1 /* 2 * Copyright (c) 1999, 2011, 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 public: 69 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic) 70 : GraphKit(caller), 71 _intrinsic(intrinsic) 72 { 73 } 74 75 ciMethod* caller() const { return jvms()->method(); } 76 int bci() const { return jvms()->bci(); } 77 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 79 ciMethod* callee() const { return _intrinsic->method(); } 80 ciSignature* signature() const { return callee()->signature(); } 81 int arg_size() const { return callee()->arg_size(); } 82 83 bool try_to_inline(); 84 85 // Helper functions to inline natives 86 void push_result(RegionNode* region, PhiNode* value); 87 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 88 Node* generate_slow_guard(Node* test, RegionNode* region); 89 Node* generate_fair_guard(Node* test, RegionNode* region); 90 Node* generate_negative_guard(Node* index, RegionNode* region, 91 // resulting CastII of index: 92 Node* *pos_index = NULL); 93 Node* generate_nonpositive_guard(Node* index, bool never_negative, 94 // resulting CastII of index: 95 Node* *pos_index = NULL); 96 Node* generate_limit_guard(Node* offset, Node* subseq_length, 97 Node* array_length, 98 RegionNode* region); 99 Node* generate_current_thread(Node* &tls_output); 100 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 101 bool disjoint_bases, const char* &name, bool dest_uninitialized); 102 Node* load_mirror_from_klass(Node* klass); 103 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 104 int nargs, 105 RegionNode* region, int null_path, 106 int offset); 107 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs, 108 RegionNode* region, int null_path) { 109 int offset = java_lang_Class::klass_offset_in_bytes(); 110 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 111 region, null_path, 112 offset); 113 } 114 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 115 int nargs, 116 RegionNode* region, int null_path) { 117 int offset = java_lang_Class::array_klass_offset_in_bytes(); 118 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 119 region, null_path, 120 offset); 121 } 122 Node* generate_access_flags_guard(Node* kls, 123 int modifier_mask, int modifier_bits, 124 RegionNode* region); 125 Node* generate_interface_guard(Node* kls, RegionNode* region); 126 Node* generate_array_guard(Node* kls, RegionNode* region) { 127 return generate_array_guard_common(kls, region, false, false); 128 } 129 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 130 return generate_array_guard_common(kls, region, false, true); 131 } 132 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 133 return generate_array_guard_common(kls, region, true, false); 134 } 135 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 136 return generate_array_guard_common(kls, region, true, true); 137 } 138 Node* generate_array_guard_common(Node* kls, RegionNode* region, 139 bool obj_array, bool not_array); 140 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 141 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 142 bool is_virtual = false, bool is_static = false); 143 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 144 return generate_method_call(method_id, false, true); 145 } 146 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 147 return generate_method_call(method_id, true, false); 148 } 149 150 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2); 151 bool inline_string_compareTo(); 152 bool inline_string_indexOf(); 153 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 154 bool inline_string_equals(); 155 Node* pop_math_arg(); 156 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 157 bool inline_math_native(vmIntrinsics::ID id); 158 bool inline_trig(vmIntrinsics::ID id); 159 bool inline_trans(vmIntrinsics::ID id); 160 bool inline_abs(vmIntrinsics::ID id); 161 bool inline_sqrt(vmIntrinsics::ID id); 162 bool inline_pow(vmIntrinsics::ID id); 163 bool inline_exp(vmIntrinsics::ID id); 164 bool inline_min_max(vmIntrinsics::ID id); 165 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 166 // This returns Type::AnyPtr, RawPtr, or OopPtr. 167 int classify_unsafe_addr(Node* &base, Node* &offset); 168 Node* make_unsafe_address(Node* base, Node* offset); 169 // Helper for inline_unsafe_access. 170 // Generates the guards that check whether the result of 171 // Unsafe.getObject should be recorded in an SATB log buffer. 172 void insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val); 173 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 174 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 175 bool inline_unsafe_allocate(); 176 bool inline_unsafe_copyMemory(); 177 bool inline_native_currentThread(); 178 bool inline_native_time_funcs(bool isNano); 179 bool inline_native_isInterrupted(); 180 bool inline_native_Class_query(vmIntrinsics::ID id); 181 bool inline_native_subtype_check(); 182 183 bool inline_native_newArray(); 184 bool inline_native_getLength(); 185 bool inline_array_copyOf(bool is_copyOfRange); 186 bool inline_array_equals(); 187 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 188 bool inline_native_clone(bool is_virtual); 189 bool inline_native_Reflection_getCallerClass(); 190 bool inline_native_AtomicLong_get(); 191 bool inline_native_AtomicLong_attemptUpdate(); 192 bool is_method_invoke_or_aux_frame(JVMState* jvms); 193 // Helper function for inlining native object hash method 194 bool inline_native_hashcode(bool is_virtual, bool is_static); 195 bool inline_native_getClass(); 196 197 // Helper functions for inlining arraycopy 198 bool inline_arraycopy(); 199 void generate_arraycopy(const TypePtr* adr_type, 200 BasicType basic_elem_type, 201 Node* src, Node* src_offset, 202 Node* dest, Node* dest_offset, 203 Node* copy_length, 204 bool disjoint_bases = false, 205 bool length_never_negative = false, 206 RegionNode* slow_region = NULL); 207 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 208 RegionNode* slow_region); 209 void generate_clear_array(const TypePtr* adr_type, 210 Node* dest, 211 BasicType basic_elem_type, 212 Node* slice_off, 213 Node* slice_len, 214 Node* slice_end); 215 bool generate_block_arraycopy(const TypePtr* adr_type, 216 BasicType basic_elem_type, 217 AllocateNode* alloc, 218 Node* src, Node* src_offset, 219 Node* dest, Node* dest_offset, 220 Node* dest_size, bool dest_uninitialized); 221 void generate_slow_arraycopy(const TypePtr* adr_type, 222 Node* src, Node* src_offset, 223 Node* dest, Node* dest_offset, 224 Node* copy_length, bool dest_uninitialized); 225 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 226 Node* dest_elem_klass, 227 Node* src, Node* src_offset, 228 Node* dest, Node* dest_offset, 229 Node* copy_length, bool dest_uninitialized); 230 Node* generate_generic_arraycopy(const TypePtr* adr_type, 231 Node* src, Node* src_offset, 232 Node* dest, Node* dest_offset, 233 Node* copy_length, bool dest_uninitialized); 234 void generate_unchecked_arraycopy(const TypePtr* adr_type, 235 BasicType basic_elem_type, 236 bool disjoint_bases, 237 Node* src, Node* src_offset, 238 Node* dest, Node* dest_offset, 239 Node* copy_length, bool dest_uninitialized); 240 bool inline_unsafe_CAS(BasicType type); 241 bool inline_unsafe_ordered_store(BasicType type); 242 bool inline_fp_conversions(vmIntrinsics::ID id); 243 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id); 244 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id); 245 bool inline_bitCount(vmIntrinsics::ID id); 246 bool inline_reverseBytes(vmIntrinsics::ID id); 247 248 bool inline_reference_get(); 249 }; 250 251 252 //---------------------------make_vm_intrinsic---------------------------- 253 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 254 vmIntrinsics::ID id = m->intrinsic_id(); 255 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 256 257 if (DisableIntrinsic[0] != '\0' 258 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 259 // disabled by a user request on the command line: 260 // example: -XX:DisableIntrinsic=_hashCode,_getClass 261 return NULL; 262 } 263 264 if (!m->is_loaded()) { 265 // do not attempt to inline unloaded methods 266 return NULL; 267 } 268 269 // Only a few intrinsics implement a virtual dispatch. 270 // They are expensive calls which are also frequently overridden. 271 if (is_virtual) { 272 switch (id) { 273 case vmIntrinsics::_hashCode: 274 case vmIntrinsics::_clone: 275 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 276 break; 277 default: 278 return NULL; 279 } 280 } 281 282 // -XX:-InlineNatives disables nearly all intrinsics: 283 if (!InlineNatives) { 284 switch (id) { 285 case vmIntrinsics::_indexOf: 286 case vmIntrinsics::_compareTo: 287 case vmIntrinsics::_equals: 288 case vmIntrinsics::_equalsC: 289 break; // InlineNatives does not control String.compareTo 290 default: 291 return NULL; 292 } 293 } 294 295 switch (id) { 296 case vmIntrinsics::_compareTo: 297 if (!SpecialStringCompareTo) return NULL; 298 break; 299 case vmIntrinsics::_indexOf: 300 if (!SpecialStringIndexOf) return NULL; 301 break; 302 case vmIntrinsics::_equals: 303 if (!SpecialStringEquals) return NULL; 304 break; 305 case vmIntrinsics::_equalsC: 306 if (!SpecialArraysEquals) return NULL; 307 break; 308 case vmIntrinsics::_arraycopy: 309 if (!InlineArrayCopy) return NULL; 310 break; 311 case vmIntrinsics::_copyMemory: 312 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 313 if (!InlineArrayCopy) return NULL; 314 break; 315 case vmIntrinsics::_hashCode: 316 if (!InlineObjectHash) return NULL; 317 break; 318 case vmIntrinsics::_clone: 319 case vmIntrinsics::_copyOf: 320 case vmIntrinsics::_copyOfRange: 321 if (!InlineObjectCopy) return NULL; 322 // These also use the arraycopy intrinsic mechanism: 323 if (!InlineArrayCopy) return NULL; 324 break; 325 case vmIntrinsics::_checkIndex: 326 // We do not intrinsify this. The optimizer does fine with it. 327 return NULL; 328 329 case vmIntrinsics::_get_AtomicLong: 330 case vmIntrinsics::_attemptUpdate: 331 if (!InlineAtomicLong) return NULL; 332 break; 333 334 case vmIntrinsics::_getCallerClass: 335 if (!UseNewReflection) return NULL; 336 if (!InlineReflectionGetCallerClass) return NULL; 337 if (!JDK_Version::is_gte_jdk14x_version()) return NULL; 338 break; 339 340 case vmIntrinsics::_bitCount_i: 341 case vmIntrinsics::_bitCount_l: 342 if (!UsePopCountInstruction) return NULL; 343 break; 344 345 case vmIntrinsics::_Reference_get: 346 // It is only when G1 is enabled that we absolutely 347 // need to use the intrinsic version of Reference.get() 348 // so that the value in the referent field, if necessary, 349 // can be registered by the pre-barrier code. 350 if (!UseG1GC) return NULL; 351 break; 352 353 default: 354 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 355 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 356 break; 357 } 358 359 // -XX:-InlineClassNatives disables natives from the Class class. 360 // The flag applies to all reflective calls, notably Array.newArray 361 // (visible to Java programmers as Array.newInstance). 362 if (m->holder()->name() == ciSymbol::java_lang_Class() || 363 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 364 if (!InlineClassNatives) return NULL; 365 } 366 367 // -XX:-InlineThreadNatives disables natives from the Thread class. 368 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 369 if (!InlineThreadNatives) return NULL; 370 } 371 372 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 373 if (m->holder()->name() == ciSymbol::java_lang_Math() || 374 m->holder()->name() == ciSymbol::java_lang_Float() || 375 m->holder()->name() == ciSymbol::java_lang_Double()) { 376 if (!InlineMathNatives) return NULL; 377 } 378 379 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 380 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 381 if (!InlineUnsafeOps) return NULL; 382 } 383 384 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); 385 } 386 387 //----------------------register_library_intrinsics----------------------- 388 // Initialize this file's data structures, for each Compile instance. 389 void Compile::register_library_intrinsics() { 390 // Nothing to do here. 391 } 392 393 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 394 LibraryCallKit kit(jvms, this); 395 Compile* C = kit.C; 396 int nodes = C->unique(); 397 #ifndef PRODUCT 398 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 399 char buf[1000]; 400 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 401 tty->print_cr("Intrinsic %s", str); 402 } 403 #endif 404 405 if (kit.try_to_inline()) { 406 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 407 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 408 } 409 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 410 if (C->log()) { 411 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 412 vmIntrinsics::name_at(intrinsic_id()), 413 (is_virtual() ? " virtual='1'" : ""), 414 C->unique() - nodes); 415 } 416 return kit.transfer_exceptions_into_jvms(); 417 } 418 419 if (PrintIntrinsics) { 420 if (jvms->has_method()) { 421 // Not a root compile. 422 tty->print("Did not inline intrinsic %s%s at bci:%d in", 423 vmIntrinsics::name_at(intrinsic_id()), 424 (is_virtual() ? " (virtual)" : ""), kit.bci()); 425 kit.caller()->print_short_name(tty); 426 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); 427 } else { 428 // Root compile 429 tty->print("Did not generate intrinsic %s%s at bci:%d in", 430 vmIntrinsics::name_at(intrinsic_id()), 431 (is_virtual() ? " (virtual)" : ""), kit.bci()); 432 } 433 } 434 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 435 return NULL; 436 } 437 438 bool LibraryCallKit::try_to_inline() { 439 // Handle symbolic names for otherwise undistinguished boolean switches: 440 const bool is_store = true; 441 const bool is_native_ptr = true; 442 const bool is_static = true; 443 444 if (!jvms()->has_method()) { 445 // Root JVMState has a null method. 446 assert(map()->memory()->Opcode() == Op_Parm, ""); 447 // Insert the memory aliasing node 448 set_all_memory(reset_memory()); 449 } 450 assert(merged_memory(), ""); 451 452 switch (intrinsic_id()) { 453 case vmIntrinsics::_hashCode: 454 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 455 case vmIntrinsics::_identityHashCode: 456 return inline_native_hashcode(/*!virtual*/ false, is_static); 457 case vmIntrinsics::_getClass: 458 return inline_native_getClass(); 459 460 case vmIntrinsics::_dsin: 461 case vmIntrinsics::_dcos: 462 case vmIntrinsics::_dtan: 463 case vmIntrinsics::_dabs: 464 case vmIntrinsics::_datan2: 465 case vmIntrinsics::_dsqrt: 466 case vmIntrinsics::_dexp: 467 case vmIntrinsics::_dlog: 468 case vmIntrinsics::_dlog10: 469 case vmIntrinsics::_dpow: 470 return inline_math_native(intrinsic_id()); 471 472 case vmIntrinsics::_min: 473 case vmIntrinsics::_max: 474 return inline_min_max(intrinsic_id()); 475 476 case vmIntrinsics::_arraycopy: 477 return inline_arraycopy(); 478 479 case vmIntrinsics::_compareTo: 480 return inline_string_compareTo(); 481 case vmIntrinsics::_indexOf: 482 return inline_string_indexOf(); 483 case vmIntrinsics::_equals: 484 return inline_string_equals(); 485 486 case vmIntrinsics::_getObject: 487 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); 488 case vmIntrinsics::_getBoolean: 489 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); 490 case vmIntrinsics::_getByte: 491 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); 492 case vmIntrinsics::_getShort: 493 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); 494 case vmIntrinsics::_getChar: 495 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); 496 case vmIntrinsics::_getInt: 497 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); 498 case vmIntrinsics::_getLong: 499 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); 500 case vmIntrinsics::_getFloat: 501 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); 502 case vmIntrinsics::_getDouble: 503 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); 504 505 case vmIntrinsics::_putObject: 506 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); 507 case vmIntrinsics::_putBoolean: 508 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); 509 case vmIntrinsics::_putByte: 510 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); 511 case vmIntrinsics::_putShort: 512 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); 513 case vmIntrinsics::_putChar: 514 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); 515 case vmIntrinsics::_putInt: 516 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); 517 case vmIntrinsics::_putLong: 518 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); 519 case vmIntrinsics::_putFloat: 520 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); 521 case vmIntrinsics::_putDouble: 522 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); 523 524 case vmIntrinsics::_getByte_raw: 525 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); 526 case vmIntrinsics::_getShort_raw: 527 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); 528 case vmIntrinsics::_getChar_raw: 529 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); 530 case vmIntrinsics::_getInt_raw: 531 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); 532 case vmIntrinsics::_getLong_raw: 533 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); 534 case vmIntrinsics::_getFloat_raw: 535 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); 536 case vmIntrinsics::_getDouble_raw: 537 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); 538 case vmIntrinsics::_getAddress_raw: 539 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); 540 541 case vmIntrinsics::_putByte_raw: 542 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); 543 case vmIntrinsics::_putShort_raw: 544 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); 545 case vmIntrinsics::_putChar_raw: 546 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); 547 case vmIntrinsics::_putInt_raw: 548 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); 549 case vmIntrinsics::_putLong_raw: 550 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); 551 case vmIntrinsics::_putFloat_raw: 552 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); 553 case vmIntrinsics::_putDouble_raw: 554 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); 555 case vmIntrinsics::_putAddress_raw: 556 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); 557 558 case vmIntrinsics::_getObjectVolatile: 559 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); 560 case vmIntrinsics::_getBooleanVolatile: 561 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); 562 case vmIntrinsics::_getByteVolatile: 563 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); 564 case vmIntrinsics::_getShortVolatile: 565 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); 566 case vmIntrinsics::_getCharVolatile: 567 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); 568 case vmIntrinsics::_getIntVolatile: 569 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); 570 case vmIntrinsics::_getLongVolatile: 571 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); 572 case vmIntrinsics::_getFloatVolatile: 573 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); 574 case vmIntrinsics::_getDoubleVolatile: 575 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); 576 577 case vmIntrinsics::_putObjectVolatile: 578 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); 579 case vmIntrinsics::_putBooleanVolatile: 580 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); 581 case vmIntrinsics::_putByteVolatile: 582 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); 583 case vmIntrinsics::_putShortVolatile: 584 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); 585 case vmIntrinsics::_putCharVolatile: 586 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); 587 case vmIntrinsics::_putIntVolatile: 588 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); 589 case vmIntrinsics::_putLongVolatile: 590 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); 591 case vmIntrinsics::_putFloatVolatile: 592 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); 593 case vmIntrinsics::_putDoubleVolatile: 594 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); 595 596 case vmIntrinsics::_prefetchRead: 597 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 598 case vmIntrinsics::_prefetchWrite: 599 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 600 case vmIntrinsics::_prefetchReadStatic: 601 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 602 case vmIntrinsics::_prefetchWriteStatic: 603 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 604 605 case vmIntrinsics::_compareAndSwapObject: 606 return inline_unsafe_CAS(T_OBJECT); 607 case vmIntrinsics::_compareAndSwapInt: 608 return inline_unsafe_CAS(T_INT); 609 case vmIntrinsics::_compareAndSwapLong: 610 return inline_unsafe_CAS(T_LONG); 611 612 case vmIntrinsics::_putOrderedObject: 613 return inline_unsafe_ordered_store(T_OBJECT); 614 case vmIntrinsics::_putOrderedInt: 615 return inline_unsafe_ordered_store(T_INT); 616 case vmIntrinsics::_putOrderedLong: 617 return inline_unsafe_ordered_store(T_LONG); 618 619 case vmIntrinsics::_currentThread: 620 return inline_native_currentThread(); 621 case vmIntrinsics::_isInterrupted: 622 return inline_native_isInterrupted(); 623 624 case vmIntrinsics::_currentTimeMillis: 625 return inline_native_time_funcs(false); 626 case vmIntrinsics::_nanoTime: 627 return inline_native_time_funcs(true); 628 case vmIntrinsics::_allocateInstance: 629 return inline_unsafe_allocate(); 630 case vmIntrinsics::_copyMemory: 631 return inline_unsafe_copyMemory(); 632 case vmIntrinsics::_newArray: 633 return inline_native_newArray(); 634 case vmIntrinsics::_getLength: 635 return inline_native_getLength(); 636 case vmIntrinsics::_copyOf: 637 return inline_array_copyOf(false); 638 case vmIntrinsics::_copyOfRange: 639 return inline_array_copyOf(true); 640 case vmIntrinsics::_equalsC: 641 return inline_array_equals(); 642 case vmIntrinsics::_clone: 643 return inline_native_clone(intrinsic()->is_virtual()); 644 645 case vmIntrinsics::_isAssignableFrom: 646 return inline_native_subtype_check(); 647 648 case vmIntrinsics::_isInstance: 649 case vmIntrinsics::_getModifiers: 650 case vmIntrinsics::_isInterface: 651 case vmIntrinsics::_isArray: 652 case vmIntrinsics::_isPrimitive: 653 case vmIntrinsics::_getSuperclass: 654 case vmIntrinsics::_getComponentType: 655 case vmIntrinsics::_getClassAccessFlags: 656 return inline_native_Class_query(intrinsic_id()); 657 658 case vmIntrinsics::_floatToRawIntBits: 659 case vmIntrinsics::_floatToIntBits: 660 case vmIntrinsics::_intBitsToFloat: 661 case vmIntrinsics::_doubleToRawLongBits: 662 case vmIntrinsics::_doubleToLongBits: 663 case vmIntrinsics::_longBitsToDouble: 664 return inline_fp_conversions(intrinsic_id()); 665 666 case vmIntrinsics::_numberOfLeadingZeros_i: 667 case vmIntrinsics::_numberOfLeadingZeros_l: 668 return inline_numberOfLeadingZeros(intrinsic_id()); 669 670 case vmIntrinsics::_numberOfTrailingZeros_i: 671 case vmIntrinsics::_numberOfTrailingZeros_l: 672 return inline_numberOfTrailingZeros(intrinsic_id()); 673 674 case vmIntrinsics::_bitCount_i: 675 case vmIntrinsics::_bitCount_l: 676 return inline_bitCount(intrinsic_id()); 677 678 case vmIntrinsics::_reverseBytes_i: 679 case vmIntrinsics::_reverseBytes_l: 680 case vmIntrinsics::_reverseBytes_s: 681 case vmIntrinsics::_reverseBytes_c: 682 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); 683 684 case vmIntrinsics::_get_AtomicLong: 685 return inline_native_AtomicLong_get(); 686 case vmIntrinsics::_attemptUpdate: 687 return inline_native_AtomicLong_attemptUpdate(); 688 689 case vmIntrinsics::_getCallerClass: 690 return inline_native_Reflection_getCallerClass(); 691 692 case vmIntrinsics::_Reference_get: 693 return inline_reference_get(); 694 695 default: 696 // If you get here, it may be that someone has added a new intrinsic 697 // to the list in vmSymbols.hpp without implementing it here. 698 #ifndef PRODUCT 699 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 700 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 701 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 702 } 703 #endif 704 return false; 705 } 706 } 707 708 //------------------------------push_result------------------------------ 709 // Helper function for finishing intrinsics. 710 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { 711 record_for_igvn(region); 712 set_control(_gvn.transform(region)); 713 BasicType value_type = value->type()->basic_type(); 714 push_node(value_type, _gvn.transform(value)); 715 } 716 717 //------------------------------generate_guard--------------------------- 718 // Helper function for generating guarded fast-slow graph structures. 719 // The given 'test', if true, guards a slow path. If the test fails 720 // then a fast path can be taken. (We generally hope it fails.) 721 // In all cases, GraphKit::control() is updated to the fast path. 722 // The returned value represents the control for the slow path. 723 // The return value is never 'top'; it is either a valid control 724 // or NULL if it is obvious that the slow path can never be taken. 725 // Also, if region and the slow control are not NULL, the slow edge 726 // is appended to the region. 727 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 728 if (stopped()) { 729 // Already short circuited. 730 return NULL; 731 } 732 733 // Build an if node and its projections. 734 // If test is true we take the slow path, which we assume is uncommon. 735 if (_gvn.type(test) == TypeInt::ZERO) { 736 // The slow branch is never taken. No need to build this guard. 737 return NULL; 738 } 739 740 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 741 742 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); 743 if (if_slow == top()) { 744 // The slow branch is never taken. No need to build this guard. 745 return NULL; 746 } 747 748 if (region != NULL) 749 region->add_req(if_slow); 750 751 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); 752 set_control(if_fast); 753 754 return if_slow; 755 } 756 757 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 758 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 759 } 760 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 761 return generate_guard(test, region, PROB_FAIR); 762 } 763 764 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 765 Node* *pos_index) { 766 if (stopped()) 767 return NULL; // already stopped 768 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 769 return NULL; // index is already adequately typed 770 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 771 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 772 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 773 if (is_neg != NULL && pos_index != NULL) { 774 // Emulate effect of Parse::adjust_map_after_if. 775 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); 776 ccast->set_req(0, control()); 777 (*pos_index) = _gvn.transform(ccast); 778 } 779 return is_neg; 780 } 781 782 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 783 Node* *pos_index) { 784 if (stopped()) 785 return NULL; // already stopped 786 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 787 return NULL; // index is already adequately typed 788 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 789 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 790 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) ); 791 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 792 if (is_notp != NULL && pos_index != NULL) { 793 // Emulate effect of Parse::adjust_map_after_if. 794 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1); 795 ccast->set_req(0, control()); 796 (*pos_index) = _gvn.transform(ccast); 797 } 798 return is_notp; 799 } 800 801 // Make sure that 'position' is a valid limit index, in [0..length]. 802 // There are two equivalent plans for checking this: 803 // A. (offset + copyLength) unsigned<= arrayLength 804 // B. offset <= (arrayLength - copyLength) 805 // We require that all of the values above, except for the sum and 806 // difference, are already known to be non-negative. 807 // Plan A is robust in the face of overflow, if offset and copyLength 808 // are both hugely positive. 809 // 810 // Plan B is less direct and intuitive, but it does not overflow at 811 // all, since the difference of two non-negatives is always 812 // representable. Whenever Java methods must perform the equivalent 813 // check they generally use Plan B instead of Plan A. 814 // For the moment we use Plan A. 815 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 816 Node* subseq_length, 817 Node* array_length, 818 RegionNode* region) { 819 if (stopped()) 820 return NULL; // already stopped 821 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 822 if (zero_offset && subseq_length->eqv_uncast(array_length)) 823 return NULL; // common case of whole-array copy 824 Node* last = subseq_length; 825 if (!zero_offset) // last += offset 826 last = _gvn.transform( new (C, 3) AddINode(last, offset)); 827 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); 828 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 829 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 830 return is_over; 831 } 832 833 834 //--------------------------generate_current_thread-------------------- 835 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 836 ciKlass* thread_klass = env()->Thread_klass(); 837 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 838 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); 839 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 840 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 841 tls_output = thread; 842 return threadObj; 843 } 844 845 846 //------------------------------make_string_method_node------------------------ 847 // Helper method for String intrinsic finctions. 848 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) { 849 const int value_offset = java_lang_String::value_offset_in_bytes(); 850 const int count_offset = java_lang_String::count_offset_in_bytes(); 851 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 852 853 Node* no_ctrl = NULL; 854 855 ciInstanceKlass* klass = env()->String_klass(); 856 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 857 858 const TypeAryPtr* value_type = 859 TypeAryPtr::make(TypePtr::NotNull, 860 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 861 ciTypeArrayKlass::make(T_CHAR), true, 0); 862 863 // Get start addr of string and substring 864 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset); 865 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset)); 866 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset); 867 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 868 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 869 870 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset); 871 Node* str2_value = make_load(no_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset)); 872 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset); 873 Node* str2_offset = make_load(no_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 874 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 875 876 Node* result = NULL; 877 switch (opcode) { 878 case Op_StrIndexOf: 879 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 880 str1_start, cnt1, str2_start, cnt2); 881 break; 882 case Op_StrComp: 883 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS), 884 str1_start, cnt1, str2_start, cnt2); 885 break; 886 case Op_StrEquals: 887 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 888 str1_start, str2_start, cnt1); 889 break; 890 default: 891 ShouldNotReachHere(); 892 return NULL; 893 } 894 895 // All these intrinsics have checks. 896 C->set_has_split_ifs(true); // Has chance for split-if optimization 897 898 return _gvn.transform(result); 899 } 900 901 //------------------------------inline_string_compareTo------------------------ 902 bool LibraryCallKit::inline_string_compareTo() { 903 904 if (!Matcher::has_match_rule(Op_StrComp)) return false; 905 906 const int value_offset = java_lang_String::value_offset_in_bytes(); 907 const int count_offset = java_lang_String::count_offset_in_bytes(); 908 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 909 910 _sp += 2; 911 Node *argument = pop(); // pop non-receiver first: it was pushed second 912 Node *receiver = pop(); 913 914 // Null check on self without removing any arguments. The argument 915 // null check technically happens in the wrong place, which can lead to 916 // invalid stack traces when string compare is inlined into a method 917 // which handles NullPointerExceptions. 918 _sp += 2; 919 receiver = do_null_check(receiver, T_OBJECT); 920 argument = do_null_check(argument, T_OBJECT); 921 _sp -= 2; 922 if (stopped()) { 923 return true; 924 } 925 926 ciInstanceKlass* klass = env()->String_klass(); 927 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 928 Node* no_ctrl = NULL; 929 930 // Get counts for string and argument 931 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset); 932 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 933 934 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset); 935 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 936 937 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt); 938 push(compare); 939 return true; 940 } 941 942 //------------------------------inline_string_equals------------------------ 943 bool LibraryCallKit::inline_string_equals() { 944 945 if (!Matcher::has_match_rule(Op_StrEquals)) return false; 946 947 const int value_offset = java_lang_String::value_offset_in_bytes(); 948 const int count_offset = java_lang_String::count_offset_in_bytes(); 949 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 950 951 int nargs = 2; 952 _sp += nargs; 953 Node* argument = pop(); // pop non-receiver first: it was pushed second 954 Node* receiver = pop(); 955 956 // Null check on self without removing any arguments. The argument 957 // null check technically happens in the wrong place, which can lead to 958 // invalid stack traces when string compare is inlined into a method 959 // which handles NullPointerExceptions. 960 _sp += nargs; 961 receiver = do_null_check(receiver, T_OBJECT); 962 //should not do null check for argument for String.equals(), because spec 963 //allows to specify NULL as argument. 964 _sp -= nargs; 965 966 if (stopped()) { 967 return true; 968 } 969 970 // paths (plus control) merge 971 RegionNode* region = new (C, 5) RegionNode(5); 972 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL); 973 974 // does source == target string? 975 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument)); 976 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq)); 977 978 Node* if_eq = generate_slow_guard(bol, NULL); 979 if (if_eq != NULL) { 980 // receiver == argument 981 phi->init_req(2, intcon(1)); 982 region->init_req(2, if_eq); 983 } 984 985 // get String klass for instanceOf 986 ciInstanceKlass* klass = env()->String_klass(); 987 988 if (!stopped()) { 989 _sp += nargs; // gen_instanceof might do an uncommon trap 990 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 991 _sp -= nargs; 992 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1))); 993 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne)); 994 995 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 996 //instanceOf == true, fallthrough 997 998 if (inst_false != NULL) { 999 phi->init_req(3, intcon(0)); 1000 region->init_req(3, inst_false); 1001 } 1002 } 1003 1004 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1005 1006 Node* no_ctrl = NULL; 1007 Node* receiver_cnt; 1008 Node* argument_cnt; 1009 1010 if (!stopped()) { 1011 // Properly cast the argument to String 1012 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type)); 1013 // This path is taken only when argument's type is String:NotNull. 1014 argument = cast_not_null(argument, false); 1015 1016 // Get counts for string and argument 1017 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset); 1018 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1019 1020 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset); 1021 argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1022 1023 // Check for receiver count != argument count 1024 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) ); 1025 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) ); 1026 Node* if_ne = generate_slow_guard(bol, NULL); 1027 if (if_ne != NULL) { 1028 phi->init_req(4, intcon(0)); 1029 region->init_req(4, if_ne); 1030 } 1031 } 1032 1033 // Check for count == 0 is done by mach node StrEquals. 1034 1035 if (!stopped()) { 1036 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt); 1037 phi->init_req(1, equals); 1038 region->init_req(1, control()); 1039 } 1040 1041 // post merge 1042 set_control(_gvn.transform(region)); 1043 record_for_igvn(region); 1044 1045 push(_gvn.transform(phi)); 1046 1047 return true; 1048 } 1049 1050 //------------------------------inline_array_equals---------------------------- 1051 bool LibraryCallKit::inline_array_equals() { 1052 1053 if (!Matcher::has_match_rule(Op_AryEq)) return false; 1054 1055 _sp += 2; 1056 Node *argument2 = pop(); 1057 Node *argument1 = pop(); 1058 1059 Node* equals = 1060 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS), 1061 argument1, argument2) ); 1062 push(equals); 1063 return true; 1064 } 1065 1066 // Java version of String.indexOf(constant string) 1067 // class StringDecl { 1068 // StringDecl(char[] ca) { 1069 // offset = 0; 1070 // count = ca.length; 1071 // value = ca; 1072 // } 1073 // int offset; 1074 // int count; 1075 // char[] value; 1076 // } 1077 // 1078 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1079 // int targetOffset, int cache_i, int md2) { 1080 // int cache = cache_i; 1081 // int sourceOffset = string_object.offset; 1082 // int sourceCount = string_object.count; 1083 // int targetCount = target_object.length; 1084 // 1085 // int targetCountLess1 = targetCount - 1; 1086 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1087 // 1088 // char[] source = string_object.value; 1089 // char[] target = target_object; 1090 // int lastChar = target[targetCountLess1]; 1091 // 1092 // outer_loop: 1093 // for (int i = sourceOffset; i < sourceEnd; ) { 1094 // int src = source[i + targetCountLess1]; 1095 // if (src == lastChar) { 1096 // // With random strings and a 4-character alphabet, 1097 // // reverse matching at this point sets up 0.8% fewer 1098 // // frames, but (paradoxically) makes 0.3% more probes. 1099 // // Since those probes are nearer the lastChar probe, 1100 // // there is may be a net D$ win with reverse matching. 1101 // // But, reversing loop inhibits unroll of inner loop 1102 // // for unknown reason. So, does running outer loop from 1103 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1104 // for (int j = 0; j < targetCountLess1; j++) { 1105 // if (target[targetOffset + j] != source[i+j]) { 1106 // if ((cache & (1 << source[i+j])) == 0) { 1107 // if (md2 < j+1) { 1108 // i += j+1; 1109 // continue outer_loop; 1110 // } 1111 // } 1112 // i += md2; 1113 // continue outer_loop; 1114 // } 1115 // } 1116 // return i - sourceOffset; 1117 // } 1118 // if ((cache & (1 << src)) == 0) { 1119 // i += targetCountLess1; 1120 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1121 // i++; 1122 // } 1123 // return -1; 1124 // } 1125 1126 //------------------------------string_indexOf------------------------ 1127 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1128 jint cache_i, jint md2_i) { 1129 1130 Node* no_ctrl = NULL; 1131 float likely = PROB_LIKELY(0.9); 1132 float unlikely = PROB_UNLIKELY(0.9); 1133 1134 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate 1135 1136 const int value_offset = java_lang_String::value_offset_in_bytes(); 1137 const int count_offset = java_lang_String::count_offset_in_bytes(); 1138 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1139 1140 ciInstanceKlass* klass = env()->String_klass(); 1141 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1142 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0); 1143 1144 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset); 1145 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 1146 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset); 1147 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1148 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset); 1149 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset)); 1150 1151 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) ); 1152 jint target_length = target_array->length(); 1153 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1154 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1155 1156 IdealKit kit(this, false, true); 1157 #define __ kit. 1158 Node* zero = __ ConI(0); 1159 Node* one = __ ConI(1); 1160 Node* cache = __ ConI(cache_i); 1161 Node* md2 = __ ConI(md2_i); 1162 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1163 Node* targetCount = __ ConI(target_length); 1164 Node* targetCountLess1 = __ ConI(target_length - 1); 1165 Node* targetOffset = __ ConI(targetOffset_i); 1166 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1167 1168 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1169 Node* outer_loop = __ make_label(2 /* goto */); 1170 Node* return_ = __ make_label(1); 1171 1172 __ set(rtn,__ ConI(-1)); 1173 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1174 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1175 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1176 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1177 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1178 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1179 Node* tpj = __ AddI(targetOffset, __ value(j)); 1180 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1181 Node* ipj = __ AddI(__ value(i), __ value(j)); 1182 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1183 __ if_then(targ, BoolTest::ne, src2); { 1184 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1185 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1186 __ increment(i, __ AddI(__ value(j), one)); 1187 __ goto_(outer_loop); 1188 } __ end_if(); __ dead(j); 1189 }__ end_if(); __ dead(j); 1190 __ increment(i, md2); 1191 __ goto_(outer_loop); 1192 }__ end_if(); 1193 __ increment(j, one); 1194 }__ end_loop(); __ dead(j); 1195 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1196 __ goto_(return_); 1197 }__ end_if(); 1198 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1199 __ increment(i, targetCountLess1); 1200 }__ end_if(); 1201 __ increment(i, one); 1202 __ bind(outer_loop); 1203 }__ end_loop(); __ dead(i); 1204 __ bind(return_); 1205 1206 // Final sync IdealKit and GraphKit. 1207 final_sync(kit); 1208 Node* result = __ value(rtn); 1209 #undef __ 1210 C->set_has_loops(true); 1211 return result; 1212 } 1213 1214 //------------------------------inline_string_indexOf------------------------ 1215 bool LibraryCallKit::inline_string_indexOf() { 1216 1217 const int value_offset = java_lang_String::value_offset_in_bytes(); 1218 const int count_offset = java_lang_String::count_offset_in_bytes(); 1219 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1220 1221 _sp += 2; 1222 Node *argument = pop(); // pop non-receiver first: it was pushed second 1223 Node *receiver = pop(); 1224 1225 Node* result; 1226 // Disable the use of pcmpestri until it can be guaranteed that 1227 // the load doesn't cross into the uncommited space. 1228 if (Matcher::has_match_rule(Op_StrIndexOf) && 1229 UseSSE42Intrinsics) { 1230 // Generate SSE4.2 version of indexOf 1231 // We currently only have match rules that use SSE4.2 1232 1233 // Null check on self without removing any arguments. The argument 1234 // null check technically happens in the wrong place, which can lead to 1235 // invalid stack traces when string compare is inlined into a method 1236 // which handles NullPointerExceptions. 1237 _sp += 2; 1238 receiver = do_null_check(receiver, T_OBJECT); 1239 argument = do_null_check(argument, T_OBJECT); 1240 _sp -= 2; 1241 1242 if (stopped()) { 1243 return true; 1244 } 1245 1246 ciInstanceKlass* str_klass = env()->String_klass(); 1247 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1248 1249 // Make the merge point 1250 RegionNode* result_rgn = new (C, 4) RegionNode(4); 1251 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT); 1252 Node* no_ctrl = NULL; 1253 1254 // Get counts for string and substr 1255 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset); 1256 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1257 1258 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset); 1259 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1260 1261 // Check for substr count > string count 1262 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) ); 1263 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) ); 1264 Node* if_gt = generate_slow_guard(bol, NULL); 1265 if (if_gt != NULL) { 1266 result_phi->init_req(2, intcon(-1)); 1267 result_rgn->init_req(2, if_gt); 1268 } 1269 1270 if (!stopped()) { 1271 // Check for substr count == 0 1272 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) ); 1273 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 1274 Node* if_zero = generate_slow_guard(bol, NULL); 1275 if (if_zero != NULL) { 1276 result_phi->init_req(3, intcon(0)); 1277 result_rgn->init_req(3, if_zero); 1278 } 1279 } 1280 1281 if (!stopped()) { 1282 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt); 1283 result_phi->init_req(1, result); 1284 result_rgn->init_req(1, control()); 1285 } 1286 set_control(_gvn.transform(result_rgn)); 1287 record_for_igvn(result_rgn); 1288 result = _gvn.transform(result_phi); 1289 1290 } else { // Use LibraryCallKit::string_indexOf 1291 // don't intrinsify if argument isn't a constant string. 1292 if (!argument->is_Con()) { 1293 return false; 1294 } 1295 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); 1296 if (str_type == NULL) { 1297 return false; 1298 } 1299 ciInstanceKlass* klass = env()->String_klass(); 1300 ciObject* str_const = str_type->const_oop(); 1301 if (str_const == NULL || str_const->klass() != klass) { 1302 return false; 1303 } 1304 ciInstance* str = str_const->as_instance(); 1305 assert(str != NULL, "must be instance"); 1306 1307 ciObject* v = str->field_value_by_offset(value_offset).as_object(); 1308 int o = str->field_value_by_offset(offset_offset).as_int(); 1309 int c = str->field_value_by_offset(count_offset).as_int(); 1310 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1311 1312 // constant strings have no offset and count == length which 1313 // simplifies the resulting code somewhat so lets optimize for that. 1314 if (o != 0 || c != pat->length()) { 1315 return false; 1316 } 1317 1318 // Null check on self without removing any arguments. The argument 1319 // null check technically happens in the wrong place, which can lead to 1320 // invalid stack traces when string compare is inlined into a method 1321 // which handles NullPointerExceptions. 1322 _sp += 2; 1323 receiver = do_null_check(receiver, T_OBJECT); 1324 // No null check on the argument is needed since it's a constant String oop. 1325 _sp -= 2; 1326 if (stopped()) { 1327 return true; 1328 } 1329 1330 // The null string as a pattern always returns 0 (match at beginning of string) 1331 if (c == 0) { 1332 push(intcon(0)); 1333 return true; 1334 } 1335 1336 // Generate default indexOf 1337 jchar lastChar = pat->char_at(o + (c - 1)); 1338 int cache = 0; 1339 int i; 1340 for (i = 0; i < c - 1; i++) { 1341 assert(i < pat->length(), "out of range"); 1342 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1343 } 1344 1345 int md2 = c; 1346 for (i = 0; i < c - 1; i++) { 1347 assert(i < pat->length(), "out of range"); 1348 if (pat->char_at(o + i) == lastChar) { 1349 md2 = (c - 1) - i; 1350 } 1351 } 1352 1353 result = string_indexOf(receiver, pat, o, cache, md2); 1354 } 1355 1356 push(result); 1357 return true; 1358 } 1359 1360 //--------------------------pop_math_arg-------------------------------- 1361 // Pop a double argument to a math function from the stack 1362 // rounding it if necessary. 1363 Node * LibraryCallKit::pop_math_arg() { 1364 Node *arg = pop_pair(); 1365 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) 1366 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); 1367 return arg; 1368 } 1369 1370 //------------------------------inline_trig---------------------------------- 1371 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1372 // argument reduction which will turn into a fast/slow diamond. 1373 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1374 _sp += arg_size(); // restore stack pointer 1375 Node* arg = pop_math_arg(); 1376 Node* trig = NULL; 1377 1378 switch (id) { 1379 case vmIntrinsics::_dsin: 1380 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); 1381 break; 1382 case vmIntrinsics::_dcos: 1383 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); 1384 break; 1385 case vmIntrinsics::_dtan: 1386 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); 1387 break; 1388 default: 1389 assert(false, "bad intrinsic was passed in"); 1390 return false; 1391 } 1392 1393 // Rounding required? Check for argument reduction! 1394 if( Matcher::strict_fp_requires_explicit_rounding ) { 1395 1396 static const double pi_4 = 0.7853981633974483; 1397 static const double neg_pi_4 = -0.7853981633974483; 1398 // pi/2 in 80-bit extended precision 1399 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1400 // -pi/2 in 80-bit extended precision 1401 // 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}; 1402 // Cutoff value for using this argument reduction technique 1403 //static const double pi_2_minus_epsilon = 1.564660403643354; 1404 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1405 1406 // Pseudocode for sin: 1407 // if (x <= Math.PI / 4.0) { 1408 // if (x >= -Math.PI / 4.0) return fsin(x); 1409 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1410 // } else { 1411 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1412 // } 1413 // return StrictMath.sin(x); 1414 1415 // Pseudocode for cos: 1416 // if (x <= Math.PI / 4.0) { 1417 // if (x >= -Math.PI / 4.0) return fcos(x); 1418 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1419 // } else { 1420 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1421 // } 1422 // return StrictMath.cos(x); 1423 1424 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1425 // requires a special machine instruction to load it. Instead we'll try 1426 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1427 // probably do the math inside the SIN encoding. 1428 1429 // Make the merge point 1430 RegionNode *r = new (C, 3) RegionNode(3); 1431 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); 1432 1433 // Flatten arg so we need only 1 test 1434 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); 1435 // Node for PI/4 constant 1436 Node *pi4 = makecon(TypeD::make(pi_4)); 1437 // Check PI/4 : abs(arg) 1438 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); 1439 // Check: If PI/4 < abs(arg) then go slow 1440 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); 1441 // Branch either way 1442 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1443 set_control(opt_iff(r,iff)); 1444 1445 // Set fast path result 1446 phi->init_req(2,trig); 1447 1448 // Slow path - non-blocking leaf call 1449 Node* call = NULL; 1450 switch (id) { 1451 case vmIntrinsics::_dsin: 1452 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1453 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1454 "Sin", NULL, arg, top()); 1455 break; 1456 case vmIntrinsics::_dcos: 1457 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1458 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1459 "Cos", NULL, arg, top()); 1460 break; 1461 case vmIntrinsics::_dtan: 1462 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1463 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1464 "Tan", NULL, arg, top()); 1465 break; 1466 } 1467 assert(control()->in(0) == call, ""); 1468 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); 1469 r->init_req(1,control()); 1470 phi->init_req(1,slow_result); 1471 1472 // Post-merge 1473 set_control(_gvn.transform(r)); 1474 record_for_igvn(r); 1475 trig = _gvn.transform(phi); 1476 1477 C->set_has_split_ifs(true); // Has chance for split-if optimization 1478 } 1479 // Push result back on JVM stack 1480 push_pair(trig); 1481 return true; 1482 } 1483 1484 //------------------------------inline_sqrt------------------------------------- 1485 // Inline square root instruction, if possible. 1486 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { 1487 assert(id == vmIntrinsics::_dsqrt, "Not square root"); 1488 _sp += arg_size(); // restore stack pointer 1489 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); 1490 return true; 1491 } 1492 1493 //------------------------------inline_abs------------------------------------- 1494 // Inline absolute value instruction, if possible. 1495 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { 1496 assert(id == vmIntrinsics::_dabs, "Not absolute value"); 1497 _sp += arg_size(); // restore stack pointer 1498 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); 1499 return true; 1500 } 1501 1502 //------------------------------inline_exp------------------------------------- 1503 // Inline exp instructions, if possible. The Intel hardware only misses 1504 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1505 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { 1506 assert(id == vmIntrinsics::_dexp, "Not exp"); 1507 1508 // If this inlining ever returned NaN in the past, we do not intrinsify it 1509 // every again. NaN results requires StrictMath.exp handling. 1510 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1511 1512 // Do not intrinsify on older platforms which lack cmove. 1513 if (ConditionalMoveLimit == 0) return false; 1514 1515 _sp += arg_size(); // restore stack pointer 1516 Node *x = pop_math_arg(); 1517 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); 1518 1519 //------------------- 1520 //result=(result.isNaN())? StrictMath::exp():result; 1521 // Check: If isNaN() by checking result!=result? then go to Strict Math 1522 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1523 // Build the boolean node 1524 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1525 1526 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1527 // End the current control-flow path 1528 push_pair(x); 1529 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp 1530 // to handle. Recompile without intrinsifying Math.exp 1531 uncommon_trap(Deoptimization::Reason_intrinsic, 1532 Deoptimization::Action_make_not_entrant); 1533 } 1534 1535 C->set_has_split_ifs(true); // Has chance for split-if optimization 1536 1537 push_pair(result); 1538 1539 return true; 1540 } 1541 1542 //------------------------------inline_pow------------------------------------- 1543 // Inline power instructions, if possible. 1544 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { 1545 assert(id == vmIntrinsics::_dpow, "Not pow"); 1546 1547 // If this inlining ever returned NaN in the past, we do not intrinsify it 1548 // every again. NaN results requires StrictMath.pow handling. 1549 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1550 1551 // Do not intrinsify on older platforms which lack cmove. 1552 if (ConditionalMoveLimit == 0) return false; 1553 1554 // Pseudocode for pow 1555 // if (x <= 0.0) { 1556 // if ((double)((int)y)==y) { // if y is int 1557 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y) 1558 // } else { 1559 // result = NaN; 1560 // } 1561 // } else { 1562 // result = DPow(x,y); 1563 // } 1564 // if (result != result)? { 1565 // uncommon_trap(); 1566 // } 1567 // return result; 1568 1569 _sp += arg_size(); // restore stack pointer 1570 Node* y = pop_math_arg(); 1571 Node* x = pop_math_arg(); 1572 1573 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); 1574 1575 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow 1576 // inside of something) then skip the fancy tests and just check for 1577 // NaN result. 1578 Node *result = NULL; 1579 if( jvms()->depth() >= 1 ) { 1580 result = fast_result; 1581 } else { 1582 1583 // Set the merge point for If node with condition of (x <= 0.0) 1584 // There are four possible paths to region node and phi node 1585 RegionNode *r = new (C, 4) RegionNode(4); 1586 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); 1587 1588 // Build the first if node: if (x <= 0.0) 1589 // Node for 0 constant 1590 Node *zeronode = makecon(TypeD::ZERO); 1591 // Check x:0 1592 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); 1593 // Check: If (x<=0) then go complex path 1594 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); 1595 // Branch either way 1596 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1597 Node *opt_test = _gvn.transform(if1); 1598 //assert( opt_test->is_If(), "Expect an IfNode"); 1599 IfNode *opt_if1 = (IfNode*)opt_test; 1600 // Fast path taken; set region slot 3 1601 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) ); 1602 r->init_req(3,fast_taken); // Capture fast-control 1603 1604 // Fast path not-taken, i.e. slow path 1605 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) ); 1606 1607 // Set fast path result 1608 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) ); 1609 phi->init_req(3, fast_result); 1610 1611 // Complex path 1612 // Build the second if node (if y is int) 1613 // Node for (int)y 1614 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y)); 1615 // Node for (double)((int) y) 1616 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty)); 1617 // Check (double)((int) y) : y 1618 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y)); 1619 // Check if (y isn't int) then go to slow path 1620 1621 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) ); 1622 // Branch either way 1623 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1624 Node *slow_path = opt_iff(r,if2); // Set region path 2 1625 1626 // Calculate DPow(abs(x), y)*(1 & (int)y) 1627 // Node for constant 1 1628 Node *conone = intcon(1); 1629 // 1& (int)y 1630 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) ); 1631 // zero node 1632 Node *conzero = intcon(0); 1633 // Check (1&(int)y)==0? 1634 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero)); 1635 // Check if (1&(int)y)!=0?, if so the result is negative 1636 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); 1637 // abs(x) 1638 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); 1639 // abs(x)^y 1640 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) ); 1641 // -abs(x)^y 1642 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); 1643 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1644 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1645 // Set complex path fast result 1646 phi->init_req(2, signresult); 1647 1648 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1649 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1650 r->init_req(1,slow_path); 1651 phi->init_req(1,slow_result); 1652 1653 // Post merge 1654 set_control(_gvn.transform(r)); 1655 record_for_igvn(r); 1656 result=_gvn.transform(phi); 1657 } 1658 1659 //------------------- 1660 //result=(result.isNaN())? uncommon_trap():result; 1661 // Check: If isNaN() by checking result!=result? then go to Strict Math 1662 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1663 // Build the boolean node 1664 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1665 1666 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1667 // End the current control-flow path 1668 push_pair(x); 1669 push_pair(y); 1670 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow 1671 // to handle. Recompile without intrinsifying Math.pow. 1672 uncommon_trap(Deoptimization::Reason_intrinsic, 1673 Deoptimization::Action_make_not_entrant); 1674 } 1675 1676 C->set_has_split_ifs(true); // Has chance for split-if optimization 1677 1678 push_pair(result); 1679 1680 return true; 1681 } 1682 1683 //------------------------------inline_trans------------------------------------- 1684 // Inline transcendental instructions, if possible. The Intel hardware gets 1685 // these right, no funny corner cases missed. 1686 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { 1687 _sp += arg_size(); // restore stack pointer 1688 Node* arg = pop_math_arg(); 1689 Node* trans = NULL; 1690 1691 switch (id) { 1692 case vmIntrinsics::_dlog: 1693 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); 1694 break; 1695 case vmIntrinsics::_dlog10: 1696 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); 1697 break; 1698 default: 1699 assert(false, "bad intrinsic was passed in"); 1700 return false; 1701 } 1702 1703 // Push result back on JVM stack 1704 push_pair(trans); 1705 return true; 1706 } 1707 1708 //------------------------------runtime_math----------------------------- 1709 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1710 Node* a = NULL; 1711 Node* b = NULL; 1712 1713 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1714 "must be (DD)D or (D)D type"); 1715 1716 // Inputs 1717 _sp += arg_size(); // restore stack pointer 1718 if (call_type == OptoRuntime::Math_DD_D_Type()) { 1719 b = pop_math_arg(); 1720 } 1721 a = pop_math_arg(); 1722 1723 const TypePtr* no_memory_effects = NULL; 1724 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1725 no_memory_effects, 1726 a, top(), b, b ? top() : NULL); 1727 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); 1728 #ifdef ASSERT 1729 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); 1730 assert(value_top == top(), "second value must be top"); 1731 #endif 1732 1733 push_pair(value); 1734 return true; 1735 } 1736 1737 //------------------------------inline_math_native----------------------------- 1738 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1739 switch (id) { 1740 // These intrinsics are not properly supported on all hardware 1741 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1742 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); 1743 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1744 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); 1745 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1746 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); 1747 1748 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : 1749 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); 1750 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : 1751 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); 1752 1753 // These intrinsics are supported on all hardware 1754 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; 1755 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; 1756 1757 // These intrinsics don't work on X86. The ad implementation doesn't 1758 // handle NaN's properly. Instead of returning infinity, the ad 1759 // implementation returns a NaN on overflow. See bug: 6304089 1760 // Once the ad implementations are fixed, change the code below 1761 // to match the intrinsics above 1762 1763 case vmIntrinsics::_dexp: return 1764 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1765 case vmIntrinsics::_dpow: return 1766 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1767 1768 // These intrinsics are not yet correctly implemented 1769 case vmIntrinsics::_datan2: 1770 return false; 1771 1772 default: 1773 ShouldNotReachHere(); 1774 return false; 1775 } 1776 } 1777 1778 static bool is_simple_name(Node* n) { 1779 return (n->req() == 1 // constant 1780 || (n->is_Type() && n->as_Type()->type()->singleton()) 1781 || n->is_Proj() // parameter or return value 1782 || n->is_Phi() // local of some sort 1783 ); 1784 } 1785 1786 //----------------------------inline_min_max----------------------------------- 1787 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1788 push(generate_min_max(id, argument(0), argument(1))); 1789 1790 return true; 1791 } 1792 1793 Node* 1794 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1795 // These are the candidate return value: 1796 Node* xvalue = x0; 1797 Node* yvalue = y0; 1798 1799 if (xvalue == yvalue) { 1800 return xvalue; 1801 } 1802 1803 bool want_max = (id == vmIntrinsics::_max); 1804 1805 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1806 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1807 if (txvalue == NULL || tyvalue == NULL) return top(); 1808 // This is not really necessary, but it is consistent with a 1809 // hypothetical MaxINode::Value method: 1810 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1811 1812 // %%% This folding logic should (ideally) be in a different place. 1813 // Some should be inside IfNode, and there to be a more reliable 1814 // transformation of ?: style patterns into cmoves. We also want 1815 // more powerful optimizations around cmove and min/max. 1816 1817 // Try to find a dominating comparison of these guys. 1818 // It can simplify the index computation for Arrays.copyOf 1819 // and similar uses of System.arraycopy. 1820 // First, compute the normalized version of CmpI(x, y). 1821 int cmp_op = Op_CmpI; 1822 Node* xkey = xvalue; 1823 Node* ykey = yvalue; 1824 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); 1825 if (ideal_cmpxy->is_Cmp()) { 1826 // E.g., if we have CmpI(length - offset, count), 1827 // it might idealize to CmpI(length, count + offset) 1828 cmp_op = ideal_cmpxy->Opcode(); 1829 xkey = ideal_cmpxy->in(1); 1830 ykey = ideal_cmpxy->in(2); 1831 } 1832 1833 // Start by locating any relevant comparisons. 1834 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1835 Node* cmpxy = NULL; 1836 Node* cmpyx = NULL; 1837 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1838 Node* cmp = start_from->fast_out(k); 1839 if (cmp->outcnt() > 0 && // must have prior uses 1840 cmp->in(0) == NULL && // must be context-independent 1841 cmp->Opcode() == cmp_op) { // right kind of compare 1842 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1843 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1844 } 1845 } 1846 1847 const int NCMPS = 2; 1848 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1849 int cmpn; 1850 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1851 if (cmps[cmpn] != NULL) break; // find a result 1852 } 1853 if (cmpn < NCMPS) { 1854 // Look for a dominating test that tells us the min and max. 1855 int depth = 0; // Limit search depth for speed 1856 Node* dom = control(); 1857 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1858 if (++depth >= 100) break; 1859 Node* ifproj = dom; 1860 if (!ifproj->is_Proj()) continue; 1861 Node* iff = ifproj->in(0); 1862 if (!iff->is_If()) continue; 1863 Node* bol = iff->in(1); 1864 if (!bol->is_Bool()) continue; 1865 Node* cmp = bol->in(1); 1866 if (cmp == NULL) continue; 1867 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1868 if (cmps[cmpn] == cmp) break; 1869 if (cmpn == NCMPS) continue; 1870 BoolTest::mask btest = bol->as_Bool()->_test._test; 1871 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1872 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1873 // At this point, we know that 'x btest y' is true. 1874 switch (btest) { 1875 case BoolTest::eq: 1876 // They are proven equal, so we can collapse the min/max. 1877 // Either value is the answer. Choose the simpler. 1878 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1879 return yvalue; 1880 return xvalue; 1881 case BoolTest::lt: // x < y 1882 case BoolTest::le: // x <= y 1883 return (want_max ? yvalue : xvalue); 1884 case BoolTest::gt: // x > y 1885 case BoolTest::ge: // x >= y 1886 return (want_max ? xvalue : yvalue); 1887 } 1888 } 1889 } 1890 1891 // We failed to find a dominating test. 1892 // Let's pick a test that might GVN with prior tests. 1893 Node* best_bol = NULL; 1894 BoolTest::mask best_btest = BoolTest::illegal; 1895 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1896 Node* cmp = cmps[cmpn]; 1897 if (cmp == NULL) continue; 1898 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1899 Node* bol = cmp->fast_out(j); 1900 if (!bol->is_Bool()) continue; 1901 BoolTest::mask btest = bol->as_Bool()->_test._test; 1902 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1903 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1904 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1905 best_bol = bol->as_Bool(); 1906 best_btest = btest; 1907 } 1908 } 1909 } 1910 1911 Node* answer_if_true = NULL; 1912 Node* answer_if_false = NULL; 1913 switch (best_btest) { 1914 default: 1915 if (cmpxy == NULL) 1916 cmpxy = ideal_cmpxy; 1917 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); 1918 // and fall through: 1919 case BoolTest::lt: // x < y 1920 case BoolTest::le: // x <= y 1921 answer_if_true = (want_max ? yvalue : xvalue); 1922 answer_if_false = (want_max ? xvalue : yvalue); 1923 break; 1924 case BoolTest::gt: // x > y 1925 case BoolTest::ge: // x >= y 1926 answer_if_true = (want_max ? xvalue : yvalue); 1927 answer_if_false = (want_max ? yvalue : xvalue); 1928 break; 1929 } 1930 1931 jint hi, lo; 1932 if (want_max) { 1933 // We can sharpen the minimum. 1934 hi = MAX2(txvalue->_hi, tyvalue->_hi); 1935 lo = MAX2(txvalue->_lo, tyvalue->_lo); 1936 } else { 1937 // We can sharpen the maximum. 1938 hi = MIN2(txvalue->_hi, tyvalue->_hi); 1939 lo = MIN2(txvalue->_lo, tyvalue->_lo); 1940 } 1941 1942 // Use a flow-free graph structure, to avoid creating excess control edges 1943 // which could hinder other optimizations. 1944 // Since Math.min/max is often used with arraycopy, we want 1945 // tightly_coupled_allocation to be able to see beyond min/max expressions. 1946 Node* cmov = CMoveNode::make(C, NULL, best_bol, 1947 answer_if_false, answer_if_true, 1948 TypeInt::make(lo, hi, widen)); 1949 1950 return _gvn.transform(cmov); 1951 1952 /* 1953 // This is not as desirable as it may seem, since Min and Max 1954 // nodes do not have a full set of optimizations. 1955 // And they would interfere, anyway, with 'if' optimizations 1956 // and with CMoveI canonical forms. 1957 switch (id) { 1958 case vmIntrinsics::_min: 1959 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 1960 case vmIntrinsics::_max: 1961 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 1962 default: 1963 ShouldNotReachHere(); 1964 } 1965 */ 1966 } 1967 1968 inline int 1969 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 1970 const TypePtr* base_type = TypePtr::NULL_PTR; 1971 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 1972 if (base_type == NULL) { 1973 // Unknown type. 1974 return Type::AnyPtr; 1975 } else if (base_type == TypePtr::NULL_PTR) { 1976 // Since this is a NULL+long form, we have to switch to a rawptr. 1977 base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); 1978 offset = MakeConX(0); 1979 return Type::RawPtr; 1980 } else if (base_type->base() == Type::RawPtr) { 1981 return Type::RawPtr; 1982 } else if (base_type->isa_oopptr()) { 1983 // Base is never null => always a heap address. 1984 if (base_type->ptr() == TypePtr::NotNull) { 1985 return Type::OopPtr; 1986 } 1987 // Offset is small => always a heap address. 1988 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 1989 if (offset_type != NULL && 1990 base_type->offset() == 0 && // (should always be?) 1991 offset_type->_lo >= 0 && 1992 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 1993 return Type::OopPtr; 1994 } 1995 // Otherwise, it might either be oop+off or NULL+addr. 1996 return Type::AnyPtr; 1997 } else { 1998 // No information: 1999 return Type::AnyPtr; 2000 } 2001 } 2002 2003 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2004 int kind = classify_unsafe_addr(base, offset); 2005 if (kind == Type::RawPtr) { 2006 return basic_plus_adr(top(), base, offset); 2007 } else { 2008 return basic_plus_adr(base, offset); 2009 } 2010 } 2011 2012 //-------------------inline_numberOfLeadingZeros_int/long----------------------- 2013 // inline int Integer.numberOfLeadingZeros(int) 2014 // inline int Long.numberOfLeadingZeros(long) 2015 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) { 2016 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros"); 2017 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false; 2018 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false; 2019 _sp += arg_size(); // restore stack pointer 2020 switch (id) { 2021 case vmIntrinsics::_numberOfLeadingZeros_i: 2022 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop()))); 2023 break; 2024 case vmIntrinsics::_numberOfLeadingZeros_l: 2025 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair()))); 2026 break; 2027 default: 2028 ShouldNotReachHere(); 2029 } 2030 return true; 2031 } 2032 2033 //-------------------inline_numberOfTrailingZeros_int/long---------------------- 2034 // inline int Integer.numberOfTrailingZeros(int) 2035 // inline int Long.numberOfTrailingZeros(long) 2036 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) { 2037 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros"); 2038 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false; 2039 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false; 2040 _sp += arg_size(); // restore stack pointer 2041 switch (id) { 2042 case vmIntrinsics::_numberOfTrailingZeros_i: 2043 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop()))); 2044 break; 2045 case vmIntrinsics::_numberOfTrailingZeros_l: 2046 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair()))); 2047 break; 2048 default: 2049 ShouldNotReachHere(); 2050 } 2051 return true; 2052 } 2053 2054 //----------------------------inline_bitCount_int/long----------------------- 2055 // inline int Integer.bitCount(int) 2056 // inline int Long.bitCount(long) 2057 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) { 2058 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount"); 2059 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false; 2060 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false; 2061 _sp += arg_size(); // restore stack pointer 2062 switch (id) { 2063 case vmIntrinsics::_bitCount_i: 2064 push(_gvn.transform(new (C, 2) PopCountINode(pop()))); 2065 break; 2066 case vmIntrinsics::_bitCount_l: 2067 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair()))); 2068 break; 2069 default: 2070 ShouldNotReachHere(); 2071 } 2072 return true; 2073 } 2074 2075 //----------------------------inline_reverseBytes_int/long/char/short------------------- 2076 // inline Integer.reverseBytes(int) 2077 // inline Long.reverseBytes(long) 2078 // inline Character.reverseBytes(char) 2079 // inline Short.reverseBytes(short) 2080 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { 2081 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l || 2082 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s, 2083 "not reverse Bytes"); 2084 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; 2085 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; 2086 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false; 2087 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false; 2088 _sp += arg_size(); // restore stack pointer 2089 switch (id) { 2090 case vmIntrinsics::_reverseBytes_i: 2091 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); 2092 break; 2093 case vmIntrinsics::_reverseBytes_l: 2094 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); 2095 break; 2096 case vmIntrinsics::_reverseBytes_c: 2097 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop()))); 2098 break; 2099 case vmIntrinsics::_reverseBytes_s: 2100 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop()))); 2101 break; 2102 default: 2103 ; 2104 } 2105 return true; 2106 } 2107 2108 //----------------------------inline_unsafe_access---------------------------- 2109 2110 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2111 2112 // Helper that guards and inserts a G1 pre-barrier. 2113 void LibraryCallKit::insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val) { 2114 assert(UseG1GC, "should not call this otherwise"); 2115 2116 // We could be accessing the referent field of a reference object. If so, when G1 2117 // is enabled, we need to log the value in the referent field in an SATB buffer. 2118 // This routine performs some compile time filters and generates suitable 2119 // runtime filters that guard the pre-barrier code. 2120 2121 // Some compile time checks. 2122 2123 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2124 const TypeX* otype = offset->find_intptr_t_type(); 2125 if (otype != NULL && otype->is_con() && 2126 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2127 // Constant offset but not the reference_offset so just return 2128 return; 2129 } 2130 2131 // We only need to generate the runtime guards for instances. 2132 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2133 if (btype != NULL) { 2134 if (btype->isa_aryptr()) { 2135 // Array type so nothing to do 2136 return; 2137 } 2138 2139 const TypeInstPtr* itype = btype->isa_instptr(); 2140 if (itype != NULL) { 2141 // Can the klass of base_oop be statically determined 2142 // to be _not_ a sub-class of Reference? 2143 ciKlass* klass = itype->klass(); 2144 if (klass->is_subtype_of(env()->Reference_klass()) && 2145 !env()->Reference_klass()->is_subtype_of(klass)) { 2146 return; 2147 } 2148 } 2149 } 2150 2151 // The compile time filters did not reject base_oop/offset so 2152 // we need to generate the following runtime filters 2153 // 2154 // if (offset == java_lang_ref_Reference::_reference_offset) { 2155 // if (base != null) { 2156 // if (instance_of(base, java.lang.ref.Reference)) { 2157 // pre_barrier(_, pre_val, ...); 2158 // } 2159 // } 2160 // } 2161 2162 float likely = PROB_LIKELY(0.999); 2163 float unlikely = PROB_UNLIKELY(0.999); 2164 2165 IdealKit ideal(this); 2166 #define __ ideal. 2167 2168 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2169 2170 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2171 __ if_then(base_oop, BoolTest::ne, null(), likely); { 2172 2173 // Update graphKit memory and control from IdealKit. 2174 sync_kit(ideal); 2175 2176 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2177 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2178 2179 // Update IdealKit memory and control from graphKit. 2180 __ sync_kit(this); 2181 2182 Node* one = __ ConI(1); 2183 2184 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2185 2186 // Update graphKit from IdeakKit. 2187 sync_kit(ideal); 2188 2189 // Use the pre-barrier to record the value in the referent field 2190 pre_barrier(false /* do_load */, 2191 __ ctrl(), 2192 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2193 pre_val /* pre_val */, 2194 T_OBJECT); 2195 2196 // Update IdealKit from graphKit. 2197 __ sync_kit(this); 2198 2199 } __ end_if(); // _ref_type != ref_none 2200 } __ end_if(); // base != NULL 2201 } __ end_if(); // offset == referent_offset 2202 2203 // Final sync IdealKit and GraphKit. 2204 final_sync(ideal); 2205 #undef __ 2206 } 2207 2208 2209 // Interpret Unsafe.fieldOffset cookies correctly: 2210 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2211 2212 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2213 if (callee()->is_static()) return false; // caller must have the capability! 2214 2215 #ifndef PRODUCT 2216 { 2217 ResourceMark rm; 2218 // Check the signatures. 2219 ciSignature* sig = signature(); 2220 #ifdef ASSERT 2221 if (!is_store) { 2222 // Object getObject(Object base, int/long offset), etc. 2223 BasicType rtype = sig->return_type()->basic_type(); 2224 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2225 rtype = T_ADDRESS; // it is really a C void* 2226 assert(rtype == type, "getter must return the expected value"); 2227 if (!is_native_ptr) { 2228 assert(sig->count() == 2, "oop getter has 2 arguments"); 2229 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2230 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2231 } else { 2232 assert(sig->count() == 1, "native getter has 1 argument"); 2233 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2234 } 2235 } else { 2236 // void putObject(Object base, int/long offset, Object x), etc. 2237 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2238 if (!is_native_ptr) { 2239 assert(sig->count() == 3, "oop putter has 3 arguments"); 2240 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2241 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2242 } else { 2243 assert(sig->count() == 2, "native putter has 2 arguments"); 2244 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2245 } 2246 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2247 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2248 vtype = T_ADDRESS; // it is really a C void* 2249 assert(vtype == type, "putter must accept the expected value"); 2250 } 2251 #endif // ASSERT 2252 } 2253 #endif //PRODUCT 2254 2255 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2256 2257 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; 2258 2259 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words 2260 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); 2261 2262 debug_only(int saved_sp = _sp); 2263 _sp += nargs; 2264 2265 Node* val; 2266 debug_only(val = (Node*)(uintptr_t)-1); 2267 2268 2269 if (is_store) { 2270 // Get the value being stored. (Pop it first; it was pushed last.) 2271 switch (type) { 2272 case T_DOUBLE: 2273 case T_LONG: 2274 case T_ADDRESS: 2275 val = pop_pair(); 2276 break; 2277 default: 2278 val = pop(); 2279 } 2280 } 2281 2282 // Build address expression. See the code in inline_unsafe_prefetch. 2283 Node *adr; 2284 Node *heap_base_oop = top(); 2285 Node* offset = top(); 2286 2287 if (!is_native_ptr) { 2288 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2289 offset = pop_pair(); 2290 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2291 Node* base = pop(); 2292 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2293 // to be plain byte offsets, which are also the same as those accepted 2294 // by oopDesc::field_base. 2295 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2296 "fieldOffset must be byte-scaled"); 2297 // 32-bit machines ignore the high half! 2298 offset = ConvL2X(offset); 2299 adr = make_unsafe_address(base, offset); 2300 heap_base_oop = base; 2301 } else { 2302 Node* ptr = pop_pair(); 2303 // Adjust Java long to machine word: 2304 ptr = ConvL2X(ptr); 2305 adr = make_unsafe_address(NULL, ptr); 2306 } 2307 2308 // Pop receiver last: it was pushed first. 2309 Node *receiver = pop(); 2310 2311 assert(saved_sp == _sp, "must have correct argument count"); 2312 2313 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2314 2315 // First guess at the value type. 2316 const Type *value_type = Type::get_const_basic_type(type); 2317 2318 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2319 // there was not enough information to nail it down. 2320 Compile::AliasType* alias_type = C->alias_type(adr_type); 2321 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2322 2323 // We will need memory barriers unless we can determine a unique 2324 // alias category for this reference. (Note: If for some reason 2325 // the barriers get omitted and the unsafe reference begins to "pollute" 2326 // the alias analysis of the rest of the graph, either Compile::can_alias 2327 // or Compile::must_alias will throw a diagnostic assert.) 2328 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2329 2330 // If we are reading the value of the referent field of a Reference 2331 // object (either by using Unsafe directly or through reflection) 2332 // then, if G1 is enabled, we need to record the referent in an 2333 // SATB log buffer using the pre-barrier mechanism. 2334 bool need_read_barrier = UseG1GC && !is_native_ptr && !is_store && 2335 offset != top() && heap_base_oop != top(); 2336 2337 if (!is_store && type == T_OBJECT) { 2338 // Attempt to infer a sharper value type from the offset and base type. 2339 ciKlass* sharpened_klass = NULL; 2340 2341 // See if it is an instance field, with an object type. 2342 if (alias_type->field() != NULL) { 2343 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2344 if (alias_type->field()->type()->is_klass()) { 2345 sharpened_klass = alias_type->field()->type()->as_klass(); 2346 } 2347 } 2348 2349 // See if it is a narrow oop array. 2350 if (adr_type->isa_aryptr()) { 2351 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2352 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2353 if (elem_type != NULL) { 2354 sharpened_klass = elem_type->klass(); 2355 } 2356 } 2357 } 2358 2359 if (sharpened_klass != NULL) { 2360 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2361 2362 // Sharpen the value type. 2363 value_type = tjp; 2364 2365 #ifndef PRODUCT 2366 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2367 tty->print(" from base type: "); adr_type->dump(); 2368 tty->print(" sharpened value: "); value_type->dump(); 2369 } 2370 #endif 2371 } 2372 } 2373 2374 // Null check on self without removing any arguments. The argument 2375 // null check technically happens in the wrong place, which can lead to 2376 // invalid stack traces when the primitive is inlined into a method 2377 // which handles NullPointerExceptions. 2378 _sp += nargs; 2379 do_null_check(receiver, T_OBJECT); 2380 _sp -= nargs; 2381 if (stopped()) { 2382 return true; 2383 } 2384 // Heap pointers get a null-check from the interpreter, 2385 // as a courtesy. However, this is not guaranteed by Unsafe, 2386 // and it is not possible to fully distinguish unintended nulls 2387 // from intended ones in this API. 2388 2389 if (is_volatile) { 2390 // We need to emit leading and trailing CPU membars (see below) in 2391 // addition to memory membars when is_volatile. This is a little 2392 // too strong, but avoids the need to insert per-alias-type 2393 // volatile membars (for stores; compare Parse::do_put_xxx), which 2394 // we cannot do effectively here because we probably only have a 2395 // rough approximation of type. 2396 need_mem_bar = true; 2397 // For Stores, place a memory ordering barrier now. 2398 if (is_store) 2399 insert_mem_bar(Op_MemBarRelease); 2400 } 2401 2402 // Memory barrier to prevent normal and 'unsafe' accesses from 2403 // bypassing each other. Happens after null checks, so the 2404 // exception paths do not take memory state from the memory barrier, 2405 // so there's no problems making a strong assert about mixing users 2406 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2407 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2408 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2409 2410 if (!is_store) { 2411 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2412 // load value and push onto stack 2413 switch (type) { 2414 case T_BOOLEAN: 2415 case T_CHAR: 2416 case T_BYTE: 2417 case T_SHORT: 2418 case T_INT: 2419 case T_FLOAT: 2420 push(p); 2421 break; 2422 case T_OBJECT: 2423 if (need_read_barrier) { 2424 insert_g1_pre_barrier(heap_base_oop, offset, p); 2425 } 2426 push(p); 2427 break; 2428 case T_ADDRESS: 2429 // Cast to an int type. 2430 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); 2431 p = ConvX2L(p); 2432 push_pair(p); 2433 break; 2434 case T_DOUBLE: 2435 case T_LONG: 2436 push_pair( p ); 2437 break; 2438 default: ShouldNotReachHere(); 2439 } 2440 } else { 2441 // place effect of store into memory 2442 switch (type) { 2443 case T_DOUBLE: 2444 val = dstore_rounding(val); 2445 break; 2446 case T_ADDRESS: 2447 // Repackage the long as a pointer. 2448 val = ConvL2X(val); 2449 val = _gvn.transform( new (C, 2) CastX2PNode(val) ); 2450 break; 2451 } 2452 2453 if (type != T_OBJECT ) { 2454 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2455 } else { 2456 // Possibly an oop being stored to Java heap or native memory 2457 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2458 // oop to Java heap. 2459 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2460 } else { 2461 // We can't tell at compile time if we are storing in the Java heap or outside 2462 // of it. So we need to emit code to conditionally do the proper type of 2463 // store. 2464 2465 IdealKit ideal(this); 2466 #define __ ideal. 2467 // QQQ who knows what probability is here?? 2468 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2469 // Sync IdealKit and graphKit. 2470 sync_kit(ideal); 2471 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2472 // Update IdealKit memory. 2473 __ sync_kit(this); 2474 } __ else_(); { 2475 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile); 2476 } __ end_if(); 2477 // Final sync IdealKit and GraphKit. 2478 final_sync(ideal); 2479 #undef __ 2480 } 2481 } 2482 } 2483 2484 if (is_volatile) { 2485 if (!is_store) 2486 insert_mem_bar(Op_MemBarAcquire); 2487 else 2488 insert_mem_bar(Op_MemBarVolatile); 2489 } 2490 2491 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2492 2493 return true; 2494 } 2495 2496 //----------------------------inline_unsafe_prefetch---------------------------- 2497 2498 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2499 #ifndef PRODUCT 2500 { 2501 ResourceMark rm; 2502 // Check the signatures. 2503 ciSignature* sig = signature(); 2504 #ifdef ASSERT 2505 // Object getObject(Object base, int/long offset), etc. 2506 BasicType rtype = sig->return_type()->basic_type(); 2507 if (!is_native_ptr) { 2508 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2509 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2510 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2511 } else { 2512 assert(sig->count() == 1, "native prefetch has 1 argument"); 2513 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2514 } 2515 #endif // ASSERT 2516 } 2517 #endif // !PRODUCT 2518 2519 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2520 2521 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args 2522 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); 2523 2524 debug_only(int saved_sp = _sp); 2525 _sp += nargs; 2526 2527 // Build address expression. See the code in inline_unsafe_access. 2528 Node *adr; 2529 if (!is_native_ptr) { 2530 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2531 Node* offset = pop_pair(); 2532 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2533 Node* base = pop(); 2534 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2535 // to be plain byte offsets, which are also the same as those accepted 2536 // by oopDesc::field_base. 2537 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2538 "fieldOffset must be byte-scaled"); 2539 // 32-bit machines ignore the high half! 2540 offset = ConvL2X(offset); 2541 adr = make_unsafe_address(base, offset); 2542 } else { 2543 Node* ptr = pop_pair(); 2544 // Adjust Java long to machine word: 2545 ptr = ConvL2X(ptr); 2546 adr = make_unsafe_address(NULL, ptr); 2547 } 2548 2549 if (is_static) { 2550 assert(saved_sp == _sp, "must have correct argument count"); 2551 } else { 2552 // Pop receiver last: it was pushed first. 2553 Node *receiver = pop(); 2554 assert(saved_sp == _sp, "must have correct argument count"); 2555 2556 // Null check on self without removing any arguments. The argument 2557 // null check technically happens in the wrong place, which can lead to 2558 // invalid stack traces when the primitive is inlined into a method 2559 // which handles NullPointerExceptions. 2560 _sp += nargs; 2561 do_null_check(receiver, T_OBJECT); 2562 _sp -= nargs; 2563 if (stopped()) { 2564 return true; 2565 } 2566 } 2567 2568 // Generate the read or write prefetch 2569 Node *prefetch; 2570 if (is_store) { 2571 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); 2572 } else { 2573 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); 2574 } 2575 prefetch->init_req(0, control()); 2576 set_i_o(_gvn.transform(prefetch)); 2577 2578 return true; 2579 } 2580 2581 //----------------------------inline_unsafe_CAS---------------------------- 2582 2583 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { 2584 // This basic scheme here is the same as inline_unsafe_access, but 2585 // differs in enough details that combining them would make the code 2586 // overly confusing. (This is a true fact! I originally combined 2587 // them, but even I was confused by it!) As much code/comments as 2588 // possible are retained from inline_unsafe_access though to make 2589 // the correspondences clearer. - dl 2590 2591 if (callee()->is_static()) return false; // caller must have the capability! 2592 2593 #ifndef PRODUCT 2594 { 2595 ResourceMark rm; 2596 // Check the signatures. 2597 ciSignature* sig = signature(); 2598 #ifdef ASSERT 2599 BasicType rtype = sig->return_type()->basic_type(); 2600 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2601 assert(sig->count() == 4, "CAS has 4 arguments"); 2602 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2603 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2604 #endif // ASSERT 2605 } 2606 #endif //PRODUCT 2607 2608 // number of stack slots per value argument (1 or 2) 2609 int type_words = type2size[type]; 2610 2611 // Cannot inline wide CAS on machines that don't support it natively 2612 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8()) 2613 return false; 2614 2615 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2616 2617 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; 2618 int nargs = 1 + 1 + 2 + type_words + type_words; 2619 2620 // pop arguments: newval, oldval, offset, base, and receiver 2621 debug_only(int saved_sp = _sp); 2622 _sp += nargs; 2623 Node* newval = (type_words == 1) ? pop() : pop_pair(); 2624 Node* oldval = (type_words == 1) ? pop() : pop_pair(); 2625 Node *offset = pop_pair(); 2626 Node *base = pop(); 2627 Node *receiver = pop(); 2628 assert(saved_sp == _sp, "must have correct argument count"); 2629 2630 // Null check receiver. 2631 _sp += nargs; 2632 do_null_check(receiver, T_OBJECT); 2633 _sp -= nargs; 2634 if (stopped()) { 2635 return true; 2636 } 2637 2638 // Build field offset expression. 2639 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2640 // to be plain byte offsets, which are also the same as those accepted 2641 // by oopDesc::field_base. 2642 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2643 // 32-bit machines ignore the high half of long offsets 2644 offset = ConvL2X(offset); 2645 Node* adr = make_unsafe_address(base, offset); 2646 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2647 2648 // (Unlike inline_unsafe_access, there seems no point in trying 2649 // to refine types. Just use the coarse types here. 2650 const Type *value_type = Type::get_const_basic_type(type); 2651 Compile::AliasType* alias_type = C->alias_type(adr_type); 2652 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2653 int alias_idx = C->get_alias_index(adr_type); 2654 2655 // Memory-model-wise, a CAS acts like a little synchronized block, 2656 // so needs barriers on each side. These don't translate into 2657 // actual barriers on most machines, but we still need rest of 2658 // compiler to respect ordering. 2659 2660 insert_mem_bar(Op_MemBarRelease); 2661 insert_mem_bar(Op_MemBarCPUOrder); 2662 2663 // 4984716: MemBars must be inserted before this 2664 // memory node in order to avoid a false 2665 // dependency which will confuse the scheduler. 2666 Node *mem = memory(alias_idx); 2667 2668 // For now, we handle only those cases that actually exist: ints, 2669 // longs, and Object. Adding others should be straightforward. 2670 Node* cas; 2671 switch(type) { 2672 case T_INT: 2673 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2674 break; 2675 case T_LONG: 2676 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2677 break; 2678 case T_OBJECT: 2679 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2680 // could be delayed during Parse (for example, in adjust_map_after_if()). 2681 // Execute transformation here to avoid barrier generation in such case. 2682 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2683 newval = _gvn.makecon(TypePtr::NULL_PTR); 2684 2685 // Reference stores need a store barrier. 2686 // (They don't if CAS fails, but it isn't worth checking.) 2687 pre_barrier(true /* do_load*/, 2688 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2689 NULL /* pre_val*/, 2690 T_OBJECT); 2691 #ifdef _LP64 2692 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2693 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2694 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2695 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr, 2696 newval_enc, oldval_enc)); 2697 } else 2698 #endif 2699 { 2700 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2701 } 2702 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); 2703 break; 2704 default: 2705 ShouldNotReachHere(); 2706 break; 2707 } 2708 2709 // SCMemProjNodes represent the memory state of CAS. Their main 2710 // role is to prevent CAS nodes from being optimized away when their 2711 // results aren't used. 2712 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 2713 set_memory(proj, alias_idx); 2714 2715 // Add the trailing membar surrounding the access 2716 insert_mem_bar(Op_MemBarCPUOrder); 2717 insert_mem_bar(Op_MemBarAcquire); 2718 2719 push(cas); 2720 return true; 2721 } 2722 2723 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2724 // This is another variant of inline_unsafe_access, differing in 2725 // that it always issues store-store ("release") barrier and ensures 2726 // store-atomicity (which only matters for "long"). 2727 2728 if (callee()->is_static()) return false; // caller must have the capability! 2729 2730 #ifndef PRODUCT 2731 { 2732 ResourceMark rm; 2733 // Check the signatures. 2734 ciSignature* sig = signature(); 2735 #ifdef ASSERT 2736 BasicType rtype = sig->return_type()->basic_type(); 2737 assert(rtype == T_VOID, "must return void"); 2738 assert(sig->count() == 3, "has 3 arguments"); 2739 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2740 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2741 #endif // ASSERT 2742 } 2743 #endif //PRODUCT 2744 2745 // number of stack slots per value argument (1 or 2) 2746 int type_words = type2size[type]; 2747 2748 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2749 2750 // Argument words: "this" plus oop plus offset plus value; 2751 int nargs = 1 + 1 + 2 + type_words; 2752 2753 // pop arguments: val, offset, base, and receiver 2754 debug_only(int saved_sp = _sp); 2755 _sp += nargs; 2756 Node* val = (type_words == 1) ? pop() : pop_pair(); 2757 Node *offset = pop_pair(); 2758 Node *base = pop(); 2759 Node *receiver = pop(); 2760 assert(saved_sp == _sp, "must have correct argument count"); 2761 2762 // Null check receiver. 2763 _sp += nargs; 2764 do_null_check(receiver, T_OBJECT); 2765 _sp -= nargs; 2766 if (stopped()) { 2767 return true; 2768 } 2769 2770 // Build field offset expression. 2771 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2772 // 32-bit machines ignore the high half of long offsets 2773 offset = ConvL2X(offset); 2774 Node* adr = make_unsafe_address(base, offset); 2775 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2776 const Type *value_type = Type::get_const_basic_type(type); 2777 Compile::AliasType* alias_type = C->alias_type(adr_type); 2778 2779 insert_mem_bar(Op_MemBarRelease); 2780 insert_mem_bar(Op_MemBarCPUOrder); 2781 // Ensure that the store is atomic for longs: 2782 bool require_atomic_access = true; 2783 Node* store; 2784 if (type == T_OBJECT) // reference stores need a store barrier. 2785 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2786 else { 2787 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2788 } 2789 insert_mem_bar(Op_MemBarCPUOrder); 2790 return true; 2791 } 2792 2793 bool LibraryCallKit::inline_unsafe_allocate() { 2794 if (callee()->is_static()) return false; // caller must have the capability! 2795 int nargs = 1 + 1; 2796 assert(signature()->size() == nargs-1, "alloc has 1 argument"); 2797 null_check_receiver(callee()); // check then ignore argument(0) 2798 _sp += nargs; // set original stack for use by uncommon_trap 2799 Node* cls = do_null_check(argument(1), T_OBJECT); 2800 _sp -= nargs; 2801 if (stopped()) return true; 2802 2803 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 2804 _sp += nargs; // set original stack for use by uncommon_trap 2805 kls = do_null_check(kls, T_OBJECT); 2806 _sp -= nargs; 2807 if (stopped()) return true; // argument was like int.class 2808 2809 // Note: The argument might still be an illegal value like 2810 // Serializable.class or Object[].class. The runtime will handle it. 2811 // But we must make an explicit check for initialization. 2812 Node* insp = basic_plus_adr(kls, in_bytes(instanceKlass::init_state_offset())); 2813 // Use T_BOOLEAN for instanceKlass::_init_state so the compiler 2814 // can generate code to load it as unsigned byte. 2815 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN); 2816 Node* bits = intcon(instanceKlass::fully_initialized); 2817 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); 2818 // The 'test' is non-zero if we need to take a slow path. 2819 2820 Node* obj = new_instance(kls, test); 2821 push(obj); 2822 2823 return true; 2824 } 2825 2826 //------------------------inline_native_time_funcs-------------- 2827 // inline code for System.currentTimeMillis() and System.nanoTime() 2828 // these have the same type and signature 2829 bool LibraryCallKit::inline_native_time_funcs(bool isNano) { 2830 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) : 2831 CAST_FROM_FN_PTR(address, os::javaTimeMillis); 2832 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis"; 2833 const TypeFunc *tf = OptoRuntime::current_time_millis_Type(); 2834 const TypePtr* no_memory_effects = NULL; 2835 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2836 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); 2837 #ifdef ASSERT 2838 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); 2839 assert(value_top == top(), "second value must be top"); 2840 #endif 2841 push_pair(value); 2842 return true; 2843 } 2844 2845 //------------------------inline_native_currentThread------------------ 2846 bool LibraryCallKit::inline_native_currentThread() { 2847 Node* junk = NULL; 2848 push(generate_current_thread(junk)); 2849 return true; 2850 } 2851 2852 //------------------------inline_native_isInterrupted------------------ 2853 bool LibraryCallKit::inline_native_isInterrupted() { 2854 const int nargs = 1+1; // receiver + boolean 2855 assert(nargs == arg_size(), "sanity"); 2856 // Add a fast path to t.isInterrupted(clear_int): 2857 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 2858 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 2859 // So, in the common case that the interrupt bit is false, 2860 // we avoid making a call into the VM. Even if the interrupt bit 2861 // is true, if the clear_int argument is false, we avoid the VM call. 2862 // However, if the receiver is not currentThread, we must call the VM, 2863 // because there must be some locking done around the operation. 2864 2865 // We only go to the fast case code if we pass two guards. 2866 // Paths which do not pass are accumulated in the slow_region. 2867 RegionNode* slow_region = new (C, 1) RegionNode(1); 2868 record_for_igvn(slow_region); 2869 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow 2870 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); 2871 enum { no_int_result_path = 1, 2872 no_clear_result_path = 2, 2873 slow_result_path = 3 2874 }; 2875 2876 // (a) Receiving thread must be the current thread. 2877 Node* rec_thr = argument(0); 2878 Node* tls_ptr = NULL; 2879 Node* cur_thr = generate_current_thread(tls_ptr); 2880 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); 2881 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); 2882 2883 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); 2884 if (!known_current_thread) 2885 generate_slow_guard(bol_thr, slow_region); 2886 2887 // (b) Interrupt bit on TLS must be false. 2888 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2889 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 2890 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 2891 // Set the control input on the field _interrupted read to prevent it floating up. 2892 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 2893 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); 2894 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); 2895 2896 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2897 2898 // First fast path: if (!TLS._interrupted) return false; 2899 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); 2900 result_rgn->init_req(no_int_result_path, false_bit); 2901 result_val->init_req(no_int_result_path, intcon(0)); 2902 2903 // drop through to next case 2904 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); 2905 2906 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 2907 Node* clr_arg = argument(1); 2908 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); 2909 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); 2910 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 2911 2912 // Second fast path: ... else if (!clear_int) return true; 2913 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); 2914 result_rgn->init_req(no_clear_result_path, false_arg); 2915 result_val->init_req(no_clear_result_path, intcon(1)); 2916 2917 // drop through to next case 2918 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); 2919 2920 // (d) Otherwise, go to the slow path. 2921 slow_region->add_req(control()); 2922 set_control( _gvn.transform(slow_region) ); 2923 2924 if (stopped()) { 2925 // There is no slow path. 2926 result_rgn->init_req(slow_result_path, top()); 2927 result_val->init_req(slow_result_path, top()); 2928 } else { 2929 // non-virtual because it is a private non-static 2930 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 2931 2932 Node* slow_val = set_results_for_java_call(slow_call); 2933 // this->control() comes from set_results_for_java_call 2934 2935 // If we know that the result of the slow call will be true, tell the optimizer! 2936 if (known_current_thread) slow_val = intcon(1); 2937 2938 Node* fast_io = slow_call->in(TypeFunc::I_O); 2939 Node* fast_mem = slow_call->in(TypeFunc::Memory); 2940 // These two phis are pre-filled with copies of of the fast IO and Memory 2941 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 2942 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 2943 2944 result_rgn->init_req(slow_result_path, control()); 2945 io_phi ->init_req(slow_result_path, i_o()); 2946 mem_phi ->init_req(slow_result_path, reset_memory()); 2947 result_val->init_req(slow_result_path, slow_val); 2948 2949 set_all_memory( _gvn.transform(mem_phi) ); 2950 set_i_o( _gvn.transform(io_phi) ); 2951 } 2952 2953 push_result(result_rgn, result_val); 2954 C->set_has_split_ifs(true); // Has chance for split-if optimization 2955 2956 return true; 2957 } 2958 2959 //---------------------------load_mirror_from_klass---------------------------- 2960 // Given a klass oop, load its java mirror (a java.lang.Class oop). 2961 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 2962 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 2963 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 2964 } 2965 2966 //-----------------------load_klass_from_mirror_common------------------------- 2967 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 2968 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 2969 // and branch to the given path on the region. 2970 // If never_see_null, take an uncommon trap on null, so we can optimistically 2971 // compile for the non-null case. 2972 // If the region is NULL, force never_see_null = true. 2973 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 2974 bool never_see_null, 2975 int nargs, 2976 RegionNode* region, 2977 int null_path, 2978 int offset) { 2979 if (region == NULL) never_see_null = true; 2980 Node* p = basic_plus_adr(mirror, offset); 2981 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 2982 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); 2983 _sp += nargs; // any deopt will start just before call to enclosing method 2984 Node* null_ctl = top(); 2985 kls = null_check_oop(kls, &null_ctl, never_see_null); 2986 if (region != NULL) { 2987 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 2988 region->init_req(null_path, null_ctl); 2989 } else { 2990 assert(null_ctl == top(), "no loose ends"); 2991 } 2992 _sp -= nargs; 2993 return kls; 2994 } 2995 2996 //--------------------(inline_native_Class_query helpers)--------------------- 2997 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 2998 // Fall through if (mods & mask) == bits, take the guard otherwise. 2999 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3000 // Branch around if the given klass has the given modifier bit set. 3001 // Like generate_guard, adds a new path onto the region. 3002 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3003 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 3004 Node* mask = intcon(modifier_mask); 3005 Node* bits = intcon(modifier_bits); 3006 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); 3007 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); 3008 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); 3009 return generate_fair_guard(bol, region); 3010 } 3011 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3012 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3013 } 3014 3015 //-------------------------inline_native_Class_query------------------- 3016 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3017 int nargs = 1+0; // just the Class mirror, in most cases 3018 const Type* return_type = TypeInt::BOOL; 3019 Node* prim_return_value = top(); // what happens if it's a primitive class? 3020 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3021 bool expect_prim = false; // most of these guys expect to work on refs 3022 3023 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3024 3025 switch (id) { 3026 case vmIntrinsics::_isInstance: 3027 nargs = 1+1; // the Class mirror, plus the object getting queried about 3028 // nothing is an instance of a primitive type 3029 prim_return_value = intcon(0); 3030 break; 3031 case vmIntrinsics::_getModifiers: 3032 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3033 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3034 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3035 break; 3036 case vmIntrinsics::_isInterface: 3037 prim_return_value = intcon(0); 3038 break; 3039 case vmIntrinsics::_isArray: 3040 prim_return_value = intcon(0); 3041 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3042 break; 3043 case vmIntrinsics::_isPrimitive: 3044 prim_return_value = intcon(1); 3045 expect_prim = true; // obviously 3046 break; 3047 case vmIntrinsics::_getSuperclass: 3048 prim_return_value = null(); 3049 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3050 break; 3051 case vmIntrinsics::_getComponentType: 3052 prim_return_value = null(); 3053 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3054 break; 3055 case vmIntrinsics::_getClassAccessFlags: 3056 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3057 return_type = TypeInt::INT; // not bool! 6297094 3058 break; 3059 default: 3060 ShouldNotReachHere(); 3061 } 3062 3063 Node* mirror = argument(0); 3064 Node* obj = (nargs <= 1)? top(): argument(1); 3065 3066 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3067 if (mirror_con == NULL) return false; // cannot happen? 3068 3069 #ifndef PRODUCT 3070 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 3071 ciType* k = mirror_con->java_mirror_type(); 3072 if (k) { 3073 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3074 k->print_name(); 3075 tty->cr(); 3076 } 3077 } 3078 #endif 3079 3080 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3081 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3082 record_for_igvn(region); 3083 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); 3084 3085 // The mirror will never be null of Reflection.getClassAccessFlags, however 3086 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3087 // if it is. See bug 4774291. 3088 3089 // For Reflection.getClassAccessFlags(), the null check occurs in 3090 // the wrong place; see inline_unsafe_access(), above, for a similar 3091 // situation. 3092 _sp += nargs; // set original stack for use by uncommon_trap 3093 mirror = do_null_check(mirror, T_OBJECT); 3094 _sp -= nargs; 3095 // If mirror or obj is dead, only null-path is taken. 3096 if (stopped()) return true; 3097 3098 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3099 3100 // Now load the mirror's klass metaobject, and null-check it. 3101 // Side-effects region with the control path if the klass is null. 3102 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, 3103 region, _prim_path); 3104 // If kls is null, we have a primitive mirror. 3105 phi->init_req(_prim_path, prim_return_value); 3106 if (stopped()) { push_result(region, phi); return true; } 3107 3108 Node* p; // handy temp 3109 Node* null_ctl; 3110 3111 // Now that we have the non-null klass, we can perform the real query. 3112 // For constant classes, the query will constant-fold in LoadNode::Value. 3113 Node* query_value = top(); 3114 switch (id) { 3115 case vmIntrinsics::_isInstance: 3116 // nothing is an instance of a primitive type 3117 _sp += nargs; // gen_instanceof might do an uncommon trap 3118 query_value = gen_instanceof(obj, kls); 3119 _sp -= nargs; 3120 break; 3121 3122 case vmIntrinsics::_getModifiers: 3123 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3124 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3125 break; 3126 3127 case vmIntrinsics::_isInterface: 3128 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3129 if (generate_interface_guard(kls, region) != NULL) 3130 // A guard was added. If the guard is taken, it was an interface. 3131 phi->add_req(intcon(1)); 3132 // If we fall through, it's a plain class. 3133 query_value = intcon(0); 3134 break; 3135 3136 case vmIntrinsics::_isArray: 3137 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3138 if (generate_array_guard(kls, region) != NULL) 3139 // A guard was added. If the guard is taken, it was an array. 3140 phi->add_req(intcon(1)); 3141 // If we fall through, it's a plain class. 3142 query_value = intcon(0); 3143 break; 3144 3145 case vmIntrinsics::_isPrimitive: 3146 query_value = intcon(0); // "normal" path produces false 3147 break; 3148 3149 case vmIntrinsics::_getSuperclass: 3150 // The rules here are somewhat unfortunate, but we can still do better 3151 // with random logic than with a JNI call. 3152 // Interfaces store null or Object as _super, but must report null. 3153 // Arrays store an intermediate super as _super, but must report Object. 3154 // Other types can report the actual _super. 3155 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3156 if (generate_interface_guard(kls, region) != NULL) 3157 // A guard was added. If the guard is taken, it was an interface. 3158 phi->add_req(null()); 3159 if (generate_array_guard(kls, region) != NULL) 3160 // A guard was added. If the guard is taken, it was an array. 3161 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3162 // If we fall through, it's a plain class. Get its _super. 3163 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3164 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); 3165 null_ctl = top(); 3166 kls = null_check_oop(kls, &null_ctl); 3167 if (null_ctl != top()) { 3168 // If the guard is taken, Object.superClass is null (both klass and mirror). 3169 region->add_req(null_ctl); 3170 phi ->add_req(null()); 3171 } 3172 if (!stopped()) { 3173 query_value = load_mirror_from_klass(kls); 3174 } 3175 break; 3176 3177 case vmIntrinsics::_getComponentType: 3178 if (generate_array_guard(kls, region) != NULL) { 3179 // Be sure to pin the oop load to the guard edge just created: 3180 Node* is_array_ctrl = region->in(region->req()-1); 3181 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset())); 3182 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 3183 phi->add_req(cmo); 3184 } 3185 query_value = null(); // non-array case is null 3186 break; 3187 3188 case vmIntrinsics::_getClassAccessFlags: 3189 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3190 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3191 break; 3192 3193 default: 3194 ShouldNotReachHere(); 3195 } 3196 3197 // Fall-through is the normal case of a query to a real class. 3198 phi->init_req(1, query_value); 3199 region->init_req(1, control()); 3200 3201 push_result(region, phi); 3202 C->set_has_split_ifs(true); // Has chance for split-if optimization 3203 3204 return true; 3205 } 3206 3207 //--------------------------inline_native_subtype_check------------------------ 3208 // This intrinsic takes the JNI calls out of the heart of 3209 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3210 bool LibraryCallKit::inline_native_subtype_check() { 3211 int nargs = 1+1; // the Class mirror, plus the other class getting examined 3212 3213 // Pull both arguments off the stack. 3214 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3215 args[0] = argument(0); 3216 args[1] = argument(1); 3217 Node* klasses[2]; // corresponding Klasses: superk, subk 3218 klasses[0] = klasses[1] = top(); 3219 3220 enum { 3221 // A full decision tree on {superc is prim, subc is prim}: 3222 _prim_0_path = 1, // {P,N} => false 3223 // {P,P} & superc!=subc => false 3224 _prim_same_path, // {P,P} & superc==subc => true 3225 _prim_1_path, // {N,P} => false 3226 _ref_subtype_path, // {N,N} & subtype check wins => true 3227 _both_ref_path, // {N,N} & subtype check loses => false 3228 PATH_LIMIT 3229 }; 3230 3231 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3232 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 3233 record_for_igvn(region); 3234 3235 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3236 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3237 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3238 3239 // First null-check both mirrors and load each mirror's klass metaobject. 3240 int which_arg; 3241 for (which_arg = 0; which_arg <= 1; which_arg++) { 3242 Node* arg = args[which_arg]; 3243 _sp += nargs; // set original stack for use by uncommon_trap 3244 arg = do_null_check(arg, T_OBJECT); 3245 _sp -= nargs; 3246 if (stopped()) break; 3247 args[which_arg] = _gvn.transform(arg); 3248 3249 Node* p = basic_plus_adr(arg, class_klass_offset); 3250 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3251 klasses[which_arg] = _gvn.transform(kls); 3252 } 3253 3254 // Having loaded both klasses, test each for null. 3255 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3256 for (which_arg = 0; which_arg <= 1; which_arg++) { 3257 Node* kls = klasses[which_arg]; 3258 Node* null_ctl = top(); 3259 _sp += nargs; // set original stack for use by uncommon_trap 3260 kls = null_check_oop(kls, &null_ctl, never_see_null); 3261 _sp -= nargs; 3262 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3263 region->init_req(prim_path, null_ctl); 3264 if (stopped()) break; 3265 klasses[which_arg] = kls; 3266 } 3267 3268 if (!stopped()) { 3269 // now we have two reference types, in klasses[0..1] 3270 Node* subk = klasses[1]; // the argument to isAssignableFrom 3271 Node* superk = klasses[0]; // the receiver 3272 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3273 // now we have a successful reference subtype check 3274 region->set_req(_ref_subtype_path, control()); 3275 } 3276 3277 // If both operands are primitive (both klasses null), then 3278 // we must return true when they are identical primitives. 3279 // It is convenient to test this after the first null klass check. 3280 set_control(region->in(_prim_0_path)); // go back to first null check 3281 if (!stopped()) { 3282 // Since superc is primitive, make a guard for the superc==subc case. 3283 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); 3284 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); 3285 generate_guard(bol_eq, region, PROB_FAIR); 3286 if (region->req() == PATH_LIMIT+1) { 3287 // A guard was added. If the added guard is taken, superc==subc. 3288 region->swap_edges(PATH_LIMIT, _prim_same_path); 3289 region->del_req(PATH_LIMIT); 3290 } 3291 region->set_req(_prim_0_path, control()); // Not equal after all. 3292 } 3293 3294 // these are the only paths that produce 'true': 3295 phi->set_req(_prim_same_path, intcon(1)); 3296 phi->set_req(_ref_subtype_path, intcon(1)); 3297 3298 // pull together the cases: 3299 assert(region->req() == PATH_LIMIT, "sane region"); 3300 for (uint i = 1; i < region->req(); i++) { 3301 Node* ctl = region->in(i); 3302 if (ctl == NULL || ctl == top()) { 3303 region->set_req(i, top()); 3304 phi ->set_req(i, top()); 3305 } else if (phi->in(i) == NULL) { 3306 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3307 } 3308 } 3309 3310 set_control(_gvn.transform(region)); 3311 push(_gvn.transform(phi)); 3312 3313 return true; 3314 } 3315 3316 //---------------------generate_array_guard_common------------------------ 3317 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3318 bool obj_array, bool not_array) { 3319 // If obj_array/non_array==false/false: 3320 // Branch around if the given klass is in fact an array (either obj or prim). 3321 // If obj_array/non_array==false/true: 3322 // Branch around if the given klass is not an array klass of any kind. 3323 // If obj_array/non_array==true/true: 3324 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3325 // If obj_array/non_array==true/false: 3326 // Branch around if the kls is an oop array (Object[] or subtype) 3327 // 3328 // Like generate_guard, adds a new path onto the region. 3329 jint layout_con = 0; 3330 Node* layout_val = get_layout_helper(kls, layout_con); 3331 if (layout_val == NULL) { 3332 bool query = (obj_array 3333 ? Klass::layout_helper_is_objArray(layout_con) 3334 : Klass::layout_helper_is_javaArray(layout_con)); 3335 if (query == not_array) { 3336 return NULL; // never a branch 3337 } else { // always a branch 3338 Node* always_branch = control(); 3339 if (region != NULL) 3340 region->add_req(always_branch); 3341 set_control(top()); 3342 return always_branch; 3343 } 3344 } 3345 // Now test the correct condition. 3346 jint nval = (obj_array 3347 ? ((jint)Klass::_lh_array_tag_type_value 3348 << Klass::_lh_array_tag_shift) 3349 : Klass::_lh_neutral_value); 3350 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); 3351 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3352 // invert the test if we are looking for a non-array 3353 if (not_array) btest = BoolTest(btest).negate(); 3354 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); 3355 return generate_fair_guard(bol, region); 3356 } 3357 3358 3359 //-----------------------inline_native_newArray-------------------------- 3360 bool LibraryCallKit::inline_native_newArray() { 3361 int nargs = 2; 3362 Node* mirror = argument(0); 3363 Node* count_val = argument(1); 3364 3365 _sp += nargs; // set original stack for use by uncommon_trap 3366 mirror = do_null_check(mirror, T_OBJECT); 3367 _sp -= nargs; 3368 // If mirror or obj is dead, only null-path is taken. 3369 if (stopped()) return true; 3370 3371 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3372 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3373 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3374 TypeInstPtr::NOTNULL); 3375 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3376 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3377 TypePtr::BOTTOM); 3378 3379 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3380 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3381 nargs, 3382 result_reg, _slow_path); 3383 Node* normal_ctl = control(); 3384 Node* no_array_ctl = result_reg->in(_slow_path); 3385 3386 // Generate code for the slow case. We make a call to newArray(). 3387 set_control(no_array_ctl); 3388 if (!stopped()) { 3389 // Either the input type is void.class, or else the 3390 // array klass has not yet been cached. Either the 3391 // ensuing call will throw an exception, or else it 3392 // will cache the array klass for next time. 3393 PreserveJVMState pjvms(this); 3394 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3395 Node* slow_result = set_results_for_java_call(slow_call); 3396 // this->control() comes from set_results_for_java_call 3397 result_reg->set_req(_slow_path, control()); 3398 result_val->set_req(_slow_path, slow_result); 3399 result_io ->set_req(_slow_path, i_o()); 3400 result_mem->set_req(_slow_path, reset_memory()); 3401 } 3402 3403 set_control(normal_ctl); 3404 if (!stopped()) { 3405 // Normal case: The array type has been cached in the java.lang.Class. 3406 // The following call works fine even if the array type is polymorphic. 3407 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3408 Node* obj = new_array(klass_node, count_val, nargs); 3409 result_reg->init_req(_normal_path, control()); 3410 result_val->init_req(_normal_path, obj); 3411 result_io ->init_req(_normal_path, i_o()); 3412 result_mem->init_req(_normal_path, reset_memory()); 3413 } 3414 3415 // Return the combined state. 3416 set_i_o( _gvn.transform(result_io) ); 3417 set_all_memory( _gvn.transform(result_mem) ); 3418 push_result(result_reg, result_val); 3419 C->set_has_split_ifs(true); // Has chance for split-if optimization 3420 3421 return true; 3422 } 3423 3424 //----------------------inline_native_getLength-------------------------- 3425 bool LibraryCallKit::inline_native_getLength() { 3426 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3427 3428 int nargs = 1; 3429 Node* array = argument(0); 3430 3431 _sp += nargs; // set original stack for use by uncommon_trap 3432 array = do_null_check(array, T_OBJECT); 3433 _sp -= nargs; 3434 3435 // If array is dead, only null-path is taken. 3436 if (stopped()) return true; 3437 3438 // Deoptimize if it is a non-array. 3439 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3440 3441 if (non_array != NULL) { 3442 PreserveJVMState pjvms(this); 3443 set_control(non_array); 3444 _sp += nargs; // push the arguments back on the stack 3445 uncommon_trap(Deoptimization::Reason_intrinsic, 3446 Deoptimization::Action_maybe_recompile); 3447 } 3448 3449 // If control is dead, only non-array-path is taken. 3450 if (stopped()) return true; 3451 3452 // The works fine even if the array type is polymorphic. 3453 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3454 push( load_array_length(array) ); 3455 3456 C->set_has_split_ifs(true); // Has chance for split-if optimization 3457 3458 return true; 3459 } 3460 3461 //------------------------inline_array_copyOf---------------------------- 3462 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3463 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3464 3465 // Restore the stack and pop off the arguments. 3466 int nargs = 3 + (is_copyOfRange? 1: 0); 3467 Node* original = argument(0); 3468 Node* start = is_copyOfRange? argument(1): intcon(0); 3469 Node* end = is_copyOfRange? argument(2): argument(1); 3470 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3471 3472 Node* newcopy; 3473 3474 //set the original stack and the reexecute bit for the interpreter to reexecute 3475 //the bytecode that invokes Arrays.copyOf if deoptimization happens 3476 { PreserveReexecuteState preexecs(this); 3477 _sp += nargs; 3478 jvms()->set_should_reexecute(true); 3479 3480 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); 3481 original = do_null_check(original, T_OBJECT); 3482 3483 // Check if a null path was taken unconditionally. 3484 if (stopped()) return true; 3485 3486 Node* orig_length = load_array_length(original); 3487 3488 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0, 3489 NULL, 0); 3490 klass_node = do_null_check(klass_node, T_OBJECT); 3491 3492 RegionNode* bailout = new (C, 1) RegionNode(1); 3493 record_for_igvn(bailout); 3494 3495 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3496 // Bail out if that is so. 3497 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3498 if (not_objArray != NULL) { 3499 // Improve the klass node's type from the new optimistic assumption: 3500 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3501 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3502 Node* cast = new (C, 2) CastPPNode(klass_node, akls); 3503 cast->init_req(0, control()); 3504 klass_node = _gvn.transform(cast); 3505 } 3506 3507 // Bail out if either start or end is negative. 3508 generate_negative_guard(start, bailout, &start); 3509 generate_negative_guard(end, bailout, &end); 3510 3511 Node* length = end; 3512 if (_gvn.type(start) != TypeInt::ZERO) { 3513 length = _gvn.transform( new (C, 3) SubINode(end, start) ); 3514 } 3515 3516 // Bail out if length is negative. 3517 // ...Not needed, since the new_array will throw the right exception. 3518 //generate_negative_guard(length, bailout, &length); 3519 3520 if (bailout->req() > 1) { 3521 PreserveJVMState pjvms(this); 3522 set_control( _gvn.transform(bailout) ); 3523 uncommon_trap(Deoptimization::Reason_intrinsic, 3524 Deoptimization::Action_maybe_recompile); 3525 } 3526 3527 if (!stopped()) { 3528 3529 // How many elements will we copy from the original? 3530 // The answer is MinI(orig_length - start, length). 3531 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); 3532 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3533 3534 newcopy = new_array(klass_node, length, 0); 3535 3536 // Generate a direct call to the right arraycopy function(s). 3537 // We know the copy is disjoint but we might not know if the 3538 // oop stores need checking. 3539 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3540 // This will fail a store-check if x contains any non-nulls. 3541 bool disjoint_bases = true; 3542 bool length_never_negative = true; 3543 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3544 original, start, newcopy, intcon(0), moved, 3545 disjoint_bases, length_never_negative); 3546 } 3547 } //original reexecute and sp are set back here 3548 3549 if(!stopped()) { 3550 push(newcopy); 3551 } 3552 3553 C->set_has_split_ifs(true); // Has chance for split-if optimization 3554 3555 return true; 3556 } 3557 3558 3559 //----------------------generate_virtual_guard--------------------------- 3560 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3561 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3562 RegionNode* slow_region) { 3563 ciMethod* method = callee(); 3564 int vtable_index = method->vtable_index(); 3565 // Get the methodOop out of the appropriate vtable entry. 3566 int entry_offset = (instanceKlass::vtable_start_offset() + 3567 vtable_index*vtableEntry::size()) * wordSize + 3568 vtableEntry::method_offset_in_bytes(); 3569 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3570 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); 3571 3572 // Compare the target method with the expected method (e.g., Object.hashCode). 3573 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method); 3574 3575 Node* native_call = makecon(native_call_addr); 3576 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); 3577 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); 3578 3579 return generate_slow_guard(test_native, slow_region); 3580 } 3581 3582 //-----------------------generate_method_call---------------------------- 3583 // Use generate_method_call to make a slow-call to the real 3584 // method if the fast path fails. An alternative would be to 3585 // use a stub like OptoRuntime::slow_arraycopy_Java. 3586 // This only works for expanding the current library call, 3587 // not another intrinsic. (E.g., don't use this for making an 3588 // arraycopy call inside of the copyOf intrinsic.) 3589 CallJavaNode* 3590 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3591 // When compiling the intrinsic method itself, do not use this technique. 3592 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3593 3594 ciMethod* method = callee(); 3595 // ensure the JVMS we have will be correct for this call 3596 guarantee(method_id == method->intrinsic_id(), "must match"); 3597 3598 const TypeFunc* tf = TypeFunc::make(method); 3599 int tfdc = tf->domain()->cnt(); 3600 CallJavaNode* slow_call; 3601 if (is_static) { 3602 assert(!is_virtual, ""); 3603 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3604 SharedRuntime::get_resolve_static_call_stub(), 3605 method, bci()); 3606 } else if (is_virtual) { 3607 null_check_receiver(method); 3608 int vtable_index = methodOopDesc::invalid_vtable_index; 3609 if (UseInlineCaches) { 3610 // Suppress the vtable call 3611 } else { 3612 // hashCode and clone are not a miranda methods, 3613 // so the vtable index is fixed. 3614 // No need to use the linkResolver to get it. 3615 vtable_index = method->vtable_index(); 3616 } 3617 slow_call = new(C, tfdc) CallDynamicJavaNode(tf, 3618 SharedRuntime::get_resolve_virtual_call_stub(), 3619 method, vtable_index, bci()); 3620 } else { // neither virtual nor static: opt_virtual 3621 null_check_receiver(method); 3622 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3623 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3624 method, bci()); 3625 slow_call->set_optimized_virtual(true); 3626 } 3627 set_arguments_for_java_call(slow_call); 3628 set_edges_for_java_call(slow_call); 3629 return slow_call; 3630 } 3631 3632 3633 //------------------------------inline_native_hashcode-------------------- 3634 // Build special case code for calls to hashCode on an object. 3635 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3636 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3637 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3638 3639 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3640 3641 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3642 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3643 TypeInt::INT); 3644 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3645 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3646 TypePtr::BOTTOM); 3647 Node* obj = NULL; 3648 if (!is_static) { 3649 // Check for hashing null object 3650 obj = null_check_receiver(callee()); 3651 if (stopped()) return true; // unconditionally null 3652 result_reg->init_req(_null_path, top()); 3653 result_val->init_req(_null_path, top()); 3654 } else { 3655 // Do a null check, and return zero if null. 3656 // System.identityHashCode(null) == 0 3657 obj = argument(0); 3658 Node* null_ctl = top(); 3659 obj = null_check_oop(obj, &null_ctl); 3660 result_reg->init_req(_null_path, null_ctl); 3661 result_val->init_req(_null_path, _gvn.intcon(0)); 3662 } 3663 3664 // Unconditionally null? Then return right away. 3665 if (stopped()) { 3666 set_control( result_reg->in(_null_path) ); 3667 if (!stopped()) 3668 push( result_val ->in(_null_path) ); 3669 return true; 3670 } 3671 3672 // After null check, get the object's klass. 3673 Node* obj_klass = load_object_klass(obj); 3674 3675 // This call may be virtual (invokevirtual) or bound (invokespecial). 3676 // For each case we generate slightly different code. 3677 3678 // We only go to the fast case code if we pass a number of guards. The 3679 // paths which do not pass are accumulated in the slow_region. 3680 RegionNode* slow_region = new (C, 1) RegionNode(1); 3681 record_for_igvn(slow_region); 3682 3683 // If this is a virtual call, we generate a funny guard. We pull out 3684 // the vtable entry corresponding to hashCode() from the target object. 3685 // If the target method which we are calling happens to be the native 3686 // Object hashCode() method, we pass the guard. We do not need this 3687 // guard for non-virtual calls -- the caller is known to be the native 3688 // Object hashCode(). 3689 if (is_virtual) { 3690 generate_virtual_guard(obj_klass, slow_region); 3691 } 3692 3693 // Get the header out of the object, use LoadMarkNode when available 3694 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3695 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type()); 3696 3697 // Test the header to see if it is unlocked. 3698 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3699 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); 3700 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3701 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); 3702 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); 3703 3704 generate_slow_guard(test_unlocked, slow_region); 3705 3706 // Get the hash value and check to see that it has been properly assigned. 3707 // We depend on hash_mask being at most 32 bits and avoid the use of 3708 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3709 // vm: see markOop.hpp. 3710 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3711 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3712 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); 3713 // This hack lets the hash bits live anywhere in the mark object now, as long 3714 // as the shift drops the relevant bits into the low 32 bits. Note that 3715 // Java spec says that HashCode is an int so there's no point in capturing 3716 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3717 hshifted_header = ConvX2I(hshifted_header); 3718 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); 3719 3720 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3721 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); 3722 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); 3723 3724 generate_slow_guard(test_assigned, slow_region); 3725 3726 Node* init_mem = reset_memory(); 3727 // fill in the rest of the null path: 3728 result_io ->init_req(_null_path, i_o()); 3729 result_mem->init_req(_null_path, init_mem); 3730 3731 result_val->init_req(_fast_path, hash_val); 3732 result_reg->init_req(_fast_path, control()); 3733 result_io ->init_req(_fast_path, i_o()); 3734 result_mem->init_req(_fast_path, init_mem); 3735 3736 // Generate code for the slow case. We make a call to hashCode(). 3737 set_control(_gvn.transform(slow_region)); 3738 if (!stopped()) { 3739 // No need for PreserveJVMState, because we're using up the present state. 3740 set_all_memory(init_mem); 3741 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; 3742 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; 3743 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3744 Node* slow_result = set_results_for_java_call(slow_call); 3745 // this->control() comes from set_results_for_java_call 3746 result_reg->init_req(_slow_path, control()); 3747 result_val->init_req(_slow_path, slow_result); 3748 result_io ->set_req(_slow_path, i_o()); 3749 result_mem ->set_req(_slow_path, reset_memory()); 3750 } 3751 3752 // Return the combined state. 3753 set_i_o( _gvn.transform(result_io) ); 3754 set_all_memory( _gvn.transform(result_mem) ); 3755 push_result(result_reg, result_val); 3756 3757 return true; 3758 } 3759 3760 //---------------------------inline_native_getClass---------------------------- 3761 // Build special case code for calls to getClass on an object. 3762 bool LibraryCallKit::inline_native_getClass() { 3763 Node* obj = null_check_receiver(callee()); 3764 if (stopped()) return true; 3765 push( load_mirror_from_klass(load_object_klass(obj)) ); 3766 return true; 3767 } 3768 3769 //-----------------inline_native_Reflection_getCallerClass--------------------- 3770 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3771 // 3772 // NOTE that this code must perform the same logic as 3773 // vframeStream::security_get_caller_frame in that it must skip 3774 // Method.invoke() and auxiliary frames. 3775 3776 3777 3778 3779 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3780 ciMethod* method = callee(); 3781 3782 #ifndef PRODUCT 3783 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3784 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3785 } 3786 #endif 3787 3788 debug_only(int saved_sp = _sp); 3789 3790 // Argument words: (int depth) 3791 int nargs = 1; 3792 3793 _sp += nargs; 3794 Node* caller_depth_node = pop(); 3795 3796 assert(saved_sp == _sp, "must have correct argument count"); 3797 3798 // The depth value must be a constant in order for the runtime call 3799 // to be eliminated. 3800 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); 3801 if (caller_depth_type == NULL || !caller_depth_type->is_con()) { 3802 #ifndef PRODUCT 3803 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3804 tty->print_cr(" Bailing out because caller depth was not a constant"); 3805 } 3806 #endif 3807 return false; 3808 } 3809 // Note that the JVM state at this point does not include the 3810 // getCallerClass() frame which we are trying to inline. The 3811 // semantics of getCallerClass(), however, are that the "first" 3812 // frame is the getCallerClass() frame, so we subtract one from the 3813 // requested depth before continuing. We don't inline requests of 3814 // getCallerClass(0). 3815 int caller_depth = caller_depth_type->get_con() - 1; 3816 if (caller_depth < 0) { 3817 #ifndef PRODUCT 3818 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3819 tty->print_cr(" Bailing out because caller depth was %d", caller_depth); 3820 } 3821 #endif 3822 return false; 3823 } 3824 3825 if (!jvms()->has_method()) { 3826 #ifndef PRODUCT 3827 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3828 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3829 } 3830 #endif 3831 return false; 3832 } 3833 int _depth = jvms()->depth(); // cache call chain depth 3834 3835 // Walk back up the JVM state to find the caller at the required 3836 // depth. NOTE that this code must perform the same logic as 3837 // vframeStream::security_get_caller_frame in that it must skip 3838 // Method.invoke() and auxiliary frames. Note also that depth is 3839 // 1-based (1 is the bottom of the inlining). 3840 int inlining_depth = _depth; 3841 JVMState* caller_jvms = NULL; 3842 3843 if (inlining_depth > 0) { 3844 caller_jvms = jvms(); 3845 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); 3846 do { 3847 // The following if-tests should be performed in this order 3848 if (is_method_invoke_or_aux_frame(caller_jvms)) { 3849 // Skip a Method.invoke() or auxiliary frame 3850 } else if (caller_depth > 0) { 3851 // Skip real frame 3852 --caller_depth; 3853 } else { 3854 // We're done: reached desired caller after skipping. 3855 break; 3856 } 3857 caller_jvms = caller_jvms->caller(); 3858 --inlining_depth; 3859 } while (inlining_depth > 0); 3860 } 3861 3862 if (inlining_depth == 0) { 3863 #ifndef PRODUCT 3864 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3865 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); 3866 tty->print_cr(" JVM state at this point:"); 3867 for (int i = _depth; i >= 1; i--) { 3868 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3869 } 3870 } 3871 #endif 3872 return false; // Reached end of inlining 3873 } 3874 3875 // Acquire method holder as java.lang.Class 3876 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 3877 ciInstance* caller_mirror = caller_klass->java_mirror(); 3878 // Push this as a constant 3879 push(makecon(TypeInstPtr::make(caller_mirror))); 3880 #ifndef PRODUCT 3881 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3882 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); 3883 tty->print_cr(" JVM state at this point:"); 3884 for (int i = _depth; i >= 1; i--) { 3885 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3886 } 3887 } 3888 #endif 3889 return true; 3890 } 3891 3892 // Helper routine for above 3893 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { 3894 ciMethod* method = jvms->method(); 3895 3896 // Is this the Method.invoke method itself? 3897 if (method->intrinsic_id() == vmIntrinsics::_invoke) 3898 return true; 3899 3900 // Is this a helper, defined somewhere underneath MethodAccessorImpl. 3901 ciKlass* k = method->holder(); 3902 if (k->is_instance_klass()) { 3903 ciInstanceKlass* ik = k->as_instance_klass(); 3904 for (; ik != NULL; ik = ik->super()) { 3905 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && 3906 ik == env()->find_system_klass(ik->name())) { 3907 return true; 3908 } 3909 } 3910 } 3911 else if (method->is_method_handle_adapter()) { 3912 // This is an internal adapter frame from the MethodHandleCompiler -- skip it 3913 return true; 3914 } 3915 3916 return false; 3917 } 3918 3919 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by 3920 // inline_native_AtomicLong_attemptUpdate() but it has no way of 3921 // computing it since there is no lookup field by name function in the 3922 // CI interface. This is computed and set by inline_native_AtomicLong_get(). 3923 // Using a static variable here is safe even if we have multiple compilation 3924 // threads because the offset is constant. At worst the same offset will be 3925 // computed and stored multiple 3926 3927 bool LibraryCallKit::inline_native_AtomicLong_get() { 3928 // Restore the stack and pop off the argument 3929 _sp+=1; 3930 Node *obj = pop(); 3931 3932 // get the offset of the "value" field. Since the CI interfaces 3933 // does not provide a way to look up a field by name, we scan the bytecodes 3934 // to get the field index. We expect the first 2 instructions of the method 3935 // to be: 3936 // 0 aload_0 3937 // 1 getfield "value" 3938 ciMethod* method = callee(); 3939 if (value_field_offset == -1) 3940 { 3941 ciField* value_field; 3942 ciBytecodeStream iter(method); 3943 Bytecodes::Code bc = iter.next(); 3944 3945 if ((bc != Bytecodes::_aload_0) && 3946 ((bc != Bytecodes::_aload) || (iter.get_index() != 0))) 3947 return false; 3948 bc = iter.next(); 3949 if (bc != Bytecodes::_getfield) 3950 return false; 3951 bool ignore; 3952 value_field = iter.get_field(ignore); 3953 value_field_offset = value_field->offset_in_bytes(); 3954 } 3955 3956 // Null check without removing any arguments. 3957 _sp++; 3958 obj = do_null_check(obj, T_OBJECT); 3959 _sp--; 3960 // Check for locking null object 3961 if (stopped()) return true; 3962 3963 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3964 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3965 int alias_idx = C->get_alias_index(adr_type); 3966 3967 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr)); 3968 3969 push_pair(result); 3970 3971 return true; 3972 } 3973 3974 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() { 3975 // Restore the stack and pop off the arguments 3976 _sp+=5; 3977 Node *newVal = pop_pair(); 3978 Node *oldVal = pop_pair(); 3979 Node *obj = pop(); 3980 3981 // we need the offset of the "value" field which was computed when 3982 // inlining the get() method. Give up if we don't have it. 3983 if (value_field_offset == -1) 3984 return false; 3985 3986 // Null check without removing any arguments. 3987 _sp+=5; 3988 obj = do_null_check(obj, T_OBJECT); 3989 _sp-=5; 3990 // Check for locking null object 3991 if (stopped()) return true; 3992 3993 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3994 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3995 int alias_idx = C->get_alias_index(adr_type); 3996 3997 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal)); 3998 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 3999 set_memory(store_proj, alias_idx); 4000 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) ); 4001 4002 Node *result; 4003 // CMove node is not used to be able fold a possible check code 4004 // after attemptUpdate() call. This code could be transformed 4005 // into CMove node by loop optimizations. 4006 { 4007 RegionNode *r = new (C, 3) RegionNode(3); 4008 result = new (C, 3) PhiNode(r, TypeInt::BOOL); 4009 4010 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 4011 Node *iftrue = opt_iff(r, iff); 4012 r->init_req(1, iftrue); 4013 result->init_req(1, intcon(1)); 4014 result->init_req(2, intcon(0)); 4015 4016 set_control(_gvn.transform(r)); 4017 record_for_igvn(r); 4018 4019 C->set_has_split_ifs(true); // Has chance for split-if optimization 4020 } 4021 4022 push(_gvn.transform(result)); 4023 return true; 4024 } 4025 4026 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4027 // restore the arguments 4028 _sp += arg_size(); 4029 4030 switch (id) { 4031 case vmIntrinsics::_floatToRawIntBits: 4032 push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); 4033 break; 4034 4035 case vmIntrinsics::_intBitsToFloat: 4036 push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); 4037 break; 4038 4039 case vmIntrinsics::_doubleToRawLongBits: 4040 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); 4041 break; 4042 4043 case vmIntrinsics::_longBitsToDouble: 4044 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); 4045 break; 4046 4047 case vmIntrinsics::_doubleToLongBits: { 4048 Node* value = pop_pair(); 4049 4050 // two paths (plus control) merge in a wood 4051 RegionNode *r = new (C, 3) RegionNode(3); 4052 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); 4053 4054 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); 4055 // Build the boolean node 4056 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 4057 4058 // Branch either way. 4059 // NaN case is less traveled, which makes all the difference. 4060 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4061 Node *opt_isnan = _gvn.transform(ifisnan); 4062 assert( opt_isnan->is_If(), "Expect an IfNode"); 4063 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4064 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 4065 4066 set_control(iftrue); 4067 4068 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4069 Node *slow_result = longcon(nan_bits); // return NaN 4070 phi->init_req(1, _gvn.transform( slow_result )); 4071 r->init_req(1, iftrue); 4072 4073 // Else fall through 4074 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 4075 set_control(iffalse); 4076 4077 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); 4078 r->init_req(2, iffalse); 4079 4080 // Post merge 4081 set_control(_gvn.transform(r)); 4082 record_for_igvn(r); 4083 4084 Node* result = _gvn.transform(phi); 4085 assert(result->bottom_type()->isa_long(), "must be"); 4086 push_pair(result); 4087 4088 C->set_has_split_ifs(true); // Has chance for split-if optimization 4089 4090 break; 4091 } 4092 4093 case vmIntrinsics::_floatToIntBits: { 4094 Node* value = pop(); 4095 4096 // two paths (plus control) merge in a wood 4097 RegionNode *r = new (C, 3) RegionNode(3); 4098 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); 4099 4100 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); 4101 // Build the boolean node 4102 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 4103 4104 // Branch either way. 4105 // NaN case is less traveled, which makes all the difference. 4106 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4107 Node *opt_isnan = _gvn.transform(ifisnan); 4108 assert( opt_isnan->is_If(), "Expect an IfNode"); 4109 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4110 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 4111 4112 set_control(iftrue); 4113 4114 static const jint nan_bits = 0x7fc00000; 4115 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4116 phi->init_req(1, _gvn.transform( slow_result )); 4117 r->init_req(1, iftrue); 4118 4119 // Else fall through 4120 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 4121 set_control(iffalse); 4122 4123 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); 4124 r->init_req(2, iffalse); 4125 4126 // Post merge 4127 set_control(_gvn.transform(r)); 4128 record_for_igvn(r); 4129 4130 Node* result = _gvn.transform(phi); 4131 assert(result->bottom_type()->isa_int(), "must be"); 4132 push(result); 4133 4134 C->set_has_split_ifs(true); // Has chance for split-if optimization 4135 4136 break; 4137 } 4138 4139 default: 4140 ShouldNotReachHere(); 4141 } 4142 4143 return true; 4144 } 4145 4146 #ifdef _LP64 4147 #define XTOP ,top() /*additional argument*/ 4148 #else //_LP64 4149 #define XTOP /*no additional argument*/ 4150 #endif //_LP64 4151 4152 //----------------------inline_unsafe_copyMemory------------------------- 4153 bool LibraryCallKit::inline_unsafe_copyMemory() { 4154 if (callee()->is_static()) return false; // caller must have the capability! 4155 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) 4156 assert(signature()->size() == nargs-1, "copy has 5 arguments"); 4157 null_check_receiver(callee()); // check then ignore argument(0) 4158 if (stopped()) return true; 4159 4160 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4161 4162 Node* src_ptr = argument(1); 4163 Node* src_off = ConvL2X(argument(2)); 4164 assert(argument(3)->is_top(), "2nd half of long"); 4165 Node* dst_ptr = argument(4); 4166 Node* dst_off = ConvL2X(argument(5)); 4167 assert(argument(6)->is_top(), "2nd half of long"); 4168 Node* size = ConvL2X(argument(7)); 4169 assert(argument(8)->is_top(), "2nd half of long"); 4170 4171 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4172 "fieldOffset must be byte-scaled"); 4173 4174 Node* src = make_unsafe_address(src_ptr, src_off); 4175 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4176 4177 // Conservatively insert a memory barrier on all memory slices. 4178 // Do not let writes of the copy source or destination float below the copy. 4179 insert_mem_bar(Op_MemBarCPUOrder); 4180 4181 // Call it. Note that the length argument is not scaled. 4182 make_runtime_call(RC_LEAF|RC_NO_FP, 4183 OptoRuntime::fast_arraycopy_Type(), 4184 StubRoutines::unsafe_arraycopy(), 4185 "unsafe_arraycopy", 4186 TypeRawPtr::BOTTOM, 4187 src, dst, size XTOP); 4188 4189 // Do not let reads of the copy destination float above the copy. 4190 insert_mem_bar(Op_MemBarCPUOrder); 4191 4192 return true; 4193 } 4194 4195 //------------------------clone_coping----------------------------------- 4196 // Helper function for inline_native_clone. 4197 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4198 assert(obj_size != NULL, ""); 4199 Node* raw_obj = alloc_obj->in(1); 4200 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4201 4202 AllocateNode* alloc = NULL; 4203 if (ReduceBulkZeroing) { 4204 // We will be completely responsible for initializing this object - 4205 // mark Initialize node as complete. 4206 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4207 // The object was just allocated - there should be no any stores! 4208 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4209 // Mark as complete_with_arraycopy so that on AllocateNode 4210 // expansion, we know this AllocateNode is initialized by an array 4211 // copy and a StoreStore barrier exists after the array copy. 4212 alloc->initialization()->set_complete_with_arraycopy(); 4213 } 4214 4215 // Copy the fastest available way. 4216 // TODO: generate fields copies for small objects instead. 4217 Node* src = obj; 4218 Node* dest = alloc_obj; 4219 Node* size = _gvn.transform(obj_size); 4220 4221 // Exclude the header but include array length to copy by 8 bytes words. 4222 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4223 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4224 instanceOopDesc::base_offset_in_bytes(); 4225 // base_off: 4226 // 8 - 32-bit VM 4227 // 12 - 64-bit VM, compressed oops 4228 // 16 - 64-bit VM, normal oops 4229 if (base_off % BytesPerLong != 0) { 4230 assert(UseCompressedOops, ""); 4231 if (is_array) { 4232 // Exclude length to copy by 8 bytes words. 4233 base_off += sizeof(int); 4234 } else { 4235 // Include klass to copy by 8 bytes words. 4236 base_off = instanceOopDesc::klass_offset_in_bytes(); 4237 } 4238 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4239 } 4240 src = basic_plus_adr(src, base_off); 4241 dest = basic_plus_adr(dest, base_off); 4242 4243 // Compute the length also, if needed: 4244 Node* countx = size; 4245 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); 4246 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4247 4248 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4249 bool disjoint_bases = true; 4250 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4251 src, NULL, dest, NULL, countx, 4252 /*dest_uninitialized*/true); 4253 4254 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4255 if (card_mark) { 4256 assert(!is_array, ""); 4257 // Put in store barrier for any and all oops we are sticking 4258 // into this object. (We could avoid this if we could prove 4259 // that the object type contains no oop fields at all.) 4260 Node* no_particular_value = NULL; 4261 Node* no_particular_field = NULL; 4262 int raw_adr_idx = Compile::AliasIdxRaw; 4263 post_barrier(control(), 4264 memory(raw_adr_type), 4265 alloc_obj, 4266 no_particular_field, 4267 raw_adr_idx, 4268 no_particular_value, 4269 T_OBJECT, 4270 false); 4271 } 4272 4273 // Do not let reads from the cloned object float above the arraycopy. 4274 if (alloc != NULL) { 4275 // Do not let stores that initialize this object be reordered with 4276 // a subsequent store that would make this object accessible by 4277 // other threads. 4278 // Record what AllocateNode this StoreStore protects so that 4279 // escape analysis can go from the MemBarStoreStoreNode to the 4280 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4281 // based on the escape status of the AllocateNode. 4282 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4283 } else { 4284 insert_mem_bar(Op_MemBarCPUOrder); 4285 } 4286 } 4287 4288 //------------------------inline_native_clone---------------------------- 4289 // Here are the simple edge cases: 4290 // null receiver => normal trap 4291 // virtual and clone was overridden => slow path to out-of-line clone 4292 // not cloneable or finalizer => slow path to out-of-line Object.clone 4293 // 4294 // The general case has two steps, allocation and copying. 4295 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4296 // 4297 // Copying also has two cases, oop arrays and everything else. 4298 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4299 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4300 // 4301 // These steps fold up nicely if and when the cloned object's klass 4302 // can be sharply typed as an object array, a type array, or an instance. 4303 // 4304 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4305 int nargs = 1; 4306 PhiNode* result_val; 4307 4308 //set the original stack and the reexecute bit for the interpreter to reexecute 4309 //the bytecode that invokes Object.clone if deoptimization happens 4310 { PreserveReexecuteState preexecs(this); 4311 jvms()->set_should_reexecute(true); 4312 4313 //null_check_receiver will adjust _sp (push and pop) 4314 Node* obj = null_check_receiver(callee()); 4315 if (stopped()) return true; 4316 4317 _sp += nargs; 4318 4319 Node* obj_klass = load_object_klass(obj); 4320 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4321 const TypeOopPtr* toop = ((tklass != NULL) 4322 ? tklass->as_instance_type() 4323 : TypeInstPtr::NOTNULL); 4324 4325 // Conservatively insert a memory barrier on all memory slices. 4326 // Do not let writes into the original float below the clone. 4327 insert_mem_bar(Op_MemBarCPUOrder); 4328 4329 // paths into result_reg: 4330 enum { 4331 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4332 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4333 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4334 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4335 PATH_LIMIT 4336 }; 4337 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4338 result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 4339 TypeInstPtr::NOTNULL); 4340 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 4341 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 4342 TypePtr::BOTTOM); 4343 record_for_igvn(result_reg); 4344 4345 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4346 int raw_adr_idx = Compile::AliasIdxRaw; 4347 4348 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4349 if (array_ctl != NULL) { 4350 // It's an array. 4351 PreserveJVMState pjvms(this); 4352 set_control(array_ctl); 4353 Node* obj_length = load_array_length(obj); 4354 Node* obj_size = NULL; 4355 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); 4356 4357 if (!use_ReduceInitialCardMarks()) { 4358 // If it is an oop array, it requires very special treatment, 4359 // because card marking is required on each card of the array. 4360 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4361 if (is_obja != NULL) { 4362 PreserveJVMState pjvms2(this); 4363 set_control(is_obja); 4364 // Generate a direct call to the right arraycopy function(s). 4365 bool disjoint_bases = true; 4366 bool length_never_negative = true; 4367 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4368 obj, intcon(0), alloc_obj, intcon(0), 4369 obj_length, 4370 disjoint_bases, length_never_negative); 4371 result_reg->init_req(_objArray_path, control()); 4372 result_val->init_req(_objArray_path, alloc_obj); 4373 result_i_o ->set_req(_objArray_path, i_o()); 4374 result_mem ->set_req(_objArray_path, reset_memory()); 4375 } 4376 } 4377 // Otherwise, there are no card marks to worry about. 4378 // (We can dispense with card marks if we know the allocation 4379 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4380 // causes the non-eden paths to take compensating steps to 4381 // simulate a fresh allocation, so that no further 4382 // card marks are required in compiled code to initialize 4383 // the object.) 4384 4385 if (!stopped()) { 4386 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4387 4388 // Present the results of the copy. 4389 result_reg->init_req(_array_path, control()); 4390 result_val->init_req(_array_path, alloc_obj); 4391 result_i_o ->set_req(_array_path, i_o()); 4392 result_mem ->set_req(_array_path, reset_memory()); 4393 } 4394 } 4395 4396 // We only go to the instance fast case code if we pass a number of guards. 4397 // The paths which do not pass are accumulated in the slow_region. 4398 RegionNode* slow_region = new (C, 1) RegionNode(1); 4399 record_for_igvn(slow_region); 4400 if (!stopped()) { 4401 // It's an instance (we did array above). Make the slow-path tests. 4402 // If this is a virtual call, we generate a funny guard. We grab 4403 // the vtable entry corresponding to clone() from the target object. 4404 // If the target method which we are calling happens to be the 4405 // Object clone() method, we pass the guard. We do not need this 4406 // guard for non-virtual calls; the caller is known to be the native 4407 // Object clone(). 4408 if (is_virtual) { 4409 generate_virtual_guard(obj_klass, slow_region); 4410 } 4411 4412 // The object must be cloneable and must not have a finalizer. 4413 // Both of these conditions may be checked in a single test. 4414 // We could optimize the cloneable test further, but we don't care. 4415 generate_access_flags_guard(obj_klass, 4416 // Test both conditions: 4417 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4418 // Must be cloneable but not finalizer: 4419 JVM_ACC_IS_CLONEABLE, 4420 slow_region); 4421 } 4422 4423 if (!stopped()) { 4424 // It's an instance, and it passed the slow-path tests. 4425 PreserveJVMState pjvms(this); 4426 Node* obj_size = NULL; 4427 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size); 4428 4429 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4430 4431 // Present the results of the slow call. 4432 result_reg->init_req(_instance_path, control()); 4433 result_val->init_req(_instance_path, alloc_obj); 4434 result_i_o ->set_req(_instance_path, i_o()); 4435 result_mem ->set_req(_instance_path, reset_memory()); 4436 } 4437 4438 // Generate code for the slow case. We make a call to clone(). 4439 set_control(_gvn.transform(slow_region)); 4440 if (!stopped()) { 4441 PreserveJVMState pjvms(this); 4442 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4443 Node* slow_result = set_results_for_java_call(slow_call); 4444 // this->control() comes from set_results_for_java_call 4445 result_reg->init_req(_slow_path, control()); 4446 result_val->init_req(_slow_path, slow_result); 4447 result_i_o ->set_req(_slow_path, i_o()); 4448 result_mem ->set_req(_slow_path, reset_memory()); 4449 } 4450 4451 // Return the combined state. 4452 set_control( _gvn.transform(result_reg) ); 4453 set_i_o( _gvn.transform(result_i_o) ); 4454 set_all_memory( _gvn.transform(result_mem) ); 4455 } //original reexecute and sp are set back here 4456 4457 push(_gvn.transform(result_val)); 4458 4459 return true; 4460 } 4461 4462 //------------------------------basictype2arraycopy---------------------------- 4463 address LibraryCallKit::basictype2arraycopy(BasicType t, 4464 Node* src_offset, 4465 Node* dest_offset, 4466 bool disjoint_bases, 4467 const char* &name, 4468 bool dest_uninitialized) { 4469 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4470 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4471 4472 bool aligned = false; 4473 bool disjoint = disjoint_bases; 4474 4475 // if the offsets are the same, we can treat the memory regions as 4476 // disjoint, because either the memory regions are in different arrays, 4477 // or they are identical (which we can treat as disjoint.) We can also 4478 // treat a copy with a destination index less that the source index 4479 // as disjoint since a low->high copy will work correctly in this case. 4480 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4481 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4482 // both indices are constants 4483 int s_offs = src_offset_inttype->get_con(); 4484 int d_offs = dest_offset_inttype->get_con(); 4485 int element_size = type2aelembytes(t); 4486 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4487 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4488 if (s_offs >= d_offs) disjoint = true; 4489 } else if (src_offset == dest_offset && src_offset != NULL) { 4490 // This can occur if the offsets are identical non-constants. 4491 disjoint = true; 4492 } 4493 4494 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4495 } 4496 4497 4498 //------------------------------inline_arraycopy----------------------- 4499 bool LibraryCallKit::inline_arraycopy() { 4500 // Restore the stack and pop off the arguments. 4501 int nargs = 5; // 2 oops, 3 ints, no size_t or long 4502 assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); 4503 4504 Node *src = argument(0); 4505 Node *src_offset = argument(1); 4506 Node *dest = argument(2); 4507 Node *dest_offset = argument(3); 4508 Node *length = argument(4); 4509 4510 // Compile time checks. If any of these checks cannot be verified at compile time, 4511 // we do not make a fast path for this call. Instead, we let the call remain as it 4512 // is. The checks we choose to mandate at compile time are: 4513 // 4514 // (1) src and dest are arrays. 4515 const Type* src_type = src->Value(&_gvn); 4516 const Type* dest_type = dest->Value(&_gvn); 4517 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4518 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4519 if (top_src == NULL || top_src->klass() == NULL || 4520 top_dest == NULL || top_dest->klass() == NULL) { 4521 // Conservatively insert a memory barrier on all memory slices. 4522 // Do not let writes into the source float below the arraycopy. 4523 insert_mem_bar(Op_MemBarCPUOrder); 4524 4525 // Call StubRoutines::generic_arraycopy stub. 4526 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4527 src, src_offset, dest, dest_offset, length); 4528 4529 // Do not let reads from the destination float above the arraycopy. 4530 // Since we cannot type the arrays, we don't know which slices 4531 // might be affected. We could restrict this barrier only to those 4532 // memory slices which pertain to array elements--but don't bother. 4533 if (!InsertMemBarAfterArraycopy) 4534 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4535 insert_mem_bar(Op_MemBarCPUOrder); 4536 return true; 4537 } 4538 4539 // (2) src and dest arrays must have elements of the same BasicType 4540 // Figure out the size and type of the elements we will be copying. 4541 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4542 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4543 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4544 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4545 4546 if (src_elem != dest_elem || dest_elem == T_VOID) { 4547 // The component types are not the same or are not recognized. Punt. 4548 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4549 generate_slow_arraycopy(TypePtr::BOTTOM, 4550 src, src_offset, dest, dest_offset, length, 4551 /*dest_uninitialized*/false); 4552 return true; 4553 } 4554 4555 //--------------------------------------------------------------------------- 4556 // We will make a fast path for this call to arraycopy. 4557 4558 // We have the following tests left to perform: 4559 // 4560 // (3) src and dest must not be null. 4561 // (4) src_offset must not be negative. 4562 // (5) dest_offset must not be negative. 4563 // (6) length must not be negative. 4564 // (7) src_offset + length must not exceed length of src. 4565 // (8) dest_offset + length must not exceed length of dest. 4566 // (9) each element of an oop array must be assignable 4567 4568 RegionNode* slow_region = new (C, 1) RegionNode(1); 4569 record_for_igvn(slow_region); 4570 4571 // (3) operands must not be null 4572 // We currently perform our null checks with the do_null_check routine. 4573 // This means that the null exceptions will be reported in the caller 4574 // rather than (correctly) reported inside of the native arraycopy call. 4575 // This should be corrected, given time. We do our null check with the 4576 // stack pointer restored. 4577 _sp += nargs; 4578 src = do_null_check(src, T_ARRAY); 4579 dest = do_null_check(dest, T_ARRAY); 4580 _sp -= nargs; 4581 4582 // (4) src_offset must not be negative. 4583 generate_negative_guard(src_offset, slow_region); 4584 4585 // (5) dest_offset must not be negative. 4586 generate_negative_guard(dest_offset, slow_region); 4587 4588 // (6) length must not be negative (moved to generate_arraycopy()). 4589 // generate_negative_guard(length, slow_region); 4590 4591 // (7) src_offset + length must not exceed length of src. 4592 generate_limit_guard(src_offset, length, 4593 load_array_length(src), 4594 slow_region); 4595 4596 // (8) dest_offset + length must not exceed length of dest. 4597 generate_limit_guard(dest_offset, length, 4598 load_array_length(dest), 4599 slow_region); 4600 4601 // (9) each element of an oop array must be assignable 4602 // The generate_arraycopy subroutine checks this. 4603 4604 // This is where the memory effects are placed: 4605 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4606 generate_arraycopy(adr_type, dest_elem, 4607 src, src_offset, dest, dest_offset, length, 4608 false, false, slow_region); 4609 4610 return true; 4611 } 4612 4613 //-----------------------------generate_arraycopy---------------------- 4614 // Generate an optimized call to arraycopy. 4615 // Caller must guard against non-arrays. 4616 // Caller must determine a common array basic-type for both arrays. 4617 // Caller must validate offsets against array bounds. 4618 // The slow_region has already collected guard failure paths 4619 // (such as out of bounds length or non-conformable array types). 4620 // The generated code has this shape, in general: 4621 // 4622 // if (length == 0) return // via zero_path 4623 // slowval = -1 4624 // if (types unknown) { 4625 // slowval = call generic copy loop 4626 // if (slowval == 0) return // via checked_path 4627 // } else if (indexes in bounds) { 4628 // if ((is object array) && !(array type check)) { 4629 // slowval = call checked copy loop 4630 // if (slowval == 0) return // via checked_path 4631 // } else { 4632 // call bulk copy loop 4633 // return // via fast_path 4634 // } 4635 // } 4636 // // adjust params for remaining work: 4637 // if (slowval != -1) { 4638 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4639 // } 4640 // slow_region: 4641 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4642 // return // via slow_call_path 4643 // 4644 // This routine is used from several intrinsics: System.arraycopy, 4645 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4646 // 4647 void 4648 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4649 BasicType basic_elem_type, 4650 Node* src, Node* src_offset, 4651 Node* dest, Node* dest_offset, 4652 Node* copy_length, 4653 bool disjoint_bases, 4654 bool length_never_negative, 4655 RegionNode* slow_region) { 4656 4657 if (slow_region == NULL) { 4658 slow_region = new(C,1) RegionNode(1); 4659 record_for_igvn(slow_region); 4660 } 4661 4662 Node* original_dest = dest; 4663 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4664 bool dest_uninitialized = false; 4665 4666 // See if this is the initialization of a newly-allocated array. 4667 // If so, we will take responsibility here for initializing it to zero. 4668 // (Note: Because tightly_coupled_allocation performs checks on the 4669 // out-edges of the dest, we need to avoid making derived pointers 4670 // from it until we have checked its uses.) 4671 if (ReduceBulkZeroing 4672 && !ZeroTLAB // pointless if already zeroed 4673 && basic_elem_type != T_CONFLICT // avoid corner case 4674 && !src->eqv_uncast(dest) 4675 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4676 != NULL) 4677 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4678 && alloc->maybe_set_complete(&_gvn)) { 4679 // "You break it, you buy it." 4680 InitializeNode* init = alloc->initialization(); 4681 assert(init->is_complete(), "we just did this"); 4682 init->set_complete_with_arraycopy(); 4683 assert(dest->is_CheckCastPP(), "sanity"); 4684 assert(dest->in(0)->in(0) == init, "dest pinned"); 4685 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4686 // From this point on, every exit path is responsible for 4687 // initializing any non-copied parts of the object to zero. 4688 // Also, if this flag is set we make sure that arraycopy interacts properly 4689 // with G1, eliding pre-barriers. See CR 6627983. 4690 dest_uninitialized = true; 4691 } else { 4692 // No zeroing elimination here. 4693 alloc = NULL; 4694 //original_dest = dest; 4695 //dest_uninitialized = false; 4696 } 4697 4698 // Results are placed here: 4699 enum { fast_path = 1, // normal void-returning assembly stub 4700 checked_path = 2, // special assembly stub with cleanup 4701 slow_call_path = 3, // something went wrong; call the VM 4702 zero_path = 4, // bypass when length of copy is zero 4703 bcopy_path = 5, // copy primitive array by 64-bit blocks 4704 PATH_LIMIT = 6 4705 }; 4706 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4707 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO); 4708 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type); 4709 record_for_igvn(result_region); 4710 _gvn.set_type_bottom(result_i_o); 4711 _gvn.set_type_bottom(result_memory); 4712 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4713 4714 // The slow_control path: 4715 Node* slow_control; 4716 Node* slow_i_o = i_o(); 4717 Node* slow_mem = memory(adr_type); 4718 debug_only(slow_control = (Node*) badAddress); 4719 4720 // Checked control path: 4721 Node* checked_control = top(); 4722 Node* checked_mem = NULL; 4723 Node* checked_i_o = NULL; 4724 Node* checked_value = NULL; 4725 4726 if (basic_elem_type == T_CONFLICT) { 4727 assert(!dest_uninitialized, ""); 4728 Node* cv = generate_generic_arraycopy(adr_type, 4729 src, src_offset, dest, dest_offset, 4730 copy_length, dest_uninitialized); 4731 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4732 checked_control = control(); 4733 checked_i_o = i_o(); 4734 checked_mem = memory(adr_type); 4735 checked_value = cv; 4736 set_control(top()); // no fast path 4737 } 4738 4739 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4740 if (not_pos != NULL) { 4741 PreserveJVMState pjvms(this); 4742 set_control(not_pos); 4743 4744 // (6) length must not be negative. 4745 if (!length_never_negative) { 4746 generate_negative_guard(copy_length, slow_region); 4747 } 4748 4749 // copy_length is 0. 4750 if (!stopped() && dest_uninitialized) { 4751 Node* dest_length = alloc->in(AllocateNode::ALength); 4752 if (copy_length->eqv_uncast(dest_length) 4753 || _gvn.find_int_con(dest_length, 1) <= 0) { 4754 // There is no zeroing to do. No need for a secondary raw memory barrier. 4755 } else { 4756 // Clear the whole thing since there are no source elements to copy. 4757 generate_clear_array(adr_type, dest, basic_elem_type, 4758 intcon(0), NULL, 4759 alloc->in(AllocateNode::AllocSize)); 4760 // Use a secondary InitializeNode as raw memory barrier. 4761 // Currently it is needed only on this path since other 4762 // paths have stub or runtime calls as raw memory barriers. 4763 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4764 Compile::AliasIdxRaw, 4765 top())->as_Initialize(); 4766 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4767 } 4768 } 4769 4770 // Present the results of the fast call. 4771 result_region->init_req(zero_path, control()); 4772 result_i_o ->init_req(zero_path, i_o()); 4773 result_memory->init_req(zero_path, memory(adr_type)); 4774 } 4775 4776 if (!stopped() && dest_uninitialized) { 4777 // We have to initialize the *uncopied* part of the array to zero. 4778 // The copy destination is the slice dest[off..off+len]. The other slices 4779 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4780 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4781 Node* dest_length = alloc->in(AllocateNode::ALength); 4782 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset, 4783 copy_length) ); 4784 4785 // If there is a head section that needs zeroing, do it now. 4786 if (find_int_con(dest_offset, -1) != 0) { 4787 generate_clear_array(adr_type, dest, basic_elem_type, 4788 intcon(0), dest_offset, 4789 NULL); 4790 } 4791 4792 // Next, perform a dynamic check on the tail length. 4793 // It is often zero, and we can win big if we prove this. 4794 // There are two wins: Avoid generating the ClearArray 4795 // with its attendant messy index arithmetic, and upgrade 4796 // the copy to a more hardware-friendly word size of 64 bits. 4797 Node* tail_ctl = NULL; 4798 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 4799 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) ); 4800 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) ); 4801 tail_ctl = generate_slow_guard(bol_lt, NULL); 4802 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4803 } 4804 4805 // At this point, let's assume there is no tail. 4806 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4807 // There is no tail. Try an upgrade to a 64-bit copy. 4808 bool didit = false; 4809 { PreserveJVMState pjvms(this); 4810 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4811 src, src_offset, dest, dest_offset, 4812 dest_size, dest_uninitialized); 4813 if (didit) { 4814 // Present the results of the block-copying fast call. 4815 result_region->init_req(bcopy_path, control()); 4816 result_i_o ->init_req(bcopy_path, i_o()); 4817 result_memory->init_req(bcopy_path, memory(adr_type)); 4818 } 4819 } 4820 if (didit) 4821 set_control(top()); // no regular fast path 4822 } 4823 4824 // Clear the tail, if any. 4825 if (tail_ctl != NULL) { 4826 Node* notail_ctl = stopped() ? NULL : control(); 4827 set_control(tail_ctl); 4828 if (notail_ctl == NULL) { 4829 generate_clear_array(adr_type, dest, basic_elem_type, 4830 dest_tail, NULL, 4831 dest_size); 4832 } else { 4833 // Make a local merge. 4834 Node* done_ctl = new(C,3) RegionNode(3); 4835 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type); 4836 done_ctl->init_req(1, notail_ctl); 4837 done_mem->init_req(1, memory(adr_type)); 4838 generate_clear_array(adr_type, dest, basic_elem_type, 4839 dest_tail, NULL, 4840 dest_size); 4841 done_ctl->init_req(2, control()); 4842 done_mem->init_req(2, memory(adr_type)); 4843 set_control( _gvn.transform(done_ctl) ); 4844 set_memory( _gvn.transform(done_mem), adr_type ); 4845 } 4846 } 4847 } 4848 4849 BasicType copy_type = basic_elem_type; 4850 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 4851 if (!stopped() && copy_type == T_OBJECT) { 4852 // If src and dest have compatible element types, we can copy bits. 4853 // Types S[] and D[] are compatible if D is a supertype of S. 4854 // 4855 // If they are not, we will use checked_oop_disjoint_arraycopy, 4856 // which performs a fast optimistic per-oop check, and backs off 4857 // further to JVM_ArrayCopy on the first per-oop check that fails. 4858 // (Actually, we don't move raw bits only; the GC requires card marks.) 4859 4860 // Get the klassOop for both src and dest 4861 Node* src_klass = load_object_klass(src); 4862 Node* dest_klass = load_object_klass(dest); 4863 4864 // Generate the subtype check. 4865 // This might fold up statically, or then again it might not. 4866 // 4867 // Non-static example: Copying List<String>.elements to a new String[]. 4868 // The backing store for a List<String> is always an Object[], 4869 // but its elements are always type String, if the generic types 4870 // are correct at the source level. 4871 // 4872 // Test S[] against D[], not S against D, because (probably) 4873 // the secondary supertype cache is less busy for S[] than S. 4874 // This usually only matters when D is an interface. 4875 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4876 // Plug failing path into checked_oop_disjoint_arraycopy 4877 if (not_subtype_ctrl != top()) { 4878 PreserveJVMState pjvms(this); 4879 set_control(not_subtype_ctrl); 4880 // (At this point we can assume disjoint_bases, since types differ.) 4881 int ek_offset = in_bytes(objArrayKlass::element_klass_offset()); 4882 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 4883 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 4884 Node* dest_elem_klass = _gvn.transform(n1); 4885 Node* cv = generate_checkcast_arraycopy(adr_type, 4886 dest_elem_klass, 4887 src, src_offset, dest, dest_offset, 4888 ConvI2X(copy_length), dest_uninitialized); 4889 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4890 checked_control = control(); 4891 checked_i_o = i_o(); 4892 checked_mem = memory(adr_type); 4893 checked_value = cv; 4894 } 4895 // At this point we know we do not need type checks on oop stores. 4896 4897 // Let's see if we need card marks: 4898 if (alloc != NULL && use_ReduceInitialCardMarks()) { 4899 // If we do not need card marks, copy using the jint or jlong stub. 4900 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 4901 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 4902 "sizes agree"); 4903 } 4904 } 4905 4906 if (!stopped()) { 4907 // Generate the fast path, if possible. 4908 PreserveJVMState pjvms(this); 4909 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 4910 src, src_offset, dest, dest_offset, 4911 ConvI2X(copy_length), dest_uninitialized); 4912 4913 // Present the results of the fast call. 4914 result_region->init_req(fast_path, control()); 4915 result_i_o ->init_req(fast_path, i_o()); 4916 result_memory->init_req(fast_path, memory(adr_type)); 4917 } 4918 4919 // Here are all the slow paths up to this point, in one bundle: 4920 slow_control = top(); 4921 if (slow_region != NULL) 4922 slow_control = _gvn.transform(slow_region); 4923 debug_only(slow_region = (RegionNode*)badAddress); 4924 4925 set_control(checked_control); 4926 if (!stopped()) { 4927 // Clean up after the checked call. 4928 // The returned value is either 0 or -1^K, 4929 // where K = number of partially transferred array elements. 4930 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) ); 4931 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 4932 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4933 4934 // If it is 0, we are done, so transfer to the end. 4935 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) ); 4936 result_region->init_req(checked_path, checks_done); 4937 result_i_o ->init_req(checked_path, checked_i_o); 4938 result_memory->init_req(checked_path, checked_mem); 4939 4940 // If it is not zero, merge into the slow call. 4941 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) )); 4942 RegionNode* slow_reg2 = new(C, 3) RegionNode(3); 4943 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO); 4944 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type); 4945 record_for_igvn(slow_reg2); 4946 slow_reg2 ->init_req(1, slow_control); 4947 slow_i_o2 ->init_req(1, slow_i_o); 4948 slow_mem2 ->init_req(1, slow_mem); 4949 slow_reg2 ->init_req(2, control()); 4950 slow_i_o2 ->init_req(2, checked_i_o); 4951 slow_mem2 ->init_req(2, checked_mem); 4952 4953 slow_control = _gvn.transform(slow_reg2); 4954 slow_i_o = _gvn.transform(slow_i_o2); 4955 slow_mem = _gvn.transform(slow_mem2); 4956 4957 if (alloc != NULL) { 4958 // We'll restart from the very beginning, after zeroing the whole thing. 4959 // This can cause double writes, but that's OK since dest is brand new. 4960 // So we ignore the low 31 bits of the value returned from the stub. 4961 } else { 4962 // We must continue the copy exactly where it failed, or else 4963 // another thread might see the wrong number of writes to dest. 4964 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) ); 4965 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT); 4966 slow_offset->init_req(1, intcon(0)); 4967 slow_offset->init_req(2, checked_offset); 4968 slow_offset = _gvn.transform(slow_offset); 4969 4970 // Adjust the arguments by the conditionally incoming offset. 4971 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) ); 4972 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) ); 4973 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) ); 4974 4975 // Tweak the node variables to adjust the code produced below: 4976 src_offset = src_off_plus; 4977 dest_offset = dest_off_plus; 4978 copy_length = length_minus; 4979 } 4980 } 4981 4982 set_control(slow_control); 4983 if (!stopped()) { 4984 // Generate the slow path, if needed. 4985 PreserveJVMState pjvms(this); // replace_in_map may trash the map 4986 4987 set_memory(slow_mem, adr_type); 4988 set_i_o(slow_i_o); 4989 4990 if (dest_uninitialized) { 4991 generate_clear_array(adr_type, dest, basic_elem_type, 4992 intcon(0), NULL, 4993 alloc->in(AllocateNode::AllocSize)); 4994 } 4995 4996 generate_slow_arraycopy(adr_type, 4997 src, src_offset, dest, dest_offset, 4998 copy_length, /*dest_uninitialized*/false); 4999 5000 result_region->init_req(slow_call_path, control()); 5001 result_i_o ->init_req(slow_call_path, i_o()); 5002 result_memory->init_req(slow_call_path, memory(adr_type)); 5003 } 5004 5005 // Remove unused edges. 5006 for (uint i = 1; i < result_region->req(); i++) { 5007 if (result_region->in(i) == NULL) 5008 result_region->init_req(i, top()); 5009 } 5010 5011 // Finished; return the combined state. 5012 set_control( _gvn.transform(result_region) ); 5013 set_i_o( _gvn.transform(result_i_o) ); 5014 set_memory( _gvn.transform(result_memory), adr_type ); 5015 5016 // The memory edges above are precise in order to model effects around 5017 // array copies accurately to allow value numbering of field loads around 5018 // arraycopy. Such field loads, both before and after, are common in Java 5019 // collections and similar classes involving header/array data structures. 5020 // 5021 // But with low number of register or when some registers are used or killed 5022 // by arraycopy calls it causes registers spilling on stack. See 6544710. 5023 // The next memory barrier is added to avoid it. If the arraycopy can be 5024 // optimized away (which it can, sometimes) then we can manually remove 5025 // the membar also. 5026 // 5027 // Do not let reads from the cloned object float above the arraycopy. 5028 if (alloc != NULL) { 5029 // Do not let stores that initialize this object be reordered with 5030 // a subsequent store that would make this object accessible by 5031 // other threads. 5032 // Record what AllocateNode this StoreStore protects so that 5033 // escape analysis can go from the MemBarStoreStoreNode to the 5034 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5035 // based on the escape status of the AllocateNode. 5036 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5037 } else if (InsertMemBarAfterArraycopy) 5038 insert_mem_bar(Op_MemBarCPUOrder); 5039 } 5040 5041 5042 // Helper function which determines if an arraycopy immediately follows 5043 // an allocation, with no intervening tests or other escapes for the object. 5044 AllocateArrayNode* 5045 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5046 RegionNode* slow_region) { 5047 if (stopped()) return NULL; // no fast path 5048 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5049 5050 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5051 if (alloc == NULL) return NULL; 5052 5053 Node* rawmem = memory(Compile::AliasIdxRaw); 5054 // Is the allocation's memory state untouched? 5055 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5056 // Bail out if there have been raw-memory effects since the allocation. 5057 // (Example: There might have been a call or safepoint.) 5058 return NULL; 5059 } 5060 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5061 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5062 return NULL; 5063 } 5064 5065 // There must be no unexpected observers of this allocation. 5066 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5067 Node* obs = ptr->fast_out(i); 5068 if (obs != this->map()) { 5069 return NULL; 5070 } 5071 } 5072 5073 // This arraycopy must unconditionally follow the allocation of the ptr. 5074 Node* alloc_ctl = ptr->in(0); 5075 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5076 5077 Node* ctl = control(); 5078 while (ctl != alloc_ctl) { 5079 // There may be guards which feed into the slow_region. 5080 // Any other control flow means that we might not get a chance 5081 // to finish initializing the allocated object. 5082 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5083 IfNode* iff = ctl->in(0)->as_If(); 5084 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5085 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5086 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5087 ctl = iff->in(0); // This test feeds the known slow_region. 5088 continue; 5089 } 5090 // One more try: Various low-level checks bottom out in 5091 // uncommon traps. If the debug-info of the trap omits 5092 // any reference to the allocation, as we've already 5093 // observed, then there can be no objection to the trap. 5094 bool found_trap = false; 5095 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5096 Node* obs = not_ctl->fast_out(j); 5097 if (obs->in(0) == not_ctl && obs->is_Call() && 5098 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5099 found_trap = true; break; 5100 } 5101 } 5102 if (found_trap) { 5103 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5104 continue; 5105 } 5106 } 5107 return NULL; 5108 } 5109 5110 // If we get this far, we have an allocation which immediately 5111 // precedes the arraycopy, and we can take over zeroing the new object. 5112 // The arraycopy will finish the initialization, and provide 5113 // a new control state to which we will anchor the destination pointer. 5114 5115 return alloc; 5116 } 5117 5118 // Helper for initialization of arrays, creating a ClearArray. 5119 // It writes zero bits in [start..end), within the body of an array object. 5120 // The memory effects are all chained onto the 'adr_type' alias category. 5121 // 5122 // Since the object is otherwise uninitialized, we are free 5123 // to put a little "slop" around the edges of the cleared area, 5124 // as long as it does not go back into the array's header, 5125 // or beyond the array end within the heap. 5126 // 5127 // The lower edge can be rounded down to the nearest jint and the 5128 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5129 // 5130 // Arguments: 5131 // adr_type memory slice where writes are generated 5132 // dest oop of the destination array 5133 // basic_elem_type element type of the destination 5134 // slice_idx array index of first element to store 5135 // slice_len number of elements to store (or NULL) 5136 // dest_size total size in bytes of the array object 5137 // 5138 // Exactly one of slice_len or dest_size must be non-NULL. 5139 // If dest_size is non-NULL, zeroing extends to the end of the object. 5140 // If slice_len is non-NULL, the slice_idx value must be a constant. 5141 void 5142 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5143 Node* dest, 5144 BasicType basic_elem_type, 5145 Node* slice_idx, 5146 Node* slice_len, 5147 Node* dest_size) { 5148 // one or the other but not both of slice_len and dest_size: 5149 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5150 if (slice_len == NULL) slice_len = top(); 5151 if (dest_size == NULL) dest_size = top(); 5152 5153 // operate on this memory slice: 5154 Node* mem = memory(adr_type); // memory slice to operate on 5155 5156 // scaling and rounding of indexes: 5157 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5158 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5159 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5160 int bump_bit = (-1 << scale) & BytesPerInt; 5161 5162 // determine constant starts and ends 5163 const intptr_t BIG_NEG = -128; 5164 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5165 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5166 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5167 if (slice_len_con == 0) { 5168 return; // nothing to do here 5169 } 5170 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5171 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5172 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5173 assert(end_con < 0, "not two cons"); 5174 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5175 BytesPerLong); 5176 } 5177 5178 if (start_con >= 0 && end_con >= 0) { 5179 // Constant start and end. Simple. 5180 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5181 start_con, end_con, &_gvn); 5182 } else if (start_con >= 0 && dest_size != top()) { 5183 // Constant start, pre-rounded end after the tail of the array. 5184 Node* end = dest_size; 5185 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5186 start_con, end, &_gvn); 5187 } else if (start_con >= 0 && slice_len != top()) { 5188 // Constant start, non-constant end. End needs rounding up. 5189 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5190 intptr_t end_base = abase + (slice_idx_con << scale); 5191 int end_round = (-1 << scale) & (BytesPerLong - 1); 5192 Node* end = ConvI2X(slice_len); 5193 if (scale != 0) 5194 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) )); 5195 end_base += end_round; 5196 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) ); 5197 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) ); 5198 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5199 start_con, end, &_gvn); 5200 } else if (start_con < 0 && dest_size != top()) { 5201 // Non-constant start, pre-rounded end after the tail of the array. 5202 // This is almost certainly a "round-to-end" operation. 5203 Node* start = slice_idx; 5204 start = ConvI2X(start); 5205 if (scale != 0) 5206 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) )); 5207 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) ); 5208 if ((bump_bit | clear_low) != 0) { 5209 int to_clear = (bump_bit | clear_low); 5210 // Align up mod 8, then store a jint zero unconditionally 5211 // just before the mod-8 boundary. 5212 if (((abase + bump_bit) & ~to_clear) - bump_bit 5213 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5214 bump_bit = 0; 5215 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5216 } else { 5217 // Bump 'start' up to (or past) the next jint boundary: 5218 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) ); 5219 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5220 } 5221 // Round bumped 'start' down to jlong boundary in body of array. 5222 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) ); 5223 if (bump_bit != 0) { 5224 // Store a zero to the immediately preceding jint: 5225 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) ); 5226 Node* p1 = basic_plus_adr(dest, x1); 5227 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 5228 mem = _gvn.transform(mem); 5229 } 5230 } 5231 Node* end = dest_size; // pre-rounded 5232 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5233 start, end, &_gvn); 5234 } else { 5235 // Non-constant start, unrounded non-constant end. 5236 // (Nobody zeroes a random midsection of an array using this routine.) 5237 ShouldNotReachHere(); // fix caller 5238 } 5239 5240 // Done. 5241 set_memory(mem, adr_type); 5242 } 5243 5244 5245 bool 5246 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5247 BasicType basic_elem_type, 5248 AllocateNode* alloc, 5249 Node* src, Node* src_offset, 5250 Node* dest, Node* dest_offset, 5251 Node* dest_size, bool dest_uninitialized) { 5252 // See if there is an advantage from block transfer. 5253 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5254 if (scale >= LogBytesPerLong) 5255 return false; // it is already a block transfer 5256 5257 // Look at the alignment of the starting offsets. 5258 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5259 5260 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5261 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5262 if (src_off_con < 0 || dest_off_con < 0) 5263 // At present, we can only understand constants. 5264 return false; 5265 5266 intptr_t src_off = abase + (src_off_con << scale); 5267 intptr_t dest_off = abase + (dest_off_con << scale); 5268 5269 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5270 // Non-aligned; too bad. 5271 // One more chance: Pick off an initial 32-bit word. 5272 // This is a common case, since abase can be odd mod 8. 5273 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5274 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5275 Node* sptr = basic_plus_adr(src, src_off); 5276 Node* dptr = basic_plus_adr(dest, dest_off); 5277 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 5278 store_to_memory(control(), dptr, sval, T_INT, adr_type); 5279 src_off += BytesPerInt; 5280 dest_off += BytesPerInt; 5281 } else { 5282 return false; 5283 } 5284 } 5285 assert(src_off % BytesPerLong == 0, ""); 5286 assert(dest_off % BytesPerLong == 0, ""); 5287 5288 // Do this copy by giant steps. 5289 Node* sptr = basic_plus_adr(src, src_off); 5290 Node* dptr = basic_plus_adr(dest, dest_off); 5291 Node* countx = dest_size; 5292 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) ); 5293 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) ); 5294 5295 bool disjoint_bases = true; // since alloc != NULL 5296 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5297 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5298 5299 return true; 5300 } 5301 5302 5303 // Helper function; generates code for the slow case. 5304 // We make a call to a runtime method which emulates the native method, 5305 // but without the native wrapper overhead. 5306 void 5307 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5308 Node* src, Node* src_offset, 5309 Node* dest, Node* dest_offset, 5310 Node* copy_length, bool dest_uninitialized) { 5311 assert(!dest_uninitialized, "Invariant"); 5312 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5313 OptoRuntime::slow_arraycopy_Type(), 5314 OptoRuntime::slow_arraycopy_Java(), 5315 "slow_arraycopy", adr_type, 5316 src, src_offset, dest, dest_offset, 5317 copy_length); 5318 5319 // Handle exceptions thrown by this fellow: 5320 make_slow_call_ex(call, env()->Throwable_klass(), false); 5321 } 5322 5323 // Helper function; generates code for cases requiring runtime checks. 5324 Node* 5325 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5326 Node* dest_elem_klass, 5327 Node* src, Node* src_offset, 5328 Node* dest, Node* dest_offset, 5329 Node* copy_length, bool dest_uninitialized) { 5330 if (stopped()) return NULL; 5331 5332 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5333 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5334 return NULL; 5335 } 5336 5337 // Pick out the parameters required to perform a store-check 5338 // for the target array. This is an optimistic check. It will 5339 // look in each non-null element's class, at the desired klass's 5340 // super_check_offset, for the desired klass. 5341 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5342 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5343 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr()); 5344 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5345 Node* check_value = dest_elem_klass; 5346 5347 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5348 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5349 5350 // (We know the arrays are never conjoint, because their types differ.) 5351 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5352 OptoRuntime::checkcast_arraycopy_Type(), 5353 copyfunc_addr, "checkcast_arraycopy", adr_type, 5354 // five arguments, of which two are 5355 // intptr_t (jlong in LP64) 5356 src_start, dest_start, 5357 copy_length XTOP, 5358 check_offset XTOP, 5359 check_value); 5360 5361 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5362 } 5363 5364 5365 // Helper function; generates code for cases requiring runtime checks. 5366 Node* 5367 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5368 Node* src, Node* src_offset, 5369 Node* dest, Node* dest_offset, 5370 Node* copy_length, bool dest_uninitialized) { 5371 assert(!dest_uninitialized, "Invariant"); 5372 if (stopped()) return NULL; 5373 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5374 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5375 return NULL; 5376 } 5377 5378 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5379 OptoRuntime::generic_arraycopy_Type(), 5380 copyfunc_addr, "generic_arraycopy", adr_type, 5381 src, src_offset, dest, dest_offset, copy_length); 5382 5383 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5384 } 5385 5386 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5387 void 5388 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5389 BasicType basic_elem_type, 5390 bool disjoint_bases, 5391 Node* src, Node* src_offset, 5392 Node* dest, Node* dest_offset, 5393 Node* copy_length, bool dest_uninitialized) { 5394 if (stopped()) return; // nothing to do 5395 5396 Node* src_start = src; 5397 Node* dest_start = dest; 5398 if (src_offset != NULL || dest_offset != NULL) { 5399 assert(src_offset != NULL && dest_offset != NULL, ""); 5400 src_start = array_element_address(src, src_offset, basic_elem_type); 5401 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5402 } 5403 5404 // Figure out which arraycopy runtime method to call. 5405 const char* copyfunc_name = "arraycopy"; 5406 address copyfunc_addr = 5407 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5408 disjoint_bases, copyfunc_name, dest_uninitialized); 5409 5410 // Call it. Note that the count_ix value is not scaled to a byte-size. 5411 make_runtime_call(RC_LEAF|RC_NO_FP, 5412 OptoRuntime::fast_arraycopy_Type(), 5413 copyfunc_addr, copyfunc_name, adr_type, 5414 src_start, dest_start, copy_length XTOP); 5415 } 5416 5417 //----------------------------inline_reference_get---------------------------- 5418 5419 bool LibraryCallKit::inline_reference_get() { 5420 const int nargs = 1; // self 5421 5422 guarantee(java_lang_ref_Reference::referent_offset > 0, 5423 "should have already been set"); 5424 5425 int referent_offset = java_lang_ref_Reference::referent_offset; 5426 5427 // Restore the stack and pop off the argument 5428 _sp += nargs; 5429 Node *reference_obj = pop(); 5430 5431 // Null check on self without removing any arguments. 5432 _sp += nargs; 5433 reference_obj = do_null_check(reference_obj, T_OBJECT); 5434 _sp -= nargs;; 5435 5436 if (stopped()) return true; 5437 5438 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5439 5440 ciInstanceKlass* klass = env()->Object_klass(); 5441 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5442 5443 Node* no_ctrl = NULL; 5444 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT); 5445 5446 // Use the pre-barrier to record the value in the referent field 5447 pre_barrier(false /* do_load */, 5448 control(), 5449 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5450 result /* pre_val */, 5451 T_OBJECT); 5452 5453 push(result); 5454 return true; 5455 } 5456