1 /* 2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "classfile/vmSymbols.hpp" 28 #include "compiler/compileLog.hpp" 29 #include "oops/objArrayKlass.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/callGenerator.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/idealKit.hpp" 34 #include "opto/mulnode.hpp" 35 #include "opto/parse.hpp" 36 #include "opto/runtime.hpp" 37 #include "opto/subnode.hpp" 38 #include "prims/nativeLookup.hpp" 39 #include "runtime/sharedRuntime.hpp" 40 41 class LibraryIntrinsic : public InlineCallGenerator { 42 // Extend the set of intrinsics known to the runtime: 43 public: 44 private: 45 bool _is_virtual; 46 vmIntrinsics::ID _intrinsic_id; 47 48 public: 49 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id) 50 : InlineCallGenerator(m), 51 _is_virtual(is_virtual), 52 _intrinsic_id(id) 53 { 54 } 55 virtual bool is_intrinsic() const { return true; } 56 virtual bool is_virtual() const { return _is_virtual; } 57 virtual JVMState* generate(JVMState* jvms); 58 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 59 }; 60 61 // Local helper class for LibraryIntrinsic: 62 class LibraryCallKit : public GraphKit { 63 private: 64 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 65 66 public: 67 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic) 68 : GraphKit(caller), 69 _intrinsic(intrinsic) 70 { 71 } 72 73 ciMethod* caller() const { return jvms()->method(); } 74 int bci() const { return jvms()->bci(); } 75 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 76 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 77 ciMethod* callee() const { return _intrinsic->method(); } 78 ciSignature* signature() const { return callee()->signature(); } 79 int arg_size() const { return callee()->arg_size(); } 80 81 bool try_to_inline(); 82 83 // Helper functions to inline natives 84 void push_result(RegionNode* region, PhiNode* value); 85 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 86 Node* generate_slow_guard(Node* test, RegionNode* region); 87 Node* generate_fair_guard(Node* test, RegionNode* region); 88 Node* generate_negative_guard(Node* index, RegionNode* region, 89 // resulting CastII of index: 90 Node* *pos_index = NULL); 91 Node* generate_nonpositive_guard(Node* index, bool never_negative, 92 // resulting CastII of index: 93 Node* *pos_index = NULL); 94 Node* generate_limit_guard(Node* offset, Node* subseq_length, 95 Node* array_length, 96 RegionNode* region); 97 Node* generate_current_thread(Node* &tls_output); 98 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 99 bool disjoint_bases, const char* &name); 100 Node* load_mirror_from_klass(Node* klass); 101 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 102 int nargs, 103 RegionNode* region, int null_path, 104 int offset); 105 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs, 106 RegionNode* region, int null_path) { 107 int offset = java_lang_Class::klass_offset_in_bytes(); 108 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 109 region, null_path, 110 offset); 111 } 112 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 113 int nargs, 114 RegionNode* region, int null_path) { 115 int offset = java_lang_Class::array_klass_offset_in_bytes(); 116 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 117 region, null_path, 118 offset); 119 } 120 Node* generate_access_flags_guard(Node* kls, 121 int modifier_mask, int modifier_bits, 122 RegionNode* region); 123 Node* generate_interface_guard(Node* kls, RegionNode* region); 124 Node* generate_array_guard(Node* kls, RegionNode* region) { 125 return generate_array_guard_common(kls, region, false, false); 126 } 127 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 128 return generate_array_guard_common(kls, region, false, true); 129 } 130 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 131 return generate_array_guard_common(kls, region, true, false); 132 } 133 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 134 return generate_array_guard_common(kls, region, true, true); 135 } 136 Node* generate_array_guard_common(Node* kls, RegionNode* region, 137 bool obj_array, bool not_array); 138 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 139 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 140 bool is_virtual = false, bool is_static = false); 141 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 142 return generate_method_call(method_id, false, true); 143 } 144 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 145 return generate_method_call(method_id, true, false); 146 } 147 148 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2); 149 bool inline_string_compareTo(); 150 bool inline_string_indexOf(); 151 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 152 bool inline_string_equals(); 153 Node* pop_math_arg(); 154 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 155 bool inline_math_native(vmIntrinsics::ID id); 156 bool inline_trig(vmIntrinsics::ID id); 157 bool inline_trans(vmIntrinsics::ID id); 158 bool inline_abs(vmIntrinsics::ID id); 159 bool inline_sqrt(vmIntrinsics::ID id); 160 bool inline_pow(vmIntrinsics::ID id); 161 bool inline_exp(vmIntrinsics::ID id); 162 bool inline_min_max(vmIntrinsics::ID id); 163 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 164 // This returns Type::AnyPtr, RawPtr, or OopPtr. 165 int classify_unsafe_addr(Node* &base, Node* &offset); 166 Node* make_unsafe_address(Node* base, Node* offset); 167 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 168 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 169 bool inline_unsafe_allocate(); 170 bool inline_unsafe_copyMemory(); 171 bool inline_native_currentThread(); 172 bool inline_native_time_funcs(bool isNano); 173 bool inline_native_isInterrupted(); 174 bool inline_native_Class_query(vmIntrinsics::ID id); 175 bool inline_native_subtype_check(); 176 177 bool inline_native_newArray(); 178 bool inline_native_getLength(); 179 bool inline_array_copyOf(bool is_copyOfRange); 180 bool inline_array_equals(); 181 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 182 bool inline_native_clone(bool is_virtual); 183 bool inline_native_Reflection_getCallerClass(); 184 bool inline_native_AtomicLong_get(); 185 bool inline_native_AtomicLong_attemptUpdate(); 186 bool is_method_invoke_or_aux_frame(JVMState* jvms); 187 // Helper function for inlining native object hash method 188 bool inline_native_hashcode(bool is_virtual, bool is_static); 189 bool inline_native_getClass(); 190 191 // Helper functions for inlining arraycopy 192 bool inline_arraycopy(); 193 void generate_arraycopy(const TypePtr* adr_type, 194 BasicType basic_elem_type, 195 Node* src, Node* src_offset, 196 Node* dest, Node* dest_offset, 197 Node* copy_length, 198 bool disjoint_bases = false, 199 bool length_never_negative = false, 200 RegionNode* slow_region = NULL); 201 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 202 RegionNode* slow_region); 203 void generate_clear_array(const TypePtr* adr_type, 204 Node* dest, 205 BasicType basic_elem_type, 206 Node* slice_off, 207 Node* slice_len, 208 Node* slice_end); 209 bool generate_block_arraycopy(const TypePtr* adr_type, 210 BasicType basic_elem_type, 211 AllocateNode* alloc, 212 Node* src, Node* src_offset, 213 Node* dest, Node* dest_offset, 214 Node* dest_size); 215 void generate_slow_arraycopy(const TypePtr* adr_type, 216 Node* src, Node* src_offset, 217 Node* dest, Node* dest_offset, 218 Node* copy_length); 219 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 220 Node* dest_elem_klass, 221 Node* src, Node* src_offset, 222 Node* dest, Node* dest_offset, 223 Node* copy_length); 224 Node* generate_generic_arraycopy(const TypePtr* adr_type, 225 Node* src, Node* src_offset, 226 Node* dest, Node* dest_offset, 227 Node* copy_length); 228 void generate_unchecked_arraycopy(const TypePtr* adr_type, 229 BasicType basic_elem_type, 230 bool disjoint_bases, 231 Node* src, Node* src_offset, 232 Node* dest, Node* dest_offset, 233 Node* copy_length); 234 bool inline_unsafe_CAS(BasicType type); 235 bool inline_unsafe_ordered_store(BasicType type); 236 bool inline_fp_conversions(vmIntrinsics::ID id); 237 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id); 238 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id); 239 bool inline_bitCount(vmIntrinsics::ID id); 240 bool inline_reverseBytes(vmIntrinsics::ID id); 241 242 bool inline_reference_get(); 243 }; 244 245 246 //---------------------------make_vm_intrinsic---------------------------- 247 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 248 vmIntrinsics::ID id = m->intrinsic_id(); 249 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 250 251 if (DisableIntrinsic[0] != '\0' 252 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 253 // disabled by a user request on the command line: 254 // example: -XX:DisableIntrinsic=_hashCode,_getClass 255 return NULL; 256 } 257 258 if (!m->is_loaded()) { 259 // do not attempt to inline unloaded methods 260 return NULL; 261 } 262 263 // Only a few intrinsics implement a virtual dispatch. 264 // They are expensive calls which are also frequently overridden. 265 if (is_virtual) { 266 switch (id) { 267 case vmIntrinsics::_hashCode: 268 case vmIntrinsics::_clone: 269 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 270 break; 271 default: 272 return NULL; 273 } 274 } 275 276 // -XX:-InlineNatives disables nearly all intrinsics: 277 if (!InlineNatives) { 278 switch (id) { 279 case vmIntrinsics::_indexOf: 280 case vmIntrinsics::_compareTo: 281 case vmIntrinsics::_equals: 282 case vmIntrinsics::_equalsC: 283 break; // InlineNatives does not control String.compareTo 284 default: 285 return NULL; 286 } 287 } 288 289 switch (id) { 290 case vmIntrinsics::_compareTo: 291 if (!SpecialStringCompareTo) return NULL; 292 break; 293 case vmIntrinsics::_indexOf: 294 if (!SpecialStringIndexOf) return NULL; 295 break; 296 case vmIntrinsics::_equals: 297 if (!SpecialStringEquals) return NULL; 298 break; 299 case vmIntrinsics::_equalsC: 300 if (!SpecialArraysEquals) return NULL; 301 break; 302 case vmIntrinsics::_arraycopy: 303 if (!InlineArrayCopy) return NULL; 304 break; 305 case vmIntrinsics::_copyMemory: 306 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 307 if (!InlineArrayCopy) return NULL; 308 break; 309 case vmIntrinsics::_hashCode: 310 if (!InlineObjectHash) return NULL; 311 break; 312 case vmIntrinsics::_clone: 313 case vmIntrinsics::_copyOf: 314 case vmIntrinsics::_copyOfRange: 315 if (!InlineObjectCopy) return NULL; 316 // These also use the arraycopy intrinsic mechanism: 317 if (!InlineArrayCopy) return NULL; 318 break; 319 case vmIntrinsics::_checkIndex: 320 // We do not intrinsify this. The optimizer does fine with it. 321 return NULL; 322 323 case vmIntrinsics::_get_AtomicLong: 324 case vmIntrinsics::_attemptUpdate: 325 if (!InlineAtomicLong) return NULL; 326 break; 327 328 case vmIntrinsics::_getCallerClass: 329 if (!UseNewReflection) return NULL; 330 if (!InlineReflectionGetCallerClass) return NULL; 331 if (!JDK_Version::is_gte_jdk14x_version()) return NULL; 332 break; 333 334 case vmIntrinsics::_bitCount_i: 335 case vmIntrinsics::_bitCount_l: 336 if (!UsePopCountInstruction) return NULL; 337 break; 338 339 case vmIntrinsics::_Reference_get: 340 // It is only when G1 is enabled that we absolutely 341 // need to use the intrinsic version of Reference.get() 342 // so that the value in the referent field, if necessary, 343 // can be registered by the pre-barrier code. 344 if (!UseG1GC) return NULL; 345 break; 346 347 default: 348 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 349 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 350 break; 351 } 352 353 // -XX:-InlineClassNatives disables natives from the Class class. 354 // The flag applies to all reflective calls, notably Array.newArray 355 // (visible to Java programmers as Array.newInstance). 356 if (m->holder()->name() == ciSymbol::java_lang_Class() || 357 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 358 if (!InlineClassNatives) return NULL; 359 } 360 361 // -XX:-InlineThreadNatives disables natives from the Thread class. 362 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 363 if (!InlineThreadNatives) return NULL; 364 } 365 366 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 367 if (m->holder()->name() == ciSymbol::java_lang_Math() || 368 m->holder()->name() == ciSymbol::java_lang_Float() || 369 m->holder()->name() == ciSymbol::java_lang_Double()) { 370 if (!InlineMathNatives) return NULL; 371 } 372 373 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 374 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 375 if (!InlineUnsafeOps) return NULL; 376 } 377 378 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); 379 } 380 381 //----------------------register_library_intrinsics----------------------- 382 // Initialize this file's data structures, for each Compile instance. 383 void Compile::register_library_intrinsics() { 384 // Nothing to do here. 385 } 386 387 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 388 LibraryCallKit kit(jvms, this); 389 Compile* C = kit.C; 390 int nodes = C->unique(); 391 #ifndef PRODUCT 392 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 393 char buf[1000]; 394 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 395 tty->print_cr("Intrinsic %s", str); 396 } 397 #endif 398 399 if (kit.try_to_inline()) { 400 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 401 if (jvms->has_method()) { 402 // Not a root compile. 403 tty->print("Inlining intrinsic %s%s at bci:%d in", 404 vmIntrinsics::name_at(intrinsic_id()), 405 (is_virtual() ? " (virtual)" : ""), kit.bci()); 406 kit.caller()->print_short_name(tty); 407 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); 408 } else { 409 // Root compile 410 tty->print_cr("Generating intrinsic %s%s at bci:%d", 411 vmIntrinsics::name_at(intrinsic_id()), 412 (is_virtual() ? " (virtual)" : ""), kit.bci()); 413 } 414 } 415 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 416 if (C->log()) { 417 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 418 vmIntrinsics::name_at(intrinsic_id()), 419 (is_virtual() ? " virtual='1'" : ""), 420 C->unique() - nodes); 421 } 422 return kit.transfer_exceptions_into_jvms(); 423 } 424 425 if (PrintIntrinsics) { 426 if (jvms->has_method()) { 427 // Not a root compile. 428 tty->print("Did not inline intrinsic %s%s at bci:%d in", 429 vmIntrinsics::name_at(intrinsic_id()), 430 (is_virtual() ? " (virtual)" : ""), kit.bci()); 431 kit.caller()->print_short_name(tty); 432 tty->print_cr(" (%d bytes)", kit.caller()->code_size()); 433 } else { 434 // Root compile 435 tty->print("Did not generate intrinsic %s%s at bci:%d in", 436 vmIntrinsics::name_at(intrinsic_id()), 437 (is_virtual() ? " (virtual)" : ""), kit.bci()); 438 } 439 } 440 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 441 return NULL; 442 } 443 444 bool LibraryCallKit::try_to_inline() { 445 // Handle symbolic names for otherwise undistinguished boolean switches: 446 const bool is_store = true; 447 const bool is_native_ptr = true; 448 const bool is_static = true; 449 450 if (!jvms()->has_method()) { 451 // Root JVMState has a null method. 452 assert(map()->memory()->Opcode() == Op_Parm, ""); 453 // Insert the memory aliasing node 454 set_all_memory(reset_memory()); 455 } 456 assert(merged_memory(), ""); 457 458 switch (intrinsic_id()) { 459 case vmIntrinsics::_hashCode: 460 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 461 case vmIntrinsics::_identityHashCode: 462 return inline_native_hashcode(/*!virtual*/ false, is_static); 463 case vmIntrinsics::_getClass: 464 return inline_native_getClass(); 465 466 case vmIntrinsics::_dsin: 467 case vmIntrinsics::_dcos: 468 case vmIntrinsics::_dtan: 469 case vmIntrinsics::_dabs: 470 case vmIntrinsics::_datan2: 471 case vmIntrinsics::_dsqrt: 472 case vmIntrinsics::_dexp: 473 case vmIntrinsics::_dlog: 474 case vmIntrinsics::_dlog10: 475 case vmIntrinsics::_dpow: 476 return inline_math_native(intrinsic_id()); 477 478 case vmIntrinsics::_min: 479 case vmIntrinsics::_max: 480 return inline_min_max(intrinsic_id()); 481 482 case vmIntrinsics::_arraycopy: 483 return inline_arraycopy(); 484 485 case vmIntrinsics::_compareTo: 486 return inline_string_compareTo(); 487 case vmIntrinsics::_indexOf: 488 return inline_string_indexOf(); 489 case vmIntrinsics::_equals: 490 return inline_string_equals(); 491 492 case vmIntrinsics::_getObject: 493 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); 494 case vmIntrinsics::_getBoolean: 495 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); 496 case vmIntrinsics::_getByte: 497 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); 498 case vmIntrinsics::_getShort: 499 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); 500 case vmIntrinsics::_getChar: 501 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); 502 case vmIntrinsics::_getInt: 503 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); 504 case vmIntrinsics::_getLong: 505 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); 506 case vmIntrinsics::_getFloat: 507 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); 508 case vmIntrinsics::_getDouble: 509 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); 510 511 case vmIntrinsics::_putObject: 512 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); 513 case vmIntrinsics::_putBoolean: 514 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); 515 case vmIntrinsics::_putByte: 516 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); 517 case vmIntrinsics::_putShort: 518 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); 519 case vmIntrinsics::_putChar: 520 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); 521 case vmIntrinsics::_putInt: 522 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); 523 case vmIntrinsics::_putLong: 524 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); 525 case vmIntrinsics::_putFloat: 526 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); 527 case vmIntrinsics::_putDouble: 528 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); 529 530 case vmIntrinsics::_getByte_raw: 531 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); 532 case vmIntrinsics::_getShort_raw: 533 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); 534 case vmIntrinsics::_getChar_raw: 535 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); 536 case vmIntrinsics::_getInt_raw: 537 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); 538 case vmIntrinsics::_getLong_raw: 539 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); 540 case vmIntrinsics::_getFloat_raw: 541 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); 542 case vmIntrinsics::_getDouble_raw: 543 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); 544 case vmIntrinsics::_getAddress_raw: 545 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); 546 547 case vmIntrinsics::_putByte_raw: 548 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); 549 case vmIntrinsics::_putShort_raw: 550 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); 551 case vmIntrinsics::_putChar_raw: 552 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); 553 case vmIntrinsics::_putInt_raw: 554 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); 555 case vmIntrinsics::_putLong_raw: 556 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); 557 case vmIntrinsics::_putFloat_raw: 558 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); 559 case vmIntrinsics::_putDouble_raw: 560 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); 561 case vmIntrinsics::_putAddress_raw: 562 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); 563 564 case vmIntrinsics::_getObjectVolatile: 565 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); 566 case vmIntrinsics::_getBooleanVolatile: 567 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); 568 case vmIntrinsics::_getByteVolatile: 569 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); 570 case vmIntrinsics::_getShortVolatile: 571 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); 572 case vmIntrinsics::_getCharVolatile: 573 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); 574 case vmIntrinsics::_getIntVolatile: 575 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); 576 case vmIntrinsics::_getLongVolatile: 577 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); 578 case vmIntrinsics::_getFloatVolatile: 579 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); 580 case vmIntrinsics::_getDoubleVolatile: 581 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); 582 583 case vmIntrinsics::_putObjectVolatile: 584 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); 585 case vmIntrinsics::_putBooleanVolatile: 586 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); 587 case vmIntrinsics::_putByteVolatile: 588 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); 589 case vmIntrinsics::_putShortVolatile: 590 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); 591 case vmIntrinsics::_putCharVolatile: 592 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); 593 case vmIntrinsics::_putIntVolatile: 594 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); 595 case vmIntrinsics::_putLongVolatile: 596 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); 597 case vmIntrinsics::_putFloatVolatile: 598 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); 599 case vmIntrinsics::_putDoubleVolatile: 600 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); 601 602 case vmIntrinsics::_prefetchRead: 603 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 604 case vmIntrinsics::_prefetchWrite: 605 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 606 case vmIntrinsics::_prefetchReadStatic: 607 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 608 case vmIntrinsics::_prefetchWriteStatic: 609 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 610 611 case vmIntrinsics::_compareAndSwapObject: 612 return inline_unsafe_CAS(T_OBJECT); 613 case vmIntrinsics::_compareAndSwapInt: 614 return inline_unsafe_CAS(T_INT); 615 case vmIntrinsics::_compareAndSwapLong: 616 return inline_unsafe_CAS(T_LONG); 617 618 case vmIntrinsics::_putOrderedObject: 619 return inline_unsafe_ordered_store(T_OBJECT); 620 case vmIntrinsics::_putOrderedInt: 621 return inline_unsafe_ordered_store(T_INT); 622 case vmIntrinsics::_putOrderedLong: 623 return inline_unsafe_ordered_store(T_LONG); 624 625 case vmIntrinsics::_currentThread: 626 return inline_native_currentThread(); 627 case vmIntrinsics::_isInterrupted: 628 return inline_native_isInterrupted(); 629 630 case vmIntrinsics::_currentTimeMillis: 631 return inline_native_time_funcs(false); 632 case vmIntrinsics::_nanoTime: 633 return inline_native_time_funcs(true); 634 case vmIntrinsics::_allocateInstance: 635 return inline_unsafe_allocate(); 636 case vmIntrinsics::_copyMemory: 637 return inline_unsafe_copyMemory(); 638 case vmIntrinsics::_newArray: 639 return inline_native_newArray(); 640 case vmIntrinsics::_getLength: 641 return inline_native_getLength(); 642 case vmIntrinsics::_copyOf: 643 return inline_array_copyOf(false); 644 case vmIntrinsics::_copyOfRange: 645 return inline_array_copyOf(true); 646 case vmIntrinsics::_equalsC: 647 return inline_array_equals(); 648 case vmIntrinsics::_clone: 649 return inline_native_clone(intrinsic()->is_virtual()); 650 651 case vmIntrinsics::_isAssignableFrom: 652 return inline_native_subtype_check(); 653 654 case vmIntrinsics::_isInstance: 655 case vmIntrinsics::_getModifiers: 656 case vmIntrinsics::_isInterface: 657 case vmIntrinsics::_isArray: 658 case vmIntrinsics::_isPrimitive: 659 case vmIntrinsics::_getSuperclass: 660 case vmIntrinsics::_getComponentType: 661 case vmIntrinsics::_getClassAccessFlags: 662 return inline_native_Class_query(intrinsic_id()); 663 664 case vmIntrinsics::_floatToRawIntBits: 665 case vmIntrinsics::_floatToIntBits: 666 case vmIntrinsics::_intBitsToFloat: 667 case vmIntrinsics::_doubleToRawLongBits: 668 case vmIntrinsics::_doubleToLongBits: 669 case vmIntrinsics::_longBitsToDouble: 670 return inline_fp_conversions(intrinsic_id()); 671 672 case vmIntrinsics::_numberOfLeadingZeros_i: 673 case vmIntrinsics::_numberOfLeadingZeros_l: 674 return inline_numberOfLeadingZeros(intrinsic_id()); 675 676 case vmIntrinsics::_numberOfTrailingZeros_i: 677 case vmIntrinsics::_numberOfTrailingZeros_l: 678 return inline_numberOfTrailingZeros(intrinsic_id()); 679 680 case vmIntrinsics::_bitCount_i: 681 case vmIntrinsics::_bitCount_l: 682 return inline_bitCount(intrinsic_id()); 683 684 case vmIntrinsics::_reverseBytes_i: 685 case vmIntrinsics::_reverseBytes_l: 686 case vmIntrinsics::_reverseBytes_s: 687 case vmIntrinsics::_reverseBytes_c: 688 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); 689 690 case vmIntrinsics::_get_AtomicLong: 691 return inline_native_AtomicLong_get(); 692 case vmIntrinsics::_attemptUpdate: 693 return inline_native_AtomicLong_attemptUpdate(); 694 695 case vmIntrinsics::_getCallerClass: 696 return inline_native_Reflection_getCallerClass(); 697 698 case vmIntrinsics::_Reference_get: 699 return inline_reference_get(); 700 701 default: 702 // If you get here, it may be that someone has added a new intrinsic 703 // to the list in vmSymbols.hpp without implementing it here. 704 #ifndef PRODUCT 705 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 706 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 707 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 708 } 709 #endif 710 return false; 711 } 712 } 713 714 //------------------------------push_result------------------------------ 715 // Helper function for finishing intrinsics. 716 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { 717 record_for_igvn(region); 718 set_control(_gvn.transform(region)); 719 BasicType value_type = value->type()->basic_type(); 720 push_node(value_type, _gvn.transform(value)); 721 } 722 723 //------------------------------generate_guard--------------------------- 724 // Helper function for generating guarded fast-slow graph structures. 725 // The given 'test', if true, guards a slow path. If the test fails 726 // then a fast path can be taken. (We generally hope it fails.) 727 // In all cases, GraphKit::control() is updated to the fast path. 728 // The returned value represents the control for the slow path. 729 // The return value is never 'top'; it is either a valid control 730 // or NULL if it is obvious that the slow path can never be taken. 731 // Also, if region and the slow control are not NULL, the slow edge 732 // is appended to the region. 733 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 734 if (stopped()) { 735 // Already short circuited. 736 return NULL; 737 } 738 739 // Build an if node and its projections. 740 // If test is true we take the slow path, which we assume is uncommon. 741 if (_gvn.type(test) == TypeInt::ZERO) { 742 // The slow branch is never taken. No need to build this guard. 743 return NULL; 744 } 745 746 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 747 748 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); 749 if (if_slow == top()) { 750 // The slow branch is never taken. No need to build this guard. 751 return NULL; 752 } 753 754 if (region != NULL) 755 region->add_req(if_slow); 756 757 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); 758 set_control(if_fast); 759 760 return if_slow; 761 } 762 763 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 764 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 765 } 766 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 767 return generate_guard(test, region, PROB_FAIR); 768 } 769 770 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 771 Node* *pos_index) { 772 if (stopped()) 773 return NULL; // already stopped 774 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 775 return NULL; // index is already adequately typed 776 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 777 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 778 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 779 if (is_neg != NULL && pos_index != NULL) { 780 // Emulate effect of Parse::adjust_map_after_if. 781 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); 782 ccast->set_req(0, control()); 783 (*pos_index) = _gvn.transform(ccast); 784 } 785 return is_neg; 786 } 787 788 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 789 Node* *pos_index) { 790 if (stopped()) 791 return NULL; // already stopped 792 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 793 return NULL; // index is already adequately typed 794 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 795 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 796 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) ); 797 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 798 if (is_notp != NULL && pos_index != NULL) { 799 // Emulate effect of Parse::adjust_map_after_if. 800 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1); 801 ccast->set_req(0, control()); 802 (*pos_index) = _gvn.transform(ccast); 803 } 804 return is_notp; 805 } 806 807 // Make sure that 'position' is a valid limit index, in [0..length]. 808 // There are two equivalent plans for checking this: 809 // A. (offset + copyLength) unsigned<= arrayLength 810 // B. offset <= (arrayLength - copyLength) 811 // We require that all of the values above, except for the sum and 812 // difference, are already known to be non-negative. 813 // Plan A is robust in the face of overflow, if offset and copyLength 814 // are both hugely positive. 815 // 816 // Plan B is less direct and intuitive, but it does not overflow at 817 // all, since the difference of two non-negatives is always 818 // representable. Whenever Java methods must perform the equivalent 819 // check they generally use Plan B instead of Plan A. 820 // For the moment we use Plan A. 821 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 822 Node* subseq_length, 823 Node* array_length, 824 RegionNode* region) { 825 if (stopped()) 826 return NULL; // already stopped 827 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 828 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length)) 829 return NULL; // common case of whole-array copy 830 Node* last = subseq_length; 831 if (!zero_offset) // last += offset 832 last = _gvn.transform( new (C, 3) AddINode(last, offset)); 833 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); 834 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 835 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 836 return is_over; 837 } 838 839 840 //--------------------------generate_current_thread-------------------- 841 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 842 ciKlass* thread_klass = env()->Thread_klass(); 843 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 844 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); 845 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 846 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 847 tls_output = thread; 848 return threadObj; 849 } 850 851 852 //------------------------------make_string_method_node------------------------ 853 // Helper method for String intrinsic finctions. 854 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) { 855 const int value_offset = java_lang_String::value_offset_in_bytes(); 856 const int count_offset = java_lang_String::count_offset_in_bytes(); 857 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 858 859 Node* no_ctrl = NULL; 860 861 ciInstanceKlass* klass = env()->String_klass(); 862 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 863 864 const TypeAryPtr* value_type = 865 TypeAryPtr::make(TypePtr::NotNull, 866 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 867 ciTypeArrayKlass::make(T_CHAR), true, 0); 868 869 // Get start addr of string and substring 870 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset); 871 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset)); 872 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset); 873 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 874 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 875 876 // Pin loads from String::equals() argument since it could be NULL. 877 Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl; 878 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset); 879 Node* str2_value = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset)); 880 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset); 881 Node* str2_offset = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 882 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 883 884 Node* result = NULL; 885 switch (opcode) { 886 case Op_StrIndexOf: 887 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 888 str1_start, cnt1, str2_start, cnt2); 889 break; 890 case Op_StrComp: 891 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS), 892 str1_start, cnt1, str2_start, cnt2); 893 break; 894 case Op_StrEquals: 895 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 896 str1_start, str2_start, cnt1); 897 break; 898 default: 899 ShouldNotReachHere(); 900 return NULL; 901 } 902 903 // All these intrinsics have checks. 904 C->set_has_split_ifs(true); // Has chance for split-if optimization 905 906 return _gvn.transform(result); 907 } 908 909 //------------------------------inline_string_compareTo------------------------ 910 bool LibraryCallKit::inline_string_compareTo() { 911 912 if (!Matcher::has_match_rule(Op_StrComp)) return false; 913 914 const int value_offset = java_lang_String::value_offset_in_bytes(); 915 const int count_offset = java_lang_String::count_offset_in_bytes(); 916 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 917 918 _sp += 2; 919 Node *argument = pop(); // pop non-receiver first: it was pushed second 920 Node *receiver = pop(); 921 922 // Null check on self without removing any arguments. The argument 923 // null check technically happens in the wrong place, which can lead to 924 // invalid stack traces when string compare is inlined into a method 925 // which handles NullPointerExceptions. 926 _sp += 2; 927 receiver = do_null_check(receiver, T_OBJECT); 928 argument = do_null_check(argument, T_OBJECT); 929 _sp -= 2; 930 if (stopped()) { 931 return true; 932 } 933 934 ciInstanceKlass* klass = env()->String_klass(); 935 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 936 Node* no_ctrl = NULL; 937 938 // Get counts for string and argument 939 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset); 940 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 941 942 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset); 943 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 944 945 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt); 946 push(compare); 947 return true; 948 } 949 950 //------------------------------inline_string_equals------------------------ 951 bool LibraryCallKit::inline_string_equals() { 952 953 if (!Matcher::has_match_rule(Op_StrEquals)) return false; 954 955 const int value_offset = java_lang_String::value_offset_in_bytes(); 956 const int count_offset = java_lang_String::count_offset_in_bytes(); 957 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 958 959 int nargs = 2; 960 _sp += nargs; 961 Node* argument = pop(); // pop non-receiver first: it was pushed second 962 Node* receiver = pop(); 963 964 // Null check on self without removing any arguments. The argument 965 // null check technically happens in the wrong place, which can lead to 966 // invalid stack traces when string compare is inlined into a method 967 // which handles NullPointerExceptions. 968 _sp += nargs; 969 receiver = do_null_check(receiver, T_OBJECT); 970 //should not do null check for argument for String.equals(), because spec 971 //allows to specify NULL as argument. 972 _sp -= nargs; 973 974 if (stopped()) { 975 return true; 976 } 977 978 // paths (plus control) merge 979 RegionNode* region = new (C, 5) RegionNode(5); 980 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL); 981 982 // does source == target string? 983 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument)); 984 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq)); 985 986 Node* if_eq = generate_slow_guard(bol, NULL); 987 if (if_eq != NULL) { 988 // receiver == argument 989 phi->init_req(2, intcon(1)); 990 region->init_req(2, if_eq); 991 } 992 993 // get String klass for instanceOf 994 ciInstanceKlass* klass = env()->String_klass(); 995 996 if (!stopped()) { 997 _sp += nargs; // gen_instanceof might do an uncommon trap 998 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 999 _sp -= nargs; 1000 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1))); 1001 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne)); 1002 1003 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1004 //instanceOf == true, fallthrough 1005 1006 if (inst_false != NULL) { 1007 phi->init_req(3, intcon(0)); 1008 region->init_req(3, inst_false); 1009 } 1010 } 1011 1012 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1013 1014 Node* no_ctrl = NULL; 1015 Node* receiver_cnt; 1016 Node* argument_cnt; 1017 1018 if (!stopped()) { 1019 // Properly cast the argument to String 1020 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type)); 1021 1022 // Get counts for string and argument 1023 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset); 1024 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1025 1026 // Pin load from argument string since it could be NULL. 1027 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset); 1028 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1029 1030 // Check for receiver count != argument count 1031 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) ); 1032 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) ); 1033 Node* if_ne = generate_slow_guard(bol, NULL); 1034 if (if_ne != NULL) { 1035 phi->init_req(4, intcon(0)); 1036 region->init_req(4, if_ne); 1037 } 1038 } 1039 1040 // Check for count == 0 is done by mach node StrEquals. 1041 1042 if (!stopped()) { 1043 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt); 1044 phi->init_req(1, equals); 1045 region->init_req(1, control()); 1046 } 1047 1048 // post merge 1049 set_control(_gvn.transform(region)); 1050 record_for_igvn(region); 1051 1052 push(_gvn.transform(phi)); 1053 1054 return true; 1055 } 1056 1057 //------------------------------inline_array_equals---------------------------- 1058 bool LibraryCallKit::inline_array_equals() { 1059 1060 if (!Matcher::has_match_rule(Op_AryEq)) return false; 1061 1062 _sp += 2; 1063 Node *argument2 = pop(); 1064 Node *argument1 = pop(); 1065 1066 Node* equals = 1067 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS), 1068 argument1, argument2) ); 1069 push(equals); 1070 return true; 1071 } 1072 1073 // Java version of String.indexOf(constant string) 1074 // class StringDecl { 1075 // StringDecl(char[] ca) { 1076 // offset = 0; 1077 // count = ca.length; 1078 // value = ca; 1079 // } 1080 // int offset; 1081 // int count; 1082 // char[] value; 1083 // } 1084 // 1085 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1086 // int targetOffset, int cache_i, int md2) { 1087 // int cache = cache_i; 1088 // int sourceOffset = string_object.offset; 1089 // int sourceCount = string_object.count; 1090 // int targetCount = target_object.length; 1091 // 1092 // int targetCountLess1 = targetCount - 1; 1093 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1094 // 1095 // char[] source = string_object.value; 1096 // char[] target = target_object; 1097 // int lastChar = target[targetCountLess1]; 1098 // 1099 // outer_loop: 1100 // for (int i = sourceOffset; i < sourceEnd; ) { 1101 // int src = source[i + targetCountLess1]; 1102 // if (src == lastChar) { 1103 // // With random strings and a 4-character alphabet, 1104 // // reverse matching at this point sets up 0.8% fewer 1105 // // frames, but (paradoxically) makes 0.3% more probes. 1106 // // Since those probes are nearer the lastChar probe, 1107 // // there is may be a net D$ win with reverse matching. 1108 // // But, reversing loop inhibits unroll of inner loop 1109 // // for unknown reason. So, does running outer loop from 1110 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1111 // for (int j = 0; j < targetCountLess1; j++) { 1112 // if (target[targetOffset + j] != source[i+j]) { 1113 // if ((cache & (1 << source[i+j])) == 0) { 1114 // if (md2 < j+1) { 1115 // i += j+1; 1116 // continue outer_loop; 1117 // } 1118 // } 1119 // i += md2; 1120 // continue outer_loop; 1121 // } 1122 // } 1123 // return i - sourceOffset; 1124 // } 1125 // if ((cache & (1 << src)) == 0) { 1126 // i += targetCountLess1; 1127 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1128 // i++; 1129 // } 1130 // return -1; 1131 // } 1132 1133 //------------------------------string_indexOf------------------------ 1134 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1135 jint cache_i, jint md2_i) { 1136 1137 Node* no_ctrl = NULL; 1138 float likely = PROB_LIKELY(0.9); 1139 float unlikely = PROB_UNLIKELY(0.9); 1140 1141 const int value_offset = java_lang_String::value_offset_in_bytes(); 1142 const int count_offset = java_lang_String::count_offset_in_bytes(); 1143 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1144 1145 ciInstanceKlass* klass = env()->String_klass(); 1146 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1147 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0); 1148 1149 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset); 1150 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset)); 1151 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset); 1152 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1153 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset); 1154 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset)); 1155 1156 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array)) ); 1157 jint target_length = target_array->length(); 1158 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1159 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1160 1161 IdealKit kit(gvn(), control(), merged_memory(), false, true); 1162 #define __ kit. 1163 Node* zero = __ ConI(0); 1164 Node* one = __ ConI(1); 1165 Node* cache = __ ConI(cache_i); 1166 Node* md2 = __ ConI(md2_i); 1167 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1168 Node* targetCount = __ ConI(target_length); 1169 Node* targetCountLess1 = __ ConI(target_length - 1); 1170 Node* targetOffset = __ ConI(targetOffset_i); 1171 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1172 1173 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1174 Node* outer_loop = __ make_label(2 /* goto */); 1175 Node* return_ = __ make_label(1); 1176 1177 __ set(rtn,__ ConI(-1)); 1178 __ loop(i, sourceOffset, BoolTest::lt, sourceEnd); { 1179 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1180 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1181 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1182 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1183 __ loop(j, zero, BoolTest::lt, targetCountLess1); { 1184 Node* tpj = __ AddI(targetOffset, __ value(j)); 1185 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1186 Node* ipj = __ AddI(__ value(i), __ value(j)); 1187 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1188 __ if_then(targ, BoolTest::ne, src2); { 1189 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1190 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1191 __ increment(i, __ AddI(__ value(j), one)); 1192 __ goto_(outer_loop); 1193 } __ end_if(); __ dead(j); 1194 }__ end_if(); __ dead(j); 1195 __ increment(i, md2); 1196 __ goto_(outer_loop); 1197 }__ end_if(); 1198 __ increment(j, one); 1199 }__ end_loop(); __ dead(j); 1200 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1201 __ goto_(return_); 1202 }__ end_if(); 1203 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1204 __ increment(i, targetCountLess1); 1205 }__ end_if(); 1206 __ increment(i, one); 1207 __ bind(outer_loop); 1208 }__ end_loop(); __ dead(i); 1209 __ bind(return_); 1210 1211 // Final sync IdealKit and GraphKit. 1212 sync_kit(kit); 1213 Node* result = __ value(rtn); 1214 #undef __ 1215 C->set_has_loops(true); 1216 return result; 1217 } 1218 1219 //------------------------------inline_string_indexOf------------------------ 1220 bool LibraryCallKit::inline_string_indexOf() { 1221 1222 const int value_offset = java_lang_String::value_offset_in_bytes(); 1223 const int count_offset = java_lang_String::count_offset_in_bytes(); 1224 const int offset_offset = java_lang_String::offset_offset_in_bytes(); 1225 1226 _sp += 2; 1227 Node *argument = pop(); // pop non-receiver first: it was pushed second 1228 Node *receiver = pop(); 1229 1230 Node* result; 1231 // Disable the use of pcmpestri until it can be guaranteed that 1232 // the load doesn't cross into the uncommited space. 1233 if (false && Matcher::has_match_rule(Op_StrIndexOf) && 1234 UseSSE42Intrinsics) { 1235 // Generate SSE4.2 version of indexOf 1236 // We currently only have match rules that use SSE4.2 1237 1238 // Null check on self without removing any arguments. The argument 1239 // null check technically happens in the wrong place, which can lead to 1240 // invalid stack traces when string compare is inlined into a method 1241 // which handles NullPointerExceptions. 1242 _sp += 2; 1243 receiver = do_null_check(receiver, T_OBJECT); 1244 argument = do_null_check(argument, T_OBJECT); 1245 _sp -= 2; 1246 1247 if (stopped()) { 1248 return true; 1249 } 1250 1251 // Make the merge point 1252 RegionNode* result_rgn = new (C, 3) RegionNode(3); 1253 Node* result_phi = new (C, 3) PhiNode(result_rgn, TypeInt::INT); 1254 Node* no_ctrl = NULL; 1255 1256 ciInstanceKlass* klass = env()->String_klass(); 1257 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1258 1259 // Get counts for string and substr 1260 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset); 1261 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1262 1263 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset); 1264 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset)); 1265 1266 // Check for substr count > string count 1267 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) ); 1268 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) ); 1269 Node* if_gt = generate_slow_guard(bol, NULL); 1270 if (if_gt != NULL) { 1271 result_phi->init_req(2, intcon(-1)); 1272 result_rgn->init_req(2, if_gt); 1273 } 1274 1275 if (!stopped()) { 1276 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt); 1277 result_phi->init_req(1, result); 1278 result_rgn->init_req(1, control()); 1279 } 1280 set_control(_gvn.transform(result_rgn)); 1281 record_for_igvn(result_rgn); 1282 result = _gvn.transform(result_phi); 1283 1284 } else { //Use LibraryCallKit::string_indexOf 1285 // don't intrinsify is argument isn't a constant string. 1286 if (!argument->is_Con()) { 1287 return false; 1288 } 1289 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); 1290 if (str_type == NULL) { 1291 return false; 1292 } 1293 ciInstanceKlass* klass = env()->String_klass(); 1294 ciObject* str_const = str_type->const_oop(); 1295 if (str_const == NULL || str_const->klass() != klass) { 1296 return false; 1297 } 1298 ciInstance* str = str_const->as_instance(); 1299 assert(str != NULL, "must be instance"); 1300 1301 ciObject* v = str->field_value_by_offset(value_offset).as_object(); 1302 int o = str->field_value_by_offset(offset_offset).as_int(); 1303 int c = str->field_value_by_offset(count_offset).as_int(); 1304 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1305 1306 // constant strings have no offset and count == length which 1307 // simplifies the resulting code somewhat so lets optimize for that. 1308 if (o != 0 || c != pat->length()) { 1309 return false; 1310 } 1311 1312 // Null check on self without removing any arguments. The argument 1313 // null check technically happens in the wrong place, which can lead to 1314 // invalid stack traces when string compare is inlined into a method 1315 // which handles NullPointerExceptions. 1316 _sp += 2; 1317 receiver = do_null_check(receiver, T_OBJECT); 1318 // No null check on the argument is needed since it's a constant String oop. 1319 _sp -= 2; 1320 if (stopped()) { 1321 return true; 1322 } 1323 1324 // The null string as a pattern always returns 0 (match at beginning of string) 1325 if (c == 0) { 1326 push(intcon(0)); 1327 return true; 1328 } 1329 1330 // Generate default indexOf 1331 jchar lastChar = pat->char_at(o + (c - 1)); 1332 int cache = 0; 1333 int i; 1334 for (i = 0; i < c - 1; i++) { 1335 assert(i < pat->length(), "out of range"); 1336 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1337 } 1338 1339 int md2 = c; 1340 for (i = 0; i < c - 1; i++) { 1341 assert(i < pat->length(), "out of range"); 1342 if (pat->char_at(o + i) == lastChar) { 1343 md2 = (c - 1) - i; 1344 } 1345 } 1346 1347 result = string_indexOf(receiver, pat, o, cache, md2); 1348 } 1349 1350 push(result); 1351 return true; 1352 } 1353 1354 //--------------------------pop_math_arg-------------------------------- 1355 // Pop a double argument to a math function from the stack 1356 // rounding it if necessary. 1357 Node * LibraryCallKit::pop_math_arg() { 1358 Node *arg = pop_pair(); 1359 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) 1360 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); 1361 return arg; 1362 } 1363 1364 //------------------------------inline_trig---------------------------------- 1365 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1366 // argument reduction which will turn into a fast/slow diamond. 1367 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1368 _sp += arg_size(); // restore stack pointer 1369 Node* arg = pop_math_arg(); 1370 Node* trig = NULL; 1371 1372 switch (id) { 1373 case vmIntrinsics::_dsin: 1374 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); 1375 break; 1376 case vmIntrinsics::_dcos: 1377 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); 1378 break; 1379 case vmIntrinsics::_dtan: 1380 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); 1381 break; 1382 default: 1383 assert(false, "bad intrinsic was passed in"); 1384 return false; 1385 } 1386 1387 // Rounding required? Check for argument reduction! 1388 if( Matcher::strict_fp_requires_explicit_rounding ) { 1389 1390 static const double pi_4 = 0.7853981633974483; 1391 static const double neg_pi_4 = -0.7853981633974483; 1392 // pi/2 in 80-bit extended precision 1393 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1394 // -pi/2 in 80-bit extended precision 1395 // 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}; 1396 // Cutoff value for using this argument reduction technique 1397 //static const double pi_2_minus_epsilon = 1.564660403643354; 1398 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1399 1400 // Pseudocode for sin: 1401 // if (x <= Math.PI / 4.0) { 1402 // if (x >= -Math.PI / 4.0) return fsin(x); 1403 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1404 // } else { 1405 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1406 // } 1407 // return StrictMath.sin(x); 1408 1409 // Pseudocode for cos: 1410 // if (x <= Math.PI / 4.0) { 1411 // if (x >= -Math.PI / 4.0) return fcos(x); 1412 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1413 // } else { 1414 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1415 // } 1416 // return StrictMath.cos(x); 1417 1418 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1419 // requires a special machine instruction to load it. Instead we'll try 1420 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1421 // probably do the math inside the SIN encoding. 1422 1423 // Make the merge point 1424 RegionNode *r = new (C, 3) RegionNode(3); 1425 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); 1426 1427 // Flatten arg so we need only 1 test 1428 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); 1429 // Node for PI/4 constant 1430 Node *pi4 = makecon(TypeD::make(pi_4)); 1431 // Check PI/4 : abs(arg) 1432 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); 1433 // Check: If PI/4 < abs(arg) then go slow 1434 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); 1435 // Branch either way 1436 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1437 set_control(opt_iff(r,iff)); 1438 1439 // Set fast path result 1440 phi->init_req(2,trig); 1441 1442 // Slow path - non-blocking leaf call 1443 Node* call = NULL; 1444 switch (id) { 1445 case vmIntrinsics::_dsin: 1446 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1447 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1448 "Sin", NULL, arg, top()); 1449 break; 1450 case vmIntrinsics::_dcos: 1451 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1452 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1453 "Cos", NULL, arg, top()); 1454 break; 1455 case vmIntrinsics::_dtan: 1456 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1457 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1458 "Tan", NULL, arg, top()); 1459 break; 1460 } 1461 assert(control()->in(0) == call, ""); 1462 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); 1463 r->init_req(1,control()); 1464 phi->init_req(1,slow_result); 1465 1466 // Post-merge 1467 set_control(_gvn.transform(r)); 1468 record_for_igvn(r); 1469 trig = _gvn.transform(phi); 1470 1471 C->set_has_split_ifs(true); // Has chance for split-if optimization 1472 } 1473 // Push result back on JVM stack 1474 push_pair(trig); 1475 return true; 1476 } 1477 1478 //------------------------------inline_sqrt------------------------------------- 1479 // Inline square root instruction, if possible. 1480 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { 1481 assert(id == vmIntrinsics::_dsqrt, "Not square root"); 1482 _sp += arg_size(); // restore stack pointer 1483 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); 1484 return true; 1485 } 1486 1487 //------------------------------inline_abs------------------------------------- 1488 // Inline absolute value instruction, if possible. 1489 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { 1490 assert(id == vmIntrinsics::_dabs, "Not absolute value"); 1491 _sp += arg_size(); // restore stack pointer 1492 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); 1493 return true; 1494 } 1495 1496 //------------------------------inline_exp------------------------------------- 1497 // Inline exp instructions, if possible. The Intel hardware only misses 1498 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1499 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { 1500 assert(id == vmIntrinsics::_dexp, "Not exp"); 1501 1502 // If this inlining ever returned NaN in the past, we do not intrinsify it 1503 // every again. NaN results requires StrictMath.exp handling. 1504 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1505 1506 // Do not intrinsify on older platforms which lack cmove. 1507 if (ConditionalMoveLimit == 0) return false; 1508 1509 _sp += arg_size(); // restore stack pointer 1510 Node *x = pop_math_arg(); 1511 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); 1512 1513 //------------------- 1514 //result=(result.isNaN())? StrictMath::exp():result; 1515 // Check: If isNaN() by checking result!=result? then go to Strict Math 1516 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1517 // Build the boolean node 1518 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1519 1520 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1521 // End the current control-flow path 1522 push_pair(x); 1523 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp 1524 // to handle. Recompile without intrinsifying Math.exp 1525 uncommon_trap(Deoptimization::Reason_intrinsic, 1526 Deoptimization::Action_make_not_entrant); 1527 } 1528 1529 C->set_has_split_ifs(true); // Has chance for split-if optimization 1530 1531 push_pair(result); 1532 1533 return true; 1534 } 1535 1536 //------------------------------inline_pow------------------------------------- 1537 // Inline power instructions, if possible. 1538 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { 1539 assert(id == vmIntrinsics::_dpow, "Not pow"); 1540 1541 // If this inlining ever returned NaN in the past, we do not intrinsify it 1542 // every again. NaN results requires StrictMath.pow handling. 1543 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1544 1545 // Do not intrinsify on older platforms which lack cmove. 1546 if (ConditionalMoveLimit == 0) return false; 1547 1548 // Pseudocode for pow 1549 // if (x <= 0.0) { 1550 // if ((double)((int)y)==y) { // if y is int 1551 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y) 1552 // } else { 1553 // result = NaN; 1554 // } 1555 // } else { 1556 // result = DPow(x,y); 1557 // } 1558 // if (result != result)? { 1559 // uncommon_trap(); 1560 // } 1561 // return result; 1562 1563 _sp += arg_size(); // restore stack pointer 1564 Node* y = pop_math_arg(); 1565 Node* x = pop_math_arg(); 1566 1567 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); 1568 1569 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow 1570 // inside of something) then skip the fancy tests and just check for 1571 // NaN result. 1572 Node *result = NULL; 1573 if( jvms()->depth() >= 1 ) { 1574 result = fast_result; 1575 } else { 1576 1577 // Set the merge point for If node with condition of (x <= 0.0) 1578 // There are four possible paths to region node and phi node 1579 RegionNode *r = new (C, 4) RegionNode(4); 1580 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); 1581 1582 // Build the first if node: if (x <= 0.0) 1583 // Node for 0 constant 1584 Node *zeronode = makecon(TypeD::ZERO); 1585 // Check x:0 1586 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); 1587 // Check: If (x<=0) then go complex path 1588 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); 1589 // Branch either way 1590 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1591 Node *opt_test = _gvn.transform(if1); 1592 //assert( opt_test->is_If(), "Expect an IfNode"); 1593 IfNode *opt_if1 = (IfNode*)opt_test; 1594 // Fast path taken; set region slot 3 1595 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) ); 1596 r->init_req(3,fast_taken); // Capture fast-control 1597 1598 // Fast path not-taken, i.e. slow path 1599 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) ); 1600 1601 // Set fast path result 1602 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) ); 1603 phi->init_req(3, fast_result); 1604 1605 // Complex path 1606 // Build the second if node (if y is int) 1607 // Node for (int)y 1608 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y)); 1609 // Node for (double)((int) y) 1610 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty)); 1611 // Check (double)((int) y) : y 1612 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y)); 1613 // Check if (y isn't int) then go to slow path 1614 1615 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) ); 1616 // Branch either way 1617 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1618 Node *slow_path = opt_iff(r,if2); // Set region path 2 1619 1620 // Calculate DPow(abs(x), y)*(1 & (int)y) 1621 // Node for constant 1 1622 Node *conone = intcon(1); 1623 // 1& (int)y 1624 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) ); 1625 // zero node 1626 Node *conzero = intcon(0); 1627 // Check (1&(int)y)==0? 1628 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero)); 1629 // Check if (1&(int)y)!=0?, if so the result is negative 1630 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); 1631 // abs(x) 1632 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); 1633 // abs(x)^y 1634 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) ); 1635 // -abs(x)^y 1636 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); 1637 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1638 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1639 // Set complex path fast result 1640 phi->init_req(2, signresult); 1641 1642 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1643 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1644 r->init_req(1,slow_path); 1645 phi->init_req(1,slow_result); 1646 1647 // Post merge 1648 set_control(_gvn.transform(r)); 1649 record_for_igvn(r); 1650 result=_gvn.transform(phi); 1651 } 1652 1653 //------------------- 1654 //result=(result.isNaN())? uncommon_trap():result; 1655 // Check: If isNaN() by checking result!=result? then go to Strict Math 1656 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1657 // Build the boolean node 1658 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1659 1660 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1661 // End the current control-flow path 1662 push_pair(x); 1663 push_pair(y); 1664 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow 1665 // to handle. Recompile without intrinsifying Math.pow. 1666 uncommon_trap(Deoptimization::Reason_intrinsic, 1667 Deoptimization::Action_make_not_entrant); 1668 } 1669 1670 C->set_has_split_ifs(true); // Has chance for split-if optimization 1671 1672 push_pair(result); 1673 1674 return true; 1675 } 1676 1677 //------------------------------inline_trans------------------------------------- 1678 // Inline transcendental instructions, if possible. The Intel hardware gets 1679 // these right, no funny corner cases missed. 1680 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { 1681 _sp += arg_size(); // restore stack pointer 1682 Node* arg = pop_math_arg(); 1683 Node* trans = NULL; 1684 1685 switch (id) { 1686 case vmIntrinsics::_dlog: 1687 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); 1688 break; 1689 case vmIntrinsics::_dlog10: 1690 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); 1691 break; 1692 default: 1693 assert(false, "bad intrinsic was passed in"); 1694 return false; 1695 } 1696 1697 // Push result back on JVM stack 1698 push_pair(trans); 1699 return true; 1700 } 1701 1702 //------------------------------runtime_math----------------------------- 1703 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1704 Node* a = NULL; 1705 Node* b = NULL; 1706 1707 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1708 "must be (DD)D or (D)D type"); 1709 1710 // Inputs 1711 _sp += arg_size(); // restore stack pointer 1712 if (call_type == OptoRuntime::Math_DD_D_Type()) { 1713 b = pop_math_arg(); 1714 } 1715 a = pop_math_arg(); 1716 1717 const TypePtr* no_memory_effects = NULL; 1718 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1719 no_memory_effects, 1720 a, top(), b, b ? top() : NULL); 1721 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); 1722 #ifdef ASSERT 1723 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); 1724 assert(value_top == top(), "second value must be top"); 1725 #endif 1726 1727 push_pair(value); 1728 return true; 1729 } 1730 1731 //------------------------------inline_math_native----------------------------- 1732 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1733 switch (id) { 1734 // These intrinsics are not properly supported on all hardware 1735 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1736 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); 1737 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1738 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); 1739 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1740 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); 1741 1742 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : 1743 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); 1744 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : 1745 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); 1746 1747 // These intrinsics are supported on all hardware 1748 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; 1749 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; 1750 1751 // These intrinsics don't work on X86. The ad implementation doesn't 1752 // handle NaN's properly. Instead of returning infinity, the ad 1753 // implementation returns a NaN on overflow. See bug: 6304089 1754 // Once the ad implementations are fixed, change the code below 1755 // to match the intrinsics above 1756 1757 case vmIntrinsics::_dexp: return 1758 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1759 case vmIntrinsics::_dpow: return 1760 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1761 1762 // These intrinsics are not yet correctly implemented 1763 case vmIntrinsics::_datan2: 1764 return false; 1765 1766 default: 1767 ShouldNotReachHere(); 1768 return false; 1769 } 1770 } 1771 1772 static bool is_simple_name(Node* n) { 1773 return (n->req() == 1 // constant 1774 || (n->is_Type() && n->as_Type()->type()->singleton()) 1775 || n->is_Proj() // parameter or return value 1776 || n->is_Phi() // local of some sort 1777 ); 1778 } 1779 1780 //----------------------------inline_min_max----------------------------------- 1781 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1782 push(generate_min_max(id, argument(0), argument(1))); 1783 1784 return true; 1785 } 1786 1787 Node* 1788 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1789 // These are the candidate return value: 1790 Node* xvalue = x0; 1791 Node* yvalue = y0; 1792 1793 if (xvalue == yvalue) { 1794 return xvalue; 1795 } 1796 1797 bool want_max = (id == vmIntrinsics::_max); 1798 1799 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1800 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1801 if (txvalue == NULL || tyvalue == NULL) return top(); 1802 // This is not really necessary, but it is consistent with a 1803 // hypothetical MaxINode::Value method: 1804 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1805 1806 // %%% This folding logic should (ideally) be in a different place. 1807 // Some should be inside IfNode, and there to be a more reliable 1808 // transformation of ?: style patterns into cmoves. We also want 1809 // more powerful optimizations around cmove and min/max. 1810 1811 // Try to find a dominating comparison of these guys. 1812 // It can simplify the index computation for Arrays.copyOf 1813 // and similar uses of System.arraycopy. 1814 // First, compute the normalized version of CmpI(x, y). 1815 int cmp_op = Op_CmpI; 1816 Node* xkey = xvalue; 1817 Node* ykey = yvalue; 1818 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); 1819 if (ideal_cmpxy->is_Cmp()) { 1820 // E.g., if we have CmpI(length - offset, count), 1821 // it might idealize to CmpI(length, count + offset) 1822 cmp_op = ideal_cmpxy->Opcode(); 1823 xkey = ideal_cmpxy->in(1); 1824 ykey = ideal_cmpxy->in(2); 1825 } 1826 1827 // Start by locating any relevant comparisons. 1828 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1829 Node* cmpxy = NULL; 1830 Node* cmpyx = NULL; 1831 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1832 Node* cmp = start_from->fast_out(k); 1833 if (cmp->outcnt() > 0 && // must have prior uses 1834 cmp->in(0) == NULL && // must be context-independent 1835 cmp->Opcode() == cmp_op) { // right kind of compare 1836 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1837 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1838 } 1839 } 1840 1841 const int NCMPS = 2; 1842 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1843 int cmpn; 1844 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1845 if (cmps[cmpn] != NULL) break; // find a result 1846 } 1847 if (cmpn < NCMPS) { 1848 // Look for a dominating test that tells us the min and max. 1849 int depth = 0; // Limit search depth for speed 1850 Node* dom = control(); 1851 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1852 if (++depth >= 100) break; 1853 Node* ifproj = dom; 1854 if (!ifproj->is_Proj()) continue; 1855 Node* iff = ifproj->in(0); 1856 if (!iff->is_If()) continue; 1857 Node* bol = iff->in(1); 1858 if (!bol->is_Bool()) continue; 1859 Node* cmp = bol->in(1); 1860 if (cmp == NULL) continue; 1861 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1862 if (cmps[cmpn] == cmp) break; 1863 if (cmpn == NCMPS) continue; 1864 BoolTest::mask btest = bol->as_Bool()->_test._test; 1865 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1866 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1867 // At this point, we know that 'x btest y' is true. 1868 switch (btest) { 1869 case BoolTest::eq: 1870 // They are proven equal, so we can collapse the min/max. 1871 // Either value is the answer. Choose the simpler. 1872 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1873 return yvalue; 1874 return xvalue; 1875 case BoolTest::lt: // x < y 1876 case BoolTest::le: // x <= y 1877 return (want_max ? yvalue : xvalue); 1878 case BoolTest::gt: // x > y 1879 case BoolTest::ge: // x >= y 1880 return (want_max ? xvalue : yvalue); 1881 } 1882 } 1883 } 1884 1885 // We failed to find a dominating test. 1886 // Let's pick a test that might GVN with prior tests. 1887 Node* best_bol = NULL; 1888 BoolTest::mask best_btest = BoolTest::illegal; 1889 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1890 Node* cmp = cmps[cmpn]; 1891 if (cmp == NULL) continue; 1892 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1893 Node* bol = cmp->fast_out(j); 1894 if (!bol->is_Bool()) continue; 1895 BoolTest::mask btest = bol->as_Bool()->_test._test; 1896 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1897 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1898 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1899 best_bol = bol->as_Bool(); 1900 best_btest = btest; 1901 } 1902 } 1903 } 1904 1905 Node* answer_if_true = NULL; 1906 Node* answer_if_false = NULL; 1907 switch (best_btest) { 1908 default: 1909 if (cmpxy == NULL) 1910 cmpxy = ideal_cmpxy; 1911 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); 1912 // and fall through: 1913 case BoolTest::lt: // x < y 1914 case BoolTest::le: // x <= y 1915 answer_if_true = (want_max ? yvalue : xvalue); 1916 answer_if_false = (want_max ? xvalue : yvalue); 1917 break; 1918 case BoolTest::gt: // x > y 1919 case BoolTest::ge: // x >= y 1920 answer_if_true = (want_max ? xvalue : yvalue); 1921 answer_if_false = (want_max ? yvalue : xvalue); 1922 break; 1923 } 1924 1925 jint hi, lo; 1926 if (want_max) { 1927 // We can sharpen the minimum. 1928 hi = MAX2(txvalue->_hi, tyvalue->_hi); 1929 lo = MAX2(txvalue->_lo, tyvalue->_lo); 1930 } else { 1931 // We can sharpen the maximum. 1932 hi = MIN2(txvalue->_hi, tyvalue->_hi); 1933 lo = MIN2(txvalue->_lo, tyvalue->_lo); 1934 } 1935 1936 // Use a flow-free graph structure, to avoid creating excess control edges 1937 // which could hinder other optimizations. 1938 // Since Math.min/max is often used with arraycopy, we want 1939 // tightly_coupled_allocation to be able to see beyond min/max expressions. 1940 Node* cmov = CMoveNode::make(C, NULL, best_bol, 1941 answer_if_false, answer_if_true, 1942 TypeInt::make(lo, hi, widen)); 1943 1944 return _gvn.transform(cmov); 1945 1946 /* 1947 // This is not as desirable as it may seem, since Min and Max 1948 // nodes do not have a full set of optimizations. 1949 // And they would interfere, anyway, with 'if' optimizations 1950 // and with CMoveI canonical forms. 1951 switch (id) { 1952 case vmIntrinsics::_min: 1953 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 1954 case vmIntrinsics::_max: 1955 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 1956 default: 1957 ShouldNotReachHere(); 1958 } 1959 */ 1960 } 1961 1962 inline int 1963 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 1964 const TypePtr* base_type = TypePtr::NULL_PTR; 1965 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 1966 if (base_type == NULL) { 1967 // Unknown type. 1968 return Type::AnyPtr; 1969 } else if (base_type == TypePtr::NULL_PTR) { 1970 // Since this is a NULL+long form, we have to switch to a rawptr. 1971 base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); 1972 offset = MakeConX(0); 1973 return Type::RawPtr; 1974 } else if (base_type->base() == Type::RawPtr) { 1975 return Type::RawPtr; 1976 } else if (base_type->isa_oopptr()) { 1977 // Base is never null => always a heap address. 1978 if (base_type->ptr() == TypePtr::NotNull) { 1979 return Type::OopPtr; 1980 } 1981 // Offset is small => always a heap address. 1982 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 1983 if (offset_type != NULL && 1984 base_type->offset() == 0 && // (should always be?) 1985 offset_type->_lo >= 0 && 1986 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 1987 return Type::OopPtr; 1988 } 1989 // Otherwise, it might either be oop+off or NULL+addr. 1990 return Type::AnyPtr; 1991 } else { 1992 // No information: 1993 return Type::AnyPtr; 1994 } 1995 } 1996 1997 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 1998 int kind = classify_unsafe_addr(base, offset); 1999 if (kind == Type::RawPtr) { 2000 return basic_plus_adr(top(), base, offset); 2001 } else { 2002 return basic_plus_adr(base, offset); 2003 } 2004 } 2005 2006 //-------------------inline_numberOfLeadingZeros_int/long----------------------- 2007 // inline int Integer.numberOfLeadingZeros(int) 2008 // inline int Long.numberOfLeadingZeros(long) 2009 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) { 2010 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros"); 2011 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false; 2012 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false; 2013 _sp += arg_size(); // restore stack pointer 2014 switch (id) { 2015 case vmIntrinsics::_numberOfLeadingZeros_i: 2016 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop()))); 2017 break; 2018 case vmIntrinsics::_numberOfLeadingZeros_l: 2019 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair()))); 2020 break; 2021 default: 2022 ShouldNotReachHere(); 2023 } 2024 return true; 2025 } 2026 2027 //-------------------inline_numberOfTrailingZeros_int/long---------------------- 2028 // inline int Integer.numberOfTrailingZeros(int) 2029 // inline int Long.numberOfTrailingZeros(long) 2030 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) { 2031 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros"); 2032 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false; 2033 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false; 2034 _sp += arg_size(); // restore stack pointer 2035 switch (id) { 2036 case vmIntrinsics::_numberOfTrailingZeros_i: 2037 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop()))); 2038 break; 2039 case vmIntrinsics::_numberOfTrailingZeros_l: 2040 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair()))); 2041 break; 2042 default: 2043 ShouldNotReachHere(); 2044 } 2045 return true; 2046 } 2047 2048 //----------------------------inline_bitCount_int/long----------------------- 2049 // inline int Integer.bitCount(int) 2050 // inline int Long.bitCount(long) 2051 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) { 2052 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount"); 2053 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false; 2054 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false; 2055 _sp += arg_size(); // restore stack pointer 2056 switch (id) { 2057 case vmIntrinsics::_bitCount_i: 2058 push(_gvn.transform(new (C, 2) PopCountINode(pop()))); 2059 break; 2060 case vmIntrinsics::_bitCount_l: 2061 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair()))); 2062 break; 2063 default: 2064 ShouldNotReachHere(); 2065 } 2066 return true; 2067 } 2068 2069 //----------------------------inline_reverseBytes_int/long/char/short------------------- 2070 // inline Integer.reverseBytes(int) 2071 // inline Long.reverseBytes(long) 2072 // inline Character.reverseBytes(char) 2073 // inline Short.reverseBytes(short) 2074 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { 2075 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l || 2076 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s, 2077 "not reverse Bytes"); 2078 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; 2079 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; 2080 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false; 2081 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false; 2082 _sp += arg_size(); // restore stack pointer 2083 switch (id) { 2084 case vmIntrinsics::_reverseBytes_i: 2085 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); 2086 break; 2087 case vmIntrinsics::_reverseBytes_l: 2088 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); 2089 break; 2090 case vmIntrinsics::_reverseBytes_c: 2091 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop()))); 2092 break; 2093 case vmIntrinsics::_reverseBytes_s: 2094 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop()))); 2095 break; 2096 default: 2097 ; 2098 } 2099 return true; 2100 } 2101 2102 //----------------------------inline_unsafe_access---------------------------- 2103 2104 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2105 2106 // Interpret Unsafe.fieldOffset cookies correctly: 2107 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2108 2109 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2110 if (callee()->is_static()) return false; // caller must have the capability! 2111 2112 #ifndef PRODUCT 2113 { 2114 ResourceMark rm; 2115 // Check the signatures. 2116 ciSignature* sig = signature(); 2117 #ifdef ASSERT 2118 if (!is_store) { 2119 // Object getObject(Object base, int/long offset), etc. 2120 BasicType rtype = sig->return_type()->basic_type(); 2121 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2122 rtype = T_ADDRESS; // it is really a C void* 2123 assert(rtype == type, "getter must return the expected value"); 2124 if (!is_native_ptr) { 2125 assert(sig->count() == 2, "oop getter has 2 arguments"); 2126 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2127 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2128 } else { 2129 assert(sig->count() == 1, "native getter has 1 argument"); 2130 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2131 } 2132 } else { 2133 // void putObject(Object base, int/long offset, Object x), etc. 2134 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2135 if (!is_native_ptr) { 2136 assert(sig->count() == 3, "oop putter has 3 arguments"); 2137 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2138 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2139 } else { 2140 assert(sig->count() == 2, "native putter has 2 arguments"); 2141 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2142 } 2143 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2144 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2145 vtype = T_ADDRESS; // it is really a C void* 2146 assert(vtype == type, "putter must accept the expected value"); 2147 } 2148 #endif // ASSERT 2149 } 2150 #endif //PRODUCT 2151 2152 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2153 2154 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; 2155 2156 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words 2157 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); 2158 2159 debug_only(int saved_sp = _sp); 2160 _sp += nargs; 2161 2162 Node* val; 2163 debug_only(val = (Node*)(uintptr_t)-1); 2164 2165 2166 if (is_store) { 2167 // Get the value being stored. (Pop it first; it was pushed last.) 2168 switch (type) { 2169 case T_DOUBLE: 2170 case T_LONG: 2171 case T_ADDRESS: 2172 val = pop_pair(); 2173 break; 2174 default: 2175 val = pop(); 2176 } 2177 } 2178 2179 // Build address expression. See the code in inline_unsafe_prefetch. 2180 Node *adr; 2181 Node *heap_base_oop = top(); 2182 if (!is_native_ptr) { 2183 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2184 Node* offset = pop_pair(); 2185 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2186 Node* base = pop(); 2187 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2188 // to be plain byte offsets, which are also the same as those accepted 2189 // by oopDesc::field_base. 2190 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2191 "fieldOffset must be byte-scaled"); 2192 // 32-bit machines ignore the high half! 2193 offset = ConvL2X(offset); 2194 adr = make_unsafe_address(base, offset); 2195 heap_base_oop = base; 2196 } else { 2197 Node* ptr = pop_pair(); 2198 // Adjust Java long to machine word: 2199 ptr = ConvL2X(ptr); 2200 adr = make_unsafe_address(NULL, ptr); 2201 } 2202 2203 // Pop receiver last: it was pushed first. 2204 Node *receiver = pop(); 2205 2206 assert(saved_sp == _sp, "must have correct argument count"); 2207 2208 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2209 2210 // First guess at the value type. 2211 const Type *value_type = Type::get_const_basic_type(type); 2212 2213 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2214 // there was not enough information to nail it down. 2215 Compile::AliasType* alias_type = C->alias_type(adr_type); 2216 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2217 2218 // We will need memory barriers unless we can determine a unique 2219 // alias category for this reference. (Note: If for some reason 2220 // the barriers get omitted and the unsafe reference begins to "pollute" 2221 // the alias analysis of the rest of the graph, either Compile::can_alias 2222 // or Compile::must_alias will throw a diagnostic assert.) 2223 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2224 2225 if (!is_store && type == T_OBJECT) { 2226 // Attempt to infer a sharper value type from the offset and base type. 2227 ciKlass* sharpened_klass = NULL; 2228 2229 // See if it is an instance field, with an object type. 2230 if (alias_type->field() != NULL) { 2231 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2232 if (alias_type->field()->type()->is_klass()) { 2233 sharpened_klass = alias_type->field()->type()->as_klass(); 2234 } 2235 } 2236 2237 // See if it is a narrow oop array. 2238 if (adr_type->isa_aryptr()) { 2239 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2240 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2241 if (elem_type != NULL) { 2242 sharpened_klass = elem_type->klass(); 2243 } 2244 } 2245 } 2246 2247 if (sharpened_klass != NULL) { 2248 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2249 2250 // Sharpen the value type. 2251 value_type = tjp; 2252 2253 #ifndef PRODUCT 2254 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2255 tty->print(" from base type: "); adr_type->dump(); 2256 tty->print(" sharpened value: "); value_type->dump(); 2257 } 2258 #endif 2259 } 2260 } 2261 2262 // Null check on self without removing any arguments. The argument 2263 // null check technically happens in the wrong place, which can lead to 2264 // invalid stack traces when the primitive is inlined into a method 2265 // which handles NullPointerExceptions. 2266 _sp += nargs; 2267 do_null_check(receiver, T_OBJECT); 2268 _sp -= nargs; 2269 if (stopped()) { 2270 return true; 2271 } 2272 // Heap pointers get a null-check from the interpreter, 2273 // as a courtesy. However, this is not guaranteed by Unsafe, 2274 // and it is not possible to fully distinguish unintended nulls 2275 // from intended ones in this API. 2276 2277 if (is_volatile) { 2278 // We need to emit leading and trailing CPU membars (see below) in 2279 // addition to memory membars when is_volatile. This is a little 2280 // too strong, but avoids the need to insert per-alias-type 2281 // volatile membars (for stores; compare Parse::do_put_xxx), which 2282 // we cannot do effectively here because we probably only have a 2283 // rough approximation of type. 2284 need_mem_bar = true; 2285 // For Stores, place a memory ordering barrier now. 2286 if (is_store) 2287 insert_mem_bar(Op_MemBarRelease); 2288 } 2289 2290 // Memory barrier to prevent normal and 'unsafe' accesses from 2291 // bypassing each other. Happens after null checks, so the 2292 // exception paths do not take memory state from the memory barrier, 2293 // so there's no problems making a strong assert about mixing users 2294 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2295 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2296 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2297 2298 if (!is_store) { 2299 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2300 // load value and push onto stack 2301 switch (type) { 2302 case T_BOOLEAN: 2303 case T_CHAR: 2304 case T_BYTE: 2305 case T_SHORT: 2306 case T_INT: 2307 case T_FLOAT: 2308 case T_OBJECT: 2309 push( p ); 2310 break; 2311 case T_ADDRESS: 2312 // Cast to an int type. 2313 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); 2314 p = ConvX2L(p); 2315 push_pair(p); 2316 break; 2317 case T_DOUBLE: 2318 case T_LONG: 2319 push_pair( p ); 2320 break; 2321 default: ShouldNotReachHere(); 2322 } 2323 } else { 2324 // place effect of store into memory 2325 switch (type) { 2326 case T_DOUBLE: 2327 val = dstore_rounding(val); 2328 break; 2329 case T_ADDRESS: 2330 // Repackage the long as a pointer. 2331 val = ConvL2X(val); 2332 val = _gvn.transform( new (C, 2) CastX2PNode(val) ); 2333 break; 2334 } 2335 2336 if (type != T_OBJECT ) { 2337 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2338 } else { 2339 // Possibly an oop being stored to Java heap or native memory 2340 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2341 // oop to Java heap. 2342 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2343 } else { 2344 // We can't tell at compile time if we are storing in the Java heap or outside 2345 // of it. So we need to emit code to conditionally do the proper type of 2346 // store. 2347 2348 IdealKit ideal(gvn(), control(), merged_memory()); 2349 #define __ ideal. 2350 // QQQ who knows what probability is here?? 2351 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2352 // Sync IdealKit and graphKit. 2353 set_all_memory( __ merged_memory()); 2354 set_control(__ ctrl()); 2355 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2356 // Update IdealKit memory. 2357 __ set_all_memory(merged_memory()); 2358 __ set_ctrl(control()); 2359 } __ else_(); { 2360 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile); 2361 } __ end_if(); 2362 // Final sync IdealKit and GraphKit. 2363 sync_kit(ideal); 2364 #undef __ 2365 } 2366 } 2367 } 2368 2369 if (is_volatile) { 2370 if (!is_store) 2371 insert_mem_bar(Op_MemBarAcquire); 2372 else 2373 insert_mem_bar(Op_MemBarVolatile); 2374 } 2375 2376 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2377 2378 return true; 2379 } 2380 2381 //----------------------------inline_unsafe_prefetch---------------------------- 2382 2383 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2384 #ifndef PRODUCT 2385 { 2386 ResourceMark rm; 2387 // Check the signatures. 2388 ciSignature* sig = signature(); 2389 #ifdef ASSERT 2390 // Object getObject(Object base, int/long offset), etc. 2391 BasicType rtype = sig->return_type()->basic_type(); 2392 if (!is_native_ptr) { 2393 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2394 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2395 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2396 } else { 2397 assert(sig->count() == 1, "native prefetch has 1 argument"); 2398 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2399 } 2400 #endif // ASSERT 2401 } 2402 #endif // !PRODUCT 2403 2404 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2405 2406 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args 2407 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); 2408 2409 debug_only(int saved_sp = _sp); 2410 _sp += nargs; 2411 2412 // Build address expression. See the code in inline_unsafe_access. 2413 Node *adr; 2414 if (!is_native_ptr) { 2415 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2416 Node* offset = pop_pair(); 2417 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2418 Node* base = pop(); 2419 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2420 // to be plain byte offsets, which are also the same as those accepted 2421 // by oopDesc::field_base. 2422 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2423 "fieldOffset must be byte-scaled"); 2424 // 32-bit machines ignore the high half! 2425 offset = ConvL2X(offset); 2426 adr = make_unsafe_address(base, offset); 2427 } else { 2428 Node* ptr = pop_pair(); 2429 // Adjust Java long to machine word: 2430 ptr = ConvL2X(ptr); 2431 adr = make_unsafe_address(NULL, ptr); 2432 } 2433 2434 if (is_static) { 2435 assert(saved_sp == _sp, "must have correct argument count"); 2436 } else { 2437 // Pop receiver last: it was pushed first. 2438 Node *receiver = pop(); 2439 assert(saved_sp == _sp, "must have correct argument count"); 2440 2441 // Null check on self without removing any arguments. The argument 2442 // null check technically happens in the wrong place, which can lead to 2443 // invalid stack traces when the primitive is inlined into a method 2444 // which handles NullPointerExceptions. 2445 _sp += nargs; 2446 do_null_check(receiver, T_OBJECT); 2447 _sp -= nargs; 2448 if (stopped()) { 2449 return true; 2450 } 2451 } 2452 2453 // Generate the read or write prefetch 2454 Node *prefetch; 2455 if (is_store) { 2456 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); 2457 } else { 2458 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); 2459 } 2460 prefetch->init_req(0, control()); 2461 set_i_o(_gvn.transform(prefetch)); 2462 2463 return true; 2464 } 2465 2466 //----------------------------inline_unsafe_CAS---------------------------- 2467 2468 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { 2469 // This basic scheme here is the same as inline_unsafe_access, but 2470 // differs in enough details that combining them would make the code 2471 // overly confusing. (This is a true fact! I originally combined 2472 // them, but even I was confused by it!) As much code/comments as 2473 // possible are retained from inline_unsafe_access though to make 2474 // the correspondences clearer. - dl 2475 2476 if (callee()->is_static()) return false; // caller must have the capability! 2477 2478 #ifndef PRODUCT 2479 { 2480 ResourceMark rm; 2481 // Check the signatures. 2482 ciSignature* sig = signature(); 2483 #ifdef ASSERT 2484 BasicType rtype = sig->return_type()->basic_type(); 2485 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2486 assert(sig->count() == 4, "CAS has 4 arguments"); 2487 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2488 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2489 #endif // ASSERT 2490 } 2491 #endif //PRODUCT 2492 2493 // number of stack slots per value argument (1 or 2) 2494 int type_words = type2size[type]; 2495 2496 // Cannot inline wide CAS on machines that don't support it natively 2497 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8()) 2498 return false; 2499 2500 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2501 2502 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; 2503 int nargs = 1 + 1 + 2 + type_words + type_words; 2504 2505 // pop arguments: newval, oldval, offset, base, and receiver 2506 debug_only(int saved_sp = _sp); 2507 _sp += nargs; 2508 Node* newval = (type_words == 1) ? pop() : pop_pair(); 2509 Node* oldval = (type_words == 1) ? pop() : pop_pair(); 2510 Node *offset = pop_pair(); 2511 Node *base = pop(); 2512 Node *receiver = pop(); 2513 assert(saved_sp == _sp, "must have correct argument count"); 2514 2515 // Null check receiver. 2516 _sp += nargs; 2517 do_null_check(receiver, T_OBJECT); 2518 _sp -= nargs; 2519 if (stopped()) { 2520 return true; 2521 } 2522 2523 // Build field offset expression. 2524 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2525 // to be plain byte offsets, which are also the same as those accepted 2526 // by oopDesc::field_base. 2527 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2528 // 32-bit machines ignore the high half of long offsets 2529 offset = ConvL2X(offset); 2530 Node* adr = make_unsafe_address(base, offset); 2531 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2532 2533 // (Unlike inline_unsafe_access, there seems no point in trying 2534 // to refine types. Just use the coarse types here. 2535 const Type *value_type = Type::get_const_basic_type(type); 2536 Compile::AliasType* alias_type = C->alias_type(adr_type); 2537 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2538 int alias_idx = C->get_alias_index(adr_type); 2539 2540 // Memory-model-wise, a CAS acts like a little synchronized block, 2541 // so needs barriers on each side. These don't translate into 2542 // actual barriers on most machines, but we still need rest of 2543 // compiler to respect ordering. 2544 2545 insert_mem_bar(Op_MemBarRelease); 2546 insert_mem_bar(Op_MemBarCPUOrder); 2547 2548 // 4984716: MemBars must be inserted before this 2549 // memory node in order to avoid a false 2550 // dependency which will confuse the scheduler. 2551 Node *mem = memory(alias_idx); 2552 2553 // For now, we handle only those cases that actually exist: ints, 2554 // longs, and Object. Adding others should be straightforward. 2555 Node* cas; 2556 switch(type) { 2557 case T_INT: 2558 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2559 break; 2560 case T_LONG: 2561 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2562 break; 2563 case T_OBJECT: 2564 // reference stores need a store barrier. 2565 // (They don't if CAS fails, but it isn't worth checking.) 2566 pre_barrier(true /* do_load*/, 2567 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2568 NULL /* pre_val*/, 2569 T_OBJECT); 2570 #ifdef _LP64 2571 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2572 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2573 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2574 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr, 2575 newval_enc, oldval_enc)); 2576 } else 2577 #endif 2578 { 2579 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2580 } 2581 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); 2582 break; 2583 default: 2584 ShouldNotReachHere(); 2585 break; 2586 } 2587 2588 // SCMemProjNodes represent the memory state of CAS. Their main 2589 // role is to prevent CAS nodes from being optimized away when their 2590 // results aren't used. 2591 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 2592 set_memory(proj, alias_idx); 2593 2594 // Add the trailing membar surrounding the access 2595 insert_mem_bar(Op_MemBarCPUOrder); 2596 insert_mem_bar(Op_MemBarAcquire); 2597 2598 push(cas); 2599 return true; 2600 } 2601 2602 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2603 // This is another variant of inline_unsafe_access, differing in 2604 // that it always issues store-store ("release") barrier and ensures 2605 // store-atomicity (which only matters for "long"). 2606 2607 if (callee()->is_static()) return false; // caller must have the capability! 2608 2609 #ifndef PRODUCT 2610 { 2611 ResourceMark rm; 2612 // Check the signatures. 2613 ciSignature* sig = signature(); 2614 #ifdef ASSERT 2615 BasicType rtype = sig->return_type()->basic_type(); 2616 assert(rtype == T_VOID, "must return void"); 2617 assert(sig->count() == 3, "has 3 arguments"); 2618 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2619 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2620 #endif // ASSERT 2621 } 2622 #endif //PRODUCT 2623 2624 // number of stack slots per value argument (1 or 2) 2625 int type_words = type2size[type]; 2626 2627 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2628 2629 // Argument words: "this" plus oop plus offset plus value; 2630 int nargs = 1 + 1 + 2 + type_words; 2631 2632 // pop arguments: val, offset, base, and receiver 2633 debug_only(int saved_sp = _sp); 2634 _sp += nargs; 2635 Node* val = (type_words == 1) ? pop() : pop_pair(); 2636 Node *offset = pop_pair(); 2637 Node *base = pop(); 2638 Node *receiver = pop(); 2639 assert(saved_sp == _sp, "must have correct argument count"); 2640 2641 // Null check receiver. 2642 _sp += nargs; 2643 do_null_check(receiver, T_OBJECT); 2644 _sp -= nargs; 2645 if (stopped()) { 2646 return true; 2647 } 2648 2649 // Build field offset expression. 2650 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2651 // 32-bit machines ignore the high half of long offsets 2652 offset = ConvL2X(offset); 2653 Node* adr = make_unsafe_address(base, offset); 2654 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2655 const Type *value_type = Type::get_const_basic_type(type); 2656 Compile::AliasType* alias_type = C->alias_type(adr_type); 2657 2658 insert_mem_bar(Op_MemBarRelease); 2659 insert_mem_bar(Op_MemBarCPUOrder); 2660 // Ensure that the store is atomic for longs: 2661 bool require_atomic_access = true; 2662 Node* store; 2663 if (type == T_OBJECT) // reference stores need a store barrier. 2664 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2665 else { 2666 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2667 } 2668 insert_mem_bar(Op_MemBarCPUOrder); 2669 return true; 2670 } 2671 2672 bool LibraryCallKit::inline_unsafe_allocate() { 2673 if (callee()->is_static()) return false; // caller must have the capability! 2674 int nargs = 1 + 1; 2675 assert(signature()->size() == nargs-1, "alloc has 1 argument"); 2676 null_check_receiver(callee()); // check then ignore argument(0) 2677 _sp += nargs; // set original stack for use by uncommon_trap 2678 Node* cls = do_null_check(argument(1), T_OBJECT); 2679 _sp -= nargs; 2680 if (stopped()) return true; 2681 2682 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 2683 _sp += nargs; // set original stack for use by uncommon_trap 2684 kls = do_null_check(kls, T_OBJECT); 2685 _sp -= nargs; 2686 if (stopped()) return true; // argument was like int.class 2687 2688 // Note: The argument might still be an illegal value like 2689 // Serializable.class or Object[].class. The runtime will handle it. 2690 // But we must make an explicit check for initialization. 2691 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc)); 2692 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT); 2693 Node* bits = intcon(instanceKlass::fully_initialized); 2694 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); 2695 // The 'test' is non-zero if we need to take a slow path. 2696 2697 Node* obj = new_instance(kls, test); 2698 push(obj); 2699 2700 return true; 2701 } 2702 2703 //------------------------inline_native_time_funcs-------------- 2704 // inline code for System.currentTimeMillis() and System.nanoTime() 2705 // these have the same type and signature 2706 bool LibraryCallKit::inline_native_time_funcs(bool isNano) { 2707 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) : 2708 CAST_FROM_FN_PTR(address, os::javaTimeMillis); 2709 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis"; 2710 const TypeFunc *tf = OptoRuntime::current_time_millis_Type(); 2711 const TypePtr* no_memory_effects = NULL; 2712 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2713 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); 2714 #ifdef ASSERT 2715 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); 2716 assert(value_top == top(), "second value must be top"); 2717 #endif 2718 push_pair(value); 2719 return true; 2720 } 2721 2722 //------------------------inline_native_currentThread------------------ 2723 bool LibraryCallKit::inline_native_currentThread() { 2724 Node* junk = NULL; 2725 push(generate_current_thread(junk)); 2726 return true; 2727 } 2728 2729 //------------------------inline_native_isInterrupted------------------ 2730 bool LibraryCallKit::inline_native_isInterrupted() { 2731 const int nargs = 1+1; // receiver + boolean 2732 assert(nargs == arg_size(), "sanity"); 2733 // Add a fast path to t.isInterrupted(clear_int): 2734 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 2735 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 2736 // So, in the common case that the interrupt bit is false, 2737 // we avoid making a call into the VM. Even if the interrupt bit 2738 // is true, if the clear_int argument is false, we avoid the VM call. 2739 // However, if the receiver is not currentThread, we must call the VM, 2740 // because there must be some locking done around the operation. 2741 2742 // We only go to the fast case code if we pass two guards. 2743 // Paths which do not pass are accumulated in the slow_region. 2744 RegionNode* slow_region = new (C, 1) RegionNode(1); 2745 record_for_igvn(slow_region); 2746 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow 2747 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); 2748 enum { no_int_result_path = 1, 2749 no_clear_result_path = 2, 2750 slow_result_path = 3 2751 }; 2752 2753 // (a) Receiving thread must be the current thread. 2754 Node* rec_thr = argument(0); 2755 Node* tls_ptr = NULL; 2756 Node* cur_thr = generate_current_thread(tls_ptr); 2757 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); 2758 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); 2759 2760 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); 2761 if (!known_current_thread) 2762 generate_slow_guard(bol_thr, slow_region); 2763 2764 // (b) Interrupt bit on TLS must be false. 2765 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2766 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 2767 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 2768 // Set the control input on the field _interrupted read to prevent it floating up. 2769 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 2770 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); 2771 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); 2772 2773 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2774 2775 // First fast path: if (!TLS._interrupted) return false; 2776 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); 2777 result_rgn->init_req(no_int_result_path, false_bit); 2778 result_val->init_req(no_int_result_path, intcon(0)); 2779 2780 // drop through to next case 2781 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); 2782 2783 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 2784 Node* clr_arg = argument(1); 2785 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); 2786 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); 2787 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 2788 2789 // Second fast path: ... else if (!clear_int) return true; 2790 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); 2791 result_rgn->init_req(no_clear_result_path, false_arg); 2792 result_val->init_req(no_clear_result_path, intcon(1)); 2793 2794 // drop through to next case 2795 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); 2796 2797 // (d) Otherwise, go to the slow path. 2798 slow_region->add_req(control()); 2799 set_control( _gvn.transform(slow_region) ); 2800 2801 if (stopped()) { 2802 // There is no slow path. 2803 result_rgn->init_req(slow_result_path, top()); 2804 result_val->init_req(slow_result_path, top()); 2805 } else { 2806 // non-virtual because it is a private non-static 2807 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 2808 2809 Node* slow_val = set_results_for_java_call(slow_call); 2810 // this->control() comes from set_results_for_java_call 2811 2812 // If we know that the result of the slow call will be true, tell the optimizer! 2813 if (known_current_thread) slow_val = intcon(1); 2814 2815 Node* fast_io = slow_call->in(TypeFunc::I_O); 2816 Node* fast_mem = slow_call->in(TypeFunc::Memory); 2817 // These two phis are pre-filled with copies of of the fast IO and Memory 2818 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 2819 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 2820 2821 result_rgn->init_req(slow_result_path, control()); 2822 io_phi ->init_req(slow_result_path, i_o()); 2823 mem_phi ->init_req(slow_result_path, reset_memory()); 2824 result_val->init_req(slow_result_path, slow_val); 2825 2826 set_all_memory( _gvn.transform(mem_phi) ); 2827 set_i_o( _gvn.transform(io_phi) ); 2828 } 2829 2830 push_result(result_rgn, result_val); 2831 C->set_has_split_ifs(true); // Has chance for split-if optimization 2832 2833 return true; 2834 } 2835 2836 //---------------------------load_mirror_from_klass---------------------------- 2837 // Given a klass oop, load its java mirror (a java.lang.Class oop). 2838 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 2839 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc)); 2840 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 2841 } 2842 2843 //-----------------------load_klass_from_mirror_common------------------------- 2844 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 2845 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 2846 // and branch to the given path on the region. 2847 // If never_see_null, take an uncommon trap on null, so we can optimistically 2848 // compile for the non-null case. 2849 // If the region is NULL, force never_see_null = true. 2850 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 2851 bool never_see_null, 2852 int nargs, 2853 RegionNode* region, 2854 int null_path, 2855 int offset) { 2856 if (region == NULL) never_see_null = true; 2857 Node* p = basic_plus_adr(mirror, offset); 2858 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 2859 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); 2860 _sp += nargs; // any deopt will start just before call to enclosing method 2861 Node* null_ctl = top(); 2862 kls = null_check_oop(kls, &null_ctl, never_see_null); 2863 if (region != NULL) { 2864 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 2865 region->init_req(null_path, null_ctl); 2866 } else { 2867 assert(null_ctl == top(), "no loose ends"); 2868 } 2869 _sp -= nargs; 2870 return kls; 2871 } 2872 2873 //--------------------(inline_native_Class_query helpers)--------------------- 2874 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 2875 // Fall through if (mods & mask) == bits, take the guard otherwise. 2876 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 2877 // Branch around if the given klass has the given modifier bit set. 2878 // Like generate_guard, adds a new path onto the region. 2879 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); 2880 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 2881 Node* mask = intcon(modifier_mask); 2882 Node* bits = intcon(modifier_bits); 2883 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); 2884 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); 2885 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); 2886 return generate_fair_guard(bol, region); 2887 } 2888 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 2889 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 2890 } 2891 2892 //-------------------------inline_native_Class_query------------------- 2893 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 2894 int nargs = 1+0; // just the Class mirror, in most cases 2895 const Type* return_type = TypeInt::BOOL; 2896 Node* prim_return_value = top(); // what happens if it's a primitive class? 2897 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 2898 bool expect_prim = false; // most of these guys expect to work on refs 2899 2900 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 2901 2902 switch (id) { 2903 case vmIntrinsics::_isInstance: 2904 nargs = 1+1; // the Class mirror, plus the object getting queried about 2905 // nothing is an instance of a primitive type 2906 prim_return_value = intcon(0); 2907 break; 2908 case vmIntrinsics::_getModifiers: 2909 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2910 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 2911 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 2912 break; 2913 case vmIntrinsics::_isInterface: 2914 prim_return_value = intcon(0); 2915 break; 2916 case vmIntrinsics::_isArray: 2917 prim_return_value = intcon(0); 2918 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 2919 break; 2920 case vmIntrinsics::_isPrimitive: 2921 prim_return_value = intcon(1); 2922 expect_prim = true; // obviously 2923 break; 2924 case vmIntrinsics::_getSuperclass: 2925 prim_return_value = null(); 2926 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 2927 break; 2928 case vmIntrinsics::_getComponentType: 2929 prim_return_value = null(); 2930 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 2931 break; 2932 case vmIntrinsics::_getClassAccessFlags: 2933 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 2934 return_type = TypeInt::INT; // not bool! 6297094 2935 break; 2936 default: 2937 ShouldNotReachHere(); 2938 } 2939 2940 Node* mirror = argument(0); 2941 Node* obj = (nargs <= 1)? top(): argument(1); 2942 2943 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 2944 if (mirror_con == NULL) return false; // cannot happen? 2945 2946 #ifndef PRODUCT 2947 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2948 ciType* k = mirror_con->java_mirror_type(); 2949 if (k) { 2950 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 2951 k->print_name(); 2952 tty->cr(); 2953 } 2954 } 2955 #endif 2956 2957 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 2958 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2959 record_for_igvn(region); 2960 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); 2961 2962 // The mirror will never be null of Reflection.getClassAccessFlags, however 2963 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 2964 // if it is. See bug 4774291. 2965 2966 // For Reflection.getClassAccessFlags(), the null check occurs in 2967 // the wrong place; see inline_unsafe_access(), above, for a similar 2968 // situation. 2969 _sp += nargs; // set original stack for use by uncommon_trap 2970 mirror = do_null_check(mirror, T_OBJECT); 2971 _sp -= nargs; 2972 // If mirror or obj is dead, only null-path is taken. 2973 if (stopped()) return true; 2974 2975 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 2976 2977 // Now load the mirror's klass metaobject, and null-check it. 2978 // Side-effects region with the control path if the klass is null. 2979 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, 2980 region, _prim_path); 2981 // If kls is null, we have a primitive mirror. 2982 phi->init_req(_prim_path, prim_return_value); 2983 if (stopped()) { push_result(region, phi); return true; } 2984 2985 Node* p; // handy temp 2986 Node* null_ctl; 2987 2988 // Now that we have the non-null klass, we can perform the real query. 2989 // For constant classes, the query will constant-fold in LoadNode::Value. 2990 Node* query_value = top(); 2991 switch (id) { 2992 case vmIntrinsics::_isInstance: 2993 // nothing is an instance of a primitive type 2994 _sp += nargs; // gen_instanceof might do an uncommon trap 2995 query_value = gen_instanceof(obj, kls); 2996 _sp -= nargs; 2997 break; 2998 2999 case vmIntrinsics::_getModifiers: 3000 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc)); 3001 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3002 break; 3003 3004 case vmIntrinsics::_isInterface: 3005 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3006 if (generate_interface_guard(kls, region) != NULL) 3007 // A guard was added. If the guard is taken, it was an interface. 3008 phi->add_req(intcon(1)); 3009 // If we fall through, it's a plain class. 3010 query_value = intcon(0); 3011 break; 3012 3013 case vmIntrinsics::_isArray: 3014 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3015 if (generate_array_guard(kls, region) != NULL) 3016 // A guard was added. If the guard is taken, it was an array. 3017 phi->add_req(intcon(1)); 3018 // If we fall through, it's a plain class. 3019 query_value = intcon(0); 3020 break; 3021 3022 case vmIntrinsics::_isPrimitive: 3023 query_value = intcon(0); // "normal" path produces false 3024 break; 3025 3026 case vmIntrinsics::_getSuperclass: 3027 // The rules here are somewhat unfortunate, but we can still do better 3028 // with random logic than with a JNI call. 3029 // Interfaces store null or Object as _super, but must report null. 3030 // Arrays store an intermediate super as _super, but must report Object. 3031 // Other types can report the actual _super. 3032 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3033 if (generate_interface_guard(kls, region) != NULL) 3034 // A guard was added. If the guard is taken, it was an interface. 3035 phi->add_req(null()); 3036 if (generate_array_guard(kls, region) != NULL) 3037 // A guard was added. If the guard is taken, it was an array. 3038 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3039 // If we fall through, it's a plain class. Get its _super. 3040 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc)); 3041 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); 3042 null_ctl = top(); 3043 kls = null_check_oop(kls, &null_ctl); 3044 if (null_ctl != top()) { 3045 // If the guard is taken, Object.superClass is null (both klass and mirror). 3046 region->add_req(null_ctl); 3047 phi ->add_req(null()); 3048 } 3049 if (!stopped()) { 3050 query_value = load_mirror_from_klass(kls); 3051 } 3052 break; 3053 3054 case vmIntrinsics::_getComponentType: 3055 if (generate_array_guard(kls, region) != NULL) { 3056 // Be sure to pin the oop load to the guard edge just created: 3057 Node* is_array_ctrl = region->in(region->req()-1); 3058 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc)); 3059 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 3060 phi->add_req(cmo); 3061 } 3062 query_value = null(); // non-array case is null 3063 break; 3064 3065 case vmIntrinsics::_getClassAccessFlags: 3066 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc)); 3067 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3068 break; 3069 3070 default: 3071 ShouldNotReachHere(); 3072 } 3073 3074 // Fall-through is the normal case of a query to a real class. 3075 phi->init_req(1, query_value); 3076 region->init_req(1, control()); 3077 3078 push_result(region, phi); 3079 C->set_has_split_ifs(true); // Has chance for split-if optimization 3080 3081 return true; 3082 } 3083 3084 //--------------------------inline_native_subtype_check------------------------ 3085 // This intrinsic takes the JNI calls out of the heart of 3086 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3087 bool LibraryCallKit::inline_native_subtype_check() { 3088 int nargs = 1+1; // the Class mirror, plus the other class getting examined 3089 3090 // Pull both arguments off the stack. 3091 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3092 args[0] = argument(0); 3093 args[1] = argument(1); 3094 Node* klasses[2]; // corresponding Klasses: superk, subk 3095 klasses[0] = klasses[1] = top(); 3096 3097 enum { 3098 // A full decision tree on {superc is prim, subc is prim}: 3099 _prim_0_path = 1, // {P,N} => false 3100 // {P,P} & superc!=subc => false 3101 _prim_same_path, // {P,P} & superc==subc => true 3102 _prim_1_path, // {N,P} => false 3103 _ref_subtype_path, // {N,N} & subtype check wins => true 3104 _both_ref_path, // {N,N} & subtype check loses => false 3105 PATH_LIMIT 3106 }; 3107 3108 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3109 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 3110 record_for_igvn(region); 3111 3112 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3113 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3114 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3115 3116 // First null-check both mirrors and load each mirror's klass metaobject. 3117 int which_arg; 3118 for (which_arg = 0; which_arg <= 1; which_arg++) { 3119 Node* arg = args[which_arg]; 3120 _sp += nargs; // set original stack for use by uncommon_trap 3121 arg = do_null_check(arg, T_OBJECT); 3122 _sp -= nargs; 3123 if (stopped()) break; 3124 args[which_arg] = _gvn.transform(arg); 3125 3126 Node* p = basic_plus_adr(arg, class_klass_offset); 3127 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3128 klasses[which_arg] = _gvn.transform(kls); 3129 } 3130 3131 // Having loaded both klasses, test each for null. 3132 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3133 for (which_arg = 0; which_arg <= 1; which_arg++) { 3134 Node* kls = klasses[which_arg]; 3135 Node* null_ctl = top(); 3136 _sp += nargs; // set original stack for use by uncommon_trap 3137 kls = null_check_oop(kls, &null_ctl, never_see_null); 3138 _sp -= nargs; 3139 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3140 region->init_req(prim_path, null_ctl); 3141 if (stopped()) break; 3142 klasses[which_arg] = kls; 3143 } 3144 3145 if (!stopped()) { 3146 // now we have two reference types, in klasses[0..1] 3147 Node* subk = klasses[1]; // the argument to isAssignableFrom 3148 Node* superk = klasses[0]; // the receiver 3149 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3150 // now we have a successful reference subtype check 3151 region->set_req(_ref_subtype_path, control()); 3152 } 3153 3154 // If both operands are primitive (both klasses null), then 3155 // we must return true when they are identical primitives. 3156 // It is convenient to test this after the first null klass check. 3157 set_control(region->in(_prim_0_path)); // go back to first null check 3158 if (!stopped()) { 3159 // Since superc is primitive, make a guard for the superc==subc case. 3160 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); 3161 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); 3162 generate_guard(bol_eq, region, PROB_FAIR); 3163 if (region->req() == PATH_LIMIT+1) { 3164 // A guard was added. If the added guard is taken, superc==subc. 3165 region->swap_edges(PATH_LIMIT, _prim_same_path); 3166 region->del_req(PATH_LIMIT); 3167 } 3168 region->set_req(_prim_0_path, control()); // Not equal after all. 3169 } 3170 3171 // these are the only paths that produce 'true': 3172 phi->set_req(_prim_same_path, intcon(1)); 3173 phi->set_req(_ref_subtype_path, intcon(1)); 3174 3175 // pull together the cases: 3176 assert(region->req() == PATH_LIMIT, "sane region"); 3177 for (uint i = 1; i < region->req(); i++) { 3178 Node* ctl = region->in(i); 3179 if (ctl == NULL || ctl == top()) { 3180 region->set_req(i, top()); 3181 phi ->set_req(i, top()); 3182 } else if (phi->in(i) == NULL) { 3183 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3184 } 3185 } 3186 3187 set_control(_gvn.transform(region)); 3188 push(_gvn.transform(phi)); 3189 3190 return true; 3191 } 3192 3193 //---------------------generate_array_guard_common------------------------ 3194 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3195 bool obj_array, bool not_array) { 3196 // If obj_array/non_array==false/false: 3197 // Branch around if the given klass is in fact an array (either obj or prim). 3198 // If obj_array/non_array==false/true: 3199 // Branch around if the given klass is not an array klass of any kind. 3200 // If obj_array/non_array==true/true: 3201 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3202 // If obj_array/non_array==true/false: 3203 // Branch around if the kls is an oop array (Object[] or subtype) 3204 // 3205 // Like generate_guard, adds a new path onto the region. 3206 jint layout_con = 0; 3207 Node* layout_val = get_layout_helper(kls, layout_con); 3208 if (layout_val == NULL) { 3209 bool query = (obj_array 3210 ? Klass::layout_helper_is_objArray(layout_con) 3211 : Klass::layout_helper_is_javaArray(layout_con)); 3212 if (query == not_array) { 3213 return NULL; // never a branch 3214 } else { // always a branch 3215 Node* always_branch = control(); 3216 if (region != NULL) 3217 region->add_req(always_branch); 3218 set_control(top()); 3219 return always_branch; 3220 } 3221 } 3222 // Now test the correct condition. 3223 jint nval = (obj_array 3224 ? ((jint)Klass::_lh_array_tag_type_value 3225 << Klass::_lh_array_tag_shift) 3226 : Klass::_lh_neutral_value); 3227 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); 3228 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3229 // invert the test if we are looking for a non-array 3230 if (not_array) btest = BoolTest(btest).negate(); 3231 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); 3232 return generate_fair_guard(bol, region); 3233 } 3234 3235 3236 //-----------------------inline_native_newArray-------------------------- 3237 bool LibraryCallKit::inline_native_newArray() { 3238 int nargs = 2; 3239 Node* mirror = argument(0); 3240 Node* count_val = argument(1); 3241 3242 _sp += nargs; // set original stack for use by uncommon_trap 3243 mirror = do_null_check(mirror, T_OBJECT); 3244 _sp -= nargs; 3245 // If mirror or obj is dead, only null-path is taken. 3246 if (stopped()) return true; 3247 3248 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3249 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3250 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3251 TypeInstPtr::NOTNULL); 3252 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3253 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3254 TypePtr::BOTTOM); 3255 3256 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3257 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3258 nargs, 3259 result_reg, _slow_path); 3260 Node* normal_ctl = control(); 3261 Node* no_array_ctl = result_reg->in(_slow_path); 3262 3263 // Generate code for the slow case. We make a call to newArray(). 3264 set_control(no_array_ctl); 3265 if (!stopped()) { 3266 // Either the input type is void.class, or else the 3267 // array klass has not yet been cached. Either the 3268 // ensuing call will throw an exception, or else it 3269 // will cache the array klass for next time. 3270 PreserveJVMState pjvms(this); 3271 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3272 Node* slow_result = set_results_for_java_call(slow_call); 3273 // this->control() comes from set_results_for_java_call 3274 result_reg->set_req(_slow_path, control()); 3275 result_val->set_req(_slow_path, slow_result); 3276 result_io ->set_req(_slow_path, i_o()); 3277 result_mem->set_req(_slow_path, reset_memory()); 3278 } 3279 3280 set_control(normal_ctl); 3281 if (!stopped()) { 3282 // Normal case: The array type has been cached in the java.lang.Class. 3283 // The following call works fine even if the array type is polymorphic. 3284 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3285 Node* obj = new_array(klass_node, count_val, nargs); 3286 result_reg->init_req(_normal_path, control()); 3287 result_val->init_req(_normal_path, obj); 3288 result_io ->init_req(_normal_path, i_o()); 3289 result_mem->init_req(_normal_path, reset_memory()); 3290 } 3291 3292 // Return the combined state. 3293 set_i_o( _gvn.transform(result_io) ); 3294 set_all_memory( _gvn.transform(result_mem) ); 3295 push_result(result_reg, result_val); 3296 C->set_has_split_ifs(true); // Has chance for split-if optimization 3297 3298 return true; 3299 } 3300 3301 //----------------------inline_native_getLength-------------------------- 3302 bool LibraryCallKit::inline_native_getLength() { 3303 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3304 3305 int nargs = 1; 3306 Node* array = argument(0); 3307 3308 _sp += nargs; // set original stack for use by uncommon_trap 3309 array = do_null_check(array, T_OBJECT); 3310 _sp -= nargs; 3311 3312 // If array is dead, only null-path is taken. 3313 if (stopped()) return true; 3314 3315 // Deoptimize if it is a non-array. 3316 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3317 3318 if (non_array != NULL) { 3319 PreserveJVMState pjvms(this); 3320 set_control(non_array); 3321 _sp += nargs; // push the arguments back on the stack 3322 uncommon_trap(Deoptimization::Reason_intrinsic, 3323 Deoptimization::Action_maybe_recompile); 3324 } 3325 3326 // If control is dead, only non-array-path is taken. 3327 if (stopped()) return true; 3328 3329 // The works fine even if the array type is polymorphic. 3330 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3331 push( load_array_length(array) ); 3332 3333 C->set_has_split_ifs(true); // Has chance for split-if optimization 3334 3335 return true; 3336 } 3337 3338 //------------------------inline_array_copyOf---------------------------- 3339 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3340 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3341 3342 // Restore the stack and pop off the arguments. 3343 int nargs = 3 + (is_copyOfRange? 1: 0); 3344 Node* original = argument(0); 3345 Node* start = is_copyOfRange? argument(1): intcon(0); 3346 Node* end = is_copyOfRange? argument(2): argument(1); 3347 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3348 3349 Node* newcopy; 3350 3351 //set the original stack and the reexecute bit for the interpreter to reexecute 3352 //the bytecode that invokes Arrays.copyOf if deoptimization happens 3353 { PreserveReexecuteState preexecs(this); 3354 _sp += nargs; 3355 jvms()->set_should_reexecute(true); 3356 3357 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); 3358 original = do_null_check(original, T_OBJECT); 3359 3360 // Check if a null path was taken unconditionally. 3361 if (stopped()) return true; 3362 3363 Node* orig_length = load_array_length(original); 3364 3365 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0, 3366 NULL, 0); 3367 klass_node = do_null_check(klass_node, T_OBJECT); 3368 3369 RegionNode* bailout = new (C, 1) RegionNode(1); 3370 record_for_igvn(bailout); 3371 3372 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3373 // Bail out if that is so. 3374 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3375 if (not_objArray != NULL) { 3376 // Improve the klass node's type from the new optimistic assumption: 3377 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3378 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3379 Node* cast = new (C, 2) CastPPNode(klass_node, akls); 3380 cast->init_req(0, control()); 3381 klass_node = _gvn.transform(cast); 3382 } 3383 3384 // Bail out if either start or end is negative. 3385 generate_negative_guard(start, bailout, &start); 3386 generate_negative_guard(end, bailout, &end); 3387 3388 Node* length = end; 3389 if (_gvn.type(start) != TypeInt::ZERO) { 3390 length = _gvn.transform( new (C, 3) SubINode(end, start) ); 3391 } 3392 3393 // Bail out if length is negative. 3394 // ...Not needed, since the new_array will throw the right exception. 3395 //generate_negative_guard(length, bailout, &length); 3396 3397 if (bailout->req() > 1) { 3398 PreserveJVMState pjvms(this); 3399 set_control( _gvn.transform(bailout) ); 3400 uncommon_trap(Deoptimization::Reason_intrinsic, 3401 Deoptimization::Action_maybe_recompile); 3402 } 3403 3404 if (!stopped()) { 3405 3406 // How many elements will we copy from the original? 3407 // The answer is MinI(orig_length - start, length). 3408 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); 3409 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3410 3411 const bool raw_mem_only = true; 3412 newcopy = new_array(klass_node, length, 0, raw_mem_only); 3413 3414 // Generate a direct call to the right arraycopy function(s). 3415 // We know the copy is disjoint but we might not know if the 3416 // oop stores need checking. 3417 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3418 // This will fail a store-check if x contains any non-nulls. 3419 bool disjoint_bases = true; 3420 bool length_never_negative = true; 3421 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3422 original, start, newcopy, intcon(0), moved, 3423 disjoint_bases, length_never_negative); 3424 } 3425 } //original reexecute and sp are set back here 3426 3427 if(!stopped()) { 3428 push(newcopy); 3429 } 3430 3431 C->set_has_split_ifs(true); // Has chance for split-if optimization 3432 3433 return true; 3434 } 3435 3436 3437 //----------------------generate_virtual_guard--------------------------- 3438 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3439 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3440 RegionNode* slow_region) { 3441 ciMethod* method = callee(); 3442 int vtable_index = method->vtable_index(); 3443 // Get the methodOop out of the appropriate vtable entry. 3444 int entry_offset = (instanceKlass::vtable_start_offset() + 3445 vtable_index*vtableEntry::size()) * wordSize + 3446 vtableEntry::method_offset_in_bytes(); 3447 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3448 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); 3449 3450 // Compare the target method with the expected method (e.g., Object.hashCode). 3451 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method); 3452 3453 Node* native_call = makecon(native_call_addr); 3454 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); 3455 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); 3456 3457 return generate_slow_guard(test_native, slow_region); 3458 } 3459 3460 //-----------------------generate_method_call---------------------------- 3461 // Use generate_method_call to make a slow-call to the real 3462 // method if the fast path fails. An alternative would be to 3463 // use a stub like OptoRuntime::slow_arraycopy_Java. 3464 // This only works for expanding the current library call, 3465 // not another intrinsic. (E.g., don't use this for making an 3466 // arraycopy call inside of the copyOf intrinsic.) 3467 CallJavaNode* 3468 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3469 // When compiling the intrinsic method itself, do not use this technique. 3470 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3471 3472 ciMethod* method = callee(); 3473 // ensure the JVMS we have will be correct for this call 3474 guarantee(method_id == method->intrinsic_id(), "must match"); 3475 3476 const TypeFunc* tf = TypeFunc::make(method); 3477 int tfdc = tf->domain()->cnt(); 3478 CallJavaNode* slow_call; 3479 if (is_static) { 3480 assert(!is_virtual, ""); 3481 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3482 SharedRuntime::get_resolve_static_call_stub(), 3483 method, bci()); 3484 } else if (is_virtual) { 3485 null_check_receiver(method); 3486 int vtable_index = methodOopDesc::invalid_vtable_index; 3487 if (UseInlineCaches) { 3488 // Suppress the vtable call 3489 } else { 3490 // hashCode and clone are not a miranda methods, 3491 // so the vtable index is fixed. 3492 // No need to use the linkResolver to get it. 3493 vtable_index = method->vtable_index(); 3494 } 3495 slow_call = new(C, tfdc) CallDynamicJavaNode(tf, 3496 SharedRuntime::get_resolve_virtual_call_stub(), 3497 method, vtable_index, bci()); 3498 } else { // neither virtual nor static: opt_virtual 3499 null_check_receiver(method); 3500 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3501 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3502 method, bci()); 3503 slow_call->set_optimized_virtual(true); 3504 } 3505 set_arguments_for_java_call(slow_call); 3506 set_edges_for_java_call(slow_call); 3507 return slow_call; 3508 } 3509 3510 3511 //------------------------------inline_native_hashcode-------------------- 3512 // Build special case code for calls to hashCode on an object. 3513 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3514 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3515 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3516 3517 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3518 3519 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3520 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3521 TypeInt::INT); 3522 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3523 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3524 TypePtr::BOTTOM); 3525 Node* obj = NULL; 3526 if (!is_static) { 3527 // Check for hashing null object 3528 obj = null_check_receiver(callee()); 3529 if (stopped()) return true; // unconditionally null 3530 result_reg->init_req(_null_path, top()); 3531 result_val->init_req(_null_path, top()); 3532 } else { 3533 // Do a null check, and return zero if null. 3534 // System.identityHashCode(null) == 0 3535 obj = argument(0); 3536 Node* null_ctl = top(); 3537 obj = null_check_oop(obj, &null_ctl); 3538 result_reg->init_req(_null_path, null_ctl); 3539 result_val->init_req(_null_path, _gvn.intcon(0)); 3540 } 3541 3542 // Unconditionally null? Then return right away. 3543 if (stopped()) { 3544 set_control( result_reg->in(_null_path) ); 3545 if (!stopped()) 3546 push( result_val ->in(_null_path) ); 3547 return true; 3548 } 3549 3550 // After null check, get the object's klass. 3551 Node* obj_klass = load_object_klass(obj); 3552 3553 // This call may be virtual (invokevirtual) or bound (invokespecial). 3554 // For each case we generate slightly different code. 3555 3556 // We only go to the fast case code if we pass a number of guards. The 3557 // paths which do not pass are accumulated in the slow_region. 3558 RegionNode* slow_region = new (C, 1) RegionNode(1); 3559 record_for_igvn(slow_region); 3560 3561 // If this is a virtual call, we generate a funny guard. We pull out 3562 // the vtable entry corresponding to hashCode() from the target object. 3563 // If the target method which we are calling happens to be the native 3564 // Object hashCode() method, we pass the guard. We do not need this 3565 // guard for non-virtual calls -- the caller is known to be the native 3566 // Object hashCode(). 3567 if (is_virtual) { 3568 generate_virtual_guard(obj_klass, slow_region); 3569 } 3570 3571 // Get the header out of the object, use LoadMarkNode when available 3572 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3573 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type()); 3574 3575 // Test the header to see if it is unlocked. 3576 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3577 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); 3578 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3579 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); 3580 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); 3581 3582 generate_slow_guard(test_unlocked, slow_region); 3583 3584 // Get the hash value and check to see that it has been properly assigned. 3585 // We depend on hash_mask being at most 32 bits and avoid the use of 3586 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3587 // vm: see markOop.hpp. 3588 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3589 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3590 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); 3591 // This hack lets the hash bits live anywhere in the mark object now, as long 3592 // as the shift drops the relevant bits into the low 32 bits. Note that 3593 // Java spec says that HashCode is an int so there's no point in capturing 3594 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3595 hshifted_header = ConvX2I(hshifted_header); 3596 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); 3597 3598 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3599 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); 3600 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); 3601 3602 generate_slow_guard(test_assigned, slow_region); 3603 3604 Node* init_mem = reset_memory(); 3605 // fill in the rest of the null path: 3606 result_io ->init_req(_null_path, i_o()); 3607 result_mem->init_req(_null_path, init_mem); 3608 3609 result_val->init_req(_fast_path, hash_val); 3610 result_reg->init_req(_fast_path, control()); 3611 result_io ->init_req(_fast_path, i_o()); 3612 result_mem->init_req(_fast_path, init_mem); 3613 3614 // Generate code for the slow case. We make a call to hashCode(). 3615 set_control(_gvn.transform(slow_region)); 3616 if (!stopped()) { 3617 // No need for PreserveJVMState, because we're using up the present state. 3618 set_all_memory(init_mem); 3619 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; 3620 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; 3621 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3622 Node* slow_result = set_results_for_java_call(slow_call); 3623 // this->control() comes from set_results_for_java_call 3624 result_reg->init_req(_slow_path, control()); 3625 result_val->init_req(_slow_path, slow_result); 3626 result_io ->set_req(_slow_path, i_o()); 3627 result_mem ->set_req(_slow_path, reset_memory()); 3628 } 3629 3630 // Return the combined state. 3631 set_i_o( _gvn.transform(result_io) ); 3632 set_all_memory( _gvn.transform(result_mem) ); 3633 push_result(result_reg, result_val); 3634 3635 return true; 3636 } 3637 3638 //---------------------------inline_native_getClass---------------------------- 3639 // Build special case code for calls to getClass on an object. 3640 bool LibraryCallKit::inline_native_getClass() { 3641 Node* obj = null_check_receiver(callee()); 3642 if (stopped()) return true; 3643 push( load_mirror_from_klass(load_object_klass(obj)) ); 3644 return true; 3645 } 3646 3647 //-----------------inline_native_Reflection_getCallerClass--------------------- 3648 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3649 // 3650 // NOTE that this code must perform the same logic as 3651 // vframeStream::security_get_caller_frame in that it must skip 3652 // Method.invoke() and auxiliary frames. 3653 3654 3655 3656 3657 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3658 ciMethod* method = callee(); 3659 3660 #ifndef PRODUCT 3661 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3662 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3663 } 3664 #endif 3665 3666 debug_only(int saved_sp = _sp); 3667 3668 // Argument words: (int depth) 3669 int nargs = 1; 3670 3671 _sp += nargs; 3672 Node* caller_depth_node = pop(); 3673 3674 assert(saved_sp == _sp, "must have correct argument count"); 3675 3676 // The depth value must be a constant in order for the runtime call 3677 // to be eliminated. 3678 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); 3679 if (caller_depth_type == NULL || !caller_depth_type->is_con()) { 3680 #ifndef PRODUCT 3681 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3682 tty->print_cr(" Bailing out because caller depth was not a constant"); 3683 } 3684 #endif 3685 return false; 3686 } 3687 // Note that the JVM state at this point does not include the 3688 // getCallerClass() frame which we are trying to inline. The 3689 // semantics of getCallerClass(), however, are that the "first" 3690 // frame is the getCallerClass() frame, so we subtract one from the 3691 // requested depth before continuing. We don't inline requests of 3692 // getCallerClass(0). 3693 int caller_depth = caller_depth_type->get_con() - 1; 3694 if (caller_depth < 0) { 3695 #ifndef PRODUCT 3696 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3697 tty->print_cr(" Bailing out because caller depth was %d", caller_depth); 3698 } 3699 #endif 3700 return false; 3701 } 3702 3703 if (!jvms()->has_method()) { 3704 #ifndef PRODUCT 3705 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3706 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3707 } 3708 #endif 3709 return false; 3710 } 3711 int _depth = jvms()->depth(); // cache call chain depth 3712 3713 // Walk back up the JVM state to find the caller at the required 3714 // depth. NOTE that this code must perform the same logic as 3715 // vframeStream::security_get_caller_frame in that it must skip 3716 // Method.invoke() and auxiliary frames. Note also that depth is 3717 // 1-based (1 is the bottom of the inlining). 3718 int inlining_depth = _depth; 3719 JVMState* caller_jvms = NULL; 3720 3721 if (inlining_depth > 0) { 3722 caller_jvms = jvms(); 3723 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); 3724 do { 3725 // The following if-tests should be performed in this order 3726 if (is_method_invoke_or_aux_frame(caller_jvms)) { 3727 // Skip a Method.invoke() or auxiliary frame 3728 } else if (caller_depth > 0) { 3729 // Skip real frame 3730 --caller_depth; 3731 } else { 3732 // We're done: reached desired caller after skipping. 3733 break; 3734 } 3735 caller_jvms = caller_jvms->caller(); 3736 --inlining_depth; 3737 } while (inlining_depth > 0); 3738 } 3739 3740 if (inlining_depth == 0) { 3741 #ifndef PRODUCT 3742 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3743 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); 3744 tty->print_cr(" JVM state at this point:"); 3745 for (int i = _depth; i >= 1; i--) { 3746 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3747 } 3748 } 3749 #endif 3750 return false; // Reached end of inlining 3751 } 3752 3753 // Acquire method holder as java.lang.Class 3754 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 3755 ciInstance* caller_mirror = caller_klass->java_mirror(); 3756 // Push this as a constant 3757 push(makecon(TypeInstPtr::make(caller_mirror))); 3758 #ifndef PRODUCT 3759 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3760 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); 3761 tty->print_cr(" JVM state at this point:"); 3762 for (int i = _depth; i >= 1; i--) { 3763 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3764 } 3765 } 3766 #endif 3767 return true; 3768 } 3769 3770 // Helper routine for above 3771 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { 3772 ciMethod* method = jvms->method(); 3773 3774 // Is this the Method.invoke method itself? 3775 if (method->intrinsic_id() == vmIntrinsics::_invoke) 3776 return true; 3777 3778 // Is this a helper, defined somewhere underneath MethodAccessorImpl. 3779 ciKlass* k = method->holder(); 3780 if (k->is_instance_klass()) { 3781 ciInstanceKlass* ik = k->as_instance_klass(); 3782 for (; ik != NULL; ik = ik->super()) { 3783 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && 3784 ik == env()->find_system_klass(ik->name())) { 3785 return true; 3786 } 3787 } 3788 } 3789 else if (method->is_method_handle_adapter()) { 3790 // This is an internal adapter frame from the MethodHandleCompiler -- skip it 3791 return true; 3792 } 3793 3794 return false; 3795 } 3796 3797 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by 3798 // inline_native_AtomicLong_attemptUpdate() but it has no way of 3799 // computing it since there is no lookup field by name function in the 3800 // CI interface. This is computed and set by inline_native_AtomicLong_get(). 3801 // Using a static variable here is safe even if we have multiple compilation 3802 // threads because the offset is constant. At worst the same offset will be 3803 // computed and stored multiple 3804 3805 bool LibraryCallKit::inline_native_AtomicLong_get() { 3806 // Restore the stack and pop off the argument 3807 _sp+=1; 3808 Node *obj = pop(); 3809 3810 // get the offset of the "value" field. Since the CI interfaces 3811 // does not provide a way to look up a field by name, we scan the bytecodes 3812 // to get the field index. We expect the first 2 instructions of the method 3813 // to be: 3814 // 0 aload_0 3815 // 1 getfield "value" 3816 ciMethod* method = callee(); 3817 if (value_field_offset == -1) 3818 { 3819 ciField* value_field; 3820 ciBytecodeStream iter(method); 3821 Bytecodes::Code bc = iter.next(); 3822 3823 if ((bc != Bytecodes::_aload_0) && 3824 ((bc != Bytecodes::_aload) || (iter.get_index() != 0))) 3825 return false; 3826 bc = iter.next(); 3827 if (bc != Bytecodes::_getfield) 3828 return false; 3829 bool ignore; 3830 value_field = iter.get_field(ignore); 3831 value_field_offset = value_field->offset_in_bytes(); 3832 } 3833 3834 // Null check without removing any arguments. 3835 _sp++; 3836 obj = do_null_check(obj, T_OBJECT); 3837 _sp--; 3838 // Check for locking null object 3839 if (stopped()) return true; 3840 3841 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3842 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3843 int alias_idx = C->get_alias_index(adr_type); 3844 3845 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr)); 3846 3847 push_pair(result); 3848 3849 return true; 3850 } 3851 3852 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() { 3853 // Restore the stack and pop off the arguments 3854 _sp+=5; 3855 Node *newVal = pop_pair(); 3856 Node *oldVal = pop_pair(); 3857 Node *obj = pop(); 3858 3859 // we need the offset of the "value" field which was computed when 3860 // inlining the get() method. Give up if we don't have it. 3861 if (value_field_offset == -1) 3862 return false; 3863 3864 // Null check without removing any arguments. 3865 _sp+=5; 3866 obj = do_null_check(obj, T_OBJECT); 3867 _sp-=5; 3868 // Check for locking null object 3869 if (stopped()) return true; 3870 3871 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3872 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3873 int alias_idx = C->get_alias_index(adr_type); 3874 3875 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal)); 3876 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 3877 set_memory(store_proj, alias_idx); 3878 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) ); 3879 3880 Node *result; 3881 // CMove node is not used to be able fold a possible check code 3882 // after attemptUpdate() call. This code could be transformed 3883 // into CMove node by loop optimizations. 3884 { 3885 RegionNode *r = new (C, 3) RegionNode(3); 3886 result = new (C, 3) PhiNode(r, TypeInt::BOOL); 3887 3888 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 3889 Node *iftrue = opt_iff(r, iff); 3890 r->init_req(1, iftrue); 3891 result->init_req(1, intcon(1)); 3892 result->init_req(2, intcon(0)); 3893 3894 set_control(_gvn.transform(r)); 3895 record_for_igvn(r); 3896 3897 C->set_has_split_ifs(true); // Has chance for split-if optimization 3898 } 3899 3900 push(_gvn.transform(result)); 3901 return true; 3902 } 3903 3904 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 3905 // restore the arguments 3906 _sp += arg_size(); 3907 3908 switch (id) { 3909 case vmIntrinsics::_floatToRawIntBits: 3910 push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); 3911 break; 3912 3913 case vmIntrinsics::_intBitsToFloat: 3914 push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); 3915 break; 3916 3917 case vmIntrinsics::_doubleToRawLongBits: 3918 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); 3919 break; 3920 3921 case vmIntrinsics::_longBitsToDouble: 3922 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); 3923 break; 3924 3925 case vmIntrinsics::_doubleToLongBits: { 3926 Node* value = pop_pair(); 3927 3928 // two paths (plus control) merge in a wood 3929 RegionNode *r = new (C, 3) RegionNode(3); 3930 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); 3931 3932 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); 3933 // Build the boolean node 3934 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 3935 3936 // Branch either way. 3937 // NaN case is less traveled, which makes all the difference. 3938 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3939 Node *opt_isnan = _gvn.transform(ifisnan); 3940 assert( opt_isnan->is_If(), "Expect an IfNode"); 3941 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3942 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 3943 3944 set_control(iftrue); 3945 3946 static const jlong nan_bits = CONST64(0x7ff8000000000000); 3947 Node *slow_result = longcon(nan_bits); // return NaN 3948 phi->init_req(1, _gvn.transform( slow_result )); 3949 r->init_req(1, iftrue); 3950 3951 // Else fall through 3952 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 3953 set_control(iffalse); 3954 3955 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); 3956 r->init_req(2, iffalse); 3957 3958 // Post merge 3959 set_control(_gvn.transform(r)); 3960 record_for_igvn(r); 3961 3962 Node* result = _gvn.transform(phi); 3963 assert(result->bottom_type()->isa_long(), "must be"); 3964 push_pair(result); 3965 3966 C->set_has_split_ifs(true); // Has chance for split-if optimization 3967 3968 break; 3969 } 3970 3971 case vmIntrinsics::_floatToIntBits: { 3972 Node* value = pop(); 3973 3974 // two paths (plus control) merge in a wood 3975 RegionNode *r = new (C, 3) RegionNode(3); 3976 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); 3977 3978 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); 3979 // Build the boolean node 3980 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 3981 3982 // Branch either way. 3983 // NaN case is less traveled, which makes all the difference. 3984 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3985 Node *opt_isnan = _gvn.transform(ifisnan); 3986 assert( opt_isnan->is_If(), "Expect an IfNode"); 3987 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 3988 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 3989 3990 set_control(iftrue); 3991 3992 static const jint nan_bits = 0x7fc00000; 3993 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 3994 phi->init_req(1, _gvn.transform( slow_result )); 3995 r->init_req(1, iftrue); 3996 3997 // Else fall through 3998 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 3999 set_control(iffalse); 4000 4001 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); 4002 r->init_req(2, iffalse); 4003 4004 // Post merge 4005 set_control(_gvn.transform(r)); 4006 record_for_igvn(r); 4007 4008 Node* result = _gvn.transform(phi); 4009 assert(result->bottom_type()->isa_int(), "must be"); 4010 push(result); 4011 4012 C->set_has_split_ifs(true); // Has chance for split-if optimization 4013 4014 break; 4015 } 4016 4017 default: 4018 ShouldNotReachHere(); 4019 } 4020 4021 return true; 4022 } 4023 4024 #ifdef _LP64 4025 #define XTOP ,top() /*additional argument*/ 4026 #else //_LP64 4027 #define XTOP /*no additional argument*/ 4028 #endif //_LP64 4029 4030 //----------------------inline_unsafe_copyMemory------------------------- 4031 bool LibraryCallKit::inline_unsafe_copyMemory() { 4032 if (callee()->is_static()) return false; // caller must have the capability! 4033 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) 4034 assert(signature()->size() == nargs-1, "copy has 5 arguments"); 4035 null_check_receiver(callee()); // check then ignore argument(0) 4036 if (stopped()) return true; 4037 4038 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4039 4040 Node* src_ptr = argument(1); 4041 Node* src_off = ConvL2X(argument(2)); 4042 assert(argument(3)->is_top(), "2nd half of long"); 4043 Node* dst_ptr = argument(4); 4044 Node* dst_off = ConvL2X(argument(5)); 4045 assert(argument(6)->is_top(), "2nd half of long"); 4046 Node* size = ConvL2X(argument(7)); 4047 assert(argument(8)->is_top(), "2nd half of long"); 4048 4049 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4050 "fieldOffset must be byte-scaled"); 4051 4052 Node* src = make_unsafe_address(src_ptr, src_off); 4053 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4054 4055 // Conservatively insert a memory barrier on all memory slices. 4056 // Do not let writes of the copy source or destination float below the copy. 4057 insert_mem_bar(Op_MemBarCPUOrder); 4058 4059 // Call it. Note that the length argument is not scaled. 4060 make_runtime_call(RC_LEAF|RC_NO_FP, 4061 OptoRuntime::fast_arraycopy_Type(), 4062 StubRoutines::unsafe_arraycopy(), 4063 "unsafe_arraycopy", 4064 TypeRawPtr::BOTTOM, 4065 src, dst, size XTOP); 4066 4067 // Do not let reads of the copy destination float above the copy. 4068 insert_mem_bar(Op_MemBarCPUOrder); 4069 4070 return true; 4071 } 4072 4073 //------------------------clone_coping----------------------------------- 4074 // Helper function for inline_native_clone. 4075 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4076 assert(obj_size != NULL, ""); 4077 Node* raw_obj = alloc_obj->in(1); 4078 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4079 4080 if (ReduceBulkZeroing) { 4081 // We will be completely responsible for initializing this object - 4082 // mark Initialize node as complete. 4083 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4084 // The object was just allocated - there should be no any stores! 4085 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4086 } 4087 4088 // Copy the fastest available way. 4089 // TODO: generate fields copies for small objects instead. 4090 Node* src = obj; 4091 Node* dest = alloc_obj; 4092 Node* size = _gvn.transform(obj_size); 4093 4094 // Exclude the header but include array length to copy by 8 bytes words. 4095 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4096 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4097 instanceOopDesc::base_offset_in_bytes(); 4098 // base_off: 4099 // 8 - 32-bit VM 4100 // 12 - 64-bit VM, compressed oops 4101 // 16 - 64-bit VM, normal oops 4102 if (base_off % BytesPerLong != 0) { 4103 assert(UseCompressedOops, ""); 4104 if (is_array) { 4105 // Exclude length to copy by 8 bytes words. 4106 base_off += sizeof(int); 4107 } else { 4108 // Include klass to copy by 8 bytes words. 4109 base_off = instanceOopDesc::klass_offset_in_bytes(); 4110 } 4111 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4112 } 4113 src = basic_plus_adr(src, base_off); 4114 dest = basic_plus_adr(dest, base_off); 4115 4116 // Compute the length also, if needed: 4117 Node* countx = size; 4118 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); 4119 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4120 4121 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4122 bool disjoint_bases = true; 4123 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4124 src, NULL, dest, NULL, countx); 4125 4126 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4127 if (card_mark) { 4128 assert(!is_array, ""); 4129 // Put in store barrier for any and all oops we are sticking 4130 // into this object. (We could avoid this if we could prove 4131 // that the object type contains no oop fields at all.) 4132 Node* no_particular_value = NULL; 4133 Node* no_particular_field = NULL; 4134 int raw_adr_idx = Compile::AliasIdxRaw; 4135 post_barrier(control(), 4136 memory(raw_adr_type), 4137 alloc_obj, 4138 no_particular_field, 4139 raw_adr_idx, 4140 no_particular_value, 4141 T_OBJECT, 4142 false); 4143 } 4144 4145 // Do not let reads from the cloned object float above the arraycopy. 4146 insert_mem_bar(Op_MemBarCPUOrder); 4147 } 4148 4149 //------------------------inline_native_clone---------------------------- 4150 // Here are the simple edge cases: 4151 // null receiver => normal trap 4152 // virtual and clone was overridden => slow path to out-of-line clone 4153 // not cloneable or finalizer => slow path to out-of-line Object.clone 4154 // 4155 // The general case has two steps, allocation and copying. 4156 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4157 // 4158 // Copying also has two cases, oop arrays and everything else. 4159 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4160 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4161 // 4162 // These steps fold up nicely if and when the cloned object's klass 4163 // can be sharply typed as an object array, a type array, or an instance. 4164 // 4165 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4166 int nargs = 1; 4167 PhiNode* result_val; 4168 4169 //set the original stack and the reexecute bit for the interpreter to reexecute 4170 //the bytecode that invokes Object.clone if deoptimization happens 4171 { PreserveReexecuteState preexecs(this); 4172 jvms()->set_should_reexecute(true); 4173 4174 //null_check_receiver will adjust _sp (push and pop) 4175 Node* obj = null_check_receiver(callee()); 4176 if (stopped()) return true; 4177 4178 _sp += nargs; 4179 4180 Node* obj_klass = load_object_klass(obj); 4181 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4182 const TypeOopPtr* toop = ((tklass != NULL) 4183 ? tklass->as_instance_type() 4184 : TypeInstPtr::NOTNULL); 4185 4186 // Conservatively insert a memory barrier on all memory slices. 4187 // Do not let writes into the original float below the clone. 4188 insert_mem_bar(Op_MemBarCPUOrder); 4189 4190 // paths into result_reg: 4191 enum { 4192 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4193 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4194 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4195 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4196 PATH_LIMIT 4197 }; 4198 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4199 result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 4200 TypeInstPtr::NOTNULL); 4201 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 4202 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 4203 TypePtr::BOTTOM); 4204 record_for_igvn(result_reg); 4205 4206 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4207 int raw_adr_idx = Compile::AliasIdxRaw; 4208 const bool raw_mem_only = true; 4209 4210 4211 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4212 if (array_ctl != NULL) { 4213 // It's an array. 4214 PreserveJVMState pjvms(this); 4215 set_control(array_ctl); 4216 Node* obj_length = load_array_length(obj); 4217 Node* obj_size = NULL; 4218 Node* alloc_obj = new_array(obj_klass, obj_length, 0, 4219 raw_mem_only, &obj_size); 4220 4221 if (!use_ReduceInitialCardMarks()) { 4222 // If it is an oop array, it requires very special treatment, 4223 // because card marking is required on each card of the array. 4224 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4225 if (is_obja != NULL) { 4226 PreserveJVMState pjvms2(this); 4227 set_control(is_obja); 4228 // Generate a direct call to the right arraycopy function(s). 4229 bool disjoint_bases = true; 4230 bool length_never_negative = true; 4231 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4232 obj, intcon(0), alloc_obj, intcon(0), 4233 obj_length, 4234 disjoint_bases, length_never_negative); 4235 result_reg->init_req(_objArray_path, control()); 4236 result_val->init_req(_objArray_path, alloc_obj); 4237 result_i_o ->set_req(_objArray_path, i_o()); 4238 result_mem ->set_req(_objArray_path, reset_memory()); 4239 } 4240 } 4241 // Otherwise, there are no card marks to worry about. 4242 // (We can dispense with card marks if we know the allocation 4243 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4244 // causes the non-eden paths to take compensating steps to 4245 // simulate a fresh allocation, so that no further 4246 // card marks are required in compiled code to initialize 4247 // the object.) 4248 4249 if (!stopped()) { 4250 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4251 4252 // Present the results of the copy. 4253 result_reg->init_req(_array_path, control()); 4254 result_val->init_req(_array_path, alloc_obj); 4255 result_i_o ->set_req(_array_path, i_o()); 4256 result_mem ->set_req(_array_path, reset_memory()); 4257 } 4258 } 4259 4260 // We only go to the instance fast case code if we pass a number of guards. 4261 // The paths which do not pass are accumulated in the slow_region. 4262 RegionNode* slow_region = new (C, 1) RegionNode(1); 4263 record_for_igvn(slow_region); 4264 if (!stopped()) { 4265 // It's an instance (we did array above). Make the slow-path tests. 4266 // If this is a virtual call, we generate a funny guard. We grab 4267 // the vtable entry corresponding to clone() from the target object. 4268 // If the target method which we are calling happens to be the 4269 // Object clone() method, we pass the guard. We do not need this 4270 // guard for non-virtual calls; the caller is known to be the native 4271 // Object clone(). 4272 if (is_virtual) { 4273 generate_virtual_guard(obj_klass, slow_region); 4274 } 4275 4276 // The object must be cloneable and must not have a finalizer. 4277 // Both of these conditions may be checked in a single test. 4278 // We could optimize the cloneable test further, but we don't care. 4279 generate_access_flags_guard(obj_klass, 4280 // Test both conditions: 4281 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4282 // Must be cloneable but not finalizer: 4283 JVM_ACC_IS_CLONEABLE, 4284 slow_region); 4285 } 4286 4287 if (!stopped()) { 4288 // It's an instance, and it passed the slow-path tests. 4289 PreserveJVMState pjvms(this); 4290 Node* obj_size = NULL; 4291 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size); 4292 4293 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4294 4295 // Present the results of the slow call. 4296 result_reg->init_req(_instance_path, control()); 4297 result_val->init_req(_instance_path, alloc_obj); 4298 result_i_o ->set_req(_instance_path, i_o()); 4299 result_mem ->set_req(_instance_path, reset_memory()); 4300 } 4301 4302 // Generate code for the slow case. We make a call to clone(). 4303 set_control(_gvn.transform(slow_region)); 4304 if (!stopped()) { 4305 PreserveJVMState pjvms(this); 4306 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4307 Node* slow_result = set_results_for_java_call(slow_call); 4308 // this->control() comes from set_results_for_java_call 4309 result_reg->init_req(_slow_path, control()); 4310 result_val->init_req(_slow_path, slow_result); 4311 result_i_o ->set_req(_slow_path, i_o()); 4312 result_mem ->set_req(_slow_path, reset_memory()); 4313 } 4314 4315 // Return the combined state. 4316 set_control( _gvn.transform(result_reg) ); 4317 set_i_o( _gvn.transform(result_i_o) ); 4318 set_all_memory( _gvn.transform(result_mem) ); 4319 } //original reexecute and sp are set back here 4320 4321 push(_gvn.transform(result_val)); 4322 4323 return true; 4324 } 4325 4326 4327 // constants for computing the copy function 4328 enum { 4329 COPYFUNC_UNALIGNED = 0, 4330 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize 4331 COPYFUNC_CONJOINT = 0, 4332 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend 4333 }; 4334 4335 // Note: The condition "disjoint" applies also for overlapping copies 4336 // where an descending copy is permitted (i.e., dest_offset <= src_offset). 4337 static address 4338 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name) { 4339 int selector = 4340 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) + 4341 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT); 4342 4343 #define RETURN_STUB(xxx_arraycopy) { \ 4344 name = #xxx_arraycopy; \ 4345 return StubRoutines::xxx_arraycopy(); } 4346 4347 switch (t) { 4348 case T_BYTE: 4349 case T_BOOLEAN: 4350 switch (selector) { 4351 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy); 4352 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy); 4353 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy); 4354 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy); 4355 } 4356 case T_CHAR: 4357 case T_SHORT: 4358 switch (selector) { 4359 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy); 4360 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy); 4361 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy); 4362 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy); 4363 } 4364 case T_INT: 4365 case T_FLOAT: 4366 switch (selector) { 4367 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy); 4368 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy); 4369 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy); 4370 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy); 4371 } 4372 case T_DOUBLE: 4373 case T_LONG: 4374 switch (selector) { 4375 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy); 4376 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy); 4377 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy); 4378 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy); 4379 } 4380 case T_ARRAY: 4381 case T_OBJECT: 4382 switch (selector) { 4383 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_arraycopy); 4384 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_arraycopy); 4385 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(oop_disjoint_arraycopy); 4386 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_oop_disjoint_arraycopy); 4387 } 4388 default: 4389 ShouldNotReachHere(); 4390 return NULL; 4391 } 4392 4393 #undef RETURN_STUB 4394 } 4395 4396 //------------------------------basictype2arraycopy---------------------------- 4397 address LibraryCallKit::basictype2arraycopy(BasicType t, 4398 Node* src_offset, 4399 Node* dest_offset, 4400 bool disjoint_bases, 4401 const char* &name) { 4402 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4403 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4404 4405 bool aligned = false; 4406 bool disjoint = disjoint_bases; 4407 4408 // if the offsets are the same, we can treat the memory regions as 4409 // disjoint, because either the memory regions are in different arrays, 4410 // or they are identical (which we can treat as disjoint.) We can also 4411 // treat a copy with a destination index less that the source index 4412 // as disjoint since a low->high copy will work correctly in this case. 4413 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4414 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4415 // both indices are constants 4416 int s_offs = src_offset_inttype->get_con(); 4417 int d_offs = dest_offset_inttype->get_con(); 4418 int element_size = type2aelembytes(t); 4419 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4420 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4421 if (s_offs >= d_offs) disjoint = true; 4422 } else if (src_offset == dest_offset && src_offset != NULL) { 4423 // This can occur if the offsets are identical non-constants. 4424 disjoint = true; 4425 } 4426 4427 return select_arraycopy_function(t, aligned, disjoint, name); 4428 } 4429 4430 4431 //------------------------------inline_arraycopy----------------------- 4432 bool LibraryCallKit::inline_arraycopy() { 4433 // Restore the stack and pop off the arguments. 4434 int nargs = 5; // 2 oops, 3 ints, no size_t or long 4435 assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); 4436 4437 Node *src = argument(0); 4438 Node *src_offset = argument(1); 4439 Node *dest = argument(2); 4440 Node *dest_offset = argument(3); 4441 Node *length = argument(4); 4442 4443 // Compile time checks. If any of these checks cannot be verified at compile time, 4444 // we do not make a fast path for this call. Instead, we let the call remain as it 4445 // is. The checks we choose to mandate at compile time are: 4446 // 4447 // (1) src and dest are arrays. 4448 const Type* src_type = src->Value(&_gvn); 4449 const Type* dest_type = dest->Value(&_gvn); 4450 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4451 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4452 if (top_src == NULL || top_src->klass() == NULL || 4453 top_dest == NULL || top_dest->klass() == NULL) { 4454 // Conservatively insert a memory barrier on all memory slices. 4455 // Do not let writes into the source float below the arraycopy. 4456 insert_mem_bar(Op_MemBarCPUOrder); 4457 4458 // Call StubRoutines::generic_arraycopy stub. 4459 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4460 src, src_offset, dest, dest_offset, length); 4461 4462 // Do not let reads from the destination float above the arraycopy. 4463 // Since we cannot type the arrays, we don't know which slices 4464 // might be affected. We could restrict this barrier only to those 4465 // memory slices which pertain to array elements--but don't bother. 4466 if (!InsertMemBarAfterArraycopy) 4467 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4468 insert_mem_bar(Op_MemBarCPUOrder); 4469 return true; 4470 } 4471 4472 // (2) src and dest arrays must have elements of the same BasicType 4473 // Figure out the size and type of the elements we will be copying. 4474 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4475 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4476 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4477 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4478 4479 if (src_elem != dest_elem || dest_elem == T_VOID) { 4480 // The component types are not the same or are not recognized. Punt. 4481 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4482 generate_slow_arraycopy(TypePtr::BOTTOM, 4483 src, src_offset, dest, dest_offset, length); 4484 return true; 4485 } 4486 4487 //--------------------------------------------------------------------------- 4488 // We will make a fast path for this call to arraycopy. 4489 4490 // We have the following tests left to perform: 4491 // 4492 // (3) src and dest must not be null. 4493 // (4) src_offset must not be negative. 4494 // (5) dest_offset must not be negative. 4495 // (6) length must not be negative. 4496 // (7) src_offset + length must not exceed length of src. 4497 // (8) dest_offset + length must not exceed length of dest. 4498 // (9) each element of an oop array must be assignable 4499 4500 RegionNode* slow_region = new (C, 1) RegionNode(1); 4501 record_for_igvn(slow_region); 4502 4503 // (3) operands must not be null 4504 // We currently perform our null checks with the do_null_check routine. 4505 // This means that the null exceptions will be reported in the caller 4506 // rather than (correctly) reported inside of the native arraycopy call. 4507 // This should be corrected, given time. We do our null check with the 4508 // stack pointer restored. 4509 _sp += nargs; 4510 src = do_null_check(src, T_ARRAY); 4511 dest = do_null_check(dest, T_ARRAY); 4512 _sp -= nargs; 4513 4514 // (4) src_offset must not be negative. 4515 generate_negative_guard(src_offset, slow_region); 4516 4517 // (5) dest_offset must not be negative. 4518 generate_negative_guard(dest_offset, slow_region); 4519 4520 // (6) length must not be negative (moved to generate_arraycopy()). 4521 // generate_negative_guard(length, slow_region); 4522 4523 // (7) src_offset + length must not exceed length of src. 4524 generate_limit_guard(src_offset, length, 4525 load_array_length(src), 4526 slow_region); 4527 4528 // (8) dest_offset + length must not exceed length of dest. 4529 generate_limit_guard(dest_offset, length, 4530 load_array_length(dest), 4531 slow_region); 4532 4533 // (9) each element of an oop array must be assignable 4534 // The generate_arraycopy subroutine checks this. 4535 4536 // This is where the memory effects are placed: 4537 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4538 generate_arraycopy(adr_type, dest_elem, 4539 src, src_offset, dest, dest_offset, length, 4540 false, false, slow_region); 4541 4542 return true; 4543 } 4544 4545 //-----------------------------generate_arraycopy---------------------- 4546 // Generate an optimized call to arraycopy. 4547 // Caller must guard against non-arrays. 4548 // Caller must determine a common array basic-type for both arrays. 4549 // Caller must validate offsets against array bounds. 4550 // The slow_region has already collected guard failure paths 4551 // (such as out of bounds length or non-conformable array types). 4552 // The generated code has this shape, in general: 4553 // 4554 // if (length == 0) return // via zero_path 4555 // slowval = -1 4556 // if (types unknown) { 4557 // slowval = call generic copy loop 4558 // if (slowval == 0) return // via checked_path 4559 // } else if (indexes in bounds) { 4560 // if ((is object array) && !(array type check)) { 4561 // slowval = call checked copy loop 4562 // if (slowval == 0) return // via checked_path 4563 // } else { 4564 // call bulk copy loop 4565 // return // via fast_path 4566 // } 4567 // } 4568 // // adjust params for remaining work: 4569 // if (slowval != -1) { 4570 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4571 // } 4572 // slow_region: 4573 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 4574 // return // via slow_call_path 4575 // 4576 // This routine is used from several intrinsics: System.arraycopy, 4577 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 4578 // 4579 void 4580 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4581 BasicType basic_elem_type, 4582 Node* src, Node* src_offset, 4583 Node* dest, Node* dest_offset, 4584 Node* copy_length, 4585 bool disjoint_bases, 4586 bool length_never_negative, 4587 RegionNode* slow_region) { 4588 4589 if (slow_region == NULL) { 4590 slow_region = new(C,1) RegionNode(1); 4591 record_for_igvn(slow_region); 4592 } 4593 4594 Node* original_dest = dest; 4595 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4596 bool must_clear_dest = false; 4597 4598 // See if this is the initialization of a newly-allocated array. 4599 // If so, we will take responsibility here for initializing it to zero. 4600 // (Note: Because tightly_coupled_allocation performs checks on the 4601 // out-edges of the dest, we need to avoid making derived pointers 4602 // from it until we have checked its uses.) 4603 if (ReduceBulkZeroing 4604 && !ZeroTLAB // pointless if already zeroed 4605 && basic_elem_type != T_CONFLICT // avoid corner case 4606 && !_gvn.eqv_uncast(src, dest) 4607 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4608 != NULL) 4609 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4610 && alloc->maybe_set_complete(&_gvn)) { 4611 // "You break it, you buy it." 4612 InitializeNode* init = alloc->initialization(); 4613 assert(init->is_complete(), "we just did this"); 4614 assert(dest->is_CheckCastPP(), "sanity"); 4615 assert(dest->in(0)->in(0) == init, "dest pinned"); 4616 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4617 // From this point on, every exit path is responsible for 4618 // initializing any non-copied parts of the object to zero. 4619 must_clear_dest = true; 4620 } else { 4621 // No zeroing elimination here. 4622 alloc = NULL; 4623 //original_dest = dest; 4624 //must_clear_dest = false; 4625 } 4626 4627 // Results are placed here: 4628 enum { fast_path = 1, // normal void-returning assembly stub 4629 checked_path = 2, // special assembly stub with cleanup 4630 slow_call_path = 3, // something went wrong; call the VM 4631 zero_path = 4, // bypass when length of copy is zero 4632 bcopy_path = 5, // copy primitive array by 64-bit blocks 4633 PATH_LIMIT = 6 4634 }; 4635 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4636 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO); 4637 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type); 4638 record_for_igvn(result_region); 4639 _gvn.set_type_bottom(result_i_o); 4640 _gvn.set_type_bottom(result_memory); 4641 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4642 4643 // The slow_control path: 4644 Node* slow_control; 4645 Node* slow_i_o = i_o(); 4646 Node* slow_mem = memory(adr_type); 4647 debug_only(slow_control = (Node*) badAddress); 4648 4649 // Checked control path: 4650 Node* checked_control = top(); 4651 Node* checked_mem = NULL; 4652 Node* checked_i_o = NULL; 4653 Node* checked_value = NULL; 4654 4655 if (basic_elem_type == T_CONFLICT) { 4656 assert(!must_clear_dest, ""); 4657 Node* cv = generate_generic_arraycopy(adr_type, 4658 src, src_offset, dest, dest_offset, 4659 copy_length); 4660 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4661 checked_control = control(); 4662 checked_i_o = i_o(); 4663 checked_mem = memory(adr_type); 4664 checked_value = cv; 4665 set_control(top()); // no fast path 4666 } 4667 4668 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4669 if (not_pos != NULL) { 4670 PreserveJVMState pjvms(this); 4671 set_control(not_pos); 4672 4673 // (6) length must not be negative. 4674 if (!length_never_negative) { 4675 generate_negative_guard(copy_length, slow_region); 4676 } 4677 4678 // copy_length is 0. 4679 if (!stopped() && must_clear_dest) { 4680 Node* dest_length = alloc->in(AllocateNode::ALength); 4681 if (_gvn.eqv_uncast(copy_length, dest_length) 4682 || _gvn.find_int_con(dest_length, 1) <= 0) { 4683 // There is no zeroing to do. No need for a secondary raw memory barrier. 4684 } else { 4685 // Clear the whole thing since there are no source elements to copy. 4686 generate_clear_array(adr_type, dest, basic_elem_type, 4687 intcon(0), NULL, 4688 alloc->in(AllocateNode::AllocSize)); 4689 // Use a secondary InitializeNode as raw memory barrier. 4690 // Currently it is needed only on this path since other 4691 // paths have stub or runtime calls as raw memory barriers. 4692 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4693 Compile::AliasIdxRaw, 4694 top())->as_Initialize(); 4695 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4696 } 4697 } 4698 4699 // Present the results of the fast call. 4700 result_region->init_req(zero_path, control()); 4701 result_i_o ->init_req(zero_path, i_o()); 4702 result_memory->init_req(zero_path, memory(adr_type)); 4703 } 4704 4705 if (!stopped() && must_clear_dest) { 4706 // We have to initialize the *uncopied* part of the array to zero. 4707 // The copy destination is the slice dest[off..off+len]. The other slices 4708 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4709 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4710 Node* dest_length = alloc->in(AllocateNode::ALength); 4711 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset, 4712 copy_length) ); 4713 4714 // If there is a head section that needs zeroing, do it now. 4715 if (find_int_con(dest_offset, -1) != 0) { 4716 generate_clear_array(adr_type, dest, basic_elem_type, 4717 intcon(0), dest_offset, 4718 NULL); 4719 } 4720 4721 // Next, perform a dynamic check on the tail length. 4722 // It is often zero, and we can win big if we prove this. 4723 // There are two wins: Avoid generating the ClearArray 4724 // with its attendant messy index arithmetic, and upgrade 4725 // the copy to a more hardware-friendly word size of 64 bits. 4726 Node* tail_ctl = NULL; 4727 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) { 4728 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) ); 4729 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) ); 4730 tail_ctl = generate_slow_guard(bol_lt, NULL); 4731 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4732 } 4733 4734 // At this point, let's assume there is no tail. 4735 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4736 // There is no tail. Try an upgrade to a 64-bit copy. 4737 bool didit = false; 4738 { PreserveJVMState pjvms(this); 4739 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4740 src, src_offset, dest, dest_offset, 4741 dest_size); 4742 if (didit) { 4743 // Present the results of the block-copying fast call. 4744 result_region->init_req(bcopy_path, control()); 4745 result_i_o ->init_req(bcopy_path, i_o()); 4746 result_memory->init_req(bcopy_path, memory(adr_type)); 4747 } 4748 } 4749 if (didit) 4750 set_control(top()); // no regular fast path 4751 } 4752 4753 // Clear the tail, if any. 4754 if (tail_ctl != NULL) { 4755 Node* notail_ctl = stopped() ? NULL : control(); 4756 set_control(tail_ctl); 4757 if (notail_ctl == NULL) { 4758 generate_clear_array(adr_type, dest, basic_elem_type, 4759 dest_tail, NULL, 4760 dest_size); 4761 } else { 4762 // Make a local merge. 4763 Node* done_ctl = new(C,3) RegionNode(3); 4764 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type); 4765 done_ctl->init_req(1, notail_ctl); 4766 done_mem->init_req(1, memory(adr_type)); 4767 generate_clear_array(adr_type, dest, basic_elem_type, 4768 dest_tail, NULL, 4769 dest_size); 4770 done_ctl->init_req(2, control()); 4771 done_mem->init_req(2, memory(adr_type)); 4772 set_control( _gvn.transform(done_ctl) ); 4773 set_memory( _gvn.transform(done_mem), adr_type ); 4774 } 4775 } 4776 } 4777 4778 BasicType copy_type = basic_elem_type; 4779 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 4780 if (!stopped() && copy_type == T_OBJECT) { 4781 // If src and dest have compatible element types, we can copy bits. 4782 // Types S[] and D[] are compatible if D is a supertype of S. 4783 // 4784 // If they are not, we will use checked_oop_disjoint_arraycopy, 4785 // which performs a fast optimistic per-oop check, and backs off 4786 // further to JVM_ArrayCopy on the first per-oop check that fails. 4787 // (Actually, we don't move raw bits only; the GC requires card marks.) 4788 4789 // Get the klassOop for both src and dest 4790 Node* src_klass = load_object_klass(src); 4791 Node* dest_klass = load_object_klass(dest); 4792 4793 // Generate the subtype check. 4794 // This might fold up statically, or then again it might not. 4795 // 4796 // Non-static example: Copying List<String>.elements to a new String[]. 4797 // The backing store for a List<String> is always an Object[], 4798 // but its elements are always type String, if the generic types 4799 // are correct at the source level. 4800 // 4801 // Test S[] against D[], not S against D, because (probably) 4802 // the secondary supertype cache is less busy for S[] than S. 4803 // This usually only matters when D is an interface. 4804 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4805 // Plug failing path into checked_oop_disjoint_arraycopy 4806 if (not_subtype_ctrl != top()) { 4807 PreserveJVMState pjvms(this); 4808 set_control(not_subtype_ctrl); 4809 // (At this point we can assume disjoint_bases, since types differ.) 4810 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc); 4811 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 4812 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 4813 Node* dest_elem_klass = _gvn.transform(n1); 4814 Node* cv = generate_checkcast_arraycopy(adr_type, 4815 dest_elem_klass, 4816 src, src_offset, dest, dest_offset, 4817 ConvI2X(copy_length)); 4818 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4819 checked_control = control(); 4820 checked_i_o = i_o(); 4821 checked_mem = memory(adr_type); 4822 checked_value = cv; 4823 } 4824 // At this point we know we do not need type checks on oop stores. 4825 4826 // Let's see if we need card marks: 4827 if (alloc != NULL && use_ReduceInitialCardMarks()) { 4828 // If we do not need card marks, copy using the jint or jlong stub. 4829 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 4830 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 4831 "sizes agree"); 4832 } 4833 } 4834 4835 if (!stopped()) { 4836 // Generate the fast path, if possible. 4837 PreserveJVMState pjvms(this); 4838 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 4839 src, src_offset, dest, dest_offset, 4840 ConvI2X(copy_length)); 4841 4842 // Present the results of the fast call. 4843 result_region->init_req(fast_path, control()); 4844 result_i_o ->init_req(fast_path, i_o()); 4845 result_memory->init_req(fast_path, memory(adr_type)); 4846 } 4847 4848 // Here are all the slow paths up to this point, in one bundle: 4849 slow_control = top(); 4850 if (slow_region != NULL) 4851 slow_control = _gvn.transform(slow_region); 4852 debug_only(slow_region = (RegionNode*)badAddress); 4853 4854 set_control(checked_control); 4855 if (!stopped()) { 4856 // Clean up after the checked call. 4857 // The returned value is either 0 or -1^K, 4858 // where K = number of partially transferred array elements. 4859 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) ); 4860 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 4861 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4862 4863 // If it is 0, we are done, so transfer to the end. 4864 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) ); 4865 result_region->init_req(checked_path, checks_done); 4866 result_i_o ->init_req(checked_path, checked_i_o); 4867 result_memory->init_req(checked_path, checked_mem); 4868 4869 // If it is not zero, merge into the slow call. 4870 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) )); 4871 RegionNode* slow_reg2 = new(C, 3) RegionNode(3); 4872 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO); 4873 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type); 4874 record_for_igvn(slow_reg2); 4875 slow_reg2 ->init_req(1, slow_control); 4876 slow_i_o2 ->init_req(1, slow_i_o); 4877 slow_mem2 ->init_req(1, slow_mem); 4878 slow_reg2 ->init_req(2, control()); 4879 slow_i_o2 ->init_req(2, checked_i_o); 4880 slow_mem2 ->init_req(2, checked_mem); 4881 4882 slow_control = _gvn.transform(slow_reg2); 4883 slow_i_o = _gvn.transform(slow_i_o2); 4884 slow_mem = _gvn.transform(slow_mem2); 4885 4886 if (alloc != NULL) { 4887 // We'll restart from the very beginning, after zeroing the whole thing. 4888 // This can cause double writes, but that's OK since dest is brand new. 4889 // So we ignore the low 31 bits of the value returned from the stub. 4890 } else { 4891 // We must continue the copy exactly where it failed, or else 4892 // another thread might see the wrong number of writes to dest. 4893 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) ); 4894 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT); 4895 slow_offset->init_req(1, intcon(0)); 4896 slow_offset->init_req(2, checked_offset); 4897 slow_offset = _gvn.transform(slow_offset); 4898 4899 // Adjust the arguments by the conditionally incoming offset. 4900 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) ); 4901 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) ); 4902 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) ); 4903 4904 // Tweak the node variables to adjust the code produced below: 4905 src_offset = src_off_plus; 4906 dest_offset = dest_off_plus; 4907 copy_length = length_minus; 4908 } 4909 } 4910 4911 set_control(slow_control); 4912 if (!stopped()) { 4913 // Generate the slow path, if needed. 4914 PreserveJVMState pjvms(this); // replace_in_map may trash the map 4915 4916 set_memory(slow_mem, adr_type); 4917 set_i_o(slow_i_o); 4918 4919 if (must_clear_dest) { 4920 generate_clear_array(adr_type, dest, basic_elem_type, 4921 intcon(0), NULL, 4922 alloc->in(AllocateNode::AllocSize)); 4923 } 4924 4925 generate_slow_arraycopy(adr_type, 4926 src, src_offset, dest, dest_offset, 4927 copy_length); 4928 4929 result_region->init_req(slow_call_path, control()); 4930 result_i_o ->init_req(slow_call_path, i_o()); 4931 result_memory->init_req(slow_call_path, memory(adr_type)); 4932 } 4933 4934 // Remove unused edges. 4935 for (uint i = 1; i < result_region->req(); i++) { 4936 if (result_region->in(i) == NULL) 4937 result_region->init_req(i, top()); 4938 } 4939 4940 // Finished; return the combined state. 4941 set_control( _gvn.transform(result_region) ); 4942 set_i_o( _gvn.transform(result_i_o) ); 4943 set_memory( _gvn.transform(result_memory), adr_type ); 4944 4945 // The memory edges above are precise in order to model effects around 4946 // array copies accurately to allow value numbering of field loads around 4947 // arraycopy. Such field loads, both before and after, are common in Java 4948 // collections and similar classes involving header/array data structures. 4949 // 4950 // But with low number of register or when some registers are used or killed 4951 // by arraycopy calls it causes registers spilling on stack. See 6544710. 4952 // The next memory barrier is added to avoid it. If the arraycopy can be 4953 // optimized away (which it can, sometimes) then we can manually remove 4954 // the membar also. 4955 // 4956 // Do not let reads from the cloned object float above the arraycopy. 4957 if (InsertMemBarAfterArraycopy || alloc != NULL) 4958 insert_mem_bar(Op_MemBarCPUOrder); 4959 } 4960 4961 4962 // Helper function which determines if an arraycopy immediately follows 4963 // an allocation, with no intervening tests or other escapes for the object. 4964 AllocateArrayNode* 4965 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4966 RegionNode* slow_region) { 4967 if (stopped()) return NULL; // no fast path 4968 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4969 4970 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4971 if (alloc == NULL) return NULL; 4972 4973 Node* rawmem = memory(Compile::AliasIdxRaw); 4974 // Is the allocation's memory state untouched? 4975 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4976 // Bail out if there have been raw-memory effects since the allocation. 4977 // (Example: There might have been a call or safepoint.) 4978 return NULL; 4979 } 4980 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4981 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4982 return NULL; 4983 } 4984 4985 // There must be no unexpected observers of this allocation. 4986 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4987 Node* obs = ptr->fast_out(i); 4988 if (obs != this->map()) { 4989 return NULL; 4990 } 4991 } 4992 4993 // This arraycopy must unconditionally follow the allocation of the ptr. 4994 Node* alloc_ctl = ptr->in(0); 4995 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4996 4997 Node* ctl = control(); 4998 while (ctl != alloc_ctl) { 4999 // There may be guards which feed into the slow_region. 5000 // Any other control flow means that we might not get a chance 5001 // to finish initializing the allocated object. 5002 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5003 IfNode* iff = ctl->in(0)->as_If(); 5004 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5005 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5006 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5007 ctl = iff->in(0); // This test feeds the known slow_region. 5008 continue; 5009 } 5010 // One more try: Various low-level checks bottom out in 5011 // uncommon traps. If the debug-info of the trap omits 5012 // any reference to the allocation, as we've already 5013 // observed, then there can be no objection to the trap. 5014 bool found_trap = false; 5015 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5016 Node* obs = not_ctl->fast_out(j); 5017 if (obs->in(0) == not_ctl && obs->is_Call() && 5018 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5019 found_trap = true; break; 5020 } 5021 } 5022 if (found_trap) { 5023 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5024 continue; 5025 } 5026 } 5027 return NULL; 5028 } 5029 5030 // If we get this far, we have an allocation which immediately 5031 // precedes the arraycopy, and we can take over zeroing the new object. 5032 // The arraycopy will finish the initialization, and provide 5033 // a new control state to which we will anchor the destination pointer. 5034 5035 return alloc; 5036 } 5037 5038 // Helper for initialization of arrays, creating a ClearArray. 5039 // It writes zero bits in [start..end), within the body of an array object. 5040 // The memory effects are all chained onto the 'adr_type' alias category. 5041 // 5042 // Since the object is otherwise uninitialized, we are free 5043 // to put a little "slop" around the edges of the cleared area, 5044 // as long as it does not go back into the array's header, 5045 // or beyond the array end within the heap. 5046 // 5047 // The lower edge can be rounded down to the nearest jint and the 5048 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5049 // 5050 // Arguments: 5051 // adr_type memory slice where writes are generated 5052 // dest oop of the destination array 5053 // basic_elem_type element type of the destination 5054 // slice_idx array index of first element to store 5055 // slice_len number of elements to store (or NULL) 5056 // dest_size total size in bytes of the array object 5057 // 5058 // Exactly one of slice_len or dest_size must be non-NULL. 5059 // If dest_size is non-NULL, zeroing extends to the end of the object. 5060 // If slice_len is non-NULL, the slice_idx value must be a constant. 5061 void 5062 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5063 Node* dest, 5064 BasicType basic_elem_type, 5065 Node* slice_idx, 5066 Node* slice_len, 5067 Node* dest_size) { 5068 // one or the other but not both of slice_len and dest_size: 5069 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5070 if (slice_len == NULL) slice_len = top(); 5071 if (dest_size == NULL) dest_size = top(); 5072 5073 // operate on this memory slice: 5074 Node* mem = memory(adr_type); // memory slice to operate on 5075 5076 // scaling and rounding of indexes: 5077 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5078 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5079 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5080 int bump_bit = (-1 << scale) & BytesPerInt; 5081 5082 // determine constant starts and ends 5083 const intptr_t BIG_NEG = -128; 5084 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5085 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5086 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5087 if (slice_len_con == 0) { 5088 return; // nothing to do here 5089 } 5090 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5091 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5092 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5093 assert(end_con < 0, "not two cons"); 5094 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5095 BytesPerLong); 5096 } 5097 5098 if (start_con >= 0 && end_con >= 0) { 5099 // Constant start and end. Simple. 5100 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5101 start_con, end_con, &_gvn); 5102 } else if (start_con >= 0 && dest_size != top()) { 5103 // Constant start, pre-rounded end after the tail of the array. 5104 Node* end = dest_size; 5105 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5106 start_con, end, &_gvn); 5107 } else if (start_con >= 0 && slice_len != top()) { 5108 // Constant start, non-constant end. End needs rounding up. 5109 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5110 intptr_t end_base = abase + (slice_idx_con << scale); 5111 int end_round = (-1 << scale) & (BytesPerLong - 1); 5112 Node* end = ConvI2X(slice_len); 5113 if (scale != 0) 5114 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) )); 5115 end_base += end_round; 5116 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) ); 5117 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) ); 5118 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5119 start_con, end, &_gvn); 5120 } else if (start_con < 0 && dest_size != top()) { 5121 // Non-constant start, pre-rounded end after the tail of the array. 5122 // This is almost certainly a "round-to-end" operation. 5123 Node* start = slice_idx; 5124 start = ConvI2X(start); 5125 if (scale != 0) 5126 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) )); 5127 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) ); 5128 if ((bump_bit | clear_low) != 0) { 5129 int to_clear = (bump_bit | clear_low); 5130 // Align up mod 8, then store a jint zero unconditionally 5131 // just before the mod-8 boundary. 5132 if (((abase + bump_bit) & ~to_clear) - bump_bit 5133 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5134 bump_bit = 0; 5135 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5136 } else { 5137 // Bump 'start' up to (or past) the next jint boundary: 5138 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) ); 5139 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5140 } 5141 // Round bumped 'start' down to jlong boundary in body of array. 5142 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) ); 5143 if (bump_bit != 0) { 5144 // Store a zero to the immediately preceding jint: 5145 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) ); 5146 Node* p1 = basic_plus_adr(dest, x1); 5147 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 5148 mem = _gvn.transform(mem); 5149 } 5150 } 5151 Node* end = dest_size; // pre-rounded 5152 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5153 start, end, &_gvn); 5154 } else { 5155 // Non-constant start, unrounded non-constant end. 5156 // (Nobody zeroes a random midsection of an array using this routine.) 5157 ShouldNotReachHere(); // fix caller 5158 } 5159 5160 // Done. 5161 set_memory(mem, adr_type); 5162 } 5163 5164 5165 bool 5166 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5167 BasicType basic_elem_type, 5168 AllocateNode* alloc, 5169 Node* src, Node* src_offset, 5170 Node* dest, Node* dest_offset, 5171 Node* dest_size) { 5172 // See if there is an advantage from block transfer. 5173 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5174 if (scale >= LogBytesPerLong) 5175 return false; // it is already a block transfer 5176 5177 // Look at the alignment of the starting offsets. 5178 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5179 const intptr_t BIG_NEG = -128; 5180 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5181 5182 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale); 5183 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale); 5184 if (src_off < 0 || dest_off < 0) 5185 // At present, we can only understand constants. 5186 return false; 5187 5188 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5189 // Non-aligned; too bad. 5190 // One more chance: Pick off an initial 32-bit word. 5191 // This is a common case, since abase can be odd mod 8. 5192 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5193 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5194 Node* sptr = basic_plus_adr(src, src_off); 5195 Node* dptr = basic_plus_adr(dest, dest_off); 5196 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 5197 store_to_memory(control(), dptr, sval, T_INT, adr_type); 5198 src_off += BytesPerInt; 5199 dest_off += BytesPerInt; 5200 } else { 5201 return false; 5202 } 5203 } 5204 assert(src_off % BytesPerLong == 0, ""); 5205 assert(dest_off % BytesPerLong == 0, ""); 5206 5207 // Do this copy by giant steps. 5208 Node* sptr = basic_plus_adr(src, src_off); 5209 Node* dptr = basic_plus_adr(dest, dest_off); 5210 Node* countx = dest_size; 5211 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) ); 5212 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) ); 5213 5214 bool disjoint_bases = true; // since alloc != NULL 5215 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5216 sptr, NULL, dptr, NULL, countx); 5217 5218 return true; 5219 } 5220 5221 5222 // Helper function; generates code for the slow case. 5223 // We make a call to a runtime method which emulates the native method, 5224 // but without the native wrapper overhead. 5225 void 5226 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5227 Node* src, Node* src_offset, 5228 Node* dest, Node* dest_offset, 5229 Node* copy_length) { 5230 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5231 OptoRuntime::slow_arraycopy_Type(), 5232 OptoRuntime::slow_arraycopy_Java(), 5233 "slow_arraycopy", adr_type, 5234 src, src_offset, dest, dest_offset, 5235 copy_length); 5236 5237 // Handle exceptions thrown by this fellow: 5238 make_slow_call_ex(call, env()->Throwable_klass(), false); 5239 } 5240 5241 // Helper function; generates code for cases requiring runtime checks. 5242 Node* 5243 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5244 Node* dest_elem_klass, 5245 Node* src, Node* src_offset, 5246 Node* dest, Node* dest_offset, 5247 Node* copy_length) { 5248 if (stopped()) return NULL; 5249 5250 address copyfunc_addr = StubRoutines::checkcast_arraycopy(); 5251 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5252 return NULL; 5253 } 5254 5255 // Pick out the parameters required to perform a store-check 5256 // for the target array. This is an optimistic check. It will 5257 // look in each non-null element's class, at the desired klass's 5258 // super_check_offset, for the desired klass. 5259 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc); 5260 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5261 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr()); 5262 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5263 Node* check_value = dest_elem_klass; 5264 5265 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5266 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5267 5268 // (We know the arrays are never conjoint, because their types differ.) 5269 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5270 OptoRuntime::checkcast_arraycopy_Type(), 5271 copyfunc_addr, "checkcast_arraycopy", adr_type, 5272 // five arguments, of which two are 5273 // intptr_t (jlong in LP64) 5274 src_start, dest_start, 5275 copy_length XTOP, 5276 check_offset XTOP, 5277 check_value); 5278 5279 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5280 } 5281 5282 5283 // Helper function; generates code for cases requiring runtime checks. 5284 Node* 5285 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5286 Node* src, Node* src_offset, 5287 Node* dest, Node* dest_offset, 5288 Node* copy_length) { 5289 if (stopped()) return NULL; 5290 5291 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5292 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5293 return NULL; 5294 } 5295 5296 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5297 OptoRuntime::generic_arraycopy_Type(), 5298 copyfunc_addr, "generic_arraycopy", adr_type, 5299 src, src_offset, dest, dest_offset, copy_length); 5300 5301 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5302 } 5303 5304 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5305 void 5306 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5307 BasicType basic_elem_type, 5308 bool disjoint_bases, 5309 Node* src, Node* src_offset, 5310 Node* dest, Node* dest_offset, 5311 Node* copy_length) { 5312 if (stopped()) return; // nothing to do 5313 5314 Node* src_start = src; 5315 Node* dest_start = dest; 5316 if (src_offset != NULL || dest_offset != NULL) { 5317 assert(src_offset != NULL && dest_offset != NULL, ""); 5318 src_start = array_element_address(src, src_offset, basic_elem_type); 5319 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5320 } 5321 5322 // Figure out which arraycopy runtime method to call. 5323 const char* copyfunc_name = "arraycopy"; 5324 address copyfunc_addr = 5325 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5326 disjoint_bases, copyfunc_name); 5327 5328 // Call it. Note that the count_ix value is not scaled to a byte-size. 5329 make_runtime_call(RC_LEAF|RC_NO_FP, 5330 OptoRuntime::fast_arraycopy_Type(), 5331 copyfunc_addr, copyfunc_name, adr_type, 5332 src_start, dest_start, copy_length XTOP); 5333 } 5334 5335 //----------------------------inline_reference_get---------------------------- 5336 5337 bool LibraryCallKit::inline_reference_get() { 5338 const int nargs = 1; // self 5339 5340 guarantee(java_lang_ref_Reference::referent_offset > 0, 5341 "should have already been set"); 5342 5343 int referent_offset = java_lang_ref_Reference::referent_offset; 5344 5345 // Validate the offset of the "referent" field. 5346 // Since the CI interface does not provide a way to look up a field by name, 5347 // we scan the bytecodes to get the field index. We expect the first 2 5348 // instructions of the method to be: 5349 // 0 aload_0 5350 // 1 getfield "referent" 5351 5352 ciMethod* method = callee(); 5353 5354 ciBytecodeStream iter(method); 5355 Bytecodes::Code bc = iter.next(); 5356 5357 if (bc != Bytecodes::_aload_0 && 5358 (bc != Bytecodes::_aload || iter.get_index() != 0)) { 5359 // We have an unexpected bytecode at bci 0. We're expecting aload_0 or aload 0 5360 return false; 5361 } 5362 5363 bc = iter.next(); 5364 if (bc != Bytecodes::_getfield) { 5365 // We have an unexpected bytecode at bci 1. We're expecting getfield. 5366 return false; 5367 } 5368 5369 // Restore the stack and pop off the argument 5370 _sp += nargs; 5371 Node *reference_obj = pop(); 5372 5373 // Null check on self without removing any arguments. 5374 _sp += nargs; 5375 reference_obj = do_null_check(reference_obj, T_OBJECT); 5376 _sp -= nargs;; 5377 5378 if (stopped()) return true; 5379 5380 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5381 5382 ciInstanceKlass* klass = env()->Object_klass(); 5383 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5384 5385 Node* no_ctrl = NULL; 5386 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT); 5387 5388 // Use the pre-barrier to record the value in the referent field 5389 pre_barrier(false /* do_load */, 5390 control(), 5391 NULL /* obj */, NULL /* adr */, -1 /* alias_idx */, NULL /* val */, NULL /* val_type */, 5392 result /* pre_val */, 5393 T_OBJECT); 5394 5395 push(result); 5396 return true; 5397 } 5398