1 /* 2 * Copyright 1999-2009 Sun Microsystems, Inc. All Rights Reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_library_call.cpp.incl" 27 28 class LibraryIntrinsic : public InlineCallGenerator { 29 // Extend the set of intrinsics known to the runtime: 30 public: 31 private: 32 bool _is_virtual; 33 vmIntrinsics::ID _intrinsic_id; 34 35 public: 36 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id) 37 : InlineCallGenerator(m), 38 _is_virtual(is_virtual), 39 _intrinsic_id(id) 40 { 41 } 42 virtual bool is_intrinsic() const { return true; } 43 virtual bool is_virtual() const { return _is_virtual; } 44 virtual JVMState* generate(JVMState* jvms); 45 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 46 }; 47 48 49 // Local helper class for LibraryIntrinsic: 50 class LibraryCallKit : public GraphKit { 51 private: 52 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 53 54 public: 55 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic) 56 : GraphKit(caller), 57 _intrinsic(intrinsic) 58 { 59 } 60 61 ciMethod* caller() const { return jvms()->method(); } 62 int bci() const { return jvms()->bci(); } 63 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 65 ciMethod* callee() const { return _intrinsic->method(); } 66 ciSignature* signature() const { return callee()->signature(); } 67 int arg_size() const { return callee()->arg_size(); } 68 69 bool try_to_inline(); 70 71 // Helper functions to inline natives 72 void push_result(RegionNode* region, PhiNode* value); 73 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 74 Node* generate_slow_guard(Node* test, RegionNode* region); 75 Node* generate_fair_guard(Node* test, RegionNode* region); 76 Node* generate_negative_guard(Node* index, RegionNode* region, 77 // resulting CastII of index: 78 Node* *pos_index = NULL); 79 Node* generate_nonpositive_guard(Node* index, bool never_negative, 80 // resulting CastII of index: 81 Node* *pos_index = NULL); 82 Node* generate_limit_guard(Node* offset, Node* subseq_length, 83 Node* array_length, 84 RegionNode* region); 85 Node* generate_current_thread(Node* &tls_output); 86 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 87 bool disjoint_bases, const char* &name); 88 Node* load_mirror_from_klass(Node* klass); 89 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 90 int nargs, 91 RegionNode* region, int null_path, 92 int offset); 93 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs, 94 RegionNode* region, int null_path) { 95 int offset = java_lang_Class::klass_offset_in_bytes(); 96 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 97 region, null_path, 98 offset); 99 } 100 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 101 int nargs, 102 RegionNode* region, int null_path) { 103 int offset = java_lang_Class::array_klass_offset_in_bytes(); 104 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 105 region, null_path, 106 offset); 107 } 108 Node* generate_access_flags_guard(Node* kls, 109 int modifier_mask, int modifier_bits, 110 RegionNode* region); 111 Node* generate_interface_guard(Node* kls, RegionNode* region); 112 Node* generate_array_guard(Node* kls, RegionNode* region) { 113 return generate_array_guard_common(kls, region, false, false); 114 } 115 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 116 return generate_array_guard_common(kls, region, false, true); 117 } 118 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 119 return generate_array_guard_common(kls, region, true, false); 120 } 121 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 122 return generate_array_guard_common(kls, region, true, true); 123 } 124 Node* generate_array_guard_common(Node* kls, RegionNode* region, 125 bool obj_array, bool not_array); 126 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 127 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 128 bool is_virtual = false, bool is_static = false); 129 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 130 return generate_method_call(method_id, false, true); 131 } 132 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 133 return generate_method_call(method_id, true, false); 134 } 135 136 bool inline_string_compareTo(); 137 bool inline_string_indexOf(); 138 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 139 bool inline_string_equals(); 140 Node* pop_math_arg(); 141 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 142 bool inline_math_native(vmIntrinsics::ID id); 143 bool inline_trig(vmIntrinsics::ID id); 144 bool inline_trans(vmIntrinsics::ID id); 145 bool inline_abs(vmIntrinsics::ID id); 146 bool inline_sqrt(vmIntrinsics::ID id); 147 bool inline_pow(vmIntrinsics::ID id); 148 bool inline_exp(vmIntrinsics::ID id); 149 bool inline_min_max(vmIntrinsics::ID id); 150 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 151 // This returns Type::AnyPtr, RawPtr, or OopPtr. 152 int classify_unsafe_addr(Node* &base, Node* &offset); 153 Node* make_unsafe_address(Node* base, Node* offset); 154 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 155 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 156 bool inline_unsafe_allocate(); 157 bool inline_unsafe_copyMemory(); 158 bool inline_native_currentThread(); 159 bool inline_native_time_funcs(bool isNano); 160 bool inline_native_isInterrupted(); 161 bool inline_native_Class_query(vmIntrinsics::ID id); 162 bool inline_native_subtype_check(); 163 164 bool inline_native_newArray(); 165 bool inline_native_getLength(); 166 bool inline_array_copyOf(bool is_copyOfRange); 167 bool inline_array_equals(); 168 bool inline_native_clone(bool is_virtual); 169 bool inline_native_Reflection_getCallerClass(); 170 bool inline_native_AtomicLong_get(); 171 bool inline_native_AtomicLong_attemptUpdate(); 172 bool is_method_invoke_or_aux_frame(JVMState* jvms); 173 // Helper function for inlining native object hash method 174 bool inline_native_hashcode(bool is_virtual, bool is_static); 175 bool inline_native_getClass(); 176 177 // Helper functions for inlining arraycopy 178 bool inline_arraycopy(); 179 void generate_arraycopy(const TypePtr* adr_type, 180 BasicType basic_elem_type, 181 Node* src, Node* src_offset, 182 Node* dest, Node* dest_offset, 183 Node* copy_length, 184 int nargs, // arguments on stack for debug info 185 bool disjoint_bases = false, 186 bool length_never_negative = false, 187 RegionNode* slow_region = NULL); 188 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 189 RegionNode* slow_region); 190 void generate_clear_array(const TypePtr* adr_type, 191 Node* dest, 192 BasicType basic_elem_type, 193 Node* slice_off, 194 Node* slice_len, 195 Node* slice_end); 196 bool generate_block_arraycopy(const TypePtr* adr_type, 197 BasicType basic_elem_type, 198 AllocateNode* alloc, 199 Node* src, Node* src_offset, 200 Node* dest, Node* dest_offset, 201 Node* dest_size); 202 void generate_slow_arraycopy(const TypePtr* adr_type, 203 Node* src, Node* src_offset, 204 Node* dest, Node* dest_offset, 205 Node* copy_length, 206 int nargs); 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, int nargs); 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, int nargs); 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_bitCount(vmIntrinsics::ID id); 226 bool inline_reverseBytes(vmIntrinsics::ID id); 227 }; 228 229 230 //---------------------------make_vm_intrinsic---------------------------- 231 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 232 vmIntrinsics::ID id = m->intrinsic_id(); 233 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 234 235 if (DisableIntrinsic[0] != '\0' 236 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 237 // disabled by a user request on the command line: 238 // example: -XX:DisableIntrinsic=_hashCode,_getClass 239 return NULL; 240 } 241 242 if (!m->is_loaded()) { 243 // do not attempt to inline unloaded methods 244 return NULL; 245 } 246 247 // Only a few intrinsics implement a virtual dispatch. 248 // They are expensive calls which are also frequently overridden. 249 if (is_virtual) { 250 switch (id) { 251 case vmIntrinsics::_hashCode: 252 case vmIntrinsics::_clone: 253 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 254 break; 255 default: 256 return NULL; 257 } 258 } 259 260 // -XX:-InlineNatives disables nearly all intrinsics: 261 if (!InlineNatives) { 262 switch (id) { 263 case vmIntrinsics::_indexOf: 264 case vmIntrinsics::_compareTo: 265 case vmIntrinsics::_equals: 266 case vmIntrinsics::_equalsC: 267 break; // InlineNatives does not control String.compareTo 268 default: 269 return NULL; 270 } 271 } 272 273 switch (id) { 274 case vmIntrinsics::_compareTo: 275 if (!SpecialStringCompareTo) return NULL; 276 break; 277 case vmIntrinsics::_indexOf: 278 if (!SpecialStringIndexOf) return NULL; 279 break; 280 case vmIntrinsics::_equals: 281 if (!SpecialStringEquals) return NULL; 282 break; 283 case vmIntrinsics::_equalsC: 284 if (!SpecialArraysEquals) return NULL; 285 break; 286 case vmIntrinsics::_arraycopy: 287 if (!InlineArrayCopy) return NULL; 288 break; 289 case vmIntrinsics::_copyMemory: 290 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 291 if (!InlineArrayCopy) return NULL; 292 break; 293 case vmIntrinsics::_hashCode: 294 if (!InlineObjectHash) return NULL; 295 break; 296 case vmIntrinsics::_clone: 297 case vmIntrinsics::_copyOf: 298 case vmIntrinsics::_copyOfRange: 299 if (!InlineObjectCopy) return NULL; 300 // These also use the arraycopy intrinsic mechanism: 301 if (!InlineArrayCopy) return NULL; 302 break; 303 case vmIntrinsics::_checkIndex: 304 // We do not intrinsify this. The optimizer does fine with it. 305 return NULL; 306 307 case vmIntrinsics::_get_AtomicLong: 308 case vmIntrinsics::_attemptUpdate: 309 if (!InlineAtomicLong) return NULL; 310 break; 311 312 case vmIntrinsics::_Object_init: 313 case vmIntrinsics::_invoke: 314 // We do not intrinsify these; they are marked for other purposes. 315 return NULL; 316 317 case vmIntrinsics::_getCallerClass: 318 if (!UseNewReflection) return NULL; 319 if (!InlineReflectionGetCallerClass) return NULL; 320 if (!JDK_Version::is_gte_jdk14x_version()) return NULL; 321 break; 322 323 case vmIntrinsics::_bitCount_i: 324 case vmIntrinsics::_bitCount_l: 325 if (!UsePopCountInstruction) return NULL; 326 break; 327 328 default: 329 break; 330 } 331 332 // -XX:-InlineClassNatives disables natives from the Class class. 333 // The flag applies to all reflective calls, notably Array.newArray 334 // (visible to Java programmers as Array.newInstance). 335 if (m->holder()->name() == ciSymbol::java_lang_Class() || 336 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 337 if (!InlineClassNatives) return NULL; 338 } 339 340 // -XX:-InlineThreadNatives disables natives from the Thread class. 341 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 342 if (!InlineThreadNatives) return NULL; 343 } 344 345 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 346 if (m->holder()->name() == ciSymbol::java_lang_Math() || 347 m->holder()->name() == ciSymbol::java_lang_Float() || 348 m->holder()->name() == ciSymbol::java_lang_Double()) { 349 if (!InlineMathNatives) return NULL; 350 } 351 352 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 353 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 354 if (!InlineUnsafeOps) return NULL; 355 } 356 357 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); 358 } 359 360 //----------------------register_library_intrinsics----------------------- 361 // Initialize this file's data structures, for each Compile instance. 362 void Compile::register_library_intrinsics() { 363 // Nothing to do here. 364 } 365 366 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 367 LibraryCallKit kit(jvms, this); 368 Compile* C = kit.C; 369 int nodes = C->unique(); 370 #ifndef PRODUCT 371 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 372 char buf[1000]; 373 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 374 tty->print_cr("Intrinsic %s", str); 375 } 376 #endif 377 if (kit.try_to_inline()) { 378 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 379 tty->print("Inlining intrinsic %s%s at bci:%d in", 380 vmIntrinsics::name_at(intrinsic_id()), 381 (is_virtual() ? " (virtual)" : ""), kit.bci()); 382 kit.caller()->print_short_name(tty); 383 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); 384 } 385 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 386 if (C->log()) { 387 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 388 vmIntrinsics::name_at(intrinsic_id()), 389 (is_virtual() ? " virtual='1'" : ""), 390 C->unique() - nodes); 391 } 392 return kit.transfer_exceptions_into_jvms(); 393 } 394 395 if (PrintIntrinsics) { 396 switch (intrinsic_id()) { 397 case vmIntrinsics::_invoke: 398 case vmIntrinsics::_Object_init: 399 // We do not expect to inline these, so do not produce any noise about them. 400 break; 401 default: 402 tty->print("Did not inline intrinsic %s%s at bci:%d in", 403 vmIntrinsics::name_at(intrinsic_id()), 404 (is_virtual() ? " (virtual)" : ""), kit.bci()); 405 kit.caller()->print_short_name(tty); 406 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); 407 } 408 } 409 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 410 return NULL; 411 } 412 413 bool LibraryCallKit::try_to_inline() { 414 // Handle symbolic names for otherwise undistinguished boolean switches: 415 const bool is_store = true; 416 const bool is_native_ptr = true; 417 const bool is_static = true; 418 419 switch (intrinsic_id()) { 420 case vmIntrinsics::_hashCode: 421 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 422 case vmIntrinsics::_identityHashCode: 423 return inline_native_hashcode(/*!virtual*/ false, is_static); 424 case vmIntrinsics::_getClass: 425 return inline_native_getClass(); 426 427 case vmIntrinsics::_dsin: 428 case vmIntrinsics::_dcos: 429 case vmIntrinsics::_dtan: 430 case vmIntrinsics::_dabs: 431 case vmIntrinsics::_datan2: 432 case vmIntrinsics::_dsqrt: 433 case vmIntrinsics::_dexp: 434 case vmIntrinsics::_dlog: 435 case vmIntrinsics::_dlog10: 436 case vmIntrinsics::_dpow: 437 return inline_math_native(intrinsic_id()); 438 439 case vmIntrinsics::_min: 440 case vmIntrinsics::_max: 441 return inline_min_max(intrinsic_id()); 442 443 case vmIntrinsics::_arraycopy: 444 return inline_arraycopy(); 445 446 case vmIntrinsics::_compareTo: 447 return inline_string_compareTo(); 448 case vmIntrinsics::_indexOf: 449 return inline_string_indexOf(); 450 case vmIntrinsics::_equals: 451 return inline_string_equals(); 452 453 case vmIntrinsics::_getObject: 454 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); 455 case vmIntrinsics::_getBoolean: 456 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); 457 case vmIntrinsics::_getByte: 458 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); 459 case vmIntrinsics::_getShort: 460 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); 461 case vmIntrinsics::_getChar: 462 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); 463 case vmIntrinsics::_getInt: 464 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); 465 case vmIntrinsics::_getLong: 466 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); 467 case vmIntrinsics::_getFloat: 468 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); 469 case vmIntrinsics::_getDouble: 470 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); 471 472 case vmIntrinsics::_putObject: 473 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); 474 case vmIntrinsics::_putBoolean: 475 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); 476 case vmIntrinsics::_putByte: 477 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); 478 case vmIntrinsics::_putShort: 479 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); 480 case vmIntrinsics::_putChar: 481 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); 482 case vmIntrinsics::_putInt: 483 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); 484 case vmIntrinsics::_putLong: 485 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); 486 case vmIntrinsics::_putFloat: 487 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); 488 case vmIntrinsics::_putDouble: 489 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); 490 491 case vmIntrinsics::_getByte_raw: 492 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); 493 case vmIntrinsics::_getShort_raw: 494 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); 495 case vmIntrinsics::_getChar_raw: 496 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); 497 case vmIntrinsics::_getInt_raw: 498 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); 499 case vmIntrinsics::_getLong_raw: 500 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); 501 case vmIntrinsics::_getFloat_raw: 502 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); 503 case vmIntrinsics::_getDouble_raw: 504 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); 505 case vmIntrinsics::_getAddress_raw: 506 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); 507 508 case vmIntrinsics::_putByte_raw: 509 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); 510 case vmIntrinsics::_putShort_raw: 511 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); 512 case vmIntrinsics::_putChar_raw: 513 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); 514 case vmIntrinsics::_putInt_raw: 515 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); 516 case vmIntrinsics::_putLong_raw: 517 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); 518 case vmIntrinsics::_putFloat_raw: 519 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); 520 case vmIntrinsics::_putDouble_raw: 521 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); 522 case vmIntrinsics::_putAddress_raw: 523 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); 524 525 case vmIntrinsics::_getObjectVolatile: 526 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); 527 case vmIntrinsics::_getBooleanVolatile: 528 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); 529 case vmIntrinsics::_getByteVolatile: 530 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); 531 case vmIntrinsics::_getShortVolatile: 532 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); 533 case vmIntrinsics::_getCharVolatile: 534 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); 535 case vmIntrinsics::_getIntVolatile: 536 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); 537 case vmIntrinsics::_getLongVolatile: 538 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); 539 case vmIntrinsics::_getFloatVolatile: 540 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); 541 case vmIntrinsics::_getDoubleVolatile: 542 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); 543 544 case vmIntrinsics::_putObjectVolatile: 545 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); 546 case vmIntrinsics::_putBooleanVolatile: 547 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); 548 case vmIntrinsics::_putByteVolatile: 549 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); 550 case vmIntrinsics::_putShortVolatile: 551 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); 552 case vmIntrinsics::_putCharVolatile: 553 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); 554 case vmIntrinsics::_putIntVolatile: 555 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); 556 case vmIntrinsics::_putLongVolatile: 557 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); 558 case vmIntrinsics::_putFloatVolatile: 559 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); 560 case vmIntrinsics::_putDoubleVolatile: 561 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); 562 563 case vmIntrinsics::_prefetchRead: 564 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 565 case vmIntrinsics::_prefetchWrite: 566 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 567 case vmIntrinsics::_prefetchReadStatic: 568 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 569 case vmIntrinsics::_prefetchWriteStatic: 570 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 571 572 case vmIntrinsics::_compareAndSwapObject: 573 return inline_unsafe_CAS(T_OBJECT); 574 case vmIntrinsics::_compareAndSwapInt: 575 return inline_unsafe_CAS(T_INT); 576 case vmIntrinsics::_compareAndSwapLong: 577 return inline_unsafe_CAS(T_LONG); 578 579 case vmIntrinsics::_putOrderedObject: 580 return inline_unsafe_ordered_store(T_OBJECT); 581 case vmIntrinsics::_putOrderedInt: 582 return inline_unsafe_ordered_store(T_INT); 583 case vmIntrinsics::_putOrderedLong: 584 return inline_unsafe_ordered_store(T_LONG); 585 586 case vmIntrinsics::_currentThread: 587 return inline_native_currentThread(); 588 case vmIntrinsics::_isInterrupted: 589 return inline_native_isInterrupted(); 590 591 case vmIntrinsics::_currentTimeMillis: 592 return inline_native_time_funcs(false); 593 case vmIntrinsics::_nanoTime: 594 return inline_native_time_funcs(true); 595 case vmIntrinsics::_allocateInstance: 596 return inline_unsafe_allocate(); 597 case vmIntrinsics::_copyMemory: 598 return inline_unsafe_copyMemory(); 599 case vmIntrinsics::_newArray: 600 return inline_native_newArray(); 601 case vmIntrinsics::_getLength: 602 return inline_native_getLength(); 603 case vmIntrinsics::_copyOf: 604 return inline_array_copyOf(false); 605 case vmIntrinsics::_copyOfRange: 606 return inline_array_copyOf(true); 607 case vmIntrinsics::_equalsC: 608 return inline_array_equals(); 609 case vmIntrinsics::_clone: 610 return inline_native_clone(intrinsic()->is_virtual()); 611 612 case vmIntrinsics::_isAssignableFrom: 613 return inline_native_subtype_check(); 614 615 case vmIntrinsics::_isInstance: 616 case vmIntrinsics::_getModifiers: 617 case vmIntrinsics::_isInterface: 618 case vmIntrinsics::_isArray: 619 case vmIntrinsics::_isPrimitive: 620 case vmIntrinsics::_getSuperclass: 621 case vmIntrinsics::_getComponentType: 622 case vmIntrinsics::_getClassAccessFlags: 623 return inline_native_Class_query(intrinsic_id()); 624 625 case vmIntrinsics::_floatToRawIntBits: 626 case vmIntrinsics::_floatToIntBits: 627 case vmIntrinsics::_intBitsToFloat: 628 case vmIntrinsics::_doubleToRawLongBits: 629 case vmIntrinsics::_doubleToLongBits: 630 case vmIntrinsics::_longBitsToDouble: 631 return inline_fp_conversions(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 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); 640 641 case vmIntrinsics::_get_AtomicLong: 642 return inline_native_AtomicLong_get(); 643 case vmIntrinsics::_attemptUpdate: 644 return inline_native_AtomicLong_attemptUpdate(); 645 646 case vmIntrinsics::_getCallerClass: 647 return inline_native_Reflection_getCallerClass(); 648 649 default: 650 // If you get here, it may be that someone has added a new intrinsic 651 // to the list in vmSymbols.hpp without implementing it here. 652 #ifndef PRODUCT 653 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 654 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 655 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 656 } 657 #endif 658 return false; 659 } 660 } 661 662 //------------------------------push_result------------------------------ 663 // Helper function for finishing intrinsics. 664 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { 665 record_for_igvn(region); 666 set_control(_gvn.transform(region)); 667 BasicType value_type = value->type()->basic_type(); 668 push_node(value_type, _gvn.transform(value)); 669 } 670 671 //------------------------------generate_guard--------------------------- 672 // Helper function for generating guarded fast-slow graph structures. 673 // The given 'test', if true, guards a slow path. If the test fails 674 // then a fast path can be taken. (We generally hope it fails.) 675 // In all cases, GraphKit::control() is updated to the fast path. 676 // The returned value represents the control for the slow path. 677 // The return value is never 'top'; it is either a valid control 678 // or NULL if it is obvious that the slow path can never be taken. 679 // Also, if region and the slow control are not NULL, the slow edge 680 // is appended to the region. 681 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 682 if (stopped()) { 683 // Already short circuited. 684 return NULL; 685 } 686 687 // Build an if node and its projections. 688 // If test is true we take the slow path, which we assume is uncommon. 689 if (_gvn.type(test) == TypeInt::ZERO) { 690 // The slow branch is never taken. No need to build this guard. 691 return NULL; 692 } 693 694 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 695 696 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); 697 if (if_slow == top()) { 698 // The slow branch is never taken. No need to build this guard. 699 return NULL; 700 } 701 702 if (region != NULL) 703 region->add_req(if_slow); 704 705 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); 706 set_control(if_fast); 707 708 return if_slow; 709 } 710 711 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 712 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 713 } 714 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 715 return generate_guard(test, region, PROB_FAIR); 716 } 717 718 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 719 Node* *pos_index) { 720 if (stopped()) 721 return NULL; // already stopped 722 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 723 return NULL; // index is already adequately typed 724 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 725 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 726 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 727 if (is_neg != NULL && pos_index != NULL) { 728 // Emulate effect of Parse::adjust_map_after_if. 729 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); 730 ccast->set_req(0, control()); 731 (*pos_index) = _gvn.transform(ccast); 732 } 733 return is_neg; 734 } 735 736 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 737 Node* *pos_index) { 738 if (stopped()) 739 return NULL; // already stopped 740 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 741 return NULL; // index is already adequately typed 742 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 743 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 744 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) ); 745 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 746 if (is_notp != NULL && pos_index != NULL) { 747 // Emulate effect of Parse::adjust_map_after_if. 748 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1); 749 ccast->set_req(0, control()); 750 (*pos_index) = _gvn.transform(ccast); 751 } 752 return is_notp; 753 } 754 755 // Make sure that 'position' is a valid limit index, in [0..length]. 756 // There are two equivalent plans for checking this: 757 // A. (offset + copyLength) unsigned<= arrayLength 758 // B. offset <= (arrayLength - copyLength) 759 // We require that all of the values above, except for the sum and 760 // difference, are already known to be non-negative. 761 // Plan A is robust in the face of overflow, if offset and copyLength 762 // are both hugely positive. 763 // 764 // Plan B is less direct and intuitive, but it does not overflow at 765 // all, since the difference of two non-negatives is always 766 // representable. Whenever Java methods must perform the equivalent 767 // check they generally use Plan B instead of Plan A. 768 // For the moment we use Plan A. 769 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 770 Node* subseq_length, 771 Node* array_length, 772 RegionNode* region) { 773 if (stopped()) 774 return NULL; // already stopped 775 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 776 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length)) 777 return NULL; // common case of whole-array copy 778 Node* last = subseq_length; 779 if (!zero_offset) // last += offset 780 last = _gvn.transform( new (C, 3) AddINode(last, offset)); 781 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); 782 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 783 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 784 return is_over; 785 } 786 787 788 //--------------------------generate_current_thread-------------------- 789 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 790 ciKlass* thread_klass = env()->Thread_klass(); 791 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 792 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); 793 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 794 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 795 tls_output = thread; 796 return threadObj; 797 } 798 799 800 //------------------------------inline_string_compareTo------------------------ 801 bool LibraryCallKit::inline_string_compareTo() { 802 803 if (!Matcher::has_match_rule(Op_StrComp)) return false; 804 805 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 _sp += 2; 810 Node *argument = pop(); // pop non-receiver first: it was pushed second 811 Node *receiver = pop(); 812 813 // Null check on self without removing any arguments. The argument 814 // null check technically happens in the wrong place, which can lead to 815 // invalid stack traces when string compare is inlined into a method 816 // which handles NullPointerExceptions. 817 _sp += 2; 818 receiver = do_null_check(receiver, T_OBJECT); 819 argument = do_null_check(argument, T_OBJECT); 820 _sp -= 2; 821 if (stopped()) { 822 return true; 823 } 824 825 ciInstanceKlass* klass = env()->String_klass(); 826 const TypeInstPtr* string_type = 827 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 828 829 Node* compare = 830 _gvn.transform(new (C, 7) StrCompNode( 831 control(), 832 memory(TypeAryPtr::CHARS), 833 memory(string_type->add_offset(value_offset)), 834 memory(string_type->add_offset(count_offset)), 835 memory(string_type->add_offset(offset_offset)), 836 receiver, 837 argument)); 838 push(compare); 839 return true; 840 } 841 842 //------------------------------inline_string_equals------------------------ 843 bool LibraryCallKit::inline_string_equals() { 844 845 if (!Matcher::has_match_rule(Op_StrEquals)) return false; 846 847 const int value_offset = java_lang_String::value_offset_in_bytes(); 848 const int count_offset = java_lang_String::count_offset_in_bytes(); 849 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 850 851 _sp += 2; 852 Node* argument = pop(); // pop non-receiver first: it was pushed second 853 Node* receiver = pop(); 854 855 // Null check on self without removing any arguments. The argument 856 // null check technically happens in the wrong place, which can lead to 857 // invalid stack traces when string compare is inlined into a method 858 // which handles NullPointerExceptions. 859 _sp += 2; 860 receiver = do_null_check(receiver, T_OBJECT); 861 //should not do null check for argument for String.equals(), because spec 862 //allows to specify NULL as argument. 863 _sp -= 2; 864 865 if (stopped()) { 866 return true; 867 } 868 869 // get String klass for instanceOf 870 ciInstanceKlass* klass = env()->String_klass(); 871 872 // two paths (plus control) merge 873 RegionNode* region = new (C, 3) RegionNode(3); 874 Node* phi = new (C, 3) PhiNode(region, TypeInt::BOOL); 875 876 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 877 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1))); 878 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq)); 879 880 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 881 882 Node* if_true = _gvn.transform(new (C, 1) IfTrueNode(iff)); 883 set_control(if_true); 884 885 const TypeInstPtr* string_type = 886 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 887 888 // instanceOf == true 889 Node* equals = 890 _gvn.transform(new (C, 7) StrEqualsNode( 891 control(), 892 memory(TypeAryPtr::CHARS), 893 memory(string_type->add_offset(value_offset)), 894 memory(string_type->add_offset(count_offset)), 895 memory(string_type->add_offset(offset_offset)), 896 receiver, 897 argument)); 898 899 phi->init_req(1, _gvn.transform(equals)); 900 region->init_req(1, if_true); 901 902 //instanceOf == false, fallthrough 903 Node* if_false = _gvn.transform(new (C, 1) IfFalseNode(iff)); 904 set_control(if_false); 905 906 phi->init_req(2, _gvn.transform(intcon(0))); 907 region->init_req(2, if_false); 908 909 // post merge 910 set_control(_gvn.transform(region)); 911 record_for_igvn(region); 912 913 push(_gvn.transform(phi)); 914 915 return true; 916 } 917 918 //------------------------------inline_array_equals---------------------------- 919 bool LibraryCallKit::inline_array_equals() { 920 921 if (!Matcher::has_match_rule(Op_AryEq)) return false; 922 923 _sp += 2; 924 Node *argument2 = pop(); 925 Node *argument1 = pop(); 926 927 Node* equals = 928 _gvn.transform(new (C, 3) AryEqNode(control(), 929 argument1, 930 argument2) 931 ); 932 push(equals); 933 return true; 934 } 935 936 // Java version of String.indexOf(constant string) 937 // class StringDecl { 938 // StringDecl(char[] ca) { 939 // offset = 0; 940 // count = ca.length; 941 // value = ca; 942 // } 943 // int offset; 944 // int count; 945 // char[] value; 946 // } 947 // 948 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 949 // int targetOffset, int cache_i, int md2) { 950 // int cache = cache_i; 951 // int sourceOffset = string_object.offset; 952 // int sourceCount = string_object.count; 953 // int targetCount = target_object.length; 954 // 955 // int targetCountLess1 = targetCount - 1; 956 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 957 // 958 // char[] source = string_object.value; 959 // char[] target = target_object; 960 // int lastChar = target[targetCountLess1]; 961 // 962 // outer_loop: 963 // for (int i = sourceOffset; i < sourceEnd; ) { 964 // int src = source[i + targetCountLess1]; 965 // if (src == lastChar) { 966 // // With random strings and a 4-character alphabet, 967 // // reverse matching at this point sets up 0.8% fewer 968 // // frames, but (paradoxically) makes 0.3% more probes. 969 // // Since those probes are nearer the lastChar probe, 970 // // there is may be a net D$ win with reverse matching. 971 // // But, reversing loop inhibits unroll of inner loop 972 // // for unknown reason. So, does running outer loop from 973 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 974 // for (int j = 0; j < targetCountLess1; j++) { 975 // if (target[targetOffset + j] != source[i+j]) { 976 // if ((cache & (1 << source[i+j])) == 0) { 977 // if (md2 < j+1) { 978 // i += j+1; 979 // continue outer_loop; 980 // } 981 // } 982 // i += md2; 983 // continue outer_loop; 984 // } 985 // } 986 // return i - sourceOffset; 987 // } 988 // if ((cache & (1 << src)) == 0) { 989 // i += targetCountLess1; 990 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 991 // i++; 992 // } 993 // return -1; 994 // } 995 996 //------------------------------string_indexOf------------------------ 997 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 998 jint cache_i, jint md2_i) { 999 1000 Node* no_ctrl = NULL; 1001 float likely = PROB_LIKELY(0.9); 1002 float unlikely = PROB_UNLIKELY(0.9); 1003 1004 const int value_offset = java_lang_String::value_offset_in_bytes(); 1005 const int count_offset = java_lang_String::count_offset_in_bytes(); 1006 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1007 1008 ciInstanceKlass* klass = env()->String_klass(); 1009 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 1010 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0); 1011 1012 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset); 1013 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 1014 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset); 1015 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1016 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset); 1017 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset)); 1018 1019 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) ); 1020 jint target_length = target_array->length(); 1021 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1022 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1023 1024 IdealKit kit(gvn(), control(), merged_memory()); 1025 #define __ kit. 1026 Node* zero = __ ConI(0); 1027 Node* one = __ ConI(1); 1028 Node* cache = __ ConI(cache_i); 1029 Node* md2 = __ ConI(md2_i); 1030 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1031 Node* targetCount = __ ConI(target_length); 1032 Node* targetCountLess1 = __ ConI(target_length - 1); 1033 Node* targetOffset = __ ConI(targetOffset_i); 1034 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1035 1036 IdealVariable rtn(kit), i(kit), j(kit); __ declares_done(); 1037 Node* outer_loop = __ make_label(2 /* goto */); 1038 Node* return_ = __ make_label(1); 1039 1040 __ set(rtn,__ ConI(-1)); 1041 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); { 1042 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1043 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1044 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1045 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1046 __ loop(j, zero, BoolTest::lt, targetCountLess1); { 1047 Node* tpj = __ AddI(targetOffset, __ value(j)); 1048 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1049 Node* ipj = __ AddI(__ value(i), __ value(j)); 1050 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1051 __ if_then(targ, BoolTest::ne, src2); { 1052 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1053 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1054 __ increment(i, __ AddI(__ value(j), one)); 1055 __ goto_(outer_loop); 1056 } __ end_if(); __ dead(j); 1057 }__ end_if(); __ dead(j); 1058 __ increment(i, md2); 1059 __ goto_(outer_loop); 1060 }__ end_if(); 1061 __ increment(j, one); 1062 }__ end_loop(); __ dead(j); 1063 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1064 __ goto_(return_); 1065 }__ end_if(); 1066 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1067 __ increment(i, targetCountLess1); 1068 }__ end_if(); 1069 __ increment(i, one); 1070 __ bind(outer_loop); 1071 }__ end_loop(); __ dead(i); 1072 __ bind(return_); 1073 __ drain_delay_transform(); 1074 1075 set_control(__ ctrl()); 1076 Node* result = __ value(rtn); 1077 #undef __ 1078 C->set_has_loops(true); 1079 return result; 1080 } 1081 1082 //------------------------------inline_string_indexOf------------------------ 1083 bool LibraryCallKit::inline_string_indexOf() { 1084 1085 const int value_offset = java_lang_String::value_offset_in_bytes(); 1086 const int count_offset = java_lang_String::count_offset_in_bytes(); 1087 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1088 1089 _sp += 2; 1090 Node *argument = pop(); // pop non-receiver first: it was pushed second 1091 Node *receiver = pop(); 1092 1093 Node* result; 1094 if (Matcher::has_match_rule(Op_StrIndexOf) && 1095 UseSSE42Intrinsics) { 1096 // Generate SSE4.2 version of indexOf 1097 // We currently only have match rules that use SSE4.2 1098 1099 // Null check on self without removing any arguments. The argument 1100 // null check technically happens in the wrong place, which can lead to 1101 // invalid stack traces when string compare is inlined into a method 1102 // which handles NullPointerExceptions. 1103 _sp += 2; 1104 receiver = do_null_check(receiver, T_OBJECT); 1105 argument = do_null_check(argument, T_OBJECT); 1106 _sp -= 2; 1107 1108 if (stopped()) { 1109 return true; 1110 } 1111 1112 ciInstanceKlass* klass = env()->String_klass(); 1113 const TypeInstPtr* string_type = 1114 TypeInstPtr::make(TypePtr::BotPTR, klass, false, NULL, 0); 1115 1116 result = 1117 _gvn.transform(new (C, 7) 1118 StrIndexOfNode(control(), 1119 memory(TypeAryPtr::CHARS), 1120 memory(string_type->add_offset(value_offset)), 1121 memory(string_type->add_offset(count_offset)), 1122 memory(string_type->add_offset(offset_offset)), 1123 receiver, 1124 argument)); 1125 } else { //Use LibraryCallKit::string_indexOf 1126 // don't intrinsify is argument isn't a constant string. 1127 if (!argument->is_Con()) { 1128 return false; 1129 } 1130 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); 1131 if (str_type == NULL) { 1132 return false; 1133 } 1134 ciInstanceKlass* klass = env()->String_klass(); 1135 ciObject* str_const = str_type->const_oop(); 1136 if (str_const == NULL || str_const->klass() != klass) { 1137 return false; 1138 } 1139 ciInstance* str = str_const->as_instance(); 1140 assert(str != NULL, "must be instance"); 1141 1142 ciObject* v = str->field_value_by_offset(value_offset).as_object(); 1143 int o = str->field_value_by_offset(offset_offset).as_int(); 1144 int c = str->field_value_by_offset(count_offset).as_int(); 1145 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1146 1147 // constant strings have no offset and count == length which 1148 // simplifies the resulting code somewhat so lets optimize for that. 1149 if (o != 0 || c != pat->length()) { 1150 return false; 1151 } 1152 1153 // Null check on self without removing any arguments. The argument 1154 // null check technically happens in the wrong place, which can lead to 1155 // invalid stack traces when string compare is inlined into a method 1156 // which handles NullPointerExceptions. 1157 _sp += 2; 1158 receiver = do_null_check(receiver, T_OBJECT); 1159 // No null check on the argument is needed since it's a constant String oop. 1160 _sp -= 2; 1161 if (stopped()) { 1162 return true; 1163 } 1164 1165 // The null string as a pattern always returns 0 (match at beginning of string) 1166 if (c == 0) { 1167 push(intcon(0)); 1168 return true; 1169 } 1170 1171 // Generate default indexOf 1172 jchar lastChar = pat->char_at(o + (c - 1)); 1173 int cache = 0; 1174 int i; 1175 for (i = 0; i < c - 1; i++) { 1176 assert(i < pat->length(), "out of range"); 1177 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1178 } 1179 1180 int md2 = c; 1181 for (i = 0; i < c - 1; i++) { 1182 assert(i < pat->length(), "out of range"); 1183 if (pat->char_at(o + i) == lastChar) { 1184 md2 = (c - 1) - i; 1185 } 1186 } 1187 1188 result = string_indexOf(receiver, pat, o, cache, md2); 1189 } 1190 1191 push(result); 1192 return true; 1193 } 1194 1195 //--------------------------pop_math_arg-------------------------------- 1196 // Pop a double argument to a math function from the stack 1197 // rounding it if necessary. 1198 Node * LibraryCallKit::pop_math_arg() { 1199 Node *arg = pop_pair(); 1200 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) 1201 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); 1202 return arg; 1203 } 1204 1205 //------------------------------inline_trig---------------------------------- 1206 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1207 // argument reduction which will turn into a fast/slow diamond. 1208 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1209 _sp += arg_size(); // restore stack pointer 1210 Node* arg = pop_math_arg(); 1211 Node* trig = NULL; 1212 1213 switch (id) { 1214 case vmIntrinsics::_dsin: 1215 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); 1216 break; 1217 case vmIntrinsics::_dcos: 1218 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); 1219 break; 1220 case vmIntrinsics::_dtan: 1221 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); 1222 break; 1223 default: 1224 assert(false, "bad intrinsic was passed in"); 1225 return false; 1226 } 1227 1228 // Rounding required? Check for argument reduction! 1229 if( Matcher::strict_fp_requires_explicit_rounding ) { 1230 1231 static const double pi_4 = 0.7853981633974483; 1232 static const double neg_pi_4 = -0.7853981633974483; 1233 // pi/2 in 80-bit extended precision 1234 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1235 // -pi/2 in 80-bit extended precision 1236 // 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}; 1237 // Cutoff value for using this argument reduction technique 1238 //static const double pi_2_minus_epsilon = 1.564660403643354; 1239 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1240 1241 // Pseudocode for sin: 1242 // if (x <= Math.PI / 4.0) { 1243 // if (x >= -Math.PI / 4.0) return fsin(x); 1244 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1245 // } else { 1246 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1247 // } 1248 // return StrictMath.sin(x); 1249 1250 // Pseudocode for cos: 1251 // if (x <= Math.PI / 4.0) { 1252 // if (x >= -Math.PI / 4.0) return fcos(x); 1253 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1254 // } else { 1255 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1256 // } 1257 // return StrictMath.cos(x); 1258 1259 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1260 // requires a special machine instruction to load it. Instead we'll try 1261 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1262 // probably do the math inside the SIN encoding. 1263 1264 // Make the merge point 1265 RegionNode *r = new (C, 3) RegionNode(3); 1266 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); 1267 1268 // Flatten arg so we need only 1 test 1269 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); 1270 // Node for PI/4 constant 1271 Node *pi4 = makecon(TypeD::make(pi_4)); 1272 // Check PI/4 : abs(arg) 1273 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); 1274 // Check: If PI/4 < abs(arg) then go slow 1275 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); 1276 // Branch either way 1277 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1278 set_control(opt_iff(r,iff)); 1279 1280 // Set fast path result 1281 phi->init_req(2,trig); 1282 1283 // Slow path - non-blocking leaf call 1284 Node* call = NULL; 1285 switch (id) { 1286 case vmIntrinsics::_dsin: 1287 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1288 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1289 "Sin", NULL, arg, top()); 1290 break; 1291 case vmIntrinsics::_dcos: 1292 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1293 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1294 "Cos", NULL, arg, top()); 1295 break; 1296 case vmIntrinsics::_dtan: 1297 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1298 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1299 "Tan", NULL, arg, top()); 1300 break; 1301 } 1302 assert(control()->in(0) == call, ""); 1303 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); 1304 r->init_req(1,control()); 1305 phi->init_req(1,slow_result); 1306 1307 // Post-merge 1308 set_control(_gvn.transform(r)); 1309 record_for_igvn(r); 1310 trig = _gvn.transform(phi); 1311 1312 C->set_has_split_ifs(true); // Has chance for split-if optimization 1313 } 1314 // Push result back on JVM stack 1315 push_pair(trig); 1316 return true; 1317 } 1318 1319 //------------------------------inline_sqrt------------------------------------- 1320 // Inline square root instruction, if possible. 1321 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { 1322 assert(id == vmIntrinsics::_dsqrt, "Not square root"); 1323 _sp += arg_size(); // restore stack pointer 1324 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); 1325 return true; 1326 } 1327 1328 //------------------------------inline_abs------------------------------------- 1329 // Inline absolute value instruction, if possible. 1330 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { 1331 assert(id == vmIntrinsics::_dabs, "Not absolute value"); 1332 _sp += arg_size(); // restore stack pointer 1333 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); 1334 return true; 1335 } 1336 1337 //------------------------------inline_exp------------------------------------- 1338 // Inline exp instructions, if possible. The Intel hardware only misses 1339 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1340 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { 1341 assert(id == vmIntrinsics::_dexp, "Not exp"); 1342 1343 // If this inlining ever returned NaN in the past, we do not intrinsify it 1344 // every again. NaN results requires StrictMath.exp handling. 1345 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1346 1347 // Do not intrinsify on older platforms which lack cmove. 1348 if (ConditionalMoveLimit == 0) return false; 1349 1350 _sp += arg_size(); // restore stack pointer 1351 Node *x = pop_math_arg(); 1352 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); 1353 1354 //------------------- 1355 //result=(result.isNaN())? StrictMath::exp():result; 1356 // Check: If isNaN() by checking result!=result? then go to Strict Math 1357 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1358 // Build the boolean node 1359 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1360 1361 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1362 // End the current control-flow path 1363 push_pair(x); 1364 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp 1365 // to handle. Recompile without intrinsifying Math.exp 1366 uncommon_trap(Deoptimization::Reason_intrinsic, 1367 Deoptimization::Action_make_not_entrant); 1368 } 1369 1370 C->set_has_split_ifs(true); // Has chance for split-if optimization 1371 1372 push_pair(result); 1373 1374 return true; 1375 } 1376 1377 //------------------------------inline_pow------------------------------------- 1378 // Inline power instructions, if possible. 1379 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { 1380 assert(id == vmIntrinsics::_dpow, "Not pow"); 1381 1382 // If this inlining ever returned NaN in the past, we do not intrinsify it 1383 // every again. NaN results requires StrictMath.pow handling. 1384 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1385 1386 // Do not intrinsify on older platforms which lack cmove. 1387 if (ConditionalMoveLimit == 0) return false; 1388 1389 // Pseudocode for pow 1390 // if (x <= 0.0) { 1391 // if ((double)((int)y)==y) { // if y is int 1392 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y) 1393 // } else { 1394 // result = NaN; 1395 // } 1396 // } else { 1397 // result = DPow(x,y); 1398 // } 1399 // if (result != result)? { 1400 // uncommon_trap(); 1401 // } 1402 // return result; 1403 1404 _sp += arg_size(); // restore stack pointer 1405 Node* y = pop_math_arg(); 1406 Node* x = pop_math_arg(); 1407 1408 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); 1409 1410 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow 1411 // inside of something) then skip the fancy tests and just check for 1412 // NaN result. 1413 Node *result = NULL; 1414 if( jvms()->depth() >= 1 ) { 1415 result = fast_result; 1416 } else { 1417 1418 // Set the merge point for If node with condition of (x <= 0.0) 1419 // There are four possible paths to region node and phi node 1420 RegionNode *r = new (C, 4) RegionNode(4); 1421 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); 1422 1423 // Build the first if node: if (x <= 0.0) 1424 // Node for 0 constant 1425 Node *zeronode = makecon(TypeD::ZERO); 1426 // Check x:0 1427 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); 1428 // Check: If (x<=0) then go complex path 1429 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); 1430 // Branch either way 1431 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1432 Node *opt_test = _gvn.transform(if1); 1433 //assert( opt_test->is_If(), "Expect an IfNode"); 1434 IfNode *opt_if1 = (IfNode*)opt_test; 1435 // Fast path taken; set region slot 3 1436 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) ); 1437 r->init_req(3,fast_taken); // Capture fast-control 1438 1439 // Fast path not-taken, i.e. slow path 1440 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) ); 1441 1442 // Set fast path result 1443 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) ); 1444 phi->init_req(3, fast_result); 1445 1446 // Complex path 1447 // Build the second if node (if y is int) 1448 // Node for (int)y 1449 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y)); 1450 // Node for (double)((int) y) 1451 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty)); 1452 // Check (double)((int) y) : y 1453 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y)); 1454 // Check if (y isn't int) then go to slow path 1455 1456 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) ); 1457 // Branch either way 1458 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1459 Node *slow_path = opt_iff(r,if2); // Set region path 2 1460 1461 // Calculate DPow(abs(x), y)*(1 & (int)y) 1462 // Node for constant 1 1463 Node *conone = intcon(1); 1464 // 1& (int)y 1465 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) ); 1466 // zero node 1467 Node *conzero = intcon(0); 1468 // Check (1&(int)y)==0? 1469 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero)); 1470 // Check if (1&(int)y)!=0?, if so the result is negative 1471 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); 1472 // abs(x) 1473 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); 1474 // abs(x)^y 1475 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) ); 1476 // -abs(x)^y 1477 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); 1478 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1479 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1480 // Set complex path fast result 1481 phi->init_req(2, signresult); 1482 1483 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1484 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1485 r->init_req(1,slow_path); 1486 phi->init_req(1,slow_result); 1487 1488 // Post merge 1489 set_control(_gvn.transform(r)); 1490 record_for_igvn(r); 1491 result=_gvn.transform(phi); 1492 } 1493 1494 //------------------- 1495 //result=(result.isNaN())? uncommon_trap():result; 1496 // Check: If isNaN() by checking result!=result? then go to Strict Math 1497 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1498 // Build the boolean node 1499 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1500 1501 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1502 // End the current control-flow path 1503 push_pair(x); 1504 push_pair(y); 1505 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow 1506 // to handle. Recompile without intrinsifying Math.pow. 1507 uncommon_trap(Deoptimization::Reason_intrinsic, 1508 Deoptimization::Action_make_not_entrant); 1509 } 1510 1511 C->set_has_split_ifs(true); // Has chance for split-if optimization 1512 1513 push_pair(result); 1514 1515 return true; 1516 } 1517 1518 //------------------------------inline_trans------------------------------------- 1519 // Inline transcendental instructions, if possible. The Intel hardware gets 1520 // these right, no funny corner cases missed. 1521 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { 1522 _sp += arg_size(); // restore stack pointer 1523 Node* arg = pop_math_arg(); 1524 Node* trans = NULL; 1525 1526 switch (id) { 1527 case vmIntrinsics::_dlog: 1528 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); 1529 break; 1530 case vmIntrinsics::_dlog10: 1531 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); 1532 break; 1533 default: 1534 assert(false, "bad intrinsic was passed in"); 1535 return false; 1536 } 1537 1538 // Push result back on JVM stack 1539 push_pair(trans); 1540 return true; 1541 } 1542 1543 //------------------------------runtime_math----------------------------- 1544 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1545 Node* a = NULL; 1546 Node* b = NULL; 1547 1548 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1549 "must be (DD)D or (D)D type"); 1550 1551 // Inputs 1552 _sp += arg_size(); // restore stack pointer 1553 if (call_type == OptoRuntime::Math_DD_D_Type()) { 1554 b = pop_math_arg(); 1555 } 1556 a = pop_math_arg(); 1557 1558 const TypePtr* no_memory_effects = NULL; 1559 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1560 no_memory_effects, 1561 a, top(), b, b ? top() : NULL); 1562 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); 1563 #ifdef ASSERT 1564 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); 1565 assert(value_top == top(), "second value must be top"); 1566 #endif 1567 1568 push_pair(value); 1569 return true; 1570 } 1571 1572 //------------------------------inline_math_native----------------------------- 1573 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1574 switch (id) { 1575 // These intrinsics are not properly supported on all hardware 1576 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1577 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); 1578 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1579 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); 1580 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1581 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); 1582 1583 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : 1584 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); 1585 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : 1586 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); 1587 1588 // These intrinsics are supported on all hardware 1589 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; 1590 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; 1591 1592 // These intrinsics don't work on X86. The ad implementation doesn't 1593 // handle NaN's properly. Instead of returning infinity, the ad 1594 // implementation returns a NaN on overflow. See bug: 6304089 1595 // Once the ad implementations are fixed, change the code below 1596 // to match the intrinsics above 1597 1598 case vmIntrinsics::_dexp: return 1599 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1600 case vmIntrinsics::_dpow: return 1601 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1602 1603 // These intrinsics are not yet correctly implemented 1604 case vmIntrinsics::_datan2: 1605 return false; 1606 1607 default: 1608 ShouldNotReachHere(); 1609 return false; 1610 } 1611 } 1612 1613 static bool is_simple_name(Node* n) { 1614 return (n->req() == 1 // constant 1615 || (n->is_Type() && n->as_Type()->type()->singleton()) 1616 || n->is_Proj() // parameter or return value 1617 || n->is_Phi() // local of some sort 1618 ); 1619 } 1620 1621 //----------------------------inline_min_max----------------------------------- 1622 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1623 push(generate_min_max(id, argument(0), argument(1))); 1624 1625 return true; 1626 } 1627 1628 Node* 1629 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1630 // These are the candidate return value: 1631 Node* xvalue = x0; 1632 Node* yvalue = y0; 1633 1634 if (xvalue == yvalue) { 1635 return xvalue; 1636 } 1637 1638 bool want_max = (id == vmIntrinsics::_max); 1639 1640 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1641 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1642 if (txvalue == NULL || tyvalue == NULL) return top(); 1643 // This is not really necessary, but it is consistent with a 1644 // hypothetical MaxINode::Value method: 1645 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1646 1647 // %%% This folding logic should (ideally) be in a different place. 1648 // Some should be inside IfNode, and there to be a more reliable 1649 // transformation of ?: style patterns into cmoves. We also want 1650 // more powerful optimizations around cmove and min/max. 1651 1652 // Try to find a dominating comparison of these guys. 1653 // It can simplify the index computation for Arrays.copyOf 1654 // and similar uses of System.arraycopy. 1655 // First, compute the normalized version of CmpI(x, y). 1656 int cmp_op = Op_CmpI; 1657 Node* xkey = xvalue; 1658 Node* ykey = yvalue; 1659 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); 1660 if (ideal_cmpxy->is_Cmp()) { 1661 // E.g., if we have CmpI(length - offset, count), 1662 // it might idealize to CmpI(length, count + offset) 1663 cmp_op = ideal_cmpxy->Opcode(); 1664 xkey = ideal_cmpxy->in(1); 1665 ykey = ideal_cmpxy->in(2); 1666 } 1667 1668 // Start by locating any relevant comparisons. 1669 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1670 Node* cmpxy = NULL; 1671 Node* cmpyx = NULL; 1672 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1673 Node* cmp = start_from->fast_out(k); 1674 if (cmp->outcnt() > 0 && // must have prior uses 1675 cmp->in(0) == NULL && // must be context-independent 1676 cmp->Opcode() == cmp_op) { // right kind of compare 1677 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1678 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1679 } 1680 } 1681 1682 const int NCMPS = 2; 1683 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1684 int cmpn; 1685 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1686 if (cmps[cmpn] != NULL) break; // find a result 1687 } 1688 if (cmpn < NCMPS) { 1689 // Look for a dominating test that tells us the min and max. 1690 int depth = 0; // Limit search depth for speed 1691 Node* dom = control(); 1692 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1693 if (++depth >= 100) break; 1694 Node* ifproj = dom; 1695 if (!ifproj->is_Proj()) continue; 1696 Node* iff = ifproj->in(0); 1697 if (!iff->is_If()) continue; 1698 Node* bol = iff->in(1); 1699 if (!bol->is_Bool()) continue; 1700 Node* cmp = bol->in(1); 1701 if (cmp == NULL) continue; 1702 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1703 if (cmps[cmpn] == cmp) break; 1704 if (cmpn == NCMPS) continue; 1705 BoolTest::mask btest = bol->as_Bool()->_test._test; 1706 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1707 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1708 // At this point, we know that 'x btest y' is true. 1709 switch (btest) { 1710 case BoolTest::eq: 1711 // They are proven equal, so we can collapse the min/max. 1712 // Either value is the answer. Choose the simpler. 1713 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1714 return yvalue; 1715 return xvalue; 1716 case BoolTest::lt: // x < y 1717 case BoolTest::le: // x <= y 1718 return (want_max ? yvalue : xvalue); 1719 case BoolTest::gt: // x > y 1720 case BoolTest::ge: // x >= y 1721 return (want_max ? xvalue : yvalue); 1722 } 1723 } 1724 } 1725 1726 // We failed to find a dominating test. 1727 // Let's pick a test that might GVN with prior tests. 1728 Node* best_bol = NULL; 1729 BoolTest::mask best_btest = BoolTest::illegal; 1730 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1731 Node* cmp = cmps[cmpn]; 1732 if (cmp == NULL) continue; 1733 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1734 Node* bol = cmp->fast_out(j); 1735 if (!bol->is_Bool()) continue; 1736 BoolTest::mask btest = bol->as_Bool()->_test._test; 1737 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1738 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1739 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1740 best_bol = bol->as_Bool(); 1741 best_btest = btest; 1742 } 1743 } 1744 } 1745 1746 Node* answer_if_true = NULL; 1747 Node* answer_if_false = NULL; 1748 switch (best_btest) { 1749 default: 1750 if (cmpxy == NULL) 1751 cmpxy = ideal_cmpxy; 1752 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); 1753 // and fall through: 1754 case BoolTest::lt: // x < y 1755 case BoolTest::le: // x <= y 1756 answer_if_true = (want_max ? yvalue : xvalue); 1757 answer_if_false = (want_max ? xvalue : yvalue); 1758 break; 1759 case BoolTest::gt: // x > y 1760 case BoolTest::ge: // x >= y 1761 answer_if_true = (want_max ? xvalue : yvalue); 1762 answer_if_false = (want_max ? yvalue : xvalue); 1763 break; 1764 } 1765 1766 jint hi, lo; 1767 if (want_max) { 1768 // We can sharpen the minimum. 1769 hi = MAX2(txvalue->_hi, tyvalue->_hi); 1770 lo = MAX2(txvalue->_lo, tyvalue->_lo); 1771 } else { 1772 // We can sharpen the maximum. 1773 hi = MIN2(txvalue->_hi, tyvalue->_hi); 1774 lo = MIN2(txvalue->_lo, tyvalue->_lo); 1775 } 1776 1777 // Use a flow-free graph structure, to avoid creating excess control edges 1778 // which could hinder other optimizations. 1779 // Since Math.min/max is often used with arraycopy, we want 1780 // tightly_coupled_allocation to be able to see beyond min/max expressions. 1781 Node* cmov = CMoveNode::make(C, NULL, best_bol, 1782 answer_if_false, answer_if_true, 1783 TypeInt::make(lo, hi, widen)); 1784 1785 return _gvn.transform(cmov); 1786 1787 /* 1788 // This is not as desirable as it may seem, since Min and Max 1789 // nodes do not have a full set of optimizations. 1790 // And they would interfere, anyway, with 'if' optimizations 1791 // and with CMoveI canonical forms. 1792 switch (id) { 1793 case vmIntrinsics::_min: 1794 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 1795 case vmIntrinsics::_max: 1796 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 1797 default: 1798 ShouldNotReachHere(); 1799 } 1800 */ 1801 } 1802 1803 inline int 1804 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 1805 const TypePtr* base_type = TypePtr::NULL_PTR; 1806 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 1807 if (base_type == NULL) { 1808 // Unknown type. 1809 return Type::AnyPtr; 1810 } else if (base_type == TypePtr::NULL_PTR) { 1811 // Since this is a NULL+long form, we have to switch to a rawptr. 1812 base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); 1813 offset = MakeConX(0); 1814 return Type::RawPtr; 1815 } else if (base_type->base() == Type::RawPtr) { 1816 return Type::RawPtr; 1817 } else if (base_type->isa_oopptr()) { 1818 // Base is never null => always a heap address. 1819 if (base_type->ptr() == TypePtr::NotNull) { 1820 return Type::OopPtr; 1821 } 1822 // Offset is small => always a heap address. 1823 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 1824 if (offset_type != NULL && 1825 base_type->offset() == 0 && // (should always be?) 1826 offset_type->_lo >= 0 && 1827 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 1828 return Type::OopPtr; 1829 } 1830 // Otherwise, it might either be oop+off or NULL+addr. 1831 return Type::AnyPtr; 1832 } else { 1833 // No information: 1834 return Type::AnyPtr; 1835 } 1836 } 1837 1838 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 1839 int kind = classify_unsafe_addr(base, offset); 1840 if (kind == Type::RawPtr) { 1841 return basic_plus_adr(top(), base, offset); 1842 } else { 1843 return basic_plus_adr(base, offset); 1844 } 1845 } 1846 1847 //----------------------------inline_bitCount_int/long----------------------- 1848 // inline int Integer.bitCount(int) 1849 // inline int Long.bitCount(long) 1850 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) { 1851 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount"); 1852 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false; 1853 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false; 1854 _sp += arg_size(); // restore stack pointer 1855 switch (id) { 1856 case vmIntrinsics::_bitCount_i: 1857 push(_gvn.transform(new (C, 2) PopCountINode(pop()))); 1858 break; 1859 case vmIntrinsics::_bitCount_l: 1860 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair()))); 1861 break; 1862 default: 1863 ShouldNotReachHere(); 1864 } 1865 return true; 1866 } 1867 1868 //----------------------------inline_reverseBytes_int/long------------------- 1869 // inline Integer.reverseBytes(int) 1870 // inline Long.reverseBytes(long) 1871 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { 1872 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l, "not reverse Bytes"); 1873 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; 1874 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; 1875 _sp += arg_size(); // restore stack pointer 1876 switch (id) { 1877 case vmIntrinsics::_reverseBytes_i: 1878 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); 1879 break; 1880 case vmIntrinsics::_reverseBytes_l: 1881 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); 1882 break; 1883 default: 1884 ; 1885 } 1886 return true; 1887 } 1888 1889 //----------------------------inline_unsafe_access---------------------------- 1890 1891 const static BasicType T_ADDRESS_HOLDER = T_LONG; 1892 1893 // Interpret Unsafe.fieldOffset cookies correctly: 1894 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 1895 1896 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 1897 if (callee()->is_static()) return false; // caller must have the capability! 1898 1899 #ifndef PRODUCT 1900 { 1901 ResourceMark rm; 1902 // Check the signatures. 1903 ciSignature* sig = signature(); 1904 #ifdef ASSERT 1905 if (!is_store) { 1906 // Object getObject(Object base, int/long offset), etc. 1907 BasicType rtype = sig->return_type()->basic_type(); 1908 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 1909 rtype = T_ADDRESS; // it is really a C void* 1910 assert(rtype == type, "getter must return the expected value"); 1911 if (!is_native_ptr) { 1912 assert(sig->count() == 2, "oop getter has 2 arguments"); 1913 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 1914 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 1915 } else { 1916 assert(sig->count() == 1, "native getter has 1 argument"); 1917 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 1918 } 1919 } else { 1920 // void putObject(Object base, int/long offset, Object x), etc. 1921 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 1922 if (!is_native_ptr) { 1923 assert(sig->count() == 3, "oop putter has 3 arguments"); 1924 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 1925 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 1926 } else { 1927 assert(sig->count() == 2, "native putter has 2 arguments"); 1928 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 1929 } 1930 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 1931 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 1932 vtype = T_ADDRESS; // it is really a C void* 1933 assert(vtype == type, "putter must accept the expected value"); 1934 } 1935 #endif // ASSERT 1936 } 1937 #endif //PRODUCT 1938 1939 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 1940 1941 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; 1942 1943 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words 1944 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); 1945 1946 debug_only(int saved_sp = _sp); 1947 _sp += nargs; 1948 1949 Node* val; 1950 debug_only(val = (Node*)(uintptr_t)-1); 1951 1952 1953 if (is_store) { 1954 // Get the value being stored. (Pop it first; it was pushed last.) 1955 switch (type) { 1956 case T_DOUBLE: 1957 case T_LONG: 1958 case T_ADDRESS: 1959 val = pop_pair(); 1960 break; 1961 default: 1962 val = pop(); 1963 } 1964 } 1965 1966 // Build address expression. See the code in inline_unsafe_prefetch. 1967 Node *adr; 1968 Node *heap_base_oop = top(); 1969 if (!is_native_ptr) { 1970 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 1971 Node* offset = pop_pair(); 1972 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 1973 Node* base = pop(); 1974 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 1975 // to be plain byte offsets, which are also the same as those accepted 1976 // by oopDesc::field_base. 1977 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 1978 "fieldOffset must be byte-scaled"); 1979 // 32-bit machines ignore the high half! 1980 offset = ConvL2X(offset); 1981 adr = make_unsafe_address(base, offset); 1982 heap_base_oop = base; 1983 } else { 1984 Node* ptr = pop_pair(); 1985 // Adjust Java long to machine word: 1986 ptr = ConvL2X(ptr); 1987 adr = make_unsafe_address(NULL, ptr); 1988 } 1989 1990 // Pop receiver last: it was pushed first. 1991 Node *receiver = pop(); 1992 1993 assert(saved_sp == _sp, "must have correct argument count"); 1994 1995 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 1996 1997 // First guess at the value type. 1998 const Type *value_type = Type::get_const_basic_type(type); 1999 2000 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2001 // there was not enough information to nail it down. 2002 Compile::AliasType* alias_type = C->alias_type(adr_type); 2003 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2004 2005 // We will need memory barriers unless we can determine a unique 2006 // alias category for this reference. (Note: If for some reason 2007 // the barriers get omitted and the unsafe reference begins to "pollute" 2008 // the alias analysis of the rest of the graph, either Compile::can_alias 2009 // or Compile::must_alias will throw a diagnostic assert.) 2010 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2011 2012 if (!is_store && type == T_OBJECT) { 2013 // Attempt to infer a sharper value type from the offset and base type. 2014 ciKlass* sharpened_klass = NULL; 2015 2016 // See if it is an instance field, with an object type. 2017 if (alias_type->field() != NULL) { 2018 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2019 if (alias_type->field()->type()->is_klass()) { 2020 sharpened_klass = alias_type->field()->type()->as_klass(); 2021 } 2022 } 2023 2024 // See if it is a narrow oop array. 2025 if (adr_type->isa_aryptr()) { 2026 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes(type)) { 2027 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2028 if (elem_type != NULL) { 2029 sharpened_klass = elem_type->klass(); 2030 } 2031 } 2032 } 2033 2034 if (sharpened_klass != NULL) { 2035 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2036 2037 // Sharpen the value type. 2038 value_type = tjp; 2039 2040 #ifndef PRODUCT 2041 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2042 tty->print(" from base type: "); adr_type->dump(); 2043 tty->print(" sharpened value: "); value_type->dump(); 2044 } 2045 #endif 2046 } 2047 } 2048 2049 // Null check on self without removing any arguments. The argument 2050 // null check technically happens in the wrong place, which can lead to 2051 // invalid stack traces when the primitive is inlined into a method 2052 // which handles NullPointerExceptions. 2053 _sp += nargs; 2054 do_null_check(receiver, T_OBJECT); 2055 _sp -= nargs; 2056 if (stopped()) { 2057 return true; 2058 } 2059 // Heap pointers get a null-check from the interpreter, 2060 // as a courtesy. However, this is not guaranteed by Unsafe, 2061 // and it is not possible to fully distinguish unintended nulls 2062 // from intended ones in this API. 2063 2064 if (is_volatile) { 2065 // We need to emit leading and trailing CPU membars (see below) in 2066 // addition to memory membars when is_volatile. This is a little 2067 // too strong, but avoids the need to insert per-alias-type 2068 // volatile membars (for stores; compare Parse::do_put_xxx), which 2069 // we cannot do effectively here because we probably only have a 2070 // rough approximation of type. 2071 need_mem_bar = true; 2072 // For Stores, place a memory ordering barrier now. 2073 if (is_store) 2074 insert_mem_bar(Op_MemBarRelease); 2075 } 2076 2077 // Memory barrier to prevent normal and 'unsafe' accesses from 2078 // bypassing each other. Happens after null checks, so the 2079 // exception paths do not take memory state from the memory barrier, 2080 // so there's no problems making a strong assert about mixing users 2081 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2082 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2083 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2084 2085 if (!is_store) { 2086 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2087 // load value and push onto stack 2088 switch (type) { 2089 case T_BOOLEAN: 2090 case T_CHAR: 2091 case T_BYTE: 2092 case T_SHORT: 2093 case T_INT: 2094 case T_FLOAT: 2095 case T_OBJECT: 2096 push( p ); 2097 break; 2098 case T_ADDRESS: 2099 // Cast to an int type. 2100 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); 2101 p = ConvX2L(p); 2102 push_pair(p); 2103 break; 2104 case T_DOUBLE: 2105 case T_LONG: 2106 push_pair( p ); 2107 break; 2108 default: ShouldNotReachHere(); 2109 } 2110 } else { 2111 // place effect of store into memory 2112 switch (type) { 2113 case T_DOUBLE: 2114 val = dstore_rounding(val); 2115 break; 2116 case T_ADDRESS: 2117 // Repackage the long as a pointer. 2118 val = ConvL2X(val); 2119 val = _gvn.transform( new (C, 2) CastX2PNode(val) ); 2120 break; 2121 } 2122 2123 if (type != T_OBJECT ) { 2124 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2125 } else { 2126 // Possibly an oop being stored to Java heap or native memory 2127 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2128 // oop to Java heap. 2129 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type); 2130 } else { 2131 2132 // We can't tell at compile time if we are storing in the Java heap or outside 2133 // of it. So we need to emit code to conditionally do the proper type of 2134 // store. 2135 2136 IdealKit kit(gvn(), control(), merged_memory()); 2137 kit.declares_done(); 2138 // QQQ who knows what probability is here?? 2139 kit.if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2140 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, val->bottom_type(), type); 2141 } kit.else_(); { 2142 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2143 } kit.end_if(); 2144 } 2145 } 2146 } 2147 2148 if (is_volatile) { 2149 if (!is_store) 2150 insert_mem_bar(Op_MemBarAcquire); 2151 else 2152 insert_mem_bar(Op_MemBarVolatile); 2153 } 2154 2155 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2156 2157 return true; 2158 } 2159 2160 //----------------------------inline_unsafe_prefetch---------------------------- 2161 2162 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2163 #ifndef PRODUCT 2164 { 2165 ResourceMark rm; 2166 // Check the signatures. 2167 ciSignature* sig = signature(); 2168 #ifdef ASSERT 2169 // Object getObject(Object base, int/long offset), etc. 2170 BasicType rtype = sig->return_type()->basic_type(); 2171 if (!is_native_ptr) { 2172 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2173 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2174 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2175 } else { 2176 assert(sig->count() == 1, "native prefetch has 1 argument"); 2177 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2178 } 2179 #endif // ASSERT 2180 } 2181 #endif // !PRODUCT 2182 2183 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2184 2185 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args 2186 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); 2187 2188 debug_only(int saved_sp = _sp); 2189 _sp += nargs; 2190 2191 // Build address expression. See the code in inline_unsafe_access. 2192 Node *adr; 2193 if (!is_native_ptr) { 2194 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2195 Node* offset = pop_pair(); 2196 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2197 Node* base = pop(); 2198 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2199 // to be plain byte offsets, which are also the same as those accepted 2200 // by oopDesc::field_base. 2201 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2202 "fieldOffset must be byte-scaled"); 2203 // 32-bit machines ignore the high half! 2204 offset = ConvL2X(offset); 2205 adr = make_unsafe_address(base, offset); 2206 } else { 2207 Node* ptr = pop_pair(); 2208 // Adjust Java long to machine word: 2209 ptr = ConvL2X(ptr); 2210 adr = make_unsafe_address(NULL, ptr); 2211 } 2212 2213 if (is_static) { 2214 assert(saved_sp == _sp, "must have correct argument count"); 2215 } else { 2216 // Pop receiver last: it was pushed first. 2217 Node *receiver = pop(); 2218 assert(saved_sp == _sp, "must have correct argument count"); 2219 2220 // Null check on self without removing any arguments. The argument 2221 // null check technically happens in the wrong place, which can lead to 2222 // invalid stack traces when the primitive is inlined into a method 2223 // which handles NullPointerExceptions. 2224 _sp += nargs; 2225 do_null_check(receiver, T_OBJECT); 2226 _sp -= nargs; 2227 if (stopped()) { 2228 return true; 2229 } 2230 } 2231 2232 // Generate the read or write prefetch 2233 Node *prefetch; 2234 if (is_store) { 2235 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); 2236 } else { 2237 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); 2238 } 2239 prefetch->init_req(0, control()); 2240 set_i_o(_gvn.transform(prefetch)); 2241 2242 return true; 2243 } 2244 2245 //----------------------------inline_unsafe_CAS---------------------------- 2246 2247 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { 2248 // This basic scheme here is the same as inline_unsafe_access, but 2249 // differs in enough details that combining them would make the code 2250 // overly confusing. (This is a true fact! I originally combined 2251 // them, but even I was confused by it!) As much code/comments as 2252 // possible are retained from inline_unsafe_access though to make 2253 // the correspondences clearer. - dl 2254 2255 if (callee()->is_static()) return false; // caller must have the capability! 2256 2257 #ifndef PRODUCT 2258 { 2259 ResourceMark rm; 2260 // Check the signatures. 2261 ciSignature* sig = signature(); 2262 #ifdef ASSERT 2263 BasicType rtype = sig->return_type()->basic_type(); 2264 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2265 assert(sig->count() == 4, "CAS has 4 arguments"); 2266 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2267 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2268 #endif // ASSERT 2269 } 2270 #endif //PRODUCT 2271 2272 // number of stack slots per value argument (1 or 2) 2273 int type_words = type2size[type]; 2274 2275 // Cannot inline wide CAS on machines that don't support it natively 2276 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8()) 2277 return false; 2278 2279 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2280 2281 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; 2282 int nargs = 1 + 1 + 2 + type_words + type_words; 2283 2284 // pop arguments: newval, oldval, offset, base, and receiver 2285 debug_only(int saved_sp = _sp); 2286 _sp += nargs; 2287 Node* newval = (type_words == 1) ? pop() : pop_pair(); 2288 Node* oldval = (type_words == 1) ? pop() : pop_pair(); 2289 Node *offset = pop_pair(); 2290 Node *base = pop(); 2291 Node *receiver = pop(); 2292 assert(saved_sp == _sp, "must have correct argument count"); 2293 2294 // Null check receiver. 2295 _sp += nargs; 2296 do_null_check(receiver, T_OBJECT); 2297 _sp -= nargs; 2298 if (stopped()) { 2299 return true; 2300 } 2301 2302 // Build field offset expression. 2303 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2304 // to be plain byte offsets, which are also the same as those accepted 2305 // by oopDesc::field_base. 2306 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2307 // 32-bit machines ignore the high half of long offsets 2308 offset = ConvL2X(offset); 2309 Node* adr = make_unsafe_address(base, offset); 2310 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2311 2312 // (Unlike inline_unsafe_access, there seems no point in trying 2313 // to refine types. Just use the coarse types here. 2314 const Type *value_type = Type::get_const_basic_type(type); 2315 Compile::AliasType* alias_type = C->alias_type(adr_type); 2316 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2317 int alias_idx = C->get_alias_index(adr_type); 2318 2319 // Memory-model-wise, a CAS acts like a little synchronized block, 2320 // so needs barriers on each side. These don't translate into 2321 // actual barriers on most machines, but we still need rest of 2322 // compiler to respect ordering. 2323 2324 insert_mem_bar(Op_MemBarRelease); 2325 insert_mem_bar(Op_MemBarCPUOrder); 2326 2327 // 4984716: MemBars must be inserted before this 2328 // memory node in order to avoid a false 2329 // dependency which will confuse the scheduler. 2330 Node *mem = memory(alias_idx); 2331 2332 // For now, we handle only those cases that actually exist: ints, 2333 // longs, and Object. Adding others should be straightforward. 2334 Node* cas; 2335 switch(type) { 2336 case T_INT: 2337 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2338 break; 2339 case T_LONG: 2340 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2341 break; 2342 case T_OBJECT: 2343 // reference stores need a store barrier. 2344 // (They don't if CAS fails, but it isn't worth checking.) 2345 pre_barrier(control(), base, adr, alias_idx, newval, value_type, T_OBJECT); 2346 #ifdef _LP64 2347 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2348 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2349 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2350 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr, 2351 newval_enc, oldval_enc)); 2352 } else 2353 #endif 2354 { 2355 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2356 } 2357 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); 2358 break; 2359 default: 2360 ShouldNotReachHere(); 2361 break; 2362 } 2363 2364 // SCMemProjNodes represent the memory state of CAS. Their main 2365 // role is to prevent CAS nodes from being optimized away when their 2366 // results aren't used. 2367 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 2368 set_memory(proj, alias_idx); 2369 2370 // Add the trailing membar surrounding the access 2371 insert_mem_bar(Op_MemBarCPUOrder); 2372 insert_mem_bar(Op_MemBarAcquire); 2373 2374 push(cas); 2375 return true; 2376 } 2377 2378 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2379 // This is another variant of inline_unsafe_access, differing in 2380 // that it always issues store-store ("release") barrier and ensures 2381 // store-atomicity (which only matters for "long"). 2382 2383 if (callee()->is_static()) return false; // caller must have the capability! 2384 2385 #ifndef PRODUCT 2386 { 2387 ResourceMark rm; 2388 // Check the signatures. 2389 ciSignature* sig = signature(); 2390 #ifdef ASSERT 2391 BasicType rtype = sig->return_type()->basic_type(); 2392 assert(rtype == T_VOID, "must return void"); 2393 assert(sig->count() == 3, "has 3 arguments"); 2394 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2395 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2396 #endif // ASSERT 2397 } 2398 #endif //PRODUCT 2399 2400 // number of stack slots per value argument (1 or 2) 2401 int type_words = type2size[type]; 2402 2403 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2404 2405 // Argument words: "this" plus oop plus offset plus value; 2406 int nargs = 1 + 1 + 2 + type_words; 2407 2408 // pop arguments: val, offset, base, and receiver 2409 debug_only(int saved_sp = _sp); 2410 _sp += nargs; 2411 Node* val = (type_words == 1) ? pop() : pop_pair(); 2412 Node *offset = pop_pair(); 2413 Node *base = pop(); 2414 Node *receiver = pop(); 2415 assert(saved_sp == _sp, "must have correct argument count"); 2416 2417 // Null check receiver. 2418 _sp += nargs; 2419 do_null_check(receiver, T_OBJECT); 2420 _sp -= nargs; 2421 if (stopped()) { 2422 return true; 2423 } 2424 2425 // Build field offset expression. 2426 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2427 // 32-bit machines ignore the high half of long offsets 2428 offset = ConvL2X(offset); 2429 Node* adr = make_unsafe_address(base, offset); 2430 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2431 const Type *value_type = Type::get_const_basic_type(type); 2432 Compile::AliasType* alias_type = C->alias_type(adr_type); 2433 2434 insert_mem_bar(Op_MemBarRelease); 2435 insert_mem_bar(Op_MemBarCPUOrder); 2436 // Ensure that the store is atomic for longs: 2437 bool require_atomic_access = true; 2438 Node* store; 2439 if (type == T_OBJECT) // reference stores need a store barrier. 2440 store = store_oop_to_unknown(control(), base, adr, adr_type, val, value_type, type); 2441 else { 2442 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2443 } 2444 insert_mem_bar(Op_MemBarCPUOrder); 2445 return true; 2446 } 2447 2448 bool LibraryCallKit::inline_unsafe_allocate() { 2449 if (callee()->is_static()) return false; // caller must have the capability! 2450 int nargs = 1 + 1; 2451 assert(signature()->size() == nargs-1, "alloc has 1 argument"); 2452 null_check_receiver(callee()); // check then ignore argument(0) 2453 _sp += nargs; // set original stack for use by uncommon_trap 2454 Node* cls = do_null_check(argument(1), T_OBJECT); 2455 _sp -= nargs; 2456 if (stopped()) return true; 2457 2458 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 2459 _sp += nargs; // set original stack for use by uncommon_trap 2460 kls = do_null_check(kls, T_OBJECT); 2461 _sp -= nargs; 2462 if (stopped()) return true; // argument was like int.class 2463 2464 // Note: The argument might still be an illegal value like 2465 // Serializable.class or Object[].class. The runtime will handle it. 2466 // But we must make an explicit check for initialization. 2467 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc)); 2468 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT); 2469 Node* bits = intcon(instanceKlass::fully_initialized); 2470 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); 2471 // The 'test' is non-zero if we need to take a slow path. 2472 2473 Node* obj = new_instance(kls, test); 2474 push(obj); 2475 2476 return true; 2477 } 2478 2479 //------------------------inline_native_time_funcs-------------- 2480 // inline code for System.currentTimeMillis() and System.nanoTime() 2481 // these have the same type and signature 2482 bool LibraryCallKit::inline_native_time_funcs(bool isNano) { 2483 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) : 2484 CAST_FROM_FN_PTR(address, os::javaTimeMillis); 2485 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis"; 2486 const TypeFunc *tf = OptoRuntime::current_time_millis_Type(); 2487 const TypePtr* no_memory_effects = NULL; 2488 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2489 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); 2490 #ifdef ASSERT 2491 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); 2492 assert(value_top == top(), "second value must be top"); 2493 #endif 2494 push_pair(value); 2495 return true; 2496 } 2497 2498 //------------------------inline_native_currentThread------------------ 2499 bool LibraryCallKit::inline_native_currentThread() { 2500 Node* junk = NULL; 2501 push(generate_current_thread(junk)); 2502 return true; 2503 } 2504 2505 //------------------------inline_native_isInterrupted------------------ 2506 bool LibraryCallKit::inline_native_isInterrupted() { 2507 const int nargs = 1+1; // receiver + boolean 2508 assert(nargs == arg_size(), "sanity"); 2509 // Add a fast path to t.isInterrupted(clear_int): 2510 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 2511 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 2512 // So, in the common case that the interrupt bit is false, 2513 // we avoid making a call into the VM. Even if the interrupt bit 2514 // is true, if the clear_int argument is false, we avoid the VM call. 2515 // However, if the receiver is not currentThread, we must call the VM, 2516 // because there must be some locking done around the operation. 2517 2518 // We only go to the fast case code if we pass two guards. 2519 // Paths which do not pass are accumulated in the slow_region. 2520 RegionNode* slow_region = new (C, 1) RegionNode(1); 2521 record_for_igvn(slow_region); 2522 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow 2523 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); 2524 enum { no_int_result_path = 1, 2525 no_clear_result_path = 2, 2526 slow_result_path = 3 2527 }; 2528 2529 // (a) Receiving thread must be the current thread. 2530 Node* rec_thr = argument(0); 2531 Node* tls_ptr = NULL; 2532 Node* cur_thr = generate_current_thread(tls_ptr); 2533 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); 2534 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); 2535 2536 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); 2537 if (!known_current_thread) 2538 generate_slow_guard(bol_thr, slow_region); 2539 2540 // (b) Interrupt bit on TLS must be false. 2541 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2542 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 2543 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 2544 Node* int_bit = make_load(NULL, p, TypeInt::BOOL, T_INT); 2545 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); 2546 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); 2547 2548 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2549 2550 // First fast path: if (!TLS._interrupted) return false; 2551 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); 2552 result_rgn->init_req(no_int_result_path, false_bit); 2553 result_val->init_req(no_int_result_path, intcon(0)); 2554 2555 // drop through to next case 2556 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); 2557 2558 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 2559 Node* clr_arg = argument(1); 2560 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); 2561 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); 2562 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 2563 2564 // Second fast path: ... else if (!clear_int) return true; 2565 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); 2566 result_rgn->init_req(no_clear_result_path, false_arg); 2567 result_val->init_req(no_clear_result_path, intcon(1)); 2568 2569 // drop through to next case 2570 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); 2571 2572 // (d) Otherwise, go to the slow path. 2573 slow_region->add_req(control()); 2574 set_control( _gvn.transform(slow_region) ); 2575 2576 if (stopped()) { 2577 // There is no slow path. 2578 result_rgn->init_req(slow_result_path, top()); 2579 result_val->init_req(slow_result_path, top()); 2580 } else { 2581 // non-virtual because it is a private non-static 2582 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 2583 2584 Node* slow_val = set_results_for_java_call(slow_call); 2585 // this->control() comes from set_results_for_java_call 2586 2587 // If we know that the result of the slow call will be true, tell the optimizer! 2588 if (known_current_thread) slow_val = intcon(1); 2589 2590 Node* fast_io = slow_call->in(TypeFunc::I_O); 2591 Node* fast_mem = slow_call->in(TypeFunc::Memory); 2592 // These two phis are pre-filled with copies of of the fast IO and Memory 2593 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 2594 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 2595 2596 result_rgn->init_req(slow_result_path, control()); 2597 io_phi ->init_req(slow_result_path, i_o()); 2598 mem_phi ->init_req(slow_result_path, reset_memory()); 2599 result_val->init_req(slow_result_path, slow_val); 2600 2601 set_all_memory( _gvn.transform(mem_phi) ); 2602 set_i_o( _gvn.transform(io_phi) ); 2603 } 2604 2605 push_result(result_rgn, result_val); 2606 C->set_has_split_ifs(true); // Has chance for split-if optimization 2607 2608 return true; 2609 } 2610 2611 //---------------------------load_mirror_from_klass---------------------------- 2612 // Given a klass oop, load its java mirror (a java.lang.Class oop). 2613 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 2614 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc)); 2615 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 2616 } 2617 2618 //-----------------------load_klass_from_mirror_common------------------------- 2619 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 2620 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 2621 // and branch to the given path on the region. 2622 // If never_see_null, take an uncommon trap on null, so we can optimistically 2623 // compile for the non-null case. 2624 // If the region is NULL, force never_see_null = true. 2625 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 2626 bool never_see_null, 2627 int nargs, 2628 RegionNode* region, 2629 int null_path, 2630 int offset) { 2631 if (region == NULL) never_see_null = true; 2632 Node* p = basic_plus_adr(mirror, offset); 2633 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 2634 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); 2635 _sp += nargs; // any deopt will start just before call to enclosing method 2636 Node* null_ctl = top(); 2637 kls = null_check_oop(kls, &null_ctl, never_see_null); 2638 if (region != NULL) { 2639 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 2640 region->init_req(null_path, null_ctl); 2641 } else { 2642 assert(null_ctl == top(), "no loose ends"); 2643 } 2644 _sp -= nargs; 2645 return kls; 2646 } 2647 2648 //--------------------(inline_native_Class_query helpers)--------------------- 2649 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 2650 // Fall through if (mods & mask) == bits, take the guard otherwise. 2651 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 2652 // Branch around if the given klass has the given modifier bit set. 2653 // Like generate_guard, adds a new path onto the region. 2654 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); 2655 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 2656 Node* mask = intcon(modifier_mask); 2657 Node* bits = intcon(modifier_bits); 2658 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); 2659 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); 2660 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); 2661 return generate_fair_guard(bol, region); 2662 } 2663 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 2664 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 2665 } 2666 2667 //-------------------------inline_native_Class_query------------------- 2668 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 2669 int nargs = 1+0; // just the Class mirror, in most cases 2670 const Type* return_type = TypeInt::BOOL; 2671 Node* prim_return_value = top(); // what happens if it's a primitive class? 2672 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 2673 bool expect_prim = false; // most of these guys expect to work on refs 2674 2675 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 2676 2677 switch (id) { 2678 case vmIntrinsics::_isInstance: 2679 nargs = 1+1; // the Class mirror, plus the object getting queried about 2680 // nothing is an instance of a primitive type 2681 prim_return_value = intcon(0); 2682 break; 2683 case vmIntrinsics::_getModifiers: 2684 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2685 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 2686 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 2687 break; 2688 case vmIntrinsics::_isInterface: 2689 prim_return_value = intcon(0); 2690 break; 2691 case vmIntrinsics::_isArray: 2692 prim_return_value = intcon(0); 2693 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 2694 break; 2695 case vmIntrinsics::_isPrimitive: 2696 prim_return_value = intcon(1); 2697 expect_prim = true; // obviously 2698 break; 2699 case vmIntrinsics::_getSuperclass: 2700 prim_return_value = null(); 2701 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 2702 break; 2703 case vmIntrinsics::_getComponentType: 2704 prim_return_value = null(); 2705 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 2706 break; 2707 case vmIntrinsics::_getClassAccessFlags: 2708 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2709 return_type = TypeInt::INT; // not bool! 6297094 2710 break; 2711 default: 2712 ShouldNotReachHere(); 2713 } 2714 2715 Node* mirror = argument(0); 2716 Node* obj = (nargs <= 1)? top(): argument(1); 2717 2718 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 2719 if (mirror_con == NULL) return false; // cannot happen? 2720 2721 #ifndef PRODUCT 2722 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2723 ciType* k = mirror_con->java_mirror_type(); 2724 if (k) { 2725 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 2726 k->print_name(); 2727 tty->cr(); 2728 } 2729 } 2730 #endif 2731 2732 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 2733 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2734 record_for_igvn(region); 2735 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); 2736 2737 // The mirror will never be null of Reflection.getClassAccessFlags, however 2738 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 2739 // if it is. See bug 4774291. 2740 2741 // For Reflection.getClassAccessFlags(), the null check occurs in 2742 // the wrong place; see inline_unsafe_access(), above, for a similar 2743 // situation. 2744 _sp += nargs; // set original stack for use by uncommon_trap 2745 mirror = do_null_check(mirror, T_OBJECT); 2746 _sp -= nargs; 2747 // If mirror or obj is dead, only null-path is taken. 2748 if (stopped()) return true; 2749 2750 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 2751 2752 // Now load the mirror's klass metaobject, and null-check it. 2753 // Side-effects region with the control path if the klass is null. 2754 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, 2755 region, _prim_path); 2756 // If kls is null, we have a primitive mirror. 2757 phi->init_req(_prim_path, prim_return_value); 2758 if (stopped()) { push_result(region, phi); return true; } 2759 2760 Node* p; // handy temp 2761 Node* null_ctl; 2762 2763 // Now that we have the non-null klass, we can perform the real query. 2764 // For constant classes, the query will constant-fold in LoadNode::Value. 2765 Node* query_value = top(); 2766 switch (id) { 2767 case vmIntrinsics::_isInstance: 2768 // nothing is an instance of a primitive type 2769 query_value = gen_instanceof(obj, kls); 2770 break; 2771 2772 case vmIntrinsics::_getModifiers: 2773 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc)); 2774 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 2775 break; 2776 2777 case vmIntrinsics::_isInterface: 2778 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 2779 if (generate_interface_guard(kls, region) != NULL) 2780 // A guard was added. If the guard is taken, it was an interface. 2781 phi->add_req(intcon(1)); 2782 // If we fall through, it's a plain class. 2783 query_value = intcon(0); 2784 break; 2785 2786 case vmIntrinsics::_isArray: 2787 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 2788 if (generate_array_guard(kls, region) != NULL) 2789 // A guard was added. If the guard is taken, it was an array. 2790 phi->add_req(intcon(1)); 2791 // If we fall through, it's a plain class. 2792 query_value = intcon(0); 2793 break; 2794 2795 case vmIntrinsics::_isPrimitive: 2796 query_value = intcon(0); // "normal" path produces false 2797 break; 2798 2799 case vmIntrinsics::_getSuperclass: 2800 // The rules here are somewhat unfortunate, but we can still do better 2801 // with random logic than with a JNI call. 2802 // Interfaces store null or Object as _super, but must report null. 2803 // Arrays store an intermediate super as _super, but must report Object. 2804 // Other types can report the actual _super. 2805 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 2806 if (generate_interface_guard(kls, region) != NULL) 2807 // A guard was added. If the guard is taken, it was an interface. 2808 phi->add_req(null()); 2809 if (generate_array_guard(kls, region) != NULL) 2810 // A guard was added. If the guard is taken, it was an array. 2811 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 2812 // If we fall through, it's a plain class. Get its _super. 2813 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc)); 2814 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); 2815 null_ctl = top(); 2816 kls = null_check_oop(kls, &null_ctl); 2817 if (null_ctl != top()) { 2818 // If the guard is taken, Object.superClass is null (both klass and mirror). 2819 region->add_req(null_ctl); 2820 phi ->add_req(null()); 2821 } 2822 if (!stopped()) { 2823 query_value = load_mirror_from_klass(kls); 2824 } 2825 break; 2826 2827 case vmIntrinsics::_getComponentType: 2828 if (generate_array_guard(kls, region) != NULL) { 2829 // Be sure to pin the oop load to the guard edge just created: 2830 Node* is_array_ctrl = region->in(region->req()-1); 2831 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc)); 2832 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 2833 phi->add_req(cmo); 2834 } 2835 query_value = null(); // non-array case is null 2836 break; 2837 2838 case vmIntrinsics::_getClassAccessFlags: 2839 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); 2840 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 2841 break; 2842 2843 default: 2844 ShouldNotReachHere(); 2845 } 2846 2847 // Fall-through is the normal case of a query to a real class. 2848 phi->init_req(1, query_value); 2849 region->init_req(1, control()); 2850 2851 push_result(region, phi); 2852 C->set_has_split_ifs(true); // Has chance for split-if optimization 2853 2854 return true; 2855 } 2856 2857 //--------------------------inline_native_subtype_check------------------------ 2858 // This intrinsic takes the JNI calls out of the heart of 2859 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 2860 bool LibraryCallKit::inline_native_subtype_check() { 2861 int nargs = 1+1; // the Class mirror, plus the other class getting examined 2862 2863 // Pull both arguments off the stack. 2864 Node* args[2]; // two java.lang.Class mirrors: superc, subc 2865 args[0] = argument(0); 2866 args[1] = argument(1); 2867 Node* klasses[2]; // corresponding Klasses: superk, subk 2868 klasses[0] = klasses[1] = top(); 2869 2870 enum { 2871 // A full decision tree on {superc is prim, subc is prim}: 2872 _prim_0_path = 1, // {P,N} => false 2873 // {P,P} & superc!=subc => false 2874 _prim_same_path, // {P,P} & superc==subc => true 2875 _prim_1_path, // {N,P} => false 2876 _ref_subtype_path, // {N,N} & subtype check wins => true 2877 _both_ref_path, // {N,N} & subtype check loses => false 2878 PATH_LIMIT 2879 }; 2880 2881 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2882 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 2883 record_for_igvn(region); 2884 2885 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 2886 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 2887 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 2888 2889 // First null-check both mirrors and load each mirror's klass metaobject. 2890 int which_arg; 2891 for (which_arg = 0; which_arg <= 1; which_arg++) { 2892 Node* arg = args[which_arg]; 2893 _sp += nargs; // set original stack for use by uncommon_trap 2894 arg = do_null_check(arg, T_OBJECT); 2895 _sp -= nargs; 2896 if (stopped()) break; 2897 args[which_arg] = _gvn.transform(arg); 2898 2899 Node* p = basic_plus_adr(arg, class_klass_offset); 2900 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 2901 klasses[which_arg] = _gvn.transform(kls); 2902 } 2903 2904 // Having loaded both klasses, test each for null. 2905 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 2906 for (which_arg = 0; which_arg <= 1; which_arg++) { 2907 Node* kls = klasses[which_arg]; 2908 Node* null_ctl = top(); 2909 _sp += nargs; // set original stack for use by uncommon_trap 2910 kls = null_check_oop(kls, &null_ctl, never_see_null); 2911 _sp -= nargs; 2912 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 2913 region->init_req(prim_path, null_ctl); 2914 if (stopped()) break; 2915 klasses[which_arg] = kls; 2916 } 2917 2918 if (!stopped()) { 2919 // now we have two reference types, in klasses[0..1] 2920 Node* subk = klasses[1]; // the argument to isAssignableFrom 2921 Node* superk = klasses[0]; // the receiver 2922 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 2923 // now we have a successful reference subtype check 2924 region->set_req(_ref_subtype_path, control()); 2925 } 2926 2927 // If both operands are primitive (both klasses null), then 2928 // we must return true when they are identical primitives. 2929 // It is convenient to test this after the first null klass check. 2930 set_control(region->in(_prim_0_path)); // go back to first null check 2931 if (!stopped()) { 2932 // Since superc is primitive, make a guard for the superc==subc case. 2933 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); 2934 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); 2935 generate_guard(bol_eq, region, PROB_FAIR); 2936 if (region->req() == PATH_LIMIT+1) { 2937 // A guard was added. If the added guard is taken, superc==subc. 2938 region->swap_edges(PATH_LIMIT, _prim_same_path); 2939 region->del_req(PATH_LIMIT); 2940 } 2941 region->set_req(_prim_0_path, control()); // Not equal after all. 2942 } 2943 2944 // these are the only paths that produce 'true': 2945 phi->set_req(_prim_same_path, intcon(1)); 2946 phi->set_req(_ref_subtype_path, intcon(1)); 2947 2948 // pull together the cases: 2949 assert(region->req() == PATH_LIMIT, "sane region"); 2950 for (uint i = 1; i < region->req(); i++) { 2951 Node* ctl = region->in(i); 2952 if (ctl == NULL || ctl == top()) { 2953 region->set_req(i, top()); 2954 phi ->set_req(i, top()); 2955 } else if (phi->in(i) == NULL) { 2956 phi->set_req(i, intcon(0)); // all other paths produce 'false' 2957 } 2958 } 2959 2960 set_control(_gvn.transform(region)); 2961 push(_gvn.transform(phi)); 2962 2963 return true; 2964 } 2965 2966 //---------------------generate_array_guard_common------------------------ 2967 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 2968 bool obj_array, bool not_array) { 2969 // If obj_array/non_array==false/false: 2970 // Branch around if the given klass is in fact an array (either obj or prim). 2971 // If obj_array/non_array==false/true: 2972 // Branch around if the given klass is not an array klass of any kind. 2973 // If obj_array/non_array==true/true: 2974 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 2975 // If obj_array/non_array==true/false: 2976 // Branch around if the kls is an oop array (Object[] or subtype) 2977 // 2978 // Like generate_guard, adds a new path onto the region. 2979 jint layout_con = 0; 2980 Node* layout_val = get_layout_helper(kls, layout_con); 2981 if (layout_val == NULL) { 2982 bool query = (obj_array 2983 ? Klass::layout_helper_is_objArray(layout_con) 2984 : Klass::layout_helper_is_javaArray(layout_con)); 2985 if (query == not_array) { 2986 return NULL; // never a branch 2987 } else { // always a branch 2988 Node* always_branch = control(); 2989 if (region != NULL) 2990 region->add_req(always_branch); 2991 set_control(top()); 2992 return always_branch; 2993 } 2994 } 2995 // Now test the correct condition. 2996 jint nval = (obj_array 2997 ? ((jint)Klass::_lh_array_tag_type_value 2998 << Klass::_lh_array_tag_shift) 2999 : Klass::_lh_neutral_value); 3000 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); 3001 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3002 // invert the test if we are looking for a non-array 3003 if (not_array) btest = BoolTest(btest).negate(); 3004 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); 3005 return generate_fair_guard(bol, region); 3006 } 3007 3008 3009 //-----------------------inline_native_newArray-------------------------- 3010 bool LibraryCallKit::inline_native_newArray() { 3011 int nargs = 2; 3012 Node* mirror = argument(0); 3013 Node* count_val = argument(1); 3014 3015 _sp += nargs; // set original stack for use by uncommon_trap 3016 mirror = do_null_check(mirror, T_OBJECT); 3017 _sp -= nargs; 3018 // If mirror or obj is dead, only null-path is taken. 3019 if (stopped()) return true; 3020 3021 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3022 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3023 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3024 TypeInstPtr::NOTNULL); 3025 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3026 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3027 TypePtr::BOTTOM); 3028 3029 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3030 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3031 nargs, 3032 result_reg, _slow_path); 3033 Node* normal_ctl = control(); 3034 Node* no_array_ctl = result_reg->in(_slow_path); 3035 3036 // Generate code for the slow case. We make a call to newArray(). 3037 set_control(no_array_ctl); 3038 if (!stopped()) { 3039 // Either the input type is void.class, or else the 3040 // array klass has not yet been cached. Either the 3041 // ensuing call will throw an exception, or else it 3042 // will cache the array klass for next time. 3043 PreserveJVMState pjvms(this); 3044 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3045 Node* slow_result = set_results_for_java_call(slow_call); 3046 // this->control() comes from set_results_for_java_call 3047 result_reg->set_req(_slow_path, control()); 3048 result_val->set_req(_slow_path, slow_result); 3049 result_io ->set_req(_slow_path, i_o()); 3050 result_mem->set_req(_slow_path, reset_memory()); 3051 } 3052 3053 set_control(normal_ctl); 3054 if (!stopped()) { 3055 // Normal case: The array type has been cached in the java.lang.Class. 3056 // The following call works fine even if the array type is polymorphic. 3057 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3058 Node* obj = new_array(klass_node, count_val, nargs); 3059 result_reg->init_req(_normal_path, control()); 3060 result_val->init_req(_normal_path, obj); 3061 result_io ->init_req(_normal_path, i_o()); 3062 result_mem->init_req(_normal_path, reset_memory()); 3063 } 3064 3065 // Return the combined state. 3066 set_i_o( _gvn.transform(result_io) ); 3067 set_all_memory( _gvn.transform(result_mem) ); 3068 push_result(result_reg, result_val); 3069 C->set_has_split_ifs(true); // Has chance for split-if optimization 3070 3071 return true; 3072 } 3073 3074 //----------------------inline_native_getLength-------------------------- 3075 bool LibraryCallKit::inline_native_getLength() { 3076 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3077 3078 int nargs = 1; 3079 Node* array = argument(0); 3080 3081 _sp += nargs; // set original stack for use by uncommon_trap 3082 array = do_null_check(array, T_OBJECT); 3083 _sp -= nargs; 3084 3085 // If array is dead, only null-path is taken. 3086 if (stopped()) return true; 3087 3088 // Deoptimize if it is a non-array. 3089 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3090 3091 if (non_array != NULL) { 3092 PreserveJVMState pjvms(this); 3093 set_control(non_array); 3094 _sp += nargs; // push the arguments back on the stack 3095 uncommon_trap(Deoptimization::Reason_intrinsic, 3096 Deoptimization::Action_maybe_recompile); 3097 } 3098 3099 // If control is dead, only non-array-path is taken. 3100 if (stopped()) return true; 3101 3102 // The works fine even if the array type is polymorphic. 3103 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3104 push( load_array_length(array) ); 3105 3106 C->set_has_split_ifs(true); // Has chance for split-if optimization 3107 3108 return true; 3109 } 3110 3111 //------------------------inline_array_copyOf---------------------------- 3112 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3113 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3114 3115 // Restore the stack and pop off the arguments. 3116 int nargs = 3 + (is_copyOfRange? 1: 0); 3117 Node* original = argument(0); 3118 Node* start = is_copyOfRange? argument(1): intcon(0); 3119 Node* end = is_copyOfRange? argument(2): argument(1); 3120 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3121 3122 _sp += nargs; // set original stack for use by uncommon_trap 3123 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); 3124 original = do_null_check(original, T_OBJECT); 3125 _sp -= nargs; 3126 3127 // Check if a null path was taken unconditionally. 3128 if (stopped()) return true; 3129 3130 Node* orig_length = load_array_length(original); 3131 3132 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, nargs, 3133 NULL, 0); 3134 _sp += nargs; // set original stack for use by uncommon_trap 3135 klass_node = do_null_check(klass_node, T_OBJECT); 3136 _sp -= nargs; 3137 3138 RegionNode* bailout = new (C, 1) RegionNode(1); 3139 record_for_igvn(bailout); 3140 3141 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3142 // Bail out if that is so. 3143 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3144 if (not_objArray != NULL) { 3145 // Improve the klass node's type from the new optimistic assumption: 3146 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3147 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3148 Node* cast = new (C, 2) CastPPNode(klass_node, akls); 3149 cast->init_req(0, control()); 3150 klass_node = _gvn.transform(cast); 3151 } 3152 3153 // Bail out if either start or end is negative. 3154 generate_negative_guard(start, bailout, &start); 3155 generate_negative_guard(end, bailout, &end); 3156 3157 Node* length = end; 3158 if (_gvn.type(start) != TypeInt::ZERO) { 3159 length = _gvn.transform( new (C, 3) SubINode(end, start) ); 3160 } 3161 3162 // Bail out if length is negative. 3163 // ...Not needed, since the new_array will throw the right exception. 3164 //generate_negative_guard(length, bailout, &length); 3165 3166 if (bailout->req() > 1) { 3167 PreserveJVMState pjvms(this); 3168 set_control( _gvn.transform(bailout) ); 3169 _sp += nargs; // push the arguments back on the stack 3170 uncommon_trap(Deoptimization::Reason_intrinsic, 3171 Deoptimization::Action_maybe_recompile); 3172 } 3173 3174 if (!stopped()) { 3175 // How many elements will we copy from the original? 3176 // The answer is MinI(orig_length - start, length). 3177 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); 3178 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3179 3180 Node* newcopy = new_array(klass_node, length, nargs); 3181 3182 // Generate a direct call to the right arraycopy function(s). 3183 // We know the copy is disjoint but we might not know if the 3184 // oop stores need checking. 3185 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3186 // This will fail a store-check if x contains any non-nulls. 3187 bool disjoint_bases = true; 3188 bool length_never_negative = true; 3189 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3190 original, start, newcopy, intcon(0), moved, 3191 nargs, disjoint_bases, length_never_negative); 3192 3193 push(newcopy); 3194 } 3195 3196 C->set_has_split_ifs(true); // Has chance for split-if optimization 3197 3198 return true; 3199 } 3200 3201 3202 //----------------------generate_virtual_guard--------------------------- 3203 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3204 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3205 RegionNode* slow_region) { 3206 ciMethod* method = callee(); 3207 int vtable_index = method->vtable_index(); 3208 // Get the methodOop out of the appropriate vtable entry. 3209 int entry_offset = (instanceKlass::vtable_start_offset() + 3210 vtable_index*vtableEntry::size()) * wordSize + 3211 vtableEntry::method_offset_in_bytes(); 3212 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3213 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); 3214 3215 // Compare the target method with the expected method (e.g., Object.hashCode). 3216 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method); 3217 3218 Node* native_call = makecon(native_call_addr); 3219 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); 3220 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); 3221 3222 return generate_slow_guard(test_native, slow_region); 3223 } 3224 3225 //-----------------------generate_method_call---------------------------- 3226 // Use generate_method_call to make a slow-call to the real 3227 // method if the fast path fails. An alternative would be to 3228 // use a stub like OptoRuntime::slow_arraycopy_Java. 3229 // This only works for expanding the current library call, 3230 // not another intrinsic. (E.g., don't use this for making an 3231 // arraycopy call inside of the copyOf intrinsic.) 3232 CallJavaNode* 3233 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3234 // When compiling the intrinsic method itself, do not use this technique. 3235 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3236 3237 ciMethod* method = callee(); 3238 // ensure the JVMS we have will be correct for this call 3239 guarantee(method_id == method->intrinsic_id(), "must match"); 3240 3241 const TypeFunc* tf = TypeFunc::make(method); 3242 int tfdc = tf->domain()->cnt(); 3243 CallJavaNode* slow_call; 3244 if (is_static) { 3245 assert(!is_virtual, ""); 3246 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3247 SharedRuntime::get_resolve_static_call_stub(), 3248 method, bci()); 3249 } else if (is_virtual) { 3250 null_check_receiver(method); 3251 int vtable_index = methodOopDesc::invalid_vtable_index; 3252 if (UseInlineCaches) { 3253 // Suppress the vtable call 3254 } else { 3255 // hashCode and clone are not a miranda methods, 3256 // so the vtable index is fixed. 3257 // No need to use the linkResolver to get it. 3258 vtable_index = method->vtable_index(); 3259 } 3260 slow_call = new(C, tfdc) CallDynamicJavaNode(tf, 3261 SharedRuntime::get_resolve_virtual_call_stub(), 3262 method, vtable_index, bci()); 3263 } else { // neither virtual nor static: opt_virtual 3264 null_check_receiver(method); 3265 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3266 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3267 method, bci()); 3268 slow_call->set_optimized_virtual(true); 3269 } 3270 set_arguments_for_java_call(slow_call); 3271 set_edges_for_java_call(slow_call); 3272 return slow_call; 3273 } 3274 3275 3276 //------------------------------inline_native_hashcode-------------------- 3277 // Build special case code for calls to hashCode on an object. 3278 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3279 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3280 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3281 3282 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3283 3284 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3285 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3286 TypeInt::INT); 3287 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3288 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3289 TypePtr::BOTTOM); 3290 Node* obj = NULL; 3291 if (!is_static) { 3292 // Check for hashing null object 3293 obj = null_check_receiver(callee()); 3294 if (stopped()) return true; // unconditionally null 3295 result_reg->init_req(_null_path, top()); 3296 result_val->init_req(_null_path, top()); 3297 } else { 3298 // Do a null check, and return zero if null. 3299 // System.identityHashCode(null) == 0 3300 obj = argument(0); 3301 Node* null_ctl = top(); 3302 obj = null_check_oop(obj, &null_ctl); 3303 result_reg->init_req(_null_path, null_ctl); 3304 result_val->init_req(_null_path, _gvn.intcon(0)); 3305 } 3306 3307 // Unconditionally null? Then return right away. 3308 if (stopped()) { 3309 set_control( result_reg->in(_null_path) ); 3310 if (!stopped()) 3311 push( result_val ->in(_null_path) ); 3312 return true; 3313 } 3314 3315 // After null check, get the object's klass. 3316 Node* obj_klass = load_object_klass(obj); 3317 3318 // This call may be virtual (invokevirtual) or bound (invokespecial). 3319 // For each case we generate slightly different code. 3320 3321 // We only go to the fast case code if we pass a number of guards. The 3322 // paths which do not pass are accumulated in the slow_region. 3323 RegionNode* slow_region = new (C, 1) RegionNode(1); 3324 record_for_igvn(slow_region); 3325 3326 // If this is a virtual call, we generate a funny guard. We pull out 3327 // the vtable entry corresponding to hashCode() from the target object. 3328 // If the target method which we are calling happens to be the native 3329 // Object hashCode() method, we pass the guard. We do not need this 3330 // guard for non-virtual calls -- the caller is known to be the native 3331 // Object hashCode(). 3332 if (is_virtual) { 3333 generate_virtual_guard(obj_klass, slow_region); 3334 } 3335 3336 // Get the header out of the object, use LoadMarkNode when available 3337 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3338 Node* header = make_load(NULL, header_addr, TypeRawPtr::BOTTOM, T_ADDRESS); 3339 header = _gvn.transform( new (C, 2) CastP2XNode(NULL, header) ); 3340 3341 // Test the header to see if it is unlocked. 3342 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3343 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); 3344 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3345 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); 3346 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); 3347 3348 generate_slow_guard(test_unlocked, slow_region); 3349 3350 // Get the hash value and check to see that it has been properly assigned. 3351 // We depend on hash_mask being at most 32 bits and avoid the use of 3352 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3353 // vm: see markOop.hpp. 3354 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3355 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3356 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); 3357 // This hack lets the hash bits live anywhere in the mark object now, as long 3358 // as the shift drops the relevant bits into the low 32 bits. Note that 3359 // Java spec says that HashCode is an int so there's no point in capturing 3360 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3361 hshifted_header = ConvX2I(hshifted_header); 3362 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); 3363 3364 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3365 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); 3366 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); 3367 3368 generate_slow_guard(test_assigned, slow_region); 3369 3370 Node* init_mem = reset_memory(); 3371 // fill in the rest of the null path: 3372 result_io ->init_req(_null_path, i_o()); 3373 result_mem->init_req(_null_path, init_mem); 3374 3375 result_val->init_req(_fast_path, hash_val); 3376 result_reg->init_req(_fast_path, control()); 3377 result_io ->init_req(_fast_path, i_o()); 3378 result_mem->init_req(_fast_path, init_mem); 3379 3380 // Generate code for the slow case. We make a call to hashCode(). 3381 set_control(_gvn.transform(slow_region)); 3382 if (!stopped()) { 3383 // No need for PreserveJVMState, because we're using up the present state. 3384 set_all_memory(init_mem); 3385 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; 3386 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; 3387 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3388 Node* slow_result = set_results_for_java_call(slow_call); 3389 // this->control() comes from set_results_for_java_call 3390 result_reg->init_req(_slow_path, control()); 3391 result_val->init_req(_slow_path, slow_result); 3392 result_io ->set_req(_slow_path, i_o()); 3393 result_mem ->set_req(_slow_path, reset_memory()); 3394 } 3395 3396 // Return the combined state. 3397 set_i_o( _gvn.transform(result_io) ); 3398 set_all_memory( _gvn.transform(result_mem) ); 3399 push_result(result_reg, result_val); 3400 3401 return true; 3402 } 3403 3404 //---------------------------inline_native_getClass---------------------------- 3405 // Build special case code for calls to getClass on an object. 3406 bool LibraryCallKit::inline_native_getClass() { 3407 Node* obj = null_check_receiver(callee()); 3408 if (stopped()) return true; 3409 push( load_mirror_from_klass(load_object_klass(obj)) ); 3410 return true; 3411 } 3412 3413 //-----------------inline_native_Reflection_getCallerClass--------------------- 3414 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3415 // 3416 // NOTE that this code must perform the same logic as 3417 // vframeStream::security_get_caller_frame in that it must skip 3418 // Method.invoke() and auxiliary frames. 3419 3420 3421 3422 3423 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3424 ciMethod* method = callee(); 3425 3426 #ifndef PRODUCT 3427 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3428 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3429 } 3430 #endif 3431 3432 debug_only(int saved_sp = _sp); 3433 3434 // Argument words: (int depth) 3435 int nargs = 1; 3436 3437 _sp += nargs; 3438 Node* caller_depth_node = pop(); 3439 3440 assert(saved_sp == _sp, "must have correct argument count"); 3441 3442 // The depth value must be a constant in order for the runtime call 3443 // to be eliminated. 3444 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); 3445 if (caller_depth_type == NULL || !caller_depth_type->is_con()) { 3446 #ifndef PRODUCT 3447 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3448 tty->print_cr(" Bailing out because caller depth was not a constant"); 3449 } 3450 #endif 3451 return false; 3452 } 3453 // Note that the JVM state at this point does not include the 3454 // getCallerClass() frame which we are trying to inline. The 3455 // semantics of getCallerClass(), however, are that the "first" 3456 // frame is the getCallerClass() frame, so we subtract one from the 3457 // requested depth before continuing. We don't inline requests of 3458 // getCallerClass(0). 3459 int caller_depth = caller_depth_type->get_con() - 1; 3460 if (caller_depth < 0) { 3461 #ifndef PRODUCT 3462 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3463 tty->print_cr(" Bailing out because caller depth was %d", caller_depth); 3464 } 3465 #endif 3466 return false; 3467 } 3468 3469 if (!jvms()->has_method()) { 3470 #ifndef PRODUCT 3471 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3472 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3473 } 3474 #endif 3475 return false; 3476 } 3477 int _depth = jvms()->depth(); // cache call chain depth 3478 3479 // Walk back up the JVM state to find the caller at the required 3480 // depth. NOTE that this code must perform the same logic as 3481 // vframeStream::security_get_caller_frame in that it must skip 3482 // Method.invoke() and auxiliary frames. Note also that depth is 3483 // 1-based (1 is the bottom of the inlining). 3484 int inlining_depth = _depth; 3485 JVMState* caller_jvms = NULL; 3486 3487 if (inlining_depth > 0) { 3488 caller_jvms = jvms(); 3489 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); 3490 do { 3491 // The following if-tests should be performed in this order 3492 if (is_method_invoke_or_aux_frame(caller_jvms)) { 3493 // Skip a Method.invoke() or auxiliary frame 3494 } else if (caller_depth > 0) { 3495 // Skip real frame 3496 --caller_depth; 3497 } else { 3498 // We're done: reached desired caller after skipping. 3499 break; 3500 } 3501 caller_jvms = caller_jvms->caller(); 3502 --inlining_depth; 3503 } while (inlining_depth > 0); 3504 } 3505 3506 if (inlining_depth == 0) { 3507 #ifndef PRODUCT 3508 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3509 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); 3510 tty->print_cr(" JVM state at this point:"); 3511 for (int i = _depth; i >= 1; i--) { 3512 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3513 } 3514 } 3515 #endif 3516 return false; // Reached end of inlining 3517 } 3518 3519 // Acquire method holder as java.lang.Class 3520 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 3521 ciInstance* caller_mirror = caller_klass->java_mirror(); 3522 // Push this as a constant 3523 push(makecon(TypeInstPtr::make(caller_mirror))); 3524 #ifndef PRODUCT 3525 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3526 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); 3527 tty->print_cr(" JVM state at this point:"); 3528 for (int i = _depth; i >= 1; i--) { 3529 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3530 } 3531 } 3532 #endif 3533 return true; 3534 } 3535 3536 // Helper routine for above 3537 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { 3538 // Is this the Method.invoke method itself? 3539 if (jvms->method()->intrinsic_id() == vmIntrinsics::_invoke) 3540 return true; 3541 3542 // Is this a helper, defined somewhere underneath MethodAccessorImpl. 3543 ciKlass* k = jvms->method()->holder(); 3544 if (k->is_instance_klass()) { 3545 ciInstanceKlass* ik = k->as_instance_klass(); 3546 for (; ik != NULL; ik = ik->super()) { 3547 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && 3548 ik == env()->find_system_klass(ik->name())) { 3549 return true; 3550 } 3551 } 3552 } 3553 3554 return false; 3555 } 3556 3557 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by 3558 // inline_native_AtomicLong_attemptUpdate() but it has no way of 3559 // computing it since there is no lookup field by name function in the 3560 // CI interface. This is computed and set by inline_native_AtomicLong_get(). 3561 // Using a static variable here is safe even if we have multiple compilation 3562 // threads because the offset is constant. At worst the same offset will be 3563 // computed and stored multiple 3564 3565 bool LibraryCallKit::inline_native_AtomicLong_get() { 3566 // Restore the stack and pop off the argument 3567 _sp+=1; 3568 Node *obj = pop(); 3569 3570 // get the offset of the "value" field. Since the CI interfaces 3571 // does not provide a way to look up a field by name, we scan the bytecodes 3572 // to get the field index. We expect the first 2 instructions of the method 3573 // to be: 3574 // 0 aload_0 3575 // 1 getfield "value" 3576 ciMethod* method = callee(); 3577 if (value_field_offset == -1) 3578 { 3579 ciField* value_field; 3580 ciBytecodeStream iter(method); 3581 Bytecodes::Code bc = iter.next(); 3582 3583 if ((bc != Bytecodes::_aload_0) && 3584 ((bc != Bytecodes::_aload) || (iter.get_index() != 0))) 3585 return false; 3586 bc = iter.next(); 3587 if (bc != Bytecodes::_getfield) 3588 return false; 3589 bool ignore; 3590 value_field = iter.get_field(ignore); 3591 value_field_offset = value_field->offset_in_bytes(); 3592 } 3593 3594 // Null check without removing any arguments. 3595 _sp++; 3596 obj = do_null_check(obj, T_OBJECT); 3597 _sp--; 3598 // Check for locking null object 3599 if (stopped()) return true; 3600 3601 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3602 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3603 int alias_idx = C->get_alias_index(adr_type); 3604 3605 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr)); 3606 3607 push_pair(result); 3608 3609 return true; 3610 } 3611 3612 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() { 3613 // Restore the stack and pop off the arguments 3614 _sp+=5; 3615 Node *newVal = pop_pair(); 3616 Node *oldVal = pop_pair(); 3617 Node *obj = pop(); 3618 3619 // we need the offset of the "value" field which was computed when 3620 // inlining the get() method. Give up if we don't have it. 3621 if (value_field_offset == -1) 3622 return false; 3623 3624 // Null check without removing any arguments. 3625 _sp+=5; 3626 obj = do_null_check(obj, T_OBJECT); 3627 _sp-=5; 3628 // Check for locking null object 3629 if (stopped()) return true; 3630 3631 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3632 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3633 int alias_idx = C->get_alias_index(adr_type); 3634 3635 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal)); 3636 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 3637 set_memory(store_proj, alias_idx); 3638 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) ); 3639 3640 Node *result; 3641 // CMove node is not used to be able fold a possible check code 3642 // after attemptUpdate() call. This code could be transformed 3643 // into CMove node by loop optimizations. 3644 { 3645 RegionNode *r = new (C, 3) RegionNode(3); 3646 result = new (C, 3) PhiNode(r, TypeInt::BOOL); 3647 3648 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 3649 Node *iftrue = opt_iff(r, iff); 3650 r->init_req(1, iftrue); 3651 result->init_req(1, intcon(1)); 3652 result->init_req(2, intcon(0)); 3653 3654 set_control(_gvn.transform(r)); 3655 record_for_igvn(r); 3656 3657 C->set_has_split_ifs(true); // Has chance for split-if optimization 3658 } 3659 3660 push(_gvn.transform(result)); 3661 return true; 3662 } 3663 3664 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 3665 // restore the arguments 3666 _sp += arg_size(); 3667 3668 switch (id) { 3669 case vmIntrinsics::_floatToRawIntBits: 3670 push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); 3671 break; 3672 3673 case vmIntrinsics::_intBitsToFloat: 3674 push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); 3675 break; 3676 3677 case vmIntrinsics::_doubleToRawLongBits: 3678 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); 3679 break; 3680 3681 case vmIntrinsics::_longBitsToDouble: 3682 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); 3683 break; 3684 3685 case vmIntrinsics::_doubleToLongBits: { 3686 Node* value = pop_pair(); 3687 3688 // two paths (plus control) merge in a wood 3689 RegionNode *r = new (C, 3) RegionNode(3); 3690 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); 3691 3692 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); 3693 // Build the boolean node 3694 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 3695 3696 // Branch either way. 3697 // NaN case is less traveled, which makes all the difference. 3698 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3699 Node *opt_isnan = _gvn.transform(ifisnan); 3700 assert( opt_isnan->is_If(), "Expect an IfNode"); 3701 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3702 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 3703 3704 set_control(iftrue); 3705 3706 static const jlong nan_bits = CONST64(0x7ff8000000000000); 3707 Node *slow_result = longcon(nan_bits); // return NaN 3708 phi->init_req(1, _gvn.transform( slow_result )); 3709 r->init_req(1, iftrue); 3710 3711 // Else fall through 3712 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 3713 set_control(iffalse); 3714 3715 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); 3716 r->init_req(2, iffalse); 3717 3718 // Post merge 3719 set_control(_gvn.transform(r)); 3720 record_for_igvn(r); 3721 3722 Node* result = _gvn.transform(phi); 3723 assert(result->bottom_type()->isa_long(), "must be"); 3724 push_pair(result); 3725 3726 C->set_has_split_ifs(true); // Has chance for split-if optimization 3727 3728 break; 3729 } 3730 3731 case vmIntrinsics::_floatToIntBits: { 3732 Node* value = pop(); 3733 3734 // two paths (plus control) merge in a wood 3735 RegionNode *r = new (C, 3) RegionNode(3); 3736 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); 3737 3738 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); 3739 // Build the boolean node 3740 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 3741 3742 // Branch either way. 3743 // NaN case is less traveled, which makes all the difference. 3744 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3745 Node *opt_isnan = _gvn.transform(ifisnan); 3746 assert( opt_isnan->is_If(), "Expect an IfNode"); 3747 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3748 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 3749 3750 set_control(iftrue); 3751 3752 static const jint nan_bits = 0x7fc00000; 3753 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 3754 phi->init_req(1, _gvn.transform( slow_result )); 3755 r->init_req(1, iftrue); 3756 3757 // Else fall through 3758 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 3759 set_control(iffalse); 3760 3761 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); 3762 r->init_req(2, iffalse); 3763 3764 // Post merge 3765 set_control(_gvn.transform(r)); 3766 record_for_igvn(r); 3767 3768 Node* result = _gvn.transform(phi); 3769 assert(result->bottom_type()->isa_int(), "must be"); 3770 push(result); 3771 3772 C->set_has_split_ifs(true); // Has chance for split-if optimization 3773 3774 break; 3775 } 3776 3777 default: 3778 ShouldNotReachHere(); 3779 } 3780 3781 return true; 3782 } 3783 3784 #ifdef _LP64 3785 #define XTOP ,top() /*additional argument*/ 3786 #else //_LP64 3787 #define XTOP /*no additional argument*/ 3788 #endif //_LP64 3789 3790 //----------------------inline_unsafe_copyMemory------------------------- 3791 bool LibraryCallKit::inline_unsafe_copyMemory() { 3792 if (callee()->is_static()) return false; // caller must have the capability! 3793 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) 3794 assert(signature()->size() == nargs-1, "copy has 5 arguments"); 3795 null_check_receiver(callee()); // check then ignore argument(0) 3796 if (stopped()) return true; 3797 3798 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3799 3800 Node* src_ptr = argument(1); 3801 Node* src_off = ConvL2X(argument(2)); 3802 assert(argument(3)->is_top(), "2nd half of long"); 3803 Node* dst_ptr = argument(4); 3804 Node* dst_off = ConvL2X(argument(5)); 3805 assert(argument(6)->is_top(), "2nd half of long"); 3806 Node* size = ConvL2X(argument(7)); 3807 assert(argument(8)->is_top(), "2nd half of long"); 3808 3809 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 3810 "fieldOffset must be byte-scaled"); 3811 3812 Node* src = make_unsafe_address(src_ptr, src_off); 3813 Node* dst = make_unsafe_address(dst_ptr, dst_off); 3814 3815 // Conservatively insert a memory barrier on all memory slices. 3816 // Do not let writes of the copy source or destination float below the copy. 3817 insert_mem_bar(Op_MemBarCPUOrder); 3818 3819 // Call it. Note that the length argument is not scaled. 3820 make_runtime_call(RC_LEAF|RC_NO_FP, 3821 OptoRuntime::fast_arraycopy_Type(), 3822 StubRoutines::unsafe_arraycopy(), 3823 "unsafe_arraycopy", 3824 TypeRawPtr::BOTTOM, 3825 src, dst, size XTOP); 3826 3827 // Do not let reads of the copy destination float above the copy. 3828 insert_mem_bar(Op_MemBarCPUOrder); 3829 3830 return true; 3831 } 3832 3833 3834 //------------------------inline_native_clone---------------------------- 3835 // Here are the simple edge cases: 3836 // null receiver => normal trap 3837 // virtual and clone was overridden => slow path to out-of-line clone 3838 // not cloneable or finalizer => slow path to out-of-line Object.clone 3839 // 3840 // The general case has two steps, allocation and copying. 3841 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 3842 // 3843 // Copying also has two cases, oop arrays and everything else. 3844 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 3845 // Everything else uses the tight inline loop supplied by CopyArrayNode. 3846 // 3847 // These steps fold up nicely if and when the cloned object's klass 3848 // can be sharply typed as an object array, a type array, or an instance. 3849 // 3850 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 3851 int nargs = 1; 3852 Node* obj = null_check_receiver(callee()); 3853 if (stopped()) return true; 3854 Node* obj_klass = load_object_klass(obj); 3855 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 3856 const TypeOopPtr* toop = ((tklass != NULL) 3857 ? tklass->as_instance_type() 3858 : TypeInstPtr::NOTNULL); 3859 3860 // Conservatively insert a memory barrier on all memory slices. 3861 // Do not let writes into the original float below the clone. 3862 insert_mem_bar(Op_MemBarCPUOrder); 3863 3864 // paths into result_reg: 3865 enum { 3866 _slow_path = 1, // out-of-line call to clone method (virtual or not) 3867 _objArray_path, // plain allocation, plus arrayof_oop_arraycopy 3868 _fast_path, // plain allocation, plus a CopyArray operation 3869 PATH_LIMIT 3870 }; 3871 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3872 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3873 TypeInstPtr::NOTNULL); 3874 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3875 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3876 TypePtr::BOTTOM); 3877 record_for_igvn(result_reg); 3878 3879 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 3880 int raw_adr_idx = Compile::AliasIdxRaw; 3881 const bool raw_mem_only = true; 3882 3883 // paths into alloc_reg (on the fast path, just before the CopyArray): 3884 enum { _typeArray_alloc = 1, _instance_alloc, ALLOC_LIMIT }; 3885 RegionNode* alloc_reg = new(C, ALLOC_LIMIT) RegionNode(ALLOC_LIMIT); 3886 PhiNode* alloc_val = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, raw_adr_type); 3887 PhiNode* alloc_siz = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, TypeX_X); 3888 PhiNode* alloc_i_o = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::ABIO); 3889 PhiNode* alloc_mem = new(C, ALLOC_LIMIT) PhiNode(alloc_reg, Type::MEMORY, 3890 raw_adr_type); 3891 record_for_igvn(alloc_reg); 3892 3893 bool card_mark = false; // (see below) 3894 3895 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 3896 if (array_ctl != NULL) { 3897 // It's an array. 3898 PreserveJVMState pjvms(this); 3899 set_control(array_ctl); 3900 Node* obj_length = load_array_length(obj); 3901 Node* obj_size = NULL; 3902 Node* alloc_obj = new_array(obj_klass, obj_length, nargs, 3903 raw_mem_only, &obj_size); 3904 assert(obj_size != NULL, ""); 3905 Node* raw_obj = alloc_obj->in(1); 3906 assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 3907 if (ReduceBulkZeroing) { 3908 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 3909 if (alloc != NULL) { 3910 // We will be completely responsible for initializing this object. 3911 alloc->maybe_set_complete(&_gvn); 3912 } 3913 } 3914 3915 if (!use_ReduceInitialCardMarks()) { 3916 // If it is an oop array, it requires very special treatment, 3917 // because card marking is required on each card of the array. 3918 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 3919 if (is_obja != NULL) { 3920 PreserveJVMState pjvms2(this); 3921 set_control(is_obja); 3922 // Generate a direct call to the right arraycopy function(s). 3923 bool disjoint_bases = true; 3924 bool length_never_negative = true; 3925 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3926 obj, intcon(0), alloc_obj, intcon(0), 3927 obj_length, nargs, 3928 disjoint_bases, length_never_negative); 3929 result_reg->init_req(_objArray_path, control()); 3930 result_val->init_req(_objArray_path, alloc_obj); 3931 result_i_o ->set_req(_objArray_path, i_o()); 3932 result_mem ->set_req(_objArray_path, reset_memory()); 3933 } 3934 } 3935 // We can dispense with card marks if we know the allocation 3936 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 3937 // causes the non-eden paths to simulate a fresh allocation, 3938 // insofar that no further card marks are required to initialize 3939 // the object. 3940 3941 // Otherwise, there are no card marks to worry about. 3942 alloc_val->init_req(_typeArray_alloc, raw_obj); 3943 alloc_siz->init_req(_typeArray_alloc, obj_size); 3944 alloc_reg->init_req(_typeArray_alloc, control()); 3945 alloc_i_o->init_req(_typeArray_alloc, i_o()); 3946 alloc_mem->init_req(_typeArray_alloc, memory(raw_adr_type)); 3947 } 3948 3949 // We only go to the fast case code if we pass a number of guards. 3950 // The paths which do not pass are accumulated in the slow_region. 3951 RegionNode* slow_region = new (C, 1) RegionNode(1); 3952 record_for_igvn(slow_region); 3953 if (!stopped()) { 3954 // It's an instance. Make the slow-path tests. 3955 // If this is a virtual call, we generate a funny guard. We grab 3956 // the vtable entry corresponding to clone() from the target object. 3957 // If the target method which we are calling happens to be the 3958 // Object clone() method, we pass the guard. We do not need this 3959 // guard for non-virtual calls; the caller is known to be the native 3960 // Object clone(). 3961 if (is_virtual) { 3962 generate_virtual_guard(obj_klass, slow_region); 3963 } 3964 3965 // The object must be cloneable and must not have a finalizer. 3966 // Both of these conditions may be checked in a single test. 3967 // We could optimize the cloneable test further, but we don't care. 3968 generate_access_flags_guard(obj_klass, 3969 // Test both conditions: 3970 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 3971 // Must be cloneable but not finalizer: 3972 JVM_ACC_IS_CLONEABLE, 3973 slow_region); 3974 } 3975 3976 if (!stopped()) { 3977 // It's an instance, and it passed the slow-path tests. 3978 PreserveJVMState pjvms(this); 3979 Node* obj_size = NULL; 3980 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size); 3981 assert(obj_size != NULL, ""); 3982 Node* raw_obj = alloc_obj->in(1); 3983 assert(raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 3984 if (ReduceBulkZeroing) { 3985 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 3986 if (alloc != NULL && !alloc->maybe_set_complete(&_gvn)) 3987 alloc = NULL; 3988 } 3989 if (!use_ReduceInitialCardMarks()) { 3990 // Put in store barrier for any and all oops we are sticking 3991 // into this object. (We could avoid this if we could prove 3992 // that the object type contains no oop fields at all.) 3993 card_mark = true; 3994 } 3995 alloc_val->init_req(_instance_alloc, raw_obj); 3996 alloc_siz->init_req(_instance_alloc, obj_size); 3997 alloc_reg->init_req(_instance_alloc, control()); 3998 alloc_i_o->init_req(_instance_alloc, i_o()); 3999 alloc_mem->init_req(_instance_alloc, memory(raw_adr_type)); 4000 } 4001 4002 // Generate code for the slow case. We make a call to clone(). 4003 set_control(_gvn.transform(slow_region)); 4004 if (!stopped()) { 4005 PreserveJVMState pjvms(this); 4006 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4007 Node* slow_result = set_results_for_java_call(slow_call); 4008 // this->control() comes from set_results_for_java_call 4009 result_reg->init_req(_slow_path, control()); 4010 result_val->init_req(_slow_path, slow_result); 4011 result_i_o ->set_req(_slow_path, i_o()); 4012 result_mem ->set_req(_slow_path, reset_memory()); 4013 } 4014 4015 // The object is allocated, as an array and/or an instance. Now copy it. 4016 set_control( _gvn.transform(alloc_reg) ); 4017 set_i_o( _gvn.transform(alloc_i_o) ); 4018 set_memory( _gvn.transform(alloc_mem), raw_adr_type ); 4019 Node* raw_obj = _gvn.transform(alloc_val); 4020 4021 if (!stopped()) { 4022 // Copy the fastest available way. 4023 // (No need for PreserveJVMState, since we're using it all up now.) 4024 // TODO: generate fields/elements copies for small objects instead. 4025 Node* src = obj; 4026 Node* dest = raw_obj; 4027 Node* size = _gvn.transform(alloc_siz); 4028 4029 // Exclude the header. 4030 int base_off = instanceOopDesc::base_offset_in_bytes(); 4031 if (UseCompressedOops) { 4032 assert(base_off % BytesPerLong != 0, "base with compressed oops"); 4033 // With compressed oops base_offset_in_bytes is 12 which creates 4034 // the gap since countx is rounded by 8 bytes below. 4035 // Copy klass and the gap. 4036 base_off = instanceOopDesc::klass_offset_in_bytes(); 4037 } 4038 src = basic_plus_adr(src, base_off); 4039 dest = basic_plus_adr(dest, base_off); 4040 4041 // Compute the length also, if needed: 4042 Node* countx = size; 4043 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); 4044 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4045 4046 // Select an appropriate instruction to initialize the range. 4047 // The CopyArray instruction (if supported) can be optimized 4048 // into a discrete set of scalar loads and stores. 4049 bool disjoint_bases = true; 4050 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4051 src, NULL, dest, NULL, countx); 4052 4053 // Now that the object is properly initialized, type it as an oop. 4054 // Use a secondary InitializeNode memory barrier. 4055 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, raw_adr_idx, 4056 raw_obj)->as_Initialize(); 4057 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4058 Node* new_obj = new(C, 2) CheckCastPPNode(control(), raw_obj, 4059 TypeInstPtr::NOTNULL); 4060 new_obj = _gvn.transform(new_obj); 4061 4062 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4063 if (card_mark) { 4064 Node* no_particular_value = NULL; 4065 Node* no_particular_field = NULL; 4066 post_barrier(control(), 4067 memory(raw_adr_type), 4068 new_obj, 4069 no_particular_field, 4070 raw_adr_idx, 4071 no_particular_value, 4072 T_OBJECT, 4073 false); 4074 } 4075 // Present the results of the slow call. 4076 result_reg->init_req(_fast_path, control()); 4077 result_val->init_req(_fast_path, new_obj); 4078 result_i_o ->set_req(_fast_path, i_o()); 4079 result_mem ->set_req(_fast_path, reset_memory()); 4080 } 4081 4082 // Return the combined state. 4083 set_control( _gvn.transform(result_reg) ); 4084 set_i_o( _gvn.transform(result_i_o) ); 4085 set_all_memory( _gvn.transform(result_mem) ); 4086 4087 // Cast the result to a sharper type, since we know what clone does. 4088 Node* new_obj = _gvn.transform(result_val); 4089 Node* cast = new (C, 2) CheckCastPPNode(control(), new_obj, toop); 4090 push(_gvn.transform(cast)); 4091 4092 return true; 4093 } 4094 4095 4096 // constants for computing the copy function 4097 enum { 4098 COPYFUNC_UNALIGNED = 0, 4099 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize 4100 COPYFUNC_CONJOINT = 0, 4101 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend 4102 }; 4103 4104 // Note: The condition "disjoint" applies also for overlapping copies 4105 // where an descending copy is permitted (i.e., dest_offset <= src_offset). 4106 static address 4107 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) { 4108 int selector = 4109 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) + 4110 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT); 4111 4112 #define RETURN_STUB(xxx_arraycopy) { \ 4113 name = #xxx_arraycopy; \ 4114 return StubRoutines::xxx_arraycopy(); } 4115 4116 switch (t) { 4117 case T_BYTE: 4118 case T_BOOLEAN: 4119 switch (selector) { 4120 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy); 4121 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy); 4122 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy); 4123 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy); 4124 } 4125 case T_CHAR: 4126 case T_SHORT: 4127 switch (selector) { 4128 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy); 4129 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy); 4130 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy); 4131 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy); 4132 } 4133 case T_INT: 4134 case T_FLOAT: 4135 switch (selector) { 4136 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy); 4137 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy); 4138 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy); 4139 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy); 4140 } 4141 case T_DOUBLE: 4142 case T_LONG: 4143 switch (selector) { 4144 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy); 4145 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy); 4146 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy); 4147 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy); 4148 } 4149 case T_ARRAY: 4150 case T_OBJECT: 4151 switch (selector) { 4152 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy); 4153 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy); 4154 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy); 4155 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy); 4156 } 4157 default: 4158 ShouldNotReachHere(); 4159 return NULL; 4160 } 4161 4162 #undef RETURN_STUB 4163 } 4164 4165 //------------------------------basictype2arraycopy---------------------------- 4166 address LibraryCallKit::basictype2arraycopy(BasicType t, 4167 Node* src_offset, 4168 Node* dest_offset, 4169 bool disjoint_bases, 4170 const char* &name) { 4171 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4172 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4173 4174 bool aligned = false; 4175 bool disjoint = disjoint_bases; 4176 4177 // if the offsets are the same, we can treat the memory regions as 4178 // disjoint, because either the memory regions are in different arrays, 4179 // or they are identical (which we can treat as disjoint.) We can also 4180 // treat a copy with a destination index less that the source index 4181 // as disjoint since a low->high copy will work correctly in this case. 4182 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4183 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4184 // both indices are constants 4185 int s_offs = src_offset_inttype->get_con(); 4186 int d_offs = dest_offset_inttype->get_con(); 4187 int element_size = type2aelembytes(t); 4188 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4189 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4190 if (s_offs >= d_offs) disjoint = true; 4191 } else if (src_offset == dest_offset && src_offset != NULL) { 4192 // This can occur if the offsets are identical non-constants. 4193 disjoint = true; 4194 } 4195 4196 return select_arraycopy_function(t, aligned, disjoint, name); 4197 } 4198 4199 4200 //------------------------------inline_arraycopy----------------------- 4201 bool LibraryCallKit::inline_arraycopy() { 4202 // Restore the stack and pop off the arguments. 4203 int nargs = 5; // 2 oops, 3 ints, no size_t or long 4204 assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); 4205 4206 Node *src = argument(0); 4207 Node *src_offset = argument(1); 4208 Node *dest = argument(2); 4209 Node *dest_offset = argument(3); 4210 Node *length = argument(4); 4211 4212 // Compile time checks. If any of these checks cannot be verified at compile time, 4213 // we do not make a fast path for this call. Instead, we let the call remain as it 4214 // is. The checks we choose to mandate at compile time are: 4215 // 4216 // (1) src and dest are arrays. 4217 const Type* src_type = src->Value(&_gvn); 4218 const Type* dest_type = dest->Value(&_gvn); 4219 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4220 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4221 if (top_src == NULL || top_src->klass() == NULL || 4222 top_dest == NULL || top_dest->klass() == NULL) { 4223 // Conservatively insert a memory barrier on all memory slices. 4224 // Do not let writes into the source float below the arraycopy. 4225 insert_mem_bar(Op_MemBarCPUOrder); 4226 4227 // Call StubRoutines::generic_arraycopy stub. 4228 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4229 src, src_offset, dest, dest_offset, length, 4230 nargs); 4231 4232 // Do not let reads from the destination float above the arraycopy. 4233 // Since we cannot type the arrays, we don't know which slices 4234 // might be affected. We could restrict this barrier only to those 4235 // memory slices which pertain to array elements--but don't bother. 4236 if (!InsertMemBarAfterArraycopy) 4237 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4238 insert_mem_bar(Op_MemBarCPUOrder); 4239 return true; 4240 } 4241 4242 // (2) src and dest arrays must have elements of the same BasicType 4243 // Figure out the size and type of the elements we will be copying. 4244 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4245 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4246 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4247 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4248 4249 if (src_elem != dest_elem || dest_elem == T_VOID) { 4250 // The component types are not the same or are not recognized. Punt. 4251 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4252 generate_slow_arraycopy(TypePtr::BOTTOM, 4253 src, src_offset, dest, dest_offset, length, 4254 nargs); 4255 return true; 4256 } 4257 4258 //--------------------------------------------------------------------------- 4259 // We will make a fast path for this call to arraycopy. 4260 4261 // We have the following tests left to perform: 4262 // 4263 // (3) src and dest must not be null. 4264 // (4) src_offset must not be negative. 4265 // (5) dest_offset must not be negative. 4266 // (6) length must not be negative. 4267 // (7) src_offset + length must not exceed length of src. 4268 // (8) dest_offset + length must not exceed length of dest. 4269 // (9) each element of an oop array must be assignable 4270 4271 RegionNode* slow_region = new (C, 1) RegionNode(1); 4272 record_for_igvn(slow_region); 4273 4274 // (3) operands must not be null 4275 // We currently perform our null checks with the do_null_check routine. 4276 // This means that the null exceptions will be reported in the caller 4277 // rather than (correctly) reported inside of the native arraycopy call. 4278 // This should be corrected, given time. We do our null check with the 4279 // stack pointer restored. 4280 _sp += nargs; 4281 src = do_null_check(src, T_ARRAY); 4282 dest = do_null_check(dest, T_ARRAY); 4283 _sp -= nargs; 4284 4285 // (4) src_offset must not be negative. 4286 generate_negative_guard(src_offset, slow_region); 4287 4288 // (5) dest_offset must not be negative. 4289 generate_negative_guard(dest_offset, slow_region); 4290 4291 // (6) length must not be negative (moved to generate_arraycopy()). 4292 // generate_negative_guard(length, slow_region); 4293 4294 // (7) src_offset + length must not exceed length of src. 4295 generate_limit_guard(src_offset, length, 4296 load_array_length(src), 4297 slow_region); 4298 4299 // (8) dest_offset + length must not exceed length of dest. 4300 generate_limit_guard(dest_offset, length, 4301 load_array_length(dest), 4302 slow_region); 4303 4304 // (9) each element of an oop array must be assignable 4305 // The generate_arraycopy subroutine checks this. 4306 4307 // This is where the memory effects are placed: 4308 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4309 generate_arraycopy(adr_type, dest_elem, 4310 src, src_offset, dest, dest_offset, length, 4311 nargs, false, false, slow_region); 4312 4313 return true; 4314 } 4315 4316 //-----------------------------generate_arraycopy---------------------- 4317 // Generate an optimized call to arraycopy. 4318 // Caller must guard against non-arrays. 4319 // Caller must determine a common array basic-type for both arrays. 4320 // Caller must validate offsets against array bounds. 4321 // The slow_region has already collected guard failure paths 4322 // (such as out of bounds length or non-conformable array types). 4323 // The generated code has this shape, in general: 4324 // 4325 // if (length == 0) return // via zero_path 4326 // slowval = -1 4327 // if (types unknown) { 4328 // slowval = call generic copy loop 4329 // if (slowval == 0) return // via checked_path 4330 // } else if (indexes in bounds) { 4331 // if ((is object array) && !(array type check)) { 4332 // slowval = call checked copy loop 4333 // if (slowval == 0) return // via checked_path 4334 // } else { 4335 // call bulk copy loop 4336 // return // via fast_path 4337 // } 4338 // } 4339 // // adjust params for remaining work: 4340 // if (slowval != -1) { 4341 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4342 // } 4343 // slow_region: 4344 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4345 // return // via slow_call_path 4346 // 4347 // This routine is used from several intrinsics: System.arraycopy, 4348 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4349 // 4350 void 4351 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4352 BasicType basic_elem_type, 4353 Node* src, Node* src_offset, 4354 Node* dest, Node* dest_offset, 4355 Node* copy_length, 4356 int nargs, 4357 bool disjoint_bases, 4358 bool length_never_negative, 4359 RegionNode* slow_region) { 4360 4361 if (slow_region == NULL) { 4362 slow_region = new(C,1) RegionNode(1); 4363 record_for_igvn(slow_region); 4364 } 4365 4366 Node* original_dest = dest; 4367 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4368 Node* raw_dest = NULL; // used before zeroing, if needed 4369 bool must_clear_dest = false; 4370 4371 // See if this is the initialization of a newly-allocated array. 4372 // If so, we will take responsibility here for initializing it to zero. 4373 // (Note: Because tightly_coupled_allocation performs checks on the 4374 // out-edges of the dest, we need to avoid making derived pointers 4375 // from it until we have checked its uses.) 4376 if (ReduceBulkZeroing 4377 && !ZeroTLAB // pointless if already zeroed 4378 && basic_elem_type != T_CONFLICT // avoid corner case 4379 && !_gvn.eqv_uncast(src, dest) 4380 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4381 != NULL) 4382 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4383 && alloc->maybe_set_complete(&_gvn)) { 4384 // "You break it, you buy it." 4385 InitializeNode* init = alloc->initialization(); 4386 assert(init->is_complete(), "we just did this"); 4387 assert(dest->Opcode() == Op_CheckCastPP, "sanity"); 4388 assert(dest->in(0)->in(0) == init, "dest pinned"); 4389 raw_dest = dest->in(1); // grab the raw pointer! 4390 original_dest = dest; 4391 dest = raw_dest; 4392 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4393 // Decouple the original InitializeNode, turning it into a simple membar. 4394 // We will build a new one at the end of this routine. 4395 init->set_req(InitializeNode::RawAddress, top()); 4396 // From this point on, every exit path is responsible for 4397 // initializing any non-copied parts of the object to zero. 4398 must_clear_dest = true; 4399 } else { 4400 // No zeroing elimination here. 4401 alloc = NULL; 4402 //original_dest = dest; 4403 //must_clear_dest = false; 4404 } 4405 4406 // Results are placed here: 4407 enum { fast_path = 1, // normal void-returning assembly stub 4408 checked_path = 2, // special assembly stub with cleanup 4409 slow_call_path = 3, // something went wrong; call the VM 4410 zero_path = 4, // bypass when length of copy is zero 4411 bcopy_path = 5, // copy primitive array by 64-bit blocks 4412 PATH_LIMIT = 6 4413 }; 4414 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4415 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO); 4416 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type); 4417 record_for_igvn(result_region); 4418 _gvn.set_type_bottom(result_i_o); 4419 _gvn.set_type_bottom(result_memory); 4420 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4421 4422 // The slow_control path: 4423 Node* slow_control; 4424 Node* slow_i_o = i_o(); 4425 Node* slow_mem = memory(adr_type); 4426 debug_only(slow_control = (Node*) badAddress); 4427 4428 // Checked control path: 4429 Node* checked_control = top(); 4430 Node* checked_mem = NULL; 4431 Node* checked_i_o = NULL; 4432 Node* checked_value = NULL; 4433 4434 if (basic_elem_type == T_CONFLICT) { 4435 assert(!must_clear_dest, ""); 4436 Node* cv = generate_generic_arraycopy(adr_type, 4437 src, src_offset, dest, dest_offset, 4438 copy_length, nargs); 4439 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4440 checked_control = control(); 4441 checked_i_o = i_o(); 4442 checked_mem = memory(adr_type); 4443 checked_value = cv; 4444 set_control(top()); // no fast path 4445 } 4446 4447 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4448 if (not_pos != NULL) { 4449 PreserveJVMState pjvms(this); 4450 set_control(not_pos); 4451 4452 // (6) length must not be negative. 4453 if (!length_never_negative) { 4454 generate_negative_guard(copy_length, slow_region); 4455 } 4456 4457 if (!stopped() && must_clear_dest) { 4458 Node* dest_length = alloc->in(AllocateNode::ALength); 4459 if (_gvn.eqv_uncast(copy_length, dest_length) 4460 || _gvn.find_int_con(dest_length, 1) <= 0) { 4461 // There is no zeroing to do. 4462 } else { 4463 // Clear the whole thing since there are no source elements to copy. 4464 generate_clear_array(adr_type, dest, basic_elem_type, 4465 intcon(0), NULL, 4466 alloc->in(AllocateNode::AllocSize)); 4467 } 4468 } 4469 4470 // Present the results of the fast call. 4471 result_region->init_req(zero_path, control()); 4472 result_i_o ->init_req(zero_path, i_o()); 4473 result_memory->init_req(zero_path, memory(adr_type)); 4474 } 4475 4476 if (!stopped() && must_clear_dest) { 4477 // We have to initialize the *uncopied* part of the array to zero. 4478 // The copy destination is the slice dest[off..off+len]. The other slices 4479 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4480 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4481 Node* dest_length = alloc->in(AllocateNode::ALength); 4482 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset, 4483 copy_length) ); 4484 4485 // If there is a head section that needs zeroing, do it now. 4486 if (find_int_con(dest_offset, -1) != 0) { 4487 generate_clear_array(adr_type, dest, basic_elem_type, 4488 intcon(0), dest_offset, 4489 NULL); 4490 } 4491 4492 // Next, perform a dynamic check on the tail length. 4493 // It is often zero, and we can win big if we prove this. 4494 // There are two wins: Avoid generating the ClearArray 4495 // with its attendant messy index arithmetic, and upgrade 4496 // the copy to a more hardware-friendly word size of 64 bits. 4497 Node* tail_ctl = NULL; 4498 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) { 4499 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) ); 4500 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) ); 4501 tail_ctl = generate_slow_guard(bol_lt, NULL); 4502 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4503 } 4504 4505 // At this point, let's assume there is no tail. 4506 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4507 // There is no tail. Try an upgrade to a 64-bit copy. 4508 bool didit = false; 4509 { PreserveJVMState pjvms(this); 4510 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4511 src, src_offset, dest, dest_offset, 4512 dest_size); 4513 if (didit) { 4514 // Present the results of the block-copying fast call. 4515 result_region->init_req(bcopy_path, control()); 4516 result_i_o ->init_req(bcopy_path, i_o()); 4517 result_memory->init_req(bcopy_path, memory(adr_type)); 4518 } 4519 } 4520 if (didit) 4521 set_control(top()); // no regular fast path 4522 } 4523 4524 // Clear the tail, if any. 4525 if (tail_ctl != NULL) { 4526 Node* notail_ctl = stopped() ? NULL : control(); 4527 set_control(tail_ctl); 4528 if (notail_ctl == NULL) { 4529 generate_clear_array(adr_type, dest, basic_elem_type, 4530 dest_tail, NULL, 4531 dest_size); 4532 } else { 4533 // Make a local merge. 4534 Node* done_ctl = new(C,3) RegionNode(3); 4535 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type); 4536 done_ctl->init_req(1, notail_ctl); 4537 done_mem->init_req(1, memory(adr_type)); 4538 generate_clear_array(adr_type, dest, basic_elem_type, 4539 dest_tail, NULL, 4540 dest_size); 4541 done_ctl->init_req(2, control()); 4542 done_mem->init_req(2, memory(adr_type)); 4543 set_control( _gvn.transform(done_ctl) ); 4544 set_memory( _gvn.transform(done_mem), adr_type ); 4545 } 4546 } 4547 } 4548 4549 BasicType copy_type = basic_elem_type; 4550 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 4551 if (!stopped() && copy_type == T_OBJECT) { 4552 // If src and dest have compatible element types, we can copy bits. 4553 // Types S[] and D[] are compatible if D is a supertype of S. 4554 // 4555 // If they are not, we will use checked_oop_disjoint_arraycopy, 4556 // which performs a fast optimistic per-oop check, and backs off 4557 // further to JVM_ArrayCopy on the first per-oop check that fails. 4558 // (Actually, we don't move raw bits only; the GC requires card marks.) 4559 4560 // Get the klassOop for both src and dest 4561 Node* src_klass = load_object_klass(src); 4562 Node* dest_klass = load_object_klass(dest); 4563 4564 // Generate the subtype check. 4565 // This might fold up statically, or then again it might not. 4566 // 4567 // Non-static example: Copying List<String>.elements to a new String[]. 4568 // The backing store for a List<String> is always an Object[], 4569 // but its elements are always type String, if the generic types 4570 // are correct at the source level. 4571 // 4572 // Test S[] against D[], not S against D, because (probably) 4573 // the secondary supertype cache is less busy for S[] than S. 4574 // This usually only matters when D is an interface. 4575 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4576 // Plug failing path into checked_oop_disjoint_arraycopy 4577 if (not_subtype_ctrl != top()) { 4578 PreserveJVMState pjvms(this); 4579 set_control(not_subtype_ctrl); 4580 // (At this point we can assume disjoint_bases, since types differ.) 4581 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc); 4582 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 4583 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 4584 Node* dest_elem_klass = _gvn.transform(n1); 4585 Node* cv = generate_checkcast_arraycopy(adr_type, 4586 dest_elem_klass, 4587 src, src_offset, dest, dest_offset, 4588 copy_length, 4589 nargs); 4590 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4591 checked_control = control(); 4592 checked_i_o = i_o(); 4593 checked_mem = memory(adr_type); 4594 checked_value = cv; 4595 } 4596 // At this point we know we do not need type checks on oop stores. 4597 4598 // Let's see if we need card marks: 4599 if (alloc != NULL && use_ReduceInitialCardMarks()) { 4600 // If we do not need card marks, copy using the jint or jlong stub. 4601 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 4602 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 4603 "sizes agree"); 4604 } 4605 } 4606 4607 if (!stopped()) { 4608 // Generate the fast path, if possible. 4609 PreserveJVMState pjvms(this); 4610 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 4611 src, src_offset, dest, dest_offset, 4612 ConvI2X(copy_length)); 4613 4614 // Present the results of the fast call. 4615 result_region->init_req(fast_path, control()); 4616 result_i_o ->init_req(fast_path, i_o()); 4617 result_memory->init_req(fast_path, memory(adr_type)); 4618 } 4619 4620 // Here are all the slow paths up to this point, in one bundle: 4621 slow_control = top(); 4622 if (slow_region != NULL) 4623 slow_control = _gvn.transform(slow_region); 4624 debug_only(slow_region = (RegionNode*)badAddress); 4625 4626 set_control(checked_control); 4627 if (!stopped()) { 4628 // Clean up after the checked call. 4629 // The returned value is either 0 or -1^K, 4630 // where K = number of partially transferred array elements. 4631 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) ); 4632 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 4633 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4634 4635 // If it is 0, we are done, so transfer to the end. 4636 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) ); 4637 result_region->init_req(checked_path, checks_done); 4638 result_i_o ->init_req(checked_path, checked_i_o); 4639 result_memory->init_req(checked_path, checked_mem); 4640 4641 // If it is not zero, merge into the slow call. 4642 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) )); 4643 RegionNode* slow_reg2 = new(C, 3) RegionNode(3); 4644 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO); 4645 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type); 4646 record_for_igvn(slow_reg2); 4647 slow_reg2 ->init_req(1, slow_control); 4648 slow_i_o2 ->init_req(1, slow_i_o); 4649 slow_mem2 ->init_req(1, slow_mem); 4650 slow_reg2 ->init_req(2, control()); 4651 slow_i_o2 ->init_req(2, i_o()); 4652 slow_mem2 ->init_req(2, memory(adr_type)); 4653 4654 slow_control = _gvn.transform(slow_reg2); 4655 slow_i_o = _gvn.transform(slow_i_o2); 4656 slow_mem = _gvn.transform(slow_mem2); 4657 4658 if (alloc != NULL) { 4659 // We'll restart from the very beginning, after zeroing the whole thing. 4660 // This can cause double writes, but that's OK since dest is brand new. 4661 // So we ignore the low 31 bits of the value returned from the stub. 4662 } else { 4663 // We must continue the copy exactly where it failed, or else 4664 // another thread might see the wrong number of writes to dest. 4665 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) ); 4666 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT); 4667 slow_offset->init_req(1, intcon(0)); 4668 slow_offset->init_req(2, checked_offset); 4669 slow_offset = _gvn.transform(slow_offset); 4670 4671 // Adjust the arguments by the conditionally incoming offset. 4672 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) ); 4673 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) ); 4674 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) ); 4675 4676 // Tweak the node variables to adjust the code produced below: 4677 src_offset = src_off_plus; 4678 dest_offset = dest_off_plus; 4679 copy_length = length_minus; 4680 } 4681 } 4682 4683 set_control(slow_control); 4684 if (!stopped()) { 4685 // Generate the slow path, if needed. 4686 PreserveJVMState pjvms(this); // replace_in_map may trash the map 4687 4688 set_memory(slow_mem, adr_type); 4689 set_i_o(slow_i_o); 4690 4691 if (must_clear_dest) { 4692 generate_clear_array(adr_type, dest, basic_elem_type, 4693 intcon(0), NULL, 4694 alloc->in(AllocateNode::AllocSize)); 4695 } 4696 4697 if (dest != original_dest) { 4698 // Promote from rawptr to oop, so it looks right in the call's GC map. 4699 dest = _gvn.transform( new(C,2) CheckCastPPNode(control(), dest, 4700 TypeInstPtr::NOTNULL) ); 4701 4702 // Edit the call's debug-info to avoid referring to original_dest. 4703 // (The problem with original_dest is that it isn't ready until 4704 // after the InitializeNode completes, but this stuff is before.) 4705 // Substitute in the locally valid dest_oop. 4706 replace_in_map(original_dest, dest); 4707 } 4708 4709 generate_slow_arraycopy(adr_type, 4710 src, src_offset, dest, dest_offset, 4711 copy_length, nargs); 4712 4713 result_region->init_req(slow_call_path, control()); 4714 result_i_o ->init_req(slow_call_path, i_o()); 4715 result_memory->init_req(slow_call_path, memory(adr_type)); 4716 } 4717 4718 // Remove unused edges. 4719 for (uint i = 1; i < result_region->req(); i++) { 4720 if (result_region->in(i) == NULL) 4721 result_region->init_req(i, top()); 4722 } 4723 4724 // Finished; return the combined state. 4725 set_control( _gvn.transform(result_region) ); 4726 set_i_o( _gvn.transform(result_i_o) ); 4727 set_memory( _gvn.transform(result_memory), adr_type ); 4728 4729 if (dest != original_dest) { 4730 // Pin the "finished" array node after the arraycopy/zeroing operations. 4731 // Use a secondary InitializeNode memory barrier. 4732 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4733 Compile::AliasIdxRaw, 4734 raw_dest)->as_Initialize(); 4735 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4736 _gvn.hash_delete(original_dest); 4737 original_dest->set_req(0, control()); 4738 _gvn.hash_find_insert(original_dest); // put back into GVN table 4739 } 4740 4741 // The memory edges above are precise in order to model effects around 4742 // array copies accurately to allow value numbering of field loads around 4743 // arraycopy. Such field loads, both before and after, are common in Java 4744 // collections and similar classes involving header/array data structures. 4745 // 4746 // But with low number of register or when some registers are used or killed 4747 // by arraycopy calls it causes registers spilling on stack. See 6544710. 4748 // The next memory barrier is added to avoid it. If the arraycopy can be 4749 // optimized away (which it can, sometimes) then we can manually remove 4750 // the membar also. 4751 if (InsertMemBarAfterArraycopy) 4752 insert_mem_bar(Op_MemBarCPUOrder); 4753 } 4754 4755 4756 // Helper function which determines if an arraycopy immediately follows 4757 // an allocation, with no intervening tests or other escapes for the object. 4758 AllocateArrayNode* 4759 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4760 RegionNode* slow_region) { 4761 if (stopped()) return NULL; // no fast path 4762 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4763 4764 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4765 if (alloc == NULL) return NULL; 4766 4767 Node* rawmem = memory(Compile::AliasIdxRaw); 4768 // Is the allocation's memory state untouched? 4769 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4770 // Bail out if there have been raw-memory effects since the allocation. 4771 // (Example: There might have been a call or safepoint.) 4772 return NULL; 4773 } 4774 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4775 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4776 return NULL; 4777 } 4778 4779 // There must be no unexpected observers of this allocation. 4780 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4781 Node* obs = ptr->fast_out(i); 4782 if (obs != this->map()) { 4783 return NULL; 4784 } 4785 } 4786 4787 // This arraycopy must unconditionally follow the allocation of the ptr. 4788 Node* alloc_ctl = ptr->in(0); 4789 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4790 4791 Node* ctl = control(); 4792 while (ctl != alloc_ctl) { 4793 // There may be guards which feed into the slow_region. 4794 // Any other control flow means that we might not get a chance 4795 // to finish initializing the allocated object. 4796 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4797 IfNode* iff = ctl->in(0)->as_If(); 4798 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 4799 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4800 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4801 ctl = iff->in(0); // This test feeds the known slow_region. 4802 continue; 4803 } 4804 // One more try: Various low-level checks bottom out in 4805 // uncommon traps. If the debug-info of the trap omits 4806 // any reference to the allocation, as we've already 4807 // observed, then there can be no objection to the trap. 4808 bool found_trap = false; 4809 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 4810 Node* obs = not_ctl->fast_out(j); 4811 if (obs->in(0) == not_ctl && obs->is_Call() && 4812 (obs->as_Call()->entry_point() == 4813 SharedRuntime::uncommon_trap_blob()->instructions_begin())) { 4814 found_trap = true; break; 4815 } 4816 } 4817 if (found_trap) { 4818 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 4819 continue; 4820 } 4821 } 4822 return NULL; 4823 } 4824 4825 // If we get this far, we have an allocation which immediately 4826 // precedes the arraycopy, and we can take over zeroing the new object. 4827 // The arraycopy will finish the initialization, and provide 4828 // a new control state to which we will anchor the destination pointer. 4829 4830 return alloc; 4831 } 4832 4833 // Helper for initialization of arrays, creating a ClearArray. 4834 // It writes zero bits in [start..end), within the body of an array object. 4835 // The memory effects are all chained onto the 'adr_type' alias category. 4836 // 4837 // Since the object is otherwise uninitialized, we are free 4838 // to put a little "slop" around the edges of the cleared area, 4839 // as long as it does not go back into the array's header, 4840 // or beyond the array end within the heap. 4841 // 4842 // The lower edge can be rounded down to the nearest jint and the 4843 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 4844 // 4845 // Arguments: 4846 // adr_type memory slice where writes are generated 4847 // dest oop of the destination array 4848 // basic_elem_type element type of the destination 4849 // slice_idx array index of first element to store 4850 // slice_len number of elements to store (or NULL) 4851 // dest_size total size in bytes of the array object 4852 // 4853 // Exactly one of slice_len or dest_size must be non-NULL. 4854 // If dest_size is non-NULL, zeroing extends to the end of the object. 4855 // If slice_len is non-NULL, the slice_idx value must be a constant. 4856 void 4857 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 4858 Node* dest, 4859 BasicType basic_elem_type, 4860 Node* slice_idx, 4861 Node* slice_len, 4862 Node* dest_size) { 4863 // one or the other but not both of slice_len and dest_size: 4864 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 4865 if (slice_len == NULL) slice_len = top(); 4866 if (dest_size == NULL) dest_size = top(); 4867 4868 // operate on this memory slice: 4869 Node* mem = memory(adr_type); // memory slice to operate on 4870 4871 // scaling and rounding of indexes: 4872 int scale = exact_log2(type2aelembytes(basic_elem_type)); 4873 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 4874 int clear_low = (-1 << scale) & (BytesPerInt - 1); 4875 int bump_bit = (-1 << scale) & BytesPerInt; 4876 4877 // determine constant starts and ends 4878 const intptr_t BIG_NEG = -128; 4879 assert(BIG_NEG + 2*abase < 0, "neg enough"); 4880 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 4881 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 4882 if (slice_len_con == 0) { 4883 return; // nothing to do here 4884 } 4885 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 4886 intptr_t end_con = find_intptr_t_con(dest_size, -1); 4887 if (slice_idx_con >= 0 && slice_len_con >= 0) { 4888 assert(end_con < 0, "not two cons"); 4889 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 4890 BytesPerLong); 4891 } 4892 4893 if (start_con >= 0 && end_con >= 0) { 4894 // Constant start and end. Simple. 4895 mem = ClearArrayNode::clear_memory(control(), mem, dest, 4896 start_con, end_con, &_gvn); 4897 } else if (start_con >= 0 && dest_size != top()) { 4898 // Constant start, pre-rounded end after the tail of the array. 4899 Node* end = dest_size; 4900 mem = ClearArrayNode::clear_memory(control(), mem, dest, 4901 start_con, end, &_gvn); 4902 } else if (start_con >= 0 && slice_len != top()) { 4903 // Constant start, non-constant end. End needs rounding up. 4904 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 4905 intptr_t end_base = abase + (slice_idx_con << scale); 4906 int end_round = (-1 << scale) & (BytesPerLong - 1); 4907 Node* end = ConvI2X(slice_len); 4908 if (scale != 0) 4909 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) )); 4910 end_base += end_round; 4911 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) ); 4912 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) ); 4913 mem = ClearArrayNode::clear_memory(control(), mem, dest, 4914 start_con, end, &_gvn); 4915 } else if (start_con < 0 && dest_size != top()) { 4916 // Non-constant start, pre-rounded end after the tail of the array. 4917 // This is almost certainly a "round-to-end" operation. 4918 Node* start = slice_idx; 4919 start = ConvI2X(start); 4920 if (scale != 0) 4921 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) )); 4922 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) ); 4923 if ((bump_bit | clear_low) != 0) { 4924 int to_clear = (bump_bit | clear_low); 4925 // Align up mod 8, then store a jint zero unconditionally 4926 // just before the mod-8 boundary. 4927 if (((abase + bump_bit) & ~to_clear) - bump_bit 4928 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 4929 bump_bit = 0; 4930 assert((abase & to_clear) == 0, "array base must be long-aligned"); 4931 } else { 4932 // Bump 'start' up to (or past) the next jint boundary: 4933 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) ); 4934 assert((abase & clear_low) == 0, "array base must be int-aligned"); 4935 } 4936 // Round bumped 'start' down to jlong boundary in body of array. 4937 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) ); 4938 if (bump_bit != 0) { 4939 // Store a zero to the immediately preceding jint: 4940 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) ); 4941 Node* p1 = basic_plus_adr(dest, x1); 4942 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 4943 mem = _gvn.transform(mem); 4944 } 4945 } 4946 Node* end = dest_size; // pre-rounded 4947 mem = ClearArrayNode::clear_memory(control(), mem, dest, 4948 start, end, &_gvn); 4949 } else { 4950 // Non-constant start, unrounded non-constant end. 4951 // (Nobody zeroes a random midsection of an array using this routine.) 4952 ShouldNotReachHere(); // fix caller 4953 } 4954 4955 // Done. 4956 set_memory(mem, adr_type); 4957 } 4958 4959 4960 bool 4961 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 4962 BasicType basic_elem_type, 4963 AllocateNode* alloc, 4964 Node* src, Node* src_offset, 4965 Node* dest, Node* dest_offset, 4966 Node* dest_size) { 4967 // See if there is an advantage from block transfer. 4968 int scale = exact_log2(type2aelembytes(basic_elem_type)); 4969 if (scale >= LogBytesPerLong) 4970 return false; // it is already a block transfer 4971 4972 // Look at the alignment of the starting offsets. 4973 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 4974 const intptr_t BIG_NEG = -128; 4975 assert(BIG_NEG + 2*abase < 0, "neg enough"); 4976 4977 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale); 4978 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale); 4979 if (src_off < 0 || dest_off < 0) 4980 // At present, we can only understand constants. 4981 return false; 4982 4983 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 4984 // Non-aligned; too bad. 4985 // One more chance: Pick off an initial 32-bit word. 4986 // This is a common case, since abase can be odd mod 8. 4987 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 4988 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 4989 Node* sptr = basic_plus_adr(src, src_off); 4990 Node* dptr = basic_plus_adr(dest, dest_off); 4991 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 4992 store_to_memory(control(), dptr, sval, T_INT, adr_type); 4993 src_off += BytesPerInt; 4994 dest_off += BytesPerInt; 4995 } else { 4996 return false; 4997 } 4998 } 4999 assert(src_off % BytesPerLong == 0, ""); 5000 assert(dest_off % BytesPerLong == 0, ""); 5001 5002 // Do this copy by giant steps. 5003 Node* sptr = basic_plus_adr(src, src_off); 5004 Node* dptr = basic_plus_adr(dest, dest_off); 5005 Node* countx = dest_size; 5006 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) ); 5007 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) ); 5008 5009 bool disjoint_bases = true; // since alloc != NULL 5010 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5011 sptr, NULL, dptr, NULL, countx); 5012 5013 return true; 5014 } 5015 5016 5017 // Helper function; generates code for the slow case. 5018 // We make a call to a runtime method which emulates the native method, 5019 // but without the native wrapper overhead. 5020 void 5021 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5022 Node* src, Node* src_offset, 5023 Node* dest, Node* dest_offset, 5024 Node* copy_length, 5025 int nargs) { 5026 _sp += nargs; // any deopt will start just before call to enclosing method 5027 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5028 OptoRuntime::slow_arraycopy_Type(), 5029 OptoRuntime::slow_arraycopy_Java(), 5030 "slow_arraycopy", adr_type, 5031 src, src_offset, dest, dest_offset, 5032 copy_length); 5033 _sp -= nargs; 5034 5035 // Handle exceptions thrown by this fellow: 5036 make_slow_call_ex(call, env()->Throwable_klass(), false); 5037 } 5038 5039 // Helper function; generates code for cases requiring runtime checks. 5040 Node* 5041 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5042 Node* dest_elem_klass, 5043 Node* src, Node* src_offset, 5044 Node* dest, Node* dest_offset, 5045 Node* copy_length, 5046 int nargs) { 5047 if (stopped()) return NULL; 5048 5049 address copyfunc_addr = StubRoutines::checkcast_arraycopy(); 5050 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5051 return NULL; 5052 } 5053 5054 // Pick out the parameters required to perform a store-check 5055 // for the target array. This is an optimistic check. It will 5056 // look in each non-null element's class, at the desired klass's 5057 // super_check_offset, for the desired klass. 5058 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc); 5059 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5060 Node* n3 = new(C, 3) LoadINode(NULL, immutable_memory(), p3, TypeRawPtr::BOTTOM); 5061 Node* check_offset = _gvn.transform(n3); 5062 Node* check_value = dest_elem_klass; 5063 5064 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5065 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5066 5067 // (We know the arrays are never conjoint, because their types differ.) 5068 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5069 OptoRuntime::checkcast_arraycopy_Type(), 5070 copyfunc_addr, "checkcast_arraycopy", adr_type, 5071 // five arguments, of which two are 5072 // intptr_t (jlong in LP64) 5073 src_start, dest_start, 5074 copy_length XTOP, 5075 check_offset XTOP, 5076 check_value); 5077 5078 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5079 } 5080 5081 5082 // Helper function; generates code for cases requiring runtime checks. 5083 Node* 5084 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5085 Node* src, Node* src_offset, 5086 Node* dest, Node* dest_offset, 5087 Node* copy_length, 5088 int nargs) { 5089 if (stopped()) return NULL; 5090 5091 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5092 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5093 return NULL; 5094 } 5095 5096 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5097 OptoRuntime::generic_arraycopy_Type(), 5098 copyfunc_addr, "generic_arraycopy", adr_type, 5099 src, src_offset, dest, dest_offset, copy_length); 5100 5101 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5102 } 5103 5104 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5105 void 5106 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5107 BasicType basic_elem_type, 5108 bool disjoint_bases, 5109 Node* src, Node* src_offset, 5110 Node* dest, Node* dest_offset, 5111 Node* copy_length) { 5112 if (stopped()) return; // nothing to do 5113 5114 Node* src_start = src; 5115 Node* dest_start = dest; 5116 if (src_offset != NULL || dest_offset != NULL) { 5117 assert(src_offset != NULL && dest_offset != NULL, ""); 5118 src_start = array_element_address(src, src_offset, basic_elem_type); 5119 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5120 } 5121 5122 // Figure out which arraycopy runtime method to call. 5123 const char* copyfunc_name = "arraycopy"; 5124 address copyfunc_addr = 5125 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5126 disjoint_bases, copyfunc_name); 5127 5128 // Call it. Note that the count_ix value is not scaled to a byte-size. 5129 make_runtime_call(RC_LEAF|RC_NO_FP, 5130 OptoRuntime::fast_arraycopy_Type(), 5131 copyfunc_addr, copyfunc_name, adr_type, 5132 src_start, dest_start, copy_length XTOP); 5133 }