1 /* 2 * Copyright 1999-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_library_call.cpp.incl" 27 28 class LibraryIntrinsic : public InlineCallGenerator { 29 // Extend the set of intrinsics known to the runtime: 30 public: 31 private: 32 bool _is_virtual; 33 vmIntrinsics::ID _intrinsic_id; 34 35 public: 36 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id) 37 : InlineCallGenerator(m), 38 _is_virtual(is_virtual), 39 _intrinsic_id(id) 40 { 41 } 42 virtual bool is_intrinsic() const { return true; } 43 virtual bool is_virtual() const { return _is_virtual; } 44 virtual JVMState* generate(JVMState* jvms); 45 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 46 }; 47 48 49 // Local helper class for LibraryIntrinsic: 50 class LibraryCallKit : public GraphKit { 51 private: 52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 53 54 public: 55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic) 56 : GraphKit(caller), 57 _intrinsic(intrinsic) 58 { 59 } 60 61 ciMethod* caller() const { return jvms()->method(); } 62 int bci() const { return jvms()->bci(); } 63 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 65 ciMethod* callee() const { return _intrinsic->method(); } 66 ciSignature* signature() const { return callee()->signature(); } 67 int arg_size() const { return callee()->arg_size(); } 68 69 bool try_to_inline(); 70 71 // Helper functions to inline natives 72 void push_result(RegionNode* region, PhiNode* value); 73 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 74 Node* generate_slow_guard(Node* test, RegionNode* region); 75 Node* generate_fair_guard(Node* test, RegionNode* region); 76 Node* generate_negative_guard(Node* index, RegionNode* region, 77 // resulting CastII of index: 78 Node* *pos_index = NULL); 79 Node* generate_nonpositive_guard(Node* index, bool never_negative, 80 // resulting CastII of index: 81 Node* *pos_index = NULL); 82 Node* generate_limit_guard(Node* offset, Node* subseq_length, 83 Node* array_length, 84 RegionNode* region); 85 Node* generate_current_thread(Node* &tls_output); 86 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 87 bool disjoint_bases, const char* &name); 88 Node* load_mirror_from_klass(Node* klass); 89 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 90 int nargs, 91 RegionNode* region, int null_path, 92 int offset); 93 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs, 94 RegionNode* region, int null_path) { 95 int offset = java_lang_Class::klass_offset_in_bytes(); 96 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 97 region, null_path, 98 offset); 99 } 100 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 101 int nargs, 102 RegionNode* region, int null_path) { 103 int offset = java_lang_Class::array_klass_offset_in_bytes(); 104 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 105 region, null_path, 106 offset); 107 } 108 Node* generate_access_flags_guard(Node* kls, 109 int modifier_mask, int modifier_bits, 110 RegionNode* region); 111 Node* generate_interface_guard(Node* kls, RegionNode* region); 112 Node* generate_array_guard(Node* kls, RegionNode* region) { 113 return generate_array_guard_common(kls, region, false, false); 114 } 115 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 116 return generate_array_guard_common(kls, region, false, true); 117 } 118 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 119 return generate_array_guard_common(kls, region, true, false); 120 } 121 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 122 return generate_array_guard_common(kls, region, true, true); 123 } 124 Node* generate_array_guard_common(Node* kls, RegionNode* region, 125 bool obj_array, bool not_array); 126 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 127 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 128 bool is_virtual = false, bool is_static = false); 129 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 130 return generate_method_call(method_id, false, true); 131 } 132 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 133 return generate_method_call(method_id, true, false); 134 } 135 136 bool inline_string_compareTo(); 137 bool inline_string_indexOf(); 138 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 139 bool inline_string_equals(); 140 Node* pop_math_arg(); 141 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 142 bool inline_math_native(vmIntrinsics::ID id); 143 bool inline_trig(vmIntrinsics::ID id); 144 bool inline_trans(vmIntrinsics::ID id); 145 bool inline_abs(vmIntrinsics::ID id); 146 bool inline_sqrt(vmIntrinsics::ID id); 147 bool inline_pow(vmIntrinsics::ID id); 148 bool inline_exp(vmIntrinsics::ID id); 149 bool inline_min_max(vmIntrinsics::ID id); 150 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 151 // This returns Type::AnyPtr, RawPtr, or OopPtr. 152 int classify_unsafe_addr(Node* &base, Node* &offset); 153 Node* make_unsafe_address(Node* base, Node* offset); 154 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 155 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 156 bool inline_unsafe_allocate(); 157 bool inline_unsafe_copyMemory(); 158 bool inline_native_currentThread(); 159 bool inline_native_time_funcs(bool isNano); 160 bool inline_native_isInterrupted(); 161 bool inline_native_Class_query(vmIntrinsics::ID id); 162 bool inline_native_subtype_check(); 163 164 bool inline_native_newArray(); 165 bool inline_native_getLength(); 166 bool inline_array_copyOf(bool is_copyOfRange); 167 bool inline_array_equals(); 168 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 169 bool inline_native_clone(bool is_virtual); 170 bool inline_native_Reflection_getCallerClass(); 171 bool inline_native_AtomicLong_get(); 172 bool inline_native_AtomicLong_attemptUpdate(); 173 bool is_method_invoke_or_aux_frame(JVMState* jvms); 174 // Helper function for inlining native object hash method 175 bool inline_native_hashcode(bool is_virtual, bool is_static); 176 bool inline_native_getClass(); 177 178 // Helper functions for inlining arraycopy 179 bool inline_arraycopy(); 180 void generate_arraycopy(const TypePtr* adr_type, 181 BasicType basic_elem_type, 182 Node* src, Node* src_offset, 183 Node* dest, Node* dest_offset, 184 Node* copy_length, 185 bool disjoint_bases = false, 186 bool length_never_negative = false, 187 RegionNode* slow_region = NULL); 188 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 189 RegionNode* slow_region); 190 void generate_clear_array(const TypePtr* adr_type, 191 Node* dest, 192 BasicType basic_elem_type, 193 Node* slice_off, 194 Node* slice_len, 195 Node* slice_end); 196 bool generate_block_arraycopy(const TypePtr* adr_type, 197 BasicType basic_elem_type, 198 AllocateNode* alloc, 199 Node* src, Node* src_offset, 200 Node* dest, Node* dest_offset, 201 Node* dest_size); 202 void generate_slow_arraycopy(const TypePtr* adr_type, 203 Node* src, Node* src_offset, 204 Node* dest, Node* dest_offset, 205 Node* copy_length); 206 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 207 Node* dest_elem_klass, 208 Node* src, Node* src_offset, 209 Node* dest, Node* dest_offset, 210 Node* copy_length); 211 Node* generate_generic_arraycopy(const TypePtr* adr_type, 212 Node* src, Node* src_offset, 213 Node* dest, Node* dest_offset, 214 Node* copy_length); 215 void generate_unchecked_arraycopy(const TypePtr* adr_type, 216 BasicType basic_elem_type, 217 bool disjoint_bases, 218 Node* src, Node* src_offset, 219 Node* dest, Node* dest_offset, 220 Node* copy_length); 221 bool inline_unsafe_CAS(BasicType type); 222 bool inline_unsafe_ordered_store(BasicType type); 223 bool inline_fp_conversions(vmIntrinsics::ID id); 224 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id); 225 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id); 226 bool inline_bitCount(vmIntrinsics::ID id); 227 bool inline_reverseBytes(vmIntrinsics::ID id); 228 }; 229 230 231 //---------------------------make_vm_intrinsic---------------------------- 232 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 233 vmIntrinsics::ID id = m->intrinsic_id(); 234 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 235 236 if (DisableIntrinsic[0] != '\0' 237 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 238 // disabled by a user request on the command line: 239 // example: -XX:DisableIntrinsic=_hashCode,_getClass 240 return NULL; 241 } 242 243 if (!m->is_loaded()) { 244 // do not attempt to inline unloaded methods 245 return NULL; 246 } 247 248 // Only a few intrinsics implement a virtual dispatch. 249 // They are expensive calls which are also frequently overridden. 250 if (is_virtual) { 251 switch (id) { 252 case vmIntrinsics::_hashCode: 253 case vmIntrinsics::_clone: 254 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 255 break; 256 default: 257 return NULL; 258 } 259 } 260 261 // -XX:-InlineNatives disables nearly all intrinsics: 262 if (!InlineNatives) { 263 switch (id) { 264 case vmIntrinsics::_indexOf: 265 case vmIntrinsics::_compareTo: 266 case vmIntrinsics::_equals: 267 case vmIntrinsics::_equalsC: 268 break; // InlineNatives does not control String.compareTo 269 default: 270 return NULL; 271 } 272 } 273 274 switch (id) { 275 case vmIntrinsics::_compareTo: 276 if (!SpecialStringCompareTo) return NULL; 277 break; 278 case vmIntrinsics::_indexOf: 279 if (!SpecialStringIndexOf) return NULL; 280 break; 281 case vmIntrinsics::_equals: 282 if (!SpecialStringEquals) return NULL; 283 break; 284 case vmIntrinsics::_equalsC: 285 if (!SpecialArraysEquals) return NULL; 286 break; 287 case vmIntrinsics::_arraycopy: 288 if (!InlineArrayCopy) return NULL; 289 break; 290 case vmIntrinsics::_copyMemory: 291 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 292 if (!InlineArrayCopy) return NULL; 293 break; 294 case vmIntrinsics::_hashCode: 295 if (!InlineObjectHash) return NULL; 296 break; 297 case vmIntrinsics::_clone: 298 case vmIntrinsics::_copyOf: 299 case vmIntrinsics::_copyOfRange: 300 if (!InlineObjectCopy) return NULL; 301 // These also use the arraycopy intrinsic mechanism: 302 if (!InlineArrayCopy) return NULL; 303 break; 304 case vmIntrinsics::_checkIndex: 305 // We do not intrinsify this. The optimizer does fine with it. 306 return NULL; 307 308 case vmIntrinsics::_get_AtomicLong: 309 case vmIntrinsics::_attemptUpdate: 310 if (!InlineAtomicLong) return NULL; 311 break; 312 313 case vmIntrinsics::_Object_init: 314 case vmIntrinsics::_invoke: 315 // We do not intrinsify these; they are marked for other purposes. 316 return NULL; 317 318 case vmIntrinsics::_getCallerClass: 319 if (!UseNewReflection) return NULL; 320 if (!InlineReflectionGetCallerClass) return NULL; 321 if (!JDK_Version::is_gte_jdk14x_version()) return NULL; 322 break; 323 324 case vmIntrinsics::_bitCount_i: 325 case vmIntrinsics::_bitCount_l: 326 if (!UsePopCountInstruction) return NULL; 327 break; 328 329 default: 330 break; 331 } 332 333 // -XX:-InlineClassNatives disables natives from the Class class. 334 // The flag applies to all reflective calls, notably Array.newArray 335 // (visible to Java programmers as Array.newInstance). 336 if (m->holder()->name() == ciSymbol::java_lang_Class() || 337 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 338 if (!InlineClassNatives) return NULL; 339 } 340 341 // -XX:-InlineThreadNatives disables natives from the Thread class. 342 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 343 if (!InlineThreadNatives) return NULL; 344 } 345 346 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 347 if (m->holder()->name() == ciSymbol::java_lang_Math() || 348 m->holder()->name() == ciSymbol::java_lang_Float() || 349 m->holder()->name() == ciSymbol::java_lang_Double()) { 350 if (!InlineMathNatives) return NULL; 351 } 352 353 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 354 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 355 if (!InlineUnsafeOps) return NULL; 356 } 357 358 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); 359 } 360 361 //----------------------register_library_intrinsics----------------------- 362 // Initialize this file's data structures, for each Compile instance. 363 void Compile::register_library_intrinsics() { 364 // Nothing to do here. 365 } 366 367 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 368 LibraryCallKit kit(jvms, this); 369 Compile* C = kit.C; 370 int nodes = C->unique(); 371 #ifndef PRODUCT 372 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 373 char buf[1000]; 374 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 375 tty->print_cr("Intrinsic %s", str); 376 } 377 #endif 378 if (kit.try_to_inline()) { 379 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 380 tty->print("Inlining intrinsic %s%s at bci:%d in", 381 vmIntrinsics::name_at(intrinsic_id()), 382 (is_virtual() ? " (virtual)" : ""), kit.bci()); 383 kit.caller()->print_short_name(tty); 384 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); 385 } 386 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 387 if (C->log()) { 388 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 389 vmIntrinsics::name_at(intrinsic_id()), 390 (is_virtual() ? " virtual='1'" : ""), 391 C->unique() - nodes); 392 } 393 return kit.transfer_exceptions_into_jvms(); 394 } 395 396 if (PrintIntrinsics) { 397 switch (intrinsic_id()) { 398 case vmIntrinsics::_invoke: 399 case vmIntrinsics::_Object_init: 400 // We do not expect to inline these, so do not produce any noise about them. 401 break; 402 default: 403 tty->print("Did not inline intrinsic %s%s at bci:%d in", 404 vmIntrinsics::name_at(intrinsic_id()), 405 (is_virtual() ? " (virtual)" : ""), kit.bci()); 406 kit.caller()->print_short_name(tty); 407 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); 408 } 409 } 410 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 411 return NULL; 412 } 413 414 bool LibraryCallKit::try_to_inline() { 415 // Handle symbolic names for otherwise undistinguished boolean switches: 416 const bool is_store = true; 417 const bool is_native_ptr = true; 418 const bool is_static = true; 419 420 switch (intrinsic_id()) { 421 case vmIntrinsics::_hashCode: 422 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 423 case vmIntrinsics::_identityHashCode: 424 return inline_native_hashcode(/*!virtual*/ false, is_static); 425 case vmIntrinsics::_getClass: 426 return inline_native_getClass(); 427 428 case vmIntrinsics::_dsin: 429 case vmIntrinsics::_dcos: 430 case vmIntrinsics::_dtan: 431 case vmIntrinsics::_dabs: 432 case vmIntrinsics::_datan2: 433 case vmIntrinsics::_dsqrt: 434 case vmIntrinsics::_dexp: 435 case vmIntrinsics::_dlog: 436 case vmIntrinsics::_dlog10: 437 case vmIntrinsics::_dpow: 438 return inline_math_native(intrinsic_id()); 439 440 case vmIntrinsics::_min: 441 case vmIntrinsics::_max: 442 return inline_min_max(intrinsic_id()); 443 444 case vmIntrinsics::_arraycopy: 445 return inline_arraycopy(); 446 447 case vmIntrinsics::_compareTo: 448 return inline_string_compareTo(); 449 case vmIntrinsics::_indexOf: 450 return inline_string_indexOf(); 451 case vmIntrinsics::_equals: 452 return inline_string_equals(); 453 454 case vmIntrinsics::_getObject: 455 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); 456 case vmIntrinsics::_getBoolean: 457 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); 458 case vmIntrinsics::_getByte: 459 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); 460 case vmIntrinsics::_getShort: 461 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); 462 case vmIntrinsics::_getChar: 463 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); 464 case vmIntrinsics::_getInt: 465 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); 466 case vmIntrinsics::_getLong: 467 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); 468 case vmIntrinsics::_getFloat: 469 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); 470 case vmIntrinsics::_getDouble: 471 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); 472 473 case vmIntrinsics::_putObject: 474 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); 475 case vmIntrinsics::_putBoolean: 476 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); 477 case vmIntrinsics::_putByte: 478 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); 479 case vmIntrinsics::_putShort: 480 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); 481 case vmIntrinsics::_putChar: 482 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); 483 case vmIntrinsics::_putInt: 484 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); 485 case vmIntrinsics::_putLong: 486 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); 487 case vmIntrinsics::_putFloat: 488 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); 489 case vmIntrinsics::_putDouble: 490 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); 491 492 case vmIntrinsics::_getByte_raw: 493 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); 494 case vmIntrinsics::_getShort_raw: 495 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); 496 case vmIntrinsics::_getChar_raw: 497 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); 498 case vmIntrinsics::_getInt_raw: 499 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); 500 case vmIntrinsics::_getLong_raw: 501 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); 502 case vmIntrinsics::_getFloat_raw: 503 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); 504 case vmIntrinsics::_getDouble_raw: 505 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); 506 case vmIntrinsics::_getAddress_raw: 507 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); 508 509 case vmIntrinsics::_putByte_raw: 510 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); 511 case vmIntrinsics::_putShort_raw: 512 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); 513 case vmIntrinsics::_putChar_raw: 514 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); 515 case vmIntrinsics::_putInt_raw: 516 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); 517 case vmIntrinsics::_putLong_raw: 518 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); 519 case vmIntrinsics::_putFloat_raw: 520 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); 521 case vmIntrinsics::_putDouble_raw: 522 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); 523 case vmIntrinsics::_putAddress_raw: 524 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); 525 526 case vmIntrinsics::_getObjectVolatile: 527 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); 528 case vmIntrinsics::_getBooleanVolatile: 529 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); 530 case vmIntrinsics::_getByteVolatile: 531 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); 532 case vmIntrinsics::_getShortVolatile: 533 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); 534 case vmIntrinsics::_getCharVolatile: 535 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); 536 case vmIntrinsics::_getIntVolatile: 537 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); 538 case vmIntrinsics::_getLongVolatile: 539 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); 540 case vmIntrinsics::_getFloatVolatile: 541 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); 542 case vmIntrinsics::_getDoubleVolatile: 543 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); 544 545 case vmIntrinsics::_putObjectVolatile: 546 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); 547 case vmIntrinsics::_putBooleanVolatile: 548 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); 549 case vmIntrinsics::_putByteVolatile: 550 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); 551 case vmIntrinsics::_putShortVolatile: 552 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); 553 case vmIntrinsics::_putCharVolatile: 554 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); 555 case vmIntrinsics::_putIntVolatile: 556 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); 557 case vmIntrinsics::_putLongVolatile: 558 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); 559 case vmIntrinsics::_putFloatVolatile: 560 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); 561 case vmIntrinsics::_putDoubleVolatile: 562 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); 563 564 case vmIntrinsics::_prefetchRead: 565 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 566 case vmIntrinsics::_prefetchWrite: 567 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 568 case vmIntrinsics::_prefetchReadStatic: 569 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 570 case vmIntrinsics::_prefetchWriteStatic: 571 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 572 573 case vmIntrinsics::_compareAndSwapObject: 574 return inline_unsafe_CAS(T_OBJECT); 575 case vmIntrinsics::_compareAndSwapInt: 576 return inline_unsafe_CAS(T_INT); 577 case vmIntrinsics::_compareAndSwapLong: 578 return inline_unsafe_CAS(T_LONG); 579 580 case vmIntrinsics::_putOrderedObject: 581 return inline_unsafe_ordered_store(T_OBJECT); 582 case vmIntrinsics::_putOrderedInt: 583 return inline_unsafe_ordered_store(T_INT); 584 case vmIntrinsics::_putOrderedLong: 585 return inline_unsafe_ordered_store(T_LONG); 586 587 case vmIntrinsics::_currentThread: 588 return inline_native_currentThread(); 589 case vmIntrinsics::_isInterrupted: 590 return inline_native_isInterrupted(); 591 592 case vmIntrinsics::_currentTimeMillis: 593 return inline_native_time_funcs(false); 594 case vmIntrinsics::_nanoTime: 595 return inline_native_time_funcs(true); 596 case vmIntrinsics::_allocateInstance: 597 return inline_unsafe_allocate(); 598 case vmIntrinsics::_copyMemory: 599 return inline_unsafe_copyMemory(); 600 case vmIntrinsics::_newArray: 601 return inline_native_newArray(); 602 case vmIntrinsics::_getLength: 603 return inline_native_getLength(); 604 case vmIntrinsics::_copyOf: 605 return inline_array_copyOf(false); 606 case vmIntrinsics::_copyOfRange: 607 return inline_array_copyOf(true); 608 case vmIntrinsics::_equalsC: 609 return inline_array_equals(); 610 case vmIntrinsics::_clone: 611 return inline_native_clone(intrinsic()->is_virtual()); 612 613 case vmIntrinsics::_isAssignableFrom: 614 return inline_native_subtype_check(); 615 616 case vmIntrinsics::_isInstance: 617 case vmIntrinsics::_getModifiers: 618 case vmIntrinsics::_isInterface: 619 case vmIntrinsics::_isArray: 620 case vmIntrinsics::_isPrimitive: 621 case vmIntrinsics::_getSuperclass: 622 case vmIntrinsics::_getComponentType: 623 case vmIntrinsics::_getClassAccessFlags: 624 return inline_native_Class_query(intrinsic_id()); 625 626 case vmIntrinsics::_floatToRawIntBits: 627 case vmIntrinsics::_floatToIntBits: 628 case vmIntrinsics::_intBitsToFloat: 629 case vmIntrinsics::_doubleToRawLongBits: 630 case vmIntrinsics::_doubleToLongBits: 631 case vmIntrinsics::_longBitsToDouble: 632 return inline_fp_conversions(intrinsic_id()); 633 634 case vmIntrinsics::_numberOfLeadingZeros_i: 635 case vmIntrinsics::_numberOfLeadingZeros_l: 636 return inline_numberOfLeadingZeros(intrinsic_id()); 637 638 case vmIntrinsics::_numberOfTrailingZeros_i: 639 case vmIntrinsics::_numberOfTrailingZeros_l: 640 return inline_numberOfTrailingZeros(intrinsic_id()); 641 642 case vmIntrinsics::_bitCount_i: 643 case vmIntrinsics::_bitCount_l: 644 return inline_bitCount(intrinsic_id()); 645 646 case vmIntrinsics::_reverseBytes_i: 647 case vmIntrinsics::_reverseBytes_l: 648 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); 649 650 case vmIntrinsics::_get_AtomicLong: 651 return inline_native_AtomicLong_get(); 652 case vmIntrinsics::_attemptUpdate: 653 return inline_native_AtomicLong_attemptUpdate(); 654 655 case vmIntrinsics::_getCallerClass: 656 return inline_native_Reflection_getCallerClass(); 657 658 default: 659 // If you get here, it may be that someone has added a new intrinsic 660 // to the list in vmSymbols.hpp without implementing it here. 661 #ifndef PRODUCT 662 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 663 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 664 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 665 } 666 #endif 667 return false; 668 } 669 } 670 671 //------------------------------push_result------------------------------ 672 // Helper function for finishing intrinsics. 673 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { 674 record_for_igvn(region); 675 set_control(_gvn.transform(region)); 676 BasicType value_type = value->type()->basic_type(); 677 push_node(value_type, _gvn.transform(value)); 678 } 679 680 //------------------------------generate_guard--------------------------- 681 // Helper function for generating guarded fast-slow graph structures. 682 // The given 'test', if true, guards a slow path. If the test fails 683 // then a fast path can be taken. (We generally hope it fails.) 684 // In all cases, GraphKit::control() is updated to the fast path. 685 // The returned value represents the control for the slow path. 686 // The return value is never 'top'; it is either a valid control 687 // or NULL if it is obvious that the slow path can never be taken. 688 // Also, if region and the slow control are not NULL, the slow edge 689 // is appended to the region. 690 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 691 if (stopped()) { 692 // Already short circuited. 693 return NULL; 694 } 695 696 // Build an if node and its projections. 697 // If test is true we take the slow path, which we assume is uncommon. 698 if (_gvn.type(test) == TypeInt::ZERO) { 699 // The slow branch is never taken. No need to build this guard. 700 return NULL; 701 } 702 703 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 704 705 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); 706 if (if_slow == top()) { 707 // The slow branch is never taken. No need to build this guard. 708 return NULL; 709 } 710 711 if (region != NULL) 712 region->add_req(if_slow); 713 714 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); 715 set_control(if_fast); 716 717 return if_slow; 718 } 719 720 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 721 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 722 } 723 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 724 return generate_guard(test, region, PROB_FAIR); 725 } 726 727 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 728 Node* *pos_index) { 729 if (stopped()) 730 return NULL; // already stopped 731 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 732 return NULL; // index is already adequately typed 733 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 734 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 735 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 736 if (is_neg != NULL && pos_index != NULL) { 737 // Emulate effect of Parse::adjust_map_after_if. 738 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); 739 ccast->set_req(0, control()); 740 (*pos_index) = _gvn.transform(ccast); 741 } 742 return is_neg; 743 } 744 745 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 746 Node* *pos_index) { 747 if (stopped()) 748 return NULL; // already stopped 749 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 750 return NULL; // index is already adequately typed 751 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 752 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 753 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) ); 754 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 755 if (is_notp != NULL && pos_index != NULL) { 756 // Emulate effect of Parse::adjust_map_after_if. 757 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1); 758 ccast->set_req(0, control()); 759 (*pos_index) = _gvn.transform(ccast); 760 } 761 return is_notp; 762 } 763 764 // Make sure that 'position' is a valid limit index, in [0..length]. 765 // There are two equivalent plans for checking this: 766 // A. (offset + copyLength) unsigned<= arrayLength 767 // B. offset <= (arrayLength - copyLength) 768 // We require that all of the values above, except for the sum and 769 // difference, are already known to be non-negative. 770 // Plan A is robust in the face of overflow, if offset and copyLength 771 // are both hugely positive. 772 // 773 // Plan B is less direct and intuitive, but it does not overflow at 774 // all, since the difference of two non-negatives is always 775 // representable. Whenever Java methods must perform the equivalent 776 // check they generally use Plan B instead of Plan A. 777 // For the moment we use Plan A. 778 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 779 Node* subseq_length, 780 Node* array_length, 781 RegionNode* region) { 782 if (stopped()) 783 return NULL; // already stopped 784 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 785 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length)) 786 return NULL; // common case of whole-array copy 787 Node* last = subseq_length; 788 if (!zero_offset) // last += offset 789 last = _gvn.transform( new (C, 3) AddINode(last, offset)); 790 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); 791 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 792 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 793 return is_over; 794 } 795 796 797 //--------------------------generate_current_thread-------------------- 798 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 799 ciKlass* thread_klass = env()->Thread_klass(); 800 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 801 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); 802 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 803 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 804 tls_output = thread; 805 return threadObj; 806 } 807 808 809 //------------------------------inline_string_compareTo------------------------ 810 bool LibraryCallKit::inline_string_compareTo() { 811 812 if (!Matcher::has_match_rule(Op_StrComp)) return false; 813 814 const int value_offset = java_lang_String::value_offset_in_bytes(); 815 const int count_offset = java_lang_String::count_offset_in_bytes(); 816 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 817 818 _sp += 2; 819 Node *argument = pop(); // pop non-receiver first: it was pushed second 820 Node *receiver = pop(); 821 822 // Null check on self without removing any arguments. The argument 823 // null check technically happens in the wrong place, which can lead to 824 // invalid stack traces when string compare is inlined into a method 825 // which handles NullPointerExceptions. 826 _sp += 2; 827 receiver = do_null_check(receiver, T_OBJECT); 828 argument = do_null_check(argument, T_OBJECT); 829 _sp -= 2; 830 if (stopped()) { 831 return true; 832 } 833 834 ciInstanceKlass* klass = env()->String_klass(); 835 const TypeInstPtr* string_type = 836 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 837 838 Node* compare = 839 _gvn.transform(new (C, 7) StrCompNode( 840 control(), 841 memory(TypeAryPtr::CHARS), 842 memory(string_type->add_offset(value_offset)), 843 memory(string_type->add_offset(count_offset)), 844 memory(string_type->add_offset(offset_offset)), 845 receiver, 846 argument)); 847 push(compare); 848 return true; 849 } 850 851 //------------------------------inline_string_equals------------------------ 852 bool LibraryCallKit::inline_string_equals() { 853 854 if (!Matcher::has_match_rule(Op_StrEquals)) return false; 855 856 const int value_offset = java_lang_String::value_offset_in_bytes(); 857 const int count_offset = java_lang_String::count_offset_in_bytes(); 858 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 859 860 _sp += 2; 861 Node* argument = pop(); // pop non-receiver first: it was pushed second 862 Node* receiver = pop(); 863 864 // Null check on self without removing any arguments. The argument 865 // null check technically happens in the wrong place, which can lead to 866 // invalid stack traces when string compare is inlined into a method 867 // which handles NullPointerExceptions. 868 _sp += 2; 869 receiver = do_null_check(receiver, T_OBJECT); 870 //should not do null check for argument for String.equals(), because spec 871 //allows to specify NULL as argument. 872 _sp -= 2; 873 874 if (stopped()) { 875 return true; 876 } 877 878 // get String klass for instanceOf 879 ciInstanceKlass* klass = env()->String_klass(); 880 881 // two paths (plus control) merge 882 RegionNode* region = new (C, 3) RegionNode(3); 883 Node* phi = new (C, 3) PhiNode(region, TypeInt::BOOL); 884 885 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 886 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1))); 887 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq)); 888 889 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 890 891 Node* if_true = _gvn.transform(new (C, 1) IfTrueNode(iff)); 892 set_control(if_true); 893 894 const TypeInstPtr* string_type = 895 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 896 897 // instanceOf == true 898 Node* equals = 899 _gvn.transform(new (C, 7) StrEqualsNode( 900 control(), 901 memory(TypeAryPtr::CHARS), 902 memory(string_type->add_offset(value_offset)), 903 memory(string_type->add_offset(count_offset)), 904 memory(string_type->add_offset(offset_offset)), 905 receiver, 906 argument)); 907 908 phi->init_req(1, _gvn.transform(equals)); 909 region->init_req(1, if_true); 910 911 //instanceOf == false, fallthrough 912 Node* if_false = _gvn.transform(new (C, 1) IfFalseNode(iff)); 913 set_control(if_false); 914 915 phi->init_req(2, _gvn.transform(intcon(0))); 916 region->init_req(2, if_false); 917 918 // post merge 919 set_control(_gvn.transform(region)); 920 record_for_igvn(region); 921 922 push(_gvn.transform(phi)); 923 924 return true; 925 } 926 927 //------------------------------inline_array_equals---------------------------- 928 bool LibraryCallKit::inline_array_equals() { 929 930 if (!Matcher::has_match_rule(Op_AryEq)) return false; 931 932 _sp += 2; 933 Node *argument2 = pop(); 934 Node *argument1 = pop(); 935 936 Node* equals = 937 _gvn.transform(new (C, 3) AryEqNode(control(), 938 argument1, 939 argument2) 940 ); 941 push(equals); 942 return true; 943 } 944 945 // Java version of String.indexOf(constant string) 946 // class StringDecl { 947 // StringDecl(char[] ca) { 948 // offset = 0; 949 // count = ca.length; 950 // value = ca; 951 // } 952 // int offset; 953 // int count; 954 // char[] value; 955 // } 956 // 957 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 958 // int targetOffset, int cache_i, int md2) { 959 // int cache = cache_i; 960 // int sourceOffset = string_object.offset; 961 // int sourceCount = string_object.count; 962 // int targetCount = target_object.length; 963 // 964 // int targetCountLess1 = targetCount - 1; 965 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 966 // 967 // char[] source = string_object.value; 968 // char[] target = target_object; 969 // int lastChar = target[targetCountLess1]; 970 // 971 // outer_loop: 972 // for (int i = sourceOffset; i < sourceEnd; ) { 973 // int src = source[i + targetCountLess1]; 974 // if (src == lastChar) { 975 // // With random strings and a 4-character alphabet, 976 // // reverse matching at this point sets up 0.8% fewer 977 // // frames, but (paradoxically) makes 0.3% more probes. 978 // // Since those probes are nearer the lastChar probe, 979 // // there is may be a net D$ win with reverse matching. 980 // // But, reversing loop inhibits unroll of inner loop 981 // // for unknown reason. So, does running outer loop from 982 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 983 // for (int j = 0; j < targetCountLess1; j++) { 984 // if (target[targetOffset + j] != source[i+j]) { 985 // if ((cache & (1 << source[i+j])) == 0) { 986 // if (md2 < j+1) { 987 // i += j+1; 988 // continue outer_loop; 989 // } 990 // } 991 // i += md2; 992 // continue outer_loop; 993 // } 994 // } 995 // return i - sourceOffset; 996 // } 997 // if ((cache & (1 << src)) == 0) { 998 // i += targetCountLess1; 999 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1000 // i++; 1001 // } 1002 // return -1; 1003 // } 1004 1005 //------------------------------string_indexOf------------------------ 1006 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1007 jint cache_i, jint md2_i) { 1008 1009 Node* no_ctrl = NULL; 1010 float likely = PROB_LIKELY(0.9); 1011 float unlikely = PROB_UNLIKELY(0.9); 1012 1013 const int value_offset = java_lang_String::value_offset_in_bytes(); 1014 const int count_offset = java_lang_String::count_offset_in_bytes(); 1015 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1016 1017 ciInstanceKlass* klass = env()->String_klass(); 1018 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 1019 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0); 1020 1021 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset); 1022 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 1023 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset); 1024 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1025 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset); 1026 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset)); 1027 1028 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) ); 1029 jint target_length = target_array->length(); 1030 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1031 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1032 1033 IdealKit kit(gvn(), control(), merged_memory()); 1034 #define __ kit. 1035 Node* zero = __ ConI(0); 1036 Node* one = __ ConI(1); 1037 Node* cache = __ ConI(cache_i); 1038 Node* md2 = __ ConI(md2_i); 1039 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1040 Node* targetCount = __ ConI(target_length); 1041 Node* targetCountLess1 = __ ConI(target_length - 1); 1042 Node* targetOffset = __ ConI(targetOffset_i); 1043 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1044 1045 IdealVariable rtn(kit), i(kit), j(kit); __ declares_done(); 1046 Node* outer_loop = __ make_label(2 /* goto */); 1047 Node* return_ = __ make_label(1); 1048 1049 __ set(rtn,__ ConI(-1)); 1050 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); { 1051 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1052 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1053 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1054 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1055 __ loop(j, zero, BoolTest::lt, targetCountLess1); { 1056 Node* tpj = __ AddI(targetOffset, __ value(j)); 1057 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1058 Node* ipj = __ AddI(__ value(i), __ value(j)); 1059 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1060 __ if_then(targ, BoolTest::ne, src2); { 1061 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1062 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1063 __ increment(i, __ AddI(__ value(j), one)); 1064 __ goto_(outer_loop); 1065 } __ end_if(); __ dead(j); 1066 }__ end_if(); __ dead(j); 1067 __ increment(i, md2); 1068 __ goto_(outer_loop); 1069 }__ end_if(); 1070 __ increment(j, one); 1071 }__ end_loop(); __ dead(j); 1072 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1073 __ goto_(return_); 1074 }__ end_if(); 1075 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1076 __ increment(i, targetCountLess1); 1077 }__ end_if(); 1078 __ increment(i, one); 1079 __ bind(outer_loop); 1080 }__ end_loop(); __ dead(i); 1081 __ bind(return_); 1082 __ drain_delay_transform(); 1083 1084 set_control(__ ctrl()); 1085 Node* result = __ value(rtn); 1086 #undef __ 1087 C->set_has_loops(true); 1088 return result; 1089 } 1090 1091 //------------------------------inline_string_indexOf------------------------ 1092 bool LibraryCallKit::inline_string_indexOf() { 1093 1094 const int value_offset = java_lang_String::value_offset_in_bytes(); 1095 const int count_offset = java_lang_String::count_offset_in_bytes(); 1096 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1097 1098 _sp += 2; 1099 Node *argument = pop(); // pop non-receiver first: it was pushed second 1100 Node *receiver = pop(); 1101 1102 Node* result; 1103 if (Matcher::has_match_rule(Op_StrIndexOf) && 1104 UseSSE42Intrinsics) { 1105 // Generate SSE4.2 version of indexOf 1106 // We currently only have match rules that use SSE4.2 1107 1108 // Null check on self without removing any arguments. The argument 1109 // null check technically happens in the wrong place, which can lead to 1110 // invalid stack traces when string compare is inlined into a method 1111 // which handles NullPointerExceptions. 1112 _sp += 2; 1113 receiver = do_null_check(receiver, T_OBJECT); 1114 argument = do_null_check(argument, T_OBJECT); 1115 _sp -= 2; 1116 1117 if (stopped()) { 1118 return true; 1119 } 1120 1121 ciInstanceKlass* klass = env()->String_klass(); 1122 const TypeInstPtr* string_type = 1123 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 1124 1125 result = 1126 _gvn.transform(new (C, 7) 1127 StrIndexOfNode(control(), 1128 memory(TypeAryPtr::CHARS), 1129 memory(string_type->add_offset(value_offset)), 1130 memory(string_type->add_offset(count_offset)), 1131 memory(string_type->add_offset(offset_offset)), 1132 receiver, 1133 argument)); 1134 } else { //Use LibraryCallKit::string_indexOf 1135 // don't intrinsify is argument isn't a constant string. 1136 if (!argument->is_Con()) { 1137 return false; 1138 } 1139 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); 1140 if (str_type == NULL) { 1141 return false; 1142 } 1143 ciInstanceKlass* klass = env()->String_klass(); 1144 ciObject* str_const = str_type->const_oop(); 1145 if (str_const == NULL || str_const->klass() != klass) { 1146 return false; 1147 } 1148 ciInstance* str = str_const->as_instance(); 1149 assert(str != NULL, "must be instance"); 1150 1151 ciObject* v = str->field_value_by_offset(value_offset).as_object(); 1152 int o = str->field_value_by_offset(offset_offset).as_int(); 1153 int c = str->field_value_by_offset(count_offset).as_int(); 1154 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1155 1156 // constant strings have no offset and count == length which 1157 // simplifies the resulting code somewhat so lets optimize for that. 1158 if (o != 0 || c != pat->length()) { 1159 return false; 1160 } 1161 1162 // Null check on self without removing any arguments. The argument 1163 // null check technically happens in the wrong place, which can lead to 1164 // invalid stack traces when string compare is inlined into a method 1165 // which handles NullPointerExceptions. 1166 _sp += 2; 1167 receiver = do_null_check(receiver, T_OBJECT); 1168 // No null check on the argument is needed since it's a constant String oop. 1169 _sp -= 2; 1170 if (stopped()) { 1171 return true; 1172 } 1173 1174 // The null string as a pattern always returns 0 (match at beginning of string) 1175 if (c == 0) { 1176 push(intcon(0)); 1177 return true; 1178 } 1179 1180 // Generate default indexOf 1181 jchar lastChar = pat->char_at(o + (c - 1)); 1182 int cache = 0; 1183 int i; 1184 for (i = 0; i < c - 1; i++) { 1185 assert(i < pat->length(), "out of range"); 1186 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1187 } 1188 1189 int md2 = c; 1190 for (i = 0; i < c - 1; i++) { 1191 assert(i < pat->length(), "out of range"); 1192 if (pat->char_at(o + i) == lastChar) { 1193 md2 = (c - 1) - i; 1194 } 1195 } 1196 1197 result = string_indexOf(receiver, pat, o, cache, md2); 1198 } 1199 1200 push(result); 1201 return true; 1202 } 1203 1204 //--------------------------pop_math_arg-------------------------------- 1205 // Pop a double argument to a math function from the stack 1206 // rounding it if necessary. 1207 Node * LibraryCallKit::pop_math_arg() { 1208 Node *arg = pop_pair(); 1209 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) 1210 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); 1211 return arg; 1212 } 1213 1214 //------------------------------inline_trig---------------------------------- 1215 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1216 // argument reduction which will turn into a fast/slow diamond. 1217 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1218 _sp += arg_size(); // restore stack pointer 1219 Node* arg = pop_math_arg(); 1220 Node* trig = NULL; 1221 1222 switch (id) { 1223 case vmIntrinsics::_dsin: 1224 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); 1225 break; 1226 case vmIntrinsics::_dcos: 1227 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); 1228 break; 1229 case vmIntrinsics::_dtan: 1230 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); 1231 break; 1232 default: 1233 assert(false, "bad intrinsic was passed in"); 1234 return false; 1235 } 1236 1237 // Rounding required? Check for argument reduction! 1238 if( Matcher::strict_fp_requires_explicit_rounding ) { 1239 1240 static const double pi_4 = 0.7853981633974483; 1241 static const double neg_pi_4 = -0.7853981633974483; 1242 // pi/2 in 80-bit extended precision 1243 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1244 // -pi/2 in 80-bit extended precision 1245 // 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}; 1246 // Cutoff value for using this argument reduction technique 1247 //static const double pi_2_minus_epsilon = 1.564660403643354; 1248 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1249 1250 // Pseudocode for sin: 1251 // if (x <= Math.PI / 4.0) { 1252 // if (x >= -Math.PI / 4.0) return fsin(x); 1253 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1254 // } else { 1255 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1256 // } 1257 // return StrictMath.sin(x); 1258 1259 // Pseudocode for cos: 1260 // if (x <= Math.PI / 4.0) { 1261 // if (x >= -Math.PI / 4.0) return fcos(x); 1262 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1263 // } else { 1264 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1265 // } 1266 // return StrictMath.cos(x); 1267 1268 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1269 // requires a special machine instruction to load it. Instead we'll try 1270 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1271 // probably do the math inside the SIN encoding. 1272 1273 // Make the merge point 1274 RegionNode *r = new (C, 3) RegionNode(3); 1275 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); 1276 1277 // Flatten arg so we need only 1 test 1278 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); 1279 // Node for PI/4 constant 1280 Node *pi4 = makecon(TypeD::make(pi_4)); 1281 // Check PI/4 : abs(arg) 1282 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); 1283 // Check: If PI/4 < abs(arg) then go slow 1284 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); 1285 // Branch either way 1286 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1287 set_control(opt_iff(r,iff)); 1288 1289 // Set fast path result 1290 phi->init_req(2,trig); 1291 1292 // Slow path - non-blocking leaf call 1293 Node* call = NULL; 1294 switch (id) { 1295 case vmIntrinsics::_dsin: 1296 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1297 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1298 "Sin", NULL, arg, top()); 1299 break; 1300 case vmIntrinsics::_dcos: 1301 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1302 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1303 "Cos", NULL, arg, top()); 1304 break; 1305 case vmIntrinsics::_dtan: 1306 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1307 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1308 "Tan", NULL, arg, top()); 1309 break; 1310 } 1311 assert(control()->in(0) == call, ""); 1312 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); 1313 r->init_req(1,control()); 1314 phi->init_req(1,slow_result); 1315 1316 // Post-merge 1317 set_control(_gvn.transform(r)); 1318 record_for_igvn(r); 1319 trig = _gvn.transform(phi); 1320 1321 C->set_has_split_ifs(true); // Has chance for split-if optimization 1322 } 1323 // Push result back on JVM stack 1324 push_pair(trig); 1325 return true; 1326 } 1327 1328 //------------------------------inline_sqrt------------------------------------- 1329 // Inline square root instruction, if possible. 1330 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { 1331 assert(id == vmIntrinsics::_dsqrt, "Not square root"); 1332 _sp += arg_size(); // restore stack pointer 1333 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); 1334 return true; 1335 } 1336 1337 //------------------------------inline_abs------------------------------------- 1338 // Inline absolute value instruction, if possible. 1339 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { 1340 assert(id == vmIntrinsics::_dabs, "Not absolute value"); 1341 _sp += arg_size(); // restore stack pointer 1342 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); 1343 return true; 1344 } 1345 1346 //------------------------------inline_exp------------------------------------- 1347 // Inline exp instructions, if possible. The Intel hardware only misses 1348 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1349 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { 1350 assert(id == vmIntrinsics::_dexp, "Not exp"); 1351 1352 // If this inlining ever returned NaN in the past, we do not intrinsify it 1353 // every again. NaN results requires StrictMath.exp handling. 1354 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1355 1356 // Do not intrinsify on older platforms which lack cmove. 1357 if (ConditionalMoveLimit == 0) return false; 1358 1359 _sp += arg_size(); // restore stack pointer 1360 Node *x = pop_math_arg(); 1361 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); 1362 1363 //------------------- 1364 //result=(result.isNaN())? StrictMath::exp():result; 1365 // Check: If isNaN() by checking result!=result? then go to Strict Math 1366 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1367 // Build the boolean node 1368 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1369 1370 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1371 // End the current control-flow path 1372 push_pair(x); 1373 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp 1374 // to handle. Recompile without intrinsifying Math.exp 1375 uncommon_trap(Deoptimization::Reason_intrinsic, 1376 Deoptimization::Action_make_not_entrant); 1377 } 1378 1379 C->set_has_split_ifs(true); // Has chance for split-if optimization 1380 1381 push_pair(result); 1382 1383 return true; 1384 } 1385 1386 //------------------------------inline_pow------------------------------------- 1387 // Inline power instructions, if possible. 1388 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { 1389 assert(id == vmIntrinsics::_dpow, "Not pow"); 1390 1391 // If this inlining ever returned NaN in the past, we do not intrinsify it 1392 // every again. NaN results requires StrictMath.pow handling. 1393 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1394 1395 // Do not intrinsify on older platforms which lack cmove. 1396 if (ConditionalMoveLimit == 0) return false; 1397 1398 // Pseudocode for pow 1399 // if (x <= 0.0) { 1400 // if ((double)((int)y)==y) { // if y is int 1401 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y) 1402 // } else { 1403 // result = NaN; 1404 // } 1405 // } else { 1406 // result = DPow(x,y); 1407 // } 1408 // if (result != result)? { 1409 // uncommon_trap(); 1410 // } 1411 // return result; 1412 1413 _sp += arg_size(); // restore stack pointer 1414 Node* y = pop_math_arg(); 1415 Node* x = pop_math_arg(); 1416 1417 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); 1418 1419 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow 1420 // inside of something) then skip the fancy tests and just check for 1421 // NaN result. 1422 Node *result = NULL; 1423 if( jvms()->depth() >= 1 ) { 1424 result = fast_result; 1425 } else { 1426 1427 // Set the merge point for If node with condition of (x <= 0.0) 1428 // There are four possible paths to region node and phi node 1429 RegionNode *r = new (C, 4) RegionNode(4); 1430 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); 1431 1432 // Build the first if node: if (x <= 0.0) 1433 // Node for 0 constant 1434 Node *zeronode = makecon(TypeD::ZERO); 1435 // Check x:0 1436 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); 1437 // Check: If (x<=0) then go complex path 1438 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); 1439 // Branch either way 1440 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1441 Node *opt_test = _gvn.transform(if1); 1442 //assert( opt_test->is_If(), "Expect an IfNode"); 1443 IfNode *opt_if1 = (IfNode*)opt_test; 1444 // Fast path taken; set region slot 3 1445 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) ); 1446 r->init_req(3,fast_taken); // Capture fast-control 1447 1448 // Fast path not-taken, i.e. slow path 1449 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) ); 1450 1451 // Set fast path result 1452 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) ); 1453 phi->init_req(3, fast_result); 1454 1455 // Complex path 1456 // Build the second if node (if y is int) 1457 // Node for (int)y 1458 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y)); 1459 // Node for (double)((int) y) 1460 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty)); 1461 // Check (double)((int) y) : y 1462 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y)); 1463 // Check if (y isn't int) then go to slow path 1464 1465 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) ); 1466 // Branch either way 1467 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1468 Node *slow_path = opt_iff(r,if2); // Set region path 2 1469 1470 // Calculate DPow(abs(x), y)*(1 & (int)y) 1471 // Node for constant 1 1472 Node *conone = intcon(1); 1473 // 1& (int)y 1474 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) ); 1475 // zero node 1476 Node *conzero = intcon(0); 1477 // Check (1&(int)y)==0? 1478 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero)); 1479 // Check if (1&(int)y)!=0?, if so the result is negative 1480 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); 1481 // abs(x) 1482 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); 1483 // abs(x)^y 1484 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) ); 1485 // -abs(x)^y 1486 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); 1487 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1488 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1489 // Set complex path fast result 1490 phi->init_req(2, signresult); 1491 1492 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1493 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1494 r->init_req(1,slow_path); 1495 phi->init_req(1,slow_result); 1496 1497 // Post merge 1498 set_control(_gvn.transform(r)); 1499 record_for_igvn(r); 1500 result=_gvn.transform(phi); 1501 } 1502 1503 //------------------- 1504 //result=(result.isNaN())? uncommon_trap():result; 1505 // Check: If isNaN() by checking result!=result? then go to Strict Math 1506 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1507 // Build the boolean node 1508 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1509 1510 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1511 // End the current control-flow path 1512 push_pair(x); 1513 push_pair(y); 1514 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow 1515 // to handle. Recompile without intrinsifying Math.pow. 1516 uncommon_trap(Deoptimization::Reason_intrinsic, 1517 Deoptimization::Action_make_not_entrant); 1518 } 1519 1520 C->set_has_split_ifs(true); // Has chance for split-if optimization 1521 1522 push_pair(result); 1523 1524 return true; 1525 } 1526 1527 //------------------------------inline_trans------------------------------------- 1528 // Inline transcendental instructions, if possible. The Intel hardware gets 1529 // these right, no funny corner cases missed. 1530 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { 1531 _sp += arg_size(); // restore stack pointer 1532 Node* arg = pop_math_arg(); 1533 Node* trans = NULL; 1534 1535 switch (id) { 1536 case vmIntrinsics::_dlog: 1537 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); 1538 break; 1539 case vmIntrinsics::_dlog10: 1540 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); 1541 break; 1542 default: 1543 assert(false, "bad intrinsic was passed in"); 1544 return false; 1545 } 1546 1547 // Push result back on JVM stack 1548 push_pair(trans); 1549 return true; 1550 } 1551 1552 //------------------------------runtime_math----------------------------- 1553 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1554 Node* a = NULL; 1555 Node* b = NULL; 1556 1557 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1558 "must be (DD)D or (D)D type"); 1559 1560 // Inputs 1561 _sp += arg_size(); // restore stack pointer 1562 if (call_type == OptoRuntime::Math_DD_D_Type()) { 1563 b = pop_math_arg(); 1564 } 1565 a = pop_math_arg(); 1566 1567 const TypePtr* no_memory_effects = NULL; 1568 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1569 no_memory_effects, 1570 a, top(), b, b ? top() : NULL); 1571 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); 1572 #ifdef ASSERT 1573 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); 1574 assert(value_top == top(), "second value must be top"); 1575 #endif 1576 1577 push_pair(value); 1578 return true; 1579 } 1580 1581 //------------------------------inline_math_native----------------------------- 1582 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1583 switch (id) { 1584 // These intrinsics are not properly supported on all hardware 1585 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1586 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); 1587 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1588 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); 1589 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1590 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); 1591 1592 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : 1593 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); 1594 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : 1595 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); 1596 1597 // These intrinsics are supported on all hardware 1598 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; 1599 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; 1600 1601 // These intrinsics don't work on X86. The ad implementation doesn't 1602 // handle NaN's properly. Instead of returning infinity, the ad 1603 // implementation returns a NaN on overflow. See bug: 6304089 1604 // Once the ad implementations are fixed, change the code below 1605 // to match the intrinsics above 1606 1607 case vmIntrinsics::_dexp: return 1608 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1609 case vmIntrinsics::_dpow: return 1610 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1611 1612 // These intrinsics are not yet correctly implemented 1613 case vmIntrinsics::_datan2: 1614 return false; 1615 1616 default: 1617 ShouldNotReachHere(); 1618 return false; 1619 } 1620 } 1621 1622 static bool is_simple_name(Node* n) { 1623 return (n->req() == 1 // constant 1624 || (n->is_Type() && n->as_Type()->type()->singleton()) 1625 || n->is_Proj() // parameter or return value 1626 || n->is_Phi() // local of some sort 1627 ); 1628 } 1629 1630 //----------------------------inline_min_max----------------------------------- 1631 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1632 push(generate_min_max(id, argument(0), argument(1))); 1633 1634 return true; 1635 } 1636 1637 Node* 1638 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1639 // These are the candidate return value: 1640 Node* xvalue = x0; 1641 Node* yvalue = y0; 1642 1643 if (xvalue == yvalue) { 1644 return xvalue; 1645 } 1646 1647 bool want_max = (id == vmIntrinsics::_max); 1648 1649 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1650 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1651 if (txvalue == NULL || tyvalue == NULL) return top(); 1652 // This is not really necessary, but it is consistent with a 1653 // hypothetical MaxINode::Value method: 1654 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1655 1656 // %%% This folding logic should (ideally) be in a different place. 1657 // Some should be inside IfNode, and there to be a more reliable 1658 // transformation of ?: style patterns into cmoves. We also want 1659 // more powerful optimizations around cmove and min/max. 1660 1661 // Try to find a dominating comparison of these guys. 1662 // It can simplify the index computation for Arrays.copyOf 1663 // and similar uses of System.arraycopy. 1664 // First, compute the normalized version of CmpI(x, y). 1665 int cmp_op = Op_CmpI; 1666 Node* xkey = xvalue; 1667 Node* ykey = yvalue; 1668 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); 1669 if (ideal_cmpxy->is_Cmp()) { 1670 // E.g., if we have CmpI(length - offset, count), 1671 // it might idealize to CmpI(length, count + offset) 1672 cmp_op = ideal_cmpxy->Opcode(); 1673 xkey = ideal_cmpxy->in(1); 1674 ykey = ideal_cmpxy->in(2); 1675 } 1676 1677 // Start by locating any relevant comparisons. 1678 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1679 Node* cmpxy = NULL; 1680 Node* cmpyx = NULL; 1681 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1682 Node* cmp = start_from->fast_out(k); 1683 if (cmp->outcnt() > 0 && // must have prior uses 1684 cmp->in(0) == NULL && // must be context-independent 1685 cmp->Opcode() == cmp_op) { // right kind of compare 1686 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1687 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1688 } 1689 } 1690 1691 const int NCMPS = 2; 1692 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1693 int cmpn; 1694 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1695 if (cmps[cmpn] != NULL) break; // find a result 1696 } 1697 if (cmpn < NCMPS) { 1698 // Look for a dominating test that tells us the min and max. 1699 int depth = 0; // Limit search depth for speed 1700 Node* dom = control(); 1701 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1702 if (++depth >= 100) break; 1703 Node* ifproj = dom; 1704 if (!ifproj->is_Proj()) continue; 1705 Node* iff = ifproj->in(0); 1706 if (!iff->is_If()) continue; 1707 Node* bol = iff->in(1); 1708 if (!bol->is_Bool()) continue; 1709 Node* cmp = bol->in(1); 1710 if (cmp == NULL) continue; 1711 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1712 if (cmps[cmpn] == cmp) break; 1713 if (cmpn == NCMPS) continue; 1714 BoolTest::mask btest = bol->as_Bool()->_test._test; 1715 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1716 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1717 // At this point, we know that 'x btest y' is true. 1718 switch (btest) { 1719 case BoolTest::eq: 1720 // They are proven equal, so we can collapse the min/max. 1721 // Either value is the answer. Choose the simpler. 1722 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1723 return yvalue; 1724 return xvalue; 1725 case BoolTest::lt: // x < y 1726 case BoolTest::le: // x <= y 1727 return (want_max ? yvalue : xvalue); 1728 case BoolTest::gt: // x > y 1729 case BoolTest::ge: // x >= y 1730 return (want_max ? xvalue : yvalue); 1731 } 1732 } 1733 } 1734 1735 // We failed to find a dominating test. 1736 // Let's pick a test that might GVN with prior tests. 1737 Node* best_bol = NULL; 1738 BoolTest::mask best_btest = BoolTest::illegal; 1739 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1740 Node* cmp = cmps[cmpn]; 1741 if (cmp == NULL) continue; 1742 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1743 Node* bol = cmp->fast_out(j); 1744 if (!bol->is_Bool()) continue; 1745 BoolTest::mask btest = bol->as_Bool()->_test._test; 1746 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1747 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1748 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1749 best_bol = bol->as_Bool(); 1750 best_btest = btest; 1751 } 1752 } 1753 } 1754 1755 Node* answer_if_true = NULL; 1756 Node* answer_if_false = NULL; 1757 switch (best_btest) { 1758 default: 1759 if (cmpxy == NULL) 1760 cmpxy = ideal_cmpxy; 1761 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); 1762 // and fall through: 1763 case BoolTest::lt: // x < y 1764 case BoolTest::le: // x <= y 1765 answer_if_true = (want_max ? yvalue : xvalue); 1766 answer_if_false = (want_max ? xvalue : yvalue); 1767 break; 1768 case BoolTest::gt: // x > y 1769 case BoolTest::ge: // x >= y 1770 answer_if_true = (want_max ? xvalue : yvalue); 1771 answer_if_false = (want_max ? yvalue : xvalue); 1772 break; 1773 } 1774 1775 jint hi, lo; 1776 if (want_max) { 1777 // We can sharpen the minimum. 1778 hi = MAX2(txvalue->_hi, tyvalue->_hi); 1779 lo = MAX2(txvalue->_lo, tyvalue->_lo); 1780 } else { 1781 // We can sharpen the maximum. 1782 hi = MIN2(txvalue->_hi, tyvalue->_hi); 1783 lo = MIN2(txvalue->_lo, tyvalue->_lo); 1784 } 1785 1786 // Use a flow-free graph structure, to avoid creating excess control edges 1787 // which could hinder other optimizations. 1788 // Since Math.min/max is often used with arraycopy, we want 1789 // tightly_coupled_allocation to be able to see beyond min/max expressions. 1790 Node* cmov = CMoveNode::make(C, NULL, best_bol, 1791 answer_if_false, answer_if_true, 1792 TypeInt::make(lo, hi, widen)); 1793 1794 return _gvn.transform(cmov); 1795 1796 /* 1797 // This is not as desirable as it may seem, since Min and Max 1798 // nodes do not have a full set of optimizations. 1799 // And they would interfere, anyway, with 'if' optimizations 1800 // and with CMoveI canonical forms. 1801 switch (id) { 1802 case vmIntrinsics::_min: 1803 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 1804 case vmIntrinsics::_max: 1805 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 1806 default: 1807 ShouldNotReachHere(); 1808 } 1809 */ 1810 } 1811 1812 inline int 1813 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 1814 const TypePtr* base_type = TypePtr::NULL_PTR; 1815 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 1816 if (base_type == NULL) { 1817 // Unknown type. 1818 return Type::AnyPtr; 1819 } else if (base_type == TypePtr::NULL_PTR) { 1820 // Since this is a NULL+long form, we have to switch to a rawptr. 1821 base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); 1822 offset = MakeConX(0); 1823 return Type::RawPtr; 1824 } else if (base_type->base() == Type::RawPtr) { 1825 return Type::RawPtr; 1826 } else if (base_type->isa_oopptr()) { 1827 // Base is never null => always a heap address. 1828 if (base_type->ptr() == TypePtr::NotNull) { 1829 return Type::OopPtr; 1830 } 1831 // Offset is small => always a heap address. 1832 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 1833 if (offset_type != NULL && 1834 base_type->offset() == 0 && // (should always be?) 1835 offset_type->_lo >= 0 && 1836 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 1837 return Type::OopPtr; 1838 } 1839 // Otherwise, it might either be oop+off or NULL+addr. 1840 return Type::AnyPtr; 1841 } else { 1842 // No information: 1843 return Type::AnyPtr; 1844 } 1845 } 1846 1847 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 1848 int kind = classify_unsafe_addr(base, offset); 1849 if (kind == Type::RawPtr) { 1850 return basic_plus_adr(top(), base, offset); 1851 } else { 1852 return basic_plus_adr(base, offset); 1853 } 1854 } 1855 1856 //-------------------inline_numberOfLeadingZeros_int/long----------------------- 1857 // inline int Integer.numberOfLeadingZeros(int) 1858 // inline int Long.numberOfLeadingZeros(long) 1859 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) { 1860 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros"); 1861 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false; 1862 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false; 1863 _sp += arg_size(); // restore stack pointer 1864 switch (id) { 1865 case vmIntrinsics::_numberOfLeadingZeros_i: 1866 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop()))); 1867 break; 1868 case vmIntrinsics::_numberOfLeadingZeros_l: 1869 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair()))); 1870 break; 1871 default: 1872 ShouldNotReachHere(); 1873 } 1874 return true; 1875 } 1876 1877 //-------------------inline_numberOfTrailingZeros_int/long---------------------- 1878 // inline int Integer.numberOfTrailingZeros(int) 1879 // inline int Long.numberOfTrailingZeros(long) 1880 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) { 1881 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros"); 1882 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false; 1883 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false; 1884 _sp += arg_size(); // restore stack pointer 1885 switch (id) { 1886 case vmIntrinsics::_numberOfTrailingZeros_i: 1887 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop()))); 1888 break; 1889 case vmIntrinsics::_numberOfTrailingZeros_l: 1890 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair()))); 1891 break; 1892 default: 1893 ShouldNotReachHere(); 1894 } 1895 return true; 1896 } 1897 1898 //----------------------------inline_bitCount_int/long----------------------- 1899 // inline int Integer.bitCount(int) 1900 // inline int Long.bitCount(long) 1901 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) { 1902 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount"); 1903 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false; 1904 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false; 1905 _sp += arg_size(); // restore stack pointer 1906 switch (id) { 1907 case vmIntrinsics::_bitCount_i: 1908 push(_gvn.transform(new (C, 2) PopCountINode(pop()))); 1909 break; 1910 case vmIntrinsics::_bitCount_l: 1911 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair()))); 1912 break; 1913 default: 1914 ShouldNotReachHere(); 1915 } 1916 return true; 1917 } 1918 1919 //----------------------------inline_reverseBytes_int/long------------------- 1920 // inline Integer.reverseBytes(int) 1921 // inline Long.reverseBytes(long) 1922 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { 1923 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes"); 1924 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; 1925 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; 1926 _sp += arg_size(); // restore stack pointer 1927 switch (id) { 1928 case vmIntrinsics::_reverseBytes_i: 1929 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); 1930 break; 1931 case vmIntrinsics::_reverseBytes_l: 1932 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); 1933 break; 1934 default: 1935 ; 1936 } 1937 return true; 1938 } 1939 1940 //----------------------------inline_unsafe_access---------------------------- 1941 1942 const static BasicType T_ADDRESS_HOLDER = T_LONG; 1943 1944 // Interpret Unsafe.fieldOffset cookies correctly: 1945 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 1946 1947 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 1948 if (callee()->is_static()) return false; // caller must have the capability! 1949 1950 #ifndef PRODUCT 1951 { 1952 ResourceMark rm; 1953 // Check the signatures. 1954 ciSignature* sig = signature(); 1955 #ifdef ASSERT 1956 if (!is_store) { 1957 // Object getObject(Object base, int/long offset), etc. 1958 BasicType rtype = sig->return_type()->basic_type(); 1959 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 1960 rtype = T_ADDRESS; // it is really a C void* 1961 assert(rtype == type, "getter must return the expected value"); 1962 if (!is_native_ptr) { 1963 assert(sig->count() == 2, "oop getter has 2 arguments"); 1964 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 1965 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 1966 } else { 1967 assert(sig->count() == 1, "native getter has 1 argument"); 1968 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 1969 } 1970 } else { 1971 // void putObject(Object base, int/long offset, Object x), etc. 1972 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 1973 if (!is_native_ptr) { 1974 assert(sig->count() == 3, "oop putter has 3 arguments"); 1975 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 1976 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 1977 } else { 1978 assert(sig->count() == 2, "native putter has 2 arguments"); 1979 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 1980 } 1981 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 1982 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 1983 vtype = T_ADDRESS; // it is really a C void* 1984 assert(vtype == type, "putter must accept the expected value"); 1985 } 1986 #endif // ASSERT 1987 } 1988 #endif //PRODUCT 1989 1990 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 1991 1992 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; 1993 1994 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words 1995 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); 1996 1997 debug_only(int saved_sp = _sp); 1998 _sp += nargs; 1999 2000 Node* val; 2001 debug_only(val = (Node*)(uintptr_t)-1); 2002 2003 2004 if (is_store) { 2005 // Get the value being stored. (Pop it first; it was pushed last.) 2006 switch (type) { 2007 case T_DOUBLE: 2008 case T_LONG: 2009 case T_ADDRESS: 2010 val = pop_pair(); 2011 break; 2012 default: 2013 val = pop(); 2014 } 2015 } 2016 2017 // Build address expression. See the code in inline_unsafe_prefetch. 2018 Node *adr; 2019 Node *heap_base_oop = top(); 2020 if (!is_native_ptr) { 2021 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2022 Node* offset = pop_pair(); 2023 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2024 Node* base = pop(); 2025 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2026 // to be plain byte offsets, which are also the same as those accepted 2027 // by oopDesc::field_base. 2028 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2029 "fieldOffset must be byte-scaled"); 2030 // 32-bit machines ignore the high half! 2031 offset = ConvL2X(offset); 2032 adr = make_unsafe_address(base, offset); 2033 heap_base_oop = base; 2034 } else { 2035 Node* ptr = pop_pair(); 2036 // Adjust Java long to machine word: 2037 ptr = ConvL2X(ptr); 2038 adr = make_unsafe_address(NULL, ptr); 2039 } 2040 2041 // Pop receiver last: it was pushed first. 2042 Node *receiver = pop(); 2043 2044 assert(saved_sp == _sp, "must have correct argument count"); 2045 2046 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2047 2048 // First guess at the value type. 2049 const Type *value_type = Type::get_const_basic_type(type); 2050 2051 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2052 // there was not enough information to nail it down. 2053 Compile::AliasType* alias_type = C->alias_type(adr_type); 2054 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2055 2056 // We will need memory barriers unless we can determine a unique 2057 // alias category for this reference. (Note: If for some reason 2058 // the barriers get omitted and the unsafe reference begins to "pollute" 2059 // the alias analysis of the rest of the graph, either Compile::can_alias 2060 // or Compile::must_alias will throw a diagnostic assert.) 2061 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2062 2063 if (!is_store && type == T_OBJECT) { 2064 // Attempt to infer a sharper value type from the offset and base type. 2065 ciKlass* sharpened_klass = NULL; 2066 2067 // See if it is an instance field, with an object type. 2068 if (alias_type->field() != NULL) { 2069 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2070 if (alias_type->field()->type()->is_klass()) { 2071 sharpened_klass = alias_type->field()->type()->as_klass(); 2072 } 2073 } 2074 2075 // See if it is a narrow oop array. 2076 if (adr_type->isa_aryptr()) { 2077 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes(type)) { 2078 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2079 if (elem_type != NULL) { 2080 sharpened_klass = elem_type->klass(); 2081 } 2082 } 2083 } 2084 2085 if (sharpened_klass != NULL) { 2086 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2087 2088 // Sharpen the value type. 2089 value_type = tjp; 2090 2091 #ifndef PRODUCT 2092 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2093 tty->print(" from base type: "); adr_type->dump(); 2094 tty->print(" sharpened value: "); value_type->dump(); 2095 } 2096 #endif 2097 } 2098 } 2099 2100 // Null check on self without removing any arguments. The argument 2101 // null check technically happens in the wrong place, which can lead to 2102 // invalid stack traces when the primitive is inlined into a method 2103 // which handles NullPointerExceptions. 2104 _sp += nargs; 2105 do_null_check(receiver, T_OBJECT); 2106 _sp -= nargs; 2107 if (stopped()) { 2108 return true; 2109 } 2110 // Heap pointers get a null-check from the interpreter, 2111 // as a courtesy. However, this is not guaranteed by Unsafe, 2112 // and it is not possible to fully distinguish unintended nulls 2113 // from intended ones in this API. 2114 2115 if (is_volatile) { 2116 // We need to emit leading and trailing CPU membars (see below) in 2117 // addition to memory membars when is_volatile. This is a little 2118 // too strong, but avoids the need to insert per-alias-type 2119 // volatile membars (for stores; compare Parse::do_put_xxx), which 2120 // we cannot do effectively here because we probably only have a 2121 // rough approximation of type. 2122 need_mem_bar = true; 2123 // For Stores, place a memory ordering barrier now. 2124 if (is_store) 2125 insert_mem_bar(Op_MemBarRelease); 2126 } 2127 2128 // Memory barrier to prevent normal and 'unsafe' accesses from 2129 // bypassing each other. Happens after null checks, so the 2130 // exception paths do not take memory state from the memory barrier, 2131 // so there's no problems making a strong assert about mixing users 2132 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2133 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2134 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2135 2136 if (!is_store) { 2137 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2138 // load value and push onto stack 2139 switch (type) { 2140 case T_BOOLEAN: 2141 case T_CHAR: 2142 case T_BYTE: 2143 case T_SHORT: 2144 case T_INT: 2145 case T_FLOAT: 2146 case T_OBJECT: 2147 push( p ); 2148 break; 2149 case T_ADDRESS: 2150 // Cast to an int type. 2151 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); 2152 p = ConvX2L(p); 2153 push_pair(p); 2154 break; 2155 case T_DOUBLE: 2156 case T_LONG: 2157 push_pair( p ); 2158 break; 2159 default: ShouldNotReachHere(); 2160 } 2161 } else { 2162 // place effect of store into memory 2163 switch (type) { 2164 case T_DOUBLE: 2165 val = dstore_rounding(val); 2166 break; 2167 case T_ADDRESS: 2168 // Repackage the long as a pointer. 2169 val = ConvL2X(val); 2170 val = _gvn.transform( new (C, 2) CastX2PNode(val) ); 2171 break; 2172 } 2173 2174 if (type != T_OBJECT ) { 2175 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2176 } else { 2177 // Possibly an oop being stored to Java heap or native memory 2178 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2179 // oop to Java heap. 2180 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2181 } else { 2182 // We can't tell at compile time if we are storing in the Java heap or outside 2183 // of it. So we need to emit code to conditionally do the proper type of 2184 // store. 2185 2186 IdealKit kit(gvn(), control(), merged_memory()); 2187 kit.declares_done(); 2188 // QQQ who knows what probability is here?? 2189 kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2190 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2191 } kit.else_(); { 2192 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2193 } kit.end_if(); 2194 } 2195 } 2196 } 2197 2198 if (is_volatile) { 2199 if (!is_store) 2200 insert_mem_bar(Op_MemBarAcquire); 2201 else 2202 insert_mem_bar(Op_MemBarVolatile); 2203 } 2204 2205 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2206 2207 return true; 2208 } 2209 2210 //----------------------------inline_unsafe_prefetch---------------------------- 2211 2212 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2213 #ifndef PRODUCT 2214 { 2215 ResourceMark rm; 2216 // Check the signatures. 2217 ciSignature* sig = signature(); 2218 #ifdef ASSERT 2219 // Object getObject(Object base, int/long offset), etc. 2220 BasicType rtype = sig->return_type()->basic_type(); 2221 if (!is_native_ptr) { 2222 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2223 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2224 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2225 } else { 2226 assert(sig->count() == 1, "native prefetch has 1 argument"); 2227 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2228 } 2229 #endif // ASSERT 2230 } 2231 #endif // !PRODUCT 2232 2233 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2234 2235 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args 2236 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); 2237 2238 debug_only(int saved_sp = _sp); 2239 _sp += nargs; 2240 2241 // Build address expression. See the code in inline_unsafe_access. 2242 Node *adr; 2243 if (!is_native_ptr) { 2244 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2245 Node* offset = pop_pair(); 2246 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2247 Node* base = pop(); 2248 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2249 // to be plain byte offsets, which are also the same as those accepted 2250 // by oopDesc::field_base. 2251 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2252 "fieldOffset must be byte-scaled"); 2253 // 32-bit machines ignore the high half! 2254 offset = ConvL2X(offset); 2255 adr = make_unsafe_address(base, offset); 2256 } else { 2257 Node* ptr = pop_pair(); 2258 // Adjust Java long to machine word: 2259 ptr = ConvL2X(ptr); 2260 adr = make_unsafe_address(NULL, ptr); 2261 } 2262 2263 if (is_static) { 2264 assert(saved_sp == _sp, "must have correct argument count"); 2265 } else { 2266 // Pop receiver last: it was pushed first. 2267 Node *receiver = pop(); 2268 assert(saved_sp == _sp, "must have correct argument count"); 2269 2270 // Null check on self without removing any arguments. The argument 2271 // null check technically happens in the wrong place, which can lead to 2272 // invalid stack traces when the primitive is inlined into a method 2273 // which handles NullPointerExceptions. 2274 _sp += nargs; 2275 do_null_check(receiver, T_OBJECT); 2276 _sp -= nargs; 2277 if (stopped()) { 2278 return true; 2279 } 2280 } 2281 2282 // Generate the read or write prefetch 2283 Node *prefetch; 2284 if (is_store) { 2285 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); 2286 } else { 2287 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); 2288 } 2289 prefetch->init_req(0, control()); 2290 set_i_o(_gvn.transform(prefetch)); 2291 2292 return true; 2293 } 2294 2295 //----------------------------inline_unsafe_CAS---------------------------- 2296 2297 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { 2298 // This basic scheme here is the same as inline_unsafe_access, but 2299 // differs in enough details that combining them would make the code 2300 // overly confusing. (This is a true fact! I originally combined 2301 // them, but even I was confused by it!) As much code/comments as 2302 // possible are retained from inline_unsafe_access though to make 2303 // the correspondences clearer. - dl 2304 2305 if (callee()->is_static()) return false; // caller must have the capability! 2306 2307 #ifndef PRODUCT 2308 { 2309 ResourceMark rm; 2310 // Check the signatures. 2311 ciSignature* sig = signature(); 2312 #ifdef ASSERT 2313 BasicType rtype = sig->return_type()->basic_type(); 2314 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2315 assert(sig->count() == 4, "CAS has 4 arguments"); 2316 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2317 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2318 #endif // ASSERT 2319 } 2320 #endif //PRODUCT 2321 2322 // number of stack slots per value argument (1 or 2) 2323 int type_words = type2size[type]; 2324 2325 // Cannot inline wide CAS on machines that don't support it natively 2326 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8()) 2327 return false; 2328 2329 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2330 2331 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; 2332 int nargs = 1 + 1 + 2 + type_words + type_words; 2333 2334 // pop arguments: newval, oldval, offset, base, and receiver 2335 debug_only(int saved_sp = _sp); 2336 _sp += nargs; 2337 Node* newval = (type_words == 1) ? pop() : pop_pair(); 2338 Node* oldval = (type_words == 1) ? pop() : pop_pair(); 2339 Node *offset = pop_pair(); 2340 Node *base = pop(); 2341 Node *receiver = pop(); 2342 assert(saved_sp == _sp, "must have correct argument count"); 2343 2344 // Null check receiver. 2345 _sp += nargs; 2346 do_null_check(receiver, T_OBJECT); 2347 _sp -= nargs; 2348 if (stopped()) { 2349 return true; 2350 } 2351 2352 // Build field offset expression. 2353 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2354 // to be plain byte offsets, which are also the same as those accepted 2355 // by oopDesc::field_base. 2356 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2357 // 32-bit machines ignore the high half of long offsets 2358 offset = ConvL2X(offset); 2359 Node* adr = make_unsafe_address(base, offset); 2360 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2361 2362 // (Unlike inline_unsafe_access, there seems no point in trying 2363 // to refine types. Just use the coarse types here. 2364 const Type *value_type = Type::get_const_basic_type(type); 2365 Compile::AliasType* alias_type = C->alias_type(adr_type); 2366 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2367 int alias_idx = C->get_alias_index(adr_type); 2368 2369 // Memory-model-wise, a CAS acts like a little synchronized block, 2370 // so needs barriers on each side. These don't translate into 2371 // actual barriers on most machines, but we still need rest of 2372 // compiler to respect ordering. 2373 2374 insert_mem_bar(Op_MemBarRelease); 2375 insert_mem_bar(Op_MemBarCPUOrder); 2376 2377 // 4984716: MemBars must be inserted before this 2378 // memory node in order to avoid a false 2379 // dependency which will confuse the scheduler. 2380 Node *mem = memory(alias_idx); 2381 2382 // For now, we handle only those cases that actually exist: ints, 2383 // longs, and Object. Adding others should be straightforward. 2384 Node* cas; 2385 switch(type) { 2386 case T_INT: 2387 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2388 break; 2389 case T_LONG: 2390 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2391 break; 2392 case T_OBJECT: 2393 // reference stores need a store barrier. 2394 // (They don't if CAS fails, but it isn't worth checking.) 2395 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT); 2396 #ifdef _LP64 2397 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2398 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2399 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2400 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr, 2401 newval_enc, oldval_enc)); 2402 } else 2403 #endif 2404 { 2405 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2406 } 2407 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); 2408 break; 2409 default: 2410 ShouldNotReachHere(); 2411 break; 2412 } 2413 2414 // SCMemProjNodes represent the memory state of CAS. Their main 2415 // role is to prevent CAS nodes from being optimized away when their 2416 // results aren't used. 2417 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 2418 set_memory(proj, alias_idx); 2419 2420 // Add the trailing membar surrounding the access 2421 insert_mem_bar(Op_MemBarCPUOrder); 2422 insert_mem_bar(Op_MemBarAcquire); 2423 2424 push(cas); 2425 return true; 2426 } 2427 2428 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2429 // This is another variant of inline_unsafe_access, differing in 2430 // that it always issues store-store ("release") barrier and ensures 2431 // store-atomicity (which only matters for "long"). 2432 2433 if (callee()->is_static()) return false; // caller must have the capability! 2434 2435 #ifndef PRODUCT 2436 { 2437 ResourceMark rm; 2438 // Check the signatures. 2439 ciSignature* sig = signature(); 2440 #ifdef ASSERT 2441 BasicType rtype = sig->return_type()->basic_type(); 2442 assert(rtype == T_VOID, "must return void"); 2443 assert(sig->count() == 3, "has 3 arguments"); 2444 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2445 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2446 #endif // ASSERT 2447 } 2448 #endif //PRODUCT 2449 2450 // number of stack slots per value argument (1 or 2) 2451 int type_words = type2size[type]; 2452 2453 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2454 2455 // Argument words: "this" plus oop plus offset plus value; 2456 int nargs = 1 + 1 + 2 + type_words; 2457 2458 // pop arguments: val, offset, base, and receiver 2459 debug_only(int saved_sp = _sp); 2460 _sp += nargs; 2461 Node* val = (type_words == 1) ? pop() : pop_pair(); 2462 Node *offset = pop_pair(); 2463 Node *base = pop(); 2464 Node *receiver = pop(); 2465 assert(saved_sp == _sp, "must have correct argument count"); 2466 2467 // Null check receiver. 2468 _sp += nargs; 2469 do_null_check(receiver, T_OBJECT); 2470 _sp -= nargs; 2471 if (stopped()) { 2472 return true; 2473 } 2474 2475 // Build field offset expression. 2476 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2477 // 32-bit machines ignore the high half of long offsets 2478 offset = ConvL2X(offset); 2479 Node* adr = make_unsafe_address(base, offset); 2480 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2481 const Type *value_type = Type::get_const_basic_type(type); 2482 Compile::AliasType* alias_type = C->alias_type(adr_type); 2483 2484 insert_mem_bar(Op_MemBarRelease); 2485 insert_mem_bar(Op_MemBarCPUOrder); 2486 // Ensure that the store is atomic for longs: 2487 bool require_atomic_access = true; 2488 Node* store; 2489 if (type == T_OBJECT) // reference stores need a store barrier. 2490 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2491 else { 2492 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2493 } 2494 insert_mem_bar(Op_MemBarCPUOrder); 2495 return true; 2496 } 2497 2498 bool LibraryCallKit::inline_unsafe_allocate() { 2499 if (callee()->is_static()) return false; // caller must have the capability! 2500 int nargs = 1 + 1; 2501 assert(signature()->size() == nargs-1, "alloc has 1 argument"); 2502 null_check_receiver(callee()); // check then ignore argument(0) 2503 _sp += nargs; // set original stack for use by uncommon_trap 2504 Node* cls = do_null_check(argument(1), T_OBJECT); 2505 _sp -= nargs; 2506 if (stopped()) return true; 2507 2508 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 2509 _sp += nargs; // set original stack for use by uncommon_trap 2510 kls = do_null_check(kls, T_OBJECT); 2511 _sp -= nargs; 2512 if (stopped()) return true; // argument was like int.class 2513 2514 // Note: The argument might still be an illegal value like 2515 // Serializable.class or Object[].class. The runtime will handle it. 2516 // But we must make an explicit check for initialization. 2517 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc)); 2518 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT); 2519 Node* bits = intcon(instanceKlass::fully_initialized); 2520 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); 2521 // The 'test' is non-zero if we need to take a slow path. 2522 2523 Node* obj = new_instance(kls, test); 2524 push(obj); 2525 2526 return true; 2527 } 2528 2529 //------------------------inline_native_time_funcs-------------- 2530 // inline code for System.currentTimeMillis() and System.nanoTime() 2531 // these have the same type and signature 2532 bool LibraryCallKit::inline_native_time_funcs(bool isNano) { 2533 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) : 2534 CAST_FROM_FN_PTR(address, os::javaTimeMillis); 2535 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis"; 2536 const TypeFunc *tf = OptoRuntime::current_time_millis_Type(); 2537 const TypePtr* no_memory_effects = NULL; 2538 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2539 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); 2540 #ifdef ASSERT 2541 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); 2542 assert(value_top == top(), "second value must be top"); 2543 #endif 2544 push_pair(value); 2545 return true; 2546 } 2547 2548 //------------------------inline_native_currentThread------------------ 2549 bool LibraryCallKit::inline_native_currentThread() { 2550 Node* junk = NULL; 2551 push(generate_current_thread(junk)); 2552 return true; 2553 } 2554 2555 //------------------------inline_native_isInterrupted------------------ 2556 bool LibraryCallKit::inline_native_isInterrupted() { 2557 const int nargs = 1+1; // receiver + boolean 2558 assert(nargs == arg_size(), "sanity"); 2559 // Add a fast path to t.isInterrupted(clear_int): 2560 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 2561 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 2562 // So, in the common case that the interrupt bit is false, 2563 // we avoid making a call into the VM. Even if the interrupt bit 2564 // is true, if the clear_int argument is false, we avoid the VM call. 2565 // However, if the receiver is not currentThread, we must call the VM, 2566 // because there must be some locking done around the operation. 2567 2568 // We only go to the fast case code if we pass two guards. 2569 // Paths which do not pass are accumulated in the slow_region. 2570 RegionNode* slow_region = new (C, 1) RegionNode(1); 2571 record_for_igvn(slow_region); 2572 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow 2573 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); 2574 enum { no_int_result_path = 1, 2575 no_clear_result_path = 2, 2576 slow_result_path = 3 2577 }; 2578 2579 // (a) Receiving thread must be the current thread. 2580 Node* rec_thr = argument(0); 2581 Node* tls_ptr = NULL; 2582 Node* cur_thr = generate_current_thread(tls_ptr); 2583 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); 2584 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); 2585 2586 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); 2587 if (!known_current_thread) 2588 generate_slow_guard(bol_thr, slow_region); 2589 2590 // (b) Interrupt bit on TLS must be false. 2591 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2592 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 2593 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 2594 // Set the control input on the field _interrupted read to prevent it floating up. 2595 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 2596 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); 2597 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); 2598 2599 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2600 2601 // First fast path: if (!TLS._interrupted) return false; 2602 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); 2603 result_rgn->init_req(no_int_result_path, false_bit); 2604 result_val->init_req(no_int_result_path, intcon(0)); 2605 2606 // drop through to next case 2607 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); 2608 2609 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 2610 Node* clr_arg = argument(1); 2611 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); 2612 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); 2613 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 2614 2615 // Second fast path: ... else if (!clear_int) return true; 2616 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); 2617 result_rgn->init_req(no_clear_result_path, false_arg); 2618 result_val->init_req(no_clear_result_path, intcon(1)); 2619 2620 // drop through to next case 2621 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); 2622 2623 // (d) Otherwise, go to the slow path. 2624 slow_region->add_req(control()); 2625 set_control( _gvn.transform(slow_region) ); 2626 2627 if (stopped()) { 2628 // There is no slow path. 2629 result_rgn->init_req(slow_result_path, top()); 2630 result_val->init_req(slow_result_path, top()); 2631 } else { 2632 // non-virtual because it is a private non-static 2633 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 2634 2635 Node* slow_val = set_results_for_java_call(slow_call); 2636 // this->control() comes from set_results_for_java_call 2637 2638 // If we know that the result of the slow call will be true, tell the optimizer! 2639 if (known_current_thread) slow_val = intcon(1); 2640 2641 Node* fast_io = slow_call->in(TypeFunc::I_O); 2642 Node* fast_mem = slow_call->in(TypeFunc::Memory); 2643 // These two phis are pre-filled with copies of of the fast IO and Memory 2644 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 2645 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 2646 2647 result_rgn->init_req(slow_result_path, control()); 2648 io_phi ->init_req(slow_result_path, i_o()); 2649 mem_phi ->init_req(slow_result_path, reset_memory()); 2650 result_val->init_req(slow_result_path, slow_val); 2651 2652 set_all_memory( _gvn.transform(mem_phi) ); 2653 set_i_o( _gvn.transform(io_phi) ); 2654 } 2655 2656 push_result(result_rgn, result_val); 2657 C->set_has_split_ifs(true); // Has chance for split-if optimization 2658 2659 return true; 2660 } 2661 2662 //---------------------------load_mirror_from_klass---------------------------- 2663 // Given a klass oop, load its java mirror (a java.lang.Class oop). 2664 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 2665 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc)); 2666 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 2667 } 2668 2669 //-----------------------load_klass_from_mirror_common------------------------- 2670 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 2671 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 2672 // and branch to the given path on the region. 2673 // If never_see_null, take an uncommon trap on null, so we can optimistically 2674 // compile for the non-null case. 2675 // If the region is NULL, force never_see_null = true. 2676 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 2677 bool never_see_null, 2678 int nargs, 2679 RegionNode* region, 2680 int null_path, 2681 int offset) { 2682 if (region == NULL) never_see_null = true; 2683 Node* p = basic_plus_adr(mirror, offset); 2684 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 2685 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); 2686 _sp += nargs; // any deopt will start just before call to enclosing method 2687 Node* null_ctl = top(); 2688 kls = null_check_oop(kls, &null_ctl, never_see_null); 2689 if (region != NULL) { 2690 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 2691 region->init_req(null_path, null_ctl); 2692 } else { 2693 assert(null_ctl == top(), "no loose ends"); 2694 } 2695 _sp -= nargs; 2696 return kls; 2697 } 2698 2699 //--------------------(inline_native_Class_query helpers)--------------------- 2700 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 2701 // Fall through if (mods & mask) == bits, take the guard otherwise. 2702 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 2703 // Branch around if the given klass has the given modifier bit set. 2704 // Like generate_guard, adds a new path onto the region. 2705 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); 2706 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 2707 Node* mask = intcon(modifier_mask); 2708 Node* bits = intcon(modifier_bits); 2709 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); 2710 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); 2711 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); 2712 return generate_fair_guard(bol, region); 2713 } 2714 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 2715 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 2716 } 2717 2718 //-------------------------inline_native_Class_query------------------- 2719 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 2720 int nargs = 1+0; // just the Class mirror, in most cases 2721 const Type* return_type = TypeInt::BOOL; 2722 Node* prim_return_value = top(); // what happens if it's a primitive class? 2723 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 2724 bool expect_prim = false; // most of these guys expect to work on refs 2725 2726 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 2727 2728 switch (id) { 2729 case vmIntrinsics::_isInstance: 2730 nargs = 1+1; // the Class mirror, plus the object getting queried about 2731 // nothing is an instance of a primitive type 2732 prim_return_value = intcon(0); 2733 break; 2734 case vmIntrinsics::_getModifiers: 2735 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2736 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 2737 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 2738 break; 2739 case vmIntrinsics::_isInterface: 2740 prim_return_value = intcon(0); 2741 break; 2742 case vmIntrinsics::_isArray: 2743 prim_return_value = intcon(0); 2744 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 2745 break; 2746 case vmIntrinsics::_isPrimitive: 2747 prim_return_value = intcon(1); 2748 expect_prim = true; // obviously 2749 break; 2750 case vmIntrinsics::_getSuperclass: 2751 prim_return_value = null(); 2752 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 2753 break; 2754 case vmIntrinsics::_getComponentType: 2755 prim_return_value = null(); 2756 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 2757 break; 2758 case vmIntrinsics::_getClassAccessFlags: 2759 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2760 return_type = TypeInt::INT; // not bool! 6297094 2761 break; 2762 default: 2763 ShouldNotReachHere(); 2764 } 2765 2766 Node* mirror = argument(0); 2767 Node* obj = (nargs <= 1)? top(): argument(1); 2768 2769 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 2770 if (mirror_con == NULL) return false; // cannot happen? 2771 2772 #ifndef PRODUCT 2773 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2774 ciType* k = mirror_con->java_mirror_type(); 2775 if (k) { 2776 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 2777 k->print_name(); 2778 tty->cr(); 2779 } 2780 } 2781 #endif 2782 2783 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 2784 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2785 record_for_igvn(region); 2786 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); 2787 2788 // The mirror will never be null of Reflection.getClassAccessFlags, however 2789 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 2790 // if it is. See bug 4774291. 2791 2792 // For Reflection.getClassAccessFlags(), the null check occurs in 2793 // the wrong place; see inline_unsafe_access(), above, for a similar 2794 // situation. 2795 _sp += nargs; // set original stack for use by uncommon_trap 2796 mirror = do_null_check(mirror, T_OBJECT); 2797 _sp -= nargs; 2798 // If mirror or obj is dead, only null-path is taken. 2799 if (stopped()) return true; 2800 2801 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 2802 2803 // Now load the mirror's klass metaobject, and null-check it. 2804 // Side-effects region with the control path if the klass is null. 2805 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, 2806 region, _prim_path); 2807 // If kls is null, we have a primitive mirror. 2808 phi->init_req(_prim_path, prim_return_value); 2809 if (stopped()) { push_result(region, phi); return true; } 2810 2811 Node* p; // handy temp 2812 Node* null_ctl; 2813 2814 // Now that we have the non-null klass, we can perform the real query. 2815 // For constant classes, the query will constant-fold in LoadNode::Value. 2816 Node* query_value = top(); 2817 switch (id) { 2818 case vmIntrinsics::_isInstance: 2819 // nothing is an instance of a primitive type 2820 query_value = gen_instanceof(obj, kls); 2821 break; 2822 2823 case vmIntrinsics::_getModifiers: 2824 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc)); 2825 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 2826 break; 2827 2828 case vmIntrinsics::_isInterface: 2829 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 2830 if (generate_interface_guard(kls, region) != NULL) 2831 // A guard was added. If the guard is taken, it was an interface. 2832 phi->add_req(intcon(1)); 2833 // If we fall through, it's a plain class. 2834 query_value = intcon(0); 2835 break; 2836 2837 case vmIntrinsics::_isArray: 2838 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 2839 if (generate_array_guard(kls, region) != NULL) 2840 // A guard was added. If the guard is taken, it was an array. 2841 phi->add_req(intcon(1)); 2842 // If we fall through, it's a plain class. 2843 query_value = intcon(0); 2844 break; 2845 2846 case vmIntrinsics::_isPrimitive: 2847 query_value = intcon(0); // "normal" path produces false 2848 break; 2849 2850 case vmIntrinsics::_getSuperclass: 2851 // The rules here are somewhat unfortunate, but we can still do better 2852 // with random logic than with a JNI call. 2853 // Interfaces store null or Object as _super, but must report null. 2854 // Arrays store an intermediate super as _super, but must report Object. 2855 // Other types can report the actual _super. 2856 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 2857 if (generate_interface_guard(kls, region) != NULL) 2858 // A guard was added. If the guard is taken, it was an interface. 2859 phi->add_req(null()); 2860 if (generate_array_guard(kls, region) != NULL) 2861 // A guard was added. If the guard is taken, it was an array. 2862 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 2863 // If we fall through, it's a plain class. Get its _super. 2864 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc)); 2865 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); 2866 null_ctl = top(); 2867 kls = null_check_oop(kls, &null_ctl); 2868 if (null_ctl != top()) { 2869 // If the guard is taken, Object.superClass is null (both klass and mirror). 2870 region->add_req(null_ctl); 2871 phi ->add_req(null()); 2872 } 2873 if (!stopped()) { 2874 query_value = load_mirror_from_klass(kls); 2875 } 2876 break; 2877 2878 case vmIntrinsics::_getComponentType: 2879 if (generate_array_guard(kls, region) != NULL) { 2880 // Be sure to pin the oop load to the guard edge just created: 2881 Node* is_array_ctrl = region->in(region->req()-1); 2882 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc)); 2883 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 2884 phi->add_req(cmo); 2885 } 2886 query_value = null(); // non-array case is null 2887 break; 2888 2889 case vmIntrinsics::_getClassAccessFlags: 2890 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); 2891 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 2892 break; 2893 2894 default: 2895 ShouldNotReachHere(); 2896 } 2897 2898 // Fall-through is the normal case of a query to a real class. 2899 phi->init_req(1, query_value); 2900 region->init_req(1, control()); 2901 2902 push_result(region, phi); 2903 C->set_has_split_ifs(true); // Has chance for split-if optimization 2904 2905 return true; 2906 } 2907 2908 //--------------------------inline_native_subtype_check------------------------ 2909 // This intrinsic takes the JNI calls out of the heart of 2910 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 2911 bool LibraryCallKit::inline_native_subtype_check() { 2912 int nargs = 1+1; // the Class mirror, plus the other class getting examined 2913 2914 // Pull both arguments off the stack. 2915 Node* args[2]; // two java.lang.Class mirrors: superc, subc 2916 args[0] = argument(0); 2917 args[1] = argument(1); 2918 Node* klasses[2]; // corresponding Klasses: superk, subk 2919 klasses[0] = klasses[1] = top(); 2920 2921 enum { 2922 // A full decision tree on {superc is prim, subc is prim}: 2923 _prim_0_path = 1, // {P,N} => false 2924 // {P,P} & superc!=subc => false 2925 _prim_same_path, // {P,P} & superc==subc => true 2926 _prim_1_path, // {N,P} => false 2927 _ref_subtype_path, // {N,N} & subtype check wins => true 2928 _both_ref_path, // {N,N} & subtype check loses => false 2929 PATH_LIMIT 2930 }; 2931 2932 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2933 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 2934 record_for_igvn(region); 2935 2936 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 2937 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 2938 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 2939 2940 // First null-check both mirrors and load each mirror's klass metaobject. 2941 int which_arg; 2942 for (which_arg = 0; which_arg <= 1; which_arg++) { 2943 Node* arg = args[which_arg]; 2944 _sp += nargs; // set original stack for use by uncommon_trap 2945 arg = do_null_check(arg, T_OBJECT); 2946 _sp -= nargs; 2947 if (stopped()) break; 2948 args[which_arg] = _gvn.transform(arg); 2949 2950 Node* p = basic_plus_adr(arg, class_klass_offset); 2951 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 2952 klasses[which_arg] = _gvn.transform(kls); 2953 } 2954 2955 // Having loaded both klasses, test each for null. 2956 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 2957 for (which_arg = 0; which_arg <= 1; which_arg++) { 2958 Node* kls = klasses[which_arg]; 2959 Node* null_ctl = top(); 2960 _sp += nargs; // set original stack for use by uncommon_trap 2961 kls = null_check_oop(kls, &null_ctl, never_see_null); 2962 _sp -= nargs; 2963 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 2964 region->init_req(prim_path, null_ctl); 2965 if (stopped()) break; 2966 klasses[which_arg] = kls; 2967 } 2968 2969 if (!stopped()) { 2970 // now we have two reference types, in klasses[0..1] 2971 Node* subk = klasses[1]; // the argument to isAssignableFrom 2972 Node* superk = klasses[0]; // the receiver 2973 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 2974 // now we have a successful reference subtype check 2975 region->set_req(_ref_subtype_path, control()); 2976 } 2977 2978 // If both operands are primitive (both klasses null), then 2979 // we must return true when they are identical primitives. 2980 // It is convenient to test this after the first null klass check. 2981 set_control(region->in(_prim_0_path)); // go back to first null check 2982 if (!stopped()) { 2983 // Since superc is primitive, make a guard for the superc==subc case. 2984 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); 2985 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); 2986 generate_guard(bol_eq, region, PROB_FAIR); 2987 if (region->req() == PATH_LIMIT+1) { 2988 // A guard was added. If the added guard is taken, superc==subc. 2989 region->swap_edges(PATH_LIMIT, _prim_same_path); 2990 region->del_req(PATH_LIMIT); 2991 } 2992 region->set_req(_prim_0_path, control()); // Not equal after all. 2993 } 2994 2995 // these are the only paths that produce 'true': 2996 phi->set_req(_prim_same_path, intcon(1)); 2997 phi->set_req(_ref_subtype_path, intcon(1)); 2998 2999 // pull together the cases: 3000 assert(region->req() == PATH_LIMIT, "sane region"); 3001 for (uint i = 1; i < region->req(); i++) { 3002 Node* ctl = region->in(i); 3003 if (ctl == NULL || ctl == top()) { 3004 region->set_req(i, top()); 3005 phi ->set_req(i, top()); 3006 } else if (phi->in(i) == NULL) { 3007 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3008 } 3009 } 3010 3011 set_control(_gvn.transform(region)); 3012 push(_gvn.transform(phi)); 3013 3014 return true; 3015 } 3016 3017 //---------------------generate_array_guard_common------------------------ 3018 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3019 bool obj_array, bool not_array) { 3020 // If obj_array/non_array==false/false: 3021 // Branch around if the given klass is in fact an array (either obj or prim). 3022 // If obj_array/non_array==false/true: 3023 // Branch around if the given klass is not an array klass of any kind. 3024 // If obj_array/non_array==true/true: 3025 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3026 // If obj_array/non_array==true/false: 3027 // Branch around if the kls is an oop array (Object[] or subtype) 3028 // 3029 // Like generate_guard, adds a new path onto the region. 3030 jint layout_con = 0; 3031 Node* layout_val = get_layout_helper(kls, layout_con); 3032 if (layout_val == NULL) { 3033 bool query = (obj_array 3034 ? Klass::layout_helper_is_objArray(layout_con) 3035 : Klass::layout_helper_is_javaArray(layout_con)); 3036 if (query == not_array) { 3037 return NULL; // never a branch 3038 } else { // always a branch 3039 Node* always_branch = control(); 3040 if (region != NULL) 3041 region->add_req(always_branch); 3042 set_control(top()); 3043 return always_branch; 3044 } 3045 } 3046 // Now test the correct condition. 3047 jint nval = (obj_array 3048 ? ((jint)Klass::_lh_array_tag_type_value 3049 << Klass::_lh_array_tag_shift) 3050 : Klass::_lh_neutral_value); 3051 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); 3052 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3053 // invert the test if we are looking for a non-array 3054 if (not_array) btest = BoolTest(btest).negate(); 3055 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); 3056 return generate_fair_guard(bol, region); 3057 } 3058 3059 3060 //-----------------------inline_native_newArray-------------------------- 3061 bool LibraryCallKit::inline_native_newArray() { 3062 int nargs = 2; 3063 Node* mirror = argument(0); 3064 Node* count_val = argument(1); 3065 3066 _sp += nargs; // set original stack for use by uncommon_trap 3067 mirror = do_null_check(mirror, T_OBJECT); 3068 _sp -= nargs; 3069 // If mirror or obj is dead, only null-path is taken. 3070 if (stopped()) return true; 3071 3072 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3073 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3074 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3075 TypeInstPtr::NOTNULL); 3076 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3077 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3078 TypePtr::BOTTOM); 3079 3080 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3081 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3082 nargs, 3083 result_reg, _slow_path); 3084 Node* normal_ctl = control(); 3085 Node* no_array_ctl = result_reg->in(_slow_path); 3086 3087 // Generate code for the slow case. We make a call to newArray(). 3088 set_control(no_array_ctl); 3089 if (!stopped()) { 3090 // Either the input type is void.class, or else the 3091 // array klass has not yet been cached. Either the 3092 // ensuing call will throw an exception, or else it 3093 // will cache the array klass for next time. 3094 PreserveJVMState pjvms(this); 3095 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3096 Node* slow_result = set_results_for_java_call(slow_call); 3097 // this->control() comes from set_results_for_java_call 3098 result_reg->set_req(_slow_path, control()); 3099 result_val->set_req(_slow_path, slow_result); 3100 result_io ->set_req(_slow_path, i_o()); 3101 result_mem->set_req(_slow_path, reset_memory()); 3102 } 3103 3104 set_control(normal_ctl); 3105 if (!stopped()) { 3106 // Normal case: The array type has been cached in the java.lang.Class. 3107 // The following call works fine even if the array type is polymorphic. 3108 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3109 Node* obj = new_array(klass_node, count_val, nargs); 3110 result_reg->init_req(_normal_path, control()); 3111 result_val->init_req(_normal_path, obj); 3112 result_io ->init_req(_normal_path, i_o()); 3113 result_mem->init_req(_normal_path, reset_memory()); 3114 } 3115 3116 // Return the combined state. 3117 set_i_o( _gvn.transform(result_io) ); 3118 set_all_memory( _gvn.transform(result_mem) ); 3119 push_result(result_reg, result_val); 3120 C->set_has_split_ifs(true); // Has chance for split-if optimization 3121 3122 return true; 3123 } 3124 3125 //----------------------inline_native_getLength-------------------------- 3126 bool LibraryCallKit::inline_native_getLength() { 3127 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3128 3129 int nargs = 1; 3130 Node* array = argument(0); 3131 3132 _sp += nargs; // set original stack for use by uncommon_trap 3133 array = do_null_check(array, T_OBJECT); 3134 _sp -= nargs; 3135 3136 // If array is dead, only null-path is taken. 3137 if (stopped()) return true; 3138 3139 // Deoptimize if it is a non-array. 3140 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3141 3142 if (non_array != NULL) { 3143 PreserveJVMState pjvms(this); 3144 set_control(non_array); 3145 _sp += nargs; // push the arguments back on the stack 3146 uncommon_trap(Deoptimization::Reason_intrinsic, 3147 Deoptimization::Action_maybe_recompile); 3148 } 3149 3150 // If control is dead, only non-array-path is taken. 3151 if (stopped()) return true; 3152 3153 // The works fine even if the array type is polymorphic. 3154 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3155 push( load_array_length(array) ); 3156 3157 C->set_has_split_ifs(true); // Has chance for split-if optimization 3158 3159 return true; 3160 } 3161 3162 //------------------------inline_array_copyOf---------------------------- 3163 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3164 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3165 3166 // Restore the stack and pop off the arguments. 3167 int nargs = 3 + (is_copyOfRange? 1: 0); 3168 Node* original = argument(0); 3169 Node* start = is_copyOfRange? argument(1): intcon(0); 3170 Node* end = is_copyOfRange? argument(2): argument(1); 3171 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3172 3173 _sp += nargs; // set original stack for use by uncommon_trap 3174 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); 3175 original = do_null_check(original, T_OBJECT); 3176 _sp -= nargs; 3177 3178 // Check if a null path was taken unconditionally. 3179 if (stopped()) return true; 3180 3181 Node* orig_length = load_array_length(original); 3182 3183 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs, 3184 NULL, 0); 3185 _sp += nargs; // set original stack for use by uncommon_trap 3186 klass_node = do_null_check(klass_node, T_OBJECT); 3187 _sp -= nargs; 3188 3189 RegionNode* bailout = new (C, 1) RegionNode(1); 3190 record_for_igvn(bailout); 3191 3192 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3193 // Bail out if that is so. 3194 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3195 if (not_objArray != NULL) { 3196 // Improve the klass node's type from the new optimistic assumption: 3197 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3198 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3199 Node* cast = new (C, 2) CastPPNode(klass_node, akls); 3200 cast->init_req(0, control()); 3201 klass_node = _gvn.transform(cast); 3202 } 3203 3204 // Bail out if either start or end is negative. 3205 generate_negative_guard(start, bailout, &start); 3206 generate_negative_guard(end, bailout, &end); 3207 3208 Node* length = end; 3209 if (_gvn.type(start) != TypeInt::ZERO) { 3210 length = _gvn.transform( new (C, 3) SubINode(end, start) ); 3211 } 3212 3213 // Bail out if length is negative. 3214 // ...Not needed, since the new_array will throw the right exception. 3215 //generate_negative_guard(length, bailout, &length); 3216 3217 if (bailout->req() > 1) { 3218 PreserveJVMState pjvms(this); 3219 set_control( _gvn.transform(bailout) ); 3220 _sp += nargs; // push the arguments back on the stack 3221 uncommon_trap(Deoptimization::Reason_intrinsic, 3222 Deoptimization::Action_maybe_recompile); 3223 } 3224 3225 if (!stopped()) { 3226 // How many elements will we copy from the original? 3227 // The answer is MinI(orig_length - start, length). 3228 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); 3229 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3230 3231 const bool raw_mem_only = true; 3232 Node* newcopy = new_array(klass_node, length, nargs, raw_mem_only); 3233 3234 // Generate a direct call to the right arraycopy function(s). 3235 // We know the copy is disjoint but we might not know if the 3236 // oop stores need checking. 3237 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3238 // This will fail a store-check if x contains any non-nulls. 3239 bool disjoint_bases = true; 3240 bool length_never_negative = true; 3241 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3242 original, start, newcopy, intcon(0), moved, 3243 disjoint_bases, length_never_negative); 3244 3245 push(newcopy); 3246 } 3247 3248 C->set_has_split_ifs(true); // Has chance for split-if optimization 3249 3250 return true; 3251 } 3252 3253 3254 //----------------------generate_virtual_guard--------------------------- 3255 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3256 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3257 RegionNode* slow_region) { 3258 ciMethod* method = callee(); 3259 int vtable_index = method->vtable_index(); 3260 // Get the methodOop out of the appropriate vtable entry. 3261 int entry_offset = (instanceKlass::vtable_start_offset() + 3262 vtable_index*vtableEntry::size()) * wordSize + 3263 vtableEntry::method_offset_in_bytes(); 3264 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3265 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); 3266 3267 // Compare the target method with the expected method (e.g., Object.hashCode). 3268 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method); 3269 3270 Node* native_call = makecon(native_call_addr); 3271 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); 3272 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); 3273 3274 return generate_slow_guard(test_native, slow_region); 3275 } 3276 3277 //-----------------------generate_method_call---------------------------- 3278 // Use generate_method_call to make a slow-call to the real 3279 // method if the fast path fails. An alternative would be to 3280 // use a stub like OptoRuntime::slow_arraycopy_Java. 3281 // This only works for expanding the current library call, 3282 // not another intrinsic. (E.g., don't use this for making an 3283 // arraycopy call inside of the copyOf intrinsic.) 3284 CallJavaNode* 3285 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3286 // When compiling the intrinsic method itself, do not use this technique. 3287 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3288 3289 ciMethod* method = callee(); 3290 // ensure the JVMS we have will be correct for this call 3291 guarantee(method_id == method->intrinsic_id(), "must match"); 3292 3293 const TypeFunc* tf = TypeFunc::make(method); 3294 int tfdc = tf->domain()->cnt(); 3295 CallJavaNode* slow_call; 3296 if (is_static) { 3297 assert(!is_virtual, ""); 3298 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3299 SharedRuntime::get_resolve_static_call_stub(), 3300 method, bci()); 3301 } else if (is_virtual) { 3302 null_check_receiver(method); 3303 int vtable_index = methodOopDesc::invalid_vtable_index; 3304 if (UseInlineCaches) { 3305 // Suppress the vtable call 3306 } else { 3307 // hashCode and clone are not a miranda methods, 3308 // so the vtable index is fixed. 3309 // No need to use the linkResolver to get it. 3310 vtable_index = method->vtable_index(); 3311 } 3312 slow_call = new(C, tfdc) CallDynamicJavaNode(tf, 3313 SharedRuntime::get_resolve_virtual_call_stub(), 3314 method, vtable_index, bci()); 3315 } else { // neither virtual nor static: opt_virtual 3316 null_check_receiver(method); 3317 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3318 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3319 method, bci()); 3320 slow_call->set_optimized_virtual(true); 3321 } 3322 set_arguments_for_java_call(slow_call); 3323 set_edges_for_java_call(slow_call); 3324 return slow_call; 3325 } 3326 3327 3328 //------------------------------inline_native_hashcode-------------------- 3329 // Build special case code for calls to hashCode on an object. 3330 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3331 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3332 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3333 3334 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3335 3336 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3337 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3338 TypeInt::INT); 3339 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3340 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3341 TypePtr::BOTTOM); 3342 Node* obj = NULL; 3343 if (!is_static) { 3344 // Check for hashing null object 3345 obj = null_check_receiver(callee()); 3346 if (stopped()) return true; // unconditionally null 3347 result_reg->init_req(_null_path, top()); 3348 result_val->init_req(_null_path, top()); 3349 } else { 3350 // Do a null check, and return zero if null. 3351 // System.identityHashCode(null) == 0 3352 obj = argument(0); 3353 Node* null_ctl = top(); 3354 obj = null_check_oop(obj, &null_ctl); 3355 result_reg->init_req(_null_path, null_ctl); 3356 result_val->init_req(_null_path, _gvn.intcon(0)); 3357 } 3358 3359 // Unconditionally null? Then return right away. 3360 if (stopped()) { 3361 set_control( result_reg->in(_null_path) ); 3362 if (!stopped()) 3363 push( result_val ->in(_null_path) ); 3364 return true; 3365 } 3366 3367 // After null check, get the object's klass. 3368 Node* obj_klass = load_object_klass(obj); 3369 3370 // This call may be virtual (invokevirtual) or bound (invokespecial). 3371 // For each case we generate slightly different code. 3372 3373 // We only go to the fast case code if we pass a number of guards. The 3374 // paths which do not pass are accumulated in the slow_region. 3375 RegionNode* slow_region = new (C, 1) RegionNode(1); 3376 record_for_igvn(slow_region); 3377 3378 // If this is a virtual call, we generate a funny guard. We pull out 3379 // the vtable entry corresponding to hashCode() from the target object. 3380 // If the target method which we are calling happens to be the native 3381 // Object hashCode() method, we pass the guard. We do not need this 3382 // guard for non-virtual calls -- the caller is known to be the native 3383 // Object hashCode(). 3384 if (is_virtual) { 3385 generate_virtual_guard(obj_klass, slow_region); 3386 } 3387 3388 // Get the header out of the object, use LoadMarkNode when available 3389 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3390 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS); 3391 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) ); 3392 3393 // Test the header to see if it is unlocked. 3394 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3395 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); 3396 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3397 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); 3398 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); 3399 3400 generate_slow_guard(test_unlocked, slow_region); 3401 3402 // Get the hash value and check to see that it has been properly assigned. 3403 // We depend on hash_mask being at most 32 bits and avoid the use of 3404 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3405 // vm: see markOop.hpp. 3406 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3407 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3408 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); 3409 // This hack lets the hash bits live anywhere in the mark object now, as long 3410 // as the shift drops the relevant bits into the low 32 bits. Note that 3411 // Java spec says that HashCode is an int so there's no point in capturing 3412 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3413 hshifted_header = ConvX2I(hshifted_header); 3414 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); 3415 3416 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3417 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); 3418 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); 3419 3420 generate_slow_guard(test_assigned, slow_region); 3421 3422 Node* init_mem = reset_memory(); 3423 // fill in the rest of the null path: 3424 result_io ->init_req(_null_path, i_o()); 3425 result_mem->init_req(_null_path, init_mem); 3426 3427 result_val->init_req(_fast_path, hash_val); 3428 result_reg->init_req(_fast_path, control()); 3429 result_io ->init_req(_fast_path, i_o()); 3430 result_mem->init_req(_fast_path, init_mem); 3431 3432 // Generate code for the slow case. We make a call to hashCode(). 3433 set_control(_gvn.transform(slow_region)); 3434 if (!stopped()) { 3435 // No need for PreserveJVMState, because we're using up the present state. 3436 set_all_memory(init_mem); 3437 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; 3438 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; 3439 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3440 Node* slow_result = set_results_for_java_call(slow_call); 3441 // this->control() comes from set_results_for_java_call 3442 result_reg->init_req(_slow_path, control()); 3443 result_val->init_req(_slow_path, slow_result); 3444 result_io ->set_req(_slow_path, i_o()); 3445 result_mem ->set_req(_slow_path, reset_memory()); 3446 } 3447 3448 // Return the combined state. 3449 set_i_o( _gvn.transform(result_io) ); 3450 set_all_memory( _gvn.transform(result_mem) ); 3451 push_result(result_reg, result_val); 3452 3453 return true; 3454 } 3455 3456 //---------------------------inline_native_getClass---------------------------- 3457 // Build special case code for calls to getClass on an object. 3458 bool LibraryCallKit::inline_native_getClass() { 3459 Node* obj = null_check_receiver(callee()); 3460 if (stopped()) return true; 3461 push( load_mirror_from_klass(load_object_klass(obj)) ); 3462 return true; 3463 } 3464 3465 //-----------------inline_native_Reflection_getCallerClass--------------------- 3466 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3467 // 3468 // NOTE that this code must perform the same logic as 3469 // vframeStream::security_get_caller_frame in that it must skip 3470 // Method.invoke() and auxiliary frames. 3471 3472 3473 3474 3475 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3476 ciMethod* method = callee(); 3477 3478 #ifndef PRODUCT 3479 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3480 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3481 } 3482 #endif 3483 3484 debug_only(int saved_sp = _sp); 3485 3486 // Argument words: (int depth) 3487 int nargs = 1; 3488 3489 _sp += nargs; 3490 Node* caller_depth_node = pop(); 3491 3492 assert(saved_sp == _sp, "must have correct argument count"); 3493 3494 // The depth value must be a constant in order for the runtime call 3495 // to be eliminated. 3496 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); 3497 if (caller_depth_type == NULL || !caller_depth_type->is_con()) { 3498 #ifndef PRODUCT 3499 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3500 tty->print_cr(" Bailing out because caller depth was not a constant"); 3501 } 3502 #endif 3503 return false; 3504 } 3505 // Note that the JVM state at this point does not include the 3506 // getCallerClass() frame which we are trying to inline. The 3507 // semantics of getCallerClass(), however, are that the "first" 3508 // frame is the getCallerClass() frame, so we subtract one from the 3509 // requested depth before continuing. We don't inline requests of 3510 // getCallerClass(0). 3511 int caller_depth = caller_depth_type->get_con() - 1; 3512 if (caller_depth < 0) { 3513 #ifndef PRODUCT 3514 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3515 tty->print_cr(" Bailing out because caller depth was %d", caller_depth); 3516 } 3517 #endif 3518 return false; 3519 } 3520 3521 if (!jvms()->has_method()) { 3522 #ifndef PRODUCT 3523 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3524 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3525 } 3526 #endif 3527 return false; 3528 } 3529 int _depth = jvms()->depth(); // cache call chain depth 3530 3531 // Walk back up the JVM state to find the caller at the required 3532 // depth. NOTE that this code must perform the same logic as 3533 // vframeStream::security_get_caller_frame in that it must skip 3534 // Method.invoke() and auxiliary frames. Note also that depth is 3535 // 1-based (1 is the bottom of the inlining). 3536 int inlining_depth = _depth; 3537 JVMState* caller_jvms = NULL; 3538 3539 if (inlining_depth > 0) { 3540 caller_jvms = jvms(); 3541 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); 3542 do { 3543 // The following if-tests should be performed in this order 3544 if (is_method_invoke_or_aux_frame(caller_jvms)) { 3545 // Skip a Method.invoke() or auxiliary frame 3546 } else if (caller_depth > 0) { 3547 // Skip real frame 3548 --caller_depth; 3549 } else { 3550 // We're done: reached desired caller after skipping. 3551 break; 3552 } 3553 caller_jvms = caller_jvms->caller(); 3554 --inlining_depth; 3555 } while (inlining_depth > 0); 3556 } 3557 3558 if (inlining_depth == 0) { 3559 #ifndef PRODUCT 3560 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3561 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); 3562 tty->print_cr(" JVM state at this point:"); 3563 for (int i = _depth; i >= 1; i--) { 3564 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3565 } 3566 } 3567 #endif 3568 return false; // Reached end of inlining 3569 } 3570 3571 // Acquire method holder as java.lang.Class 3572 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 3573 ciInstance* caller_mirror = caller_klass->java_mirror(); 3574 // Push this as a constant 3575 push(makecon(TypeInstPtr::make(caller_mirror))); 3576 #ifndef PRODUCT 3577 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3578 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); 3579 tty->print_cr(" JVM state at this point:"); 3580 for (int i = _depth; i >= 1; i--) { 3581 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3582 } 3583 } 3584 #endif 3585 return true; 3586 } 3587 3588 // Helper routine for above 3589 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { 3590 // Is this the Method.invoke method itself? 3591 if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke) 3592 return true; 3593 3594 // Is this a helper, defined somewhere underneath MethodAccessorImpl. 3595 ciKlass* k = jvms->method()->holder(); 3596 if (k->is_instance_klass()) { 3597 ciInstanceKlass* ik = k->as_instance_klass(); 3598 for (; ik != NULL; ik = ik->super()) { 3599 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && 3600 ik == env()->find_system_klass(ik->name())) { 3601 return true; 3602 } 3603 } 3604 } 3605 3606 return false; 3607 } 3608 3609 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by 3610 // inline_native_AtomicLong_attemptUpdate() but it has no way of 3611 // computing it since there is no lookup field by name function in the 3612 // CI interface. This is computed and set by inline_native_AtomicLong_get(). 3613 // Using a static variable here is safe even if we have multiple compilation 3614 // threads because the offset is constant. At worst the same offset will be 3615 // computed and stored multiple 3616 3617 bool LibraryCallKit::inline_native_AtomicLong_get() { 3618 // Restore the stack and pop off the argument 3619 _sp+=1; 3620 Node *obj = pop(); 3621 3622 // get the offset of the "value" field. Since the CI interfaces 3623 // does not provide a way to look up a field by name, we scan the bytecodes 3624 // to get the field index. We expect the first 2 instructions of the method 3625 // to be: 3626 // 0 aload_0 3627 // 1 getfield "value" 3628 ciMethod* method = callee(); 3629 if (value_field_offset == -1) 3630 { 3631 ciField* value_field; 3632 ciBytecodeStream iter(method); 3633 Bytecodes::Code bc = iter.next(); 3634 3635 if ((bc != Bytecodes::_aload_0) && 3636 ((bc != Bytecodes::_aload) || (iter.get_index() != 0))) 3637 return false; 3638 bc = iter.next(); 3639 if (bc != Bytecodes::_getfield) 3640 return false; 3641 bool ignore; 3642 value_field = iter.get_field(ignore); 3643 value_field_offset = value_field->offset_in_bytes(); 3644 } 3645 3646 // Null check without removing any arguments. 3647 _sp++; 3648 obj = do_null_check(obj, T_OBJECT); 3649 _sp--; 3650 // Check for locking null object 3651 if (stopped()) return true; 3652 3653 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3654 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3655 int alias_idx = C->get_alias_index(adr_type); 3656 3657 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr)); 3658 3659 push_pair(result); 3660 3661 return true; 3662 } 3663 3664 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() { 3665 // Restore the stack and pop off the arguments 3666 _sp+=5; 3667 Node *newVal = pop_pair(); 3668 Node *oldVal = pop_pair(); 3669 Node *obj = pop(); 3670 3671 // we need the offset of the "value" field which was computed when 3672 // inlining the get() method. Give up if we don't have it. 3673 if (value_field_offset == -1) 3674 return false; 3675 3676 // Null check without removing any arguments. 3677 _sp+=5; 3678 obj = do_null_check(obj, T_OBJECT); 3679 _sp-=5; 3680 // Check for locking null object 3681 if (stopped()) return true; 3682 3683 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3684 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3685 int alias_idx = C->get_alias_index(adr_type); 3686 3687 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal)); 3688 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 3689 set_memory(store_proj, alias_idx); 3690 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) ); 3691 3692 Node *result; 3693 // CMove node is not used to be able fold a possible check code 3694 // after attemptUpdate() call. This code could be transformed 3695 // into CMove node by loop optimizations. 3696 { 3697 RegionNode *r = new (C, 3) RegionNode(3); 3698 result = new (C, 3) PhiNode(r, TypeInt::BOOL); 3699 3700 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 3701 Node *iftrue = opt_iff(r, iff); 3702 r->init_req(1, iftrue); 3703 result->init_req(1, intcon(1)); 3704 result->init_req(2, intcon(0)); 3705 3706 set_control(_gvn.transform(r)); 3707 record_for_igvn(r); 3708 3709 C->set_has_split_ifs(true); // Has chance for split-if optimization 3710 } 3711 3712 push(_gvn.transform(result)); 3713 return true; 3714 } 3715 3716 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 3717 // restore the arguments 3718 _sp += arg_size(); 3719 3720 switch (id) { 3721 case vmIntrinsics::_floatToRawIntBits: 3722 push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); 3723 break; 3724 3725 case vmIntrinsics::_intBitsToFloat: 3726 push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); 3727 break; 3728 3729 case vmIntrinsics::_doubleToRawLongBits: 3730 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); 3731 break; 3732 3733 case vmIntrinsics::_longBitsToDouble: 3734 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); 3735 break; 3736 3737 case vmIntrinsics::_doubleToLongBits: { 3738 Node* value = pop_pair(); 3739 3740 // two paths (plus control) merge in a wood 3741 RegionNode *r = new (C, 3) RegionNode(3); 3742 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); 3743 3744 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); 3745 // Build the boolean node 3746 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 3747 3748 // Branch either way. 3749 // NaN case is less traveled, which makes all the difference. 3750 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3751 Node *opt_isnan = _gvn.transform(ifisnan); 3752 assert( opt_isnan->is_If(), "Expect an IfNode"); 3753 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3754 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 3755 3756 set_control(iftrue); 3757 3758 static const jlong nan_bits = CONST64(0x7ff8000000000000); 3759 Node *slow_result = longcon(nan_bits); // return NaN 3760 phi->init_req(1, _gvn.transform( slow_result )); 3761 r->init_req(1, iftrue); 3762 3763 // Else fall through 3764 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 3765 set_control(iffalse); 3766 3767 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); 3768 r->init_req(2, iffalse); 3769 3770 // Post merge 3771 set_control(_gvn.transform(r)); 3772 record_for_igvn(r); 3773 3774 Node* result = _gvn.transform(phi); 3775 assert(result->bottom_type()->isa_long(), "must be"); 3776 push_pair(result); 3777 3778 C->set_has_split_ifs(true); // Has chance for split-if optimization 3779 3780 break; 3781 } 3782 3783 case vmIntrinsics::_floatToIntBits: { 3784 Node* value = pop(); 3785 3786 // two paths (plus control) merge in a wood 3787 RegionNode *r = new (C, 3) RegionNode(3); 3788 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); 3789 3790 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); 3791 // Build the boolean node 3792 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 3793 3794 // Branch either way. 3795 // NaN case is less traveled, which makes all the difference. 3796 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3797 Node *opt_isnan = _gvn.transform(ifisnan); 3798 assert( opt_isnan->is_If(), "Expect an IfNode"); 3799 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3800 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 3801 3802 set_control(iftrue); 3803 3804 static const jint nan_bits = 0x7fc00000; 3805 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 3806 phi->init_req(1, _gvn.transform( slow_result )); 3807 r->init_req(1, iftrue); 3808 3809 // Else fall through 3810 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 3811 set_control(iffalse); 3812 3813 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); 3814 r->init_req(2, iffalse); 3815 3816 // Post merge 3817 set_control(_gvn.transform(r)); 3818 record_for_igvn(r); 3819 3820 Node* result = _gvn.transform(phi); 3821 assert(result->bottom_type()->isa_int(), "must be"); 3822 push(result); 3823 3824 C->set_has_split_ifs(true); // Has chance for split-if optimization 3825 3826 break; 3827 } 3828 3829 default: 3830 ShouldNotReachHere(); 3831 } 3832 3833 return true; 3834 } 3835 3836 #ifdef _LP64 3837 #define XTOP ,top() /*additional argument*/ 3838 #else //_LP64 3839 #define XTOP /*no additional argument*/ 3840 #endif //_LP64 3841 3842 //----------------------inline_unsafe_copyMemory------------------------- 3843 bool LibraryCallKit::inline_unsafe_copyMemory() { 3844 if (callee()->is_static()) return false; // caller must have the capability! 3845 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) 3846 assert(signature()->size() == nargs-1, "copy has 5 arguments"); 3847 null_check_receiver(callee()); // check then ignore argument(0) 3848 if (stopped()) return true; 3849 3850 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3851 3852 Node* src_ptr = argument(1); 3853 Node* src_off = ConvL2X(argument(2)); 3854 assert(argument(3)->is_top(), "2nd half of long"); 3855 Node* dst_ptr = argument(4); 3856 Node* dst_off = ConvL2X(argument(5)); 3857 assert(argument(6)->is_top(), "2nd half of long"); 3858 Node* size = ConvL2X(argument(7)); 3859 assert(argument(8)->is_top(), "2nd half of long"); 3860 3861 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 3862 "fieldOffset must be byte-scaled"); 3863 3864 Node* src = make_unsafe_address(src_ptr, src_off); 3865 Node* dst = make_unsafe_address(dst_ptr, dst_off); 3866 3867 // Conservatively insert a memory barrier on all memory slices. 3868 // Do not let writes of the copy source or destination float below the copy. 3869 insert_mem_bar(Op_MemBarCPUOrder); 3870 3871 // Call it. Note that the length argument is not scaled. 3872 make_runtime_call(RC_LEAF|RC_NO_FP, 3873 OptoRuntime::fast_arraycopy_Type(), 3874 StubRoutines::unsafe_arraycopy(), 3875 "unsafe_arraycopy", 3876 TypeRawPtr::BOTTOM, 3877 src, dst, size XTOP); 3878 3879 // Do not let reads of the copy destination float above the copy. 3880 insert_mem_bar(Op_MemBarCPUOrder); 3881 3882 return true; 3883 } 3884 3885 //------------------------clone_coping----------------------------------- 3886 // Helper function for inline_native_clone. 3887 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 3888 assert(obj_size != NULL, ""); 3889 Node* raw_obj = alloc_obj->in(1); 3890 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 3891 3892 if (ReduceBulkZeroing) { 3893 // We will be completely responsible for initializing this object - 3894 // mark Initialize node as complete. 3895 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 3896 // The object was just allocated - there should be no any stores! 3897 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 3898 } 3899 3900 // Cast to Object for arraycopy. 3901 // We can't use the original CheckCastPP since it should be moved 3902 // after the arraycopy to prevent stores flowing above it. 3903 Node* new_obj = new(C, 2) CheckCastPPNode(alloc_obj->in(0), raw_obj, 3904 TypeInstPtr::NOTNULL); 3905 new_obj = _gvn.transform(new_obj); 3906 // Substitute in the locally valid dest_oop. 3907 replace_in_map(alloc_obj, new_obj); 3908 3909 // Copy the fastest available way. 3910 // TODO: generate fields copies for small objects instead. 3911 Node* src = obj; 3912 Node* dest = new_obj; 3913 Node* size = _gvn.transform(obj_size); 3914 3915 // Exclude the header but include array length to copy by 8 bytes words. 3916 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 3917 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 3918 instanceOopDesc::base_offset_in_bytes(); 3919 // base_off: 3920 // 8 - 32-bit VM 3921 // 12 - 64-bit VM, compressed oops 3922 // 16 - 64-bit VM, normal oops 3923 if (base_off % BytesPerLong != 0) { 3924 assert(UseCompressedOops, ""); 3925 if (is_array) { 3926 // Exclude length to copy by 8 bytes words. 3927 base_off += sizeof(int); 3928 } else { 3929 // Include klass to copy by 8 bytes words. 3930 base_off = instanceOopDesc::klass_offset_in_bytes(); 3931 } 3932 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 3933 } 3934 src = basic_plus_adr(src, base_off); 3935 dest = basic_plus_adr(dest, base_off); 3936 3937 // Compute the length also, if needed: 3938 Node* countx = size; 3939 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); 3940 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); 3941 3942 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 3943 bool disjoint_bases = true; 3944 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 3945 src, NULL, dest, NULL, countx); 3946 3947 // If necessary, emit some card marks afterwards. (Non-arrays only.) 3948 if (card_mark) { 3949 assert(!is_array, ""); 3950 // Put in store barrier for any and all oops we are sticking 3951 // into this object. (We could avoid this if we could prove 3952 // that the object type contains no oop fields at all.) 3953 Node* no_particular_value = NULL; 3954 Node* no_particular_field = NULL; 3955 int raw_adr_idx = Compile::AliasIdxRaw; 3956 post_barrier(control(), 3957 memory(raw_adr_type), 3958 new_obj, 3959 no_particular_field, 3960 raw_adr_idx, 3961 no_particular_value, 3962 T_OBJECT, 3963 false); 3964 } 3965 3966 // Move the original CheckCastPP after arraycopy. 3967 _gvn.hash_delete(alloc_obj); 3968 alloc_obj->set_req(0, control()); 3969 // Replace raw memory edge with new CheckCastPP to have a live oop 3970 // at safepoints instead of raw value. 3971 assert(new_obj->is_CheckCastPP() && new_obj->in(1) == alloc_obj->in(1), "sanity"); 3972 alloc_obj->set_req(1, new_obj); // cast to the original type 3973 _gvn.hash_find_insert(alloc_obj); // put back into GVN table 3974 // Restore in the locally valid dest_oop. 3975 replace_in_map(new_obj, alloc_obj); 3976 } 3977 3978 //------------------------inline_native_clone---------------------------- 3979 // Here are the simple edge cases: 3980 // null receiver => normal trap 3981 // virtual and clone was overridden => slow path to out-of-line clone 3982 // not cloneable or finalizer => slow path to out-of-line Object.clone 3983 // 3984 // The general case has two steps, allocation and copying. 3985 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 3986 // 3987 // Copying also has two cases, oop arrays and everything else. 3988 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 3989 // Everything else uses the tight inline loop supplied by CopyArrayNode. 3990 // 3991 // These steps fold up nicely if and when the cloned object's klass 3992 // can be sharply typed as an object array, a type array, or an instance. 3993 // 3994 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 3995 int nargs = 1; 3996 Node* obj = null_check_receiver(callee()); 3997 if (stopped()) return true; 3998 Node* obj_klass = load_object_klass(obj); 3999 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4000 const TypeOopPtr* toop = ((tklass != NULL) 4001 ? tklass->as_instance_type() 4002 : TypeInstPtr::NOTNULL); 4003 4004 // Conservatively insert a memory barrier on all memory slices. 4005 // Do not let writes into the original float below the clone. 4006 insert_mem_bar(Op_MemBarCPUOrder); 4007 4008 // paths into result_reg: 4009 enum { 4010 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4011 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4012 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4013 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4014 PATH_LIMIT 4015 }; 4016 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4017 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 4018 TypeInstPtr::NOTNULL); 4019 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 4020 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 4021 TypePtr::BOTTOM); 4022 record_for_igvn(result_reg); 4023 4024 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4025 int raw_adr_idx = Compile::AliasIdxRaw; 4026 const bool raw_mem_only = true; 4027 4028 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4029 if (array_ctl != NULL) { 4030 // It's an array. 4031 PreserveJVMState pjvms(this); 4032 set_control(array_ctl); 4033 Node* obj_length = load_array_length(obj); 4034 Node* obj_size = NULL; 4035 Node* alloc_obj = new_array(obj_klass, obj_length, nargs, 4036 raw_mem_only, &obj_size); 4037 4038 if (!use_ReduceInitialCardMarks()) { 4039 // If it is an oop array, it requires very special treatment, 4040 // because card marking is required on each card of the array. 4041 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4042 if (is_obja != NULL) { 4043 PreserveJVMState pjvms2(this); 4044 set_control(is_obja); 4045 // Generate a direct call to the right arraycopy function(s). 4046 bool disjoint_bases = true; 4047 bool length_never_negative = true; 4048 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4049 obj, intcon(0), alloc_obj, intcon(0), 4050 obj_length, 4051 disjoint_bases, length_never_negative); 4052 result_reg->init_req(_objArray_path, control()); 4053 result_val->init_req(_objArray_path, alloc_obj); 4054 result_i_o ->set_req(_objArray_path, i_o()); 4055 result_mem ->set_req(_objArray_path, reset_memory()); 4056 } 4057 } 4058 // We can dispense with card marks if we know the allocation 4059 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4060 // causes the non-eden paths to simulate a fresh allocation, 4061 // insofar that no further card marks are required to initialize 4062 // the object. 4063 4064 // Otherwise, there are no card marks to worry about. 4065 4066 if (!stopped()) { 4067 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4068 4069 // Present the results of the copy. 4070 result_reg->init_req(_array_path, control()); 4071 result_val->init_req(_array_path, alloc_obj); 4072 result_i_o ->set_req(_array_path, i_o()); 4073 result_mem ->set_req(_array_path, reset_memory()); 4074 } 4075 } 4076 4077 // We only go to the instance fast case code if we pass a number of guards. 4078 // The paths which do not pass are accumulated in the slow_region. 4079 RegionNode* slow_region = new (C, 1) RegionNode(1); 4080 record_for_igvn(slow_region); 4081 if (!stopped()) { 4082 // It's an instance (we did array above). Make the slow-path tests. 4083 // If this is a virtual call, we generate a funny guard. We grab 4084 // the vtable entry corresponding to clone() from the target object. 4085 // If the target method which we are calling happens to be the 4086 // Object clone() method, we pass the guard. We do not need this 4087 // guard for non-virtual calls; the caller is known to be the native 4088 // Object clone(). 4089 if (is_virtual) { 4090 generate_virtual_guard(obj_klass, slow_region); 4091 } 4092 4093 // The object must be cloneable and must not have a finalizer. 4094 // Both of these conditions may be checked in a single test. 4095 // We could optimize the cloneable test further, but we don't care. 4096 generate_access_flags_guard(obj_klass, 4097 // Test both conditions: 4098 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4099 // Must be cloneable but not finalizer: 4100 JVM_ACC_IS_CLONEABLE, 4101 slow_region); 4102 } 4103 4104 if (!stopped()) { 4105 // It's an instance, and it passed the slow-path tests. 4106 PreserveJVMState pjvms(this); 4107 Node* obj_size = NULL; 4108 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size); 4109 4110 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4111 4112 // Present the results of the slow call. 4113 result_reg->init_req(_instance_path, control()); 4114 result_val->init_req(_instance_path, alloc_obj); 4115 result_i_o ->set_req(_instance_path, i_o()); 4116 result_mem ->set_req(_instance_path, reset_memory()); 4117 } 4118 4119 // Generate code for the slow case. We make a call to clone(). 4120 set_control(_gvn.transform(slow_region)); 4121 if (!stopped()) { 4122 PreserveJVMState pjvms(this); 4123 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4124 Node* slow_result = set_results_for_java_call(slow_call); 4125 // this->control() comes from set_results_for_java_call 4126 result_reg->init_req(_slow_path, control()); 4127 result_val->init_req(_slow_path, slow_result); 4128 result_i_o ->set_req(_slow_path, i_o()); 4129 result_mem ->set_req(_slow_path, reset_memory()); 4130 } 4131 4132 // Return the combined state. 4133 set_control( _gvn.transform(result_reg) ); 4134 set_i_o( _gvn.transform(result_i_o) ); 4135 set_all_memory( _gvn.transform(result_mem) ); 4136 4137 push(_gvn.transform(result_val)); 4138 4139 return true; 4140 } 4141 4142 4143 // constants for computing the copy function 4144 enum { 4145 COPYFUNC_UNALIGNED = 0, 4146 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize 4147 COPYFUNC_CONJOINT = 0, 4148 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend 4149 }; 4150 4151 // Note: The condition "disjoint" applies also for overlapping copies 4152 // where an descending copy is permitted (i.e., dest_offset <= src_offset). 4153 static address 4154 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) { 4155 int selector = 4156 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) + 4157 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT); 4158 4159 #define RETURN_STUB(xxx_arraycopy) { \ 4160 name = #xxx_arraycopy; \ 4161 return StubRoutines::xxx_arraycopy(); } 4162 4163 switch (t) { 4164 case T_BYTE: 4165 case T_BOOLEAN: 4166 switch (selector) { 4167 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy); 4168 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy); 4169 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy); 4170 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy); 4171 } 4172 case T_CHAR: 4173 case T_SHORT: 4174 switch (selector) { 4175 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy); 4176 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy); 4177 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy); 4178 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy); 4179 } 4180 case T_INT: 4181 case T_FLOAT: 4182 switch (selector) { 4183 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy); 4184 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy); 4185 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy); 4186 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy); 4187 } 4188 case T_DOUBLE: 4189 case T_LONG: 4190 switch (selector) { 4191 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy); 4192 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy); 4193 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy); 4194 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy); 4195 } 4196 case T_ARRAY: 4197 case T_OBJECT: 4198 switch (selector) { 4199 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy); 4200 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy); 4201 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy); 4202 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy); 4203 } 4204 default: 4205 ShouldNotReachHere(); 4206 return NULL; 4207 } 4208 4209 #undef RETURN_STUB 4210 } 4211 4212 //------------------------------basictype2arraycopy---------------------------- 4213 address LibraryCallKit::basictype2arraycopy(BasicType t, 4214 Node* src_offset, 4215 Node* dest_offset, 4216 bool disjoint_bases, 4217 const char* &name) { 4218 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4219 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4220 4221 bool aligned = false; 4222 bool disjoint = disjoint_bases; 4223 4224 // if the offsets are the same, we can treat the memory regions as 4225 // disjoint, because either the memory regions are in different arrays, 4226 // or they are identical (which we can treat as disjoint.) We can also 4227 // treat a copy with a destination index less that the source index 4228 // as disjoint since a low->high copy will work correctly in this case. 4229 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4230 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4231 // both indices are constants 4232 int s_offs = src_offset_inttype->get_con(); 4233 int d_offs = dest_offset_inttype->get_con(); 4234 int element_size = type2aelembytes(t); 4235 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4236 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4237 if (s_offs >= d_offs) disjoint = true; 4238 } else if (src_offset == dest_offset && src_offset != NULL) { 4239 // This can occur if the offsets are identical non-constants. 4240 disjoint = true; 4241 } 4242 4243 return select_arraycopy_function(t, aligned, disjoint, name); 4244 } 4245 4246 4247 //------------------------------inline_arraycopy----------------------- 4248 bool LibraryCallKit::inline_arraycopy() { 4249 // Restore the stack and pop off the arguments. 4250 int nargs = 5; // 2 oops, 3 ints, no size_t or long 4251 assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); 4252 4253 Node *src = argument(0); 4254 Node *src_offset = argument(1); 4255 Node *dest = argument(2); 4256 Node *dest_offset = argument(3); 4257 Node *length = argument(4); 4258 4259 // Compile time checks. If any of these checks cannot be verified at compile time, 4260 // we do not make a fast path for this call. Instead, we let the call remain as it 4261 // is. The checks we choose to mandate at compile time are: 4262 // 4263 // (1) src and dest are arrays. 4264 const Type* src_type = src->Value(&_gvn); 4265 const Type* dest_type = dest->Value(&_gvn); 4266 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4267 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4268 if (top_src == NULL || top_src->klass() == NULL || 4269 top_dest == NULL || top_dest->klass() == NULL) { 4270 // Conservatively insert a memory barrier on all memory slices. 4271 // Do not let writes into the source float below the arraycopy. 4272 insert_mem_bar(Op_MemBarCPUOrder); 4273 4274 // Call StubRoutines::generic_arraycopy stub. 4275 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4276 src, src_offset, dest, dest_offset, length); 4277 4278 // Do not let reads from the destination float above the arraycopy. 4279 // Since we cannot type the arrays, we don't know which slices 4280 // might be affected. We could restrict this barrier only to those 4281 // memory slices which pertain to array elements--but don't bother. 4282 if (!InsertMemBarAfterArraycopy) 4283 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4284 insert_mem_bar(Op_MemBarCPUOrder); 4285 return true; 4286 } 4287 4288 // (2) src and dest arrays must have elements of the same BasicType 4289 // Figure out the size and type of the elements we will be copying. 4290 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4291 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4292 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4293 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4294 4295 if (src_elem != dest_elem || dest_elem == T_VOID) { 4296 // The component types are not the same or are not recognized. Punt. 4297 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4298 generate_slow_arraycopy(TypePtr::BOTTOM, 4299 src, src_offset, dest, dest_offset, length); 4300 return true; 4301 } 4302 4303 //--------------------------------------------------------------------------- 4304 // We will make a fast path for this call to arraycopy. 4305 4306 // We have the following tests left to perform: 4307 // 4308 // (3) src and dest must not be null. 4309 // (4) src_offset must not be negative. 4310 // (5) dest_offset must not be negative. 4311 // (6) length must not be negative. 4312 // (7) src_offset + length must not exceed length of src. 4313 // (8) dest_offset + length must not exceed length of dest. 4314 // (9) each element of an oop array must be assignable 4315 4316 RegionNode* slow_region = new (C, 1) RegionNode(1); 4317 record_for_igvn(slow_region); 4318 4319 // (3) operands must not be null 4320 // We currently perform our null checks with the do_null_check routine. 4321 // This means that the null exceptions will be reported in the caller 4322 // rather than (correctly) reported inside of the native arraycopy call. 4323 // This should be corrected, given time. We do our null check with the 4324 // stack pointer restored. 4325 _sp += nargs; 4326 src = do_null_check(src, T_ARRAY); 4327 dest = do_null_check(dest, T_ARRAY); 4328 _sp -= nargs; 4329 4330 // (4) src_offset must not be negative. 4331 generate_negative_guard(src_offset, slow_region); 4332 4333 // (5) dest_offset must not be negative. 4334 generate_negative_guard(dest_offset, slow_region); 4335 4336 // (6) length must not be negative (moved to generate_arraycopy()). 4337 // generate_negative_guard(length, slow_region); 4338 4339 // (7) src_offset + length must not exceed length of src. 4340 generate_limit_guard(src_offset, length, 4341 load_array_length(src), 4342 slow_region); 4343 4344 // (8) dest_offset + length must not exceed length of dest. 4345 generate_limit_guard(dest_offset, length, 4346 load_array_length(dest), 4347 slow_region); 4348 4349 // (9) each element of an oop array must be assignable 4350 // The generate_arraycopy subroutine checks this. 4351 4352 // This is where the memory effects are placed: 4353 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4354 generate_arraycopy(adr_type, dest_elem, 4355 src, src_offset, dest, dest_offset, length, 4356 false, false, slow_region); 4357 4358 return true; 4359 } 4360 4361 //-----------------------------generate_arraycopy---------------------- 4362 // Generate an optimized call to arraycopy. 4363 // Caller must guard against non-arrays. 4364 // Caller must determine a common array basic-type for both arrays. 4365 // Caller must validate offsets against array bounds. 4366 // The slow_region has already collected guard failure paths 4367 // (such as out of bounds length or non-conformable array types). 4368 // The generated code has this shape, in general: 4369 // 4370 // if (length == 0) return // via zero_path 4371 // slowval = -1 4372 // if (types unknown) { 4373 // slowval = call generic copy loop 4374 // if (slowval == 0) return // via checked_path 4375 // } else if (indexes in bounds) { 4376 // if ((is object array) && !(array type check)) { 4377 // slowval = call checked copy loop 4378 // if (slowval == 0) return // via checked_path 4379 // } else { 4380 // call bulk copy loop 4381 // return // via fast_path 4382 // } 4383 // } 4384 // // adjust params for remaining work: 4385 // if (slowval != -1) { 4386 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4387 // } 4388 // slow_region: 4389 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4390 // return // via slow_call_path 4391 // 4392 // This routine is used from several intrinsics: System.arraycopy, 4393 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4394 // 4395 void 4396 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4397 BasicType basic_elem_type, 4398 Node* src, Node* src_offset, 4399 Node* dest, Node* dest_offset, 4400 Node* copy_length, 4401 bool disjoint_bases, 4402 bool length_never_negative, 4403 RegionNode* slow_region) { 4404 4405 if (slow_region == NULL) { 4406 slow_region = new(C,1) RegionNode(1); 4407 record_for_igvn(slow_region); 4408 } 4409 4410 Node* original_dest = dest; 4411 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4412 bool must_clear_dest = false; 4413 4414 // See if this is the initialization of a newly-allocated array. 4415 // If so, we will take responsibility here for initializing it to zero. 4416 // (Note: Because tightly_coupled_allocation performs checks on the 4417 // out-edges of the dest, we need to avoid making derived pointers 4418 // from it until we have checked its uses.) 4419 if (ReduceBulkZeroing 4420 && !ZeroTLAB // pointless if already zeroed 4421 && basic_elem_type != T_CONFLICT // avoid corner case 4422 && !_gvn.eqv_uncast(src, dest) 4423 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4424 != NULL) 4425 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4426 && alloc->maybe_set_complete(&_gvn)) { 4427 // "You break it, you buy it." 4428 InitializeNode* init = alloc->initialization(); 4429 assert(init->is_complete(), "we just did this"); 4430 assert(dest->is_CheckCastPP(), "sanity"); 4431 assert(dest->in(0)->in(0) == init, "dest pinned"); 4432 4433 // Cast to Object for arraycopy. 4434 // We can't use the original CheckCastPP since it should be moved 4435 // after the arraycopy to prevent stores flowing above it. 4436 Node* new_obj = new(C, 2) CheckCastPPNode(dest->in(0), dest->in(1), 4437 TypeInstPtr::NOTNULL); 4438 dest = _gvn.transform(new_obj); 4439 // Substitute in the locally valid dest_oop. 4440 replace_in_map(original_dest, dest); 4441 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4442 // From this point on, every exit path is responsible for 4443 // initializing any non-copied parts of the object to zero. 4444 must_clear_dest = true; 4445 } else { 4446 // No zeroing elimination here. 4447 alloc = NULL; 4448 //original_dest = dest; 4449 //must_clear_dest = false; 4450 } 4451 4452 // Results are placed here: 4453 enum { fast_path = 1, // normal void-returning assembly stub 4454 checked_path = 2, // special assembly stub with cleanup 4455 slow_call_path = 3, // something went wrong; call the VM 4456 zero_path = 4, // bypass when length of copy is zero 4457 bcopy_path = 5, // copy primitive array by 64-bit blocks 4458 PATH_LIMIT = 6 4459 }; 4460 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4461 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO); 4462 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type); 4463 record_for_igvn(result_region); 4464 _gvn.set_type_bottom(result_i_o); 4465 _gvn.set_type_bottom(result_memory); 4466 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4467 4468 // The slow_control path: 4469 Node* slow_control; 4470 Node* slow_i_o = i_o(); 4471 Node* slow_mem = memory(adr_type); 4472 debug_only(slow_control = (Node*) badAddress); 4473 4474 // Checked control path: 4475 Node* checked_control = top(); 4476 Node* checked_mem = NULL; 4477 Node* checked_i_o = NULL; 4478 Node* checked_value = NULL; 4479 4480 if (basic_elem_type == T_CONFLICT) { 4481 assert(!must_clear_dest, ""); 4482 Node* cv = generate_generic_arraycopy(adr_type, 4483 src, src_offset, dest, dest_offset, 4484 copy_length); 4485 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4486 checked_control = control(); 4487 checked_i_o = i_o(); 4488 checked_mem = memory(adr_type); 4489 checked_value = cv; 4490 set_control(top()); // no fast path 4491 } 4492 4493 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4494 if (not_pos != NULL) { 4495 PreserveJVMState pjvms(this); 4496 set_control(not_pos); 4497 4498 // (6) length must not be negative. 4499 if (!length_never_negative) { 4500 generate_negative_guard(copy_length, slow_region); 4501 } 4502 4503 if (!stopped() && must_clear_dest) { 4504 Node* dest_length = alloc->in(AllocateNode::ALength); 4505 if (_gvn.eqv_uncast(copy_length, dest_length) 4506 || _gvn.find_int_con(dest_length, 1) <= 0) { 4507 // There is no zeroing to do. 4508 } else { 4509 // Clear the whole thing since there are no source elements to copy. 4510 generate_clear_array(adr_type, dest, basic_elem_type, 4511 intcon(0), NULL, 4512 alloc->in(AllocateNode::AllocSize)); 4513 } 4514 } 4515 4516 // Use a secondary InitializeNode as raw memory barrier. 4517 // Currently it is needed only on this path since other 4518 // paths have stub or runtime calls as raw memory barriers. 4519 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4520 Compile::AliasIdxRaw, 4521 top())->as_Initialize(); 4522 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4523 4524 // Present the results of the fast call. 4525 result_region->init_req(zero_path, control()); 4526 result_i_o ->init_req(zero_path, i_o()); 4527 result_memory->init_req(zero_path, memory(adr_type)); 4528 } 4529 4530 if (!stopped() && must_clear_dest) { 4531 // We have to initialize the *uncopied* part of the array to zero. 4532 // The copy destination is the slice dest[off..off+len]. The other slices 4533 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4534 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4535 Node* dest_length = alloc->in(AllocateNode::ALength); 4536 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset, 4537 copy_length) ); 4538 4539 // If there is a head section that needs zeroing, do it now. 4540 if (find_int_con(dest_offset, -1) != 0) { 4541 generate_clear_array(adr_type, dest, basic_elem_type, 4542 intcon(0), dest_offset, 4543 NULL); 4544 } 4545 4546 // Next, perform a dynamic check on the tail length. 4547 // It is often zero, and we can win big if we prove this. 4548 // There are two wins: Avoid generating the ClearArray 4549 // with its attendant messy index arithmetic, and upgrade 4550 // the copy to a more hardware-friendly word size of 64 bits. 4551 Node* tail_ctl = NULL; 4552 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) { 4553 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) ); 4554 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) ); 4555 tail_ctl = generate_slow_guard(bol_lt, NULL); 4556 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4557 } 4558 4559 // At this point, let's assume there is no tail. 4560 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4561 // There is no tail. Try an upgrade to a 64-bit copy. 4562 bool didit = false; 4563 { PreserveJVMState pjvms(this); 4564 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4565 src, src_offset, dest, dest_offset, 4566 dest_size); 4567 if (didit) { 4568 // Present the results of the block-copying fast call. 4569 result_region->init_req(bcopy_path, control()); 4570 result_i_o ->init_req(bcopy_path, i_o()); 4571 result_memory->init_req(bcopy_path, memory(adr_type)); 4572 } 4573 } 4574 if (didit) 4575 set_control(top()); // no regular fast path 4576 } 4577 4578 // Clear the tail, if any. 4579 if (tail_ctl != NULL) { 4580 Node* notail_ctl = stopped() ? NULL : control(); 4581 set_control(tail_ctl); 4582 if (notail_ctl == NULL) { 4583 generate_clear_array(adr_type, dest, basic_elem_type, 4584 dest_tail, NULL, 4585 dest_size); 4586 } else { 4587 // Make a local merge. 4588 Node* done_ctl = new(C,3) RegionNode(3); 4589 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type); 4590 done_ctl->init_req(1, notail_ctl); 4591 done_mem->init_req(1, memory(adr_type)); 4592 generate_clear_array(adr_type, dest, basic_elem_type, 4593 dest_tail, NULL, 4594 dest_size); 4595 done_ctl->init_req(2, control()); 4596 done_mem->init_req(2, memory(adr_type)); 4597 set_control( _gvn.transform(done_ctl) ); 4598 set_memory( _gvn.transform(done_mem), adr_type ); 4599 } 4600 } 4601 } 4602 4603 BasicType copy_type = basic_elem_type; 4604 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 4605 if (!stopped() && copy_type == T_OBJECT) { 4606 // If src and dest have compatible element types, we can copy bits. 4607 // Types S[] and D[] are compatible if D is a supertype of S. 4608 // 4609 // If they are not, we will use checked_oop_disjoint_arraycopy, 4610 // which performs a fast optimistic per-oop check, and backs off 4611 // further to JVM_ArrayCopy on the first per-oop check that fails. 4612 // (Actually, we don't move raw bits only; the GC requires card marks.) 4613 4614 // Get the klassOop for both src and dest 4615 Node* src_klass = load_object_klass(src); 4616 Node* dest_klass = load_object_klass(dest); 4617 4618 // Generate the subtype check. 4619 // This might fold up statically, or then again it might not. 4620 // 4621 // Non-static example: Copying List<String>.elements to a new String[]. 4622 // The backing store for a List<String> is always an Object[], 4623 // but its elements are always type String, if the generic types 4624 // are correct at the source level. 4625 // 4626 // Test S[] against D[], not S against D, because (probably) 4627 // the secondary supertype cache is less busy for S[] than S. 4628 // This usually only matters when D is an interface. 4629 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4630 // Plug failing path into checked_oop_disjoint_arraycopy 4631 if (not_subtype_ctrl != top()) { 4632 PreserveJVMState pjvms(this); 4633 set_control(not_subtype_ctrl); 4634 // (At this point we can assume disjoint_bases, since types differ.) 4635 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc); 4636 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 4637 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 4638 Node* dest_elem_klass = _gvn.transform(n1); 4639 Node* cv = generate_checkcast_arraycopy(adr_type, 4640 dest_elem_klass, 4641 src, src_offset, dest, dest_offset, 4642 copy_length); 4643 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4644 checked_control = control(); 4645 checked_i_o = i_o(); 4646 checked_mem = memory(adr_type); 4647 checked_value = cv; 4648 } 4649 // At this point we know we do not need type checks on oop stores. 4650 4651 // Let's see if we need card marks: 4652 if (alloc != NULL && use_ReduceInitialCardMarks()) { 4653 // If we do not need card marks, copy using the jint or jlong stub. 4654 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 4655 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 4656 "sizes agree"); 4657 } 4658 } 4659 4660 if (!stopped()) { 4661 // Generate the fast path, if possible. 4662 PreserveJVMState pjvms(this); 4663 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 4664 src, src_offset, dest, dest_offset, 4665 ConvI2X(copy_length)); 4666 4667 // Present the results of the fast call. 4668 result_region->init_req(fast_path, control()); 4669 result_i_o ->init_req(fast_path, i_o()); 4670 result_memory->init_req(fast_path, memory(adr_type)); 4671 } 4672 4673 // Here are all the slow paths up to this point, in one bundle: 4674 slow_control = top(); 4675 if (slow_region != NULL) 4676 slow_control = _gvn.transform(slow_region); 4677 debug_only(slow_region = (RegionNode*)badAddress); 4678 4679 set_control(checked_control); 4680 if (!stopped()) { 4681 // Clean up after the checked call. 4682 // The returned value is either 0 or -1^K, 4683 // where K = number of partially transferred array elements. 4684 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) ); 4685 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 4686 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4687 4688 // If it is 0, we are done, so transfer to the end. 4689 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) ); 4690 result_region->init_req(checked_path, checks_done); 4691 result_i_o ->init_req(checked_path, checked_i_o); 4692 result_memory->init_req(checked_path, checked_mem); 4693 4694 // If it is not zero, merge into the slow call. 4695 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) )); 4696 RegionNode* slow_reg2 = new(C, 3) RegionNode(3); 4697 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO); 4698 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type); 4699 record_for_igvn(slow_reg2); 4700 slow_reg2 ->init_req(1, slow_control); 4701 slow_i_o2 ->init_req(1, slow_i_o); 4702 slow_mem2 ->init_req(1, slow_mem); 4703 slow_reg2 ->init_req(2, control()); 4704 slow_i_o2 ->init_req(2, checked_i_o); 4705 slow_mem2 ->init_req(2, checked_mem); 4706 4707 slow_control = _gvn.transform(slow_reg2); 4708 slow_i_o = _gvn.transform(slow_i_o2); 4709 slow_mem = _gvn.transform(slow_mem2); 4710 4711 if (alloc != NULL) { 4712 // We'll restart from the very beginning, after zeroing the whole thing. 4713 // This can cause double writes, but that's OK since dest is brand new. 4714 // So we ignore the low 31 bits of the value returned from the stub. 4715 } else { 4716 // We must continue the copy exactly where it failed, or else 4717 // another thread might see the wrong number of writes to dest. 4718 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) ); 4719 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT); 4720 slow_offset->init_req(1, intcon(0)); 4721 slow_offset->init_req(2, checked_offset); 4722 slow_offset = _gvn.transform(slow_offset); 4723 4724 // Adjust the arguments by the conditionally incoming offset. 4725 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) ); 4726 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) ); 4727 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) ); 4728 4729 // Tweak the node variables to adjust the code produced below: 4730 src_offset = src_off_plus; 4731 dest_offset = dest_off_plus; 4732 copy_length = length_minus; 4733 } 4734 } 4735 4736 set_control(slow_control); 4737 if (!stopped()) { 4738 // Generate the slow path, if needed. 4739 PreserveJVMState pjvms(this); // replace_in_map may trash the map 4740 4741 set_memory(slow_mem, adr_type); 4742 set_i_o(slow_i_o); 4743 4744 if (must_clear_dest) { 4745 generate_clear_array(adr_type, dest, basic_elem_type, 4746 intcon(0), NULL, 4747 alloc->in(AllocateNode::AllocSize)); 4748 } 4749 4750 generate_slow_arraycopy(adr_type, 4751 src, src_offset, dest, dest_offset, 4752 copy_length); 4753 4754 result_region->init_req(slow_call_path, control()); 4755 result_i_o ->init_req(slow_call_path, i_o()); 4756 result_memory->init_req(slow_call_path, memory(adr_type)); 4757 } 4758 4759 // Remove unused edges. 4760 for (uint i = 1; i < result_region->req(); i++) { 4761 if (result_region->in(i) == NULL) 4762 result_region->init_req(i, top()); 4763 } 4764 4765 // Finished; return the combined state. 4766 set_control( _gvn.transform(result_region) ); 4767 set_i_o( _gvn.transform(result_i_o) ); 4768 set_memory( _gvn.transform(result_memory), adr_type ); 4769 4770 if (dest != original_dest) { 4771 // Pin the "finished" array node after the arraycopy/zeroing operations. 4772 _gvn.hash_delete(original_dest); 4773 original_dest->set_req(0, control()); 4774 // Replace raw memory edge with new CheckCastPP to have a live oop 4775 // at safepoints instead of raw value. 4776 assert(dest->is_CheckCastPP() && dest->in(1) == original_dest->in(1), "sanity"); 4777 original_dest->set_req(1, dest); // cast to the original type 4778 _gvn.hash_find_insert(original_dest); // put back into GVN table 4779 // Restore in the locally valid dest_oop. 4780 replace_in_map(dest, original_dest); 4781 } 4782 // The memory edges above are precise in order to model effects around 4783 // array copies accurately to allow value numbering of field loads around 4784 // arraycopy. Such field loads, both before and after, are common in Java 4785 // collections and similar classes involving header/array data structures. 4786 // 4787 // But with low number of register or when some registers are used or killed 4788 // by arraycopy calls it causes registers spilling on stack. See 6544710. 4789 // The next memory barrier is added to avoid it. If the arraycopy can be 4790 // optimized away (which it can, sometimes) then we can manually remove 4791 // the membar also. 4792 if (InsertMemBarAfterArraycopy) 4793 insert_mem_bar(Op_MemBarCPUOrder); 4794 } 4795 4796 4797 // Helper function which determines if an arraycopy immediately follows 4798 // an allocation, with no intervening tests or other escapes for the object. 4799 AllocateArrayNode* 4800 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4801 RegionNode* slow_region) { 4802 if (stopped()) return NULL; // no fast path 4803 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4804 4805 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4806 if (alloc == NULL) return NULL; 4807 4808 Node* rawmem = memory(Compile::AliasIdxRaw); 4809 // Is the allocation's memory state untouched? 4810 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4811 // Bail out if there have been raw-memory effects since the allocation. 4812 // (Example: There might have been a call or safepoint.) 4813 return NULL; 4814 } 4815 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4816 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4817 return NULL; 4818 } 4819 4820 // There must be no unexpected observers of this allocation. 4821 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4822 Node* obs = ptr->fast_out(i); 4823 if (obs != this->map()) { 4824 return NULL; 4825 } 4826 } 4827 4828 // This arraycopy must unconditionally follow the allocation of the ptr. 4829 Node* alloc_ctl = ptr->in(0); 4830 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4831 4832 Node* ctl = control(); 4833 while (ctl != alloc_ctl) { 4834 // There may be guards which feed into the slow_region. 4835 // Any other control flow means that we might not get a chance 4836 // to finish initializing the allocated object. 4837 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4838 IfNode* iff = ctl->in(0)->as_If(); 4839 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 4840 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4841 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4842 ctl = iff->in(0); // This test feeds the known slow_region. 4843 continue; 4844 } 4845 // One more try: Various low-level checks bottom out in 4846 // uncommon traps. If the debug-info of the trap omits 4847 // any reference to the allocation, as we've already 4848 // observed, then there can be no objection to the trap. 4849 bool found_trap = false; 4850 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 4851 Node* obs = not_ctl->fast_out(j); 4852 if (obs->in(0) == not_ctl && obs->is_Call() && 4853 (obs->as_Call()->entry_point() == 4854 SharedRuntime::uncommon_trap_blob()->instructions_begin())) { 4855 found_trap = true; break; 4856 } 4857 } 4858 if (found_trap) { 4859 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 4860 continue; 4861 } 4862 } 4863 return NULL; 4864 } 4865 4866 // If we get this far, we have an allocation which immediately 4867 // precedes the arraycopy, and we can take over zeroing the new object. 4868 // The arraycopy will finish the initialization, and provide 4869 // a new control state to which we will anchor the destination pointer. 4870 4871 return alloc; 4872 } 4873 4874 // Helper for initialization of arrays, creating a ClearArray. 4875 // It writes zero bits in [start..end), within the body of an array object. 4876 // The memory effects are all chained onto the 'adr_type' alias category. 4877 // 4878 // Since the object is otherwise uninitialized, we are free 4879 // to put a little "slop" around the edges of the cleared area, 4880 // as long as it does not go back into the array's header, 4881 // or beyond the array end within the heap. 4882 // 4883 // The lower edge can be rounded down to the nearest jint and the 4884 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 4885 // 4886 // Arguments: 4887 // adr_type memory slice where writes are generated 4888 // dest oop of the destination array 4889 // basic_elem_type element type of the destination 4890 // slice_idx array index of first element to store 4891 // slice_len number of elements to store (or NULL) 4892 // dest_size total size in bytes of the array object 4893 // 4894 // Exactly one of slice_len or dest_size must be non-NULL. 4895 // If dest_size is non-NULL, zeroing extends to the end of the object. 4896 // If slice_len is non-NULL, the slice_idx value must be a constant. 4897 void 4898 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 4899 Node* dest, 4900 BasicType basic_elem_type, 4901 Node* slice_idx, 4902 Node* slice_len, 4903 Node* dest_size) { 4904 // one or the other but not both of slice_len and dest_size: 4905 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 4906 if (slice_len == NULL) slice_len = top(); 4907 if (dest_size == NULL) dest_size = top(); 4908 4909 // operate on this memory slice: 4910 Node* mem = memory(adr_type); // memory slice to operate on 4911 4912 // scaling and rounding of indexes: 4913 int scale = exact_log2(type2aelembytes(basic_elem_type)); 4914 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 4915 int clear_low = (-1 << scale) & (BytesPerInt - 1); 4916 int bump_bit = (-1 << scale) & BytesPerInt; 4917 4918 // determine constant starts and ends 4919 const intptr_t BIG_NEG = -128; 4920 assert(BIG_NEG + 2*abase < 0, "neg enough"); 4921 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 4922 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 4923 if (slice_len_con == 0) { 4924 return; // nothing to do here 4925 } 4926 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 4927 intptr_t end_con = find_intptr_t_con(dest_size, -1); 4928 if (slice_idx_con >= 0 && slice_len_con >= 0) { 4929 assert(end_con < 0, "not two cons"); 4930 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 4931 BytesPerLong); 4932 } 4933 4934 if (start_con >= 0 && end_con >= 0) { 4935 // Constant start and end. Simple. 4936 mem = ClearArrayNode::clear_memory(control(), mem, dest, 4937 start_con, end_con, &_gvn); 4938 } else if (start_con >= 0 && dest_size != top()) { 4939 // Constant start, pre-rounded end after the tail of the array. 4940 Node* end = dest_size; 4941 mem = ClearArrayNode::clear_memory(control(), mem, dest, 4942 start_con, end, &_gvn); 4943 } else if (start_con >= 0 && slice_len != top()) { 4944 // Constant start, non-constant end. End needs rounding up. 4945 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 4946 intptr_t end_base = abase + (slice_idx_con << scale); 4947 int end_round = (-1 << scale) & (BytesPerLong - 1); 4948 Node* end = ConvI2X(slice_len); 4949 if (scale != 0) 4950 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) )); 4951 end_base += end_round; 4952 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) ); 4953 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) ); 4954 mem = ClearArrayNode::clear_memory(control(), mem, dest, 4955 start_con, end, &_gvn); 4956 } else if (start_con < 0 && dest_size != top()) { 4957 // Non-constant start, pre-rounded end after the tail of the array. 4958 // This is almost certainly a "round-to-end" operation. 4959 Node* start = slice_idx; 4960 start = ConvI2X(start); 4961 if (scale != 0) 4962 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) )); 4963 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) ); 4964 if ((bump_bit | clear_low) != 0) { 4965 int to_clear = (bump_bit | clear_low); 4966 // Align up mod 8, then store a jint zero unconditionally 4967 // just before the mod-8 boundary. 4968 if (((abase + bump_bit) & ~to_clear) - bump_bit 4969 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 4970 bump_bit = 0; 4971 assert((abase & to_clear) == 0, "array base must be long-aligned"); 4972 } else { 4973 // Bump 'start' up to (or past) the next jint boundary: 4974 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) ); 4975 assert((abase & clear_low) == 0, "array base must be int-aligned"); 4976 } 4977 // Round bumped 'start' down to jlong boundary in body of array. 4978 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) ); 4979 if (bump_bit != 0) { 4980 // Store a zero to the immediately preceding jint: 4981 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) ); 4982 Node* p1 = basic_plus_adr(dest, x1); 4983 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 4984 mem = _gvn.transform(mem); 4985 } 4986 } 4987 Node* end = dest_size; // pre-rounded 4988 mem = ClearArrayNode::clear_memory(control(), mem, dest, 4989 start, end, &_gvn); 4990 } else { 4991 // Non-constant start, unrounded non-constant end. 4992 // (Nobody zeroes a random midsection of an array using this routine.) 4993 ShouldNotReachHere(); // fix caller 4994 } 4995 4996 // Done. 4997 set_memory(mem, adr_type); 4998 } 4999 5000 5001 bool 5002 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5003 BasicType basic_elem_type, 5004 AllocateNode* alloc, 5005 Node* src, Node* src_offset, 5006 Node* dest, Node* dest_offset, 5007 Node* dest_size) { 5008 // See if there is an advantage from block transfer. 5009 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5010 if (scale >= LogBytesPerLong) 5011 return false; // it is already a block transfer 5012 5013 // Look at the alignment of the starting offsets. 5014 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5015 const intptr_t BIG_NEG = -128; 5016 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5017 5018 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale); 5019 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale); 5020 if (src_off < 0 || dest_off < 0) 5021 // At present, we can only understand constants. 5022 return false; 5023 5024 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5025 // Non-aligned; too bad. 5026 // One more chance: Pick off an initial 32-bit word. 5027 // This is a common case, since abase can be odd mod 8. 5028 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5029 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5030 Node* sptr = basic_plus_adr(src, src_off); 5031 Node* dptr = basic_plus_adr(dest, dest_off); 5032 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 5033 store_to_memory(control(), dptr, sval, T_INT, adr_type); 5034 src_off += BytesPerInt; 5035 dest_off += BytesPerInt; 5036 } else { 5037 return false; 5038 } 5039 } 5040 assert(src_off % BytesPerLong == 0, ""); 5041 assert(dest_off % BytesPerLong == 0, ""); 5042 5043 // Do this copy by giant steps. 5044 Node* sptr = basic_plus_adr(src, src_off); 5045 Node* dptr = basic_plus_adr(dest, dest_off); 5046 Node* countx = dest_size; 5047 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) ); 5048 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) ); 5049 5050 bool disjoint_bases = true; // since alloc != NULL 5051 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5052 sptr, NULL, dptr, NULL, countx); 5053 5054 return true; 5055 } 5056 5057 5058 // Helper function; generates code for the slow case. 5059 // We make a call to a runtime method which emulates the native method, 5060 // but without the native wrapper overhead. 5061 void 5062 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5063 Node* src, Node* src_offset, 5064 Node* dest, Node* dest_offset, 5065 Node* copy_length) { 5066 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5067 OptoRuntime::slow_arraycopy_Type(), 5068 OptoRuntime::slow_arraycopy_Java(), 5069 "slow_arraycopy", adr_type, 5070 src, src_offset, dest, dest_offset, 5071 copy_length); 5072 5073 // Handle exceptions thrown by this fellow: 5074 make_slow_call_ex(call, env()->Throwable_klass(), false); 5075 } 5076 5077 // Helper function; generates code for cases requiring runtime checks. 5078 Node* 5079 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5080 Node* dest_elem_klass, 5081 Node* src, Node* src_offset, 5082 Node* dest, Node* dest_offset, 5083 Node* copy_length) { 5084 if (stopped()) return NULL; 5085 5086 address copyfunc_addr = StubRoutines::checkcast_arraycopy(); 5087 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5088 return NULL; 5089 } 5090 5091 // Pick out the parameters required to perform a store-check 5092 // for the target array. This is an optimistic check. It will 5093 // look in each non-null element's class, at the desired klass's 5094 // super_check_offset, for the desired klass. 5095 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc); 5096 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5097 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM); 5098 Node* check_offset = _gvn.transform(n3); 5099 Node* check_value = dest_elem_klass; 5100 5101 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5102 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5103 5104 // (We know the arrays are never conjoint, because their types differ.) 5105 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5106 OptoRuntime::checkcast_arraycopy_Type(), 5107 copyfunc_addr, "checkcast_arraycopy", adr_type, 5108 // five arguments, of which two are 5109 // intptr_t (jlong in LP64) 5110 src_start, dest_start, 5111 copy_length XTOP, 5112 check_offset XTOP, 5113 check_value); 5114 5115 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5116 } 5117 5118 5119 // Helper function; generates code for cases requiring runtime checks. 5120 Node* 5121 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5122 Node* src, Node* src_offset, 5123 Node* dest, Node* dest_offset, 5124 Node* copy_length) { 5125 if (stopped()) return NULL; 5126 5127 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5128 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5129 return NULL; 5130 } 5131 5132 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5133 OptoRuntime::generic_arraycopy_Type(), 5134 copyfunc_addr, "generic_arraycopy", adr_type, 5135 src, src_offset, dest, dest_offset, copy_length); 5136 5137 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5138 } 5139 5140 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5141 void 5142 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5143 BasicType basic_elem_type, 5144 bool disjoint_bases, 5145 Node* src, Node* src_offset, 5146 Node* dest, Node* dest_offset, 5147 Node* copy_length) { 5148 if (stopped()) return; // nothing to do 5149 5150 Node* src_start = src; 5151 Node* dest_start = dest; 5152 if (src_offset != NULL || dest_offset != NULL) { 5153 assert(src_offset != NULL && dest_offset != NULL, ""); 5154 src_start = array_element_address(src, src_offset, basic_elem_type); 5155 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5156 } 5157 5158 // Figure out which arraycopy runtime method to call. 5159 const char* copyfunc_name = "arraycopy"; 5160 address copyfunc_addr = 5161 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5162 disjoint_bases, copyfunc_name); 5163 5164 // Call it. Note that the count_ix value is not scaled to a byte-size. 5165 make_runtime_call(RC_LEAF|RC_NO_FP, 5166 OptoRuntime::fast_arraycopy_Type(), 5167 copyfunc_addr, copyfunc_name, adr_type, 5168 src_start, dest_start, copy_length XTOP); 5169 }