1 /* 2 * Copyright 1999-2010 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 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2); 137 bool inline_string_compareTo(); 138 bool inline_string_indexOf(); 139 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 140 bool inline_string_equals(); 141 Node* pop_math_arg(); 142 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 143 bool inline_math_native(vmIntrinsics::ID id); 144 bool inline_trig(vmIntrinsics::ID id); 145 bool inline_trans(vmIntrinsics::ID id); 146 bool inline_abs(vmIntrinsics::ID id); 147 bool inline_sqrt(vmIntrinsics::ID id); 148 bool inline_pow(vmIntrinsics::ID id); 149 bool inline_exp(vmIntrinsics::ID id); 150 bool inline_min_max(vmIntrinsics::ID id); 151 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 152 // This returns Type::AnyPtr, RawPtr, or OopPtr. 153 int classify_unsafe_addr(Node* &base, Node* &offset); 154 Node* make_unsafe_address(Node* base, Node* offset); 155 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 156 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 157 bool inline_unsafe_allocate(); 158 bool inline_unsafe_copyMemory(); 159 bool inline_native_currentThread(); 160 bool inline_native_time_funcs(bool isNano); 161 bool inline_native_isInterrupted(); 162 bool inline_native_Class_query(vmIntrinsics::ID id); 163 bool inline_native_subtype_check(); 164 165 bool inline_native_newArray(); 166 bool inline_native_getLength(); 167 bool inline_array_copyOf(bool is_copyOfRange); 168 bool inline_array_equals(); 169 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 170 bool inline_native_clone(bool is_virtual); 171 bool inline_native_Reflection_getCallerClass(); 172 bool inline_native_AtomicLong_get(); 173 bool inline_native_AtomicLong_attemptUpdate(); 174 bool is_method_invoke_or_aux_frame(JVMState* jvms); 175 // Helper function for inlining native object hash method 176 bool inline_native_hashcode(bool is_virtual, bool is_static); 177 bool inline_native_getClass(); 178 179 // Helper functions for inlining arraycopy 180 bool inline_arraycopy(); 181 void generate_arraycopy(const TypePtr* adr_type, 182 BasicType basic_elem_type, 183 Node* src, Node* src_offset, 184 Node* dest, Node* dest_offset, 185 Node* copy_length, 186 bool disjoint_bases = false, 187 bool length_never_negative = false, 188 RegionNode* slow_region = NULL); 189 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 190 RegionNode* slow_region); 191 void generate_clear_array(const TypePtr* adr_type, 192 Node* dest, 193 BasicType basic_elem_type, 194 Node* slice_off, 195 Node* slice_len, 196 Node* slice_end); 197 bool generate_block_arraycopy(const TypePtr* adr_type, 198 BasicType basic_elem_type, 199 AllocateNode* alloc, 200 Node* src, Node* src_offset, 201 Node* dest, Node* dest_offset, 202 Node* dest_size); 203 void generate_slow_arraycopy(const TypePtr* adr_type, 204 Node* src, Node* src_offset, 205 Node* dest, Node* dest_offset, 206 Node* copy_length); 207 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 208 Node* dest_elem_klass, 209 Node* src, Node* src_offset, 210 Node* dest, Node* dest_offset, 211 Node* copy_length); 212 Node* generate_generic_arraycopy(const TypePtr* adr_type, 213 Node* src, Node* src_offset, 214 Node* dest, Node* dest_offset, 215 Node* copy_length); 216 void generate_unchecked_arraycopy(const TypePtr* adr_type, 217 BasicType basic_elem_type, 218 bool disjoint_bases, 219 Node* src, Node* src_offset, 220 Node* dest, Node* dest_offset, 221 Node* copy_length); 222 bool inline_unsafe_CAS(BasicType type); 223 bool inline_unsafe_ordered_store(BasicType type); 224 bool inline_fp_conversions(vmIntrinsics::ID id); 225 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id); 226 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id); 227 bool inline_bitCount(vmIntrinsics::ID id); 228 bool inline_reverseBytes(vmIntrinsics::ID id); 229 }; 230 231 232 //---------------------------make_vm_intrinsic---------------------------- 233 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 234 vmIntrinsics::ID id = m->intrinsic_id(); 235 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 236 237 if (DisableIntrinsic[0] != '\0' 238 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 239 // disabled by a user request on the command line: 240 // example: -XX:DisableIntrinsic=_hashCode,_getClass 241 return NULL; 242 } 243 244 if (!m->is_loaded()) { 245 // do not attempt to inline unloaded methods 246 return NULL; 247 } 248 249 // Only a few intrinsics implement a virtual dispatch. 250 // They are expensive calls which are also frequently overridden. 251 if (is_virtual) { 252 switch (id) { 253 case vmIntrinsics::_hashCode: 254 case vmIntrinsics::_clone: 255 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 256 break; 257 default: 258 return NULL; 259 } 260 } 261 262 // -XX:-InlineNatives disables nearly all intrinsics: 263 if (!InlineNatives) { 264 switch (id) { 265 case vmIntrinsics::_indexOf: 266 case vmIntrinsics::_compareTo: 267 case vmIntrinsics::_equals: 268 case vmIntrinsics::_equalsC: 269 break; // InlineNatives does not control String.compareTo 270 default: 271 return NULL; 272 } 273 } 274 275 switch (id) { 276 case vmIntrinsics::_compareTo: 277 if (!SpecialStringCompareTo) return NULL; 278 break; 279 case vmIntrinsics::_indexOf: 280 if (!SpecialStringIndexOf) return NULL; 281 break; 282 case vmIntrinsics::_equals: 283 if (!SpecialStringEquals) return NULL; 284 break; 285 case vmIntrinsics::_equalsC: 286 if (!SpecialArraysEquals) return NULL; 287 break; 288 case vmIntrinsics::_arraycopy: 289 if (!InlineArrayCopy) return NULL; 290 break; 291 case vmIntrinsics::_copyMemory: 292 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 293 if (!InlineArrayCopy) return NULL; 294 break; 295 case vmIntrinsics::_hashCode: 296 if (!InlineObjectHash) return NULL; 297 break; 298 case vmIntrinsics::_clone: 299 case vmIntrinsics::_copyOf: 300 case vmIntrinsics::_copyOfRange: 301 if (!InlineObjectCopy) return NULL; 302 // These also use the arraycopy intrinsic mechanism: 303 if (!InlineArrayCopy) return NULL; 304 break; 305 case vmIntrinsics::_checkIndex: 306 // We do not intrinsify this. The optimizer does fine with it. 307 return NULL; 308 309 case vmIntrinsics::_get_AtomicLong: 310 case vmIntrinsics::_attemptUpdate: 311 if (!InlineAtomicLong) return NULL; 312 break; 313 314 case vmIntrinsics::_getCallerClass: 315 if (!UseNewReflection) return NULL; 316 if (!InlineReflectionGetCallerClass) return NULL; 317 if (!JDK_Version::is_gte_jdk14x_version()) return NULL; 318 break; 319 320 case vmIntrinsics::_bitCount_i: 321 case vmIntrinsics::_bitCount_l: 322 if (!UsePopCountInstruction) return NULL; 323 break; 324 325 default: 326 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 327 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 328 break; 329 } 330 331 // -XX:-InlineClassNatives disables natives from the Class class. 332 // The flag applies to all reflective calls, notably Array.newArray 333 // (visible to Java programmers as Array.newInstance). 334 if (m->holder()->name() == ciSymbol::java_lang_Class() || 335 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 336 if (!InlineClassNatives) return NULL; 337 } 338 339 // -XX:-InlineThreadNatives disables natives from the Thread class. 340 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 341 if (!InlineThreadNatives) return NULL; 342 } 343 344 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 345 if (m->holder()->name() == ciSymbol::java_lang_Math() || 346 m->holder()->name() == ciSymbol::java_lang_Float() || 347 m->holder()->name() == ciSymbol::java_lang_Double()) { 348 if (!InlineMathNatives) return NULL; 349 } 350 351 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 352 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 353 if (!InlineUnsafeOps) return NULL; 354 } 355 356 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); 357 } 358 359 //----------------------register_library_intrinsics----------------------- 360 // Initialize this file's data structures, for each Compile instance. 361 void Compile::register_library_intrinsics() { 362 // Nothing to do here. 363 } 364 365 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 366 LibraryCallKit kit(jvms, this); 367 Compile* C = kit.C; 368 int nodes = C->unique(); 369 #ifndef PRODUCT 370 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 371 char buf[1000]; 372 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 373 tty->print_cr("Intrinsic %s", str); 374 } 375 #endif 376 if (kit.try_to_inline()) { 377 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 378 tty->print("Inlining intrinsic %s%s at bci:%d in", 379 vmIntrinsics::name_at(intrinsic_id()), 380 (is_virtual() ? " (virtual)" : ""), kit.bci()); 381 kit.caller()->print_short_name(tty); 382 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); 383 } 384 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 385 if (C->log()) { 386 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 387 vmIntrinsics::name_at(intrinsic_id()), 388 (is_virtual() ? " virtual='1'" : ""), 389 C->unique() - nodes); 390 } 391 return kit.transfer_exceptions_into_jvms(); 392 } 393 394 if (PrintIntrinsics) { 395 tty->print("Did not inline intrinsic %s%s at bci:%d in", 396 vmIntrinsics::name_at(intrinsic_id()), 397 (is_virtual() ? " (virtual)" : ""), kit.bci()); 398 kit.caller()->print_short_name(tty); 399 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); 400 } 401 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 402 return NULL; 403 } 404 405 bool LibraryCallKit::try_to_inline() { 406 // Handle symbolic names for otherwise undistinguished boolean switches: 407 const bool is_store = true; 408 const bool is_native_ptr = true; 409 const bool is_static = true; 410 411 switch (intrinsic_id()) { 412 case vmIntrinsics::_hashCode: 413 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 414 case vmIntrinsics::_identityHashCode: 415 return inline_native_hashcode(/*!virtual*/ false, is_static); 416 case vmIntrinsics::_getClass: 417 return inline_native_getClass(); 418 419 case vmIntrinsics::_dsin: 420 case vmIntrinsics::_dcos: 421 case vmIntrinsics::_dtan: 422 case vmIntrinsics::_dabs: 423 case vmIntrinsics::_datan2: 424 case vmIntrinsics::_dsqrt: 425 case vmIntrinsics::_dexp: 426 case vmIntrinsics::_dlog: 427 case vmIntrinsics::_dlog10: 428 case vmIntrinsics::_dpow: 429 return inline_math_native(intrinsic_id()); 430 431 case vmIntrinsics::_min: 432 case vmIntrinsics::_max: 433 return inline_min_max(intrinsic_id()); 434 435 case vmIntrinsics::_arraycopy: 436 return inline_arraycopy(); 437 438 case vmIntrinsics::_compareTo: 439 return inline_string_compareTo(); 440 case vmIntrinsics::_indexOf: 441 return inline_string_indexOf(); 442 case vmIntrinsics::_equals: 443 return inline_string_equals(); 444 445 case vmIntrinsics::_getObject: 446 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); 447 case vmIntrinsics::_getBoolean: 448 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); 449 case vmIntrinsics::_getByte: 450 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); 451 case vmIntrinsics::_getShort: 452 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); 453 case vmIntrinsics::_getChar: 454 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); 455 case vmIntrinsics::_getInt: 456 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); 457 case vmIntrinsics::_getLong: 458 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); 459 case vmIntrinsics::_getFloat: 460 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); 461 case vmIntrinsics::_getDouble: 462 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); 463 464 case vmIntrinsics::_putObject: 465 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); 466 case vmIntrinsics::_putBoolean: 467 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); 468 case vmIntrinsics::_putByte: 469 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); 470 case vmIntrinsics::_putShort: 471 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); 472 case vmIntrinsics::_putChar: 473 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); 474 case vmIntrinsics::_putInt: 475 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); 476 case vmIntrinsics::_putLong: 477 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); 478 case vmIntrinsics::_putFloat: 479 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); 480 case vmIntrinsics::_putDouble: 481 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); 482 483 case vmIntrinsics::_getByte_raw: 484 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); 485 case vmIntrinsics::_getShort_raw: 486 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); 487 case vmIntrinsics::_getChar_raw: 488 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); 489 case vmIntrinsics::_getInt_raw: 490 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); 491 case vmIntrinsics::_getLong_raw: 492 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); 493 case vmIntrinsics::_getFloat_raw: 494 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); 495 case vmIntrinsics::_getDouble_raw: 496 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); 497 case vmIntrinsics::_getAddress_raw: 498 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); 499 500 case vmIntrinsics::_putByte_raw: 501 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); 502 case vmIntrinsics::_putShort_raw: 503 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); 504 case vmIntrinsics::_putChar_raw: 505 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); 506 case vmIntrinsics::_putInt_raw: 507 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); 508 case vmIntrinsics::_putLong_raw: 509 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); 510 case vmIntrinsics::_putFloat_raw: 511 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); 512 case vmIntrinsics::_putDouble_raw: 513 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); 514 case vmIntrinsics::_putAddress_raw: 515 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); 516 517 case vmIntrinsics::_getObjectVolatile: 518 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); 519 case vmIntrinsics::_getBooleanVolatile: 520 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); 521 case vmIntrinsics::_getByteVolatile: 522 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); 523 case vmIntrinsics::_getShortVolatile: 524 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); 525 case vmIntrinsics::_getCharVolatile: 526 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); 527 case vmIntrinsics::_getIntVolatile: 528 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); 529 case vmIntrinsics::_getLongVolatile: 530 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); 531 case vmIntrinsics::_getFloatVolatile: 532 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); 533 case vmIntrinsics::_getDoubleVolatile: 534 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); 535 536 case vmIntrinsics::_putObjectVolatile: 537 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); 538 case vmIntrinsics::_putBooleanVolatile: 539 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); 540 case vmIntrinsics::_putByteVolatile: 541 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); 542 case vmIntrinsics::_putShortVolatile: 543 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); 544 case vmIntrinsics::_putCharVolatile: 545 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); 546 case vmIntrinsics::_putIntVolatile: 547 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); 548 case vmIntrinsics::_putLongVolatile: 549 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); 550 case vmIntrinsics::_putFloatVolatile: 551 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); 552 case vmIntrinsics::_putDoubleVolatile: 553 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); 554 555 case vmIntrinsics::_prefetchRead: 556 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 557 case vmIntrinsics::_prefetchWrite: 558 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 559 case vmIntrinsics::_prefetchReadStatic: 560 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 561 case vmIntrinsics::_prefetchWriteStatic: 562 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 563 564 case vmIntrinsics::_compareAndSwapObject: 565 return inline_unsafe_CAS(T_OBJECT); 566 case vmIntrinsics::_compareAndSwapInt: 567 return inline_unsafe_CAS(T_INT); 568 case vmIntrinsics::_compareAndSwapLong: 569 return inline_unsafe_CAS(T_LONG); 570 571 case vmIntrinsics::_putOrderedObject: 572 return inline_unsafe_ordered_store(T_OBJECT); 573 case vmIntrinsics::_putOrderedInt: 574 return inline_unsafe_ordered_store(T_INT); 575 case vmIntrinsics::_putOrderedLong: 576 return inline_unsafe_ordered_store(T_LONG); 577 578 case vmIntrinsics::_currentThread: 579 return inline_native_currentThread(); 580 case vmIntrinsics::_isInterrupted: 581 return inline_native_isInterrupted(); 582 583 case vmIntrinsics::_currentTimeMillis: 584 return inline_native_time_funcs(false); 585 case vmIntrinsics::_nanoTime: 586 return inline_native_time_funcs(true); 587 case vmIntrinsics::_allocateInstance: 588 return inline_unsafe_allocate(); 589 case vmIntrinsics::_copyMemory: 590 return inline_unsafe_copyMemory(); 591 case vmIntrinsics::_newArray: 592 return inline_native_newArray(); 593 case vmIntrinsics::_getLength: 594 return inline_native_getLength(); 595 case vmIntrinsics::_copyOf: 596 return inline_array_copyOf(false); 597 case vmIntrinsics::_copyOfRange: 598 return inline_array_copyOf(true); 599 case vmIntrinsics::_equalsC: 600 return inline_array_equals(); 601 case vmIntrinsics::_clone: 602 return inline_native_clone(intrinsic()->is_virtual()); 603 604 case vmIntrinsics::_isAssignableFrom: 605 return inline_native_subtype_check(); 606 607 case vmIntrinsics::_isInstance: 608 case vmIntrinsics::_getModifiers: 609 case vmIntrinsics::_isInterface: 610 case vmIntrinsics::_isArray: 611 case vmIntrinsics::_isPrimitive: 612 case vmIntrinsics::_getSuperclass: 613 case vmIntrinsics::_getComponentType: 614 case vmIntrinsics::_getClassAccessFlags: 615 return inline_native_Class_query(intrinsic_id()); 616 617 case vmIntrinsics::_floatToRawIntBits: 618 case vmIntrinsics::_floatToIntBits: 619 case vmIntrinsics::_intBitsToFloat: 620 case vmIntrinsics::_doubleToRawLongBits: 621 case vmIntrinsics::_doubleToLongBits: 622 case vmIntrinsics::_longBitsToDouble: 623 return inline_fp_conversions(intrinsic_id()); 624 625 case vmIntrinsics::_numberOfLeadingZeros_i: 626 case vmIntrinsics::_numberOfLeadingZeros_l: 627 return inline_numberOfLeadingZeros(intrinsic_id()); 628 629 case vmIntrinsics::_numberOfTrailingZeros_i: 630 case vmIntrinsics::_numberOfTrailingZeros_l: 631 return inline_numberOfTrailingZeros(intrinsic_id()); 632 633 case vmIntrinsics::_bitCount_i: 634 case vmIntrinsics::_bitCount_l: 635 return inline_bitCount(intrinsic_id()); 636 637 case vmIntrinsics::_reverseBytes_i: 638 case vmIntrinsics::_reverseBytes_l: 639 case vmIntrinsics::_reverseBytes_s: 640 case vmIntrinsics::_reverseBytes_c: 641 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); 642 643 case vmIntrinsics::_get_AtomicLong: 644 return inline_native_AtomicLong_get(); 645 case vmIntrinsics::_attemptUpdate: 646 return inline_native_AtomicLong_attemptUpdate(); 647 648 case vmIntrinsics::_getCallerClass: 649 return inline_native_Reflection_getCallerClass(); 650 651 default: 652 // If you get here, it may be that someone has added a new intrinsic 653 // to the list in vmSymbols.hpp without implementing it here. 654 #ifndef PRODUCT 655 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 656 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 657 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 658 } 659 #endif 660 return false; 661 } 662 } 663 664 //------------------------------push_result------------------------------ 665 // Helper function for finishing intrinsics. 666 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { 667 record_for_igvn(region); 668 set_control(_gvn.transform(region)); 669 BasicType value_type = value->type()->basic_type(); 670 push_node(value_type, _gvn.transform(value)); 671 } 672 673 //------------------------------generate_guard--------------------------- 674 // Helper function for generating guarded fast-slow graph structures. 675 // The given 'test', if true, guards a slow path. If the test fails 676 // then a fast path can be taken. (We generally hope it fails.) 677 // In all cases, GraphKit::control() is updated to the fast path. 678 // The returned value represents the control for the slow path. 679 // The return value is never 'top'; it is either a valid control 680 // or NULL if it is obvious that the slow path can never be taken. 681 // Also, if region and the slow control are not NULL, the slow edge 682 // is appended to the region. 683 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 684 if (stopped()) { 685 // Already short circuited. 686 return NULL; 687 } 688 689 // Build an if node and its projections. 690 // If test is true we take the slow path, which we assume is uncommon. 691 if (_gvn.type(test) == TypeInt::ZERO) { 692 // The slow branch is never taken. No need to build this guard. 693 return NULL; 694 } 695 696 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 697 698 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); 699 if (if_slow == top()) { 700 // The slow branch is never taken. No need to build this guard. 701 return NULL; 702 } 703 704 if (region != NULL) 705 region->add_req(if_slow); 706 707 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); 708 set_control(if_fast); 709 710 return if_slow; 711 } 712 713 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 714 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 715 } 716 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 717 return generate_guard(test, region, PROB_FAIR); 718 } 719 720 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 721 Node* *pos_index) { 722 if (stopped()) 723 return NULL; // already stopped 724 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 725 return NULL; // index is already adequately typed 726 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 727 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 728 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 729 if (is_neg != NULL && pos_index != NULL) { 730 // Emulate effect of Parse::adjust_map_after_if. 731 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); 732 ccast->set_req(0, control()); 733 (*pos_index) = _gvn.transform(ccast); 734 } 735 return is_neg; 736 } 737 738 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 739 Node* *pos_index) { 740 if (stopped()) 741 return NULL; // already stopped 742 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 743 return NULL; // index is already adequately typed 744 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 745 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 746 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) ); 747 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 748 if (is_notp != NULL && pos_index != NULL) { 749 // Emulate effect of Parse::adjust_map_after_if. 750 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1); 751 ccast->set_req(0, control()); 752 (*pos_index) = _gvn.transform(ccast); 753 } 754 return is_notp; 755 } 756 757 // Make sure that 'position' is a valid limit index, in [0..length]. 758 // There are two equivalent plans for checking this: 759 // A. (offset + copyLength) unsigned<= arrayLength 760 // B. offset <= (arrayLength - copyLength) 761 // We require that all of the values above, except for the sum and 762 // difference, are already known to be non-negative. 763 // Plan A is robust in the face of overflow, if offset and copyLength 764 // are both hugely positive. 765 // 766 // Plan B is less direct and intuitive, but it does not overflow at 767 // all, since the difference of two non-negatives is always 768 // representable. Whenever Java methods must perform the equivalent 769 // check they generally use Plan B instead of Plan A. 770 // For the moment we use Plan A. 771 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 772 Node* subseq_length, 773 Node* array_length, 774 RegionNode* region) { 775 if (stopped()) 776 return NULL; // already stopped 777 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 778 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length)) 779 return NULL; // common case of whole-array copy 780 Node* last = subseq_length; 781 if (!zero_offset) // last += offset 782 last = _gvn.transform( new (C, 3) AddINode(last, offset)); 783 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); 784 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 785 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 786 return is_over; 787 } 788 789 790 //--------------------------generate_current_thread-------------------- 791 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 792 ciKlass* thread_klass = env()->Thread_klass(); 793 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 794 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); 795 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 796 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 797 tls_output = thread; 798 return threadObj; 799 } 800 801 802 //------------------------------make_string_method_node------------------------ 803 // Helper method for String intrinsic finctions. 804 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) { 805 const int value_offset = java_lang_String::value_offset_in_bytes(); 806 const int count_offset = java_lang_String::count_offset_in_bytes(); 807 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 808 809 Node* no_ctrl = NULL; 810 811 ciInstanceKlass* klass = env()->String_klass(); 812 const TypeInstPtr* string_type = 813 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 814 815 const TypeAryPtr* value_type = 816 TypeAryPtr::make(TypePtr::NotNull, 817 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 818 ciTypeArrayKlass::make(T_CHAR), true, 0); 819 820 // Get start addr of string and substring 821 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset); 822 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset)); 823 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset); 824 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 825 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 826 827 // Pin loads from String::equals() argument since it could be NULL. 828 Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl; 829 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset); 830 Node* str2_value = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset)); 831 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset); 832 Node* str2_offset = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 833 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 834 835 Node* result = NULL; 836 switch (opcode) { 837 case Op_StrIndexOf: 838 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 839 str1_start, cnt1, str2_start, cnt2); 840 break; 841 case Op_StrComp: 842 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS), 843 str1_start, cnt1, str2_start, cnt2); 844 break; 845 case Op_StrEquals: 846 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 847 str1_start, str2_start, cnt1); 848 break; 849 default: 850 ShouldNotReachHere(); 851 return NULL; 852 } 853 854 // All these intrinsics have checks. 855 C->set_has_split_ifs(true); // Has chance for split-if optimization 856 857 return _gvn.transform(result); 858 } 859 860 //------------------------------inline_string_compareTo------------------------ 861 bool LibraryCallKit::inline_string_compareTo() { 862 863 if (!Matcher::has_match_rule(Op_StrComp)) return false; 864 865 const int value_offset = java_lang_String::value_offset_in_bytes(); 866 const int count_offset = java_lang_String::count_offset_in_bytes(); 867 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 868 869 _sp += 2; 870 Node *argument = pop(); // pop non-receiver first: it was pushed second 871 Node *receiver = pop(); 872 873 // Null check on self without removing any arguments. The argument 874 // null check technically happens in the wrong place, which can lead to 875 // invalid stack traces when string compare is inlined into a method 876 // which handles NullPointerExceptions. 877 _sp += 2; 878 receiver = do_null_check(receiver, T_OBJECT); 879 argument = do_null_check(argument, T_OBJECT); 880 _sp -= 2; 881 if (stopped()) { 882 return true; 883 } 884 885 ciInstanceKlass* klass = env()->String_klass(); 886 const TypeInstPtr* string_type = 887 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 888 Node* no_ctrl = NULL; 889 890 // Get counts for string and argument 891 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset); 892 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 893 894 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset); 895 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 896 897 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt); 898 push(compare); 899 return true; 900 } 901 902 //------------------------------inline_string_equals------------------------ 903 bool LibraryCallKit::inline_string_equals() { 904 905 if (!Matcher::has_match_rule(Op_StrEquals)) return false; 906 907 const int value_offset = java_lang_String::value_offset_in_bytes(); 908 const int count_offset = java_lang_String::count_offset_in_bytes(); 909 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 910 911 _sp += 2; 912 Node* argument = pop(); // pop non-receiver first: it was pushed second 913 Node* receiver = pop(); 914 915 // Null check on self without removing any arguments. The argument 916 // null check technically happens in the wrong place, which can lead to 917 // invalid stack traces when string compare is inlined into a method 918 // which handles NullPointerExceptions. 919 _sp += 2; 920 receiver = do_null_check(receiver, T_OBJECT); 921 //should not do null check for argument for String.equals(), because spec 922 //allows to specify NULL as argument. 923 _sp -= 2; 924 925 if (stopped()) { 926 return true; 927 } 928 929 // paths (plus control) merge 930 RegionNode* region = new (C, 5) RegionNode(5); 931 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL); 932 933 // does source == target string? 934 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument)); 935 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq)); 936 937 Node* if_eq = generate_slow_guard(bol, NULL); 938 if (if_eq != NULL) { 939 // receiver == argument 940 phi->init_req(2, intcon(1)); 941 region->init_req(2, if_eq); 942 } 943 944 // get String klass for instanceOf 945 ciInstanceKlass* klass = env()->String_klass(); 946 947 if (!stopped()) { 948 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 949 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1))); 950 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne)); 951 952 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 953 //instanceOf == true, fallthrough 954 955 if (inst_false != NULL) { 956 phi->init_req(3, intcon(0)); 957 region->init_req(3, inst_false); 958 } 959 } 960 961 const TypeInstPtr* string_type = 962 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 963 964 Node* no_ctrl = NULL; 965 Node* receiver_cnt; 966 Node* argument_cnt; 967 968 if (!stopped()) { 969 // Get counts for string and argument 970 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset); 971 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 972 973 // Pin load from argument string since it could be NULL. 974 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset); 975 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 976 977 // Check for receiver count != argument count 978 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) ); 979 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) ); 980 Node* if_ne = generate_slow_guard(bol, NULL); 981 if (if_ne != NULL) { 982 phi->init_req(4, intcon(0)); 983 region->init_req(4, if_ne); 984 } 985 } 986 987 // Check for count == 0 is done by mach node StrEquals. 988 989 if (!stopped()) { 990 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt); 991 phi->init_req(1, equals); 992 region->init_req(1, control()); 993 } 994 995 // post merge 996 set_control(_gvn.transform(region)); 997 record_for_igvn(region); 998 999 push(_gvn.transform(phi)); 1000 1001 return true; 1002 } 1003 1004 //------------------------------inline_array_equals---------------------------- 1005 bool LibraryCallKit::inline_array_equals() { 1006 1007 if (!Matcher::has_match_rule(Op_AryEq)) return false; 1008 1009 _sp += 2; 1010 Node *argument2 = pop(); 1011 Node *argument1 = pop(); 1012 1013 Node* equals = 1014 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS), 1015 argument1, argument2) ); 1016 push(equals); 1017 return true; 1018 } 1019 1020 // Java version of String.indexOf(constant string) 1021 // class StringDecl { 1022 // StringDecl(char[] ca) { 1023 // offset = 0; 1024 // count = ca.length; 1025 // value = ca; 1026 // } 1027 // int offset; 1028 // int count; 1029 // char[] value; 1030 // } 1031 // 1032 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1033 // int targetOffset, int cache_i, int md2) { 1034 // int cache = cache_i; 1035 // int sourceOffset = string_object.offset; 1036 // int sourceCount = string_object.count; 1037 // int targetCount = target_object.length; 1038 // 1039 // int targetCountLess1 = targetCount - 1; 1040 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1041 // 1042 // char[] source = string_object.value; 1043 // char[] target = target_object; 1044 // int lastChar = target[targetCountLess1]; 1045 // 1046 // outer_loop: 1047 // for (int i = sourceOffset; i < sourceEnd; ) { 1048 // int src = source[i + targetCountLess1]; 1049 // if (src == lastChar) { 1050 // // With random strings and a 4-character alphabet, 1051 // // reverse matching at this point sets up 0.8% fewer 1052 // // frames, but (paradoxically) makes 0.3% more probes. 1053 // // Since those probes are nearer the lastChar probe, 1054 // // there is may be a net D$ win with reverse matching. 1055 // // But, reversing loop inhibits unroll of inner loop 1056 // // for unknown reason. So, does running outer loop from 1057 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1058 // for (int j = 0; j < targetCountLess1; j++) { 1059 // if (target[targetOffset + j] != source[i+j]) { 1060 // if ((cache & (1 << source[i+j])) == 0) { 1061 // if (md2 < j+1) { 1062 // i += j+1; 1063 // continue outer_loop; 1064 // } 1065 // } 1066 // i += md2; 1067 // continue outer_loop; 1068 // } 1069 // } 1070 // return i - sourceOffset; 1071 // } 1072 // if ((cache & (1 << src)) == 0) { 1073 // i += targetCountLess1; 1074 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1075 // i++; 1076 // } 1077 // return -1; 1078 // } 1079 1080 //------------------------------string_indexOf------------------------ 1081 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1082 jint cache_i, jint md2_i) { 1083 1084 Node* no_ctrl = NULL; 1085 float likely = PROB_LIKELY(0.9); 1086 float unlikely = PROB_UNLIKELY(0.9); 1087 1088 const int value_offset = java_lang_String::value_offset_in_bytes(); 1089 const int count_offset = java_lang_String::count_offset_in_bytes(); 1090 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1091 1092 ciInstanceKlass* klass = env()->String_klass(); 1093 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 1094 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0); 1095 1096 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset); 1097 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 1098 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset); 1099 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1100 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset); 1101 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset)); 1102 1103 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) ); 1104 jint target_length = target_array->length(); 1105 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1106 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1107 1108 IdealKit kit(gvn(), control(), merged_memory(), false, true); 1109 #define __ kit. 1110 Node* zero = __ ConI(0); 1111 Node* one = __ ConI(1); 1112 Node* cache = __ ConI(cache_i); 1113 Node* md2 = __ ConI(md2_i); 1114 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1115 Node* targetCount = __ ConI(target_length); 1116 Node* targetCountLess1 = __ ConI(target_length - 1); 1117 Node* targetOffset = __ ConI(targetOffset_i); 1118 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1119 1120 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1121 Node* outer_loop = __ make_label(2 /* goto */); 1122 Node* return_ = __ make_label(1); 1123 1124 __ set(rtn,__ ConI(-1)); 1125 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); { 1126 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1127 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1128 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1129 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1130 __ loop(j, zero, BoolTest::lt, targetCountLess1); { 1131 Node* tpj = __ AddI(targetOffset, __ value(j)); 1132 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1133 Node* ipj = __ AddI(__ value(i), __ value(j)); 1134 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1135 __ if_then(targ, BoolTest::ne, src2); { 1136 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1137 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1138 __ increment(i, __ AddI(__ value(j), one)); 1139 __ goto_(outer_loop); 1140 } __ end_if(); __ dead(j); 1141 }__ end_if(); __ dead(j); 1142 __ increment(i, md2); 1143 __ goto_(outer_loop); 1144 }__ end_if(); 1145 __ increment(j, one); 1146 }__ end_loop(); __ dead(j); 1147 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1148 __ goto_(return_); 1149 }__ end_if(); 1150 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1151 __ increment(i, targetCountLess1); 1152 }__ end_if(); 1153 __ increment(i, one); 1154 __ bind(outer_loop); 1155 }__ end_loop(); __ dead(i); 1156 __ bind(return_); 1157 1158 // Final sync IdealKit and GraphKit. 1159 sync_kit(kit); 1160 Node* result = __ value(rtn); 1161 #undef __ 1162 C->set_has_loops(true); 1163 return result; 1164 } 1165 1166 //------------------------------inline_string_indexOf------------------------ 1167 bool LibraryCallKit::inline_string_indexOf() { 1168 1169 const int value_offset = java_lang_String::value_offset_in_bytes(); 1170 const int count_offset = java_lang_String::count_offset_in_bytes(); 1171 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1172 1173 _sp += 2; 1174 Node *argument = pop(); // pop non-receiver first: it was pushed second 1175 Node *receiver = pop(); 1176 1177 Node* result; 1178 if (Matcher::has_match_rule(Op_StrIndexOf) && 1179 UseSSE42Intrinsics) { 1180 // Generate SSE4.2 version of indexOf 1181 // We currently only have match rules that use SSE4.2 1182 1183 // Null check on self without removing any arguments. The argument 1184 // null check technically happens in the wrong place, which can lead to 1185 // invalid stack traces when string compare is inlined into a method 1186 // which handles NullPointerExceptions. 1187 _sp += 2; 1188 receiver = do_null_check(receiver, T_OBJECT); 1189 argument = do_null_check(argument, T_OBJECT); 1190 _sp -= 2; 1191 1192 if (stopped()) { 1193 return true; 1194 } 1195 1196 // Make the merge point 1197 RegionNode* result_rgn = new (C, 3) RegionNode(3); 1198 Node* result_phi = new (C, 3) PhiNode(result_rgn, TypeInt::INT); 1199 Node* no_ctrl = NULL; 1200 1201 ciInstanceKlass* klass = env()->String_klass(); 1202 const TypeInstPtr* string_type = 1203 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 1204 1205 // Get counts for string and substr 1206 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset); 1207 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1208 1209 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset); 1210 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1211 1212 // Check for substr count > string count 1213 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) ); 1214 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) ); 1215 Node* if_gt = generate_slow_guard(bol, NULL); 1216 if (if_gt != NULL) { 1217 result_phi->init_req(2, intcon(-1)); 1218 result_rgn->init_req(2, if_gt); 1219 } 1220 1221 if (!stopped()) { 1222 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt); 1223 result_phi->init_req(1, result); 1224 result_rgn->init_req(1, control()); 1225 } 1226 set_control(_gvn.transform(result_rgn)); 1227 record_for_igvn(result_rgn); 1228 result = _gvn.transform(result_phi); 1229 1230 } else { //Use LibraryCallKit::string_indexOf 1231 // don't intrinsify is argument isn't a constant string. 1232 if (!argument->is_Con()) { 1233 return false; 1234 } 1235 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); 1236 if (str_type == NULL) { 1237 return false; 1238 } 1239 ciInstanceKlass* klass = env()->String_klass(); 1240 ciObject* str_const = str_type->const_oop(); 1241 if (str_const == NULL || str_const->klass() != klass) { 1242 return false; 1243 } 1244 ciInstance* str = str_const->as_instance(); 1245 assert(str != NULL, "must be instance"); 1246 1247 ciObject* v = str->field_value_by_offset(value_offset).as_object(); 1248 int o = str->field_value_by_offset(offset_offset).as_int(); 1249 int c = str->field_value_by_offset(count_offset).as_int(); 1250 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1251 1252 // constant strings have no offset and count == length which 1253 // simplifies the resulting code somewhat so lets optimize for that. 1254 if (o != 0 || c != pat->length()) { 1255 return false; 1256 } 1257 1258 // Null check on self without removing any arguments. The argument 1259 // null check technically happens in the wrong place, which can lead to 1260 // invalid stack traces when string compare is inlined into a method 1261 // which handles NullPointerExceptions. 1262 _sp += 2; 1263 receiver = do_null_check(receiver, T_OBJECT); 1264 // No null check on the argument is needed since it's a constant String oop. 1265 _sp -= 2; 1266 if (stopped()) { 1267 return true; 1268 } 1269 1270 // The null string as a pattern always returns 0 (match at beginning of string) 1271 if (c == 0) { 1272 push(intcon(0)); 1273 return true; 1274 } 1275 1276 // Generate default indexOf 1277 jchar lastChar = pat->char_at(o + (c - 1)); 1278 int cache = 0; 1279 int i; 1280 for (i = 0; i < c - 1; i++) { 1281 assert(i < pat->length(), "out of range"); 1282 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1283 } 1284 1285 int md2 = c; 1286 for (i = 0; i < c - 1; i++) { 1287 assert(i < pat->length(), "out of range"); 1288 if (pat->char_at(o + i) == lastChar) { 1289 md2 = (c - 1) - i; 1290 } 1291 } 1292 1293 result = string_indexOf(receiver, pat, o, cache, md2); 1294 } 1295 1296 push(result); 1297 return true; 1298 } 1299 1300 //--------------------------pop_math_arg-------------------------------- 1301 // Pop a double argument to a math function from the stack 1302 // rounding it if necessary. 1303 Node * LibraryCallKit::pop_math_arg() { 1304 Node *arg = pop_pair(); 1305 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) 1306 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); 1307 return arg; 1308 } 1309 1310 //------------------------------inline_trig---------------------------------- 1311 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1312 // argument reduction which will turn into a fast/slow diamond. 1313 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1314 _sp += arg_size(); // restore stack pointer 1315 Node* arg = pop_math_arg(); 1316 Node* trig = NULL; 1317 1318 switch (id) { 1319 case vmIntrinsics::_dsin: 1320 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); 1321 break; 1322 case vmIntrinsics::_dcos: 1323 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); 1324 break; 1325 case vmIntrinsics::_dtan: 1326 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); 1327 break; 1328 default: 1329 assert(false, "bad intrinsic was passed in"); 1330 return false; 1331 } 1332 1333 // Rounding required? Check for argument reduction! 1334 if( Matcher::strict_fp_requires_explicit_rounding ) { 1335 1336 static const double pi_4 = 0.7853981633974483; 1337 static const double neg_pi_4 = -0.7853981633974483; 1338 // pi/2 in 80-bit extended precision 1339 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1340 // -pi/2 in 80-bit extended precision 1341 // 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}; 1342 // Cutoff value for using this argument reduction technique 1343 //static const double pi_2_minus_epsilon = 1.564660403643354; 1344 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1345 1346 // Pseudocode for sin: 1347 // if (x <= Math.PI / 4.0) { 1348 // if (x >= -Math.PI / 4.0) return fsin(x); 1349 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1350 // } else { 1351 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1352 // } 1353 // return StrictMath.sin(x); 1354 1355 // Pseudocode for cos: 1356 // if (x <= Math.PI / 4.0) { 1357 // if (x >= -Math.PI / 4.0) return fcos(x); 1358 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1359 // } else { 1360 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1361 // } 1362 // return StrictMath.cos(x); 1363 1364 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1365 // requires a special machine instruction to load it. Instead we'll try 1366 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1367 // probably do the math inside the SIN encoding. 1368 1369 // Make the merge point 1370 RegionNode *r = new (C, 3) RegionNode(3); 1371 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); 1372 1373 // Flatten arg so we need only 1 test 1374 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); 1375 // Node for PI/4 constant 1376 Node *pi4 = makecon(TypeD::make(pi_4)); 1377 // Check PI/4 : abs(arg) 1378 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); 1379 // Check: If PI/4 < abs(arg) then go slow 1380 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); 1381 // Branch either way 1382 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1383 set_control(opt_iff(r,iff)); 1384 1385 // Set fast path result 1386 phi->init_req(2,trig); 1387 1388 // Slow path - non-blocking leaf call 1389 Node* call = NULL; 1390 switch (id) { 1391 case vmIntrinsics::_dsin: 1392 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1393 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1394 "Sin", NULL, arg, top()); 1395 break; 1396 case vmIntrinsics::_dcos: 1397 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1398 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1399 "Cos", NULL, arg, top()); 1400 break; 1401 case vmIntrinsics::_dtan: 1402 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1403 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1404 "Tan", NULL, arg, top()); 1405 break; 1406 } 1407 assert(control()->in(0) == call, ""); 1408 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); 1409 r->init_req(1,control()); 1410 phi->init_req(1,slow_result); 1411 1412 // Post-merge 1413 set_control(_gvn.transform(r)); 1414 record_for_igvn(r); 1415 trig = _gvn.transform(phi); 1416 1417 C->set_has_split_ifs(true); // Has chance for split-if optimization 1418 } 1419 // Push result back on JVM stack 1420 push_pair(trig); 1421 return true; 1422 } 1423 1424 //------------------------------inline_sqrt------------------------------------- 1425 // Inline square root instruction, if possible. 1426 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { 1427 assert(id == vmIntrinsics::_dsqrt, "Not square root"); 1428 _sp += arg_size(); // restore stack pointer 1429 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); 1430 return true; 1431 } 1432 1433 //------------------------------inline_abs------------------------------------- 1434 // Inline absolute value instruction, if possible. 1435 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { 1436 assert(id == vmIntrinsics::_dabs, "Not absolute value"); 1437 _sp += arg_size(); // restore stack pointer 1438 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); 1439 return true; 1440 } 1441 1442 //------------------------------inline_exp------------------------------------- 1443 // Inline exp instructions, if possible. The Intel hardware only misses 1444 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1445 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { 1446 assert(id == vmIntrinsics::_dexp, "Not exp"); 1447 1448 // If this inlining ever returned NaN in the past, we do not intrinsify it 1449 // every again. NaN results requires StrictMath.exp handling. 1450 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1451 1452 // Do not intrinsify on older platforms which lack cmove. 1453 if (ConditionalMoveLimit == 0) return false; 1454 1455 _sp += arg_size(); // restore stack pointer 1456 Node *x = pop_math_arg(); 1457 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); 1458 1459 //------------------- 1460 //result=(result.isNaN())? StrictMath::exp():result; 1461 // Check: If isNaN() by checking result!=result? then go to Strict Math 1462 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1463 // Build the boolean node 1464 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1465 1466 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1467 // End the current control-flow path 1468 push_pair(x); 1469 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp 1470 // to handle. Recompile without intrinsifying Math.exp 1471 uncommon_trap(Deoptimization::Reason_intrinsic, 1472 Deoptimization::Action_make_not_entrant); 1473 } 1474 1475 C->set_has_split_ifs(true); // Has chance for split-if optimization 1476 1477 push_pair(result); 1478 1479 return true; 1480 } 1481 1482 //------------------------------inline_pow------------------------------------- 1483 // Inline power instructions, if possible. 1484 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { 1485 assert(id == vmIntrinsics::_dpow, "Not pow"); 1486 1487 // If this inlining ever returned NaN in the past, we do not intrinsify it 1488 // every again. NaN results requires StrictMath.pow handling. 1489 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1490 1491 // Do not intrinsify on older platforms which lack cmove. 1492 if (ConditionalMoveLimit == 0) return false; 1493 1494 // Pseudocode for pow 1495 // if (x <= 0.0) { 1496 // if ((double)((int)y)==y) { // if y is int 1497 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y) 1498 // } else { 1499 // result = NaN; 1500 // } 1501 // } else { 1502 // result = DPow(x,y); 1503 // } 1504 // if (result != result)? { 1505 // uncommon_trap(); 1506 // } 1507 // return result; 1508 1509 _sp += arg_size(); // restore stack pointer 1510 Node* y = pop_math_arg(); 1511 Node* x = pop_math_arg(); 1512 1513 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); 1514 1515 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow 1516 // inside of something) then skip the fancy tests and just check for 1517 // NaN result. 1518 Node *result = NULL; 1519 if( jvms()->depth() >= 1 ) { 1520 result = fast_result; 1521 } else { 1522 1523 // Set the merge point for If node with condition of (x <= 0.0) 1524 // There are four possible paths to region node and phi node 1525 RegionNode *r = new (C, 4) RegionNode(4); 1526 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); 1527 1528 // Build the first if node: if (x <= 0.0) 1529 // Node for 0 constant 1530 Node *zeronode = makecon(TypeD::ZERO); 1531 // Check x:0 1532 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); 1533 // Check: If (x<=0) then go complex path 1534 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); 1535 // Branch either way 1536 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1537 Node *opt_test = _gvn.transform(if1); 1538 //assert( opt_test->is_If(), "Expect an IfNode"); 1539 IfNode *opt_if1 = (IfNode*)opt_test; 1540 // Fast path taken; set region slot 3 1541 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) ); 1542 r->init_req(3,fast_taken); // Capture fast-control 1543 1544 // Fast path not-taken, i.e. slow path 1545 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) ); 1546 1547 // Set fast path result 1548 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) ); 1549 phi->init_req(3, fast_result); 1550 1551 // Complex path 1552 // Build the second if node (if y is int) 1553 // Node for (int)y 1554 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y)); 1555 // Node for (double)((int) y) 1556 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty)); 1557 // Check (double)((int) y) : y 1558 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y)); 1559 // Check if (y isn't int) then go to slow path 1560 1561 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) ); 1562 // Branch either way 1563 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1564 Node *slow_path = opt_iff(r,if2); // Set region path 2 1565 1566 // Calculate DPow(abs(x), y)*(1 & (int)y) 1567 // Node for constant 1 1568 Node *conone = intcon(1); 1569 // 1& (int)y 1570 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) ); 1571 // zero node 1572 Node *conzero = intcon(0); 1573 // Check (1&(int)y)==0? 1574 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero)); 1575 // Check if (1&(int)y)!=0?, if so the result is negative 1576 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); 1577 // abs(x) 1578 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); 1579 // abs(x)^y 1580 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) ); 1581 // -abs(x)^y 1582 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); 1583 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1584 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1585 // Set complex path fast result 1586 phi->init_req(2, signresult); 1587 1588 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1589 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1590 r->init_req(1,slow_path); 1591 phi->init_req(1,slow_result); 1592 1593 // Post merge 1594 set_control(_gvn.transform(r)); 1595 record_for_igvn(r); 1596 result=_gvn.transform(phi); 1597 } 1598 1599 //------------------- 1600 //result=(result.isNaN())? uncommon_trap():result; 1601 // Check: If isNaN() by checking result!=result? then go to Strict Math 1602 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1603 // Build the boolean node 1604 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1605 1606 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1607 // End the current control-flow path 1608 push_pair(x); 1609 push_pair(y); 1610 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow 1611 // to handle. Recompile without intrinsifying Math.pow. 1612 uncommon_trap(Deoptimization::Reason_intrinsic, 1613 Deoptimization::Action_make_not_entrant); 1614 } 1615 1616 C->set_has_split_ifs(true); // Has chance for split-if optimization 1617 1618 push_pair(result); 1619 1620 return true; 1621 } 1622 1623 //------------------------------inline_trans------------------------------------- 1624 // Inline transcendental instructions, if possible. The Intel hardware gets 1625 // these right, no funny corner cases missed. 1626 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { 1627 _sp += arg_size(); // restore stack pointer 1628 Node* arg = pop_math_arg(); 1629 Node* trans = NULL; 1630 1631 switch (id) { 1632 case vmIntrinsics::_dlog: 1633 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); 1634 break; 1635 case vmIntrinsics::_dlog10: 1636 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); 1637 break; 1638 default: 1639 assert(false, "bad intrinsic was passed in"); 1640 return false; 1641 } 1642 1643 // Push result back on JVM stack 1644 push_pair(trans); 1645 return true; 1646 } 1647 1648 //------------------------------runtime_math----------------------------- 1649 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1650 Node* a = NULL; 1651 Node* b = NULL; 1652 1653 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1654 "must be (DD)D or (D)D type"); 1655 1656 // Inputs 1657 _sp += arg_size(); // restore stack pointer 1658 if (call_type == OptoRuntime::Math_DD_D_Type()) { 1659 b = pop_math_arg(); 1660 } 1661 a = pop_math_arg(); 1662 1663 const TypePtr* no_memory_effects = NULL; 1664 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1665 no_memory_effects, 1666 a, top(), b, b ? top() : NULL); 1667 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); 1668 #ifdef ASSERT 1669 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); 1670 assert(value_top == top(), "second value must be top"); 1671 #endif 1672 1673 push_pair(value); 1674 return true; 1675 } 1676 1677 //------------------------------inline_math_native----------------------------- 1678 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1679 switch (id) { 1680 // These intrinsics are not properly supported on all hardware 1681 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1682 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); 1683 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1684 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); 1685 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1686 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); 1687 1688 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : 1689 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); 1690 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : 1691 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); 1692 1693 // These intrinsics are supported on all hardware 1694 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; 1695 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; 1696 1697 // These intrinsics don't work on X86. The ad implementation doesn't 1698 // handle NaN's properly. Instead of returning infinity, the ad 1699 // implementation returns a NaN on overflow. See bug: 6304089 1700 // Once the ad implementations are fixed, change the code below 1701 // to match the intrinsics above 1702 1703 case vmIntrinsics::_dexp: return 1704 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1705 case vmIntrinsics::_dpow: return 1706 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1707 1708 // These intrinsics are not yet correctly implemented 1709 case vmIntrinsics::_datan2: 1710 return false; 1711 1712 default: 1713 ShouldNotReachHere(); 1714 return false; 1715 } 1716 } 1717 1718 static bool is_simple_name(Node* n) { 1719 return (n->req() == 1 // constant 1720 || (n->is_Type() && n->as_Type()->type()->singleton()) 1721 || n->is_Proj() // parameter or return value 1722 || n->is_Phi() // local of some sort 1723 ); 1724 } 1725 1726 //----------------------------inline_min_max----------------------------------- 1727 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1728 push(generate_min_max(id, argument(0), argument(1))); 1729 1730 return true; 1731 } 1732 1733 Node* 1734 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1735 // These are the candidate return value: 1736 Node* xvalue = x0; 1737 Node* yvalue = y0; 1738 1739 if (xvalue == yvalue) { 1740 return xvalue; 1741 } 1742 1743 bool want_max = (id == vmIntrinsics::_max); 1744 1745 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1746 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1747 if (txvalue == NULL || tyvalue == NULL) return top(); 1748 // This is not really necessary, but it is consistent with a 1749 // hypothetical MaxINode::Value method: 1750 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1751 1752 // %%% This folding logic should (ideally) be in a different place. 1753 // Some should be inside IfNode, and there to be a more reliable 1754 // transformation of ?: style patterns into cmoves. We also want 1755 // more powerful optimizations around cmove and min/max. 1756 1757 // Try to find a dominating comparison of these guys. 1758 // It can simplify the index computation for Arrays.copyOf 1759 // and similar uses of System.arraycopy. 1760 // First, compute the normalized version of CmpI(x, y). 1761 int cmp_op = Op_CmpI; 1762 Node* xkey = xvalue; 1763 Node* ykey = yvalue; 1764 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); 1765 if (ideal_cmpxy->is_Cmp()) { 1766 // E.g., if we have CmpI(length - offset, count), 1767 // it might idealize to CmpI(length, count + offset) 1768 cmp_op = ideal_cmpxy->Opcode(); 1769 xkey = ideal_cmpxy->in(1); 1770 ykey = ideal_cmpxy->in(2); 1771 } 1772 1773 // Start by locating any relevant comparisons. 1774 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1775 Node* cmpxy = NULL; 1776 Node* cmpyx = NULL; 1777 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1778 Node* cmp = start_from->fast_out(k); 1779 if (cmp->outcnt() > 0 && // must have prior uses 1780 cmp->in(0) == NULL && // must be context-independent 1781 cmp->Opcode() == cmp_op) { // right kind of compare 1782 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1783 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1784 } 1785 } 1786 1787 const int NCMPS = 2; 1788 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1789 int cmpn; 1790 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1791 if (cmps[cmpn] != NULL) break; // find a result 1792 } 1793 if (cmpn < NCMPS) { 1794 // Look for a dominating test that tells us the min and max. 1795 int depth = 0; // Limit search depth for speed 1796 Node* dom = control(); 1797 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1798 if (++depth >= 100) break; 1799 Node* ifproj = dom; 1800 if (!ifproj->is_Proj()) continue; 1801 Node* iff = ifproj->in(0); 1802 if (!iff->is_If()) continue; 1803 Node* bol = iff->in(1); 1804 if (!bol->is_Bool()) continue; 1805 Node* cmp = bol->in(1); 1806 if (cmp == NULL) continue; 1807 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1808 if (cmps[cmpn] == cmp) break; 1809 if (cmpn == NCMPS) continue; 1810 BoolTest::mask btest = bol->as_Bool()->_test._test; 1811 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1812 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1813 // At this point, we know that 'x btest y' is true. 1814 switch (btest) { 1815 case BoolTest::eq: 1816 // They are proven equal, so we can collapse the min/max. 1817 // Either value is the answer. Choose the simpler. 1818 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1819 return yvalue; 1820 return xvalue; 1821 case BoolTest::lt: // x < y 1822 case BoolTest::le: // x <= y 1823 return (want_max ? yvalue : xvalue); 1824 case BoolTest::gt: // x > y 1825 case BoolTest::ge: // x >= y 1826 return (want_max ? xvalue : yvalue); 1827 } 1828 } 1829 } 1830 1831 // We failed to find a dominating test. 1832 // Let's pick a test that might GVN with prior tests. 1833 Node* best_bol = NULL; 1834 BoolTest::mask best_btest = BoolTest::illegal; 1835 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1836 Node* cmp = cmps[cmpn]; 1837 if (cmp == NULL) continue; 1838 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1839 Node* bol = cmp->fast_out(j); 1840 if (!bol->is_Bool()) continue; 1841 BoolTest::mask btest = bol->as_Bool()->_test._test; 1842 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1843 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1844 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1845 best_bol = bol->as_Bool(); 1846 best_btest = btest; 1847 } 1848 } 1849 } 1850 1851 Node* answer_if_true = NULL; 1852 Node* answer_if_false = NULL; 1853 switch (best_btest) { 1854 default: 1855 if (cmpxy == NULL) 1856 cmpxy = ideal_cmpxy; 1857 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); 1858 // and fall through: 1859 case BoolTest::lt: // x < y 1860 case BoolTest::le: // x <= y 1861 answer_if_true = (want_max ? yvalue : xvalue); 1862 answer_if_false = (want_max ? xvalue : yvalue); 1863 break; 1864 case BoolTest::gt: // x > y 1865 case BoolTest::ge: // x >= y 1866 answer_if_true = (want_max ? xvalue : yvalue); 1867 answer_if_false = (want_max ? yvalue : xvalue); 1868 break; 1869 } 1870 1871 jint hi, lo; 1872 if (want_max) { 1873 // We can sharpen the minimum. 1874 hi = MAX2(txvalue->_hi, tyvalue->_hi); 1875 lo = MAX2(txvalue->_lo, tyvalue->_lo); 1876 } else { 1877 // We can sharpen the maximum. 1878 hi = MIN2(txvalue->_hi, tyvalue->_hi); 1879 lo = MIN2(txvalue->_lo, tyvalue->_lo); 1880 } 1881 1882 // Use a flow-free graph structure, to avoid creating excess control edges 1883 // which could hinder other optimizations. 1884 // Since Math.min/max is often used with arraycopy, we want 1885 // tightly_coupled_allocation to be able to see beyond min/max expressions. 1886 Node* cmov = CMoveNode::make(C, NULL, best_bol, 1887 answer_if_false, answer_if_true, 1888 TypeInt::make(lo, hi, widen)); 1889 1890 return _gvn.transform(cmov); 1891 1892 /* 1893 // This is not as desirable as it may seem, since Min and Max 1894 // nodes do not have a full set of optimizations. 1895 // And they would interfere, anyway, with 'if' optimizations 1896 // and with CMoveI canonical forms. 1897 switch (id) { 1898 case vmIntrinsics::_min: 1899 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 1900 case vmIntrinsics::_max: 1901 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 1902 default: 1903 ShouldNotReachHere(); 1904 } 1905 */ 1906 } 1907 1908 inline int 1909 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 1910 const TypePtr* base_type = TypePtr::NULL_PTR; 1911 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 1912 if (base_type == NULL) { 1913 // Unknown type. 1914 return Type::AnyPtr; 1915 } else if (base_type == TypePtr::NULL_PTR) { 1916 // Since this is a NULL+long form, we have to switch to a rawptr. 1917 base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); 1918 offset = MakeConX(0); 1919 return Type::RawPtr; 1920 } else if (base_type->base() == Type::RawPtr) { 1921 return Type::RawPtr; 1922 } else if (base_type->isa_oopptr()) { 1923 // Base is never null => always a heap address. 1924 if (base_type->ptr() == TypePtr::NotNull) { 1925 return Type::OopPtr; 1926 } 1927 // Offset is small => always a heap address. 1928 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 1929 if (offset_type != NULL && 1930 base_type->offset() == 0 && // (should always be?) 1931 offset_type->_lo >= 0 && 1932 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 1933 return Type::OopPtr; 1934 } 1935 // Otherwise, it might either be oop+off or NULL+addr. 1936 return Type::AnyPtr; 1937 } else { 1938 // No information: 1939 return Type::AnyPtr; 1940 } 1941 } 1942 1943 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 1944 int kind = classify_unsafe_addr(base, offset); 1945 if (kind == Type::RawPtr) { 1946 return basic_plus_adr(top(), base, offset); 1947 } else { 1948 return basic_plus_adr(base, offset); 1949 } 1950 } 1951 1952 //-------------------inline_numberOfLeadingZeros_int/long----------------------- 1953 // inline int Integer.numberOfLeadingZeros(int) 1954 // inline int Long.numberOfLeadingZeros(long) 1955 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) { 1956 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros"); 1957 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false; 1958 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false; 1959 _sp += arg_size(); // restore stack pointer 1960 switch (id) { 1961 case vmIntrinsics::_numberOfLeadingZeros_i: 1962 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop()))); 1963 break; 1964 case vmIntrinsics::_numberOfLeadingZeros_l: 1965 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair()))); 1966 break; 1967 default: 1968 ShouldNotReachHere(); 1969 } 1970 return true; 1971 } 1972 1973 //-------------------inline_numberOfTrailingZeros_int/long---------------------- 1974 // inline int Integer.numberOfTrailingZeros(int) 1975 // inline int Long.numberOfTrailingZeros(long) 1976 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) { 1977 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros"); 1978 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false; 1979 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false; 1980 _sp += arg_size(); // restore stack pointer 1981 switch (id) { 1982 case vmIntrinsics::_numberOfTrailingZeros_i: 1983 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop()))); 1984 break; 1985 case vmIntrinsics::_numberOfTrailingZeros_l: 1986 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair()))); 1987 break; 1988 default: 1989 ShouldNotReachHere(); 1990 } 1991 return true; 1992 } 1993 1994 //----------------------------inline_bitCount_int/long----------------------- 1995 // inline int Integer.bitCount(int) 1996 // inline int Long.bitCount(long) 1997 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) { 1998 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount"); 1999 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false; 2000 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false; 2001 _sp += arg_size(); // restore stack pointer 2002 switch (id) { 2003 case vmIntrinsics::_bitCount_i: 2004 push(_gvn.transform(new (C, 2) PopCountINode(pop()))); 2005 break; 2006 case vmIntrinsics::_bitCount_l: 2007 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair()))); 2008 break; 2009 default: 2010 ShouldNotReachHere(); 2011 } 2012 return true; 2013 } 2014 2015 //----------------------------inline_reverseBytes_int/long/char/short------------------- 2016 // inline Integer.reverseBytes(int) 2017 // inline Long.reverseBytes(long) 2018 // inline Character.reverseBytes(char) 2019 // inline Short.reverseBytes(short) 2020 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { 2021 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l || 2022 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s, 2023 "not reverse Bytes"); 2024 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; 2025 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; 2026 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false; 2027 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false; 2028 _sp += arg_size(); // restore stack pointer 2029 switch (id) { 2030 case vmIntrinsics::_reverseBytes_i: 2031 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); 2032 break; 2033 case vmIntrinsics::_reverseBytes_l: 2034 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); 2035 break; 2036 case vmIntrinsics::_reverseBytes_c: 2037 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop()))); 2038 break; 2039 case vmIntrinsics::_reverseBytes_s: 2040 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop()))); 2041 break; 2042 default: 2043 ; 2044 } 2045 return true; 2046 } 2047 2048 //----------------------------inline_unsafe_access---------------------------- 2049 2050 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2051 2052 // Interpret Unsafe.fieldOffset cookies correctly: 2053 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2054 2055 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2056 if (callee()->is_static()) return false; // caller must have the capability! 2057 2058 #ifndef PRODUCT 2059 { 2060 ResourceMark rm; 2061 // Check the signatures. 2062 ciSignature* sig = signature(); 2063 #ifdef ASSERT 2064 if (!is_store) { 2065 // Object getObject(Object base, int/long offset), etc. 2066 BasicType rtype = sig->return_type()->basic_type(); 2067 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2068 rtype = T_ADDRESS; // it is really a C void* 2069 assert(rtype == type, "getter must return the expected value"); 2070 if (!is_native_ptr) { 2071 assert(sig->count() == 2, "oop getter has 2 arguments"); 2072 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2073 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2074 } else { 2075 assert(sig->count() == 1, "native getter has 1 argument"); 2076 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2077 } 2078 } else { 2079 // void putObject(Object base, int/long offset, Object x), etc. 2080 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2081 if (!is_native_ptr) { 2082 assert(sig->count() == 3, "oop putter has 3 arguments"); 2083 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2084 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2085 } else { 2086 assert(sig->count() == 2, "native putter has 2 arguments"); 2087 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2088 } 2089 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2090 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2091 vtype = T_ADDRESS; // it is really a C void* 2092 assert(vtype == type, "putter must accept the expected value"); 2093 } 2094 #endif // ASSERT 2095 } 2096 #endif //PRODUCT 2097 2098 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2099 2100 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; 2101 2102 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words 2103 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); 2104 2105 debug_only(int saved_sp = _sp); 2106 _sp += nargs; 2107 2108 Node* val; 2109 debug_only(val = (Node*)(uintptr_t)-1); 2110 2111 2112 if (is_store) { 2113 // Get the value being stored. (Pop it first; it was pushed last.) 2114 switch (type) { 2115 case T_DOUBLE: 2116 case T_LONG: 2117 case T_ADDRESS: 2118 val = pop_pair(); 2119 break; 2120 default: 2121 val = pop(); 2122 } 2123 } 2124 2125 // Build address expression. See the code in inline_unsafe_prefetch. 2126 Node *adr; 2127 Node *heap_base_oop = top(); 2128 if (!is_native_ptr) { 2129 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2130 Node* offset = pop_pair(); 2131 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2132 Node* base = pop(); 2133 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2134 // to be plain byte offsets, which are also the same as those accepted 2135 // by oopDesc::field_base. 2136 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2137 "fieldOffset must be byte-scaled"); 2138 // 32-bit machines ignore the high half! 2139 offset = ConvL2X(offset); 2140 adr = make_unsafe_address(base, offset); 2141 heap_base_oop = base; 2142 } else { 2143 Node* ptr = pop_pair(); 2144 // Adjust Java long to machine word: 2145 ptr = ConvL2X(ptr); 2146 adr = make_unsafe_address(NULL, ptr); 2147 } 2148 2149 // Pop receiver last: it was pushed first. 2150 Node *receiver = pop(); 2151 2152 assert(saved_sp == _sp, "must have correct argument count"); 2153 2154 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2155 2156 // First guess at the value type. 2157 const Type *value_type = Type::get_const_basic_type(type); 2158 2159 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2160 // there was not enough information to nail it down. 2161 Compile::AliasType* alias_type = C->alias_type(adr_type); 2162 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2163 2164 // We will need memory barriers unless we can determine a unique 2165 // alias category for this reference. (Note: If for some reason 2166 // the barriers get omitted and the unsafe reference begins to "pollute" 2167 // the alias analysis of the rest of the graph, either Compile::can_alias 2168 // or Compile::must_alias will throw a diagnostic assert.) 2169 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2170 2171 if (!is_store && type == T_OBJECT) { 2172 // Attempt to infer a sharper value type from the offset and base type. 2173 ciKlass* sharpened_klass = NULL; 2174 2175 // See if it is an instance field, with an object type. 2176 if (alias_type->field() != NULL) { 2177 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2178 if (alias_type->field()->type()->is_klass()) { 2179 sharpened_klass = alias_type->field()->type()->as_klass(); 2180 } 2181 } 2182 2183 // See if it is a narrow oop array. 2184 if (adr_type->isa_aryptr()) { 2185 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2186 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2187 if (elem_type != NULL) { 2188 sharpened_klass = elem_type->klass(); 2189 } 2190 } 2191 } 2192 2193 if (sharpened_klass != NULL) { 2194 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2195 2196 // Sharpen the value type. 2197 value_type = tjp; 2198 2199 #ifndef PRODUCT 2200 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2201 tty->print(" from base type: "); adr_type->dump(); 2202 tty->print(" sharpened value: "); value_type->dump(); 2203 } 2204 #endif 2205 } 2206 } 2207 2208 // Null check on self without removing any arguments. The argument 2209 // null check technically happens in the wrong place, which can lead to 2210 // invalid stack traces when the primitive is inlined into a method 2211 // which handles NullPointerExceptions. 2212 _sp += nargs; 2213 do_null_check(receiver, T_OBJECT); 2214 _sp -= nargs; 2215 if (stopped()) { 2216 return true; 2217 } 2218 // Heap pointers get a null-check from the interpreter, 2219 // as a courtesy. However, this is not guaranteed by Unsafe, 2220 // and it is not possible to fully distinguish unintended nulls 2221 // from intended ones in this API. 2222 2223 if (is_volatile) { 2224 // We need to emit leading and trailing CPU membars (see below) in 2225 // addition to memory membars when is_volatile. This is a little 2226 // too strong, but avoids the need to insert per-alias-type 2227 // volatile membars (for stores; compare Parse::do_put_xxx), which 2228 // we cannot do effectively here because we probably only have a 2229 // rough approximation of type. 2230 need_mem_bar = true; 2231 // For Stores, place a memory ordering barrier now. 2232 if (is_store) 2233 insert_mem_bar(Op_MemBarRelease); 2234 } 2235 2236 // Memory barrier to prevent normal and 'unsafe' accesses from 2237 // bypassing each other. Happens after null checks, so the 2238 // exception paths do not take memory state from the memory barrier, 2239 // so there's no problems making a strong assert about mixing users 2240 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2241 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2242 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2243 2244 if (!is_store) { 2245 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2246 // load value and push onto stack 2247 switch (type) { 2248 case T_BOOLEAN: 2249 case T_CHAR: 2250 case T_BYTE: 2251 case T_SHORT: 2252 case T_INT: 2253 case T_FLOAT: 2254 case T_OBJECT: 2255 push( p ); 2256 break; 2257 case T_ADDRESS: 2258 // Cast to an int type. 2259 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); 2260 p = ConvX2L(p); 2261 push_pair(p); 2262 break; 2263 case T_DOUBLE: 2264 case T_LONG: 2265 push_pair( p ); 2266 break; 2267 default: ShouldNotReachHere(); 2268 } 2269 } else { 2270 // place effect of store into memory 2271 switch (type) { 2272 case T_DOUBLE: 2273 val = dstore_rounding(val); 2274 break; 2275 case T_ADDRESS: 2276 // Repackage the long as a pointer. 2277 val = ConvL2X(val); 2278 val = _gvn.transform( new (C, 2) CastX2PNode(val) ); 2279 break; 2280 } 2281 2282 if (type != T_OBJECT ) { 2283 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2284 } else { 2285 // Possibly an oop being stored to Java heap or native memory 2286 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2287 // oop to Java heap. 2288 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2289 } else { 2290 // We can't tell at compile time if we are storing in the Java heap or outside 2291 // of it. So we need to emit code to conditionally do the proper type of 2292 // store. 2293 2294 IdealKit ideal(gvn(), control(), merged_memory()); 2295 #define __ ideal. 2296 // QQQ who knows what probability is here?? 2297 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2298 // Sync IdealKit and graphKit. 2299 set_all_memory( __ merged_memory()); 2300 set_control(__ ctrl()); 2301 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2302 // Update IdealKit memory. 2303 __ set_all_memory(merged_memory()); 2304 __ set_ctrl(control()); 2305 } __ else_(); { 2306 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile); 2307 } __ end_if(); 2308 // Final sync IdealKit and GraphKit. 2309 sync_kit(ideal); 2310 #undef __ 2311 } 2312 } 2313 } 2314 2315 if (is_volatile) { 2316 if (!is_store) 2317 insert_mem_bar(Op_MemBarAcquire); 2318 else 2319 insert_mem_bar(Op_MemBarVolatile); 2320 } 2321 2322 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2323 2324 return true; 2325 } 2326 2327 //----------------------------inline_unsafe_prefetch---------------------------- 2328 2329 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2330 #ifndef PRODUCT 2331 { 2332 ResourceMark rm; 2333 // Check the signatures. 2334 ciSignature* sig = signature(); 2335 #ifdef ASSERT 2336 // Object getObject(Object base, int/long offset), etc. 2337 BasicType rtype = sig->return_type()->basic_type(); 2338 if (!is_native_ptr) { 2339 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2340 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2341 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2342 } else { 2343 assert(sig->count() == 1, "native prefetch has 1 argument"); 2344 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2345 } 2346 #endif // ASSERT 2347 } 2348 #endif // !PRODUCT 2349 2350 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2351 2352 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args 2353 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); 2354 2355 debug_only(int saved_sp = _sp); 2356 _sp += nargs; 2357 2358 // Build address expression. See the code in inline_unsafe_access. 2359 Node *adr; 2360 if (!is_native_ptr) { 2361 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2362 Node* offset = pop_pair(); 2363 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2364 Node* base = pop(); 2365 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2366 // to be plain byte offsets, which are also the same as those accepted 2367 // by oopDesc::field_base. 2368 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2369 "fieldOffset must be byte-scaled"); 2370 // 32-bit machines ignore the high half! 2371 offset = ConvL2X(offset); 2372 adr = make_unsafe_address(base, offset); 2373 } else { 2374 Node* ptr = pop_pair(); 2375 // Adjust Java long to machine word: 2376 ptr = ConvL2X(ptr); 2377 adr = make_unsafe_address(NULL, ptr); 2378 } 2379 2380 if (is_static) { 2381 assert(saved_sp == _sp, "must have correct argument count"); 2382 } else { 2383 // Pop receiver last: it was pushed first. 2384 Node *receiver = pop(); 2385 assert(saved_sp == _sp, "must have correct argument count"); 2386 2387 // Null check on self without removing any arguments. The argument 2388 // null check technically happens in the wrong place, which can lead to 2389 // invalid stack traces when the primitive is inlined into a method 2390 // which handles NullPointerExceptions. 2391 _sp += nargs; 2392 do_null_check(receiver, T_OBJECT); 2393 _sp -= nargs; 2394 if (stopped()) { 2395 return true; 2396 } 2397 } 2398 2399 // Generate the read or write prefetch 2400 Node *prefetch; 2401 if (is_store) { 2402 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); 2403 } else { 2404 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); 2405 } 2406 prefetch->init_req(0, control()); 2407 set_i_o(_gvn.transform(prefetch)); 2408 2409 return true; 2410 } 2411 2412 //----------------------------inline_unsafe_CAS---------------------------- 2413 2414 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { 2415 // This basic scheme here is the same as inline_unsafe_access, but 2416 // differs in enough details that combining them would make the code 2417 // overly confusing. (This is a true fact! I originally combined 2418 // them, but even I was confused by it!) As much code/comments as 2419 // possible are retained from inline_unsafe_access though to make 2420 // the correspondences clearer. - dl 2421 2422 if (callee()->is_static()) return false; // caller must have the capability! 2423 2424 #ifndef PRODUCT 2425 { 2426 ResourceMark rm; 2427 // Check the signatures. 2428 ciSignature* sig = signature(); 2429 #ifdef ASSERT 2430 BasicType rtype = sig->return_type()->basic_type(); 2431 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2432 assert(sig->count() == 4, "CAS has 4 arguments"); 2433 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2434 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2435 #endif // ASSERT 2436 } 2437 #endif //PRODUCT 2438 2439 // number of stack slots per value argument (1 or 2) 2440 int type_words = type2size[type]; 2441 2442 // Cannot inline wide CAS on machines that don't support it natively 2443 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8()) 2444 return false; 2445 2446 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2447 2448 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; 2449 int nargs = 1 + 1 + 2 + type_words + type_words; 2450 2451 // pop arguments: newval, oldval, offset, base, and receiver 2452 debug_only(int saved_sp = _sp); 2453 _sp += nargs; 2454 Node* newval = (type_words == 1) ? pop() : pop_pair(); 2455 Node* oldval = (type_words == 1) ? pop() : pop_pair(); 2456 Node *offset = pop_pair(); 2457 Node *base = pop(); 2458 Node *receiver = pop(); 2459 assert(saved_sp == _sp, "must have correct argument count"); 2460 2461 // Null check receiver. 2462 _sp += nargs; 2463 do_null_check(receiver, T_OBJECT); 2464 _sp -= nargs; 2465 if (stopped()) { 2466 return true; 2467 } 2468 2469 // Build field offset expression. 2470 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2471 // to be plain byte offsets, which are also the same as those accepted 2472 // by oopDesc::field_base. 2473 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2474 // 32-bit machines ignore the high half of long offsets 2475 offset = ConvL2X(offset); 2476 Node* adr = make_unsafe_address(base, offset); 2477 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2478 2479 // (Unlike inline_unsafe_access, there seems no point in trying 2480 // to refine types. Just use the coarse types here. 2481 const Type *value_type = Type::get_const_basic_type(type); 2482 Compile::AliasType* alias_type = C->alias_type(adr_type); 2483 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2484 int alias_idx = C->get_alias_index(adr_type); 2485 2486 // Memory-model-wise, a CAS acts like a little synchronized block, 2487 // so needs barriers on each side. These don't translate into 2488 // actual barriers on most machines, but we still need rest of 2489 // compiler to respect ordering. 2490 2491 insert_mem_bar(Op_MemBarRelease); 2492 insert_mem_bar(Op_MemBarCPUOrder); 2493 2494 // 4984716: MemBars must be inserted before this 2495 // memory node in order to avoid a false 2496 // dependency which will confuse the scheduler. 2497 Node *mem = memory(alias_idx); 2498 2499 // For now, we handle only those cases that actually exist: ints, 2500 // longs, and Object. Adding others should be straightforward. 2501 Node* cas; 2502 switch(type) { 2503 case T_INT: 2504 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2505 break; 2506 case T_LONG: 2507 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2508 break; 2509 case T_OBJECT: 2510 // reference stores need a store barrier. 2511 // (They don't if CAS fails, but it isn't worth checking.) 2512 pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT); 2513 #ifdef _LP64 2514 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2515 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2516 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2517 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr, 2518 newval_enc, oldval_enc)); 2519 } else 2520 #endif 2521 { 2522 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2523 } 2524 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); 2525 break; 2526 default: 2527 ShouldNotReachHere(); 2528 break; 2529 } 2530 2531 // SCMemProjNodes represent the memory state of CAS. Their main 2532 // role is to prevent CAS nodes from being optimized away when their 2533 // results aren't used. 2534 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 2535 set_memory(proj, alias_idx); 2536 2537 // Add the trailing membar surrounding the access 2538 insert_mem_bar(Op_MemBarCPUOrder); 2539 insert_mem_bar(Op_MemBarAcquire); 2540 2541 push(cas); 2542 return true; 2543 } 2544 2545 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2546 // This is another variant of inline_unsafe_access, differing in 2547 // that it always issues store-store ("release") barrier and ensures 2548 // store-atomicity (which only matters for "long"). 2549 2550 if (callee()->is_static()) return false; // caller must have the capability! 2551 2552 #ifndef PRODUCT 2553 { 2554 ResourceMark rm; 2555 // Check the signatures. 2556 ciSignature* sig = signature(); 2557 #ifdef ASSERT 2558 BasicType rtype = sig->return_type()->basic_type(); 2559 assert(rtype == T_VOID, "must return void"); 2560 assert(sig->count() == 3, "has 3 arguments"); 2561 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2562 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2563 #endif // ASSERT 2564 } 2565 #endif //PRODUCT 2566 2567 // number of stack slots per value argument (1 or 2) 2568 int type_words = type2size[type]; 2569 2570 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2571 2572 // Argument words: "this" plus oop plus offset plus value; 2573 int nargs = 1 + 1 + 2 + type_words; 2574 2575 // pop arguments: val, offset, base, and receiver 2576 debug_only(int saved_sp = _sp); 2577 _sp += nargs; 2578 Node* val = (type_words == 1) ? pop() : pop_pair(); 2579 Node *offset = pop_pair(); 2580 Node *base = pop(); 2581 Node *receiver = pop(); 2582 assert(saved_sp == _sp, "must have correct argument count"); 2583 2584 // Null check receiver. 2585 _sp += nargs; 2586 do_null_check(receiver, T_OBJECT); 2587 _sp -= nargs; 2588 if (stopped()) { 2589 return true; 2590 } 2591 2592 // Build field offset expression. 2593 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2594 // 32-bit machines ignore the high half of long offsets 2595 offset = ConvL2X(offset); 2596 Node* adr = make_unsafe_address(base, offset); 2597 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2598 const Type *value_type = Type::get_const_basic_type(type); 2599 Compile::AliasType* alias_type = C->alias_type(adr_type); 2600 2601 insert_mem_bar(Op_MemBarRelease); 2602 insert_mem_bar(Op_MemBarCPUOrder); 2603 // Ensure that the store is atomic for longs: 2604 bool require_atomic_access = true; 2605 Node* store; 2606 if (type == T_OBJECT) // reference stores need a store barrier. 2607 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2608 else { 2609 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2610 } 2611 insert_mem_bar(Op_MemBarCPUOrder); 2612 return true; 2613 } 2614 2615 bool LibraryCallKit::inline_unsafe_allocate() { 2616 if (callee()->is_static()) return false; // caller must have the capability! 2617 int nargs = 1 + 1; 2618 assert(signature()->size() == nargs-1, "alloc has 1 argument"); 2619 null_check_receiver(callee()); // check then ignore argument(0) 2620 _sp += nargs; // set original stack for use by uncommon_trap 2621 Node* cls = do_null_check(argument(1), T_OBJECT); 2622 _sp -= nargs; 2623 if (stopped()) return true; 2624 2625 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 2626 _sp += nargs; // set original stack for use by uncommon_trap 2627 kls = do_null_check(kls, T_OBJECT); 2628 _sp -= nargs; 2629 if (stopped()) return true; // argument was like int.class 2630 2631 // Note: The argument might still be an illegal value like 2632 // Serializable.class or Object[].class. The runtime will handle it. 2633 // But we must make an explicit check for initialization. 2634 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc)); 2635 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT); 2636 Node* bits = intcon(instanceKlass::fully_initialized); 2637 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); 2638 // The 'test' is non-zero if we need to take a slow path. 2639 2640 Node* obj = new_instance(kls, test); 2641 push(obj); 2642 2643 return true; 2644 } 2645 2646 //------------------------inline_native_time_funcs-------------- 2647 // inline code for System.currentTimeMillis() and System.nanoTime() 2648 // these have the same type and signature 2649 bool LibraryCallKit::inline_native_time_funcs(bool isNano) { 2650 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) : 2651 CAST_FROM_FN_PTR(address, os::javaTimeMillis); 2652 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis"; 2653 const TypeFunc *tf = OptoRuntime::current_time_millis_Type(); 2654 const TypePtr* no_memory_effects = NULL; 2655 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2656 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); 2657 #ifdef ASSERT 2658 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); 2659 assert(value_top == top(), "second value must be top"); 2660 #endif 2661 push_pair(value); 2662 return true; 2663 } 2664 2665 //------------------------inline_native_currentThread------------------ 2666 bool LibraryCallKit::inline_native_currentThread() { 2667 Node* junk = NULL; 2668 push(generate_current_thread(junk)); 2669 return true; 2670 } 2671 2672 //------------------------inline_native_isInterrupted------------------ 2673 bool LibraryCallKit::inline_native_isInterrupted() { 2674 const int nargs = 1+1; // receiver + boolean 2675 assert(nargs == arg_size(), "sanity"); 2676 // Add a fast path to t.isInterrupted(clear_int): 2677 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 2678 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 2679 // So, in the common case that the interrupt bit is false, 2680 // we avoid making a call into the VM. Even if the interrupt bit 2681 // is true, if the clear_int argument is false, we avoid the VM call. 2682 // However, if the receiver is not currentThread, we must call the VM, 2683 // because there must be some locking done around the operation. 2684 2685 // We only go to the fast case code if we pass two guards. 2686 // Paths which do not pass are accumulated in the slow_region. 2687 RegionNode* slow_region = new (C, 1) RegionNode(1); 2688 record_for_igvn(slow_region); 2689 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow 2690 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); 2691 enum { no_int_result_path = 1, 2692 no_clear_result_path = 2, 2693 slow_result_path = 3 2694 }; 2695 2696 // (a) Receiving thread must be the current thread. 2697 Node* rec_thr = argument(0); 2698 Node* tls_ptr = NULL; 2699 Node* cur_thr = generate_current_thread(tls_ptr); 2700 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); 2701 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); 2702 2703 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); 2704 if (!known_current_thread) 2705 generate_slow_guard(bol_thr, slow_region); 2706 2707 // (b) Interrupt bit on TLS must be false. 2708 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2709 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 2710 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 2711 // Set the control input on the field _interrupted read to prevent it floating up. 2712 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 2713 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); 2714 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); 2715 2716 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2717 2718 // First fast path: if (!TLS._interrupted) return false; 2719 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); 2720 result_rgn->init_req(no_int_result_path, false_bit); 2721 result_val->init_req(no_int_result_path, intcon(0)); 2722 2723 // drop through to next case 2724 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); 2725 2726 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 2727 Node* clr_arg = argument(1); 2728 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); 2729 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); 2730 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 2731 2732 // Second fast path: ... else if (!clear_int) return true; 2733 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); 2734 result_rgn->init_req(no_clear_result_path, false_arg); 2735 result_val->init_req(no_clear_result_path, intcon(1)); 2736 2737 // drop through to next case 2738 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); 2739 2740 // (d) Otherwise, go to the slow path. 2741 slow_region->add_req(control()); 2742 set_control( _gvn.transform(slow_region) ); 2743 2744 if (stopped()) { 2745 // There is no slow path. 2746 result_rgn->init_req(slow_result_path, top()); 2747 result_val->init_req(slow_result_path, top()); 2748 } else { 2749 // non-virtual because it is a private non-static 2750 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 2751 2752 Node* slow_val = set_results_for_java_call(slow_call); 2753 // this->control() comes from set_results_for_java_call 2754 2755 // If we know that the result of the slow call will be true, tell the optimizer! 2756 if (known_current_thread) slow_val = intcon(1); 2757 2758 Node* fast_io = slow_call->in(TypeFunc::I_O); 2759 Node* fast_mem = slow_call->in(TypeFunc::Memory); 2760 // These two phis are pre-filled with copies of of the fast IO and Memory 2761 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 2762 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 2763 2764 result_rgn->init_req(slow_result_path, control()); 2765 io_phi ->init_req(slow_result_path, i_o()); 2766 mem_phi ->init_req(slow_result_path, reset_memory()); 2767 result_val->init_req(slow_result_path, slow_val); 2768 2769 set_all_memory( _gvn.transform(mem_phi) ); 2770 set_i_o( _gvn.transform(io_phi) ); 2771 } 2772 2773 push_result(result_rgn, result_val); 2774 C->set_has_split_ifs(true); // Has chance for split-if optimization 2775 2776 return true; 2777 } 2778 2779 //---------------------------load_mirror_from_klass---------------------------- 2780 // Given a klass oop, load its java mirror (a java.lang.Class oop). 2781 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 2782 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc)); 2783 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 2784 } 2785 2786 //-----------------------load_klass_from_mirror_common------------------------- 2787 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 2788 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 2789 // and branch to the given path on the region. 2790 // If never_see_null, take an uncommon trap on null, so we can optimistically 2791 // compile for the non-null case. 2792 // If the region is NULL, force never_see_null = true. 2793 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 2794 bool never_see_null, 2795 int nargs, 2796 RegionNode* region, 2797 int null_path, 2798 int offset) { 2799 if (region == NULL) never_see_null = true; 2800 Node* p = basic_plus_adr(mirror, offset); 2801 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 2802 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); 2803 _sp += nargs; // any deopt will start just before call to enclosing method 2804 Node* null_ctl = top(); 2805 kls = null_check_oop(kls, &null_ctl, never_see_null); 2806 if (region != NULL) { 2807 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 2808 region->init_req(null_path, null_ctl); 2809 } else { 2810 assert(null_ctl == top(), "no loose ends"); 2811 } 2812 _sp -= nargs; 2813 return kls; 2814 } 2815 2816 //--------------------(inline_native_Class_query helpers)--------------------- 2817 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 2818 // Fall through if (mods & mask) == bits, take the guard otherwise. 2819 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 2820 // Branch around if the given klass has the given modifier bit set. 2821 // Like generate_guard, adds a new path onto the region. 2822 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); 2823 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 2824 Node* mask = intcon(modifier_mask); 2825 Node* bits = intcon(modifier_bits); 2826 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); 2827 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); 2828 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); 2829 return generate_fair_guard(bol, region); 2830 } 2831 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 2832 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 2833 } 2834 2835 //-------------------------inline_native_Class_query------------------- 2836 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 2837 int nargs = 1+0; // just the Class mirror, in most cases 2838 const Type* return_type = TypeInt::BOOL; 2839 Node* prim_return_value = top(); // what happens if it's a primitive class? 2840 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 2841 bool expect_prim = false; // most of these guys expect to work on refs 2842 2843 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 2844 2845 switch (id) { 2846 case vmIntrinsics::_isInstance: 2847 nargs = 1+1; // the Class mirror, plus the object getting queried about 2848 // nothing is an instance of a primitive type 2849 prim_return_value = intcon(0); 2850 break; 2851 case vmIntrinsics::_getModifiers: 2852 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2853 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 2854 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 2855 break; 2856 case vmIntrinsics::_isInterface: 2857 prim_return_value = intcon(0); 2858 break; 2859 case vmIntrinsics::_isArray: 2860 prim_return_value = intcon(0); 2861 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 2862 break; 2863 case vmIntrinsics::_isPrimitive: 2864 prim_return_value = intcon(1); 2865 expect_prim = true; // obviously 2866 break; 2867 case vmIntrinsics::_getSuperclass: 2868 prim_return_value = null(); 2869 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 2870 break; 2871 case vmIntrinsics::_getComponentType: 2872 prim_return_value = null(); 2873 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 2874 break; 2875 case vmIntrinsics::_getClassAccessFlags: 2876 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2877 return_type = TypeInt::INT; // not bool! 6297094 2878 break; 2879 default: 2880 ShouldNotReachHere(); 2881 } 2882 2883 Node* mirror = argument(0); 2884 Node* obj = (nargs <= 1)? top(): argument(1); 2885 2886 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 2887 if (mirror_con == NULL) return false; // cannot happen? 2888 2889 #ifndef PRODUCT 2890 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2891 ciType* k = mirror_con->java_mirror_type(); 2892 if (k) { 2893 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 2894 k->print_name(); 2895 tty->cr(); 2896 } 2897 } 2898 #endif 2899 2900 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 2901 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2902 record_for_igvn(region); 2903 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); 2904 2905 // The mirror will never be null of Reflection.getClassAccessFlags, however 2906 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 2907 // if it is. See bug 4774291. 2908 2909 // For Reflection.getClassAccessFlags(), the null check occurs in 2910 // the wrong place; see inline_unsafe_access(), above, for a similar 2911 // situation. 2912 _sp += nargs; // set original stack for use by uncommon_trap 2913 mirror = do_null_check(mirror, T_OBJECT); 2914 _sp -= nargs; 2915 // If mirror or obj is dead, only null-path is taken. 2916 if (stopped()) return true; 2917 2918 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 2919 2920 // Now load the mirror's klass metaobject, and null-check it. 2921 // Side-effects region with the control path if the klass is null. 2922 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, 2923 region, _prim_path); 2924 // If kls is null, we have a primitive mirror. 2925 phi->init_req(_prim_path, prim_return_value); 2926 if (stopped()) { push_result(region, phi); return true; } 2927 2928 Node* p; // handy temp 2929 Node* null_ctl; 2930 2931 // Now that we have the non-null klass, we can perform the real query. 2932 // For constant classes, the query will constant-fold in LoadNode::Value. 2933 Node* query_value = top(); 2934 switch (id) { 2935 case vmIntrinsics::_isInstance: 2936 // nothing is an instance of a primitive type 2937 query_value = gen_instanceof(obj, kls); 2938 break; 2939 2940 case vmIntrinsics::_getModifiers: 2941 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc)); 2942 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 2943 break; 2944 2945 case vmIntrinsics::_isInterface: 2946 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 2947 if (generate_interface_guard(kls, region) != NULL) 2948 // A guard was added. If the guard is taken, it was an interface. 2949 phi->add_req(intcon(1)); 2950 // If we fall through, it's a plain class. 2951 query_value = intcon(0); 2952 break; 2953 2954 case vmIntrinsics::_isArray: 2955 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 2956 if (generate_array_guard(kls, region) != NULL) 2957 // A guard was added. If the guard is taken, it was an array. 2958 phi->add_req(intcon(1)); 2959 // If we fall through, it's a plain class. 2960 query_value = intcon(0); 2961 break; 2962 2963 case vmIntrinsics::_isPrimitive: 2964 query_value = intcon(0); // "normal" path produces false 2965 break; 2966 2967 case vmIntrinsics::_getSuperclass: 2968 // The rules here are somewhat unfortunate, but we can still do better 2969 // with random logic than with a JNI call. 2970 // Interfaces store null or Object as _super, but must report null. 2971 // Arrays store an intermediate super as _super, but must report Object. 2972 // Other types can report the actual _super. 2973 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 2974 if (generate_interface_guard(kls, region) != NULL) 2975 // A guard was added. If the guard is taken, it was an interface. 2976 phi->add_req(null()); 2977 if (generate_array_guard(kls, region) != NULL) 2978 // A guard was added. If the guard is taken, it was an array. 2979 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 2980 // If we fall through, it's a plain class. Get its _super. 2981 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc)); 2982 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); 2983 null_ctl = top(); 2984 kls = null_check_oop(kls, &null_ctl); 2985 if (null_ctl != top()) { 2986 // If the guard is taken, Object.superClass is null (both klass and mirror). 2987 region->add_req(null_ctl); 2988 phi ->add_req(null()); 2989 } 2990 if (!stopped()) { 2991 query_value = load_mirror_from_klass(kls); 2992 } 2993 break; 2994 2995 case vmIntrinsics::_getComponentType: 2996 if (generate_array_guard(kls, region) != NULL) { 2997 // Be sure to pin the oop load to the guard edge just created: 2998 Node* is_array_ctrl = region->in(region->req()-1); 2999 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc)); 3000 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 3001 phi->add_req(cmo); 3002 } 3003 query_value = null(); // non-array case is null 3004 break; 3005 3006 case vmIntrinsics::_getClassAccessFlags: 3007 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); 3008 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3009 break; 3010 3011 default: 3012 ShouldNotReachHere(); 3013 } 3014 3015 // Fall-through is the normal case of a query to a real class. 3016 phi->init_req(1, query_value); 3017 region->init_req(1, control()); 3018 3019 push_result(region, phi); 3020 C->set_has_split_ifs(true); // Has chance for split-if optimization 3021 3022 return true; 3023 } 3024 3025 //--------------------------inline_native_subtype_check------------------------ 3026 // This intrinsic takes the JNI calls out of the heart of 3027 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3028 bool LibraryCallKit::inline_native_subtype_check() { 3029 int nargs = 1+1; // the Class mirror, plus the other class getting examined 3030 3031 // Pull both arguments off the stack. 3032 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3033 args[0] = argument(0); 3034 args[1] = argument(1); 3035 Node* klasses[2]; // corresponding Klasses: superk, subk 3036 klasses[0] = klasses[1] = top(); 3037 3038 enum { 3039 // A full decision tree on {superc is prim, subc is prim}: 3040 _prim_0_path = 1, // {P,N} => false 3041 // {P,P} & superc!=subc => false 3042 _prim_same_path, // {P,P} & superc==subc => true 3043 _prim_1_path, // {N,P} => false 3044 _ref_subtype_path, // {N,N} & subtype check wins => true 3045 _both_ref_path, // {N,N} & subtype check loses => false 3046 PATH_LIMIT 3047 }; 3048 3049 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3050 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 3051 record_for_igvn(region); 3052 3053 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3054 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3055 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3056 3057 // First null-check both mirrors and load each mirror's klass metaobject. 3058 int which_arg; 3059 for (which_arg = 0; which_arg <= 1; which_arg++) { 3060 Node* arg = args[which_arg]; 3061 _sp += nargs; // set original stack for use by uncommon_trap 3062 arg = do_null_check(arg, T_OBJECT); 3063 _sp -= nargs; 3064 if (stopped()) break; 3065 args[which_arg] = _gvn.transform(arg); 3066 3067 Node* p = basic_plus_adr(arg, class_klass_offset); 3068 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3069 klasses[which_arg] = _gvn.transform(kls); 3070 } 3071 3072 // Having loaded both klasses, test each for null. 3073 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3074 for (which_arg = 0; which_arg <= 1; which_arg++) { 3075 Node* kls = klasses[which_arg]; 3076 Node* null_ctl = top(); 3077 _sp += nargs; // set original stack for use by uncommon_trap 3078 kls = null_check_oop(kls, &null_ctl, never_see_null); 3079 _sp -= nargs; 3080 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3081 region->init_req(prim_path, null_ctl); 3082 if (stopped()) break; 3083 klasses[which_arg] = kls; 3084 } 3085 3086 if (!stopped()) { 3087 // now we have two reference types, in klasses[0..1] 3088 Node* subk = klasses[1]; // the argument to isAssignableFrom 3089 Node* superk = klasses[0]; // the receiver 3090 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3091 // now we have a successful reference subtype check 3092 region->set_req(_ref_subtype_path, control()); 3093 } 3094 3095 // If both operands are primitive (both klasses null), then 3096 // we must return true when they are identical primitives. 3097 // It is convenient to test this after the first null klass check. 3098 set_control(region->in(_prim_0_path)); // go back to first null check 3099 if (!stopped()) { 3100 // Since superc is primitive, make a guard for the superc==subc case. 3101 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); 3102 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); 3103 generate_guard(bol_eq, region, PROB_FAIR); 3104 if (region->req() == PATH_LIMIT+1) { 3105 // A guard was added. If the added guard is taken, superc==subc. 3106 region->swap_edges(PATH_LIMIT, _prim_same_path); 3107 region->del_req(PATH_LIMIT); 3108 } 3109 region->set_req(_prim_0_path, control()); // Not equal after all. 3110 } 3111 3112 // these are the only paths that produce 'true': 3113 phi->set_req(_prim_same_path, intcon(1)); 3114 phi->set_req(_ref_subtype_path, intcon(1)); 3115 3116 // pull together the cases: 3117 assert(region->req() == PATH_LIMIT, "sane region"); 3118 for (uint i = 1; i < region->req(); i++) { 3119 Node* ctl = region->in(i); 3120 if (ctl == NULL || ctl == top()) { 3121 region->set_req(i, top()); 3122 phi ->set_req(i, top()); 3123 } else if (phi->in(i) == NULL) { 3124 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3125 } 3126 } 3127 3128 set_control(_gvn.transform(region)); 3129 push(_gvn.transform(phi)); 3130 3131 return true; 3132 } 3133 3134 //---------------------generate_array_guard_common------------------------ 3135 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3136 bool obj_array, bool not_array) { 3137 // If obj_array/non_array==false/false: 3138 // Branch around if the given klass is in fact an array (either obj or prim). 3139 // If obj_array/non_array==false/true: 3140 // Branch around if the given klass is not an array klass of any kind. 3141 // If obj_array/non_array==true/true: 3142 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3143 // If obj_array/non_array==true/false: 3144 // Branch around if the kls is an oop array (Object[] or subtype) 3145 // 3146 // Like generate_guard, adds a new path onto the region. 3147 jint layout_con = 0; 3148 Node* layout_val = get_layout_helper(kls, layout_con); 3149 if (layout_val == NULL) { 3150 bool query = (obj_array 3151 ? Klass::layout_helper_is_objArray(layout_con) 3152 : Klass::layout_helper_is_javaArray(layout_con)); 3153 if (query == not_array) { 3154 return NULL; // never a branch 3155 } else { // always a branch 3156 Node* always_branch = control(); 3157 if (region != NULL) 3158 region->add_req(always_branch); 3159 set_control(top()); 3160 return always_branch; 3161 } 3162 } 3163 // Now test the correct condition. 3164 jint nval = (obj_array 3165 ? ((jint)Klass::_lh_array_tag_type_value 3166 << Klass::_lh_array_tag_shift) 3167 : Klass::_lh_neutral_value); 3168 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); 3169 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3170 // invert the test if we are looking for a non-array 3171 if (not_array) btest = BoolTest(btest).negate(); 3172 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); 3173 return generate_fair_guard(bol, region); 3174 } 3175 3176 3177 //-----------------------inline_native_newArray-------------------------- 3178 bool LibraryCallKit::inline_native_newArray() { 3179 int nargs = 2; 3180 Node* mirror = argument(0); 3181 Node* count_val = argument(1); 3182 3183 _sp += nargs; // set original stack for use by uncommon_trap 3184 mirror = do_null_check(mirror, T_OBJECT); 3185 _sp -= nargs; 3186 // If mirror or obj is dead, only null-path is taken. 3187 if (stopped()) return true; 3188 3189 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3190 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3191 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3192 TypeInstPtr::NOTNULL); 3193 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3194 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3195 TypePtr::BOTTOM); 3196 3197 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3198 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3199 nargs, 3200 result_reg, _slow_path); 3201 Node* normal_ctl = control(); 3202 Node* no_array_ctl = result_reg->in(_slow_path); 3203 3204 // Generate code for the slow case. We make a call to newArray(). 3205 set_control(no_array_ctl); 3206 if (!stopped()) { 3207 // Either the input type is void.class, or else the 3208 // array klass has not yet been cached. Either the 3209 // ensuing call will throw an exception, or else it 3210 // will cache the array klass for next time. 3211 PreserveJVMState pjvms(this); 3212 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3213 Node* slow_result = set_results_for_java_call(slow_call); 3214 // this->control() comes from set_results_for_java_call 3215 result_reg->set_req(_slow_path, control()); 3216 result_val->set_req(_slow_path, slow_result); 3217 result_io ->set_req(_slow_path, i_o()); 3218 result_mem->set_req(_slow_path, reset_memory()); 3219 } 3220 3221 set_control(normal_ctl); 3222 if (!stopped()) { 3223 // Normal case: The array type has been cached in the java.lang.Class. 3224 // The following call works fine even if the array type is polymorphic. 3225 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3226 Node* obj = new_array(klass_node, count_val, nargs); 3227 result_reg->init_req(_normal_path, control()); 3228 result_val->init_req(_normal_path, obj); 3229 result_io ->init_req(_normal_path, i_o()); 3230 result_mem->init_req(_normal_path, reset_memory()); 3231 } 3232 3233 // Return the combined state. 3234 set_i_o( _gvn.transform(result_io) ); 3235 set_all_memory( _gvn.transform(result_mem) ); 3236 push_result(result_reg, result_val); 3237 C->set_has_split_ifs(true); // Has chance for split-if optimization 3238 3239 return true; 3240 } 3241 3242 //----------------------inline_native_getLength-------------------------- 3243 bool LibraryCallKit::inline_native_getLength() { 3244 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3245 3246 int nargs = 1; 3247 Node* array = argument(0); 3248 3249 _sp += nargs; // set original stack for use by uncommon_trap 3250 array = do_null_check(array, T_OBJECT); 3251 _sp -= nargs; 3252 3253 // If array is dead, only null-path is taken. 3254 if (stopped()) return true; 3255 3256 // Deoptimize if it is a non-array. 3257 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3258 3259 if (non_array != NULL) { 3260 PreserveJVMState pjvms(this); 3261 set_control(non_array); 3262 _sp += nargs; // push the arguments back on the stack 3263 uncommon_trap(Deoptimization::Reason_intrinsic, 3264 Deoptimization::Action_maybe_recompile); 3265 } 3266 3267 // If control is dead, only non-array-path is taken. 3268 if (stopped()) return true; 3269 3270 // The works fine even if the array type is polymorphic. 3271 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3272 push( load_array_length(array) ); 3273 3274 C->set_has_split_ifs(true); // Has chance for split-if optimization 3275 3276 return true; 3277 } 3278 3279 //------------------------inline_array_copyOf---------------------------- 3280 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3281 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3282 3283 // Restore the stack and pop off the arguments. 3284 int nargs = 3 + (is_copyOfRange? 1: 0); 3285 Node* original = argument(0); 3286 Node* start = is_copyOfRange? argument(1): intcon(0); 3287 Node* end = is_copyOfRange? argument(2): argument(1); 3288 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3289 3290 Node* newcopy; 3291 3292 //set the original stack and the reexecute bit for the interpreter to reexecute 3293 //the bytecode that invokes Arrays.copyOf if deoptimization happens 3294 { PreserveReexecuteState preexecs(this); 3295 _sp += nargs; 3296 jvms()->set_should_reexecute(true); 3297 3298 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); 3299 original = do_null_check(original, T_OBJECT); 3300 3301 // Check if a null path was taken unconditionally. 3302 if (stopped()) return true; 3303 3304 Node* orig_length = load_array_length(original); 3305 3306 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0, 3307 NULL, 0); 3308 klass_node = do_null_check(klass_node, T_OBJECT); 3309 3310 RegionNode* bailout = new (C, 1) RegionNode(1); 3311 record_for_igvn(bailout); 3312 3313 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3314 // Bail out if that is so. 3315 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3316 if (not_objArray != NULL) { 3317 // Improve the klass node's type from the new optimistic assumption: 3318 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3319 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3320 Node* cast = new (C, 2) CastPPNode(klass_node, akls); 3321 cast->init_req(0, control()); 3322 klass_node = _gvn.transform(cast); 3323 } 3324 3325 // Bail out if either start or end is negative. 3326 generate_negative_guard(start, bailout, &start); 3327 generate_negative_guard(end, bailout, &end); 3328 3329 Node* length = end; 3330 if (_gvn.type(start) != TypeInt::ZERO) { 3331 length = _gvn.transform( new (C, 3) SubINode(end, start) ); 3332 } 3333 3334 // Bail out if length is negative. 3335 // ...Not needed, since the new_array will throw the right exception. 3336 //generate_negative_guard(length, bailout, &length); 3337 3338 if (bailout->req() > 1) { 3339 PreserveJVMState pjvms(this); 3340 set_control( _gvn.transform(bailout) ); 3341 uncommon_trap(Deoptimization::Reason_intrinsic, 3342 Deoptimization::Action_maybe_recompile); 3343 } 3344 3345 if (!stopped()) { 3346 3347 // How many elements will we copy from the original? 3348 // The answer is MinI(orig_length - start, length). 3349 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); 3350 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3351 3352 const bool raw_mem_only = true; 3353 newcopy = new_array(klass_node, length, 0, raw_mem_only); 3354 3355 // Generate a direct call to the right arraycopy function(s). 3356 // We know the copy is disjoint but we might not know if the 3357 // oop stores need checking. 3358 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3359 // This will fail a store-check if x contains any non-nulls. 3360 bool disjoint_bases = true; 3361 bool length_never_negative = true; 3362 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3363 original, start, newcopy, intcon(0), moved, 3364 disjoint_bases, length_never_negative); 3365 } 3366 } //original reexecute and sp are set back here 3367 3368 if(!stopped()) { 3369 push(newcopy); 3370 } 3371 3372 C->set_has_split_ifs(true); // Has chance for split-if optimization 3373 3374 return true; 3375 } 3376 3377 3378 //----------------------generate_virtual_guard--------------------------- 3379 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3380 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3381 RegionNode* slow_region) { 3382 ciMethod* method = callee(); 3383 int vtable_index = method->vtable_index(); 3384 // Get the methodOop out of the appropriate vtable entry. 3385 int entry_offset = (instanceKlass::vtable_start_offset() + 3386 vtable_index*vtableEntry::size()) * wordSize + 3387 vtableEntry::method_offset_in_bytes(); 3388 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3389 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); 3390 3391 // Compare the target method with the expected method (e.g., Object.hashCode). 3392 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method); 3393 3394 Node* native_call = makecon(native_call_addr); 3395 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); 3396 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); 3397 3398 return generate_slow_guard(test_native, slow_region); 3399 } 3400 3401 //-----------------------generate_method_call---------------------------- 3402 // Use generate_method_call to make a slow-call to the real 3403 // method if the fast path fails. An alternative would be to 3404 // use a stub like OptoRuntime::slow_arraycopy_Java. 3405 // This only works for expanding the current library call, 3406 // not another intrinsic. (E.g., don't use this for making an 3407 // arraycopy call inside of the copyOf intrinsic.) 3408 CallJavaNode* 3409 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3410 // When compiling the intrinsic method itself, do not use this technique. 3411 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3412 3413 ciMethod* method = callee(); 3414 // ensure the JVMS we have will be correct for this call 3415 guarantee(method_id == method->intrinsic_id(), "must match"); 3416 3417 const TypeFunc* tf = TypeFunc::make(method); 3418 int tfdc = tf->domain()->cnt(); 3419 CallJavaNode* slow_call; 3420 if (is_static) { 3421 assert(!is_virtual, ""); 3422 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3423 SharedRuntime::get_resolve_static_call_stub(), 3424 method, bci()); 3425 } else if (is_virtual) { 3426 null_check_receiver(method); 3427 int vtable_index = methodOopDesc::invalid_vtable_index; 3428 if (UseInlineCaches) { 3429 // Suppress the vtable call 3430 } else { 3431 // hashCode and clone are not a miranda methods, 3432 // so the vtable index is fixed. 3433 // No need to use the linkResolver to get it. 3434 vtable_index = method->vtable_index(); 3435 } 3436 slow_call = new(C, tfdc) CallDynamicJavaNode(tf, 3437 SharedRuntime::get_resolve_virtual_call_stub(), 3438 method, vtable_index, bci()); 3439 } else { // neither virtual nor static: opt_virtual 3440 null_check_receiver(method); 3441 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3442 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3443 method, bci()); 3444 slow_call->set_optimized_virtual(true); 3445 } 3446 set_arguments_for_java_call(slow_call); 3447 set_edges_for_java_call(slow_call); 3448 return slow_call; 3449 } 3450 3451 3452 //------------------------------inline_native_hashcode-------------------- 3453 // Build special case code for calls to hashCode on an object. 3454 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3455 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3456 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3457 3458 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3459 3460 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3461 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3462 TypeInt::INT); 3463 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3464 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3465 TypePtr::BOTTOM); 3466 Node* obj = NULL; 3467 if (!is_static) { 3468 // Check for hashing null object 3469 obj = null_check_receiver(callee()); 3470 if (stopped()) return true; // unconditionally null 3471 result_reg->init_req(_null_path, top()); 3472 result_val->init_req(_null_path, top()); 3473 } else { 3474 // Do a null check, and return zero if null. 3475 // System.identityHashCode(null) == 0 3476 obj = argument(0); 3477 Node* null_ctl = top(); 3478 obj = null_check_oop(obj, &null_ctl); 3479 result_reg->init_req(_null_path, null_ctl); 3480 result_val->init_req(_null_path, _gvn.intcon(0)); 3481 } 3482 3483 // Unconditionally null? Then return right away. 3484 if (stopped()) { 3485 set_control( result_reg->in(_null_path) ); 3486 if (!stopped()) 3487 push( result_val ->in(_null_path) ); 3488 return true; 3489 } 3490 3491 // After null check, get the object's klass. 3492 Node* obj_klass = load_object_klass(obj); 3493 3494 // This call may be virtual (invokevirtual) or bound (invokespecial). 3495 // For each case we generate slightly different code. 3496 3497 // We only go to the fast case code if we pass a number of guards. The 3498 // paths which do not pass are accumulated in the slow_region. 3499 RegionNode* slow_region = new (C, 1) RegionNode(1); 3500 record_for_igvn(slow_region); 3501 3502 // If this is a virtual call, we generate a funny guard. We pull out 3503 // the vtable entry corresponding to hashCode() from the target object. 3504 // If the target method which we are calling happens to be the native 3505 // Object hashCode() method, we pass the guard. We do not need this 3506 // guard for non-virtual calls -- the caller is known to be the native 3507 // Object hashCode(). 3508 if (is_virtual) { 3509 generate_virtual_guard(obj_klass, slow_region); 3510 } 3511 3512 // Get the header out of the object, use LoadMarkNode when available 3513 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3514 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS); 3515 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) ); 3516 3517 // Test the header to see if it is unlocked. 3518 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3519 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); 3520 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3521 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); 3522 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); 3523 3524 generate_slow_guard(test_unlocked, slow_region); 3525 3526 // Get the hash value and check to see that it has been properly assigned. 3527 // We depend on hash_mask being at most 32 bits and avoid the use of 3528 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3529 // vm: see markOop.hpp. 3530 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3531 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3532 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); 3533 // This hack lets the hash bits live anywhere in the mark object now, as long 3534 // as the shift drops the relevant bits into the low 32 bits. Note that 3535 // Java spec says that HashCode is an int so there's no point in capturing 3536 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3537 hshifted_header = ConvX2I(hshifted_header); 3538 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); 3539 3540 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3541 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); 3542 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); 3543 3544 generate_slow_guard(test_assigned, slow_region); 3545 3546 Node* init_mem = reset_memory(); 3547 // fill in the rest of the null path: 3548 result_io ->init_req(_null_path, i_o()); 3549 result_mem->init_req(_null_path, init_mem); 3550 3551 result_val->init_req(_fast_path, hash_val); 3552 result_reg->init_req(_fast_path, control()); 3553 result_io ->init_req(_fast_path, i_o()); 3554 result_mem->init_req(_fast_path, init_mem); 3555 3556 // Generate code for the slow case. We make a call to hashCode(). 3557 set_control(_gvn.transform(slow_region)); 3558 if (!stopped()) { 3559 // No need for PreserveJVMState, because we're using up the present state. 3560 set_all_memory(init_mem); 3561 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; 3562 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; 3563 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3564 Node* slow_result = set_results_for_java_call(slow_call); 3565 // this->control() comes from set_results_for_java_call 3566 result_reg->init_req(_slow_path, control()); 3567 result_val->init_req(_slow_path, slow_result); 3568 result_io ->set_req(_slow_path, i_o()); 3569 result_mem ->set_req(_slow_path, reset_memory()); 3570 } 3571 3572 // Return the combined state. 3573 set_i_o( _gvn.transform(result_io) ); 3574 set_all_memory( _gvn.transform(result_mem) ); 3575 push_result(result_reg, result_val); 3576 3577 return true; 3578 } 3579 3580 //---------------------------inline_native_getClass---------------------------- 3581 // Build special case code for calls to getClass on an object. 3582 bool LibraryCallKit::inline_native_getClass() { 3583 Node* obj = null_check_receiver(callee()); 3584 if (stopped()) return true; 3585 push( load_mirror_from_klass(load_object_klass(obj)) ); 3586 return true; 3587 } 3588 3589 //-----------------inline_native_Reflection_getCallerClass--------------------- 3590 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3591 // 3592 // NOTE that this code must perform the same logic as 3593 // vframeStream::security_get_caller_frame in that it must skip 3594 // Method.invoke() and auxiliary frames. 3595 3596 3597 3598 3599 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3600 ciMethod* method = callee(); 3601 3602 #ifndef PRODUCT 3603 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3604 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3605 } 3606 #endif 3607 3608 debug_only(int saved_sp = _sp); 3609 3610 // Argument words: (int depth) 3611 int nargs = 1; 3612 3613 _sp += nargs; 3614 Node* caller_depth_node = pop(); 3615 3616 assert(saved_sp == _sp, "must have correct argument count"); 3617 3618 // The depth value must be a constant in order for the runtime call 3619 // to be eliminated. 3620 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); 3621 if (caller_depth_type == NULL || !caller_depth_type->is_con()) { 3622 #ifndef PRODUCT 3623 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3624 tty->print_cr(" Bailing out because caller depth was not a constant"); 3625 } 3626 #endif 3627 return false; 3628 } 3629 // Note that the JVM state at this point does not include the 3630 // getCallerClass() frame which we are trying to inline. The 3631 // semantics of getCallerClass(), however, are that the "first" 3632 // frame is the getCallerClass() frame, so we subtract one from the 3633 // requested depth before continuing. We don't inline requests of 3634 // getCallerClass(0). 3635 int caller_depth = caller_depth_type->get_con() - 1; 3636 if (caller_depth < 0) { 3637 #ifndef PRODUCT 3638 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3639 tty->print_cr(" Bailing out because caller depth was %d", caller_depth); 3640 } 3641 #endif 3642 return false; 3643 } 3644 3645 if (!jvms()->has_method()) { 3646 #ifndef PRODUCT 3647 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3648 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3649 } 3650 #endif 3651 return false; 3652 } 3653 int _depth = jvms()->depth(); // cache call chain depth 3654 3655 // Walk back up the JVM state to find the caller at the required 3656 // depth. NOTE that this code must perform the same logic as 3657 // vframeStream::security_get_caller_frame in that it must skip 3658 // Method.invoke() and auxiliary frames. Note also that depth is 3659 // 1-based (1 is the bottom of the inlining). 3660 int inlining_depth = _depth; 3661 JVMState* caller_jvms = NULL; 3662 3663 if (inlining_depth > 0) { 3664 caller_jvms = jvms(); 3665 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); 3666 do { 3667 // The following if-tests should be performed in this order 3668 if (is_method_invoke_or_aux_frame(caller_jvms)) { 3669 // Skip a Method.invoke() or auxiliary frame 3670 } else if (caller_depth > 0) { 3671 // Skip real frame 3672 --caller_depth; 3673 } else { 3674 // We're done: reached desired caller after skipping. 3675 break; 3676 } 3677 caller_jvms = caller_jvms->caller(); 3678 --inlining_depth; 3679 } while (inlining_depth > 0); 3680 } 3681 3682 if (inlining_depth == 0) { 3683 #ifndef PRODUCT 3684 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3685 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); 3686 tty->print_cr(" JVM state at this point:"); 3687 for (int i = _depth; i >= 1; i--) { 3688 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3689 } 3690 } 3691 #endif 3692 return false; // Reached end of inlining 3693 } 3694 3695 // Acquire method holder as java.lang.Class 3696 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 3697 ciInstance* caller_mirror = caller_klass->java_mirror(); 3698 // Push this as a constant 3699 push(makecon(TypeInstPtr::make(caller_mirror))); 3700 #ifndef PRODUCT 3701 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3702 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); 3703 tty->print_cr(" JVM state at this point:"); 3704 for (int i = _depth; i >= 1; i--) { 3705 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3706 } 3707 } 3708 #endif 3709 return true; 3710 } 3711 3712 // Helper routine for above 3713 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { 3714 ciMethod* method = jvms->method(); 3715 3716 // Is this the Method.invoke method itself? 3717 if (method->intrinsic_id() == vmIntrinsics::_invoke) 3718 return true; 3719 3720 // Is this a helper, defined somewhere underneath MethodAccessorImpl. 3721 ciKlass* k = method->holder(); 3722 if (k->is_instance_klass()) { 3723 ciInstanceKlass* ik = k->as_instance_klass(); 3724 for (; ik != NULL; ik = ik->super()) { 3725 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && 3726 ik == env()->find_system_klass(ik->name())) { 3727 return true; 3728 } 3729 } 3730 } 3731 else if (method->is_method_handle_adapter()) { 3732 // This is an internal adapter frame from the MethodHandleCompiler -- skip it 3733 return true; 3734 } 3735 3736 return false; 3737 } 3738 3739 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by 3740 // inline_native_AtomicLong_attemptUpdate() but it has no way of 3741 // computing it since there is no lookup field by name function in the 3742 // CI interface. This is computed and set by inline_native_AtomicLong_get(). 3743 // Using a static variable here is safe even if we have multiple compilation 3744 // threads because the offset is constant. At worst the same offset will be 3745 // computed and stored multiple 3746 3747 bool LibraryCallKit::inline_native_AtomicLong_get() { 3748 // Restore the stack and pop off the argument 3749 _sp+=1; 3750 Node *obj = pop(); 3751 3752 // get the offset of the "value" field. Since the CI interfaces 3753 // does not provide a way to look up a field by name, we scan the bytecodes 3754 // to get the field index. We expect the first 2 instructions of the method 3755 // to be: 3756 // 0 aload_0 3757 // 1 getfield "value" 3758 ciMethod* method = callee(); 3759 if (value_field_offset == -1) 3760 { 3761 ciField* value_field; 3762 ciBytecodeStream iter(method); 3763 Bytecodes::Code bc = iter.next(); 3764 3765 if ((bc != Bytecodes::_aload_0) && 3766 ((bc != Bytecodes::_aload) || (iter.get_index() != 0))) 3767 return false; 3768 bc = iter.next(); 3769 if (bc != Bytecodes::_getfield) 3770 return false; 3771 bool ignore; 3772 value_field = iter.get_field(ignore); 3773 value_field_offset = value_field->offset_in_bytes(); 3774 } 3775 3776 // Null check without removing any arguments. 3777 _sp++; 3778 obj = do_null_check(obj, T_OBJECT); 3779 _sp--; 3780 // Check for locking null object 3781 if (stopped()) return true; 3782 3783 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3784 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3785 int alias_idx = C->get_alias_index(adr_type); 3786 3787 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr)); 3788 3789 push_pair(result); 3790 3791 return true; 3792 } 3793 3794 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() { 3795 // Restore the stack and pop off the arguments 3796 _sp+=5; 3797 Node *newVal = pop_pair(); 3798 Node *oldVal = pop_pair(); 3799 Node *obj = pop(); 3800 3801 // we need the offset of the "value" field which was computed when 3802 // inlining the get() method. Give up if we don't have it. 3803 if (value_field_offset == -1) 3804 return false; 3805 3806 // Null check without removing any arguments. 3807 _sp+=5; 3808 obj = do_null_check(obj, T_OBJECT); 3809 _sp-=5; 3810 // Check for locking null object 3811 if (stopped()) return true; 3812 3813 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3814 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3815 int alias_idx = C->get_alias_index(adr_type); 3816 3817 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal)); 3818 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 3819 set_memory(store_proj, alias_idx); 3820 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) ); 3821 3822 Node *result; 3823 // CMove node is not used to be able fold a possible check code 3824 // after attemptUpdate() call. This code could be transformed 3825 // into CMove node by loop optimizations. 3826 { 3827 RegionNode *r = new (C, 3) RegionNode(3); 3828 result = new (C, 3) PhiNode(r, TypeInt::BOOL); 3829 3830 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 3831 Node *iftrue = opt_iff(r, iff); 3832 r->init_req(1, iftrue); 3833 result->init_req(1, intcon(1)); 3834 result->init_req(2, intcon(0)); 3835 3836 set_control(_gvn.transform(r)); 3837 record_for_igvn(r); 3838 3839 C->set_has_split_ifs(true); // Has chance for split-if optimization 3840 } 3841 3842 push(_gvn.transform(result)); 3843 return true; 3844 } 3845 3846 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 3847 // restore the arguments 3848 _sp += arg_size(); 3849 3850 switch (id) { 3851 case vmIntrinsics::_floatToRawIntBits: 3852 push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); 3853 break; 3854 3855 case vmIntrinsics::_intBitsToFloat: 3856 push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); 3857 break; 3858 3859 case vmIntrinsics::_doubleToRawLongBits: 3860 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); 3861 break; 3862 3863 case vmIntrinsics::_longBitsToDouble: 3864 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); 3865 break; 3866 3867 case vmIntrinsics::_doubleToLongBits: { 3868 Node* value = pop_pair(); 3869 3870 // two paths (plus control) merge in a wood 3871 RegionNode *r = new (C, 3) RegionNode(3); 3872 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); 3873 3874 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); 3875 // Build the boolean node 3876 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 3877 3878 // Branch either way. 3879 // NaN case is less traveled, which makes all the difference. 3880 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3881 Node *opt_isnan = _gvn.transform(ifisnan); 3882 assert( opt_isnan->is_If(), "Expect an IfNode"); 3883 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3884 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 3885 3886 set_control(iftrue); 3887 3888 static const jlong nan_bits = CONST64(0x7ff8000000000000); 3889 Node *slow_result = longcon(nan_bits); // return NaN 3890 phi->init_req(1, _gvn.transform( slow_result )); 3891 r->init_req(1, iftrue); 3892 3893 // Else fall through 3894 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 3895 set_control(iffalse); 3896 3897 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); 3898 r->init_req(2, iffalse); 3899 3900 // Post merge 3901 set_control(_gvn.transform(r)); 3902 record_for_igvn(r); 3903 3904 Node* result = _gvn.transform(phi); 3905 assert(result->bottom_type()->isa_long(), "must be"); 3906 push_pair(result); 3907 3908 C->set_has_split_ifs(true); // Has chance for split-if optimization 3909 3910 break; 3911 } 3912 3913 case vmIntrinsics::_floatToIntBits: { 3914 Node* value = pop(); 3915 3916 // two paths (plus control) merge in a wood 3917 RegionNode *r = new (C, 3) RegionNode(3); 3918 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); 3919 3920 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); 3921 // Build the boolean node 3922 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 3923 3924 // Branch either way. 3925 // NaN case is less traveled, which makes all the difference. 3926 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3927 Node *opt_isnan = _gvn.transform(ifisnan); 3928 assert( opt_isnan->is_If(), "Expect an IfNode"); 3929 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3930 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 3931 3932 set_control(iftrue); 3933 3934 static const jint nan_bits = 0x7fc00000; 3935 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 3936 phi->init_req(1, _gvn.transform( slow_result )); 3937 r->init_req(1, iftrue); 3938 3939 // Else fall through 3940 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 3941 set_control(iffalse); 3942 3943 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); 3944 r->init_req(2, iffalse); 3945 3946 // Post merge 3947 set_control(_gvn.transform(r)); 3948 record_for_igvn(r); 3949 3950 Node* result = _gvn.transform(phi); 3951 assert(result->bottom_type()->isa_int(), "must be"); 3952 push(result); 3953 3954 C->set_has_split_ifs(true); // Has chance for split-if optimization 3955 3956 break; 3957 } 3958 3959 default: 3960 ShouldNotReachHere(); 3961 } 3962 3963 return true; 3964 } 3965 3966 #ifdef _LP64 3967 #define XTOP ,top() /*additional argument*/ 3968 #else //_LP64 3969 #define XTOP /*no additional argument*/ 3970 #endif //_LP64 3971 3972 //----------------------inline_unsafe_copyMemory------------------------- 3973 bool LibraryCallKit::inline_unsafe_copyMemory() { 3974 if (callee()->is_static()) return false; // caller must have the capability! 3975 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) 3976 assert(signature()->size() == nargs-1, "copy has 5 arguments"); 3977 null_check_receiver(callee()); // check then ignore argument(0) 3978 if (stopped()) return true; 3979 3980 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3981 3982 Node* src_ptr = argument(1); 3983 Node* src_off = ConvL2X(argument(2)); 3984 assert(argument(3)->is_top(), "2nd half of long"); 3985 Node* dst_ptr = argument(4); 3986 Node* dst_off = ConvL2X(argument(5)); 3987 assert(argument(6)->is_top(), "2nd half of long"); 3988 Node* size = ConvL2X(argument(7)); 3989 assert(argument(8)->is_top(), "2nd half of long"); 3990 3991 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 3992 "fieldOffset must be byte-scaled"); 3993 3994 Node* src = make_unsafe_address(src_ptr, src_off); 3995 Node* dst = make_unsafe_address(dst_ptr, dst_off); 3996 3997 // Conservatively insert a memory barrier on all memory slices. 3998 // Do not let writes of the copy source or destination float below the copy. 3999 insert_mem_bar(Op_MemBarCPUOrder); 4000 4001 // Call it. Note that the length argument is not scaled. 4002 make_runtime_call(RC_LEAF|RC_NO_FP, 4003 OptoRuntime::fast_arraycopy_Type(), 4004 StubRoutines::unsafe_arraycopy(), 4005 "unsafe_arraycopy", 4006 TypeRawPtr::BOTTOM, 4007 src, dst, size XTOP); 4008 4009 // Do not let reads of the copy destination float above the copy. 4010 insert_mem_bar(Op_MemBarCPUOrder); 4011 4012 return true; 4013 } 4014 4015 //------------------------clone_coping----------------------------------- 4016 // Helper function for inline_native_clone. 4017 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4018 assert(obj_size != NULL, ""); 4019 Node* raw_obj = alloc_obj->in(1); 4020 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4021 4022 if (ReduceBulkZeroing) { 4023 // We will be completely responsible for initializing this object - 4024 // mark Initialize node as complete. 4025 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4026 // The object was just allocated - there should be no any stores! 4027 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4028 } 4029 4030 // Copy the fastest available way. 4031 // TODO: generate fields copies for small objects instead. 4032 Node* src = obj; 4033 Node* dest = alloc_obj; 4034 Node* size = _gvn.transform(obj_size); 4035 4036 // Exclude the header but include array length to copy by 8 bytes words. 4037 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4038 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4039 instanceOopDesc::base_offset_in_bytes(); 4040 // base_off: 4041 // 8 - 32-bit VM 4042 // 12 - 64-bit VM, compressed oops 4043 // 16 - 64-bit VM, normal oops 4044 if (base_off % BytesPerLong != 0) { 4045 assert(UseCompressedOops, ""); 4046 if (is_array) { 4047 // Exclude length to copy by 8 bytes words. 4048 base_off += sizeof(int); 4049 } else { 4050 // Include klass to copy by 8 bytes words. 4051 base_off = instanceOopDesc::klass_offset_in_bytes(); 4052 } 4053 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4054 } 4055 src = basic_plus_adr(src, base_off); 4056 dest = basic_plus_adr(dest, base_off); 4057 4058 // Compute the length also, if needed: 4059 Node* countx = size; 4060 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); 4061 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4062 4063 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4064 bool disjoint_bases = true; 4065 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4066 src, NULL, dest, NULL, countx); 4067 4068 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4069 if (card_mark) { 4070 assert(!is_array, ""); 4071 // Put in store barrier for any and all oops we are sticking 4072 // into this object. (We could avoid this if we could prove 4073 // that the object type contains no oop fields at all.) 4074 Node* no_particular_value = NULL; 4075 Node* no_particular_field = NULL; 4076 int raw_adr_idx = Compile::AliasIdxRaw; 4077 post_barrier(control(), 4078 memory(raw_adr_type), 4079 alloc_obj, 4080 no_particular_field, 4081 raw_adr_idx, 4082 no_particular_value, 4083 T_OBJECT, 4084 false); 4085 } 4086 4087 // Do not let reads from the cloned object float above the arraycopy. 4088 insert_mem_bar(Op_MemBarCPUOrder); 4089 } 4090 4091 //------------------------inline_native_clone---------------------------- 4092 // Here are the simple edge cases: 4093 // null receiver => normal trap 4094 // virtual and clone was overridden => slow path to out-of-line clone 4095 // not cloneable or finalizer => slow path to out-of-line Object.clone 4096 // 4097 // The general case has two steps, allocation and copying. 4098 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4099 // 4100 // Copying also has two cases, oop arrays and everything else. 4101 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4102 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4103 // 4104 // These steps fold up nicely if and when the cloned object's klass 4105 // can be sharply typed as an object array, a type array, or an instance. 4106 // 4107 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4108 int nargs = 1; 4109 PhiNode* result_val; 4110 4111 //set the original stack and the reexecute bit for the interpreter to reexecute 4112 //the bytecode that invokes Object.clone if deoptimization happens 4113 { PreserveReexecuteState preexecs(this); 4114 jvms()->set_should_reexecute(true); 4115 4116 //null_check_receiver will adjust _sp (push and pop) 4117 Node* obj = null_check_receiver(callee()); 4118 if (stopped()) return true; 4119 4120 _sp += nargs; 4121 4122 Node* obj_klass = load_object_klass(obj); 4123 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4124 const TypeOopPtr* toop = ((tklass != NULL) 4125 ? tklass->as_instance_type() 4126 : TypeInstPtr::NOTNULL); 4127 4128 // Conservatively insert a memory barrier on all memory slices. 4129 // Do not let writes into the original float below the clone. 4130 insert_mem_bar(Op_MemBarCPUOrder); 4131 4132 // paths into result_reg: 4133 enum { 4134 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4135 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4136 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4137 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4138 PATH_LIMIT 4139 }; 4140 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4141 result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 4142 TypeInstPtr::NOTNULL); 4143 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 4144 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 4145 TypePtr::BOTTOM); 4146 record_for_igvn(result_reg); 4147 4148 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4149 int raw_adr_idx = Compile::AliasIdxRaw; 4150 const bool raw_mem_only = true; 4151 4152 4153 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4154 if (array_ctl != NULL) { 4155 // It's an array. 4156 PreserveJVMState pjvms(this); 4157 set_control(array_ctl); 4158 Node* obj_length = load_array_length(obj); 4159 Node* obj_size = NULL; 4160 Node* alloc_obj = new_array(obj_klass, obj_length, 0, 4161 raw_mem_only, &obj_size); 4162 4163 if (!use_ReduceInitialCardMarks()) { 4164 // If it is an oop array, it requires very special treatment, 4165 // because card marking is required on each card of the array. 4166 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4167 if (is_obja != NULL) { 4168 PreserveJVMState pjvms2(this); 4169 set_control(is_obja); 4170 // Generate a direct call to the right arraycopy function(s). 4171 bool disjoint_bases = true; 4172 bool length_never_negative = true; 4173 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4174 obj, intcon(0), alloc_obj, intcon(0), 4175 obj_length, 4176 disjoint_bases, length_never_negative); 4177 result_reg->init_req(_objArray_path, control()); 4178 result_val->init_req(_objArray_path, alloc_obj); 4179 result_i_o ->set_req(_objArray_path, i_o()); 4180 result_mem ->set_req(_objArray_path, reset_memory()); 4181 } 4182 } 4183 // Otherwise, there are no card marks to worry about. 4184 // (We can dispense with card marks if we know the allocation 4185 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4186 // causes the non-eden paths to take compensating steps to 4187 // simulate a fresh allocation, so that no further 4188 // card marks are required in compiled code to initialize 4189 // the object.) 4190 4191 if (!stopped()) { 4192 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4193 4194 // Present the results of the copy. 4195 result_reg->init_req(_array_path, control()); 4196 result_val->init_req(_array_path, alloc_obj); 4197 result_i_o ->set_req(_array_path, i_o()); 4198 result_mem ->set_req(_array_path, reset_memory()); 4199 } 4200 } 4201 4202 // We only go to the instance fast case code if we pass a number of guards. 4203 // The paths which do not pass are accumulated in the slow_region. 4204 RegionNode* slow_region = new (C, 1) RegionNode(1); 4205 record_for_igvn(slow_region); 4206 if (!stopped()) { 4207 // It's an instance (we did array above). Make the slow-path tests. 4208 // If this is a virtual call, we generate a funny guard. We grab 4209 // the vtable entry corresponding to clone() from the target object. 4210 // If the target method which we are calling happens to be the 4211 // Object clone() method, we pass the guard. We do not need this 4212 // guard for non-virtual calls; the caller is known to be the native 4213 // Object clone(). 4214 if (is_virtual) { 4215 generate_virtual_guard(obj_klass, slow_region); 4216 } 4217 4218 // The object must be cloneable and must not have a finalizer. 4219 // Both of these conditions may be checked in a single test. 4220 // We could optimize the cloneable test further, but we don't care. 4221 generate_access_flags_guard(obj_klass, 4222 // Test both conditions: 4223 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4224 // Must be cloneable but not finalizer: 4225 JVM_ACC_IS_CLONEABLE, 4226 slow_region); 4227 } 4228 4229 if (!stopped()) { 4230 // It's an instance, and it passed the slow-path tests. 4231 PreserveJVMState pjvms(this); 4232 Node* obj_size = NULL; 4233 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size); 4234 4235 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4236 4237 // Present the results of the slow call. 4238 result_reg->init_req(_instance_path, control()); 4239 result_val->init_req(_instance_path, alloc_obj); 4240 result_i_o ->set_req(_instance_path, i_o()); 4241 result_mem ->set_req(_instance_path, reset_memory()); 4242 } 4243 4244 // Generate code for the slow case. We make a call to clone(). 4245 set_control(_gvn.transform(slow_region)); 4246 if (!stopped()) { 4247 PreserveJVMState pjvms(this); 4248 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4249 Node* slow_result = set_results_for_java_call(slow_call); 4250 // this->control() comes from set_results_for_java_call 4251 result_reg->init_req(_slow_path, control()); 4252 result_val->init_req(_slow_path, slow_result); 4253 result_i_o ->set_req(_slow_path, i_o()); 4254 result_mem ->set_req(_slow_path, reset_memory()); 4255 } 4256 4257 // Return the combined state. 4258 set_control( _gvn.transform(result_reg) ); 4259 set_i_o( _gvn.transform(result_i_o) ); 4260 set_all_memory( _gvn.transform(result_mem) ); 4261 } //original reexecute and sp are set back here 4262 4263 push(_gvn.transform(result_val)); 4264 4265 return true; 4266 } 4267 4268 4269 // constants for computing the copy function 4270 enum { 4271 COPYFUNC_UNALIGNED = 0, 4272 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize 4273 COPYFUNC_CONJOINT = 0, 4274 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend 4275 }; 4276 4277 // Note: The condition "disjoint" applies also for overlapping copies 4278 // where an descending copy is permitted (i.e., dest_offset <= src_offset). 4279 static address 4280 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) { 4281 int selector = 4282 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) + 4283 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT); 4284 4285 #define RETURN_STUB(xxx_arraycopy) { \ 4286 name = #xxx_arraycopy; \ 4287 return StubRoutines::xxx_arraycopy(); } 4288 4289 switch (t) { 4290 case T_BYTE: 4291 case T_BOOLEAN: 4292 switch (selector) { 4293 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy); 4294 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy); 4295 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy); 4296 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy); 4297 } 4298 case T_CHAR: 4299 case T_SHORT: 4300 switch (selector) { 4301 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy); 4302 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy); 4303 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy); 4304 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy); 4305 } 4306 case T_INT: 4307 case T_FLOAT: 4308 switch (selector) { 4309 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy); 4310 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy); 4311 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy); 4312 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy); 4313 } 4314 case T_DOUBLE: 4315 case T_LONG: 4316 switch (selector) { 4317 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy); 4318 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy); 4319 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy); 4320 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy); 4321 } 4322 case T_ARRAY: 4323 case T_OBJECT: 4324 switch (selector) { 4325 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy); 4326 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy); 4327 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy); 4328 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy); 4329 } 4330 default: 4331 ShouldNotReachHere(); 4332 return NULL; 4333 } 4334 4335 #undef RETURN_STUB 4336 } 4337 4338 //------------------------------basictype2arraycopy---------------------------- 4339 address LibraryCallKit::basictype2arraycopy(BasicType t, 4340 Node* src_offset, 4341 Node* dest_offset, 4342 bool disjoint_bases, 4343 const char* &name) { 4344 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4345 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4346 4347 bool aligned = false; 4348 bool disjoint = disjoint_bases; 4349 4350 // if the offsets are the same, we can treat the memory regions as 4351 // disjoint, because either the memory regions are in different arrays, 4352 // or they are identical (which we can treat as disjoint.) We can also 4353 // treat a copy with a destination index less that the source index 4354 // as disjoint since a low->high copy will work correctly in this case. 4355 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4356 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4357 // both indices are constants 4358 int s_offs = src_offset_inttype->get_con(); 4359 int d_offs = dest_offset_inttype->get_con(); 4360 int element_size = type2aelembytes(t); 4361 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4362 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4363 if (s_offs >= d_offs) disjoint = true; 4364 } else if (src_offset == dest_offset && src_offset != NULL) { 4365 // This can occur if the offsets are identical non-constants. 4366 disjoint = true; 4367 } 4368 4369 return select_arraycopy_function(t, aligned, disjoint, name); 4370 } 4371 4372 4373 //------------------------------inline_arraycopy----------------------- 4374 bool LibraryCallKit::inline_arraycopy() { 4375 // Restore the stack and pop off the arguments. 4376 int nargs = 5; // 2 oops, 3 ints, no size_t or long 4377 assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); 4378 4379 Node *src = argument(0); 4380 Node *src_offset = argument(1); 4381 Node *dest = argument(2); 4382 Node *dest_offset = argument(3); 4383 Node *length = argument(4); 4384 4385 // Compile time checks. If any of these checks cannot be verified at compile time, 4386 // we do not make a fast path for this call. Instead, we let the call remain as it 4387 // is. The checks we choose to mandate at compile time are: 4388 // 4389 // (1) src and dest are arrays. 4390 const Type* src_type = src->Value(&_gvn); 4391 const Type* dest_type = dest->Value(&_gvn); 4392 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4393 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4394 if (top_src == NULL || top_src->klass() == NULL || 4395 top_dest == NULL || top_dest->klass() == NULL) { 4396 // Conservatively insert a memory barrier on all memory slices. 4397 // Do not let writes into the source float below the arraycopy. 4398 insert_mem_bar(Op_MemBarCPUOrder); 4399 4400 // Call StubRoutines::generic_arraycopy stub. 4401 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4402 src, src_offset, dest, dest_offset, length); 4403 4404 // Do not let reads from the destination float above the arraycopy. 4405 // Since we cannot type the arrays, we don't know which slices 4406 // might be affected. We could restrict this barrier only to those 4407 // memory slices which pertain to array elements--but don't bother. 4408 if (!InsertMemBarAfterArraycopy) 4409 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4410 insert_mem_bar(Op_MemBarCPUOrder); 4411 return true; 4412 } 4413 4414 // (2) src and dest arrays must have elements of the same BasicType 4415 // Figure out the size and type of the elements we will be copying. 4416 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4417 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4418 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4419 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4420 4421 if (src_elem != dest_elem || dest_elem == T_VOID) { 4422 // The component types are not the same or are not recognized. Punt. 4423 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4424 generate_slow_arraycopy(TypePtr::BOTTOM, 4425 src, src_offset, dest, dest_offset, length); 4426 return true; 4427 } 4428 4429 //--------------------------------------------------------------------------- 4430 // We will make a fast path for this call to arraycopy. 4431 4432 // We have the following tests left to perform: 4433 // 4434 // (3) src and dest must not be null. 4435 // (4) src_offset must not be negative. 4436 // (5) dest_offset must not be negative. 4437 // (6) length must not be negative. 4438 // (7) src_offset + length must not exceed length of src. 4439 // (8) dest_offset + length must not exceed length of dest. 4440 // (9) each element of an oop array must be assignable 4441 4442 RegionNode* slow_region = new (C, 1) RegionNode(1); 4443 record_for_igvn(slow_region); 4444 4445 // (3) operands must not be null 4446 // We currently perform our null checks with the do_null_check routine. 4447 // This means that the null exceptions will be reported in the caller 4448 // rather than (correctly) reported inside of the native arraycopy call. 4449 // This should be corrected, given time. We do our null check with the 4450 // stack pointer restored. 4451 _sp += nargs; 4452 src = do_null_check(src, T_ARRAY); 4453 dest = do_null_check(dest, T_ARRAY); 4454 _sp -= nargs; 4455 4456 // (4) src_offset must not be negative. 4457 generate_negative_guard(src_offset, slow_region); 4458 4459 // (5) dest_offset must not be negative. 4460 generate_negative_guard(dest_offset, slow_region); 4461 4462 // (6) length must not be negative (moved to generate_arraycopy()). 4463 // generate_negative_guard(length, slow_region); 4464 4465 // (7) src_offset + length must not exceed length of src. 4466 generate_limit_guard(src_offset, length, 4467 load_array_length(src), 4468 slow_region); 4469 4470 // (8) dest_offset + length must not exceed length of dest. 4471 generate_limit_guard(dest_offset, length, 4472 load_array_length(dest), 4473 slow_region); 4474 4475 // (9) each element of an oop array must be assignable 4476 // The generate_arraycopy subroutine checks this. 4477 4478 // This is where the memory effects are placed: 4479 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4480 generate_arraycopy(adr_type, dest_elem, 4481 src, src_offset, dest, dest_offset, length, 4482 false, false, slow_region); 4483 4484 return true; 4485 } 4486 4487 //-----------------------------generate_arraycopy---------------------- 4488 // Generate an optimized call to arraycopy. 4489 // Caller must guard against non-arrays. 4490 // Caller must determine a common array basic-type for both arrays. 4491 // Caller must validate offsets against array bounds. 4492 // The slow_region has already collected guard failure paths 4493 // (such as out of bounds length or non-conformable array types). 4494 // The generated code has this shape, in general: 4495 // 4496 // if (length == 0) return // via zero_path 4497 // slowval = -1 4498 // if (types unknown) { 4499 // slowval = call generic copy loop 4500 // if (slowval == 0) return // via checked_path 4501 // } else if (indexes in bounds) { 4502 // if ((is object array) && !(array type check)) { 4503 // slowval = call checked copy loop 4504 // if (slowval == 0) return // via checked_path 4505 // } else { 4506 // call bulk copy loop 4507 // return // via fast_path 4508 // } 4509 // } 4510 // // adjust params for remaining work: 4511 // if (slowval != -1) { 4512 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4513 // } 4514 // slow_region: 4515 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4516 // return // via slow_call_path 4517 // 4518 // This routine is used from several intrinsics: System.arraycopy, 4519 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4520 // 4521 void 4522 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4523 BasicType basic_elem_type, 4524 Node* src, Node* src_offset, 4525 Node* dest, Node* dest_offset, 4526 Node* copy_length, 4527 bool disjoint_bases, 4528 bool length_never_negative, 4529 RegionNode* slow_region) { 4530 4531 if (slow_region == NULL) { 4532 slow_region = new(C,1) RegionNode(1); 4533 record_for_igvn(slow_region); 4534 } 4535 4536 Node* original_dest = dest; 4537 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4538 bool must_clear_dest = false; 4539 4540 // See if this is the initialization of a newly-allocated array. 4541 // If so, we will take responsibility here for initializing it to zero. 4542 // (Note: Because tightly_coupled_allocation performs checks on the 4543 // out-edges of the dest, we need to avoid making derived pointers 4544 // from it until we have checked its uses.) 4545 if (ReduceBulkZeroing 4546 && !ZeroTLAB // pointless if already zeroed 4547 && basic_elem_type != T_CONFLICT // avoid corner case 4548 && !_gvn.eqv_uncast(src, dest) 4549 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4550 != NULL) 4551 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4552 && alloc->maybe_set_complete(&_gvn)) { 4553 // "You break it, you buy it." 4554 InitializeNode* init = alloc->initialization(); 4555 assert(init->is_complete(), "we just did this"); 4556 assert(dest->is_CheckCastPP(), "sanity"); 4557 assert(dest->in(0)->in(0) == init, "dest pinned"); 4558 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4559 // From this point on, every exit path is responsible for 4560 // initializing any non-copied parts of the object to zero. 4561 must_clear_dest = true; 4562 } else { 4563 // No zeroing elimination here. 4564 alloc = NULL; 4565 //original_dest = dest; 4566 //must_clear_dest = false; 4567 } 4568 4569 // Results are placed here: 4570 enum { fast_path = 1, // normal void-returning assembly stub 4571 checked_path = 2, // special assembly stub with cleanup 4572 slow_call_path = 3, // something went wrong; call the VM 4573 zero_path = 4, // bypass when length of copy is zero 4574 bcopy_path = 5, // copy primitive array by 64-bit blocks 4575 PATH_LIMIT = 6 4576 }; 4577 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4578 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO); 4579 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type); 4580 record_for_igvn(result_region); 4581 _gvn.set_type_bottom(result_i_o); 4582 _gvn.set_type_bottom(result_memory); 4583 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4584 4585 // The slow_control path: 4586 Node* slow_control; 4587 Node* slow_i_o = i_o(); 4588 Node* slow_mem = memory(adr_type); 4589 debug_only(slow_control = (Node*) badAddress); 4590 4591 // Checked control path: 4592 Node* checked_control = top(); 4593 Node* checked_mem = NULL; 4594 Node* checked_i_o = NULL; 4595 Node* checked_value = NULL; 4596 4597 if (basic_elem_type == T_CONFLICT) { 4598 assert(!must_clear_dest, ""); 4599 Node* cv = generate_generic_arraycopy(adr_type, 4600 src, src_offset, dest, dest_offset, 4601 copy_length); 4602 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4603 checked_control = control(); 4604 checked_i_o = i_o(); 4605 checked_mem = memory(adr_type); 4606 checked_value = cv; 4607 set_control(top()); // no fast path 4608 } 4609 4610 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4611 if (not_pos != NULL) { 4612 PreserveJVMState pjvms(this); 4613 set_control(not_pos); 4614 4615 // (6) length must not be negative. 4616 if (!length_never_negative) { 4617 generate_negative_guard(copy_length, slow_region); 4618 } 4619 4620 // copy_length is 0. 4621 if (!stopped() && must_clear_dest) { 4622 Node* dest_length = alloc->in(AllocateNode::ALength); 4623 if (_gvn.eqv_uncast(copy_length, dest_length) 4624 || _gvn.find_int_con(dest_length, 1) <= 0) { 4625 // There is no zeroing to do. No need for a secondary raw memory barrier. 4626 } else { 4627 // Clear the whole thing since there are no source elements to copy. 4628 generate_clear_array(adr_type, dest, basic_elem_type, 4629 intcon(0), NULL, 4630 alloc->in(AllocateNode::AllocSize)); 4631 // Use a secondary InitializeNode as raw memory barrier. 4632 // Currently it is needed only on this path since other 4633 // paths have stub or runtime calls as raw memory barriers. 4634 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4635 Compile::AliasIdxRaw, 4636 top())->as_Initialize(); 4637 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4638 } 4639 } 4640 4641 // Present the results of the fast call. 4642 result_region->init_req(zero_path, control()); 4643 result_i_o ->init_req(zero_path, i_o()); 4644 result_memory->init_req(zero_path, memory(adr_type)); 4645 } 4646 4647 if (!stopped() && must_clear_dest) { 4648 // We have to initialize the *uncopied* part of the array to zero. 4649 // The copy destination is the slice dest[off..off+len]. The other slices 4650 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4651 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4652 Node* dest_length = alloc->in(AllocateNode::ALength); 4653 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset, 4654 copy_length) ); 4655 4656 // If there is a head section that needs zeroing, do it now. 4657 if (find_int_con(dest_offset, -1) != 0) { 4658 generate_clear_array(adr_type, dest, basic_elem_type, 4659 intcon(0), dest_offset, 4660 NULL); 4661 } 4662 4663 // Next, perform a dynamic check on the tail length. 4664 // It is often zero, and we can win big if we prove this. 4665 // There are two wins: Avoid generating the ClearArray 4666 // with its attendant messy index arithmetic, and upgrade 4667 // the copy to a more hardware-friendly word size of 64 bits. 4668 Node* tail_ctl = NULL; 4669 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) { 4670 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) ); 4671 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) ); 4672 tail_ctl = generate_slow_guard(bol_lt, NULL); 4673 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4674 } 4675 4676 // At this point, let's assume there is no tail. 4677 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4678 // There is no tail. Try an upgrade to a 64-bit copy. 4679 bool didit = false; 4680 { PreserveJVMState pjvms(this); 4681 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4682 src, src_offset, dest, dest_offset, 4683 dest_size); 4684 if (didit) { 4685 // Present the results of the block-copying fast call. 4686 result_region->init_req(bcopy_path, control()); 4687 result_i_o ->init_req(bcopy_path, i_o()); 4688 result_memory->init_req(bcopy_path, memory(adr_type)); 4689 } 4690 } 4691 if (didit) 4692 set_control(top()); // no regular fast path 4693 } 4694 4695 // Clear the tail, if any. 4696 if (tail_ctl != NULL) { 4697 Node* notail_ctl = stopped() ? NULL : control(); 4698 set_control(tail_ctl); 4699 if (notail_ctl == NULL) { 4700 generate_clear_array(adr_type, dest, basic_elem_type, 4701 dest_tail, NULL, 4702 dest_size); 4703 } else { 4704 // Make a local merge. 4705 Node* done_ctl = new(C,3) RegionNode(3); 4706 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type); 4707 done_ctl->init_req(1, notail_ctl); 4708 done_mem->init_req(1, memory(adr_type)); 4709 generate_clear_array(adr_type, dest, basic_elem_type, 4710 dest_tail, NULL, 4711 dest_size); 4712 done_ctl->init_req(2, control()); 4713 done_mem->init_req(2, memory(adr_type)); 4714 set_control( _gvn.transform(done_ctl) ); 4715 set_memory( _gvn.transform(done_mem), adr_type ); 4716 } 4717 } 4718 } 4719 4720 BasicType copy_type = basic_elem_type; 4721 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 4722 if (!stopped() && copy_type == T_OBJECT) { 4723 // If src and dest have compatible element types, we can copy bits. 4724 // Types S[] and D[] are compatible if D is a supertype of S. 4725 // 4726 // If they are not, we will use checked_oop_disjoint_arraycopy, 4727 // which performs a fast optimistic per-oop check, and backs off 4728 // further to JVM_ArrayCopy on the first per-oop check that fails. 4729 // (Actually, we don't move raw bits only; the GC requires card marks.) 4730 4731 // Get the klassOop for both src and dest 4732 Node* src_klass = load_object_klass(src); 4733 Node* dest_klass = load_object_klass(dest); 4734 4735 // Generate the subtype check. 4736 // This might fold up statically, or then again it might not. 4737 // 4738 // Non-static example: Copying List<String>.elements to a new String[]. 4739 // The backing store for a List<String> is always an Object[], 4740 // but its elements are always type String, if the generic types 4741 // are correct at the source level. 4742 // 4743 // Test S[] against D[], not S against D, because (probably) 4744 // the secondary supertype cache is less busy for S[] than S. 4745 // This usually only matters when D is an interface. 4746 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4747 // Plug failing path into checked_oop_disjoint_arraycopy 4748 if (not_subtype_ctrl != top()) { 4749 PreserveJVMState pjvms(this); 4750 set_control(not_subtype_ctrl); 4751 // (At this point we can assume disjoint_bases, since types differ.) 4752 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc); 4753 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 4754 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 4755 Node* dest_elem_klass = _gvn.transform(n1); 4756 Node* cv = generate_checkcast_arraycopy(adr_type, 4757 dest_elem_klass, 4758 src, src_offset, dest, dest_offset, 4759 copy_length); 4760 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4761 checked_control = control(); 4762 checked_i_o = i_o(); 4763 checked_mem = memory(adr_type); 4764 checked_value = cv; 4765 } 4766 // At this point we know we do not need type checks on oop stores. 4767 4768 // Let's see if we need card marks: 4769 if (alloc != NULL && use_ReduceInitialCardMarks()) { 4770 // If we do not need card marks, copy using the jint or jlong stub. 4771 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 4772 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 4773 "sizes agree"); 4774 } 4775 } 4776 4777 if (!stopped()) { 4778 // Generate the fast path, if possible. 4779 PreserveJVMState pjvms(this); 4780 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 4781 src, src_offset, dest, dest_offset, 4782 ConvI2X(copy_length)); 4783 4784 // Present the results of the fast call. 4785 result_region->init_req(fast_path, control()); 4786 result_i_o ->init_req(fast_path, i_o()); 4787 result_memory->init_req(fast_path, memory(adr_type)); 4788 } 4789 4790 // Here are all the slow paths up to this point, in one bundle: 4791 slow_control = top(); 4792 if (slow_region != NULL) 4793 slow_control = _gvn.transform(slow_region); 4794 debug_only(slow_region = (RegionNode*)badAddress); 4795 4796 set_control(checked_control); 4797 if (!stopped()) { 4798 // Clean up after the checked call. 4799 // The returned value is either 0 or -1^K, 4800 // where K = number of partially transferred array elements. 4801 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) ); 4802 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 4803 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4804 4805 // If it is 0, we are done, so transfer to the end. 4806 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) ); 4807 result_region->init_req(checked_path, checks_done); 4808 result_i_o ->init_req(checked_path, checked_i_o); 4809 result_memory->init_req(checked_path, checked_mem); 4810 4811 // If it is not zero, merge into the slow call. 4812 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) )); 4813 RegionNode* slow_reg2 = new(C, 3) RegionNode(3); 4814 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO); 4815 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type); 4816 record_for_igvn(slow_reg2); 4817 slow_reg2 ->init_req(1, slow_control); 4818 slow_i_o2 ->init_req(1, slow_i_o); 4819 slow_mem2 ->init_req(1, slow_mem); 4820 slow_reg2 ->init_req(2, control()); 4821 slow_i_o2 ->init_req(2, checked_i_o); 4822 slow_mem2 ->init_req(2, checked_mem); 4823 4824 slow_control = _gvn.transform(slow_reg2); 4825 slow_i_o = _gvn.transform(slow_i_o2); 4826 slow_mem = _gvn.transform(slow_mem2); 4827 4828 if (alloc != NULL) { 4829 // We'll restart from the very beginning, after zeroing the whole thing. 4830 // This can cause double writes, but that's OK since dest is brand new. 4831 // So we ignore the low 31 bits of the value returned from the stub. 4832 } else { 4833 // We must continue the copy exactly where it failed, or else 4834 // another thread might see the wrong number of writes to dest. 4835 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) ); 4836 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT); 4837 slow_offset->init_req(1, intcon(0)); 4838 slow_offset->init_req(2, checked_offset); 4839 slow_offset = _gvn.transform(slow_offset); 4840 4841 // Adjust the arguments by the conditionally incoming offset. 4842 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) ); 4843 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) ); 4844 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) ); 4845 4846 // Tweak the node variables to adjust the code produced below: 4847 src_offset = src_off_plus; 4848 dest_offset = dest_off_plus; 4849 copy_length = length_minus; 4850 } 4851 } 4852 4853 set_control(slow_control); 4854 if (!stopped()) { 4855 // Generate the slow path, if needed. 4856 PreserveJVMState pjvms(this); // replace_in_map may trash the map 4857 4858 set_memory(slow_mem, adr_type); 4859 set_i_o(slow_i_o); 4860 4861 if (must_clear_dest) { 4862 generate_clear_array(adr_type, dest, basic_elem_type, 4863 intcon(0), NULL, 4864 alloc->in(AllocateNode::AllocSize)); 4865 } 4866 4867 generate_slow_arraycopy(adr_type, 4868 src, src_offset, dest, dest_offset, 4869 copy_length); 4870 4871 result_region->init_req(slow_call_path, control()); 4872 result_i_o ->init_req(slow_call_path, i_o()); 4873 result_memory->init_req(slow_call_path, memory(adr_type)); 4874 } 4875 4876 // Remove unused edges. 4877 for (uint i = 1; i < result_region->req(); i++) { 4878 if (result_region->in(i) == NULL) 4879 result_region->init_req(i, top()); 4880 } 4881 4882 // Finished; return the combined state. 4883 set_control( _gvn.transform(result_region) ); 4884 set_i_o( _gvn.transform(result_i_o) ); 4885 set_memory( _gvn.transform(result_memory), adr_type ); 4886 4887 // The memory edges above are precise in order to model effects around 4888 // array copies accurately to allow value numbering of field loads around 4889 // arraycopy. Such field loads, both before and after, are common in Java 4890 // collections and similar classes involving header/array data structures. 4891 // 4892 // But with low number of register or when some registers are used or killed 4893 // by arraycopy calls it causes registers spilling on stack. See 6544710. 4894 // The next memory barrier is added to avoid it. If the arraycopy can be 4895 // optimized away (which it can, sometimes) then we can manually remove 4896 // the membar also. 4897 // 4898 // Do not let reads from the cloned object float above the arraycopy. 4899 if (InsertMemBarAfterArraycopy || alloc != NULL) 4900 insert_mem_bar(Op_MemBarCPUOrder); 4901 } 4902 4903 4904 // Helper function which determines if an arraycopy immediately follows 4905 // an allocation, with no intervening tests or other escapes for the object. 4906 AllocateArrayNode* 4907 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4908 RegionNode* slow_region) { 4909 if (stopped()) return NULL; // no fast path 4910 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4911 4912 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4913 if (alloc == NULL) return NULL; 4914 4915 Node* rawmem = memory(Compile::AliasIdxRaw); 4916 // Is the allocation's memory state untouched? 4917 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4918 // Bail out if there have been raw-memory effects since the allocation. 4919 // (Example: There might have been a call or safepoint.) 4920 return NULL; 4921 } 4922 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4923 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4924 return NULL; 4925 } 4926 4927 // There must be no unexpected observers of this allocation. 4928 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4929 Node* obs = ptr->fast_out(i); 4930 if (obs != this->map()) { 4931 return NULL; 4932 } 4933 } 4934 4935 // This arraycopy must unconditionally follow the allocation of the ptr. 4936 Node* alloc_ctl = ptr->in(0); 4937 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4938 4939 Node* ctl = control(); 4940 while (ctl != alloc_ctl) { 4941 // There may be guards which feed into the slow_region. 4942 // Any other control flow means that we might not get a chance 4943 // to finish initializing the allocated object. 4944 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4945 IfNode* iff = ctl->in(0)->as_If(); 4946 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 4947 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4948 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4949 ctl = iff->in(0); // This test feeds the known slow_region. 4950 continue; 4951 } 4952 // One more try: Various low-level checks bottom out in 4953 // uncommon traps. If the debug-info of the trap omits 4954 // any reference to the allocation, as we've already 4955 // observed, then there can be no objection to the trap. 4956 bool found_trap = false; 4957 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 4958 Node* obs = not_ctl->fast_out(j); 4959 if (obs->in(0) == not_ctl && obs->is_Call() && 4960 (obs->as_Call()->entry_point() == 4961 SharedRuntime::uncommon_trap_blob()->instructions_begin())) { 4962 found_trap = true; break; 4963 } 4964 } 4965 if (found_trap) { 4966 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 4967 continue; 4968 } 4969 } 4970 return NULL; 4971 } 4972 4973 // If we get this far, we have an allocation which immediately 4974 // precedes the arraycopy, and we can take over zeroing the new object. 4975 // The arraycopy will finish the initialization, and provide 4976 // a new control state to which we will anchor the destination pointer. 4977 4978 return alloc; 4979 } 4980 4981 // Helper for initialization of arrays, creating a ClearArray. 4982 // It writes zero bits in [start..end), within the body of an array object. 4983 // The memory effects are all chained onto the 'adr_type' alias category. 4984 // 4985 // Since the object is otherwise uninitialized, we are free 4986 // to put a little "slop" around the edges of the cleared area, 4987 // as long as it does not go back into the array's header, 4988 // or beyond the array end within the heap. 4989 // 4990 // The lower edge can be rounded down to the nearest jint and the 4991 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 4992 // 4993 // Arguments: 4994 // adr_type memory slice where writes are generated 4995 // dest oop of the destination array 4996 // basic_elem_type element type of the destination 4997 // slice_idx array index of first element to store 4998 // slice_len number of elements to store (or NULL) 4999 // dest_size total size in bytes of the array object 5000 // 5001 // Exactly one of slice_len or dest_size must be non-NULL. 5002 // If dest_size is non-NULL, zeroing extends to the end of the object. 5003 // If slice_len is non-NULL, the slice_idx value must be a constant. 5004 void 5005 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5006 Node* dest, 5007 BasicType basic_elem_type, 5008 Node* slice_idx, 5009 Node* slice_len, 5010 Node* dest_size) { 5011 // one or the other but not both of slice_len and dest_size: 5012 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5013 if (slice_len == NULL) slice_len = top(); 5014 if (dest_size == NULL) dest_size = top(); 5015 5016 // operate on this memory slice: 5017 Node* mem = memory(adr_type); // memory slice to operate on 5018 5019 // scaling and rounding of indexes: 5020 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5021 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5022 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5023 int bump_bit = (-1 << scale) & BytesPerInt; 5024 5025 // determine constant starts and ends 5026 const intptr_t BIG_NEG = -128; 5027 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5028 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5029 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5030 if (slice_len_con == 0) { 5031 return; // nothing to do here 5032 } 5033 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5034 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5035 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5036 assert(end_con < 0, "not two cons"); 5037 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5038 BytesPerLong); 5039 } 5040 5041 if (start_con >= 0 && end_con >= 0) { 5042 // Constant start and end. Simple. 5043 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5044 start_con, end_con, &_gvn); 5045 } else if (start_con >= 0 && dest_size != top()) { 5046 // Constant start, pre-rounded end after the tail of the array. 5047 Node* end = dest_size; 5048 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5049 start_con, end, &_gvn); 5050 } else if (start_con >= 0 && slice_len != top()) { 5051 // Constant start, non-constant end. End needs rounding up. 5052 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5053 intptr_t end_base = abase + (slice_idx_con << scale); 5054 int end_round = (-1 << scale) & (BytesPerLong - 1); 5055 Node* end = ConvI2X(slice_len); 5056 if (scale != 0) 5057 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) )); 5058 end_base += end_round; 5059 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) ); 5060 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) ); 5061 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5062 start_con, end, &_gvn); 5063 } else if (start_con < 0 && dest_size != top()) { 5064 // Non-constant start, pre-rounded end after the tail of the array. 5065 // This is almost certainly a "round-to-end" operation. 5066 Node* start = slice_idx; 5067 start = ConvI2X(start); 5068 if (scale != 0) 5069 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) )); 5070 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) ); 5071 if ((bump_bit | clear_low) != 0) { 5072 int to_clear = (bump_bit | clear_low); 5073 // Align up mod 8, then store a jint zero unconditionally 5074 // just before the mod-8 boundary. 5075 if (((abase + bump_bit) & ~to_clear) - bump_bit 5076 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5077 bump_bit = 0; 5078 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5079 } else { 5080 // Bump 'start' up to (or past) the next jint boundary: 5081 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) ); 5082 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5083 } 5084 // Round bumped 'start' down to jlong boundary in body of array. 5085 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) ); 5086 if (bump_bit != 0) { 5087 // Store a zero to the immediately preceding jint: 5088 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) ); 5089 Node* p1 = basic_plus_adr(dest, x1); 5090 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 5091 mem = _gvn.transform(mem); 5092 } 5093 } 5094 Node* end = dest_size; // pre-rounded 5095 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5096 start, end, &_gvn); 5097 } else { 5098 // Non-constant start, unrounded non-constant end. 5099 // (Nobody zeroes a random midsection of an array using this routine.) 5100 ShouldNotReachHere(); // fix caller 5101 } 5102 5103 // Done. 5104 set_memory(mem, adr_type); 5105 } 5106 5107 5108 bool 5109 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5110 BasicType basic_elem_type, 5111 AllocateNode* alloc, 5112 Node* src, Node* src_offset, 5113 Node* dest, Node* dest_offset, 5114 Node* dest_size) { 5115 // See if there is an advantage from block transfer. 5116 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5117 if (scale >= LogBytesPerLong) 5118 return false; // it is already a block transfer 5119 5120 // Look at the alignment of the starting offsets. 5121 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5122 const intptr_t BIG_NEG = -128; 5123 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5124 5125 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale); 5126 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale); 5127 if (src_off < 0 || dest_off < 0) 5128 // At present, we can only understand constants. 5129 return false; 5130 5131 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5132 // Non-aligned; too bad. 5133 // One more chance: Pick off an initial 32-bit word. 5134 // This is a common case, since abase can be odd mod 8. 5135 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5136 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5137 Node* sptr = basic_plus_adr(src, src_off); 5138 Node* dptr = basic_plus_adr(dest, dest_off); 5139 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 5140 store_to_memory(control(), dptr, sval, T_INT, adr_type); 5141 src_off += BytesPerInt; 5142 dest_off += BytesPerInt; 5143 } else { 5144 return false; 5145 } 5146 } 5147 assert(src_off % BytesPerLong == 0, ""); 5148 assert(dest_off % BytesPerLong == 0, ""); 5149 5150 // Do this copy by giant steps. 5151 Node* sptr = basic_plus_adr(src, src_off); 5152 Node* dptr = basic_plus_adr(dest, dest_off); 5153 Node* countx = dest_size; 5154 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) ); 5155 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) ); 5156 5157 bool disjoint_bases = true; // since alloc != NULL 5158 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5159 sptr, NULL, dptr, NULL, countx); 5160 5161 return true; 5162 } 5163 5164 5165 // Helper function; generates code for the slow case. 5166 // We make a call to a runtime method which emulates the native method, 5167 // but without the native wrapper overhead. 5168 void 5169 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5170 Node* src, Node* src_offset, 5171 Node* dest, Node* dest_offset, 5172 Node* copy_length) { 5173 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5174 OptoRuntime::slow_arraycopy_Type(), 5175 OptoRuntime::slow_arraycopy_Java(), 5176 "slow_arraycopy", adr_type, 5177 src, src_offset, dest, dest_offset, 5178 copy_length); 5179 5180 // Handle exceptions thrown by this fellow: 5181 make_slow_call_ex(call, env()->Throwable_klass(), false); 5182 } 5183 5184 // Helper function; generates code for cases requiring runtime checks. 5185 Node* 5186 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5187 Node* dest_elem_klass, 5188 Node* src, Node* src_offset, 5189 Node* dest, Node* dest_offset, 5190 Node* copy_length) { 5191 if (stopped()) return NULL; 5192 5193 address copyfunc_addr = StubRoutines::checkcast_arraycopy(); 5194 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5195 return NULL; 5196 } 5197 5198 // Pick out the parameters required to perform a store-check 5199 // for the target array. This is an optimistic check. It will 5200 // look in each non-null element's class, at the desired klass's 5201 // super_check_offset, for the desired klass. 5202 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc); 5203 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5204 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM); 5205 Node* check_offset = _gvn.transform(n3); 5206 Node* check_value = dest_elem_klass; 5207 5208 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5209 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5210 5211 // (We know the arrays are never conjoint, because their types differ.) 5212 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5213 OptoRuntime::checkcast_arraycopy_Type(), 5214 copyfunc_addr, "checkcast_arraycopy", adr_type, 5215 // five arguments, of which two are 5216 // intptr_t (jlong in LP64) 5217 src_start, dest_start, 5218 copy_length XTOP, 5219 check_offset XTOP, 5220 check_value); 5221 5222 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5223 } 5224 5225 5226 // Helper function; generates code for cases requiring runtime checks. 5227 Node* 5228 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5229 Node* src, Node* src_offset, 5230 Node* dest, Node* dest_offset, 5231 Node* copy_length) { 5232 if (stopped()) return NULL; 5233 5234 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5235 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5236 return NULL; 5237 } 5238 5239 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5240 OptoRuntime::generic_arraycopy_Type(), 5241 copyfunc_addr, "generic_arraycopy", adr_type, 5242 src, src_offset, dest, dest_offset, copy_length); 5243 5244 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5245 } 5246 5247 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5248 void 5249 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5250 BasicType basic_elem_type, 5251 bool disjoint_bases, 5252 Node* src, Node* src_offset, 5253 Node* dest, Node* dest_offset, 5254 Node* copy_length) { 5255 if (stopped()) return; // nothing to do 5256 5257 Node* src_start = src; 5258 Node* dest_start = dest; 5259 if (src_offset != NULL || dest_offset != NULL) { 5260 assert(src_offset != NULL && dest_offset != NULL, ""); 5261 src_start = array_element_address(src, src_offset, basic_elem_type); 5262 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5263 } 5264 5265 // Figure out which arraycopy runtime method to call. 5266 const char* copyfunc_name = "arraycopy"; 5267 address copyfunc_addr = 5268 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5269 disjoint_bases, copyfunc_name); 5270 5271 // Call it. Note that the count_ix value is not scaled to a byte-size. 5272 make_runtime_call(RC_LEAF|RC_NO_FP, 5273 OptoRuntime::fast_arraycopy_Type(), 5274 copyfunc_addr, copyfunc_name, adr_type, 5275 src_start, dest_start, copy_length XTOP); 5276 }