NULL) 4478 ? tklass->as_instance_type() 4479 : TypeInstPtr::NOTNULL); 4480 4481 // Conservatively insert a memory barrier on all memory slices. 4482 // Do not let writes into the original float below the clone. 4483 insert_mem_bar(Op_MemBarCPUOrder); 4484 4485 // paths into result_reg: 4486 enum { 4487 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4488 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4489 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4490 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4491 PATH_LIMIT 4492 }; 4493 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4494 result_val = new(C) PhiNode(result_reg, 4495 TypeInstPtr::NOTNULL); 4496 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO); 4497 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 4498 TypePtr::BOTTOM); 4499 record_for_igvn(result_reg); 4500 4501 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4502 int raw_adr_idx = Compile::AliasIdxRaw; 4503 4504 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4505 if (array_ctl != NULL) { 4506 // It's an array. 4507 PreserveJVMState pjvms(this); 4508 set_control(array_ctl); 4509 Node* obj_length = load_array_length(obj); 4510 Node* obj_size = NULL; 4511 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); 4512 4513 if (!use_ReduceInitialCardMarks()) { 4514 // If it is an oop array, it requires very special treatment, 4515 // because card marking is required on each card of the array. 4516 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4517 if (is_obja != NULL) { 4518 PreserveJVMState pjvms2(this); 4519 set_control(is_obja); 4520 // Generate a direct call to the right arraycopy function(s). 4521 bool disjoint_bases = true; 4522 bool length_never_negative = true; 4523 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4524 obj, intcon(0), alloc_obj, intcon(0), 4525 obj_length, 4526 disjoint_bases, length_never_negative); 4527 result_reg->init_req(_objArray_path, control()); 4528 result_val->init_req(_objArray_path, alloc_obj); 4529 result_i_o ->set_req(_objArray_path, i_o()); 4530 result_mem ->set_req(_objArray_path, reset_memory()); 4531 } 4532 } 4533 // Otherwise, there are no card marks to worry about. 4534 // (We can dispense with card marks if we know the allocation 4535 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4536 // causes the non-eden paths to take compensating steps to 4537 // simulate a fresh allocation, so that no further 4538 // card marks are required in compiled code to initialize 4539 // the object.) 4540 4541 if (!stopped()) { 4542 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4543 4544 // Present the results of the copy. 4545 result_reg->init_req(_array_path, control()); 4546 result_val->init_req(_array_path, alloc_obj); 4547 result_i_o ->set_req(_array_path, i_o()); 4548 result_mem ->set_req(_array_path, reset_memory()); 4549 } 4550 } 4551 4552 // We only go to the instance fast case code if we pass a number of guards. 4553 // The paths which do not pass are accumulated in the slow_region. 4554 RegionNode* slow_region = new (C) RegionNode(1); 4555 record_for_igvn(slow_region); 4556 if (!stopped()) { 4557 // It's an instance (we did array above). Make the slow-path tests. 4558 // If this is a virtual call, we generate a funny guard. We grab 4559 // the vtable entry corresponding to clone() from the target object. 4560 // If the target method which we are calling happens to be the 4561 // Object clone() method, we pass the guard. We do not need this 4562 // guard for non-virtual calls; the caller is known to be the native 4563 // Object clone(). 4564 if (is_virtual) { 4565 generate_virtual_guard(obj_klass, slow_region); 4566 } 4567 4568 // The object must be cloneable and must not have a finalizer. 4569 // Both of these conditions may be checked in a single test. 4570 // We could optimize the cloneable test further, but we don't care. 4571 generate_access_flags_guard(obj_klass, 4572 // Test both conditions: 4573 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4574 // Must be cloneable but not finalizer: 4575 JVM_ACC_IS_CLONEABLE, 4576 slow_region); 4577 } 4578 4579 if (!stopped()) { 4580 // It's an instance, and it passed the slow-path tests. 4581 PreserveJVMState pjvms(this); 4582 Node* obj_size = NULL; 4583 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size); 4584 4585 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4586 4587 // Present the results of the slow call. 4588 result_reg->init_req(_instance_path, control()); 4589 result_val->init_req(_instance_path, alloc_obj); 4590 result_i_o ->set_req(_instance_path, i_o()); 4591 result_mem ->set_req(_instance_path, reset_memory()); 4592 } 4593 4594 // Generate code for the slow case. We make a call to clone(). 4595 set_control(_gvn.transform(slow_region)); 4596 if (!stopped()) { 4597 PreserveJVMState pjvms(this); 4598 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4599 Node* slow_result = set_results_for_java_call(slow_call); 4600 // this->control() comes from set_results_for_java_call 4601 result_reg->init_req(_slow_path, control()); 4602 result_val->init_req(_slow_path, slow_result); 4603 result_i_o ->set_req(_slow_path, i_o()); 4604 result_mem ->set_req(_slow_path, reset_memory()); 4605 } 4606 4607 // Return the combined state. 4608 set_control( _gvn.transform(result_reg) ); 4609 set_i_o( _gvn.transform(result_i_o) ); 4610 set_all_memory( _gvn.transform(result_mem) ); 4611 } //original reexecute and sp are set back here 4612 4613 push(_gvn.transform(result_val)); 4614 4615 return true; 4616 } 4617 4618 //------------------------------basictype2arraycopy---------------------------- 4619 address LibraryCallKit::basictype2arraycopy(BasicType t, 4620 Node* src_offset, 4621 Node* dest_offset, 4622 bool disjoint_bases, 4623 const char* &name, 4624 bool dest_uninitialized) { 4625 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4626 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4627 4628 bool aligned = false; 4629 bool disjoint = disjoint_bases; 4630 4631 // if the offsets are the same, we can treat the memory regions as 4632 // disjoint, because either the memory regions are in different arrays, 4633 // or they are identical (which we can treat as disjoint.) We can also 4634 // treat a copy with a destination index less that the source index 4635 // as disjoint since a low->high copy will work correctly in this case. 4636 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4637 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4638 // both indices are constants 4639 int s_offs = src_offset_inttype->get_con(); 4640 int d_offs = dest_offset_inttype->get_con(); 4641 int element_size = type2aelembytes(t); 4642 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4643 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4644 if (s_offs >= d_offs) disjoint = true; 4645 } else if (src_offset == dest_offset && src_offset != NULL) { 4646 // This can occur if the offsets are identical non-constants. 4647 disjoint = true; 4648 } 4649 4650 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4651 } 4652 4653 4654 //------------------------------inline_arraycopy----------------------- 4655 bool LibraryCallKit::inline_arraycopy() { 4656 // Restore the stack and pop off the arguments. 4657 int nargs = 5; // 2 oops, 3 ints, no size_t or long 4658 assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); 4659 4660 Node *src = argument(0); 4661 Node *src_offset = argument(1); 4662 Node *dest = argument(2); 4663 Node *dest_offset = argument(3); 4664 Node *length = argument(4); 4665 4666 // Compile time checks. If any of these checks cannot be verified at compile time, 4667 // we do not make a fast path for this call. Instead, we let the call remain as it 4668 // is. The checks we choose to mandate at compile time are: 4669 // 4670 // (1) src and dest are arrays. 4671 const Type* src_type = src->Value(&_gvn); 4672 const Type* dest_type = dest->Value(&_gvn); 4673 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4674 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4675 if (top_src == NULL || top_src->klass() == NULL || 4676 top_dest == NULL || top_dest->klass() == NULL) { 4677 // Conservatively insert a memory barrier on all memory slices. 4678 // Do not let writes into the source float below the arraycopy. 4679 insert_mem_bar(Op_MemBarCPUOrder); 4680 4681 // Call StubRoutines::generic_arraycopy stub. 4682 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4683 src, src_offset, dest, dest_offset, length); 4684 4685 // Do not let reads from the destination float above the arraycopy. 4686 // Since we cannot type the arrays, we don't know which slices 4687 // might be affected. We could restrict this barrier only to those 4688 // memory slices which pertain to array elements--but don't bother. 4689 if (!InsertMemBarAfterArraycopy) 4690 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4691 insert_mem_bar(Op_MemBarCPUOrder); 4692 return true; 4693 } 4694 4695 // (2) src and dest arrays must have elements of the same BasicType 4696 // Figure out the size and type of the elements we will be copying. 4697 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4698 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4699 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4700 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4701 4702 if (src_elem != dest_elem || dest_elem == T_VOID) { 4703 // The component types are not the same or are not recognized. Punt. 4704 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4705 generate_slow_arraycopy(TypePtr::BOTTOM, 4706 src, src_offset, dest, dest_offset, length, 4707 /*dest_uninitialized*/false); 4708 return true; 4709 } 4710 4711 //--------------------------------------------------------------------------- 4712 // We will make a fast path for this call to arraycopy. 4713 4714 // We have the following tests left to perform: 4715 // 4716 // (3) src and dest must not be null. 4717 // (4) src_offset must not be negative. 4718 // (5) dest_offset must not be negative. 4719 // (6) length must not be negative. 4720 // (7) src_offset + length must not exceed length of src. 4721 // (8) dest_offset + length must not exceed length of dest. 4722 // (9) each element of an oop array must be assignable 4723 4724 RegionNode* slow_region = new (C) RegionNode(1); 4725 record_for_igvn(slow_region); 4726 4727 // (3) operands must not be null 4728 // We currently perform our null checks with the do_null_check routine. 4729 // This means that the null exceptions will be reported in the caller 4730 // rather than (correctly) reported inside of the native arraycopy call. 4731 // This should be corrected, given time. We do our null check with the 4732 // stack pointer restored. 4733 _sp += nargs; 4734 src = do_null_check(src, T_ARRAY); 4735 dest = do_null_check(dest, T_ARRAY); 4736 _sp -= nargs; 4737 4738 // (4) src_offset must not be negative. 4739 generate_negative_guard(src_offset, slow_region); 4740 4741 // (5) dest_offset must not be negative. 4742 generate_negative_guard(dest_offset, slow_region); 4743 4744 // (6) length must not be negative (moved to generate_arraycopy()). 4745 // generate_negative_guard(length, slow_region); 4746 4747 // (7) src_offset + length must not exceed length of src. 4748 generate_limit_guard(src_offset, length, 4749 load_array_length(src), 4750 slow_region); 4751 4752 // (8) dest_offset + length must not exceed length of dest. 4753 generate_limit_guard(dest_offset, length, 4754 load_array_length(dest), 4755 slow_region); 4756 4757 // (9) each element of an oop array must be assignable 4758 // The generate_arraycopy subroutine checks this. 4759 4760 // This is where the memory effects are placed: 4761 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4762 generate_arraycopy(adr_type, dest_elem, 4763 src, src_offset, dest, dest_offset, length, 4764 false, false, slow_region); 4765 4766 return true; 4767 } 4768 4769 //-----------------------------generate_arraycopy---------------------- 4770 // Generate an optimized call to arraycopy. 4771 // Caller must guard against non-arrays. 4772 // Caller must determine a common array basic-type for both arrays. 4773 // Caller must validate offsets against array bounds. 4774 // The slow_region has already collected guard failure paths 4775 // (such as out of bounds length or non-conformable array types). 4776 // The generated code has this shape, in general: 4777 // 4778 // if (length == 0) return // via zero_path 4779 // slowval = -1 4780 // if (types unknown) { 4781 // slowval = call generic copy loop 4782 // if (slowval == 0) return // via checked_path 4783 // } else if (indexes in bounds) { 4784 // if ((is object array) && !(array type check)) { 4785 // slowval = call checked copy loop 4786 // if (slowval == 0) return // via checked_path 4787 // } else { 4788 // call bulk copy loop 4789 // return // via fast_path 4790 // } 4791 // } 4792 // // adjust params for remaining work: 4793 // if (slowval != -1) { 4794 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4795 // } 4796 // slow_region: 4797 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4798 // return // via slow_call_path 4799 // 4800 // This routine is used from several intrinsics: System.arraycopy, 4801 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4802 // 4803 void 4804 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4805 BasicType basic_elem_type, 4806 Node* src, Node* src_offset, 4807 Node* dest, Node* dest_offset, 4808 Node* copy_length, 4809 bool disjoint_bases, 4810 bool length_never_negative, 4811 RegionNode* slow_region) { 4812 4813 if (slow_region == NULL) { 4814 slow_region = new(C) RegionNode(1); 4815 record_for_igvn(slow_region); 4816 } 4817 4818 Node* original_dest = dest; 4819 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4820 bool dest_uninitialized = false; 4821 4822 // See if this is the initialization of a newly-allocated array. 4823 // If so, we will take responsibility here for initializing it to zero. 4824 // (Note: Because tightly_coupled_allocation performs checks on the 4825 // out-edges of the dest, we need to avoid making derived pointers 4826 // from it until we have checked its uses.) 4827 if (ReduceBulkZeroing 4828 && !ZeroTLAB // pointless if already zeroed 4829 && basic_elem_type != T_CONFLICT // avoid corner case 4830 && !src->eqv_uncast(dest) 4831 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4832 != NULL) 4833 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4834 && alloc->maybe_set_complete(&_gvn)) { 4835 // "You break it, you buy it." 4836 InitializeNode* init = alloc->initialization(); 4837 assert(init->is_complete(), "we just did this"); 4838 init->set_complete_with_arraycopy(); 4839 assert(dest->is_CheckCastPP(), "sanity"); 4840 assert(dest->in(0)->in(0) == init, "dest pinned"); 4841 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4842 // From this point on, every exit path is responsible for 4843 // initializing any non-copied parts of the object to zero. 4844 // Also, if this flag is set we make sure that arraycopy interacts properly 4845 // with G1, eliding pre-barriers. See CR 6627983. 4846 dest_uninitialized = true; 4847 } else { 4848 // No zeroing elimination here. 4849 alloc = NULL; 4850 //original_dest = dest; 4851 //dest_uninitialized = false; 4852 } 4853 4854 // Results are placed here: 4855 enum { fast_path = 1, // normal void-returning assembly stub 4856 checked_path = 2, // special assembly stub with cleanup 4857 slow_call_path = 3, // something went wrong; call the VM 4858 zero_path = 4, // bypass when length of copy is zero 4859 bcopy_path = 5, // copy primitive array by 64-bit blocks 4860 PATH_LIMIT = 6 4861 }; 4862 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT); 4863 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO); 4864 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type); 4865 record_for_igvn(result_region); 4866 _gvn.set_type_bottom(result_i_o); 4867 _gvn.set_type_bottom(result_memory); 4868 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4869 4870 // The slow_control path: 4871 Node* slow_control; 4872 Node* slow_i_o = i_o(); 4873 Node* slow_mem = memory(adr_type); 4874 debug_only(slow_control = (Node*) badAddress); 4875 4876 // Checked control path: 4877 Node* checked_control = top(); 4878 Node* checked_mem = NULL; 4879 Node* checked_i_o = NULL; 4880 Node* checked_value = NULL; 4881 4882 if (basic_elem_type == T_CONFLICT) { 4883 assert(!dest_uninitialized, ""); 4884 Node* cv = generate_generic_arraycopy(adr_type, 4885 src, src_offset, dest, dest_offset, 4886 copy_length, dest_uninitialized); 4887 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4888 checked_control = control(); 4889 checked_i_o = i_o(); 4890 checked_mem = memory(adr_type); 4891 checked_value = cv; 4892 set_control(top()); // no fast path 4893 } 4894 4895 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4896 if (not_pos != NULL) { 4897 PreserveJVMState pjvms(this); 4898 set_control(not_pos); 4899 4900 // (6) length must not be negative. 4901 if (!length_never_negative) { 4902 generate_negative_guard(copy_length, slow_region); 4903 } 4904 4905 // copy_length is 0. 4906 if (!stopped() && dest_uninitialized) { 4907 Node* dest_length = alloc->in(AllocateNode::ALength); 4908 if (copy_length->eqv_uncast(dest_length) 4909 || _gvn.find_int_con(dest_length, 1) <= 0) { 4910 // There is no zeroing to do. No need for a secondary raw memory barrier. 4911 } else { 4912 // Clear the whole thing since there are no source elements to copy. 4913 generate_clear_array(adr_type, dest, basic_elem_type, 4914 intcon(0), NULL, 4915 alloc->in(AllocateNode::AllocSize)); 4916 // Use a secondary InitializeNode as raw memory barrier. 4917 // Currently it is needed only on this path since other 4918 // paths have stub or runtime calls as raw memory barriers. 4919 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4920 Compile::AliasIdxRaw, 4921 top())->as_Initialize(); 4922 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4923 } 4924 } 4925 4926 // Present the results of the fast call. 4927 result_region->init_req(zero_path, control()); 4928 result_i_o ->init_req(zero_path, i_o()); 4929 result_memory->init_req(zero_path, memory(adr_type)); 4930 } 4931 4932 if (!stopped() && dest_uninitialized) { 4933 // We have to initialize the *uncopied* part of the array to zero. 4934 // The copy destination is the slice dest[off..off+len]. The other slices 4935 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4936 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4937 Node* dest_length = alloc->in(AllocateNode::ALength); 4938 Node* dest_tail = _gvn.transform( new(C) AddINode(dest_offset, 4939 copy_length) ); 4940 4941 // If there is a head section that needs zeroing, do it now. 4942 if (find_int_con(dest_offset, -1) != 0) { 4943 generate_clear_array(adr_type, dest, basic_elem_type, 4944 intcon(0), dest_offset, 4945 NULL); 4946 } 4947 4948 // Next, perform a dynamic check on the tail length. 4949 // It is often zero, and we can win big if we prove this. 4950 // There are two wins: Avoid generating the ClearArray 4951 // with its attendant messy index arithmetic, and upgrade 4952 // the copy to a more hardware-friendly word size of 64 bits. 4953 Node* tail_ctl = NULL; 4954 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 4955 Node* cmp_lt = _gvn.transform( new(C) CmpINode(dest_tail, dest_length) ); 4956 Node* bol_lt = _gvn.transform( new(C) BoolNode(cmp_lt, BoolTest::lt) ); 4957 tail_ctl = generate_slow_guard(bol_lt, NULL); 4958 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4959 } 4960 4961 // At this point, let's assume there is no tail. 4962 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4963 // There is no tail. Try an upgrade to a 64-bit copy. 4964 bool didit = false; 4965 { PreserveJVMState pjvms(this); 4966 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4967 src, src_offset, dest, dest_offset, 4968 dest_size, dest_uninitialized); 4969 if (didit) { 4970 // Present the results of the block-copying fast call. 4971 result_region->init_req(bcopy_path, control()); 4972 result_i_o ->init_req(bcopy_path, i_o()); 4973 result_memory->init_req(bcopy_path, memory(adr_type)); 4974 } 4975 } 4976 if (didit) 4977 set_control(top()); // no regular fast path 4978 } 4979 4980 // Clear the tail, if any. 4981 if (tail_ctl != NULL) { 4982 Node* notail_ctl = stopped() ? NULL : control(); 4983 set_control(tail_ctl); 4984 if (notail_ctl == NULL) { 4985 generate_clear_array(adr_type, dest, basic_elem_type, 4986 dest_tail, NULL, 4987 dest_size); 4988 } else { 4989 // Make a local merge. 4990 Node* done_ctl = new(C) RegionNode(3); 4991 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type); 4992 done_ctl->init_req(1, notail_ctl); 4993 done_mem->init_req(1, memory(adr_type)); 4994 generate_clear_array(adr_type, dest, basic_elem_type, 4995 dest_tail, NULL, 4996 dest_size); 4997 done_ctl->init_req(2, control()); 4998 done_mem->init_req(2, memory(adr_type)); 4999 set_control( _gvn.transform(done_ctl) ); 5000 set_memory( _gvn.transform(done_mem), adr_type ); 5001 } 5002 } 5003 } 5004 5005 BasicType copy_type = basic_elem_type; 5006 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 5007 if (!stopped() && copy_type == T_OBJECT) { 5008 // If src and dest have compatible element types, we can copy bits. 5009 // Types S[] and D[] are compatible if D is a supertype of S. 5010 // 5011 // If they are not, we will use checked_oop_disjoint_arraycopy, 5012 // which performs a fast optimistic per-oop check, and backs off 5013 // further to JVM_ArrayCopy on the first per-oop check that fails. 5014 // (Actually, we don't move raw bits only; the GC requires card marks.) 5015 5016 // Get the Klass* for both src and dest 5017 Node* src_klass = load_object_klass(src); 5018 Node* dest_klass = load_object_klass(dest); 5019 5020 // Generate the subtype check. 5021 // This might fold up statically, or then again it might not. 5022 // 5023 // Non-static example: Copying List<String>.elements to a new String[]. 5024 // The backing store for a List<String> is always an Object[], 5025 // but its elements are always type String, if the generic types 5026 // are correct at the source level. 5027 // 5028 // Test S[] against D[], not S against D, because (probably) 5029 // the secondary supertype cache is less busy for S[] than S. 5030 // This usually only matters when D is an interface. 5031 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5032 // Plug failing path into checked_oop_disjoint_arraycopy 5033 if (not_subtype_ctrl != top()) { 5034 PreserveJVMState pjvms(this); 5035 set_control(not_subtype_ctrl); 5036 // (At this point we can assume disjoint_bases, since types differ.) 5037 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 5038 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 5039 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 5040 Node* dest_elem_klass = _gvn.transform(n1); 5041 Node* cv = generate_checkcast_arraycopy(adr_type, 5042 dest_elem_klass, 5043 src, src_offset, dest, dest_offset, 5044 ConvI2X(copy_length), dest_uninitialized); 5045 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 5046 checked_control = control(); 5047 checked_i_o = i_o(); 5048 checked_mem = memory(adr_type); 5049 checked_value = cv; 5050 } 5051 // At this point we know we do not need type checks on oop stores. 5052 5053 // Let's see if we need card marks: 5054 if (alloc != NULL && use_ReduceInitialCardMarks()) { 5055 // If we do not need card marks, copy using the jint or jlong stub. 5056 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 5057 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 5058 "sizes agree"); 5059 } 5060 } 5061 5062 if (!stopped()) { 5063 // Generate the fast path, if possible. 5064 PreserveJVMState pjvms(this); 5065 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 5066 src, src_offset, dest, dest_offset, 5067 ConvI2X(copy_length), dest_uninitialized); 5068 5069 // Present the results of the fast call. 5070 result_region->init_req(fast_path, control()); 5071 result_i_o ->init_req(fast_path, i_o()); 5072 result_memory->init_req(fast_path, memory(adr_type)); 5073 } 5074 5075 // Here are all the slow paths up to this point, in one bundle: 5076 slow_control = top(); 5077 if (slow_region != NULL) 5078 slow_control = _gvn.transform(slow_region); 5079 debug_only(slow_region = (RegionNode*)badAddress); 5080 5081 set_control(checked_control); 5082 if (!stopped()) { 5083 // Clean up after the checked call. 5084 // The returned value is either 0 or -1^K, 5085 // where K = number of partially transferred array elements. 5086 Node* cmp = _gvn.transform( new(C) CmpINode(checked_value, intcon(0)) ); 5087 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 5088 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 5089 5090 // If it is 0, we are done, so transfer to the end. 5091 Node* checks_done = _gvn.transform( new(C) IfTrueNode(iff) ); 5092 result_region->init_req(checked_path, checks_done); 5093 result_i_o ->init_req(checked_path, checked_i_o); 5094 result_memory->init_req(checked_path, checked_mem); 5095 5096 // If it is not zero, merge into the slow call. 5097 set_control( _gvn.transform( new(C) IfFalseNode(iff) )); 5098 RegionNode* slow_reg2 = new(C) RegionNode(3); 5099 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO); 5100 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type); 5101 record_for_igvn(slow_reg2); 5102 slow_reg2 ->init_req(1, slow_control); 5103 slow_i_o2 ->init_req(1, slow_i_o); 5104 slow_mem2 ->init_req(1, slow_mem); 5105 slow_reg2 ->init_req(2, control()); 5106 slow_i_o2 ->init_req(2, checked_i_o); 5107 slow_mem2 ->init_req(2, checked_mem); 5108 5109 slow_control = _gvn.transform(slow_reg2); 5110 slow_i_o = _gvn.transform(slow_i_o2); 5111 slow_mem = _gvn.transform(slow_mem2); 5112 5113 if (alloc != NULL) { 5114 // We'll restart from the very beginning, after zeroing the whole thing. 5115 // This can cause double writes, but that's OK since dest is brand new. 5116 // So we ignore the low 31 bits of the value returned from the stub. 5117 } else { 5118 // We must continue the copy exactly where it failed, or else 5119 // another thread might see the wrong number of writes to dest. 5120 Node* checked_offset = _gvn.transform( new(C) XorINode(checked_value, intcon(-1)) ); 5121 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT); 5122 slow_offset->init_req(1, intcon(0)); 5123 slow_offset->init_req(2, checked_offset); 5124 slow_offset = _gvn.transform(slow_offset); 5125 5126 // Adjust the arguments by the conditionally incoming offset. 5127 Node* src_off_plus = _gvn.transform( new(C) AddINode(src_offset, slow_offset) ); 5128 Node* dest_off_plus = _gvn.transform( new(C) AddINode(dest_offset, slow_offset) ); 5129 Node* length_minus = _gvn.transform( new(C) SubINode(copy_length, slow_offset) ); 5130 5131 // Tweak the node variables to adjust the code produced below: 5132 src_offset = src_off_plus; 5133 dest_offset = dest_off_plus; 5134 copy_length = length_minus; 5135 } 5136 } 5137 5138 set_control(slow_control); 5139 if (!stopped()) { 5140 // Generate the slow path, if needed. 5141 PreserveJVMState pjvms(this); // replace_in_map may trash the map 5142 5143 set_memory(slow_mem, adr_type); 5144 set_i_o(slow_i_o); 5145 5146 if (dest_uninitialized) { 5147 generate_clear_array(adr_type, dest, basic_elem_type, 5148 intcon(0), NULL, 5149 alloc->in(AllocateNode::AllocSize)); 5150 } 5151 5152 generate_slow_arraycopy(adr_type, 5153 src, src_offset, dest, dest_offset, 5154 copy_length, /*dest_uninitialized*/false); 5155 5156 result_region->init_req(slow_call_path, control()); 5157 result_i_o ->init_req(slow_call_path, i_o()); 5158 result_memory->init_req(slow_call_path, memory(adr_type)); 5159 } 5160 5161 // Remove unused edges. 5162 for (uint i = 1; i < result_region->req(); i++) { 5163 if (result_region->in(i) == NULL) 5164 result_region->init_req(i, top()); 5165 } 5166 5167 // Finished; return the combined state. 5168 set_control( _gvn.transform(result_region) ); 5169 set_i_o( _gvn.transform(result_i_o) ); 5170 set_memory( _gvn.transform(result_memory), adr_type ); 5171 5172 // The memory edges above are precise in order to model effects around 5173 // array copies accurately to allow value numbering of field loads around 5174 // arraycopy. Such field loads, both before and after, are common in Java 5175 // collections and similar classes involving header/array data structures. 5176 // 5177 // But with low number of register or when some registers are used or killed 5178 // by arraycopy calls it causes registers spilling on stack. See 6544710. 5179 // The next memory barrier is added to avoid it. If the arraycopy can be 5180 // optimized away (which it can, sometimes) then we can manually remove 5181 // the membar also. 5182 // 5183 // Do not let reads from the cloned object float above the arraycopy. 5184 if (alloc != NULL) { 5185 // Do not let stores that initialize this object be reordered with 5186 // a subsequent store that would make this object accessible by 5187 // other threads. 5188 // Record what AllocateNode this StoreStore protects so that 5189 // escape analysis can go from the MemBarStoreStoreNode to the 5190 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5191 // based on the escape status of the AllocateNode. 5192 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5193 } else if (InsertMemBarAfterArraycopy) 5194 insert_mem_bar(Op_MemBarCPUOrder); 5195 } 5196 5197 5198 // Helper function which determines if an arraycopy immediately follows 5199 // an allocation, with no intervening tests or other escapes for the object. 5200 AllocateArrayNode* 5201 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5202 RegionNode* slow_region) { 5203 if (stopped()) return NULL; // no fast path 5204 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5205 5206 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5207 if (alloc == NULL) return NULL; 5208 5209 Node* rawmem = memory(Compile::AliasIdxRaw); 5210 // Is the allocation's memory state untouched? 5211 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5212 // Bail out if there have been raw-memory effects since the allocation. 5213 // (Example: There might have been a call or safepoint.) 5214 return NULL; 5215 } 5216 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5217 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5218 return NULL; 5219 } 5220 5221 // There must be no unexpected observers of this allocation. 5222 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5223 Node* obs = ptr->fast_out(i); 5224 if (obs != this->map()) { 5225 return NULL; 5226 } 5227 } 5228 5229 // This arraycopy must unconditionally follow the allocation of the ptr. 5230 Node* alloc_ctl = ptr->in(0); 5231 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5232 5233 Node* ctl = control(); 5234 while (ctl != alloc_ctl) { 5235 // There may be guards which feed into the slow_region. 5236 // Any other control flow means that we might not get a chance 5237 // to finish initializing the allocated object. 5238 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5239 IfNode* iff = ctl->in(0)->as_If(); 5240 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5241 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5242 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5243 ctl = iff->in(0); // This test feeds the known slow_region. 5244 continue; 5245 } 5246 // One more try: Various low-level checks bottom out in 5247 // uncommon traps. If the debug-info of the trap omits 5248 // any reference to the allocation, as we've already 5249 // observed, then there can be no objection to the trap. 5250 bool found_trap = false; 5251 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5252 Node* obs = not_ctl->fast_out(j); 5253 if (obs->in(0) == not_ctl && obs->is_Call() && 5254 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5255 found_trap = true; break; 5256 } 5257 } 5258 if (found_trap) { 5259 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5260 continue; 5261 } 5262 } 5263 return NULL; 5264 } 5265 5266 // If we get this far, we have an allocation which immediately 5267 // precedes the arraycopy, and we can take over zeroing the new object. 5268 // The arraycopy will finish the initialization, and provide 5269 // a new control state to which we will anchor the destination pointer. 5270 5271 return alloc; 5272 } 5273 5274 // Helper for initialization of arrays, creating a ClearArray. 5275 // It writes zero bits in [start..end), within the body of an array object. 5276 // The memory effects are all chained onto the 'adr_type' alias category. 5277 // 5278 // Since the object is otherwise uninitialized, we are free 5279 // to put a little "slop" around the edges of the cleared area, 5280 // as long as it does not go back into the array's header, 5281 // or beyond the array end within the heap. 5282 // 5283 // The lower edge can be rounded down to the nearest jint and the 5284 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5285 // 5286 // Arguments: 5287 // adr_type memory slice where writes are generated 5288 // dest oop of the destination array 5289 // basic_elem_type element type of the destination 5290 // slice_idx array index of first element to store 5291 // slice_len number of elements to store (or NULL) 5292 // dest_size total size in bytes of the array object 5293 // 5294 // Exactly one of slice_len or dest_size must be non-NULL. 5295 // If dest_size is non-NULL, zeroing extends to the end of the object. 5296 // If slice_len is non-NULL, the slice_idx value must be a constant. 5297 void 5298 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5299 Node* dest, 5300 BasicType basic_elem_type, 5301 Node* slice_idx, 5302 Node* slice_len, 5303 Node* dest_size) { 5304 // one or the other but not both of slice_len and dest_size: 5305 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5306 if (slice_len == NULL) slice_len = top(); 5307 if (dest_size == NULL) dest_size = top(); 5308 5309 // operate on this memory slice: 5310 Node* mem = memory(adr_type); // memory slice to operate on 5311 5312 // scaling and rounding of indexes: 5313 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5314 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5315 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5316 int bump_bit = (-1 << scale) & BytesPerInt; 5317 5318 // determine constant starts and ends 5319 const intptr_t BIG_NEG = -128; 5320 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5321 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5322 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5323 if (slice_len_con == 0) { 5324 return; // nothing to do here 5325 } 5326 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5327 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5328 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5329 assert(end_con < 0, "not two cons"); 5330 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5331 BytesPerLong); 5332 } 5333 5334 if (start_con >= 0 && end_con >= 0) { 5335 // Constant start and end. Simple. 5336 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5337 start_con, end_con, &_gvn); 5338 } else if (start_con >= 0 && dest_size != top()) { 5339 // Constant start, pre-rounded end after the tail of the array. 5340 Node* end = dest_size; 5341 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5342 start_con, end, &_gvn); 5343 } else if (start_con >= 0 && slice_len != top()) { 5344 // Constant start, non-constant end. End needs rounding up. 5345 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5346 intptr_t end_base = abase + (slice_idx_con << scale); 5347 int end_round = (-1 << scale) & (BytesPerLong - 1); 5348 Node* end = ConvI2X(slice_len); 5349 if (scale != 0) 5350 end = _gvn.transform( new(C) LShiftXNode(end, intcon(scale) )); 5351 end_base += end_round; 5352 end = _gvn.transform( new(C) AddXNode(end, MakeConX(end_base)) ); 5353 end = _gvn.transform( new(C) AndXNode(end, MakeConX(~end_round)) ); 5354 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5355 start_con, end, &_gvn); 5356 } else if (start_con < 0 && dest_size != top()) { 5357 // Non-constant start, pre-rounded end after the tail of the array. 5358 // This is almost certainly a "round-to-end" operation. 5359 Node* start = slice_idx; 5360 start = ConvI2X(start); 5361 if (scale != 0) 5362 start = _gvn.transform( new(C) LShiftXNode( start, intcon(scale) )); 5363 start = _gvn.transform( new(C) AddXNode(start, MakeConX(abase)) ); 5364 if ((bump_bit | clear_low) != 0) { 5365 int to_clear = (bump_bit | clear_low); 5366 // Align up mod 8, then store a jint zero unconditionally 5367 // just before the mod-8 boundary. 5368 if (((abase + bump_bit) & ~to_clear) - bump_bit 5369 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5370 bump_bit = 0; 5371 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5372 } else { 5373 // Bump 'start' up to (or past) the next jint boundary: 5374 start = _gvn.transform( new(C) AddXNode(start, MakeConX(bump_bit)) ); 5375 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5376 } 5377 // Round bumped 'start' down to jlong boundary in body of array. 5378 start = _gvn.transform( new(C) AndXNode(start, MakeConX(~to_clear)) ); 5379 if (bump_bit != 0) { 5380 // Store a zero to the immediately preceding jint: 5381 Node* x1 = _gvn.transform( new(C) AddXNode(start, MakeConX(-bump_bit)) ); 5382 Node* p1 = basic_plus_adr(dest, x1); 5383 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 5384 mem = _gvn.transform(mem); 5385 } 5386 } 5387 Node* end = dest_size; // pre-rounded 5388 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5389 start, end, &_gvn); 5390 } else { 5391 // Non-constant start, unrounded non-constant end. 5392 // (Nobody zeroes a random midsection of an array using this routine.) 5393 ShouldNotReachHere(); // fix caller 5394 } 5395 5396 // Done. 5397 set_memory(mem, adr_type); 5398 } 5399 5400 5401 bool 5402 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5403 BasicType basic_elem_type, 5404 AllocateNode* alloc, 5405 Node* src, Node* src_offset, 5406 Node* dest, Node* dest_offset, 5407 Node* dest_size, bool dest_uninitialized) { 5408 // See if there is an advantage from block transfer. 5409 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5410 if (scale >= LogBytesPerLong) 5411 return false; // it is already a block transfer 5412 5413 // Look at the alignment of the starting offsets. 5414 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5415 5416 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5417 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5418 if (src_off_con < 0 || dest_off_con < 0) 5419 // At present, we can only understand constants. 5420 return false; 5421 5422 intptr_t src_off = abase + (src_off_con << scale); 5423 intptr_t dest_off = abase + (dest_off_con << scale); 5424 5425 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5426 // Non-aligned; too bad. 5427 // One more chance: Pick off an initial 32-bit word. 5428 // This is a common case, since abase can be odd mod 8. 5429 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5430 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5431 Node* sptr = basic_plus_adr(src, src_off); 5432 Node* dptr = basic_plus_adr(dest, dest_off); 5433 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 5434 store_to_memory(control(), dptr, sval, T_INT, adr_type); 5435 src_off += BytesPerInt; 5436 dest_off += BytesPerInt; 5437 } else { 5438 return false; 5439 } 5440 } 5441 assert(src_off % BytesPerLong == 0, ""); 5442 assert(dest_off % BytesPerLong == 0, ""); 5443 5444 // Do this copy by giant steps. 5445 Node* sptr = basic_plus_adr(src, src_off); 5446 Node* dptr = basic_plus_adr(dest, dest_off); 5447 Node* countx = dest_size; 5448 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(dest_off)) ); 5449 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong)) ); 5450 5451 bool disjoint_bases = true; // since alloc != NULL 5452 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5453 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5454 5455 return true; 5456 } 5457 5458 5459 // Helper function; generates code for the slow case. 5460 // We make a call to a runtime method which emulates the native method, 5461 // but without the native wrapper overhead. 5462 void 5463 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5464 Node* src, Node* src_offset, 5465 Node* dest, Node* dest_offset, 5466 Node* copy_length, bool dest_uninitialized) { 5467 assert(!dest_uninitialized, "Invariant"); 5468 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5469 OptoRuntime::slow_arraycopy_Type(), 5470 OptoRuntime::slow_arraycopy_Java(), 5471 "slow_arraycopy", adr_type, 5472 src, src_offset, dest, dest_offset, 5473 copy_length); 5474 5475 // Handle exceptions thrown by this fellow: 5476 make_slow_call_ex(call, env()->Throwable_klass(), false); 5477 } 5478 5479 // Helper function; generates code for cases requiring runtime checks. 5480 Node* 5481 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5482 Node* dest_elem_klass, 5483 Node* src, Node* src_offset, 5484 Node* dest, Node* dest_offset, 5485 Node* copy_length, bool dest_uninitialized) { 5486 if (stopped()) return NULL; 5487 5488 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5489 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5490 return NULL; 5491 } 5492 5493 // Pick out the parameters required to perform a store-check 5494 // for the target array. This is an optimistic check. It will 5495 // look in each non-null element's class, at the desired klass's 5496 // super_check_offset, for the desired klass. 5497 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5498 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5499 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr()); 5500 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5501 Node* check_value = dest_elem_klass; 5502 5503 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5504 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5505 5506 // (We know the arrays are never conjoint, because their types differ.) 5507 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5508 OptoRuntime::checkcast_arraycopy_Type(), 5509 copyfunc_addr, "checkcast_arraycopy", adr_type, 5510 // five arguments, of which two are 5511 // intptr_t (jlong in LP64) 5512 src_start, dest_start, 5513 copy_length XTOP, 5514 check_offset XTOP, 5515 check_value); 5516 5517 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5518 } 5519 5520 5521 // Helper function; generates code for cases requiring runtime checks. 5522 Node* 5523 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5524 Node* src, Node* src_offset, 5525 Node* dest, Node* dest_offset, 5526 Node* copy_length, bool dest_uninitialized) { 5527 assert(!dest_uninitialized, "Invariant"); 5528 if (stopped()) return NULL; 5529 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5530 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5531 return NULL; 5532 } 5533 5534 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5535 OptoRuntime::generic_arraycopy_Type(), 5536 copyfunc_addr, "generic_arraycopy", adr_type, 5537 src, src_offset, dest, dest_offset, copy_length); 5538 5539 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5540 } 5541 5542 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5543 void 5544 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5545 BasicType basic_elem_type, 5546 bool disjoint_bases, 5547 Node* src, Node* src_offset, 5548 Node* dest, Node* dest_offset, 5549 Node* copy_length, bool dest_uninitialized) { 5550 if (stopped()) return; // nothing to do 5551 5552 Node* src_start = src; 5553 Node* dest_start = dest; 5554 if (src_offset != NULL || dest_offset != NULL) { 5555 assert(src_offset != NULL && dest_offset != NULL, ""); 5556 src_start = array_element_address(src, src_offset, basic_elem_type); 5557 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5558 } 5559 5560 // Figure out which arraycopy runtime method to call. 5561 const char* copyfunc_name = "arraycopy"; 5562 address copyfunc_addr = 5563 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5564 disjoint_bases, copyfunc_name, dest_uninitialized); 5565 5566 // Call it. Note that the count_ix value is not scaled to a byte-size. 5567 make_runtime_call(RC_LEAF|RC_NO_FP, 5568 OptoRuntime::fast_arraycopy_Type(), 5569 copyfunc_addr, copyfunc_name, adr_type, 5570 src_start, dest_start, copy_length XTOP); 5571 } 5572 5573 //----------------------------inline_reference_get---------------------------- 5574 5575 bool LibraryCallKit::inline_reference_get() { 5576 const int nargs = 1; // self 5577 5578 guarantee(java_lang_ref_Reference::referent_offset > 0, 5579 "should have already been set"); 5580 5581 int referent_offset = java_lang_ref_Reference::referent_offset; 5582 5583 // Restore the stack and pop off the argument 5584 _sp += nargs; 5585 Node *reference_obj = pop(); 5586 5587 // Null check on self without removing any arguments. 5588 _sp += nargs; 5589 reference_obj = do_null_check(reference_obj, T_OBJECT); 5590 _sp -= nargs;; 5591 5592 if (stopped()) return true; 5593 5594 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5595 5596 ciInstanceKlass* klass = env()->Object_klass(); 5597 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5598 5599 Node* no_ctrl = NULL; 5600 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT); 5601 5602 // Use the pre-barrier to record the value in the referent field 5603 pre_barrier(false /* do_load */, 5604 control(), 5605 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5606 result /* pre_val */, 5607 T_OBJECT); 5608 5609 // Add memory barrier to prevent commoning reads from this field 5610 // across safepoint since GC can change its value. 5611 insert_mem_bar(Op_MemBarCPUOrder); 5612 5613 push(result); 5614 return true; 5615 }