1 /* 2 * Copyright (c) 1999, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "classfile/vmSymbols.hpp" 28 #include "compiler/compileBroker.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "oops/objArrayKlass.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/callGenerator.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/idealKit.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/parse.hpp" 37 #include "opto/runtime.hpp" 38 #include "opto/subnode.hpp" 39 #include "prims/nativeLookup.hpp" 40 #include "runtime/sharedRuntime.hpp" 41 42 class LibraryIntrinsic : public InlineCallGenerator { 43 // Extend the set of intrinsics known to the runtime: 44 public: 45 private: 46 bool _is_virtual; 47 vmIntrinsics::ID _intrinsic_id; 48 49 public: 50 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id) 51 : InlineCallGenerator(m), 52 _is_virtual(is_virtual), 53 _intrinsic_id(id) 54 { 55 } 56 virtual bool is_intrinsic() const { return true; } 57 virtual bool is_virtual() const { return _is_virtual; } 58 virtual JVMState* generate(JVMState* jvms); 59 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 60 }; 61 62 63 // Local helper class for LibraryIntrinsic: 64 class LibraryCallKit : public GraphKit { 65 private: 66 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 67 68 public: 69 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic) 70 : GraphKit(caller), 71 _intrinsic(intrinsic) 72 { 73 } 74 75 ciMethod* caller() const { return jvms()->method(); } 76 int bci() const { return jvms()->bci(); } 77 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 79 ciMethod* callee() const { return _intrinsic->method(); } 80 ciSignature* signature() const { return callee()->signature(); } 81 int arg_size() const { return callee()->arg_size(); } 82 83 bool try_to_inline(); 84 85 // Helper functions to inline natives 86 void push_result(RegionNode* region, PhiNode* value); 87 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 88 Node* generate_slow_guard(Node* test, RegionNode* region); 89 Node* generate_fair_guard(Node* test, RegionNode* region); 90 Node* generate_negative_guard(Node* index, RegionNode* region, 91 // resulting CastII of index: 92 Node* *pos_index = NULL); 93 Node* generate_limit_guard(Node* offset, Node* subseq_length, 94 Node* array_length, 95 RegionNode* region); 96 Node* generate_current_thread(Node* &tls_output); 97 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 98 bool disjoint_bases, const char* &name, bool dest_uninitialized); 99 Node* load_mirror_from_klass(Node* klass); 100 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 101 int nargs, 102 RegionNode* region, int null_path, 103 int offset); 104 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs, 105 RegionNode* region, int null_path) { 106 int offset = java_lang_Class::klass_offset_in_bytes(); 107 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 108 region, null_path, 109 offset); 110 } 111 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 112 int nargs, 113 RegionNode* region, int null_path) { 114 int offset = java_lang_Class::array_klass_offset_in_bytes(); 115 return load_klass_from_mirror_common(mirror, never_see_null, nargs, 116 region, null_path, 117 offset); 118 } 119 Node* generate_access_flags_guard(Node* kls, 120 int modifier_mask, int modifier_bits, 121 RegionNode* region); 122 Node* generate_interface_guard(Node* kls, RegionNode* region); 123 Node* generate_array_guard(Node* kls, RegionNode* region) { 124 return generate_array_guard_common(kls, region, false, false); 125 } 126 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 127 return generate_array_guard_common(kls, region, false, true); 128 } 129 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 130 return generate_array_guard_common(kls, region, true, false); 131 } 132 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 133 return generate_array_guard_common(kls, region, true, true); 134 } 135 Node* generate_array_guard_common(Node* kls, RegionNode* region, 136 bool obj_array, bool not_array); 137 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 138 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 139 bool is_virtual = false, bool is_static = false); 140 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 141 return generate_method_call(method_id, false, true); 142 } 143 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 144 return generate_method_call(method_id, true, false); 145 } 146 147 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 148 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 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 // Helper for inline_unsafe_access. 168 // Generates the guards that check whether the result of 169 // Unsafe.getObject should be recorded in an SATB log buffer. 170 void insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val); 171 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 172 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 173 bool inline_unsafe_allocate(); 174 bool inline_unsafe_copyMemory(); 175 bool inline_native_currentThread(); 176 #ifdef TRACE_HAVE_INTRINSICS 177 bool inline_native_classID(); 178 bool inline_native_threadID(); 179 #endif 180 bool inline_native_time_funcs(address method, const char* funcName); 181 bool inline_native_isInterrupted(); 182 bool inline_native_Class_query(vmIntrinsics::ID id); 183 bool inline_native_subtype_check(); 184 185 bool inline_native_newArray(); 186 bool inline_native_getLength(); 187 bool inline_array_copyOf(bool is_copyOfRange); 188 bool inline_array_equals(); 189 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 190 bool inline_native_clone(bool is_virtual); 191 bool inline_native_Reflection_getCallerClass(); 192 bool inline_native_AtomicLong_get(); 193 bool inline_native_AtomicLong_attemptUpdate(); 194 bool is_method_invoke_or_aux_frame(JVMState* jvms); 195 // Helper function for inlining native object hash method 196 bool inline_native_hashcode(bool is_virtual, bool is_static); 197 bool inline_native_getClass(); 198 199 // Helper functions for inlining arraycopy 200 bool inline_arraycopy(); 201 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 202 RegionNode* slow_region); 203 bool inline_unsafe_CAS(BasicType type); 204 bool inline_unsafe_ordered_store(BasicType type); 205 bool inline_fp_conversions(vmIntrinsics::ID id); 206 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id); 207 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id); 208 bool inline_bitCount(vmIntrinsics::ID id); 209 bool inline_reverseBytes(vmIntrinsics::ID id); 210 211 bool inline_reference_get(); 212 }; 213 214 215 //---------------------------make_vm_intrinsic---------------------------- 216 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 217 vmIntrinsics::ID id = m->intrinsic_id(); 218 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 219 220 if (DisableIntrinsic[0] != '\0' 221 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 222 // disabled by a user request on the command line: 223 // example: -XX:DisableIntrinsic=_hashCode,_getClass 224 return NULL; 225 } 226 227 if (!m->is_loaded()) { 228 // do not attempt to inline unloaded methods 229 return NULL; 230 } 231 232 // Only a few intrinsics implement a virtual dispatch. 233 // They are expensive calls which are also frequently overridden. 234 if (is_virtual) { 235 switch (id) { 236 case vmIntrinsics::_hashCode: 237 case vmIntrinsics::_clone: 238 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 239 break; 240 default: 241 return NULL; 242 } 243 } 244 245 // -XX:-InlineNatives disables nearly all intrinsics: 246 if (!InlineNatives) { 247 switch (id) { 248 case vmIntrinsics::_indexOf: 249 case vmIntrinsics::_compareTo: 250 case vmIntrinsics::_equals: 251 case vmIntrinsics::_equalsC: 252 break; // InlineNatives does not control String.compareTo 253 default: 254 return NULL; 255 } 256 } 257 258 switch (id) { 259 case vmIntrinsics::_compareTo: 260 if (!SpecialStringCompareTo) return NULL; 261 break; 262 case vmIntrinsics::_indexOf: 263 if (!SpecialStringIndexOf) return NULL; 264 break; 265 case vmIntrinsics::_equals: 266 if (!SpecialStringEquals) return NULL; 267 break; 268 case vmIntrinsics::_equalsC: 269 if (!SpecialArraysEquals) return NULL; 270 break; 271 case vmIntrinsics::_arraycopy: 272 if (!InlineArrayCopy) return NULL; 273 break; 274 case vmIntrinsics::_copyMemory: 275 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 276 if (!InlineArrayCopy) return NULL; 277 break; 278 case vmIntrinsics::_hashCode: 279 if (!InlineObjectHash) return NULL; 280 break; 281 case vmIntrinsics::_clone: 282 case vmIntrinsics::_copyOf: 283 case vmIntrinsics::_copyOfRange: 284 if (!InlineObjectCopy) return NULL; 285 // These also use the arraycopy intrinsic mechanism: 286 if (!InlineArrayCopy) return NULL; 287 break; 288 case vmIntrinsics::_checkIndex: 289 // We do not intrinsify this. The optimizer does fine with it. 290 return NULL; 291 292 case vmIntrinsics::_get_AtomicLong: 293 case vmIntrinsics::_attemptUpdate: 294 if (!InlineAtomicLong) return NULL; 295 break; 296 297 case vmIntrinsics::_getCallerClass: 298 if (!UseNewReflection) return NULL; 299 if (!InlineReflectionGetCallerClass) return NULL; 300 if (!JDK_Version::is_gte_jdk14x_version()) return NULL; 301 break; 302 303 case vmIntrinsics::_bitCount_i: 304 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 305 break; 306 307 case vmIntrinsics::_bitCount_l: 308 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 309 break; 310 311 case vmIntrinsics::_numberOfLeadingZeros_i: 312 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 313 break; 314 315 case vmIntrinsics::_numberOfLeadingZeros_l: 316 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 317 break; 318 319 case vmIntrinsics::_numberOfTrailingZeros_i: 320 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 321 break; 322 323 case vmIntrinsics::_numberOfTrailingZeros_l: 324 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 325 break; 326 327 case vmIntrinsics::_Reference_get: 328 // It is only when G1 is enabled that we absolutely 329 // need to use the intrinsic version of Reference.get() 330 // so that the value in the referent field, if necessary, 331 // can be registered by the pre-barrier code. 332 if (!UseG1GC) return NULL; 333 break; 334 335 default: 336 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 337 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 338 break; 339 } 340 341 // -XX:-InlineClassNatives disables natives from the Class class. 342 // The flag applies to all reflective calls, notably Array.newArray 343 // (visible to Java programmers as Array.newInstance). 344 if (m->holder()->name() == ciSymbol::java_lang_Class() || 345 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 346 if (!InlineClassNatives) return NULL; 347 } 348 349 // -XX:-InlineThreadNatives disables natives from the Thread class. 350 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 351 if (!InlineThreadNatives) return NULL; 352 } 353 354 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 355 if (m->holder()->name() == ciSymbol::java_lang_Math() || 356 m->holder()->name() == ciSymbol::java_lang_Float() || 357 m->holder()->name() == ciSymbol::java_lang_Double()) { 358 if (!InlineMathNatives) return NULL; 359 } 360 361 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 362 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 363 if (!InlineUnsafeOps) return NULL; 364 } 365 366 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); 367 } 368 369 //----------------------register_library_intrinsics----------------------- 370 // Initialize this file's data structures, for each Compile instance. 371 void Compile::register_library_intrinsics() { 372 // Nothing to do here. 373 } 374 375 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 376 LibraryCallKit kit(jvms, this); 377 Compile* C = kit.C; 378 int nodes = C->unique(); 379 #ifndef PRODUCT 380 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 381 char buf[1000]; 382 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 383 tty->print_cr("Intrinsic %s", str); 384 } 385 #endif 386 387 if (kit.try_to_inline()) { 388 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 389 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 390 } 391 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 392 if (C->log()) { 393 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 394 vmIntrinsics::name_at(intrinsic_id()), 395 (is_virtual() ? " virtual='1'" : ""), 396 C->unique() - nodes); 397 } 398 return kit.transfer_exceptions_into_jvms(); 399 } 400 401 // The intrinsic bailed out 402 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 403 if (jvms->has_method()) { 404 // Not a root compile. 405 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 406 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg); 407 } else { 408 // Root compile 409 tty->print("Did not generate intrinsic %s%s at bci:%d in", 410 vmIntrinsics::name_at(intrinsic_id()), 411 (is_virtual() ? " (virtual)" : ""), kit.bci()); 412 } 413 } 414 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 415 return NULL; 416 } 417 418 bool LibraryCallKit::try_to_inline() { 419 // Handle symbolic names for otherwise undistinguished boolean switches: 420 const bool is_store = true; 421 const bool is_native_ptr = true; 422 const bool is_static = true; 423 424 if (!jvms()->has_method()) { 425 // Root JVMState has a null method. 426 assert(map()->memory()->Opcode() == Op_Parm, ""); 427 // Insert the memory aliasing node 428 set_all_memory(reset_memory()); 429 } 430 assert(merged_memory(), ""); 431 432 switch (intrinsic_id()) { 433 case vmIntrinsics::_hashCode: 434 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 435 case vmIntrinsics::_identityHashCode: 436 return inline_native_hashcode(/*!virtual*/ false, is_static); 437 case vmIntrinsics::_getClass: 438 return inline_native_getClass(); 439 440 case vmIntrinsics::_dsin: 441 case vmIntrinsics::_dcos: 442 case vmIntrinsics::_dtan: 443 case vmIntrinsics::_dabs: 444 case vmIntrinsics::_datan2: 445 case vmIntrinsics::_dsqrt: 446 case vmIntrinsics::_dexp: 447 case vmIntrinsics::_dlog: 448 case vmIntrinsics::_dlog10: 449 case vmIntrinsics::_dpow: 450 return inline_math_native(intrinsic_id()); 451 452 case vmIntrinsics::_min: 453 case vmIntrinsics::_max: 454 return inline_min_max(intrinsic_id()); 455 456 case vmIntrinsics::_arraycopy: 457 return inline_arraycopy(); 458 459 case vmIntrinsics::_compareTo: 460 return inline_string_compareTo(); 461 case vmIntrinsics::_indexOf: 462 return inline_string_indexOf(); 463 case vmIntrinsics::_equals: 464 return inline_string_equals(); 465 466 case vmIntrinsics::_getObject: 467 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); 468 case vmIntrinsics::_getBoolean: 469 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); 470 case vmIntrinsics::_getByte: 471 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); 472 case vmIntrinsics::_getShort: 473 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); 474 case vmIntrinsics::_getChar: 475 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); 476 case vmIntrinsics::_getInt: 477 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); 478 case vmIntrinsics::_getLong: 479 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); 480 case vmIntrinsics::_getFloat: 481 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); 482 case vmIntrinsics::_getDouble: 483 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); 484 485 case vmIntrinsics::_putObject: 486 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); 487 case vmIntrinsics::_putBoolean: 488 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); 489 case vmIntrinsics::_putByte: 490 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); 491 case vmIntrinsics::_putShort: 492 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); 493 case vmIntrinsics::_putChar: 494 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); 495 case vmIntrinsics::_putInt: 496 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); 497 case vmIntrinsics::_putLong: 498 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); 499 case vmIntrinsics::_putFloat: 500 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); 501 case vmIntrinsics::_putDouble: 502 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); 503 504 case vmIntrinsics::_getByte_raw: 505 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); 506 case vmIntrinsics::_getShort_raw: 507 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); 508 case vmIntrinsics::_getChar_raw: 509 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); 510 case vmIntrinsics::_getInt_raw: 511 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); 512 case vmIntrinsics::_getLong_raw: 513 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); 514 case vmIntrinsics::_getFloat_raw: 515 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); 516 case vmIntrinsics::_getDouble_raw: 517 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); 518 case vmIntrinsics::_getAddress_raw: 519 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); 520 521 case vmIntrinsics::_putByte_raw: 522 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); 523 case vmIntrinsics::_putShort_raw: 524 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); 525 case vmIntrinsics::_putChar_raw: 526 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); 527 case vmIntrinsics::_putInt_raw: 528 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); 529 case vmIntrinsics::_putLong_raw: 530 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); 531 case vmIntrinsics::_putFloat_raw: 532 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); 533 case vmIntrinsics::_putDouble_raw: 534 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); 535 case vmIntrinsics::_putAddress_raw: 536 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); 537 538 case vmIntrinsics::_getObjectVolatile: 539 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); 540 case vmIntrinsics::_getBooleanVolatile: 541 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); 542 case vmIntrinsics::_getByteVolatile: 543 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); 544 case vmIntrinsics::_getShortVolatile: 545 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); 546 case vmIntrinsics::_getCharVolatile: 547 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); 548 case vmIntrinsics::_getIntVolatile: 549 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); 550 case vmIntrinsics::_getLongVolatile: 551 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); 552 case vmIntrinsics::_getFloatVolatile: 553 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); 554 case vmIntrinsics::_getDoubleVolatile: 555 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); 556 557 case vmIntrinsics::_putObjectVolatile: 558 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); 559 case vmIntrinsics::_putBooleanVolatile: 560 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); 561 case vmIntrinsics::_putByteVolatile: 562 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); 563 case vmIntrinsics::_putShortVolatile: 564 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); 565 case vmIntrinsics::_putCharVolatile: 566 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); 567 case vmIntrinsics::_putIntVolatile: 568 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); 569 case vmIntrinsics::_putLongVolatile: 570 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); 571 case vmIntrinsics::_putFloatVolatile: 572 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); 573 case vmIntrinsics::_putDoubleVolatile: 574 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); 575 576 case vmIntrinsics::_prefetchRead: 577 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 578 case vmIntrinsics::_prefetchWrite: 579 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 580 case vmIntrinsics::_prefetchReadStatic: 581 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 582 case vmIntrinsics::_prefetchWriteStatic: 583 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 584 585 case vmIntrinsics::_compareAndSwapObject: 586 return inline_unsafe_CAS(T_OBJECT); 587 case vmIntrinsics::_compareAndSwapInt: 588 return inline_unsafe_CAS(T_INT); 589 case vmIntrinsics::_compareAndSwapLong: 590 return inline_unsafe_CAS(T_LONG); 591 592 case vmIntrinsics::_putOrderedObject: 593 return inline_unsafe_ordered_store(T_OBJECT); 594 case vmIntrinsics::_putOrderedInt: 595 return inline_unsafe_ordered_store(T_INT); 596 case vmIntrinsics::_putOrderedLong: 597 return inline_unsafe_ordered_store(T_LONG); 598 599 case vmIntrinsics::_currentThread: 600 return inline_native_currentThread(); 601 case vmIntrinsics::_isInterrupted: 602 return inline_native_isInterrupted(); 603 604 #ifdef TRACE_HAVE_INTRINSICS 605 case vmIntrinsics::_classID: 606 return inline_native_classID(); 607 case vmIntrinsics::_threadID: 608 return inline_native_threadID(); 609 case vmIntrinsics::_counterTime: 610 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 611 #endif 612 case vmIntrinsics::_currentTimeMillis: 613 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 614 case vmIntrinsics::_nanoTime: 615 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 616 case vmIntrinsics::_allocateInstance: 617 return inline_unsafe_allocate(); 618 case vmIntrinsics::_copyMemory: 619 return inline_unsafe_copyMemory(); 620 case vmIntrinsics::_newArray: 621 return inline_native_newArray(); 622 case vmIntrinsics::_getLength: 623 return inline_native_getLength(); 624 case vmIntrinsics::_copyOf: 625 return inline_array_copyOf(false); 626 case vmIntrinsics::_copyOfRange: 627 return inline_array_copyOf(true); 628 case vmIntrinsics::_equalsC: 629 return inline_array_equals(); 630 case vmIntrinsics::_clone: 631 return inline_native_clone(intrinsic()->is_virtual()); 632 633 case vmIntrinsics::_isAssignableFrom: 634 return inline_native_subtype_check(); 635 636 case vmIntrinsics::_isInstance: 637 case vmIntrinsics::_getModifiers: 638 case vmIntrinsics::_isInterface: 639 case vmIntrinsics::_isArray: 640 case vmIntrinsics::_isPrimitive: 641 case vmIntrinsics::_getSuperclass: 642 case vmIntrinsics::_getComponentType: 643 case vmIntrinsics::_getClassAccessFlags: 644 return inline_native_Class_query(intrinsic_id()); 645 646 case vmIntrinsics::_floatToRawIntBits: 647 case vmIntrinsics::_floatToIntBits: 648 case vmIntrinsics::_intBitsToFloat: 649 case vmIntrinsics::_doubleToRawLongBits: 650 case vmIntrinsics::_doubleToLongBits: 651 case vmIntrinsics::_longBitsToDouble: 652 return inline_fp_conversions(intrinsic_id()); 653 654 case vmIntrinsics::_numberOfLeadingZeros_i: 655 case vmIntrinsics::_numberOfLeadingZeros_l: 656 return inline_numberOfLeadingZeros(intrinsic_id()); 657 658 case vmIntrinsics::_numberOfTrailingZeros_i: 659 case vmIntrinsics::_numberOfTrailingZeros_l: 660 return inline_numberOfTrailingZeros(intrinsic_id()); 661 662 case vmIntrinsics::_bitCount_i: 663 case vmIntrinsics::_bitCount_l: 664 return inline_bitCount(intrinsic_id()); 665 666 case vmIntrinsics::_reverseBytes_i: 667 case vmIntrinsics::_reverseBytes_l: 668 case vmIntrinsics::_reverseBytes_s: 669 case vmIntrinsics::_reverseBytes_c: 670 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); 671 672 case vmIntrinsics::_get_AtomicLong: 673 return inline_native_AtomicLong_get(); 674 case vmIntrinsics::_attemptUpdate: 675 return inline_native_AtomicLong_attemptUpdate(); 676 677 case vmIntrinsics::_getCallerClass: 678 return inline_native_Reflection_getCallerClass(); 679 680 case vmIntrinsics::_Reference_get: 681 return inline_reference_get(); 682 683 default: 684 // If you get here, it may be that someone has added a new intrinsic 685 // to the list in vmSymbols.hpp without implementing it here. 686 #ifndef PRODUCT 687 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 688 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 689 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 690 } 691 #endif 692 return false; 693 } 694 } 695 696 //------------------------------push_result------------------------------ 697 // Helper function for finishing intrinsics. 698 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { 699 record_for_igvn(region); 700 set_control(_gvn.transform(region)); 701 BasicType value_type = value->type()->basic_type(); 702 push_node(value_type, _gvn.transform(value)); 703 } 704 705 //------------------------------generate_guard--------------------------- 706 // Helper function for generating guarded fast-slow graph structures. 707 // The given 'test', if true, guards a slow path. If the test fails 708 // then a fast path can be taken. (We generally hope it fails.) 709 // In all cases, GraphKit::control() is updated to the fast path. 710 // The returned value represents the control for the slow path. 711 // The return value is never 'top'; it is either a valid control 712 // or NULL if it is obvious that the slow path can never be taken. 713 // Also, if region and the slow control are not NULL, the slow edge 714 // is appended to the region. 715 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 716 if (stopped()) { 717 // Already short circuited. 718 return NULL; 719 } 720 721 // Build an if node and its projections. 722 // If test is true we take the slow path, which we assume is uncommon. 723 if (_gvn.type(test) == TypeInt::ZERO) { 724 // The slow branch is never taken. No need to build this guard. 725 return NULL; 726 } 727 728 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 729 730 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); 731 if (if_slow == top()) { 732 // The slow branch is never taken. No need to build this guard. 733 return NULL; 734 } 735 736 if (region != NULL) 737 region->add_req(if_slow); 738 739 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); 740 set_control(if_fast); 741 742 return if_slow; 743 } 744 745 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 746 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 747 } 748 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 749 return generate_guard(test, region, PROB_FAIR); 750 } 751 752 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 753 Node* *pos_index) { 754 if (stopped()) 755 return NULL; // already stopped 756 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 757 return NULL; // index is already adequately typed 758 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 759 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 760 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 761 if (is_neg != NULL && pos_index != NULL) { 762 // Emulate effect of Parse::adjust_map_after_if. 763 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); 764 ccast->set_req(0, control()); 765 (*pos_index) = _gvn.transform(ccast); 766 } 767 return is_neg; 768 } 769 770 // Make sure that 'position' is a valid limit index, in [0..length]. 771 // There are two equivalent plans for checking this: 772 // A. (offset + copyLength) unsigned<= arrayLength 773 // B. offset <= (arrayLength - copyLength) 774 // We require that all of the values above, except for the sum and 775 // difference, are already known to be non-negative. 776 // Plan A is robust in the face of overflow, if offset and copyLength 777 // are both hugely positive. 778 // 779 // Plan B is less direct and intuitive, but it does not overflow at 780 // all, since the difference of two non-negatives is always 781 // representable. Whenever Java methods must perform the equivalent 782 // check they generally use Plan B instead of Plan A. 783 // For the moment we use Plan A. 784 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 785 Node* subseq_length, 786 Node* array_length, 787 RegionNode* region) { 788 if (stopped()) 789 return NULL; // already stopped 790 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 791 if (zero_offset && subseq_length->eqv_uncast(array_length)) 792 return NULL; // common case of whole-array copy 793 Node* last = subseq_length; 794 if (!zero_offset) // last += offset 795 last = _gvn.transform( new (C, 3) AddINode(last, offset)); 796 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); 797 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 798 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 799 return is_over; 800 } 801 802 803 //--------------------------generate_current_thread-------------------- 804 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 805 ciKlass* thread_klass = env()->Thread_klass(); 806 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 807 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); 808 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 809 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 810 tls_output = thread; 811 return threadObj; 812 } 813 814 815 //------------------------------make_string_method_node------------------------ 816 // Helper method for String intrinsic functions. This version is called 817 // with str1 and str2 pointing to String object nodes. 818 // 819 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 820 Node* no_ctrl = NULL; 821 822 // Get start addr of string 823 Node* str1_value = load_String_value(no_ctrl, str1); 824 Node* str1_offset = load_String_offset(no_ctrl, str1); 825 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 826 827 // Get length of string 1 828 Node* str1_len = load_String_length(no_ctrl, str1); 829 830 Node* str2_value = load_String_value(no_ctrl, str2); 831 Node* str2_offset = load_String_offset(no_ctrl, str2); 832 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 833 834 Node* str2_len = NULL; 835 Node* result = NULL; 836 837 switch (opcode) { 838 case Op_StrIndexOf: 839 // Get length of string 2 840 str2_len = load_String_length(no_ctrl, str2); 841 842 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 843 str1_start, str1_len, str2_start, str2_len); 844 break; 845 case Op_StrComp: 846 // Get length of string 2 847 str2_len = load_String_length(no_ctrl, str2); 848 849 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS), 850 str1_start, str1_len, str2_start, str2_len); 851 break; 852 case Op_StrEquals: 853 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 854 str1_start, str2_start, str1_len); 855 break; 856 default: 857 ShouldNotReachHere(); 858 return NULL; 859 } 860 861 // All these intrinsics have checks. 862 C->set_has_split_ifs(true); // Has chance for split-if optimization 863 864 return _gvn.transform(result); 865 } 866 867 // Helper method for String intrinsic functions. This version is called 868 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 869 // to Int nodes containing the lenghts of str1 and str2. 870 // 871 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 872 873 Node* result = NULL; 874 switch (opcode) { 875 case Op_StrIndexOf: 876 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 877 str1_start, cnt1, str2_start, cnt2); 878 break; 879 case Op_StrComp: 880 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS), 881 str1_start, cnt1, str2_start, cnt2); 882 break; 883 case Op_StrEquals: 884 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 885 str1_start, str2_start, cnt1); 886 break; 887 default: 888 ShouldNotReachHere(); 889 return NULL; 890 } 891 892 // All these intrinsics have checks. 893 C->set_has_split_ifs(true); // Has chance for split-if optimization 894 895 return _gvn.transform(result); 896 } 897 898 //------------------------------inline_string_compareTo------------------------ 899 bool LibraryCallKit::inline_string_compareTo() { 900 901 if (!Matcher::has_match_rule(Op_StrComp)) return false; 902 903 _sp += 2; 904 Node *argument = pop(); // pop non-receiver first: it was pushed second 905 Node *receiver = pop(); 906 907 // Null check on self without removing any arguments. The argument 908 // null check technically happens in the wrong place, which can lead to 909 // invalid stack traces when string compare is inlined into a method 910 // which handles NullPointerExceptions. 911 _sp += 2; 912 receiver = do_null_check(receiver, T_OBJECT); 913 argument = do_null_check(argument, T_OBJECT); 914 _sp -= 2; 915 if (stopped()) { 916 return true; 917 } 918 919 Node* compare = make_string_method_node(Op_StrComp, receiver, argument); 920 push(compare); 921 return true; 922 } 923 924 //------------------------------inline_string_equals------------------------ 925 bool LibraryCallKit::inline_string_equals() { 926 927 if (!Matcher::has_match_rule(Op_StrEquals)) return false; 928 929 int nargs = 2; 930 _sp += nargs; 931 Node* argument = pop(); // pop non-receiver first: it was pushed second 932 Node* receiver = pop(); 933 934 // Null check on self without removing any arguments. The argument 935 // null check technically happens in the wrong place, which can lead to 936 // invalid stack traces when string compare is inlined into a method 937 // which handles NullPointerExceptions. 938 _sp += nargs; 939 receiver = do_null_check(receiver, T_OBJECT); 940 //should not do null check for argument for String.equals(), because spec 941 //allows to specify NULL as argument. 942 _sp -= nargs; 943 944 if (stopped()) { 945 return true; 946 } 947 948 // paths (plus control) merge 949 RegionNode* region = new (C, 5) RegionNode(5); 950 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL); 951 952 // does source == target string? 953 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument)); 954 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq)); 955 956 Node* if_eq = generate_slow_guard(bol, NULL); 957 if (if_eq != NULL) { 958 // receiver == argument 959 phi->init_req(2, intcon(1)); 960 region->init_req(2, if_eq); 961 } 962 963 // get String klass for instanceOf 964 ciInstanceKlass* klass = env()->String_klass(); 965 966 if (!stopped()) { 967 _sp += nargs; // gen_instanceof might do an uncommon trap 968 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 969 _sp -= nargs; 970 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1))); 971 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne)); 972 973 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 974 //instanceOf == true, fallthrough 975 976 if (inst_false != NULL) { 977 phi->init_req(3, intcon(0)); 978 region->init_req(3, inst_false); 979 } 980 } 981 982 if (!stopped()) { 983 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 984 985 // Properly cast the argument to String 986 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type)); 987 // This path is taken only when argument's type is String:NotNull. 988 argument = cast_not_null(argument, false); 989 990 Node* no_ctrl = NULL; 991 992 // Get start addr of receiver 993 Node* receiver_val = load_String_value(no_ctrl, receiver); 994 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 995 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 996 997 // Get length of receiver 998 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 999 1000 // Get start addr of argument 1001 Node* argument_val = load_String_value(no_ctrl, argument); 1002 Node* argument_offset = load_String_offset(no_ctrl, argument); 1003 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1004 1005 // Get length of argument 1006 Node* argument_cnt = load_String_length(no_ctrl, argument); 1007 1008 // Check for receiver count != argument count 1009 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) ); 1010 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) ); 1011 Node* if_ne = generate_slow_guard(bol, NULL); 1012 if (if_ne != NULL) { 1013 phi->init_req(4, intcon(0)); 1014 region->init_req(4, if_ne); 1015 } 1016 1017 // Check for count == 0 is done by assembler code for StrEquals. 1018 1019 if (!stopped()) { 1020 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1021 phi->init_req(1, equals); 1022 region->init_req(1, control()); 1023 } 1024 } 1025 1026 // post merge 1027 set_control(_gvn.transform(region)); 1028 record_for_igvn(region); 1029 1030 push(_gvn.transform(phi)); 1031 1032 return true; 1033 } 1034 1035 //------------------------------inline_array_equals---------------------------- 1036 bool LibraryCallKit::inline_array_equals() { 1037 1038 if (!Matcher::has_match_rule(Op_AryEq)) return false; 1039 1040 _sp += 2; 1041 Node *argument2 = pop(); 1042 Node *argument1 = pop(); 1043 1044 Node* equals = 1045 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS), 1046 argument1, argument2) ); 1047 push(equals); 1048 return true; 1049 } 1050 1051 // Java version of String.indexOf(constant string) 1052 // class StringDecl { 1053 // StringDecl(char[] ca) { 1054 // offset = 0; 1055 // count = ca.length; 1056 // value = ca; 1057 // } 1058 // int offset; 1059 // int count; 1060 // char[] value; 1061 // } 1062 // 1063 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1064 // int targetOffset, int cache_i, int md2) { 1065 // int cache = cache_i; 1066 // int sourceOffset = string_object.offset; 1067 // int sourceCount = string_object.count; 1068 // int targetCount = target_object.length; 1069 // 1070 // int targetCountLess1 = targetCount - 1; 1071 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1072 // 1073 // char[] source = string_object.value; 1074 // char[] target = target_object; 1075 // int lastChar = target[targetCountLess1]; 1076 // 1077 // outer_loop: 1078 // for (int i = sourceOffset; i < sourceEnd; ) { 1079 // int src = source[i + targetCountLess1]; 1080 // if (src == lastChar) { 1081 // // With random strings and a 4-character alphabet, 1082 // // reverse matching at this point sets up 0.8% fewer 1083 // // frames, but (paradoxically) makes 0.3% more probes. 1084 // // Since those probes are nearer the lastChar probe, 1085 // // there is may be a net D$ win with reverse matching. 1086 // // But, reversing loop inhibits unroll of inner loop 1087 // // for unknown reason. So, does running outer loop from 1088 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1089 // for (int j = 0; j < targetCountLess1; j++) { 1090 // if (target[targetOffset + j] != source[i+j]) { 1091 // if ((cache & (1 << source[i+j])) == 0) { 1092 // if (md2 < j+1) { 1093 // i += j+1; 1094 // continue outer_loop; 1095 // } 1096 // } 1097 // i += md2; 1098 // continue outer_loop; 1099 // } 1100 // } 1101 // return i - sourceOffset; 1102 // } 1103 // if ((cache & (1 << src)) == 0) { 1104 // i += targetCountLess1; 1105 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1106 // i++; 1107 // } 1108 // return -1; 1109 // } 1110 1111 //------------------------------string_indexOf------------------------ 1112 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1113 jint cache_i, jint md2_i) { 1114 1115 Node* no_ctrl = NULL; 1116 float likely = PROB_LIKELY(0.9); 1117 float unlikely = PROB_UNLIKELY(0.9); 1118 1119 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate 1120 1121 Node* source = load_String_value(no_ctrl, string_object); 1122 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1123 Node* sourceCount = load_String_length(no_ctrl, string_object); 1124 1125 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) ); 1126 jint target_length = target_array->length(); 1127 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1128 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1129 1130 IdealKit kit(this, false, true); 1131 #define __ kit. 1132 Node* zero = __ ConI(0); 1133 Node* one = __ ConI(1); 1134 Node* cache = __ ConI(cache_i); 1135 Node* md2 = __ ConI(md2_i); 1136 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1137 Node* targetCount = __ ConI(target_length); 1138 Node* targetCountLess1 = __ ConI(target_length - 1); 1139 Node* targetOffset = __ ConI(targetOffset_i); 1140 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1141 1142 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1143 Node* outer_loop = __ make_label(2 /* goto */); 1144 Node* return_ = __ make_label(1); 1145 1146 __ set(rtn,__ ConI(-1)); 1147 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1148 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1149 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1150 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1151 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1152 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1153 Node* tpj = __ AddI(targetOffset, __ value(j)); 1154 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1155 Node* ipj = __ AddI(__ value(i), __ value(j)); 1156 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1157 __ if_then(targ, BoolTest::ne, src2); { 1158 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1159 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1160 __ increment(i, __ AddI(__ value(j), one)); 1161 __ goto_(outer_loop); 1162 } __ end_if(); __ dead(j); 1163 }__ end_if(); __ dead(j); 1164 __ increment(i, md2); 1165 __ goto_(outer_loop); 1166 }__ end_if(); 1167 __ increment(j, one); 1168 }__ end_loop(); __ dead(j); 1169 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1170 __ goto_(return_); 1171 }__ end_if(); 1172 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1173 __ increment(i, targetCountLess1); 1174 }__ end_if(); 1175 __ increment(i, one); 1176 __ bind(outer_loop); 1177 }__ end_loop(); __ dead(i); 1178 __ bind(return_); 1179 1180 // Final sync IdealKit and GraphKit. 1181 final_sync(kit); 1182 Node* result = __ value(rtn); 1183 #undef __ 1184 C->set_has_loops(true); 1185 return result; 1186 } 1187 1188 //------------------------------inline_string_indexOf------------------------ 1189 bool LibraryCallKit::inline_string_indexOf() { 1190 1191 _sp += 2; 1192 Node *argument = pop(); // pop non-receiver first: it was pushed second 1193 Node *receiver = pop(); 1194 1195 Node* result; 1196 // Disable the use of pcmpestri until it can be guaranteed that 1197 // the load doesn't cross into the uncommited space. 1198 if (Matcher::has_match_rule(Op_StrIndexOf) && 1199 UseSSE42Intrinsics) { 1200 // Generate SSE4.2 version of indexOf 1201 // We currently only have match rules that use SSE4.2 1202 1203 // Null check on self without removing any arguments. The argument 1204 // null check technically happens in the wrong place, which can lead to 1205 // invalid stack traces when string compare is inlined into a method 1206 // which handles NullPointerExceptions. 1207 _sp += 2; 1208 receiver = do_null_check(receiver, T_OBJECT); 1209 argument = do_null_check(argument, T_OBJECT); 1210 _sp -= 2; 1211 1212 if (stopped()) { 1213 return true; 1214 } 1215 1216 ciInstanceKlass* str_klass = env()->String_klass(); 1217 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1218 1219 // Make the merge point 1220 RegionNode* result_rgn = new (C, 4) RegionNode(4); 1221 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT); 1222 Node* no_ctrl = NULL; 1223 1224 // Get start addr of source string 1225 Node* source = load_String_value(no_ctrl, receiver); 1226 Node* source_offset = load_String_offset(no_ctrl, receiver); 1227 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1228 1229 // Get length of source string 1230 Node* source_cnt = load_String_length(no_ctrl, receiver); 1231 1232 // Get start addr of substring 1233 Node* substr = load_String_value(no_ctrl, argument); 1234 Node* substr_offset = load_String_offset(no_ctrl, argument); 1235 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1236 1237 // Get length of source string 1238 Node* substr_cnt = load_String_length(no_ctrl, argument); 1239 1240 // Check for substr count > string count 1241 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) ); 1242 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) ); 1243 Node* if_gt = generate_slow_guard(bol, NULL); 1244 if (if_gt != NULL) { 1245 result_phi->init_req(2, intcon(-1)); 1246 result_rgn->init_req(2, if_gt); 1247 } 1248 1249 if (!stopped()) { 1250 // Check for substr count == 0 1251 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) ); 1252 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 1253 Node* if_zero = generate_slow_guard(bol, NULL); 1254 if (if_zero != NULL) { 1255 result_phi->init_req(3, intcon(0)); 1256 result_rgn->init_req(3, if_zero); 1257 } 1258 } 1259 1260 if (!stopped()) { 1261 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1262 result_phi->init_req(1, result); 1263 result_rgn->init_req(1, control()); 1264 } 1265 set_control(_gvn.transform(result_rgn)); 1266 record_for_igvn(result_rgn); 1267 result = _gvn.transform(result_phi); 1268 1269 } else { // Use LibraryCallKit::string_indexOf 1270 // don't intrinsify if argument isn't a constant string. 1271 if (!argument->is_Con()) { 1272 return false; 1273 } 1274 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); 1275 if (str_type == NULL) { 1276 return false; 1277 } 1278 ciInstanceKlass* klass = env()->String_klass(); 1279 ciObject* str_const = str_type->const_oop(); 1280 if (str_const == NULL || str_const->klass() != klass) { 1281 return false; 1282 } 1283 ciInstance* str = str_const->as_instance(); 1284 assert(str != NULL, "must be instance"); 1285 1286 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1287 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1288 1289 int o; 1290 int c; 1291 if (java_lang_String::has_offset_field()) { 1292 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1293 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1294 } else { 1295 o = 0; 1296 c = pat->length(); 1297 } 1298 1299 // constant strings have no offset and count == length which 1300 // simplifies the resulting code somewhat so lets optimize for that. 1301 if (o != 0 || c != pat->length()) { 1302 return false; 1303 } 1304 1305 // Null check on self without removing any arguments. The argument 1306 // null check technically happens in the wrong place, which can lead to 1307 // invalid stack traces when string compare is inlined into a method 1308 // which handles NullPointerExceptions. 1309 _sp += 2; 1310 receiver = do_null_check(receiver, T_OBJECT); 1311 // No null check on the argument is needed since it's a constant String oop. 1312 _sp -= 2; 1313 if (stopped()) { 1314 return true; 1315 } 1316 1317 // The null string as a pattern always returns 0 (match at beginning of string) 1318 if (c == 0) { 1319 push(intcon(0)); 1320 return true; 1321 } 1322 1323 // Generate default indexOf 1324 jchar lastChar = pat->char_at(o + (c - 1)); 1325 int cache = 0; 1326 int i; 1327 for (i = 0; i < c - 1; i++) { 1328 assert(i < pat->length(), "out of range"); 1329 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1330 } 1331 1332 int md2 = c; 1333 for (i = 0; i < c - 1; i++) { 1334 assert(i < pat->length(), "out of range"); 1335 if (pat->char_at(o + i) == lastChar) { 1336 md2 = (c - 1) - i; 1337 } 1338 } 1339 1340 result = string_indexOf(receiver, pat, o, cache, md2); 1341 } 1342 1343 push(result); 1344 return true; 1345 } 1346 1347 //--------------------------pop_math_arg-------------------------------- 1348 // Pop a double argument to a math function from the stack 1349 // rounding it if necessary. 1350 Node * LibraryCallKit::pop_math_arg() { 1351 Node *arg = pop_pair(); 1352 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) 1353 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); 1354 return arg; 1355 } 1356 1357 //------------------------------inline_trig---------------------------------- 1358 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1359 // argument reduction which will turn into a fast/slow diamond. 1360 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1361 _sp += arg_size(); // restore stack pointer 1362 Node* arg = pop_math_arg(); 1363 Node* trig = NULL; 1364 1365 switch (id) { 1366 case vmIntrinsics::_dsin: 1367 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); 1368 break; 1369 case vmIntrinsics::_dcos: 1370 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); 1371 break; 1372 case vmIntrinsics::_dtan: 1373 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); 1374 break; 1375 default: 1376 assert(false, "bad intrinsic was passed in"); 1377 return false; 1378 } 1379 1380 // Rounding required? Check for argument reduction! 1381 if( Matcher::strict_fp_requires_explicit_rounding ) { 1382 1383 static const double pi_4 = 0.7853981633974483; 1384 static const double neg_pi_4 = -0.7853981633974483; 1385 // pi/2 in 80-bit extended precision 1386 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1387 // -pi/2 in 80-bit extended precision 1388 // 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}; 1389 // Cutoff value for using this argument reduction technique 1390 //static const double pi_2_minus_epsilon = 1.564660403643354; 1391 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1392 1393 // Pseudocode for sin: 1394 // if (x <= Math.PI / 4.0) { 1395 // if (x >= -Math.PI / 4.0) return fsin(x); 1396 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1397 // } else { 1398 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1399 // } 1400 // return StrictMath.sin(x); 1401 1402 // Pseudocode for cos: 1403 // if (x <= Math.PI / 4.0) { 1404 // if (x >= -Math.PI / 4.0) return fcos(x); 1405 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1406 // } else { 1407 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1408 // } 1409 // return StrictMath.cos(x); 1410 1411 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1412 // requires a special machine instruction to load it. Instead we'll try 1413 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1414 // probably do the math inside the SIN encoding. 1415 1416 // Make the merge point 1417 RegionNode *r = new (C, 3) RegionNode(3); 1418 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); 1419 1420 // Flatten arg so we need only 1 test 1421 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); 1422 // Node for PI/4 constant 1423 Node *pi4 = makecon(TypeD::make(pi_4)); 1424 // Check PI/4 : abs(arg) 1425 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); 1426 // Check: If PI/4 < abs(arg) then go slow 1427 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); 1428 // Branch either way 1429 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1430 set_control(opt_iff(r,iff)); 1431 1432 // Set fast path result 1433 phi->init_req(2,trig); 1434 1435 // Slow path - non-blocking leaf call 1436 Node* call = NULL; 1437 switch (id) { 1438 case vmIntrinsics::_dsin: 1439 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1440 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1441 "Sin", NULL, arg, top()); 1442 break; 1443 case vmIntrinsics::_dcos: 1444 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1445 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1446 "Cos", NULL, arg, top()); 1447 break; 1448 case vmIntrinsics::_dtan: 1449 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1450 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1451 "Tan", NULL, arg, top()); 1452 break; 1453 } 1454 assert(control()->in(0) == call, ""); 1455 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); 1456 r->init_req(1,control()); 1457 phi->init_req(1,slow_result); 1458 1459 // Post-merge 1460 set_control(_gvn.transform(r)); 1461 record_for_igvn(r); 1462 trig = _gvn.transform(phi); 1463 1464 C->set_has_split_ifs(true); // Has chance for split-if optimization 1465 } 1466 // Push result back on JVM stack 1467 push_pair(trig); 1468 return true; 1469 } 1470 1471 //------------------------------inline_sqrt------------------------------------- 1472 // Inline square root instruction, if possible. 1473 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { 1474 assert(id == vmIntrinsics::_dsqrt, "Not square root"); 1475 _sp += arg_size(); // restore stack pointer 1476 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); 1477 return true; 1478 } 1479 1480 //------------------------------inline_abs------------------------------------- 1481 // Inline absolute value instruction, if possible. 1482 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { 1483 assert(id == vmIntrinsics::_dabs, "Not absolute value"); 1484 _sp += arg_size(); // restore stack pointer 1485 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); 1486 return true; 1487 } 1488 1489 //------------------------------inline_exp------------------------------------- 1490 // Inline exp instructions, if possible. The Intel hardware only misses 1491 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1492 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { 1493 assert(id == vmIntrinsics::_dexp, "Not exp"); 1494 1495 // If this inlining ever returned NaN in the past, we do not intrinsify it 1496 // every again. NaN results requires StrictMath.exp handling. 1497 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1498 1499 _sp += arg_size(); // restore stack pointer 1500 Node *x = pop_math_arg(); 1501 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); 1502 1503 //------------------- 1504 //result=(result.isNaN())? StrictMath::exp():result; 1505 // Check: If isNaN() by checking result!=result? then go to Strict Math 1506 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1507 // Build the boolean node 1508 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1509 1510 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1511 // End the current control-flow path 1512 push_pair(x); 1513 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp 1514 // to handle. Recompile without intrinsifying Math.exp 1515 uncommon_trap(Deoptimization::Reason_intrinsic, 1516 Deoptimization::Action_make_not_entrant); 1517 } 1518 1519 C->set_has_split_ifs(true); // Has chance for split-if optimization 1520 1521 push_pair(result); 1522 1523 return true; 1524 } 1525 1526 //------------------------------inline_pow------------------------------------- 1527 // Inline power instructions, if possible. 1528 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { 1529 assert(id == vmIntrinsics::_dpow, "Not pow"); 1530 1531 // If this inlining ever returned NaN in the past, we do not intrinsify it 1532 // every again. NaN results requires StrictMath.pow handling. 1533 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 1534 1535 // Do not intrinsify on older platforms which lack cmove. 1536 if (ConditionalMoveLimit == 0) return false; 1537 1538 // Pseudocode for pow 1539 // if (x <= 0.0) { 1540 // if ((double)((int)y)==y) { // if y is int 1541 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y) 1542 // } else { 1543 // result = NaN; 1544 // } 1545 // } else { 1546 // result = DPow(x,y); 1547 // } 1548 // if (result != result)? { 1549 // uncommon_trap(); 1550 // } 1551 // return result; 1552 1553 _sp += arg_size(); // restore stack pointer 1554 Node* y = pop_math_arg(); 1555 Node* x = pop_math_arg(); 1556 1557 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); 1558 1559 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow 1560 // inside of something) then skip the fancy tests and just check for 1561 // NaN result. 1562 Node *result = NULL; 1563 if( jvms()->depth() >= 1 ) { 1564 result = fast_result; 1565 } else { 1566 1567 // Set the merge point for If node with condition of (x <= 0.0) 1568 // There are four possible paths to region node and phi node 1569 RegionNode *r = new (C, 4) RegionNode(4); 1570 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); 1571 1572 // Build the first if node: if (x <= 0.0) 1573 // Node for 0 constant 1574 Node *zeronode = makecon(TypeD::ZERO); 1575 // Check x:0 1576 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); 1577 // Check: If (x<=0) then go complex path 1578 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); 1579 // Branch either way 1580 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1581 Node *opt_test = _gvn.transform(if1); 1582 //assert( opt_test->is_If(), "Expect an IfNode"); 1583 IfNode *opt_if1 = (IfNode*)opt_test; 1584 // Fast path taken; set region slot 3 1585 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) ); 1586 r->init_req(3,fast_taken); // Capture fast-control 1587 1588 // Fast path not-taken, i.e. slow path 1589 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) ); 1590 1591 // Set fast path result 1592 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) ); 1593 phi->init_req(3, fast_result); 1594 1595 // Complex path 1596 // Build the second if node (if y is int) 1597 // Node for (int)y 1598 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y)); 1599 // Node for (double)((int) y) 1600 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty)); 1601 // Check (double)((int) y) : y 1602 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y)); 1603 // Check if (y isn't int) then go to slow path 1604 1605 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) ); 1606 // Branch either way 1607 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1608 Node *slow_path = opt_iff(r,if2); // Set region path 2 1609 1610 // Calculate DPow(abs(x), y)*(1 & (int)y) 1611 // Node for constant 1 1612 Node *conone = intcon(1); 1613 // 1& (int)y 1614 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) ); 1615 // zero node 1616 Node *conzero = intcon(0); 1617 // Check (1&(int)y)==0? 1618 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero)); 1619 // Check if (1&(int)y)!=0?, if so the result is negative 1620 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); 1621 // abs(x) 1622 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); 1623 // abs(x)^y 1624 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) ); 1625 // -abs(x)^y 1626 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); 1627 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1628 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1629 // Set complex path fast result 1630 phi->init_req(2, signresult); 1631 1632 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1633 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1634 r->init_req(1,slow_path); 1635 phi->init_req(1,slow_result); 1636 1637 // Post merge 1638 set_control(_gvn.transform(r)); 1639 record_for_igvn(r); 1640 result=_gvn.transform(phi); 1641 } 1642 1643 //------------------- 1644 //result=(result.isNaN())? uncommon_trap():result; 1645 // Check: If isNaN() by checking result!=result? then go to Strict Math 1646 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1647 // Build the boolean node 1648 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1649 1650 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1651 // End the current control-flow path 1652 push_pair(x); 1653 push_pair(y); 1654 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow 1655 // to handle. Recompile without intrinsifying Math.pow. 1656 uncommon_trap(Deoptimization::Reason_intrinsic, 1657 Deoptimization::Action_make_not_entrant); 1658 } 1659 1660 C->set_has_split_ifs(true); // Has chance for split-if optimization 1661 1662 push_pair(result); 1663 1664 return true; 1665 } 1666 1667 //------------------------------inline_trans------------------------------------- 1668 // Inline transcendental instructions, if possible. The Intel hardware gets 1669 // these right, no funny corner cases missed. 1670 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { 1671 _sp += arg_size(); // restore stack pointer 1672 Node* arg = pop_math_arg(); 1673 Node* trans = NULL; 1674 1675 switch (id) { 1676 case vmIntrinsics::_dlog: 1677 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); 1678 break; 1679 case vmIntrinsics::_dlog10: 1680 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); 1681 break; 1682 default: 1683 assert(false, "bad intrinsic was passed in"); 1684 return false; 1685 } 1686 1687 // Push result back on JVM stack 1688 push_pair(trans); 1689 return true; 1690 } 1691 1692 //------------------------------runtime_math----------------------------- 1693 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1694 Node* a = NULL; 1695 Node* b = NULL; 1696 1697 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1698 "must be (DD)D or (D)D type"); 1699 1700 // Inputs 1701 _sp += arg_size(); // restore stack pointer 1702 if (call_type == OptoRuntime::Math_DD_D_Type()) { 1703 b = pop_math_arg(); 1704 } 1705 a = pop_math_arg(); 1706 1707 const TypePtr* no_memory_effects = NULL; 1708 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1709 no_memory_effects, 1710 a, top(), b, b ? top() : NULL); 1711 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); 1712 #ifdef ASSERT 1713 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); 1714 assert(value_top == top(), "second value must be top"); 1715 #endif 1716 1717 push_pair(value); 1718 return true; 1719 } 1720 1721 //------------------------------inline_math_native----------------------------- 1722 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1723 switch (id) { 1724 // These intrinsics are not properly supported on all hardware 1725 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1726 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); 1727 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1728 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); 1729 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1730 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); 1731 1732 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : 1733 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); 1734 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : 1735 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); 1736 1737 // These intrinsics are supported on all hardware 1738 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; 1739 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; 1740 1741 case vmIntrinsics::_dexp: return 1742 Matcher::has_match_rule(Op_ExpD) ? inline_exp(id) : 1743 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1744 case vmIntrinsics::_dpow: return 1745 Matcher::has_match_rule(Op_PowD) ? inline_pow(id) : 1746 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1747 1748 // These intrinsics are not yet correctly implemented 1749 case vmIntrinsics::_datan2: 1750 return false; 1751 1752 default: 1753 ShouldNotReachHere(); 1754 return false; 1755 } 1756 } 1757 1758 static bool is_simple_name(Node* n) { 1759 return (n->req() == 1 // constant 1760 || (n->is_Type() && n->as_Type()->type()->singleton()) 1761 || n->is_Proj() // parameter or return value 1762 || n->is_Phi() // local of some sort 1763 ); 1764 } 1765 1766 //----------------------------inline_min_max----------------------------------- 1767 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1768 push(generate_min_max(id, argument(0), argument(1))); 1769 1770 return true; 1771 } 1772 1773 Node* 1774 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1775 // These are the candidate return value: 1776 Node* xvalue = x0; 1777 Node* yvalue = y0; 1778 1779 if (xvalue == yvalue) { 1780 return xvalue; 1781 } 1782 1783 bool want_max = (id == vmIntrinsics::_max); 1784 1785 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1786 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1787 if (txvalue == NULL || tyvalue == NULL) return top(); 1788 // This is not really necessary, but it is consistent with a 1789 // hypothetical MaxINode::Value method: 1790 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1791 1792 // %%% This folding logic should (ideally) be in a different place. 1793 // Some should be inside IfNode, and there to be a more reliable 1794 // transformation of ?: style patterns into cmoves. We also want 1795 // more powerful optimizations around cmove and min/max. 1796 1797 // Try to find a dominating comparison of these guys. 1798 // It can simplify the index computation for Arrays.copyOf 1799 // and similar uses of System.arraycopy. 1800 // First, compute the normalized version of CmpI(x, y). 1801 int cmp_op = Op_CmpI; 1802 Node* xkey = xvalue; 1803 Node* ykey = yvalue; 1804 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); 1805 if (ideal_cmpxy->is_Cmp()) { 1806 // E.g., if we have CmpI(length - offset, count), 1807 // it might idealize to CmpI(length, count + offset) 1808 cmp_op = ideal_cmpxy->Opcode(); 1809 xkey = ideal_cmpxy->in(1); 1810 ykey = ideal_cmpxy->in(2); 1811 } 1812 1813 // Start by locating any relevant comparisons. 1814 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1815 Node* cmpxy = NULL; 1816 Node* cmpyx = NULL; 1817 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1818 Node* cmp = start_from->fast_out(k); 1819 if (cmp->outcnt() > 0 && // must have prior uses 1820 cmp->in(0) == NULL && // must be context-independent 1821 cmp->Opcode() == cmp_op) { // right kind of compare 1822 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1823 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1824 } 1825 } 1826 1827 const int NCMPS = 2; 1828 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1829 int cmpn; 1830 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1831 if (cmps[cmpn] != NULL) break; // find a result 1832 } 1833 if (cmpn < NCMPS) { 1834 // Look for a dominating test that tells us the min and max. 1835 int depth = 0; // Limit search depth for speed 1836 Node* dom = control(); 1837 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1838 if (++depth >= 100) break; 1839 Node* ifproj = dom; 1840 if (!ifproj->is_Proj()) continue; 1841 Node* iff = ifproj->in(0); 1842 if (!iff->is_If()) continue; 1843 Node* bol = iff->in(1); 1844 if (!bol->is_Bool()) continue; 1845 Node* cmp = bol->in(1); 1846 if (cmp == NULL) continue; 1847 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1848 if (cmps[cmpn] == cmp) break; 1849 if (cmpn == NCMPS) continue; 1850 BoolTest::mask btest = bol->as_Bool()->_test._test; 1851 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1852 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1853 // At this point, we know that 'x btest y' is true. 1854 switch (btest) { 1855 case BoolTest::eq: 1856 // They are proven equal, so we can collapse the min/max. 1857 // Either value is the answer. Choose the simpler. 1858 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1859 return yvalue; 1860 return xvalue; 1861 case BoolTest::lt: // x < y 1862 case BoolTest::le: // x <= y 1863 return (want_max ? yvalue : xvalue); 1864 case BoolTest::gt: // x > y 1865 case BoolTest::ge: // x >= y 1866 return (want_max ? xvalue : yvalue); 1867 } 1868 } 1869 } 1870 1871 // We failed to find a dominating test. 1872 // Let's pick a test that might GVN with prior tests. 1873 Node* best_bol = NULL; 1874 BoolTest::mask best_btest = BoolTest::illegal; 1875 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1876 Node* cmp = cmps[cmpn]; 1877 if (cmp == NULL) continue; 1878 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1879 Node* bol = cmp->fast_out(j); 1880 if (!bol->is_Bool()) continue; 1881 BoolTest::mask btest = bol->as_Bool()->_test._test; 1882 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1883 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1884 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1885 best_bol = bol->as_Bool(); 1886 best_btest = btest; 1887 } 1888 } 1889 } 1890 1891 Node* answer_if_true = NULL; 1892 Node* answer_if_false = NULL; 1893 switch (best_btest) { 1894 default: 1895 if (cmpxy == NULL) 1896 cmpxy = ideal_cmpxy; 1897 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); 1898 // and fall through: 1899 case BoolTest::lt: // x < y 1900 case BoolTest::le: // x <= y 1901 answer_if_true = (want_max ? yvalue : xvalue); 1902 answer_if_false = (want_max ? xvalue : yvalue); 1903 break; 1904 case BoolTest::gt: // x > y 1905 case BoolTest::ge: // x >= y 1906 answer_if_true = (want_max ? xvalue : yvalue); 1907 answer_if_false = (want_max ? yvalue : xvalue); 1908 break; 1909 } 1910 1911 jint hi, lo; 1912 if (want_max) { 1913 // We can sharpen the minimum. 1914 hi = MAX2(txvalue->_hi, tyvalue->_hi); 1915 lo = MAX2(txvalue->_lo, tyvalue->_lo); 1916 } else { 1917 // We can sharpen the maximum. 1918 hi = MIN2(txvalue->_hi, tyvalue->_hi); 1919 lo = MIN2(txvalue->_lo, tyvalue->_lo); 1920 } 1921 1922 // Use a flow-free graph structure, to avoid creating excess control edges 1923 // which could hinder other optimizations. 1924 // Since Math.min/max is often used with arraycopy, we want 1925 // tightly_coupled_allocation to be able to see beyond min/max expressions. 1926 Node* cmov = CMoveNode::make(C, NULL, best_bol, 1927 answer_if_false, answer_if_true, 1928 TypeInt::make(lo, hi, widen)); 1929 1930 return _gvn.transform(cmov); 1931 1932 /* 1933 // This is not as desirable as it may seem, since Min and Max 1934 // nodes do not have a full set of optimizations. 1935 // And they would interfere, anyway, with 'if' optimizations 1936 // and with CMoveI canonical forms. 1937 switch (id) { 1938 case vmIntrinsics::_min: 1939 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 1940 case vmIntrinsics::_max: 1941 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 1942 default: 1943 ShouldNotReachHere(); 1944 } 1945 */ 1946 } 1947 1948 inline int 1949 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 1950 const TypePtr* base_type = TypePtr::NULL_PTR; 1951 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 1952 if (base_type == NULL) { 1953 // Unknown type. 1954 return Type::AnyPtr; 1955 } else if (base_type == TypePtr::NULL_PTR) { 1956 // Since this is a NULL+long form, we have to switch to a rawptr. 1957 base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); 1958 offset = MakeConX(0); 1959 return Type::RawPtr; 1960 } else if (base_type->base() == Type::RawPtr) { 1961 return Type::RawPtr; 1962 } else if (base_type->isa_oopptr()) { 1963 // Base is never null => always a heap address. 1964 if (base_type->ptr() == TypePtr::NotNull) { 1965 return Type::OopPtr; 1966 } 1967 // Offset is small => always a heap address. 1968 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 1969 if (offset_type != NULL && 1970 base_type->offset() == 0 && // (should always be?) 1971 offset_type->_lo >= 0 && 1972 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 1973 return Type::OopPtr; 1974 } 1975 // Otherwise, it might either be oop+off or NULL+addr. 1976 return Type::AnyPtr; 1977 } else { 1978 // No information: 1979 return Type::AnyPtr; 1980 } 1981 } 1982 1983 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 1984 int kind = classify_unsafe_addr(base, offset); 1985 if (kind == Type::RawPtr) { 1986 return basic_plus_adr(top(), base, offset); 1987 } else { 1988 return basic_plus_adr(base, offset); 1989 } 1990 } 1991 1992 //-------------------inline_numberOfLeadingZeros_int/long----------------------- 1993 // inline int Integer.numberOfLeadingZeros(int) 1994 // inline int Long.numberOfLeadingZeros(long) 1995 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) { 1996 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros"); 1997 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false; 1998 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false; 1999 _sp += arg_size(); // restore stack pointer 2000 switch (id) { 2001 case vmIntrinsics::_numberOfLeadingZeros_i: 2002 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop()))); 2003 break; 2004 case vmIntrinsics::_numberOfLeadingZeros_l: 2005 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair()))); 2006 break; 2007 default: 2008 ShouldNotReachHere(); 2009 } 2010 return true; 2011 } 2012 2013 //-------------------inline_numberOfTrailingZeros_int/long---------------------- 2014 // inline int Integer.numberOfTrailingZeros(int) 2015 // inline int Long.numberOfTrailingZeros(long) 2016 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) { 2017 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros"); 2018 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false; 2019 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false; 2020 _sp += arg_size(); // restore stack pointer 2021 switch (id) { 2022 case vmIntrinsics::_numberOfTrailingZeros_i: 2023 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop()))); 2024 break; 2025 case vmIntrinsics::_numberOfTrailingZeros_l: 2026 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair()))); 2027 break; 2028 default: 2029 ShouldNotReachHere(); 2030 } 2031 return true; 2032 } 2033 2034 //----------------------------inline_bitCount_int/long----------------------- 2035 // inline int Integer.bitCount(int) 2036 // inline int Long.bitCount(long) 2037 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) { 2038 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount"); 2039 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false; 2040 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false; 2041 _sp += arg_size(); // restore stack pointer 2042 switch (id) { 2043 case vmIntrinsics::_bitCount_i: 2044 push(_gvn.transform(new (C, 2) PopCountINode(pop()))); 2045 break; 2046 case vmIntrinsics::_bitCount_l: 2047 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair()))); 2048 break; 2049 default: 2050 ShouldNotReachHere(); 2051 } 2052 return true; 2053 } 2054 2055 //----------------------------inline_reverseBytes_int/long/char/short------------------- 2056 // inline Integer.reverseBytes(int) 2057 // inline Long.reverseBytes(long) 2058 // inline Character.reverseBytes(char) 2059 // inline Short.reverseBytes(short) 2060 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { 2061 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l || 2062 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s, 2063 "not reverse Bytes"); 2064 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; 2065 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; 2066 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false; 2067 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false; 2068 _sp += arg_size(); // restore stack pointer 2069 switch (id) { 2070 case vmIntrinsics::_reverseBytes_i: 2071 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); 2072 break; 2073 case vmIntrinsics::_reverseBytes_l: 2074 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); 2075 break; 2076 case vmIntrinsics::_reverseBytes_c: 2077 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop()))); 2078 break; 2079 case vmIntrinsics::_reverseBytes_s: 2080 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop()))); 2081 break; 2082 default: 2083 ; 2084 } 2085 return true; 2086 } 2087 2088 //----------------------------inline_unsafe_access---------------------------- 2089 2090 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2091 2092 // Helper that guards and inserts a G1 pre-barrier. 2093 void LibraryCallKit::insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val) { 2094 assert(UseG1GC, "should not call this otherwise"); 2095 2096 // We could be accessing the referent field of a reference object. If so, when G1 2097 // is enabled, we need to log the value in the referent field in an SATB buffer. 2098 // This routine performs some compile time filters and generates suitable 2099 // runtime filters that guard the pre-barrier code. 2100 2101 // Some compile time checks. 2102 2103 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2104 const TypeX* otype = offset->find_intptr_t_type(); 2105 if (otype != NULL && otype->is_con() && 2106 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2107 // Constant offset but not the reference_offset so just return 2108 return; 2109 } 2110 2111 // We only need to generate the runtime guards for instances. 2112 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2113 if (btype != NULL) { 2114 if (btype->isa_aryptr()) { 2115 // Array type so nothing to do 2116 return; 2117 } 2118 2119 const TypeInstPtr* itype = btype->isa_instptr(); 2120 if (itype != NULL) { 2121 // Can the klass of base_oop be statically determined 2122 // to be _not_ a sub-class of Reference? 2123 ciKlass* klass = itype->klass(); 2124 if (klass->is_subtype_of(env()->Reference_klass()) && 2125 !env()->Reference_klass()->is_subtype_of(klass)) { 2126 return; 2127 } 2128 } 2129 } 2130 2131 // The compile time filters did not reject base_oop/offset so 2132 // we need to generate the following runtime filters 2133 // 2134 // if (offset == java_lang_ref_Reference::_reference_offset) { 2135 // if (base != null) { 2136 // if (instance_of(base, java.lang.ref.Reference)) { 2137 // pre_barrier(_, pre_val, ...); 2138 // } 2139 // } 2140 // } 2141 2142 float likely = PROB_LIKELY(0.999); 2143 float unlikely = PROB_UNLIKELY(0.999); 2144 2145 IdealKit ideal(this); 2146 #define __ ideal. 2147 2148 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2149 2150 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2151 __ if_then(base_oop, BoolTest::ne, null(), likely); { 2152 2153 // Update graphKit memory and control from IdealKit. 2154 sync_kit(ideal); 2155 2156 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2157 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2158 2159 // Update IdealKit memory and control from graphKit. 2160 __ sync_kit(this); 2161 2162 Node* one = __ ConI(1); 2163 2164 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2165 2166 // Update graphKit from IdeakKit. 2167 sync_kit(ideal); 2168 2169 // Use the pre-barrier to record the value in the referent field 2170 pre_barrier(false /* do_load */, 2171 __ ctrl(), 2172 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2173 pre_val /* pre_val */, 2174 T_OBJECT); 2175 2176 // Update IdealKit from graphKit. 2177 __ sync_kit(this); 2178 2179 } __ end_if(); // _ref_type != ref_none 2180 } __ end_if(); // base != NULL 2181 } __ end_if(); // offset == referent_offset 2182 2183 // Final sync IdealKit and GraphKit. 2184 final_sync(ideal); 2185 #undef __ 2186 } 2187 2188 2189 // Interpret Unsafe.fieldOffset cookies correctly: 2190 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2191 2192 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2193 if (callee()->is_static()) return false; // caller must have the capability! 2194 2195 #ifndef PRODUCT 2196 { 2197 ResourceMark rm; 2198 // Check the signatures. 2199 ciSignature* sig = signature(); 2200 #ifdef ASSERT 2201 if (!is_store) { 2202 // Object getObject(Object base, int/long offset), etc. 2203 BasicType rtype = sig->return_type()->basic_type(); 2204 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2205 rtype = T_ADDRESS; // it is really a C void* 2206 assert(rtype == type, "getter must return the expected value"); 2207 if (!is_native_ptr) { 2208 assert(sig->count() == 2, "oop getter has 2 arguments"); 2209 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2210 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2211 } else { 2212 assert(sig->count() == 1, "native getter has 1 argument"); 2213 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2214 } 2215 } else { 2216 // void putObject(Object base, int/long offset, Object x), etc. 2217 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2218 if (!is_native_ptr) { 2219 assert(sig->count() == 3, "oop putter has 3 arguments"); 2220 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2221 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2222 } else { 2223 assert(sig->count() == 2, "native putter has 2 arguments"); 2224 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2225 } 2226 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2227 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2228 vtype = T_ADDRESS; // it is really a C void* 2229 assert(vtype == type, "putter must accept the expected value"); 2230 } 2231 #endif // ASSERT 2232 } 2233 #endif //PRODUCT 2234 2235 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2236 2237 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; 2238 2239 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words 2240 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); 2241 2242 debug_only(int saved_sp = _sp); 2243 _sp += nargs; 2244 2245 Node* val; 2246 debug_only(val = (Node*)(uintptr_t)-1); 2247 2248 2249 if (is_store) { 2250 // Get the value being stored. (Pop it first; it was pushed last.) 2251 switch (type) { 2252 case T_DOUBLE: 2253 case T_LONG: 2254 case T_ADDRESS: 2255 val = pop_pair(); 2256 break; 2257 default: 2258 val = pop(); 2259 } 2260 } 2261 2262 // Build address expression. See the code in inline_unsafe_prefetch. 2263 Node *adr; 2264 Node *heap_base_oop = top(); 2265 Node* offset = top(); 2266 2267 if (!is_native_ptr) { 2268 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2269 offset = pop_pair(); 2270 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2271 Node* base = pop(); 2272 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2273 // to be plain byte offsets, which are also the same as those accepted 2274 // by oopDesc::field_base. 2275 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2276 "fieldOffset must be byte-scaled"); 2277 // 32-bit machines ignore the high half! 2278 offset = ConvL2X(offset); 2279 adr = make_unsafe_address(base, offset); 2280 heap_base_oop = base; 2281 } else { 2282 Node* ptr = pop_pair(); 2283 // Adjust Java long to machine word: 2284 ptr = ConvL2X(ptr); 2285 adr = make_unsafe_address(NULL, ptr); 2286 } 2287 2288 // Pop receiver last: it was pushed first. 2289 Node *receiver = pop(); 2290 2291 assert(saved_sp == _sp, "must have correct argument count"); 2292 2293 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2294 2295 // First guess at the value type. 2296 const Type *value_type = Type::get_const_basic_type(type); 2297 2298 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2299 // there was not enough information to nail it down. 2300 Compile::AliasType* alias_type = C->alias_type(adr_type); 2301 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2302 2303 // We will need memory barriers unless we can determine a unique 2304 // alias category for this reference. (Note: If for some reason 2305 // the barriers get omitted and the unsafe reference begins to "pollute" 2306 // the alias analysis of the rest of the graph, either Compile::can_alias 2307 // or Compile::must_alias will throw a diagnostic assert.) 2308 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2309 2310 // If we are reading the value of the referent field of a Reference 2311 // object (either by using Unsafe directly or through reflection) 2312 // then, if G1 is enabled, we need to record the referent in an 2313 // SATB log buffer using the pre-barrier mechanism. 2314 bool need_read_barrier = UseG1GC && !is_native_ptr && !is_store && 2315 offset != top() && heap_base_oop != top(); 2316 2317 if (!is_store && type == T_OBJECT) { 2318 // Attempt to infer a sharper value type from the offset and base type. 2319 ciKlass* sharpened_klass = NULL; 2320 2321 // See if it is an instance field, with an object type. 2322 if (alias_type->field() != NULL) { 2323 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2324 if (alias_type->field()->type()->is_klass()) { 2325 sharpened_klass = alias_type->field()->type()->as_klass(); 2326 } 2327 } 2328 2329 // See if it is a narrow oop array. 2330 if (adr_type->isa_aryptr()) { 2331 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2332 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2333 if (elem_type != NULL) { 2334 sharpened_klass = elem_type->klass(); 2335 } 2336 } 2337 } 2338 2339 if (sharpened_klass != NULL) { 2340 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2341 2342 // Sharpen the value type. 2343 value_type = tjp; 2344 2345 #ifndef PRODUCT 2346 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2347 tty->print(" from base type: "); adr_type->dump(); 2348 tty->print(" sharpened value: "); value_type->dump(); 2349 } 2350 #endif 2351 } 2352 } 2353 2354 // Null check on self without removing any arguments. The argument 2355 // null check technically happens in the wrong place, which can lead to 2356 // invalid stack traces when the primitive is inlined into a method 2357 // which handles NullPointerExceptions. 2358 _sp += nargs; 2359 do_null_check(receiver, T_OBJECT); 2360 _sp -= nargs; 2361 if (stopped()) { 2362 return true; 2363 } 2364 // Heap pointers get a null-check from the interpreter, 2365 // as a courtesy. However, this is not guaranteed by Unsafe, 2366 // and it is not possible to fully distinguish unintended nulls 2367 // from intended ones in this API. 2368 2369 if (is_volatile) { 2370 // We need to emit leading and trailing CPU membars (see below) in 2371 // addition to memory membars when is_volatile. This is a little 2372 // too strong, but avoids the need to insert per-alias-type 2373 // volatile membars (for stores; compare Parse::do_put_xxx), which 2374 // we cannot do effectively here because we probably only have a 2375 // rough approximation of type. 2376 need_mem_bar = true; 2377 // For Stores, place a memory ordering barrier now. 2378 if (is_store) 2379 insert_mem_bar(Op_MemBarRelease); 2380 } 2381 2382 // Memory barrier to prevent normal and 'unsafe' accesses from 2383 // bypassing each other. Happens after null checks, so the 2384 // exception paths do not take memory state from the memory barrier, 2385 // so there's no problems making a strong assert about mixing users 2386 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2387 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2388 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2389 2390 if (!is_store) { 2391 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2392 // load value and push onto stack 2393 switch (type) { 2394 case T_BOOLEAN: 2395 case T_CHAR: 2396 case T_BYTE: 2397 case T_SHORT: 2398 case T_INT: 2399 case T_FLOAT: 2400 push(p); 2401 break; 2402 case T_OBJECT: 2403 if (need_read_barrier) { 2404 insert_g1_pre_barrier(heap_base_oop, offset, p); 2405 } 2406 push(p); 2407 break; 2408 case T_ADDRESS: 2409 // Cast to an int type. 2410 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); 2411 p = ConvX2L(p); 2412 push_pair(p); 2413 break; 2414 case T_DOUBLE: 2415 case T_LONG: 2416 push_pair( p ); 2417 break; 2418 default: ShouldNotReachHere(); 2419 } 2420 } else { 2421 // place effect of store into memory 2422 switch (type) { 2423 case T_DOUBLE: 2424 val = dstore_rounding(val); 2425 break; 2426 case T_ADDRESS: 2427 // Repackage the long as a pointer. 2428 val = ConvL2X(val); 2429 val = _gvn.transform( new (C, 2) CastX2PNode(val) ); 2430 break; 2431 } 2432 2433 if (type != T_OBJECT ) { 2434 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2435 } else { 2436 // Possibly an oop being stored to Java heap or native memory 2437 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2438 // oop to Java heap. 2439 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2440 } else { 2441 // We can't tell at compile time if we are storing in the Java heap or outside 2442 // of it. So we need to emit code to conditionally do the proper type of 2443 // store. 2444 2445 IdealKit ideal(this); 2446 #define __ ideal. 2447 // QQQ who knows what probability is here?? 2448 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2449 // Sync IdealKit and graphKit. 2450 sync_kit(ideal); 2451 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2452 // Update IdealKit memory. 2453 __ sync_kit(this); 2454 } __ else_(); { 2455 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile); 2456 } __ end_if(); 2457 // Final sync IdealKit and GraphKit. 2458 final_sync(ideal); 2459 #undef __ 2460 } 2461 } 2462 } 2463 2464 if (is_volatile) { 2465 if (!is_store) 2466 insert_mem_bar(Op_MemBarAcquire); 2467 else 2468 insert_mem_bar(Op_MemBarVolatile); 2469 } 2470 2471 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2472 2473 return true; 2474 } 2475 2476 //----------------------------inline_unsafe_prefetch---------------------------- 2477 2478 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2479 #ifndef PRODUCT 2480 { 2481 ResourceMark rm; 2482 // Check the signatures. 2483 ciSignature* sig = signature(); 2484 #ifdef ASSERT 2485 // Object getObject(Object base, int/long offset), etc. 2486 BasicType rtype = sig->return_type()->basic_type(); 2487 if (!is_native_ptr) { 2488 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2489 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2490 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2491 } else { 2492 assert(sig->count() == 1, "native prefetch has 1 argument"); 2493 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2494 } 2495 #endif // ASSERT 2496 } 2497 #endif // !PRODUCT 2498 2499 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2500 2501 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args 2502 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); 2503 2504 debug_only(int saved_sp = _sp); 2505 _sp += nargs; 2506 2507 // Build address expression. See the code in inline_unsafe_access. 2508 Node *adr; 2509 if (!is_native_ptr) { 2510 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2511 Node* offset = pop_pair(); 2512 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2513 Node* base = pop(); 2514 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2515 // to be plain byte offsets, which are also the same as those accepted 2516 // by oopDesc::field_base. 2517 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2518 "fieldOffset must be byte-scaled"); 2519 // 32-bit machines ignore the high half! 2520 offset = ConvL2X(offset); 2521 adr = make_unsafe_address(base, offset); 2522 } else { 2523 Node* ptr = pop_pair(); 2524 // Adjust Java long to machine word: 2525 ptr = ConvL2X(ptr); 2526 adr = make_unsafe_address(NULL, ptr); 2527 } 2528 2529 if (is_static) { 2530 assert(saved_sp == _sp, "must have correct argument count"); 2531 } else { 2532 // Pop receiver last: it was pushed first. 2533 Node *receiver = pop(); 2534 assert(saved_sp == _sp, "must have correct argument count"); 2535 2536 // Null check on self without removing any arguments. The argument 2537 // null check technically happens in the wrong place, which can lead to 2538 // invalid stack traces when the primitive is inlined into a method 2539 // which handles NullPointerExceptions. 2540 _sp += nargs; 2541 do_null_check(receiver, T_OBJECT); 2542 _sp -= nargs; 2543 if (stopped()) { 2544 return true; 2545 } 2546 } 2547 2548 // Generate the read or write prefetch 2549 Node *prefetch; 2550 if (is_store) { 2551 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); 2552 } else { 2553 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); 2554 } 2555 prefetch->init_req(0, control()); 2556 set_i_o(_gvn.transform(prefetch)); 2557 2558 return true; 2559 } 2560 2561 //----------------------------inline_unsafe_CAS---------------------------- 2562 2563 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { 2564 // This basic scheme here is the same as inline_unsafe_access, but 2565 // differs in enough details that combining them would make the code 2566 // overly confusing. (This is a true fact! I originally combined 2567 // them, but even I was confused by it!) As much code/comments as 2568 // possible are retained from inline_unsafe_access though to make 2569 // the correspondences clearer. - dl 2570 2571 if (callee()->is_static()) return false; // caller must have the capability! 2572 2573 #ifndef PRODUCT 2574 { 2575 ResourceMark rm; 2576 // Check the signatures. 2577 ciSignature* sig = signature(); 2578 #ifdef ASSERT 2579 BasicType rtype = sig->return_type()->basic_type(); 2580 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2581 assert(sig->count() == 4, "CAS has 4 arguments"); 2582 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2583 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2584 #endif // ASSERT 2585 } 2586 #endif //PRODUCT 2587 2588 // number of stack slots per value argument (1 or 2) 2589 int type_words = type2size[type]; 2590 2591 // Cannot inline wide CAS on machines that don't support it natively 2592 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8()) 2593 return false; 2594 2595 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2596 2597 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; 2598 int nargs = 1 + 1 + 2 + type_words + type_words; 2599 2600 // pop arguments: newval, oldval, offset, base, and receiver 2601 debug_only(int saved_sp = _sp); 2602 _sp += nargs; 2603 Node* newval = (type_words == 1) ? pop() : pop_pair(); 2604 Node* oldval = (type_words == 1) ? pop() : pop_pair(); 2605 Node *offset = pop_pair(); 2606 Node *base = pop(); 2607 Node *receiver = pop(); 2608 assert(saved_sp == _sp, "must have correct argument count"); 2609 2610 // Null check receiver. 2611 _sp += nargs; 2612 do_null_check(receiver, T_OBJECT); 2613 _sp -= nargs; 2614 if (stopped()) { 2615 return true; 2616 } 2617 2618 // Build field offset expression. 2619 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2620 // to be plain byte offsets, which are also the same as those accepted 2621 // by oopDesc::field_base. 2622 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2623 // 32-bit machines ignore the high half of long offsets 2624 offset = ConvL2X(offset); 2625 Node* adr = make_unsafe_address(base, offset); 2626 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2627 2628 // (Unlike inline_unsafe_access, there seems no point in trying 2629 // to refine types. Just use the coarse types here. 2630 const Type *value_type = Type::get_const_basic_type(type); 2631 Compile::AliasType* alias_type = C->alias_type(adr_type); 2632 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2633 int alias_idx = C->get_alias_index(adr_type); 2634 2635 // Memory-model-wise, a CAS acts like a little synchronized block, 2636 // so needs barriers on each side. These don't translate into 2637 // actual barriers on most machines, but we still need rest of 2638 // compiler to respect ordering. 2639 2640 insert_mem_bar(Op_MemBarRelease); 2641 insert_mem_bar(Op_MemBarCPUOrder); 2642 2643 // 4984716: MemBars must be inserted before this 2644 // memory node in order to avoid a false 2645 // dependency which will confuse the scheduler. 2646 Node *mem = memory(alias_idx); 2647 2648 // For now, we handle only those cases that actually exist: ints, 2649 // longs, and Object. Adding others should be straightforward. 2650 Node* cas; 2651 switch(type) { 2652 case T_INT: 2653 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2654 break; 2655 case T_LONG: 2656 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2657 break; 2658 case T_OBJECT: 2659 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2660 // could be delayed during Parse (for example, in adjust_map_after_if()). 2661 // Execute transformation here to avoid barrier generation in such case. 2662 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2663 newval = _gvn.makecon(TypePtr::NULL_PTR); 2664 2665 // Reference stores need a store barrier. 2666 // (They don't if CAS fails, but it isn't worth checking.) 2667 pre_barrier(true /* do_load*/, 2668 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2669 NULL /* pre_val*/, 2670 T_OBJECT); 2671 #ifdef _LP64 2672 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2673 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2674 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2675 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr, 2676 newval_enc, oldval_enc)); 2677 } else 2678 #endif 2679 { 2680 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2681 } 2682 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); 2683 break; 2684 default: 2685 ShouldNotReachHere(); 2686 break; 2687 } 2688 2689 // SCMemProjNodes represent the memory state of CAS. Their main 2690 // role is to prevent CAS nodes from being optimized away when their 2691 // results aren't used. 2692 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 2693 set_memory(proj, alias_idx); 2694 2695 // Add the trailing membar surrounding the access 2696 insert_mem_bar(Op_MemBarCPUOrder); 2697 insert_mem_bar(Op_MemBarAcquire); 2698 2699 push(cas); 2700 return true; 2701 } 2702 2703 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2704 // This is another variant of inline_unsafe_access, differing in 2705 // that it always issues store-store ("release") barrier and ensures 2706 // store-atomicity (which only matters for "long"). 2707 2708 if (callee()->is_static()) return false; // caller must have the capability! 2709 2710 #ifndef PRODUCT 2711 { 2712 ResourceMark rm; 2713 // Check the signatures. 2714 ciSignature* sig = signature(); 2715 #ifdef ASSERT 2716 BasicType rtype = sig->return_type()->basic_type(); 2717 assert(rtype == T_VOID, "must return void"); 2718 assert(sig->count() == 3, "has 3 arguments"); 2719 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2720 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2721 #endif // ASSERT 2722 } 2723 #endif //PRODUCT 2724 2725 // number of stack slots per value argument (1 or 2) 2726 int type_words = type2size[type]; 2727 2728 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2729 2730 // Argument words: "this" plus oop plus offset plus value; 2731 int nargs = 1 + 1 + 2 + type_words; 2732 2733 // pop arguments: val, offset, base, and receiver 2734 debug_only(int saved_sp = _sp); 2735 _sp += nargs; 2736 Node* val = (type_words == 1) ? pop() : pop_pair(); 2737 Node *offset = pop_pair(); 2738 Node *base = pop(); 2739 Node *receiver = pop(); 2740 assert(saved_sp == _sp, "must have correct argument count"); 2741 2742 // Null check receiver. 2743 _sp += nargs; 2744 do_null_check(receiver, T_OBJECT); 2745 _sp -= nargs; 2746 if (stopped()) { 2747 return true; 2748 } 2749 2750 // Build field offset expression. 2751 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2752 // 32-bit machines ignore the high half of long offsets 2753 offset = ConvL2X(offset); 2754 Node* adr = make_unsafe_address(base, offset); 2755 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2756 const Type *value_type = Type::get_const_basic_type(type); 2757 Compile::AliasType* alias_type = C->alias_type(adr_type); 2758 2759 insert_mem_bar(Op_MemBarRelease); 2760 insert_mem_bar(Op_MemBarCPUOrder); 2761 // Ensure that the store is atomic for longs: 2762 bool require_atomic_access = true; 2763 Node* store; 2764 if (type == T_OBJECT) // reference stores need a store barrier. 2765 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2766 else { 2767 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2768 } 2769 insert_mem_bar(Op_MemBarCPUOrder); 2770 return true; 2771 } 2772 2773 bool LibraryCallKit::inline_unsafe_allocate() { 2774 if (callee()->is_static()) return false; // caller must have the capability! 2775 int nargs = 1 + 1; 2776 assert(signature()->size() == nargs-1, "alloc has 1 argument"); 2777 null_check_receiver(callee()); // check then ignore argument(0) 2778 _sp += nargs; // set original stack for use by uncommon_trap 2779 Node* cls = do_null_check(argument(1), T_OBJECT); 2780 _sp -= nargs; 2781 if (stopped()) return true; 2782 2783 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 2784 _sp += nargs; // set original stack for use by uncommon_trap 2785 kls = do_null_check(kls, T_OBJECT); 2786 _sp -= nargs; 2787 if (stopped()) return true; // argument was like int.class 2788 2789 // Note: The argument might still be an illegal value like 2790 // Serializable.class or Object[].class. The runtime will handle it. 2791 // But we must make an explicit check for initialization. 2792 Node* insp = basic_plus_adr(kls, in_bytes(instanceKlass::init_state_offset())); 2793 // Use T_BOOLEAN for instanceKlass::_init_state so the compiler 2794 // can generate code to load it as unsigned byte. 2795 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN); 2796 Node* bits = intcon(instanceKlass::fully_initialized); 2797 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); 2798 // The 'test' is non-zero if we need to take a slow path. 2799 2800 Node* obj = new_instance(kls, test); 2801 push(obj); 2802 2803 return true; 2804 } 2805 2806 #ifdef TRACE_HAVE_INTRINSICS 2807 /* 2808 * oop -> myklass 2809 * myklass->trace_id |= USED 2810 * return myklass->trace_id & ~0x3 2811 */ 2812 bool LibraryCallKit::inline_native_classID() { 2813 int nargs = 1 + 1; 2814 null_check_receiver(callee()); // check then ignore argument(0) 2815 _sp += nargs; 2816 Node* cls = do_null_check(argument(1), T_OBJECT); 2817 _sp -= nargs; 2818 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 2819 _sp += nargs; 2820 kls = do_null_check(kls, T_OBJECT); 2821 _sp -= nargs; 2822 ByteSize offset = TRACE_ID_OFFSET; 2823 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 2824 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG); 2825 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 2826 Node* andl = _gvn.transform(new (C, 3) AndLNode(tvalue, bits)); 2827 Node* clsused = longcon(0x01l); // set the class bit 2828 Node* orl = _gvn.transform(new (C, 3) OrLNode(tvalue, clsused)); 2829 2830 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 2831 store_to_memory(control(), insp, orl, T_LONG, adr_type); 2832 push_pair(andl); 2833 return true; 2834 } 2835 2836 bool LibraryCallKit::inline_native_threadID() { 2837 Node* tls_ptr = NULL; 2838 Node* cur_thr = generate_current_thread(tls_ptr); 2839 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2840 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 2841 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 2842 2843 Node* threadid = NULL; 2844 size_t thread_id_size = OSThread::thread_id_size(); 2845 if (thread_id_size == (size_t) BytesPerLong) { 2846 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG)); 2847 push(threadid); 2848 } else if (thread_id_size == (size_t) BytesPerInt) { 2849 threadid = make_load(control(), p, TypeInt::INT, T_INT); 2850 push(threadid); 2851 } else { 2852 ShouldNotReachHere(); 2853 } 2854 return true; 2855 } 2856 #endif 2857 2858 //------------------------inline_native_time_funcs-------------- 2859 // inline code for System.currentTimeMillis() and System.nanoTime() 2860 // these have the same type and signature 2861 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 2862 const TypeFunc *tf = OptoRuntime::void_long_Type(); 2863 const TypePtr* no_memory_effects = NULL; 2864 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2865 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); 2866 #ifdef ASSERT 2867 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); 2868 assert(value_top == top(), "second value must be top"); 2869 #endif 2870 push_pair(value); 2871 return true; 2872 } 2873 2874 //------------------------inline_native_currentThread------------------ 2875 bool LibraryCallKit::inline_native_currentThread() { 2876 Node* junk = NULL; 2877 push(generate_current_thread(junk)); 2878 return true; 2879 } 2880 2881 //------------------------inline_native_isInterrupted------------------ 2882 bool LibraryCallKit::inline_native_isInterrupted() { 2883 const int nargs = 1+1; // receiver + boolean 2884 assert(nargs == arg_size(), "sanity"); 2885 // Add a fast path to t.isInterrupted(clear_int): 2886 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 2887 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 2888 // So, in the common case that the interrupt bit is false, 2889 // we avoid making a call into the VM. Even if the interrupt bit 2890 // is true, if the clear_int argument is false, we avoid the VM call. 2891 // However, if the receiver is not currentThread, we must call the VM, 2892 // because there must be some locking done around the operation. 2893 2894 // We only go to the fast case code if we pass two guards. 2895 // Paths which do not pass are accumulated in the slow_region. 2896 RegionNode* slow_region = new (C, 1) RegionNode(1); 2897 record_for_igvn(slow_region); 2898 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow 2899 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); 2900 enum { no_int_result_path = 1, 2901 no_clear_result_path = 2, 2902 slow_result_path = 3 2903 }; 2904 2905 // (a) Receiving thread must be the current thread. 2906 Node* rec_thr = argument(0); 2907 Node* tls_ptr = NULL; 2908 Node* cur_thr = generate_current_thread(tls_ptr); 2909 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); 2910 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); 2911 2912 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); 2913 if (!known_current_thread) 2914 generate_slow_guard(bol_thr, slow_region); 2915 2916 // (b) Interrupt bit on TLS must be false. 2917 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2918 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 2919 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 2920 // Set the control input on the field _interrupted read to prevent it floating up. 2921 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 2922 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); 2923 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); 2924 2925 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2926 2927 // First fast path: if (!TLS._interrupted) return false; 2928 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); 2929 result_rgn->init_req(no_int_result_path, false_bit); 2930 result_val->init_req(no_int_result_path, intcon(0)); 2931 2932 // drop through to next case 2933 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); 2934 2935 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 2936 Node* clr_arg = argument(1); 2937 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); 2938 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); 2939 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 2940 2941 // Second fast path: ... else if (!clear_int) return true; 2942 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); 2943 result_rgn->init_req(no_clear_result_path, false_arg); 2944 result_val->init_req(no_clear_result_path, intcon(1)); 2945 2946 // drop through to next case 2947 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); 2948 2949 // (d) Otherwise, go to the slow path. 2950 slow_region->add_req(control()); 2951 set_control( _gvn.transform(slow_region) ); 2952 2953 if (stopped()) { 2954 // There is no slow path. 2955 result_rgn->init_req(slow_result_path, top()); 2956 result_val->init_req(slow_result_path, top()); 2957 } else { 2958 // non-virtual because it is a private non-static 2959 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 2960 2961 Node* slow_val = set_results_for_java_call(slow_call); 2962 // this->control() comes from set_results_for_java_call 2963 2964 // If we know that the result of the slow call will be true, tell the optimizer! 2965 if (known_current_thread) slow_val = intcon(1); 2966 2967 Node* fast_io = slow_call->in(TypeFunc::I_O); 2968 Node* fast_mem = slow_call->in(TypeFunc::Memory); 2969 // These two phis are pre-filled with copies of of the fast IO and Memory 2970 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 2971 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 2972 2973 result_rgn->init_req(slow_result_path, control()); 2974 io_phi ->init_req(slow_result_path, i_o()); 2975 mem_phi ->init_req(slow_result_path, reset_memory()); 2976 result_val->init_req(slow_result_path, slow_val); 2977 2978 set_all_memory( _gvn.transform(mem_phi) ); 2979 set_i_o( _gvn.transform(io_phi) ); 2980 } 2981 2982 push_result(result_rgn, result_val); 2983 C->set_has_split_ifs(true); // Has chance for split-if optimization 2984 2985 return true; 2986 } 2987 2988 //---------------------------load_mirror_from_klass---------------------------- 2989 // Given a klass oop, load its java mirror (a java.lang.Class oop). 2990 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 2991 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 2992 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 2993 } 2994 2995 //-----------------------load_klass_from_mirror_common------------------------- 2996 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 2997 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 2998 // and branch to the given path on the region. 2999 // If never_see_null, take an uncommon trap on null, so we can optimistically 3000 // compile for the non-null case. 3001 // If the region is NULL, force never_see_null = true. 3002 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3003 bool never_see_null, 3004 int nargs, 3005 RegionNode* region, 3006 int null_path, 3007 int offset) { 3008 if (region == NULL) never_see_null = true; 3009 Node* p = basic_plus_adr(mirror, offset); 3010 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3011 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); 3012 _sp += nargs; // any deopt will start just before call to enclosing method 3013 Node* null_ctl = top(); 3014 kls = null_check_oop(kls, &null_ctl, never_see_null); 3015 if (region != NULL) { 3016 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3017 region->init_req(null_path, null_ctl); 3018 } else { 3019 assert(null_ctl == top(), "no loose ends"); 3020 } 3021 _sp -= nargs; 3022 return kls; 3023 } 3024 3025 //--------------------(inline_native_Class_query helpers)--------------------- 3026 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3027 // Fall through if (mods & mask) == bits, take the guard otherwise. 3028 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3029 // Branch around if the given klass has the given modifier bit set. 3030 // Like generate_guard, adds a new path onto the region. 3031 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3032 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 3033 Node* mask = intcon(modifier_mask); 3034 Node* bits = intcon(modifier_bits); 3035 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); 3036 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); 3037 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); 3038 return generate_fair_guard(bol, region); 3039 } 3040 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3041 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3042 } 3043 3044 //-------------------------inline_native_Class_query------------------- 3045 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3046 int nargs = 1+0; // just the Class mirror, in most cases 3047 const Type* return_type = TypeInt::BOOL; 3048 Node* prim_return_value = top(); // what happens if it's a primitive class? 3049 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3050 bool expect_prim = false; // most of these guys expect to work on refs 3051 3052 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3053 3054 switch (id) { 3055 case vmIntrinsics::_isInstance: 3056 nargs = 1+1; // the Class mirror, plus the object getting queried about 3057 // nothing is an instance of a primitive type 3058 prim_return_value = intcon(0); 3059 break; 3060 case vmIntrinsics::_getModifiers: 3061 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3062 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3063 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3064 break; 3065 case vmIntrinsics::_isInterface: 3066 prim_return_value = intcon(0); 3067 break; 3068 case vmIntrinsics::_isArray: 3069 prim_return_value = intcon(0); 3070 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3071 break; 3072 case vmIntrinsics::_isPrimitive: 3073 prim_return_value = intcon(1); 3074 expect_prim = true; // obviously 3075 break; 3076 case vmIntrinsics::_getSuperclass: 3077 prim_return_value = null(); 3078 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3079 break; 3080 case vmIntrinsics::_getComponentType: 3081 prim_return_value = null(); 3082 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3083 break; 3084 case vmIntrinsics::_getClassAccessFlags: 3085 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3086 return_type = TypeInt::INT; // not bool! 6297094 3087 break; 3088 default: 3089 ShouldNotReachHere(); 3090 } 3091 3092 Node* mirror = argument(0); 3093 Node* obj = (nargs <= 1)? top(): argument(1); 3094 3095 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3096 if (mirror_con == NULL) return false; // cannot happen? 3097 3098 #ifndef PRODUCT 3099 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 3100 ciType* k = mirror_con->java_mirror_type(); 3101 if (k) { 3102 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3103 k->print_name(); 3104 tty->cr(); 3105 } 3106 } 3107 #endif 3108 3109 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3110 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3111 record_for_igvn(region); 3112 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); 3113 3114 // The mirror will never be null of Reflection.getClassAccessFlags, however 3115 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3116 // if it is. See bug 4774291. 3117 3118 // For Reflection.getClassAccessFlags(), the null check occurs in 3119 // the wrong place; see inline_unsafe_access(), above, for a similar 3120 // situation. 3121 _sp += nargs; // set original stack for use by uncommon_trap 3122 mirror = do_null_check(mirror, T_OBJECT); 3123 _sp -= nargs; 3124 // If mirror or obj is dead, only null-path is taken. 3125 if (stopped()) return true; 3126 3127 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3128 3129 // Now load the mirror's klass metaobject, and null-check it. 3130 // Side-effects region with the control path if the klass is null. 3131 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, 3132 region, _prim_path); 3133 // If kls is null, we have a primitive mirror. 3134 phi->init_req(_prim_path, prim_return_value); 3135 if (stopped()) { push_result(region, phi); return true; } 3136 3137 Node* p; // handy temp 3138 Node* null_ctl; 3139 3140 // Now that we have the non-null klass, we can perform the real query. 3141 // For constant classes, the query will constant-fold in LoadNode::Value. 3142 Node* query_value = top(); 3143 switch (id) { 3144 case vmIntrinsics::_isInstance: 3145 // nothing is an instance of a primitive type 3146 _sp += nargs; // gen_instanceof might do an uncommon trap 3147 query_value = gen_instanceof(obj, kls); 3148 _sp -= nargs; 3149 break; 3150 3151 case vmIntrinsics::_getModifiers: 3152 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3153 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3154 break; 3155 3156 case vmIntrinsics::_isInterface: 3157 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3158 if (generate_interface_guard(kls, region) != NULL) 3159 // A guard was added. If the guard is taken, it was an interface. 3160 phi->add_req(intcon(1)); 3161 // If we fall through, it's a plain class. 3162 query_value = intcon(0); 3163 break; 3164 3165 case vmIntrinsics::_isArray: 3166 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3167 if (generate_array_guard(kls, region) != NULL) 3168 // A guard was added. If the guard is taken, it was an array. 3169 phi->add_req(intcon(1)); 3170 // If we fall through, it's a plain class. 3171 query_value = intcon(0); 3172 break; 3173 3174 case vmIntrinsics::_isPrimitive: 3175 query_value = intcon(0); // "normal" path produces false 3176 break; 3177 3178 case vmIntrinsics::_getSuperclass: 3179 // The rules here are somewhat unfortunate, but we can still do better 3180 // with random logic than with a JNI call. 3181 // Interfaces store null or Object as _super, but must report null. 3182 // Arrays store an intermediate super as _super, but must report Object. 3183 // Other types can report the actual _super. 3184 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3185 if (generate_interface_guard(kls, region) != NULL) 3186 // A guard was added. If the guard is taken, it was an interface. 3187 phi->add_req(null()); 3188 if (generate_array_guard(kls, region) != NULL) 3189 // A guard was added. If the guard is taken, it was an array. 3190 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3191 // If we fall through, it's a plain class. Get its _super. 3192 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3193 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); 3194 null_ctl = top(); 3195 kls = null_check_oop(kls, &null_ctl); 3196 if (null_ctl != top()) { 3197 // If the guard is taken, Object.superClass is null (both klass and mirror). 3198 region->add_req(null_ctl); 3199 phi ->add_req(null()); 3200 } 3201 if (!stopped()) { 3202 query_value = load_mirror_from_klass(kls); 3203 } 3204 break; 3205 3206 case vmIntrinsics::_getComponentType: 3207 if (generate_array_guard(kls, region) != NULL) { 3208 // Be sure to pin the oop load to the guard edge just created: 3209 Node* is_array_ctrl = region->in(region->req()-1); 3210 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset())); 3211 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 3212 phi->add_req(cmo); 3213 } 3214 query_value = null(); // non-array case is null 3215 break; 3216 3217 case vmIntrinsics::_getClassAccessFlags: 3218 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3219 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3220 break; 3221 3222 default: 3223 ShouldNotReachHere(); 3224 } 3225 3226 // Fall-through is the normal case of a query to a real class. 3227 phi->init_req(1, query_value); 3228 region->init_req(1, control()); 3229 3230 push_result(region, phi); 3231 C->set_has_split_ifs(true); // Has chance for split-if optimization 3232 3233 return true; 3234 } 3235 3236 //--------------------------inline_native_subtype_check------------------------ 3237 // This intrinsic takes the JNI calls out of the heart of 3238 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3239 bool LibraryCallKit::inline_native_subtype_check() { 3240 int nargs = 1+1; // the Class mirror, plus the other class getting examined 3241 3242 // Pull both arguments off the stack. 3243 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3244 args[0] = argument(0); 3245 args[1] = argument(1); 3246 Node* klasses[2]; // corresponding Klasses: superk, subk 3247 klasses[0] = klasses[1] = top(); 3248 3249 enum { 3250 // A full decision tree on {superc is prim, subc is prim}: 3251 _prim_0_path = 1, // {P,N} => false 3252 // {P,P} & superc!=subc => false 3253 _prim_same_path, // {P,P} & superc==subc => true 3254 _prim_1_path, // {N,P} => false 3255 _ref_subtype_path, // {N,N} & subtype check wins => true 3256 _both_ref_path, // {N,N} & subtype check loses => false 3257 PATH_LIMIT 3258 }; 3259 3260 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3261 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 3262 record_for_igvn(region); 3263 3264 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3265 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3266 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3267 3268 // First null-check both mirrors and load each mirror's klass metaobject. 3269 int which_arg; 3270 for (which_arg = 0; which_arg <= 1; which_arg++) { 3271 Node* arg = args[which_arg]; 3272 _sp += nargs; // set original stack for use by uncommon_trap 3273 arg = do_null_check(arg, T_OBJECT); 3274 _sp -= nargs; 3275 if (stopped()) break; 3276 args[which_arg] = _gvn.transform(arg); 3277 3278 Node* p = basic_plus_adr(arg, class_klass_offset); 3279 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3280 klasses[which_arg] = _gvn.transform(kls); 3281 } 3282 3283 // Having loaded both klasses, test each for null. 3284 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3285 for (which_arg = 0; which_arg <= 1; which_arg++) { 3286 Node* kls = klasses[which_arg]; 3287 Node* null_ctl = top(); 3288 _sp += nargs; // set original stack for use by uncommon_trap 3289 kls = null_check_oop(kls, &null_ctl, never_see_null); 3290 _sp -= nargs; 3291 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3292 region->init_req(prim_path, null_ctl); 3293 if (stopped()) break; 3294 klasses[which_arg] = kls; 3295 } 3296 3297 if (!stopped()) { 3298 // now we have two reference types, in klasses[0..1] 3299 Node* subk = klasses[1]; // the argument to isAssignableFrom 3300 Node* superk = klasses[0]; // the receiver 3301 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3302 // now we have a successful reference subtype check 3303 region->set_req(_ref_subtype_path, control()); 3304 } 3305 3306 // If both operands are primitive (both klasses null), then 3307 // we must return true when they are identical primitives. 3308 // It is convenient to test this after the first null klass check. 3309 set_control(region->in(_prim_0_path)); // go back to first null check 3310 if (!stopped()) { 3311 // Since superc is primitive, make a guard for the superc==subc case. 3312 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); 3313 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); 3314 generate_guard(bol_eq, region, PROB_FAIR); 3315 if (region->req() == PATH_LIMIT+1) { 3316 // A guard was added. If the added guard is taken, superc==subc. 3317 region->swap_edges(PATH_LIMIT, _prim_same_path); 3318 region->del_req(PATH_LIMIT); 3319 } 3320 region->set_req(_prim_0_path, control()); // Not equal after all. 3321 } 3322 3323 // these are the only paths that produce 'true': 3324 phi->set_req(_prim_same_path, intcon(1)); 3325 phi->set_req(_ref_subtype_path, intcon(1)); 3326 3327 // pull together the cases: 3328 assert(region->req() == PATH_LIMIT, "sane region"); 3329 for (uint i = 1; i < region->req(); i++) { 3330 Node* ctl = region->in(i); 3331 if (ctl == NULL || ctl == top()) { 3332 region->set_req(i, top()); 3333 phi ->set_req(i, top()); 3334 } else if (phi->in(i) == NULL) { 3335 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3336 } 3337 } 3338 3339 set_control(_gvn.transform(region)); 3340 push(_gvn.transform(phi)); 3341 3342 return true; 3343 } 3344 3345 //---------------------generate_array_guard_common------------------------ 3346 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3347 bool obj_array, bool not_array) { 3348 // If obj_array/non_array==false/false: 3349 // Branch around if the given klass is in fact an array (either obj or prim). 3350 // If obj_array/non_array==false/true: 3351 // Branch around if the given klass is not an array klass of any kind. 3352 // If obj_array/non_array==true/true: 3353 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3354 // If obj_array/non_array==true/false: 3355 // Branch around if the kls is an oop array (Object[] or subtype) 3356 // 3357 // Like generate_guard, adds a new path onto the region. 3358 jint layout_con = 0; 3359 Node* layout_val = get_layout_helper(kls, layout_con); 3360 if (layout_val == NULL) { 3361 bool query = (obj_array 3362 ? Klass::layout_helper_is_objArray(layout_con) 3363 : Klass::layout_helper_is_javaArray(layout_con)); 3364 if (query == not_array) { 3365 return NULL; // never a branch 3366 } else { // always a branch 3367 Node* always_branch = control(); 3368 if (region != NULL) 3369 region->add_req(always_branch); 3370 set_control(top()); 3371 return always_branch; 3372 } 3373 } 3374 // Now test the correct condition. 3375 jint nval = (obj_array 3376 ? ((jint)Klass::_lh_array_tag_type_value 3377 << Klass::_lh_array_tag_shift) 3378 : Klass::_lh_neutral_value); 3379 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); 3380 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3381 // invert the test if we are looking for a non-array 3382 if (not_array) btest = BoolTest(btest).negate(); 3383 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); 3384 return generate_fair_guard(bol, region); 3385 } 3386 3387 3388 //-----------------------inline_native_newArray-------------------------- 3389 bool LibraryCallKit::inline_native_newArray() { 3390 int nargs = 2; 3391 Node* mirror = argument(0); 3392 Node* count_val = argument(1); 3393 3394 _sp += nargs; // set original stack for use by uncommon_trap 3395 mirror = do_null_check(mirror, T_OBJECT); 3396 _sp -= nargs; 3397 // If mirror or obj is dead, only null-path is taken. 3398 if (stopped()) return true; 3399 3400 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3401 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3402 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3403 TypeInstPtr::NOTNULL); 3404 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3405 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3406 TypePtr::BOTTOM); 3407 3408 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3409 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3410 nargs, 3411 result_reg, _slow_path); 3412 Node* normal_ctl = control(); 3413 Node* no_array_ctl = result_reg->in(_slow_path); 3414 3415 // Generate code for the slow case. We make a call to newArray(). 3416 set_control(no_array_ctl); 3417 if (!stopped()) { 3418 // Either the input type is void.class, or else the 3419 // array klass has not yet been cached. Either the 3420 // ensuing call will throw an exception, or else it 3421 // will cache the array klass for next time. 3422 PreserveJVMState pjvms(this); 3423 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3424 Node* slow_result = set_results_for_java_call(slow_call); 3425 // this->control() comes from set_results_for_java_call 3426 result_reg->set_req(_slow_path, control()); 3427 result_val->set_req(_slow_path, slow_result); 3428 result_io ->set_req(_slow_path, i_o()); 3429 result_mem->set_req(_slow_path, reset_memory()); 3430 } 3431 3432 set_control(normal_ctl); 3433 if (!stopped()) { 3434 // Normal case: The array type has been cached in the java.lang.Class. 3435 // The following call works fine even if the array type is polymorphic. 3436 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3437 Node* obj = new_array(klass_node, count_val, nargs); 3438 result_reg->init_req(_normal_path, control()); 3439 result_val->init_req(_normal_path, obj); 3440 result_io ->init_req(_normal_path, i_o()); 3441 result_mem->init_req(_normal_path, reset_memory()); 3442 } 3443 3444 // Return the combined state. 3445 set_i_o( _gvn.transform(result_io) ); 3446 set_all_memory( _gvn.transform(result_mem) ); 3447 push_result(result_reg, result_val); 3448 C->set_has_split_ifs(true); // Has chance for split-if optimization 3449 3450 return true; 3451 } 3452 3453 //----------------------inline_native_getLength-------------------------- 3454 bool LibraryCallKit::inline_native_getLength() { 3455 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3456 3457 int nargs = 1; 3458 Node* array = argument(0); 3459 3460 _sp += nargs; // set original stack for use by uncommon_trap 3461 array = do_null_check(array, T_OBJECT); 3462 _sp -= nargs; 3463 3464 // If array is dead, only null-path is taken. 3465 if (stopped()) return true; 3466 3467 // Deoptimize if it is a non-array. 3468 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3469 3470 if (non_array != NULL) { 3471 PreserveJVMState pjvms(this); 3472 set_control(non_array); 3473 _sp += nargs; // push the arguments back on the stack 3474 uncommon_trap(Deoptimization::Reason_intrinsic, 3475 Deoptimization::Action_maybe_recompile); 3476 } 3477 3478 // If control is dead, only non-array-path is taken. 3479 if (stopped()) return true; 3480 3481 // The works fine even if the array type is polymorphic. 3482 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3483 push( load_array_length(array) ); 3484 3485 C->set_has_split_ifs(true); // Has chance for split-if optimization 3486 3487 return true; 3488 } 3489 3490 //------------------------inline_array_copyOf---------------------------- 3491 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3492 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3493 3494 // Restore the stack and pop off the arguments. 3495 int nargs = 3 + (is_copyOfRange? 1: 0); 3496 Node* original = argument(0); 3497 Node* start = is_copyOfRange? argument(1): intcon(0); 3498 Node* end = is_copyOfRange? argument(2): argument(1); 3499 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3500 3501 Node* newcopy; 3502 3503 //set the original stack and the reexecute bit for the interpreter to reexecute 3504 //the bytecode that invokes Arrays.copyOf if deoptimization happens 3505 { PreserveReexecuteState preexecs(this); 3506 _sp += nargs; 3507 jvms()->set_should_reexecute(true); 3508 3509 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); 3510 original = do_null_check(original, T_OBJECT); 3511 3512 // Check if a null path was taken unconditionally. 3513 if (stopped()) return true; 3514 3515 Node* orig_length = load_array_length(original); 3516 3517 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0, 3518 NULL, 0); 3519 klass_node = do_null_check(klass_node, T_OBJECT); 3520 3521 RegionNode* bailout = new (C, 1) RegionNode(1); 3522 record_for_igvn(bailout); 3523 3524 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3525 // Bail out if that is so. 3526 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3527 if (not_objArray != NULL) { 3528 // Improve the klass node's type from the new optimistic assumption: 3529 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3530 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3531 Node* cast = new (C, 2) CastPPNode(klass_node, akls); 3532 cast->init_req(0, control()); 3533 klass_node = _gvn.transform(cast); 3534 } 3535 3536 // Bail out if either start or end is negative. 3537 generate_negative_guard(start, bailout, &start); 3538 generate_negative_guard(end, bailout, &end); 3539 3540 Node* length = end; 3541 if (_gvn.type(start) != TypeInt::ZERO) { 3542 length = _gvn.transform( new (C, 3) SubINode(end, start) ); 3543 } 3544 3545 // Bail out if length is negative. 3546 // ...Not needed, since the new_array will throw the right exception. 3547 //generate_negative_guard(length, bailout, &length); 3548 3549 if (bailout->req() > 1) { 3550 PreserveJVMState pjvms(this); 3551 set_control( _gvn.transform(bailout) ); 3552 uncommon_trap(Deoptimization::Reason_intrinsic, 3553 Deoptimization::Action_maybe_recompile); 3554 } 3555 3556 if (!stopped()) { 3557 3558 // How many elements will we copy from the original? 3559 // The answer is MinI(orig_length - start, length). 3560 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); 3561 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3562 3563 newcopy = new_array(klass_node, length, 0); 3564 3565 // Generate a direct call to the right arraycopy function(s). 3566 // We know the copy is disjoint but we might not know if the 3567 // oop stores need checking. 3568 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3569 // This will fail a store-check if x contains any non-nulls. 3570 3571 Node* alloc = tightly_coupled_allocation(newcopy, NULL); 3572 3573 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, alloc != NULL); 3574 ac->set_copyof(); 3575 Node* n = _gvn.transform(ac); 3576 assert(n == ac, "cannot disappear"); 3577 ac->connect_outputs(this); 3578 } 3579 } //original reexecute and sp are set back here 3580 3581 if(!stopped()) { 3582 push(newcopy); 3583 } 3584 3585 C->set_has_split_ifs(true); // Has chance for split-if optimization 3586 3587 return true; 3588 } 3589 3590 3591 //----------------------generate_virtual_guard--------------------------- 3592 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3593 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3594 RegionNode* slow_region) { 3595 ciMethod* method = callee(); 3596 int vtable_index = method->vtable_index(); 3597 // Get the methodOop out of the appropriate vtable entry. 3598 int entry_offset = (instanceKlass::vtable_start_offset() + 3599 vtable_index*vtableEntry::size()) * wordSize + 3600 vtableEntry::method_offset_in_bytes(); 3601 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3602 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); 3603 3604 // Compare the target method with the expected method (e.g., Object.hashCode). 3605 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method); 3606 3607 Node* native_call = makecon(native_call_addr); 3608 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); 3609 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); 3610 3611 return generate_slow_guard(test_native, slow_region); 3612 } 3613 3614 //-----------------------generate_method_call---------------------------- 3615 // Use generate_method_call to make a slow-call to the real 3616 // method if the fast path fails. An alternative would be to 3617 // use a stub like OptoRuntime::slow_arraycopy_Java. 3618 // This only works for expanding the current library call, 3619 // not another intrinsic. (E.g., don't use this for making an 3620 // arraycopy call inside of the copyOf intrinsic.) 3621 CallJavaNode* 3622 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3623 // When compiling the intrinsic method itself, do not use this technique. 3624 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3625 3626 ciMethod* method = callee(); 3627 // ensure the JVMS we have will be correct for this call 3628 guarantee(method_id == method->intrinsic_id(), "must match"); 3629 3630 const TypeFunc* tf = TypeFunc::make(method); 3631 int tfdc = tf->domain()->cnt(); 3632 CallJavaNode* slow_call; 3633 if (is_static) { 3634 assert(!is_virtual, ""); 3635 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3636 SharedRuntime::get_resolve_static_call_stub(), 3637 method, bci()); 3638 } else if (is_virtual) { 3639 null_check_receiver(method); 3640 int vtable_index = methodOopDesc::invalid_vtable_index; 3641 if (UseInlineCaches) { 3642 // Suppress the vtable call 3643 } else { 3644 // hashCode and clone are not a miranda methods, 3645 // so the vtable index is fixed. 3646 // No need to use the linkResolver to get it. 3647 vtable_index = method->vtable_index(); 3648 } 3649 slow_call = new(C, tfdc) CallDynamicJavaNode(tf, 3650 SharedRuntime::get_resolve_virtual_call_stub(), 3651 method, vtable_index, bci()); 3652 } else { // neither virtual nor static: opt_virtual 3653 null_check_receiver(method); 3654 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3655 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3656 method, bci()); 3657 slow_call->set_optimized_virtual(true); 3658 } 3659 set_arguments_for_java_call(slow_call); 3660 set_edges_for_java_call(slow_call); 3661 return slow_call; 3662 } 3663 3664 3665 //------------------------------inline_native_hashcode-------------------- 3666 // Build special case code for calls to hashCode on an object. 3667 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3668 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3669 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3670 3671 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3672 3673 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3674 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3675 TypeInt::INT); 3676 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3677 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3678 TypePtr::BOTTOM); 3679 Node* obj = NULL; 3680 if (!is_static) { 3681 // Check for hashing null object 3682 obj = null_check_receiver(callee()); 3683 if (stopped()) return true; // unconditionally null 3684 result_reg->init_req(_null_path, top()); 3685 result_val->init_req(_null_path, top()); 3686 } else { 3687 // Do a null check, and return zero if null. 3688 // System.identityHashCode(null) == 0 3689 obj = argument(0); 3690 Node* null_ctl = top(); 3691 obj = null_check_oop(obj, &null_ctl); 3692 result_reg->init_req(_null_path, null_ctl); 3693 result_val->init_req(_null_path, _gvn.intcon(0)); 3694 } 3695 3696 // Unconditionally null? Then return right away. 3697 if (stopped()) { 3698 set_control( result_reg->in(_null_path) ); 3699 if (!stopped()) 3700 push( result_val ->in(_null_path) ); 3701 return true; 3702 } 3703 3704 // After null check, get the object's klass. 3705 Node* obj_klass = load_object_klass(obj); 3706 3707 // This call may be virtual (invokevirtual) or bound (invokespecial). 3708 // For each case we generate slightly different code. 3709 3710 // We only go to the fast case code if we pass a number of guards. The 3711 // paths which do not pass are accumulated in the slow_region. 3712 RegionNode* slow_region = new (C, 1) RegionNode(1); 3713 record_for_igvn(slow_region); 3714 3715 // If this is a virtual call, we generate a funny guard. We pull out 3716 // the vtable entry corresponding to hashCode() from the target object. 3717 // If the target method which we are calling happens to be the native 3718 // Object hashCode() method, we pass the guard. We do not need this 3719 // guard for non-virtual calls -- the caller is known to be the native 3720 // Object hashCode(). 3721 if (is_virtual) { 3722 generate_virtual_guard(obj_klass, slow_region); 3723 } 3724 3725 // Get the header out of the object, use LoadMarkNode when available 3726 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3727 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type()); 3728 3729 // Test the header to see if it is unlocked. 3730 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3731 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); 3732 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3733 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); 3734 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); 3735 3736 generate_slow_guard(test_unlocked, slow_region); 3737 3738 // Get the hash value and check to see that it has been properly assigned. 3739 // We depend on hash_mask being at most 32 bits and avoid the use of 3740 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3741 // vm: see markOop.hpp. 3742 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3743 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3744 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); 3745 // This hack lets the hash bits live anywhere in the mark object now, as long 3746 // as the shift drops the relevant bits into the low 32 bits. Note that 3747 // Java spec says that HashCode is an int so there's no point in capturing 3748 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3749 hshifted_header = ConvX2I(hshifted_header); 3750 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); 3751 3752 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3753 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); 3754 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); 3755 3756 generate_slow_guard(test_assigned, slow_region); 3757 3758 Node* init_mem = reset_memory(); 3759 // fill in the rest of the null path: 3760 result_io ->init_req(_null_path, i_o()); 3761 result_mem->init_req(_null_path, init_mem); 3762 3763 result_val->init_req(_fast_path, hash_val); 3764 result_reg->init_req(_fast_path, control()); 3765 result_io ->init_req(_fast_path, i_o()); 3766 result_mem->init_req(_fast_path, init_mem); 3767 3768 // Generate code for the slow case. We make a call to hashCode(). 3769 set_control(_gvn.transform(slow_region)); 3770 if (!stopped()) { 3771 // No need for PreserveJVMState, because we're using up the present state. 3772 set_all_memory(init_mem); 3773 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; 3774 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; 3775 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3776 Node* slow_result = set_results_for_java_call(slow_call); 3777 // this->control() comes from set_results_for_java_call 3778 result_reg->init_req(_slow_path, control()); 3779 result_val->init_req(_slow_path, slow_result); 3780 result_io ->set_req(_slow_path, i_o()); 3781 result_mem ->set_req(_slow_path, reset_memory()); 3782 } 3783 3784 // Return the combined state. 3785 set_i_o( _gvn.transform(result_io) ); 3786 set_all_memory( _gvn.transform(result_mem) ); 3787 push_result(result_reg, result_val); 3788 3789 return true; 3790 } 3791 3792 //---------------------------inline_native_getClass---------------------------- 3793 // Build special case code for calls to getClass on an object. 3794 bool LibraryCallKit::inline_native_getClass() { 3795 Node* obj = null_check_receiver(callee()); 3796 if (stopped()) return true; 3797 push( load_mirror_from_klass(load_object_klass(obj)) ); 3798 return true; 3799 } 3800 3801 //-----------------inline_native_Reflection_getCallerClass--------------------- 3802 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3803 // 3804 // NOTE that this code must perform the same logic as 3805 // vframeStream::security_get_caller_frame in that it must skip 3806 // Method.invoke() and auxiliary frames. 3807 3808 3809 3810 3811 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3812 ciMethod* method = callee(); 3813 3814 #ifndef PRODUCT 3815 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3816 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3817 } 3818 #endif 3819 3820 debug_only(int saved_sp = _sp); 3821 3822 // Argument words: (int depth) 3823 int nargs = 1; 3824 3825 _sp += nargs; 3826 Node* caller_depth_node = pop(); 3827 3828 assert(saved_sp == _sp, "must have correct argument count"); 3829 3830 // The depth value must be a constant in order for the runtime call 3831 // to be eliminated. 3832 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); 3833 if (caller_depth_type == NULL || !caller_depth_type->is_con()) { 3834 #ifndef PRODUCT 3835 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3836 tty->print_cr(" Bailing out because caller depth was not a constant"); 3837 } 3838 #endif 3839 return false; 3840 } 3841 // Note that the JVM state at this point does not include the 3842 // getCallerClass() frame which we are trying to inline. The 3843 // semantics of getCallerClass(), however, are that the "first" 3844 // frame is the getCallerClass() frame, so we subtract one from the 3845 // requested depth before continuing. We don't inline requests of 3846 // getCallerClass(0). 3847 int caller_depth = caller_depth_type->get_con() - 1; 3848 if (caller_depth < 0) { 3849 #ifndef PRODUCT 3850 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3851 tty->print_cr(" Bailing out because caller depth was %d", caller_depth); 3852 } 3853 #endif 3854 return false; 3855 } 3856 3857 if (!jvms()->has_method()) { 3858 #ifndef PRODUCT 3859 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3860 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3861 } 3862 #endif 3863 return false; 3864 } 3865 int _depth = jvms()->depth(); // cache call chain depth 3866 3867 // Walk back up the JVM state to find the caller at the required 3868 // depth. NOTE that this code must perform the same logic as 3869 // vframeStream::security_get_caller_frame in that it must skip 3870 // Method.invoke() and auxiliary frames. Note also that depth is 3871 // 1-based (1 is the bottom of the inlining). 3872 int inlining_depth = _depth; 3873 JVMState* caller_jvms = NULL; 3874 3875 if (inlining_depth > 0) { 3876 caller_jvms = jvms(); 3877 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); 3878 do { 3879 // The following if-tests should be performed in this order 3880 if (is_method_invoke_or_aux_frame(caller_jvms)) { 3881 // Skip a Method.invoke() or auxiliary frame 3882 } else if (caller_depth > 0) { 3883 // Skip real frame 3884 --caller_depth; 3885 } else { 3886 // We're done: reached desired caller after skipping. 3887 break; 3888 } 3889 caller_jvms = caller_jvms->caller(); 3890 --inlining_depth; 3891 } while (inlining_depth > 0); 3892 } 3893 3894 if (inlining_depth == 0) { 3895 #ifndef PRODUCT 3896 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3897 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); 3898 tty->print_cr(" JVM state at this point:"); 3899 for (int i = _depth; i >= 1; i--) { 3900 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3901 } 3902 } 3903 #endif 3904 return false; // Reached end of inlining 3905 } 3906 3907 // Acquire method holder as java.lang.Class 3908 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 3909 ciInstance* caller_mirror = caller_klass->java_mirror(); 3910 // Push this as a constant 3911 push(makecon(TypeInstPtr::make(caller_mirror))); 3912 #ifndef PRODUCT 3913 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3914 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); 3915 tty->print_cr(" JVM state at this point:"); 3916 for (int i = _depth; i >= 1; i--) { 3917 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 3918 } 3919 } 3920 #endif 3921 return true; 3922 } 3923 3924 // Helper routine for above 3925 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { 3926 ciMethod* method = jvms->method(); 3927 3928 // Is this the Method.invoke method itself? 3929 if (method->intrinsic_id() == vmIntrinsics::_invoke) 3930 return true; 3931 3932 // Is this a helper, defined somewhere underneath MethodAccessorImpl. 3933 ciKlass* k = method->holder(); 3934 if (k->is_instance_klass()) { 3935 ciInstanceKlass* ik = k->as_instance_klass(); 3936 for (; ik != NULL; ik = ik->super()) { 3937 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && 3938 ik == env()->find_system_klass(ik->name())) { 3939 return true; 3940 } 3941 } 3942 } 3943 else if (method->is_method_handle_adapter()) { 3944 // This is an internal adapter frame from the MethodHandleCompiler -- skip it 3945 return true; 3946 } 3947 3948 return false; 3949 } 3950 3951 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by 3952 // inline_native_AtomicLong_attemptUpdate() but it has no way of 3953 // computing it since there is no lookup field by name function in the 3954 // CI interface. This is computed and set by inline_native_AtomicLong_get(). 3955 // Using a static variable here is safe even if we have multiple compilation 3956 // threads because the offset is constant. At worst the same offset will be 3957 // computed and stored multiple 3958 3959 bool LibraryCallKit::inline_native_AtomicLong_get() { 3960 // Restore the stack and pop off the argument 3961 _sp+=1; 3962 Node *obj = pop(); 3963 3964 // get the offset of the "value" field. Since the CI interfaces 3965 // does not provide a way to look up a field by name, we scan the bytecodes 3966 // to get the field index. We expect the first 2 instructions of the method 3967 // to be: 3968 // 0 aload_0 3969 // 1 getfield "value" 3970 ciMethod* method = callee(); 3971 if (value_field_offset == -1) 3972 { 3973 ciField* value_field; 3974 ciBytecodeStream iter(method); 3975 Bytecodes::Code bc = iter.next(); 3976 3977 if ((bc != Bytecodes::_aload_0) && 3978 ((bc != Bytecodes::_aload) || (iter.get_index() != 0))) 3979 return false; 3980 bc = iter.next(); 3981 if (bc != Bytecodes::_getfield) 3982 return false; 3983 bool ignore; 3984 value_field = iter.get_field(ignore); 3985 value_field_offset = value_field->offset_in_bytes(); 3986 } 3987 3988 // Null check without removing any arguments. 3989 _sp++; 3990 obj = do_null_check(obj, T_OBJECT); 3991 _sp--; 3992 // Check for locking null object 3993 if (stopped()) return true; 3994 3995 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 3996 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 3997 int alias_idx = C->get_alias_index(adr_type); 3998 3999 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr)); 4000 4001 push_pair(result); 4002 4003 return true; 4004 } 4005 4006 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() { 4007 // Restore the stack and pop off the arguments 4008 _sp+=5; 4009 Node *newVal = pop_pair(); 4010 Node *oldVal = pop_pair(); 4011 Node *obj = pop(); 4012 4013 // we need the offset of the "value" field which was computed when 4014 // inlining the get() method. Give up if we don't have it. 4015 if (value_field_offset == -1) 4016 return false; 4017 4018 // Null check without removing any arguments. 4019 _sp+=5; 4020 obj = do_null_check(obj, T_OBJECT); 4021 _sp-=5; 4022 // Check for locking null object 4023 if (stopped()) return true; 4024 4025 Node *adr = basic_plus_adr(obj, obj, value_field_offset); 4026 const TypePtr *adr_type = _gvn.type(adr)->is_ptr(); 4027 int alias_idx = C->get_alias_index(adr_type); 4028 4029 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal)); 4030 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 4031 set_memory(store_proj, alias_idx); 4032 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) ); 4033 4034 Node *result; 4035 // CMove node is not used to be able fold a possible check code 4036 // after attemptUpdate() call. This code could be transformed 4037 // into CMove node by loop optimizations. 4038 { 4039 RegionNode *r = new (C, 3) RegionNode(3); 4040 result = new (C, 3) PhiNode(r, TypeInt::BOOL); 4041 4042 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 4043 Node *iftrue = opt_iff(r, iff); 4044 r->init_req(1, iftrue); 4045 result->init_req(1, intcon(1)); 4046 result->init_req(2, intcon(0)); 4047 4048 set_control(_gvn.transform(r)); 4049 record_for_igvn(r); 4050 4051 C->set_has_split_ifs(true); // Has chance for split-if optimization 4052 } 4053 4054 push(_gvn.transform(result)); 4055 return true; 4056 } 4057 4058 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4059 // restore the arguments 4060 _sp += arg_size(); 4061 4062 switch (id) { 4063 case vmIntrinsics::_floatToRawIntBits: 4064 push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); 4065 break; 4066 4067 case vmIntrinsics::_intBitsToFloat: 4068 push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); 4069 break; 4070 4071 case vmIntrinsics::_doubleToRawLongBits: 4072 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); 4073 break; 4074 4075 case vmIntrinsics::_longBitsToDouble: 4076 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); 4077 break; 4078 4079 case vmIntrinsics::_doubleToLongBits: { 4080 Node* value = pop_pair(); 4081 4082 // two paths (plus control) merge in a wood 4083 RegionNode *r = new (C, 3) RegionNode(3); 4084 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); 4085 4086 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); 4087 // Build the boolean node 4088 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 4089 4090 // Branch either way. 4091 // NaN case is less traveled, which makes all the difference. 4092 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4093 Node *opt_isnan = _gvn.transform(ifisnan); 4094 assert( opt_isnan->is_If(), "Expect an IfNode"); 4095 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4096 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 4097 4098 set_control(iftrue); 4099 4100 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4101 Node *slow_result = longcon(nan_bits); // return NaN 4102 phi->init_req(1, _gvn.transform( slow_result )); 4103 r->init_req(1, iftrue); 4104 4105 // Else fall through 4106 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 4107 set_control(iffalse); 4108 4109 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); 4110 r->init_req(2, iffalse); 4111 4112 // Post merge 4113 set_control(_gvn.transform(r)); 4114 record_for_igvn(r); 4115 4116 Node* result = _gvn.transform(phi); 4117 assert(result->bottom_type()->isa_long(), "must be"); 4118 push_pair(result); 4119 4120 C->set_has_split_ifs(true); // Has chance for split-if optimization 4121 4122 break; 4123 } 4124 4125 case vmIntrinsics::_floatToIntBits: { 4126 Node* value = pop(); 4127 4128 // two paths (plus control) merge in a wood 4129 RegionNode *r = new (C, 3) RegionNode(3); 4130 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); 4131 4132 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); 4133 // Build the boolean node 4134 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 4135 4136 // Branch either way. 4137 // NaN case is less traveled, which makes all the difference. 4138 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4139 Node *opt_isnan = _gvn.transform(ifisnan); 4140 assert( opt_isnan->is_If(), "Expect an IfNode"); 4141 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4142 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 4143 4144 set_control(iftrue); 4145 4146 static const jint nan_bits = 0x7fc00000; 4147 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4148 phi->init_req(1, _gvn.transform( slow_result )); 4149 r->init_req(1, iftrue); 4150 4151 // Else fall through 4152 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 4153 set_control(iffalse); 4154 4155 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); 4156 r->init_req(2, iffalse); 4157 4158 // Post merge 4159 set_control(_gvn.transform(r)); 4160 record_for_igvn(r); 4161 4162 Node* result = _gvn.transform(phi); 4163 assert(result->bottom_type()->isa_int(), "must be"); 4164 push(result); 4165 4166 C->set_has_split_ifs(true); // Has chance for split-if optimization 4167 4168 break; 4169 } 4170 4171 default: 4172 ShouldNotReachHere(); 4173 } 4174 4175 return true; 4176 } 4177 4178 #ifdef _LP64 4179 #define XTOP ,top() /*additional argument*/ 4180 #else //_LP64 4181 #define XTOP /*no additional argument*/ 4182 #endif //_LP64 4183 4184 //----------------------inline_unsafe_copyMemory------------------------- 4185 bool LibraryCallKit::inline_unsafe_copyMemory() { 4186 if (callee()->is_static()) return false; // caller must have the capability! 4187 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) 4188 assert(signature()->size() == nargs-1, "copy has 5 arguments"); 4189 null_check_receiver(callee()); // check then ignore argument(0) 4190 if (stopped()) return true; 4191 4192 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4193 4194 Node* src_ptr = argument(1); 4195 Node* src_off = ConvL2X(argument(2)); 4196 assert(argument(3)->is_top(), "2nd half of long"); 4197 Node* dst_ptr = argument(4); 4198 Node* dst_off = ConvL2X(argument(5)); 4199 assert(argument(6)->is_top(), "2nd half of long"); 4200 Node* size = ConvL2X(argument(7)); 4201 assert(argument(8)->is_top(), "2nd half of long"); 4202 4203 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4204 "fieldOffset must be byte-scaled"); 4205 4206 Node* src = make_unsafe_address(src_ptr, src_off); 4207 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4208 4209 // Conservatively insert a memory barrier on all memory slices. 4210 // Do not let writes of the copy source or destination float below the copy. 4211 insert_mem_bar(Op_MemBarCPUOrder); 4212 4213 // Call it. Note that the length argument is not scaled. 4214 make_runtime_call(RC_LEAF|RC_NO_FP, 4215 OptoRuntime::fast_arraycopy_Type(), 4216 StubRoutines::unsafe_arraycopy(), 4217 "unsafe_arraycopy", 4218 TypeRawPtr::BOTTOM, 4219 src, dst, size XTOP); 4220 4221 // Do not let reads of the copy destination float above the copy. 4222 insert_mem_bar(Op_MemBarCPUOrder); 4223 4224 return true; 4225 } 4226 4227 //------------------------clone_coping----------------------------------- 4228 // Helper function for inline_native_clone. 4229 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4230 assert(obj_size != NULL, ""); 4231 Node* raw_obj = alloc_obj->in(1); 4232 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4233 4234 AllocateNode* alloc = NULL; 4235 if (ReduceBulkZeroing) { 4236 // We will be completely responsible for initializing this object - 4237 // mark Initialize node as complete. 4238 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4239 // The object was just allocated - there should be no any stores! 4240 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4241 // Mark as complete_with_arraycopy so that on AllocateNode 4242 // expansion, we know this AllocateNode is initialized by an array 4243 // copy and a StoreStore barrier exists after the array copy. 4244 alloc->initialization()->set_complete_with_arraycopy(); 4245 } 4246 4247 // Copy the fastest available way. 4248 // TODO: generate fields copies for small objects instead. 4249 Node* src = obj; 4250 Node* dest = alloc_obj; 4251 Node* size = _gvn.transform(obj_size); 4252 4253 // Exclude the header but include array length to copy by 8 bytes words. 4254 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4255 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4256 instanceOopDesc::base_offset_in_bytes(); 4257 // base_off: 4258 // 8 - 32-bit VM 4259 // 12 - 64-bit VM, compressed oops 4260 // 16 - 64-bit VM, normal oops 4261 if (base_off % BytesPerLong != 0) { 4262 assert(UseCompressedOops, ""); 4263 if (is_array) { 4264 // Exclude length to copy by 8 bytes words. 4265 base_off += sizeof(int); 4266 } else { 4267 // Include klass to copy by 8 bytes words. 4268 base_off = instanceOopDesc::klass_offset_in_bytes(); 4269 } 4270 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4271 } 4272 src = basic_plus_adr(src, base_off); 4273 dest = basic_plus_adr(dest, base_off); 4274 4275 // Compute the length also, if needed: 4276 Node* countx = size; 4277 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); 4278 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4279 4280 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4281 4282 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4283 ac->set_clonebasic(); 4284 Node* n = _gvn.transform(ac); 4285 assert(n == ac, "cannot disappear"); 4286 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4287 4288 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4289 if (card_mark) { 4290 assert(!is_array, ""); 4291 // Put in store barrier for any and all oops we are sticking 4292 // into this object. (We could avoid this if we could prove 4293 // that the object type contains no oop fields at all.) 4294 Node* no_particular_value = NULL; 4295 Node* no_particular_field = NULL; 4296 int raw_adr_idx = Compile::AliasIdxRaw; 4297 post_barrier(control(), 4298 memory(raw_adr_type), 4299 alloc_obj, 4300 no_particular_field, 4301 raw_adr_idx, 4302 no_particular_value, 4303 T_OBJECT, 4304 false); 4305 } 4306 4307 // Do not let reads from the cloned object float above the arraycopy. 4308 if (alloc != NULL) { 4309 // Do not let stores that initialize this object be reordered with 4310 // a subsequent store that would make this object accessible by 4311 // other threads. 4312 // Record what AllocateNode this StoreStore protects so that 4313 // escape analysis can go from the MemBarStoreStoreNode to the 4314 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4315 // based on the escape status of the AllocateNode. 4316 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4317 } else { 4318 insert_mem_bar(Op_MemBarCPUOrder); 4319 } 4320 } 4321 4322 //------------------------inline_native_clone---------------------------- 4323 // Here are the simple edge cases: 4324 // null receiver => normal trap 4325 // virtual and clone was overridden => slow path to out-of-line clone 4326 // not cloneable or finalizer => slow path to out-of-line Object.clone 4327 // 4328 // The general case has two steps, allocation and copying. 4329 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4330 // 4331 // Copying also has two cases, oop arrays and everything else. 4332 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4333 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4334 // 4335 // These steps fold up nicely if and when the cloned object's klass 4336 // can be sharply typed as an object array, a type array, or an instance. 4337 // 4338 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4339 int nargs = 1; 4340 PhiNode* result_val; 4341 4342 //set the original stack and the reexecute bit for the interpreter to reexecute 4343 //the bytecode that invokes Object.clone if deoptimization happens 4344 { PreserveReexecuteState preexecs(this); 4345 jvms()->set_should_reexecute(true); 4346 4347 //null_check_receiver will adjust _sp (push and pop) 4348 Node* obj = null_check_receiver(callee()); 4349 if (stopped()) return true; 4350 4351 _sp += nargs; 4352 4353 Node* obj_klass = load_object_klass(obj); 4354 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4355 const TypeOopPtr* toop = ((tklass != NULL) 4356 ? tklass->as_instance_type() 4357 : TypeInstPtr::NOTNULL); 4358 4359 // Conservatively insert a memory barrier on all memory slices. 4360 // Do not let writes into the original float below the clone. 4361 insert_mem_bar(Op_MemBarCPUOrder); 4362 4363 // paths into result_reg: 4364 enum { 4365 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4366 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4367 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4368 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4369 PATH_LIMIT 4370 }; 4371 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4372 result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 4373 TypeInstPtr::NOTNULL); 4374 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 4375 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 4376 TypePtr::BOTTOM); 4377 record_for_igvn(result_reg); 4378 4379 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4380 int raw_adr_idx = Compile::AliasIdxRaw; 4381 4382 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4383 if (array_ctl != NULL) { 4384 // It's an array. 4385 PreserveJVMState pjvms(this); 4386 set_control(array_ctl); 4387 Node* obj_length = load_array_length(obj); 4388 Node* obj_size = NULL; 4389 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); 4390 4391 if (!use_ReduceInitialCardMarks()) { 4392 // If it is an oop array, it requires very special treatment, 4393 // because card marking is required on each card of the array. 4394 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4395 if (is_obja != NULL) { 4396 PreserveJVMState pjvms2(this); 4397 set_control(is_obja); 4398 // Generate a direct call to the right arraycopy function(s). 4399 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4400 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4401 ac->set_cloneoop(); 4402 Node* n = _gvn.transform(ac); 4403 assert(n == ac, "cannot disappear"); 4404 ac->connect_outputs(this); 4405 4406 result_reg->init_req(_objArray_path, control()); 4407 result_val->init_req(_objArray_path, alloc_obj); 4408 result_i_o ->set_req(_objArray_path, i_o()); 4409 result_mem ->set_req(_objArray_path, reset_memory()); 4410 } 4411 } 4412 // Otherwise, there are no card marks to worry about. 4413 // (We can dispense with card marks if we know the allocation 4414 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4415 // causes the non-eden paths to take compensating steps to 4416 // simulate a fresh allocation, so that no further 4417 // card marks are required in compiled code to initialize 4418 // the object.) 4419 4420 if (!stopped()) { 4421 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4422 4423 // Present the results of the copy. 4424 result_reg->init_req(_array_path, control()); 4425 result_val->init_req(_array_path, alloc_obj); 4426 result_i_o ->set_req(_array_path, i_o()); 4427 result_mem ->set_req(_array_path, reset_memory()); 4428 } 4429 } 4430 4431 // We only go to the instance fast case code if we pass a number of guards. 4432 // The paths which do not pass are accumulated in the slow_region. 4433 RegionNode* slow_region = new (C, 1) RegionNode(1); 4434 record_for_igvn(slow_region); 4435 if (!stopped()) { 4436 // It's an instance (we did array above). Make the slow-path tests. 4437 // If this is a virtual call, we generate a funny guard. We grab 4438 // the vtable entry corresponding to clone() from the target object. 4439 // If the target method which we are calling happens to be the 4440 // Object clone() method, we pass the guard. We do not need this 4441 // guard for non-virtual calls; the caller is known to be the native 4442 // Object clone(). 4443 if (is_virtual) { 4444 generate_virtual_guard(obj_klass, slow_region); 4445 } 4446 4447 // The object must be cloneable and must not have a finalizer. 4448 // Both of these conditions may be checked in a single test. 4449 // We could optimize the cloneable test further, but we don't care. 4450 generate_access_flags_guard(obj_klass, 4451 // Test both conditions: 4452 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4453 // Must be cloneable but not finalizer: 4454 JVM_ACC_IS_CLONEABLE, 4455 slow_region); 4456 } 4457 4458 if (!stopped()) { 4459 // It's an instance, and it passed the slow-path tests. 4460 PreserveJVMState pjvms(this); 4461 Node* obj_size = NULL; 4462 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size); 4463 4464 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4465 4466 // Present the results of the slow call. 4467 result_reg->init_req(_instance_path, control()); 4468 result_val->init_req(_instance_path, alloc_obj); 4469 result_i_o ->set_req(_instance_path, i_o()); 4470 result_mem ->set_req(_instance_path, reset_memory()); 4471 } 4472 4473 // Generate code for the slow case. We make a call to clone(). 4474 set_control(_gvn.transform(slow_region)); 4475 if (!stopped()) { 4476 PreserveJVMState pjvms(this); 4477 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4478 Node* slow_result = set_results_for_java_call(slow_call); 4479 // this->control() comes from set_results_for_java_call 4480 result_reg->init_req(_slow_path, control()); 4481 result_val->init_req(_slow_path, slow_result); 4482 result_i_o ->set_req(_slow_path, i_o()); 4483 result_mem ->set_req(_slow_path, reset_memory()); 4484 } 4485 4486 // Return the combined state. 4487 set_control( _gvn.transform(result_reg) ); 4488 set_i_o( _gvn.transform(result_i_o) ); 4489 set_all_memory( _gvn.transform(result_mem) ); 4490 } //original reexecute and sp are set back here 4491 4492 push(_gvn.transform(result_val)); 4493 4494 return true; 4495 } 4496 4497 //------------------------------basictype2arraycopy---------------------------- 4498 address LibraryCallKit::basictype2arraycopy(BasicType t, 4499 Node* src_offset, 4500 Node* dest_offset, 4501 bool disjoint_bases, 4502 const char* &name, 4503 bool dest_uninitialized) { 4504 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4505 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4506 4507 bool aligned = false; 4508 bool disjoint = disjoint_bases; 4509 4510 // if the offsets are the same, we can treat the memory regions as 4511 // disjoint, because either the memory regions are in different arrays, 4512 // or they are identical (which we can treat as disjoint.) We can also 4513 // treat a copy with a destination index less that the source index 4514 // as disjoint since a low->high copy will work correctly in this case. 4515 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4516 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4517 // both indices are constants 4518 int s_offs = src_offset_inttype->get_con(); 4519 int d_offs = dest_offset_inttype->get_con(); 4520 int element_size = type2aelembytes(t); 4521 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4522 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4523 if (s_offs >= d_offs) disjoint = true; 4524 } else if (src_offset == dest_offset && src_offset != NULL) { 4525 // This can occur if the offsets are identical non-constants. 4526 disjoint = true; 4527 } 4528 4529 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4530 } 4531 4532 4533 //------------------------------inline_arraycopy----------------------- 4534 bool LibraryCallKit::inline_arraycopy() { 4535 // Restore the stack and pop off the arguments. 4536 int nargs = 5; // 2 oops, 3 ints, no size_t or long 4537 assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); 4538 4539 Node *src = argument(0); 4540 Node *src_offset = argument(1); 4541 Node *dest = argument(2); 4542 Node *dest_offset = argument(3); 4543 Node *length = argument(4); 4544 4545 // The following tests must be performed 4546 // (1) src and dest are arrays. 4547 // (2) src and dest arrays must have elements of the same BasicType 4548 // (3) src and dest must not be null. 4549 // (4) src_offset must not be negative. 4550 // (5) dest_offset must not be negative. 4551 // (6) length must not be negative. 4552 // (7) src_offset + length must not exceed length of src. 4553 // (8) dest_offset + length must not exceed length of dest. 4554 // (9) each element of an oop array must be assignable 4555 4556 // (3) src and dest must not be null. 4557 // always do this here because we need the JVM state for uncommon traps 4558 _sp += nargs; 4559 src = do_null_check(src, T_ARRAY); 4560 dest = do_null_check(dest, T_ARRAY); 4561 _sp -= nargs; 4562 4563 bool notest = false; 4564 4565 if (!too_many_traps(Deoptimization::Reason_intrinsic) && !src->is_top() && !dest->is_top()) { 4566 // validate arguments: enables transformation the ArrayCopyNode 4567 notest = true; 4568 4569 RegionNode* slow_region = new (C, 1) RegionNode(1); 4570 record_for_igvn(slow_region); 4571 4572 // (1) src and dest are arrays. 4573 generate_non_array_guard(load_object_klass(src), slow_region); 4574 generate_non_array_guard(load_object_klass(dest), slow_region); 4575 4576 // (2) src and dest arrays must have elements of the same BasicType 4577 // done at macro expansion or at Ideal transformation time 4578 4579 // (4) src_offset must not be negative. 4580 generate_negative_guard(src_offset, slow_region); 4581 4582 // (5) dest_offset must not be negative. 4583 generate_negative_guard(dest_offset, slow_region); 4584 4585 // (7) src_offset + length must not exceed length of src. 4586 generate_limit_guard(src_offset, length, 4587 load_array_length(src), 4588 slow_region); 4589 4590 // (8) dest_offset + length must not exceed length of dest. 4591 generate_limit_guard(dest_offset, length, 4592 load_array_length(dest), 4593 slow_region); 4594 4595 // (9) each element of an oop array must be assignable 4596 Node* src_klass = load_object_klass(src); 4597 Node* dest_klass = load_object_klass(dest); 4598 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4599 4600 if (not_subtype_ctrl != top()) { 4601 PreserveJVMState pjvms(this); 4602 set_control(not_subtype_ctrl); 4603 _sp += nargs; 4604 uncommon_trap(Deoptimization::Reason_intrinsic, 4605 Deoptimization::Action_make_not_entrant); 4606 _sp -= nargs; 4607 assert(stopped(), "Should be stopped"); 4608 } 4609 { 4610 PreserveJVMState pjvms(this); 4611 set_control(_gvn.transform(slow_region)); 4612 _sp += nargs; 4613 uncommon_trap(Deoptimization::Reason_intrinsic, 4614 Deoptimization::Action_make_not_entrant); 4615 _sp -= nargs; 4616 assert(stopped(), "Should be stopped"); 4617 } 4618 } 4619 4620 if (stopped()) { 4621 return true; 4622 } 4623 4624 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4625 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL); 4626 4627 if (notest) { 4628 ac->set_arraycopy_notest(); 4629 } 4630 4631 Node* n = _gvn.transform(ac); 4632 assert(n == ac, "cannot disappear"); 4633 ac->connect_outputs(this); 4634 4635 return true; 4636 } 4637 4638 4639 // Helper function which determines if an arraycopy immediately follows 4640 // an allocation, with no intervening tests or other escapes for the object. 4641 AllocateArrayNode* 4642 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4643 RegionNode* slow_region) { 4644 if (stopped()) return NULL; // no fast path 4645 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4646 4647 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4648 if (alloc == NULL) return NULL; 4649 4650 Node* rawmem = memory(Compile::AliasIdxRaw); 4651 // Is the allocation's memory state untouched? 4652 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4653 // Bail out if there have been raw-memory effects since the allocation. 4654 // (Example: There might have been a call or safepoint.) 4655 return NULL; 4656 } 4657 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4658 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4659 return NULL; 4660 } 4661 4662 // There must be no unexpected observers of this allocation. 4663 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4664 Node* obs = ptr->fast_out(i); 4665 if (obs != this->map()) { 4666 return NULL; 4667 } 4668 } 4669 4670 // This arraycopy must unconditionally follow the allocation of the ptr. 4671 Node* alloc_ctl = ptr->in(0); 4672 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4673 4674 Node* ctl = control(); 4675 while (ctl != alloc_ctl) { 4676 // There may be guards which feed into the slow_region. 4677 // Any other control flow means that we might not get a chance 4678 // to finish initializing the allocated object. 4679 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4680 IfNode* iff = ctl->in(0)->as_If(); 4681 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 4682 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4683 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4684 ctl = iff->in(0); // This test feeds the known slow_region. 4685 continue; 4686 } 4687 // One more try: Various low-level checks bottom out in 4688 // uncommon traps. If the debug-info of the trap omits 4689 // any reference to the allocation, as we've already 4690 // observed, then there can be no objection to the trap. 4691 bool found_trap = false; 4692 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 4693 Node* obs = not_ctl->fast_out(j); 4694 if (obs->in(0) == not_ctl && obs->is_Call() && 4695 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 4696 found_trap = true; break; 4697 } 4698 } 4699 if (found_trap) { 4700 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 4701 continue; 4702 } 4703 } 4704 return NULL; 4705 } 4706 4707 // If we get this far, we have an allocation which immediately 4708 // precedes the arraycopy, and we can take over zeroing the new object. 4709 // The arraycopy will finish the initialization, and provide 4710 // a new control state to which we will anchor the destination pointer. 4711 4712 return alloc; 4713 } 4714 4715 //----------------------------inline_reference_get---------------------------- 4716 4717 bool LibraryCallKit::inline_reference_get() { 4718 const int nargs = 1; // self 4719 4720 guarantee(java_lang_ref_Reference::referent_offset > 0, 4721 "should have already been set"); 4722 4723 int referent_offset = java_lang_ref_Reference::referent_offset; 4724 4725 // Restore the stack and pop off the argument 4726 _sp += nargs; 4727 Node *reference_obj = pop(); 4728 4729 // Null check on self without removing any arguments. 4730 _sp += nargs; 4731 reference_obj = do_null_check(reference_obj, T_OBJECT); 4732 _sp -= nargs;; 4733 4734 if (stopped()) return true; 4735 4736 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 4737 4738 ciInstanceKlass* klass = env()->Object_klass(); 4739 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 4740 4741 Node* no_ctrl = NULL; 4742 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT); 4743 4744 // Use the pre-barrier to record the value in the referent field 4745 pre_barrier(false /* do_load */, 4746 control(), 4747 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 4748 result /* pre_val */, 4749 T_OBJECT); 4750 4751 push(result); 4752 return true; 4753 }