library_call.cpp revision 3602:da91efe96a93
11590Srgrimes/* 21590Srgrimes * Copyright (c) 1999, 2012, Oracle and/or its affiliates. All rights reserved. 31590Srgrimes * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 41590Srgrimes * 51590Srgrimes * This code is free software; you can redistribute it and/or modify it 61590Srgrimes * under the terms of the GNU General Public License version 2 only, as 71590Srgrimes * published by the Free Software Foundation. 81590Srgrimes * 91590Srgrimes * This code is distributed in the hope that it will be useful, but WITHOUT 101590Srgrimes * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 111590Srgrimes * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 121590Srgrimes * version 2 for more details (a copy is included in the LICENSE file that 131590Srgrimes * accompanied this code). 141590Srgrimes * 151590Srgrimes * You should have received a copy of the GNU General Public License version 161590Srgrimes * 2 along with this work; if not, write to the Free Software Foundation, 171590Srgrimes * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 181590Srgrimes * 191590Srgrimes * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 201590Srgrimes * or visit www.oracle.com if you need additional information or have any 211590Srgrimes * questions. 221590Srgrimes * 231590Srgrimes */ 241590Srgrimes 251590Srgrimes#include "precompiled.hpp" 261590Srgrimes#include "classfile/systemDictionary.hpp" 271590Srgrimes#include "classfile/vmSymbols.hpp" 281590Srgrimes#include "compiler/compileBroker.hpp" 291590Srgrimes#include "compiler/compileLog.hpp" 301590Srgrimes#include "oops/objArrayKlass.hpp" 311590Srgrimes#include "opto/addnode.hpp" 321590Srgrimes#include "opto/callGenerator.hpp" 331590Srgrimes#include "opto/cfgnode.hpp" 341590Srgrimes#include "opto/idealKit.hpp" 351590Srgrimes#include "opto/mulnode.hpp" 361590Srgrimes#include "opto/parse.hpp" 371590Srgrimes#include "opto/runtime.hpp" 381590Srgrimes#include "opto/subnode.hpp" 391590Srgrimes#include "prims/nativeLookup.hpp" 401590Srgrimes#include "runtime/sharedRuntime.hpp" 411590Srgrimes 421590Srgrimesclass LibraryIntrinsic : public InlineCallGenerator { 431590Srgrimes // Extend the set of intrinsics known to the runtime: 441590Srgrimes public: 451590Srgrimes private: 461590Srgrimes bool _is_virtual; 4711914Sphk vmIntrinsics::ID _intrinsic_id; 481590Srgrimes 491590Srgrimes public: 501590Srgrimes LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id) 511590Srgrimes : InlineCallGenerator(m), 521590Srgrimes _is_virtual(is_virtual), 531590Srgrimes _intrinsic_id(id) 541590Srgrimes { 551590Srgrimes } 561590Srgrimes virtual bool is_intrinsic() const { return true; } 571590Srgrimes virtual bool is_virtual() const { return _is_virtual; } 581590Srgrimes virtual JVMState* generate(JVMState* jvms); 591590Srgrimes vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 601590Srgrimes}; 611590Srgrimes 621590Srgrimes 631590Srgrimes// Local helper class for LibraryIntrinsic: 641590Srgrimesclass LibraryCallKit : public GraphKit { 651590Srgrimes private: 661590Srgrimes LibraryIntrinsic* _intrinsic; // the library intrinsic being called 671590Srgrimes 681590Srgrimes public: 691590Srgrimes LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic) 701590Srgrimes : GraphKit(caller), 711590Srgrimes _intrinsic(intrinsic) 721590Srgrimes { 731590Srgrimes } 741590Srgrimes 751590Srgrimes ciMethod* caller() const { return jvms()->method(); } 761590Srgrimes int bci() const { return jvms()->bci(); } 771590Srgrimes LibraryIntrinsic* intrinsic() const { return _intrinsic; } 781590Srgrimes vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 791590Srgrimes ciMethod* callee() const { return _intrinsic->method(); } 801590Srgrimes ciSignature* signature() const { return callee()->signature(); } 811590Srgrimes int arg_size() const { return callee()->arg_size(); } 821590Srgrimes 831590Srgrimes bool try_to_inline(); 841590Srgrimes 851590Srgrimes // Helper functions to inline natives 861590Srgrimes void push_result(RegionNode* region, PhiNode* value); 871590Srgrimes Node* generate_guard(Node* test, RegionNode* region, float true_prob); 881590Srgrimes Node* generate_slow_guard(Node* test, RegionNode* region); 891590Srgrimes Node* generate_fair_guard(Node* test, RegionNode* region); 901590Srgrimes Node* generate_negative_guard(Node* index, RegionNode* region, 911590Srgrimes // resulting CastII of index: 921590Srgrimes Node* *pos_index = NULL); 931590Srgrimes Node* generate_nonpositive_guard(Node* index, bool never_negative, 941590Srgrimes // resulting CastII of index: 951590Srgrimes Node* *pos_index = NULL); 961590Srgrimes Node* generate_limit_guard(Node* offset, Node* subseq_length, 971590Srgrimes Node* array_length, 981590Srgrimes RegionNode* region); 991590Srgrimes Node* generate_current_thread(Node* &tls_output); 1001590Srgrimes address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 1011590Srgrimes bool disjoint_bases, const char* &name, bool dest_uninitialized); 1021590Srgrimes Node* load_mirror_from_klass(Node* klass); 1031590Srgrimes Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 1041590Srgrimes int nargs, 1051590Srgrimes RegionNode* region, int null_path, 1061590Srgrimes int offset); 1071590Srgrimes Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs, 1081590Srgrimes RegionNode* region, int null_path) { 1091590Srgrimes int offset = java_lang_Class::klass_offset_in_bytes(); 1101590Srgrimes return load_klass_from_mirror_common(mirror, never_see_null, nargs, 1111590Srgrimes region, null_path, 1121590Srgrimes offset); 1131590Srgrimes } 1141590Srgrimes Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 1151590Srgrimes int nargs, 1161590Srgrimes RegionNode* region, int null_path) { 1171590Srgrimes int offset = java_lang_Class::array_klass_offset_in_bytes(); 1181590Srgrimes return load_klass_from_mirror_common(mirror, never_see_null, nargs, 1191590Srgrimes region, null_path, 1201590Srgrimes offset); 1211590Srgrimes } 1221590Srgrimes Node* generate_access_flags_guard(Node* kls, 1231590Srgrimes int modifier_mask, int modifier_bits, 1241590Srgrimes RegionNode* region); 1251590Srgrimes Node* generate_interface_guard(Node* kls, RegionNode* region); 1261590Srgrimes Node* generate_array_guard(Node* kls, RegionNode* region) { 1271590Srgrimes return generate_array_guard_common(kls, region, false, false); 1281590Srgrimes } 1291590Srgrimes Node* generate_non_array_guard(Node* kls, RegionNode* region) { 1301590Srgrimes return generate_array_guard_common(kls, region, false, true); 1311590Srgrimes } 1321590Srgrimes Node* generate_objArray_guard(Node* kls, RegionNode* region) { 1331590Srgrimes return generate_array_guard_common(kls, region, true, false); 1341590Srgrimes } 1351590Srgrimes Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 1361590Srgrimes return generate_array_guard_common(kls, region, true, true); 1371590Srgrimes } 1381590Srgrimes Node* generate_array_guard_common(Node* kls, RegionNode* region, 1391590Srgrimes bool obj_array, bool not_array); 1401590Srgrimes Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 1411590Srgrimes CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 1421590Srgrimes bool is_virtual = false, bool is_static = false); 1431590Srgrimes CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 1441590Srgrimes return generate_method_call(method_id, false, true); 1451590Srgrimes } 1461590Srgrimes CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 1471590Srgrimes return generate_method_call(method_id, true, false); 1481590Srgrimes } 1491590Srgrimes 1501590Srgrimes Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 1511590Srgrimes Node* make_string_method_node(int opcode, Node* str1, Node* str2); 1521590Srgrimes bool inline_string_compareTo(); 1531590Srgrimes bool inline_string_indexOf(); 1541590Srgrimes Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 1551590Srgrimes bool inline_string_equals(); 1561590Srgrimes Node* pop_math_arg(); 1571590Srgrimes bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 1581590Srgrimes bool inline_math_native(vmIntrinsics::ID id); 1591590Srgrimes bool inline_trig(vmIntrinsics::ID id); 1601590Srgrimes bool inline_trans(vmIntrinsics::ID id); 1611590Srgrimes bool inline_abs(vmIntrinsics::ID id); 1621590Srgrimes bool inline_sqrt(vmIntrinsics::ID id); 1631590Srgrimes void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 1641590Srgrimes bool inline_pow(vmIntrinsics::ID id); 1651590Srgrimes bool inline_exp(vmIntrinsics::ID id); 1661590Srgrimes bool inline_min_max(vmIntrinsics::ID id); 1671590Srgrimes Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 1681590Srgrimes // This returns Type::AnyPtr, RawPtr, or OopPtr. 1691590Srgrimes int classify_unsafe_addr(Node* &base, Node* &offset); 1701590Srgrimes Node* make_unsafe_address(Node* base, Node* offset); 1711590Srgrimes // Helper for inline_unsafe_access. 1721590Srgrimes // Generates the guards that check whether the result of 1731590Srgrimes // Unsafe.getObject should be recorded in an SATB log buffer. 1741590Srgrimes void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, int nargs, bool need_mem_bar); 1751590Srgrimes bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 1761590Srgrimes bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 1771590Srgrimes bool inline_unsafe_allocate(); 1781590Srgrimes bool inline_unsafe_copyMemory(); 1791590Srgrimes bool inline_native_currentThread(); 1801590Srgrimes#ifdef TRACE_HAVE_INTRINSICS 1811590Srgrimes bool inline_native_classID(); 1821590Srgrimes bool inline_native_threadID(); 1831590Srgrimes#endif 1841590Srgrimes bool inline_native_time_funcs(address method, const char* funcName); 1851590Srgrimes bool inline_native_isInterrupted(); 1861590Srgrimes bool inline_native_Class_query(vmIntrinsics::ID id); 1871590Srgrimes bool inline_native_subtype_check(); 1881590Srgrimes 1891590Srgrimes bool inline_native_newArray(); 1901590Srgrimes bool inline_native_getLength(); 1911590Srgrimes bool inline_array_copyOf(bool is_copyOfRange); 1921590Srgrimes bool inline_array_equals(); 1931590Srgrimes void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 1941590Srgrimes bool inline_native_clone(bool is_virtual); 1951590Srgrimes bool inline_native_Reflection_getCallerClass(); 1961590Srgrimes bool is_method_invoke_or_aux_frame(JVMState* jvms); 1971590Srgrimes // Helper function for inlining native object hash method 1981590Srgrimes bool inline_native_hashcode(bool is_virtual, bool is_static); 1991590Srgrimes bool inline_native_getClass(); 2001590Srgrimes 2011590Srgrimes // Helper functions for inlining arraycopy 2021590Srgrimes bool inline_arraycopy(); 2031590Srgrimes void generate_arraycopy(const TypePtr* adr_type, 2041590Srgrimes BasicType basic_elem_type, 2051590Srgrimes Node* src, Node* src_offset, 2061590Srgrimes Node* dest, Node* dest_offset, 2071590Srgrimes Node* copy_length, 2081590Srgrimes bool disjoint_bases = false, 2091590Srgrimes bool length_never_negative = false, 2101590Srgrimes RegionNode* slow_region = NULL); 2111590Srgrimes AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 2121590Srgrimes RegionNode* slow_region); 2138874Srgrimes void generate_clear_array(const TypePtr* adr_type, 2141590Srgrimes Node* dest, 2151590Srgrimes BasicType basic_elem_type, 2161590Srgrimes Node* slice_off, 2171590Srgrimes Node* slice_len, 2181590Srgrimes Node* slice_end); 2191590Srgrimes bool generate_block_arraycopy(const TypePtr* adr_type, 2201590Srgrimes BasicType basic_elem_type, 2211590Srgrimes AllocateNode* alloc, 2221590Srgrimes Node* src, Node* src_offset, 2231590Srgrimes Node* dest, Node* dest_offset, 2241590Srgrimes Node* dest_size, bool dest_uninitialized); 2251590Srgrimes void generate_slow_arraycopy(const TypePtr* adr_type, 2261590Srgrimes Node* src, Node* src_offset, 2271590Srgrimes Node* dest, Node* dest_offset, 2281590Srgrimes Node* copy_length, bool dest_uninitialized); 2291590Srgrimes Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 2301590Srgrimes Node* dest_elem_klass, 2311590Srgrimes Node* src, Node* src_offset, 2321590Srgrimes Node* dest, Node* dest_offset, 2331590Srgrimes Node* copy_length, bool dest_uninitialized); 2341590Srgrimes Node* generate_generic_arraycopy(const TypePtr* adr_type, 2351590Srgrimes Node* src, Node* src_offset, 2361590Srgrimes Node* dest, Node* dest_offset, 2371590Srgrimes Node* copy_length, bool dest_uninitialized); 2381590Srgrimes void generate_unchecked_arraycopy(const TypePtr* adr_type, 2391590Srgrimes BasicType basic_elem_type, 2401590Srgrimes bool disjoint_bases, 2411590Srgrimes Node* src, Node* src_offset, 2421590Srgrimes Node* dest, Node* dest_offset, 2431590Srgrimes Node* copy_length, bool dest_uninitialized); 2441590Srgrimes bool inline_unsafe_CAS(BasicType type); 2451590Srgrimes bool inline_unsafe_ordered_store(BasicType type); 2461590Srgrimes bool inline_fp_conversions(vmIntrinsics::ID id); 247 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id); 248 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id); 249 bool inline_bitCount(vmIntrinsics::ID id); 250 bool inline_reverseBytes(vmIntrinsics::ID id); 251 252 bool inline_reference_get(); 253}; 254 255 256//---------------------------make_vm_intrinsic---------------------------- 257CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 258 vmIntrinsics::ID id = m->intrinsic_id(); 259 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 260 261 if (DisableIntrinsic[0] != '\0' 262 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) { 263 // disabled by a user request on the command line: 264 // example: -XX:DisableIntrinsic=_hashCode,_getClass 265 return NULL; 266 } 267 268 if (!m->is_loaded()) { 269 // do not attempt to inline unloaded methods 270 return NULL; 271 } 272 273 // Only a few intrinsics implement a virtual dispatch. 274 // They are expensive calls which are also frequently overridden. 275 if (is_virtual) { 276 switch (id) { 277 case vmIntrinsics::_hashCode: 278 case vmIntrinsics::_clone: 279 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 280 break; 281 default: 282 return NULL; 283 } 284 } 285 286 // -XX:-InlineNatives disables nearly all intrinsics: 287 if (!InlineNatives) { 288 switch (id) { 289 case vmIntrinsics::_indexOf: 290 case vmIntrinsics::_compareTo: 291 case vmIntrinsics::_equals: 292 case vmIntrinsics::_equalsC: 293 break; // InlineNatives does not control String.compareTo 294 case vmIntrinsics::_Reference_get: 295 break; // InlineNatives does not control Reference.get 296 default: 297 return NULL; 298 } 299 } 300 301 switch (id) { 302 case vmIntrinsics::_compareTo: 303 if (!SpecialStringCompareTo) return NULL; 304 break; 305 case vmIntrinsics::_indexOf: 306 if (!SpecialStringIndexOf) return NULL; 307 break; 308 case vmIntrinsics::_equals: 309 if (!SpecialStringEquals) return NULL; 310 break; 311 case vmIntrinsics::_equalsC: 312 if (!SpecialArraysEquals) return NULL; 313 break; 314 case vmIntrinsics::_arraycopy: 315 if (!InlineArrayCopy) return NULL; 316 break; 317 case vmIntrinsics::_copyMemory: 318 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 319 if (!InlineArrayCopy) return NULL; 320 break; 321 case vmIntrinsics::_hashCode: 322 if (!InlineObjectHash) return NULL; 323 break; 324 case vmIntrinsics::_clone: 325 case vmIntrinsics::_copyOf: 326 case vmIntrinsics::_copyOfRange: 327 if (!InlineObjectCopy) return NULL; 328 // These also use the arraycopy intrinsic mechanism: 329 if (!InlineArrayCopy) return NULL; 330 break; 331 case vmIntrinsics::_checkIndex: 332 // We do not intrinsify this. The optimizer does fine with it. 333 return NULL; 334 335 case vmIntrinsics::_getCallerClass: 336 if (!UseNewReflection) return NULL; 337 if (!InlineReflectionGetCallerClass) return NULL; 338 if (!JDK_Version::is_gte_jdk14x_version()) return NULL; 339 break; 340 341 case vmIntrinsics::_bitCount_i: 342 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 343 break; 344 345 case vmIntrinsics::_bitCount_l: 346 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 347 break; 348 349 case vmIntrinsics::_numberOfLeadingZeros_i: 350 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 351 break; 352 353 case vmIntrinsics::_numberOfLeadingZeros_l: 354 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 355 break; 356 357 case vmIntrinsics::_numberOfTrailingZeros_i: 358 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 359 break; 360 361 case vmIntrinsics::_numberOfTrailingZeros_l: 362 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 363 break; 364 365 case vmIntrinsics::_Reference_get: 366 // Use the intrinsic version of Reference.get() so that the value in 367 // the referent field can be registered by the G1 pre-barrier code. 368 // Also add memory barrier to prevent commoning reads from this field 369 // across safepoint since GC can change it value. 370 break; 371 372 default: 373 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 374 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 375 break; 376 } 377 378 // -XX:-InlineClassNatives disables natives from the Class class. 379 // The flag applies to all reflective calls, notably Array.newArray 380 // (visible to Java programmers as Array.newInstance). 381 if (m->holder()->name() == ciSymbol::java_lang_Class() || 382 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 383 if (!InlineClassNatives) return NULL; 384 } 385 386 // -XX:-InlineThreadNatives disables natives from the Thread class. 387 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 388 if (!InlineThreadNatives) return NULL; 389 } 390 391 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 392 if (m->holder()->name() == ciSymbol::java_lang_Math() || 393 m->holder()->name() == ciSymbol::java_lang_Float() || 394 m->holder()->name() == ciSymbol::java_lang_Double()) { 395 if (!InlineMathNatives) return NULL; 396 } 397 398 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 399 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 400 if (!InlineUnsafeOps) return NULL; 401 } 402 403 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id); 404} 405 406//----------------------register_library_intrinsics----------------------- 407// Initialize this file's data structures, for each Compile instance. 408void Compile::register_library_intrinsics() { 409 // Nothing to do here. 410} 411 412JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 413 LibraryCallKit kit(jvms, this); 414 Compile* C = kit.C; 415 int nodes = C->unique(); 416#ifndef PRODUCT 417 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) { 418 char buf[1000]; 419 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 420 tty->print_cr("Intrinsic %s", str); 421 } 422#endif 423 424 if (kit.try_to_inline()) { 425 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 426 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 427 } 428 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 429 if (C->log()) { 430 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 431 vmIntrinsics::name_at(intrinsic_id()), 432 (is_virtual() ? " virtual='1'" : ""), 433 C->unique() - nodes); 434 } 435 return kit.transfer_exceptions_into_jvms(); 436 } 437 438 // The intrinsic bailed out 439 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) { 440 if (jvms->has_method()) { 441 // Not a root compile. 442 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 443 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg); 444 } else { 445 // Root compile 446 tty->print("Did not generate intrinsic %s%s at bci:%d in", 447 vmIntrinsics::name_at(intrinsic_id()), 448 (is_virtual() ? " (virtual)" : ""), kit.bci()); 449 } 450 } 451 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 452 return NULL; 453} 454 455bool LibraryCallKit::try_to_inline() { 456 // Handle symbolic names for otherwise undistinguished boolean switches: 457 const bool is_store = true; 458 const bool is_native_ptr = true; 459 const bool is_static = true; 460 461 if (!jvms()->has_method()) { 462 // Root JVMState has a null method. 463 assert(map()->memory()->Opcode() == Op_Parm, ""); 464 // Insert the memory aliasing node 465 set_all_memory(reset_memory()); 466 } 467 assert(merged_memory(), ""); 468 469 switch (intrinsic_id()) { 470 case vmIntrinsics::_hashCode: 471 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 472 case vmIntrinsics::_identityHashCode: 473 return inline_native_hashcode(/*!virtual*/ false, is_static); 474 case vmIntrinsics::_getClass: 475 return inline_native_getClass(); 476 477 case vmIntrinsics::_dsin: 478 case vmIntrinsics::_dcos: 479 case vmIntrinsics::_dtan: 480 case vmIntrinsics::_dabs: 481 case vmIntrinsics::_datan2: 482 case vmIntrinsics::_dsqrt: 483 case vmIntrinsics::_dexp: 484 case vmIntrinsics::_dlog: 485 case vmIntrinsics::_dlog10: 486 case vmIntrinsics::_dpow: 487 return inline_math_native(intrinsic_id()); 488 489 case vmIntrinsics::_min: 490 case vmIntrinsics::_max: 491 return inline_min_max(intrinsic_id()); 492 493 case vmIntrinsics::_arraycopy: 494 return inline_arraycopy(); 495 496 case vmIntrinsics::_compareTo: 497 return inline_string_compareTo(); 498 case vmIntrinsics::_indexOf: 499 return inline_string_indexOf(); 500 case vmIntrinsics::_equals: 501 return inline_string_equals(); 502 503 case vmIntrinsics::_getObject: 504 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false); 505 case vmIntrinsics::_getBoolean: 506 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false); 507 case vmIntrinsics::_getByte: 508 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false); 509 case vmIntrinsics::_getShort: 510 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false); 511 case vmIntrinsics::_getChar: 512 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false); 513 case vmIntrinsics::_getInt: 514 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false); 515 case vmIntrinsics::_getLong: 516 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false); 517 case vmIntrinsics::_getFloat: 518 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false); 519 case vmIntrinsics::_getDouble: 520 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false); 521 522 case vmIntrinsics::_putObject: 523 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false); 524 case vmIntrinsics::_putBoolean: 525 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false); 526 case vmIntrinsics::_putByte: 527 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false); 528 case vmIntrinsics::_putShort: 529 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false); 530 case vmIntrinsics::_putChar: 531 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false); 532 case vmIntrinsics::_putInt: 533 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false); 534 case vmIntrinsics::_putLong: 535 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false); 536 case vmIntrinsics::_putFloat: 537 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false); 538 case vmIntrinsics::_putDouble: 539 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false); 540 541 case vmIntrinsics::_getByte_raw: 542 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false); 543 case vmIntrinsics::_getShort_raw: 544 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false); 545 case vmIntrinsics::_getChar_raw: 546 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false); 547 case vmIntrinsics::_getInt_raw: 548 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false); 549 case vmIntrinsics::_getLong_raw: 550 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false); 551 case vmIntrinsics::_getFloat_raw: 552 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false); 553 case vmIntrinsics::_getDouble_raw: 554 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false); 555 case vmIntrinsics::_getAddress_raw: 556 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false); 557 558 case vmIntrinsics::_putByte_raw: 559 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false); 560 case vmIntrinsics::_putShort_raw: 561 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false); 562 case vmIntrinsics::_putChar_raw: 563 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false); 564 case vmIntrinsics::_putInt_raw: 565 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false); 566 case vmIntrinsics::_putLong_raw: 567 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false); 568 case vmIntrinsics::_putFloat_raw: 569 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false); 570 case vmIntrinsics::_putDouble_raw: 571 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false); 572 case vmIntrinsics::_putAddress_raw: 573 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false); 574 575 case vmIntrinsics::_getObjectVolatile: 576 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true); 577 case vmIntrinsics::_getBooleanVolatile: 578 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true); 579 case vmIntrinsics::_getByteVolatile: 580 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true); 581 case vmIntrinsics::_getShortVolatile: 582 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true); 583 case vmIntrinsics::_getCharVolatile: 584 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true); 585 case vmIntrinsics::_getIntVolatile: 586 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true); 587 case vmIntrinsics::_getLongVolatile: 588 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true); 589 case vmIntrinsics::_getFloatVolatile: 590 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true); 591 case vmIntrinsics::_getDoubleVolatile: 592 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true); 593 594 case vmIntrinsics::_putObjectVolatile: 595 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true); 596 case vmIntrinsics::_putBooleanVolatile: 597 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true); 598 case vmIntrinsics::_putByteVolatile: 599 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true); 600 case vmIntrinsics::_putShortVolatile: 601 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true); 602 case vmIntrinsics::_putCharVolatile: 603 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true); 604 case vmIntrinsics::_putIntVolatile: 605 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true); 606 case vmIntrinsics::_putLongVolatile: 607 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true); 608 case vmIntrinsics::_putFloatVolatile: 609 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true); 610 case vmIntrinsics::_putDoubleVolatile: 611 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true); 612 613 case vmIntrinsics::_prefetchRead: 614 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 615 case vmIntrinsics::_prefetchWrite: 616 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 617 case vmIntrinsics::_prefetchReadStatic: 618 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 619 case vmIntrinsics::_prefetchWriteStatic: 620 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 621 622 case vmIntrinsics::_compareAndSwapObject: 623 return inline_unsafe_CAS(T_OBJECT); 624 case vmIntrinsics::_compareAndSwapInt: 625 return inline_unsafe_CAS(T_INT); 626 case vmIntrinsics::_compareAndSwapLong: 627 return inline_unsafe_CAS(T_LONG); 628 629 case vmIntrinsics::_putOrderedObject: 630 return inline_unsafe_ordered_store(T_OBJECT); 631 case vmIntrinsics::_putOrderedInt: 632 return inline_unsafe_ordered_store(T_INT); 633 case vmIntrinsics::_putOrderedLong: 634 return inline_unsafe_ordered_store(T_LONG); 635 636 case vmIntrinsics::_currentThread: 637 return inline_native_currentThread(); 638 case vmIntrinsics::_isInterrupted: 639 return inline_native_isInterrupted(); 640 641#ifdef TRACE_HAVE_INTRINSICS 642 case vmIntrinsics::_classID: 643 return inline_native_classID(); 644 case vmIntrinsics::_threadID: 645 return inline_native_threadID(); 646 case vmIntrinsics::_counterTime: 647 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 648#endif 649 case vmIntrinsics::_currentTimeMillis: 650 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 651 case vmIntrinsics::_nanoTime: 652 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 653 case vmIntrinsics::_allocateInstance: 654 return inline_unsafe_allocate(); 655 case vmIntrinsics::_copyMemory: 656 return inline_unsafe_copyMemory(); 657 case vmIntrinsics::_newArray: 658 return inline_native_newArray(); 659 case vmIntrinsics::_getLength: 660 return inline_native_getLength(); 661 case vmIntrinsics::_copyOf: 662 return inline_array_copyOf(false); 663 case vmIntrinsics::_copyOfRange: 664 return inline_array_copyOf(true); 665 case vmIntrinsics::_equalsC: 666 return inline_array_equals(); 667 case vmIntrinsics::_clone: 668 return inline_native_clone(intrinsic()->is_virtual()); 669 670 case vmIntrinsics::_isAssignableFrom: 671 return inline_native_subtype_check(); 672 673 case vmIntrinsics::_isInstance: 674 case vmIntrinsics::_getModifiers: 675 case vmIntrinsics::_isInterface: 676 case vmIntrinsics::_isArray: 677 case vmIntrinsics::_isPrimitive: 678 case vmIntrinsics::_getSuperclass: 679 case vmIntrinsics::_getComponentType: 680 case vmIntrinsics::_getClassAccessFlags: 681 return inline_native_Class_query(intrinsic_id()); 682 683 case vmIntrinsics::_floatToRawIntBits: 684 case vmIntrinsics::_floatToIntBits: 685 case vmIntrinsics::_intBitsToFloat: 686 case vmIntrinsics::_doubleToRawLongBits: 687 case vmIntrinsics::_doubleToLongBits: 688 case vmIntrinsics::_longBitsToDouble: 689 return inline_fp_conversions(intrinsic_id()); 690 691 case vmIntrinsics::_numberOfLeadingZeros_i: 692 case vmIntrinsics::_numberOfLeadingZeros_l: 693 return inline_numberOfLeadingZeros(intrinsic_id()); 694 695 case vmIntrinsics::_numberOfTrailingZeros_i: 696 case vmIntrinsics::_numberOfTrailingZeros_l: 697 return inline_numberOfTrailingZeros(intrinsic_id()); 698 699 case vmIntrinsics::_bitCount_i: 700 case vmIntrinsics::_bitCount_l: 701 return inline_bitCount(intrinsic_id()); 702 703 case vmIntrinsics::_reverseBytes_i: 704 case vmIntrinsics::_reverseBytes_l: 705 case vmIntrinsics::_reverseBytes_s: 706 case vmIntrinsics::_reverseBytes_c: 707 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id()); 708 709 case vmIntrinsics::_getCallerClass: 710 return inline_native_Reflection_getCallerClass(); 711 712 case vmIntrinsics::_Reference_get: 713 return inline_reference_get(); 714 715 default: 716 // If you get here, it may be that someone has added a new intrinsic 717 // to the list in vmSymbols.hpp without implementing it here. 718#ifndef PRODUCT 719 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 720 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 721 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 722 } 723#endif 724 return false; 725 } 726} 727 728//------------------------------push_result------------------------------ 729// Helper function for finishing intrinsics. 730void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) { 731 record_for_igvn(region); 732 set_control(_gvn.transform(region)); 733 BasicType value_type = value->type()->basic_type(); 734 push_node(value_type, _gvn.transform(value)); 735} 736 737//------------------------------generate_guard--------------------------- 738// Helper function for generating guarded fast-slow graph structures. 739// The given 'test', if true, guards a slow path. If the test fails 740// then a fast path can be taken. (We generally hope it fails.) 741// In all cases, GraphKit::control() is updated to the fast path. 742// The returned value represents the control for the slow path. 743// The return value is never 'top'; it is either a valid control 744// or NULL if it is obvious that the slow path can never be taken. 745// Also, if region and the slow control are not NULL, the slow edge 746// is appended to the region. 747Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 748 if (stopped()) { 749 // Already short circuited. 750 return NULL; 751 } 752 753 // Build an if node and its projections. 754 // If test is true we take the slow path, which we assume is uncommon. 755 if (_gvn.type(test) == TypeInt::ZERO) { 756 // The slow branch is never taken. No need to build this guard. 757 return NULL; 758 } 759 760 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 761 762 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) ); 763 if (if_slow == top()) { 764 // The slow branch is never taken. No need to build this guard. 765 return NULL; 766 } 767 768 if (region != NULL) 769 region->add_req(if_slow); 770 771 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) ); 772 set_control(if_fast); 773 774 return if_slow; 775} 776 777inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 778 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 779} 780inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 781 return generate_guard(test, region, PROB_FAIR); 782} 783 784inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 785 Node* *pos_index) { 786 if (stopped()) 787 return NULL; // already stopped 788 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 789 return NULL; // index is already adequately typed 790 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 791 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 792 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 793 if (is_neg != NULL && pos_index != NULL) { 794 // Emulate effect of Parse::adjust_map_after_if. 795 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS); 796 ccast->set_req(0, control()); 797 (*pos_index) = _gvn.transform(ccast); 798 } 799 return is_neg; 800} 801 802inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 803 Node* *pos_index) { 804 if (stopped()) 805 return NULL; // already stopped 806 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 807 return NULL; // index is already adequately typed 808 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) ); 809 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 810 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) ); 811 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 812 if (is_notp != NULL && pos_index != NULL) { 813 // Emulate effect of Parse::adjust_map_after_if. 814 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1); 815 ccast->set_req(0, control()); 816 (*pos_index) = _gvn.transform(ccast); 817 } 818 return is_notp; 819} 820 821// Make sure that 'position' is a valid limit index, in [0..length]. 822// There are two equivalent plans for checking this: 823// A. (offset + copyLength) unsigned<= arrayLength 824// B. offset <= (arrayLength - copyLength) 825// We require that all of the values above, except for the sum and 826// difference, are already known to be non-negative. 827// Plan A is robust in the face of overflow, if offset and copyLength 828// are both hugely positive. 829// 830// Plan B is less direct and intuitive, but it does not overflow at 831// all, since the difference of two non-negatives is always 832// representable. Whenever Java methods must perform the equivalent 833// check they generally use Plan B instead of Plan A. 834// For the moment we use Plan A. 835inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 836 Node* subseq_length, 837 Node* array_length, 838 RegionNode* region) { 839 if (stopped()) 840 return NULL; // already stopped 841 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 842 if (zero_offset && subseq_length->eqv_uncast(array_length)) 843 return NULL; // common case of whole-array copy 844 Node* last = subseq_length; 845 if (!zero_offset) // last += offset 846 last = _gvn.transform( new (C, 3) AddINode(last, offset)); 847 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) ); 848 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) ); 849 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 850 return is_over; 851} 852 853 854//--------------------------generate_current_thread-------------------- 855Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 856 ciKlass* thread_klass = env()->Thread_klass(); 857 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 858 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode()); 859 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 860 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT); 861 tls_output = thread; 862 return threadObj; 863} 864 865 866//------------------------------make_string_method_node------------------------ 867// Helper method for String intrinsic functions. This version is called 868// with str1 and str2 pointing to String object nodes. 869// 870Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 871 Node* no_ctrl = NULL; 872 873 // Get start addr of string 874 Node* str1_value = load_String_value(no_ctrl, str1); 875 Node* str1_offset = load_String_offset(no_ctrl, str1); 876 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 877 878 // Get length of string 1 879 Node* str1_len = load_String_length(no_ctrl, str1); 880 881 Node* str2_value = load_String_value(no_ctrl, str2); 882 Node* str2_offset = load_String_offset(no_ctrl, str2); 883 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 884 885 Node* str2_len = NULL; 886 Node* result = NULL; 887 888 switch (opcode) { 889 case Op_StrIndexOf: 890 // Get length of string 2 891 str2_len = load_String_length(no_ctrl, str2); 892 893 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 894 str1_start, str1_len, str2_start, str2_len); 895 break; 896 case Op_StrComp: 897 // Get length of string 2 898 str2_len = load_String_length(no_ctrl, str2); 899 900 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS), 901 str1_start, str1_len, str2_start, str2_len); 902 break; 903 case Op_StrEquals: 904 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 905 str1_start, str2_start, str1_len); 906 break; 907 default: 908 ShouldNotReachHere(); 909 return NULL; 910 } 911 912 // All these intrinsics have checks. 913 C->set_has_split_ifs(true); // Has chance for split-if optimization 914 915 return _gvn.transform(result); 916} 917 918// Helper method for String intrinsic functions. This version is called 919// with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 920// to Int nodes containing the lenghts of str1 and str2. 921// 922Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 923 924 Node* result = NULL; 925 switch (opcode) { 926 case Op_StrIndexOf: 927 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 928 str1_start, cnt1, str2_start, cnt2); 929 break; 930 case Op_StrComp: 931 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS), 932 str1_start, cnt1, str2_start, cnt2); 933 break; 934 case Op_StrEquals: 935 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 936 str1_start, str2_start, cnt1); 937 break; 938 default: 939 ShouldNotReachHere(); 940 return NULL; 941 } 942 943 // All these intrinsics have checks. 944 C->set_has_split_ifs(true); // Has chance for split-if optimization 945 946 return _gvn.transform(result); 947} 948 949//------------------------------inline_string_compareTo------------------------ 950bool LibraryCallKit::inline_string_compareTo() { 951 952 if (!Matcher::has_match_rule(Op_StrComp)) return false; 953 954 _sp += 2; 955 Node *argument = pop(); // pop non-receiver first: it was pushed second 956 Node *receiver = pop(); 957 958 // Null check on self without removing any arguments. The argument 959 // null check technically happens in the wrong place, which can lead to 960 // invalid stack traces when string compare is inlined into a method 961 // which handles NullPointerExceptions. 962 _sp += 2; 963 receiver = do_null_check(receiver, T_OBJECT); 964 argument = do_null_check(argument, T_OBJECT); 965 _sp -= 2; 966 if (stopped()) { 967 return true; 968 } 969 970 Node* compare = make_string_method_node(Op_StrComp, receiver, argument); 971 push(compare); 972 return true; 973} 974 975//------------------------------inline_string_equals------------------------ 976bool LibraryCallKit::inline_string_equals() { 977 978 if (!Matcher::has_match_rule(Op_StrEquals)) return false; 979 980 int nargs = 2; 981 _sp += nargs; 982 Node* argument = pop(); // pop non-receiver first: it was pushed second 983 Node* receiver = pop(); 984 985 // Null check on self without removing any arguments. The argument 986 // null check technically happens in the wrong place, which can lead to 987 // invalid stack traces when string compare is inlined into a method 988 // which handles NullPointerExceptions. 989 _sp += nargs; 990 receiver = do_null_check(receiver, T_OBJECT); 991 //should not do null check for argument for String.equals(), because spec 992 //allows to specify NULL as argument. 993 _sp -= nargs; 994 995 if (stopped()) { 996 return true; 997 } 998 999 // paths (plus control) merge 1000 RegionNode* region = new (C, 5) RegionNode(5); 1001 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL); 1002 1003 // does source == target string? 1004 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument)); 1005 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq)); 1006 1007 Node* if_eq = generate_slow_guard(bol, NULL); 1008 if (if_eq != NULL) { 1009 // receiver == argument 1010 phi->init_req(2, intcon(1)); 1011 region->init_req(2, if_eq); 1012 } 1013 1014 // get String klass for instanceOf 1015 ciInstanceKlass* klass = env()->String_klass(); 1016 1017 if (!stopped()) { 1018 _sp += nargs; // gen_instanceof might do an uncommon trap 1019 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1020 _sp -= nargs; 1021 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1))); 1022 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne)); 1023 1024 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1025 //instanceOf == true, fallthrough 1026 1027 if (inst_false != NULL) { 1028 phi->init_req(3, intcon(0)); 1029 region->init_req(3, inst_false); 1030 } 1031 } 1032 1033 if (!stopped()) { 1034 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1035 1036 // Properly cast the argument to String 1037 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type)); 1038 // This path is taken only when argument's type is String:NotNull. 1039 argument = cast_not_null(argument, false); 1040 1041 Node* no_ctrl = NULL; 1042 1043 // Get start addr of receiver 1044 Node* receiver_val = load_String_value(no_ctrl, receiver); 1045 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1046 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1047 1048 // Get length of receiver 1049 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1050 1051 // Get start addr of argument 1052 Node* argument_val = load_String_value(no_ctrl, argument); 1053 Node* argument_offset = load_String_offset(no_ctrl, argument); 1054 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1055 1056 // Get length of argument 1057 Node* argument_cnt = load_String_length(no_ctrl, argument); 1058 1059 // Check for receiver count != argument count 1060 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) ); 1061 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) ); 1062 Node* if_ne = generate_slow_guard(bol, NULL); 1063 if (if_ne != NULL) { 1064 phi->init_req(4, intcon(0)); 1065 region->init_req(4, if_ne); 1066 } 1067 1068 // Check for count == 0 is done by assembler code for StrEquals. 1069 1070 if (!stopped()) { 1071 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1072 phi->init_req(1, equals); 1073 region->init_req(1, control()); 1074 } 1075 } 1076 1077 // post merge 1078 set_control(_gvn.transform(region)); 1079 record_for_igvn(region); 1080 1081 push(_gvn.transform(phi)); 1082 1083 return true; 1084} 1085 1086//------------------------------inline_array_equals---------------------------- 1087bool LibraryCallKit::inline_array_equals() { 1088 1089 if (!Matcher::has_match_rule(Op_AryEq)) return false; 1090 1091 _sp += 2; 1092 Node *argument2 = pop(); 1093 Node *argument1 = pop(); 1094 1095 Node* equals = 1096 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS), 1097 argument1, argument2) ); 1098 push(equals); 1099 return true; 1100} 1101 1102// Java version of String.indexOf(constant string) 1103// class StringDecl { 1104// StringDecl(char[] ca) { 1105// offset = 0; 1106// count = ca.length; 1107// value = ca; 1108// } 1109// int offset; 1110// int count; 1111// char[] value; 1112// } 1113// 1114// static int string_indexOf_J(StringDecl string_object, char[] target_object, 1115// int targetOffset, int cache_i, int md2) { 1116// int cache = cache_i; 1117// int sourceOffset = string_object.offset; 1118// int sourceCount = string_object.count; 1119// int targetCount = target_object.length; 1120// 1121// int targetCountLess1 = targetCount - 1; 1122// int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1123// 1124// char[] source = string_object.value; 1125// char[] target = target_object; 1126// int lastChar = target[targetCountLess1]; 1127// 1128// outer_loop: 1129// for (int i = sourceOffset; i < sourceEnd; ) { 1130// int src = source[i + targetCountLess1]; 1131// if (src == lastChar) { 1132// // With random strings and a 4-character alphabet, 1133// // reverse matching at this point sets up 0.8% fewer 1134// // frames, but (paradoxically) makes 0.3% more probes. 1135// // Since those probes are nearer the lastChar probe, 1136// // there is may be a net D$ win with reverse matching. 1137// // But, reversing loop inhibits unroll of inner loop 1138// // for unknown reason. So, does running outer loop from 1139// // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1140// for (int j = 0; j < targetCountLess1; j++) { 1141// if (target[targetOffset + j] != source[i+j]) { 1142// if ((cache & (1 << source[i+j])) == 0) { 1143// if (md2 < j+1) { 1144// i += j+1; 1145// continue outer_loop; 1146// } 1147// } 1148// i += md2; 1149// continue outer_loop; 1150// } 1151// } 1152// return i - sourceOffset; 1153// } 1154// if ((cache & (1 << src)) == 0) { 1155// i += targetCountLess1; 1156// } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1157// i++; 1158// } 1159// return -1; 1160// } 1161 1162//------------------------------string_indexOf------------------------ 1163Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1164 jint cache_i, jint md2_i) { 1165 1166 Node* no_ctrl = NULL; 1167 float likely = PROB_LIKELY(0.9); 1168 float unlikely = PROB_UNLIKELY(0.9); 1169 1170 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate 1171 1172 Node* source = load_String_value(no_ctrl, string_object); 1173 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1174 Node* sourceCount = load_String_length(no_ctrl, string_object); 1175 1176 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) ); 1177 jint target_length = target_array->length(); 1178 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1179 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1180 1181 IdealKit kit(this, false, true); 1182#define __ kit. 1183 Node* zero = __ ConI(0); 1184 Node* one = __ ConI(1); 1185 Node* cache = __ ConI(cache_i); 1186 Node* md2 = __ ConI(md2_i); 1187 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1188 Node* targetCount = __ ConI(target_length); 1189 Node* targetCountLess1 = __ ConI(target_length - 1); 1190 Node* targetOffset = __ ConI(targetOffset_i); 1191 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1192 1193 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1194 Node* outer_loop = __ make_label(2 /* goto */); 1195 Node* return_ = __ make_label(1); 1196 1197 __ set(rtn,__ ConI(-1)); 1198 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1199 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1200 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1201 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1202 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1203 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1204 Node* tpj = __ AddI(targetOffset, __ value(j)); 1205 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1206 Node* ipj = __ AddI(__ value(i), __ value(j)); 1207 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1208 __ if_then(targ, BoolTest::ne, src2); { 1209 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1210 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1211 __ increment(i, __ AddI(__ value(j), one)); 1212 __ goto_(outer_loop); 1213 } __ end_if(); __ dead(j); 1214 }__ end_if(); __ dead(j); 1215 __ increment(i, md2); 1216 __ goto_(outer_loop); 1217 }__ end_if(); 1218 __ increment(j, one); 1219 }__ end_loop(); __ dead(j); 1220 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1221 __ goto_(return_); 1222 }__ end_if(); 1223 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1224 __ increment(i, targetCountLess1); 1225 }__ end_if(); 1226 __ increment(i, one); 1227 __ bind(outer_loop); 1228 }__ end_loop(); __ dead(i); 1229 __ bind(return_); 1230 1231 // Final sync IdealKit and GraphKit. 1232 final_sync(kit); 1233 Node* result = __ value(rtn); 1234#undef __ 1235 C->set_has_loops(true); 1236 return result; 1237} 1238 1239//------------------------------inline_string_indexOf------------------------ 1240bool LibraryCallKit::inline_string_indexOf() { 1241 1242 _sp += 2; 1243 Node *argument = pop(); // pop non-receiver first: it was pushed second 1244 Node *receiver = pop(); 1245 1246 Node* result; 1247 // Disable the use of pcmpestri until it can be guaranteed that 1248 // the load doesn't cross into the uncommited space. 1249 if (Matcher::has_match_rule(Op_StrIndexOf) && 1250 UseSSE42Intrinsics) { 1251 // Generate SSE4.2 version of indexOf 1252 // We currently only have match rules that use SSE4.2 1253 1254 // Null check on self without removing any arguments. The argument 1255 // null check technically happens in the wrong place, which can lead to 1256 // invalid stack traces when string compare is inlined into a method 1257 // which handles NullPointerExceptions. 1258 _sp += 2; 1259 receiver = do_null_check(receiver, T_OBJECT); 1260 argument = do_null_check(argument, T_OBJECT); 1261 _sp -= 2; 1262 1263 if (stopped()) { 1264 return true; 1265 } 1266 1267 ciInstanceKlass* str_klass = env()->String_klass(); 1268 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1269 1270 // Make the merge point 1271 RegionNode* result_rgn = new (C, 4) RegionNode(4); 1272 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT); 1273 Node* no_ctrl = NULL; 1274 1275 // Get start addr of source string 1276 Node* source = load_String_value(no_ctrl, receiver); 1277 Node* source_offset = load_String_offset(no_ctrl, receiver); 1278 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1279 1280 // Get length of source string 1281 Node* source_cnt = load_String_length(no_ctrl, receiver); 1282 1283 // Get start addr of substring 1284 Node* substr = load_String_value(no_ctrl, argument); 1285 Node* substr_offset = load_String_offset(no_ctrl, argument); 1286 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1287 1288 // Get length of source string 1289 Node* substr_cnt = load_String_length(no_ctrl, argument); 1290 1291 // Check for substr count > string count 1292 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) ); 1293 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) ); 1294 Node* if_gt = generate_slow_guard(bol, NULL); 1295 if (if_gt != NULL) { 1296 result_phi->init_req(2, intcon(-1)); 1297 result_rgn->init_req(2, if_gt); 1298 } 1299 1300 if (!stopped()) { 1301 // Check for substr count == 0 1302 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) ); 1303 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 1304 Node* if_zero = generate_slow_guard(bol, NULL); 1305 if (if_zero != NULL) { 1306 result_phi->init_req(3, intcon(0)); 1307 result_rgn->init_req(3, if_zero); 1308 } 1309 } 1310 1311 if (!stopped()) { 1312 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1313 result_phi->init_req(1, result); 1314 result_rgn->init_req(1, control()); 1315 } 1316 set_control(_gvn.transform(result_rgn)); 1317 record_for_igvn(result_rgn); 1318 result = _gvn.transform(result_phi); 1319 1320 } else { // Use LibraryCallKit::string_indexOf 1321 // don't intrinsify if argument isn't a constant string. 1322 if (!argument->is_Con()) { 1323 return false; 1324 } 1325 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr(); 1326 if (str_type == NULL) { 1327 return false; 1328 } 1329 ciInstanceKlass* klass = env()->String_klass(); 1330 ciObject* str_const = str_type->const_oop(); 1331 if (str_const == NULL || str_const->klass() != klass) { 1332 return false; 1333 } 1334 ciInstance* str = str_const->as_instance(); 1335 assert(str != NULL, "must be instance"); 1336 1337 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1338 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1339 1340 int o; 1341 int c; 1342 if (java_lang_String::has_offset_field()) { 1343 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1344 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1345 } else { 1346 o = 0; 1347 c = pat->length(); 1348 } 1349 1350 // constant strings have no offset and count == length which 1351 // simplifies the resulting code somewhat so lets optimize for that. 1352 if (o != 0 || c != pat->length()) { 1353 return false; 1354 } 1355 1356 // Null check on self without removing any arguments. The argument 1357 // null check technically happens in the wrong place, which can lead to 1358 // invalid stack traces when string compare is inlined into a method 1359 // which handles NullPointerExceptions. 1360 _sp += 2; 1361 receiver = do_null_check(receiver, T_OBJECT); 1362 // No null check on the argument is needed since it's a constant String oop. 1363 _sp -= 2; 1364 if (stopped()) { 1365 return true; 1366 } 1367 1368 // The null string as a pattern always returns 0 (match at beginning of string) 1369 if (c == 0) { 1370 push(intcon(0)); 1371 return true; 1372 } 1373 1374 // Generate default indexOf 1375 jchar lastChar = pat->char_at(o + (c - 1)); 1376 int cache = 0; 1377 int i; 1378 for (i = 0; i < c - 1; i++) { 1379 assert(i < pat->length(), "out of range"); 1380 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1381 } 1382 1383 int md2 = c; 1384 for (i = 0; i < c - 1; i++) { 1385 assert(i < pat->length(), "out of range"); 1386 if (pat->char_at(o + i) == lastChar) { 1387 md2 = (c - 1) - i; 1388 } 1389 } 1390 1391 result = string_indexOf(receiver, pat, o, cache, md2); 1392 } 1393 1394 push(result); 1395 return true; 1396} 1397 1398//--------------------------pop_math_arg-------------------------------- 1399// Pop a double argument to a math function from the stack 1400// rounding it if necessary. 1401Node * LibraryCallKit::pop_math_arg() { 1402 Node *arg = pop_pair(); 1403 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 ) 1404 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) ); 1405 return arg; 1406} 1407 1408//------------------------------inline_trig---------------------------------- 1409// Inline sin/cos/tan instructions, if possible. If rounding is required, do 1410// argument reduction which will turn into a fast/slow diamond. 1411bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1412 _sp += arg_size(); // restore stack pointer 1413 Node* arg = pop_math_arg(); 1414 Node* trig = NULL; 1415 1416 switch (id) { 1417 case vmIntrinsics::_dsin: 1418 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg)); 1419 break; 1420 case vmIntrinsics::_dcos: 1421 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg)); 1422 break; 1423 case vmIntrinsics::_dtan: 1424 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg)); 1425 break; 1426 default: 1427 assert(false, "bad intrinsic was passed in"); 1428 return false; 1429 } 1430 1431 // Rounding required? Check for argument reduction! 1432 if( Matcher::strict_fp_requires_explicit_rounding ) { 1433 1434 static const double pi_4 = 0.7853981633974483; 1435 static const double neg_pi_4 = -0.7853981633974483; 1436 // pi/2 in 80-bit extended precision 1437 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1438 // -pi/2 in 80-bit extended precision 1439 // 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}; 1440 // Cutoff value for using this argument reduction technique 1441 //static const double pi_2_minus_epsilon = 1.564660403643354; 1442 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1443 1444 // Pseudocode for sin: 1445 // if (x <= Math.PI / 4.0) { 1446 // if (x >= -Math.PI / 4.0) return fsin(x); 1447 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1448 // } else { 1449 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1450 // } 1451 // return StrictMath.sin(x); 1452 1453 // Pseudocode for cos: 1454 // if (x <= Math.PI / 4.0) { 1455 // if (x >= -Math.PI / 4.0) return fcos(x); 1456 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1457 // } else { 1458 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1459 // } 1460 // return StrictMath.cos(x); 1461 1462 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1463 // requires a special machine instruction to load it. Instead we'll try 1464 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1465 // probably do the math inside the SIN encoding. 1466 1467 // Make the merge point 1468 RegionNode *r = new (C, 3) RegionNode(3); 1469 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE); 1470 1471 // Flatten arg so we need only 1 test 1472 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg)); 1473 // Node for PI/4 constant 1474 Node *pi4 = makecon(TypeD::make(pi_4)); 1475 // Check PI/4 : abs(arg) 1476 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs)); 1477 // Check: If PI/4 < abs(arg) then go slow 1478 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) ); 1479 // Branch either way 1480 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1481 set_control(opt_iff(r,iff)); 1482 1483 // Set fast path result 1484 phi->init_req(2,trig); 1485 1486 // Slow path - non-blocking leaf call 1487 Node* call = NULL; 1488 switch (id) { 1489 case vmIntrinsics::_dsin: 1490 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1491 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1492 "Sin", NULL, arg, top()); 1493 break; 1494 case vmIntrinsics::_dcos: 1495 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1496 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1497 "Cos", NULL, arg, top()); 1498 break; 1499 case vmIntrinsics::_dtan: 1500 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1501 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1502 "Tan", NULL, arg, top()); 1503 break; 1504 } 1505 assert(control()->in(0) == call, ""); 1506 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms)); 1507 r->init_req(1,control()); 1508 phi->init_req(1,slow_result); 1509 1510 // Post-merge 1511 set_control(_gvn.transform(r)); 1512 record_for_igvn(r); 1513 trig = _gvn.transform(phi); 1514 1515 C->set_has_split_ifs(true); // Has chance for split-if optimization 1516 } 1517 // Push result back on JVM stack 1518 push_pair(trig); 1519 return true; 1520} 1521 1522//------------------------------inline_sqrt------------------------------------- 1523// Inline square root instruction, if possible. 1524bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) { 1525 assert(id == vmIntrinsics::_dsqrt, "Not square root"); 1526 _sp += arg_size(); // restore stack pointer 1527 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg()))); 1528 return true; 1529} 1530 1531//------------------------------inline_abs------------------------------------- 1532// Inline absolute value instruction, if possible. 1533bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) { 1534 assert(id == vmIntrinsics::_dabs, "Not absolute value"); 1535 _sp += arg_size(); // restore stack pointer 1536 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg()))); 1537 return true; 1538} 1539 1540void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1541 //------------------- 1542 //result=(result.isNaN())? funcAddr():result; 1543 // Check: If isNaN() by checking result!=result? then either trap 1544 // or go to runtime 1545 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result)); 1546 // Build the boolean node 1547 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) ); 1548 1549 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1550 { 1551 BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1552 // End the current control-flow path 1553 push_pair(x); 1554 if (y != NULL) { 1555 push_pair(y); 1556 } 1557 // The pow or exp intrinsic returned a NaN, which requires a call 1558 // to the runtime. Recompile with the runtime call. 1559 uncommon_trap(Deoptimization::Reason_intrinsic, 1560 Deoptimization::Action_make_not_entrant); 1561 } 1562 push_pair(result); 1563 } else { 1564 // If this inlining ever returned NaN in the past, we compile a call 1565 // to the runtime to properly handle corner cases 1566 1567 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1568 Node* if_slow = _gvn.transform( new (C, 1) IfFalseNode(iff) ); 1569 Node* if_fast = _gvn.transform( new (C, 1) IfTrueNode(iff) ); 1570 1571 if (!if_slow->is_top()) { 1572 RegionNode* result_region = new(C, 3) RegionNode(3); 1573 PhiNode* result_val = new (C, 3) PhiNode(result_region, Type::DOUBLE); 1574 1575 result_region->init_req(1, if_fast); 1576 result_val->init_req(1, result); 1577 1578 set_control(if_slow); 1579 1580 const TypePtr* no_memory_effects = NULL; 1581 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1582 no_memory_effects, 1583 x, top(), y, y ? top() : NULL); 1584 Node* value = _gvn.transform(new (C, 1) ProjNode(rt, TypeFunc::Parms+0)); 1585#ifdef ASSERT 1586 Node* value_top = _gvn.transform(new (C, 1) ProjNode(rt, TypeFunc::Parms+1)); 1587 assert(value_top == top(), "second value must be top"); 1588#endif 1589 1590 result_region->init_req(2, control()); 1591 result_val->init_req(2, value); 1592 push_result(result_region, result_val); 1593 } else { 1594 push_pair(result); 1595 } 1596 } 1597} 1598 1599//------------------------------inline_exp------------------------------------- 1600// Inline exp instructions, if possible. The Intel hardware only misses 1601// really odd corner cases (+/- Infinity). Just uncommon-trap them. 1602bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) { 1603 assert(id == vmIntrinsics::_dexp, "Not exp"); 1604 1605 _sp += arg_size(); // restore stack pointer 1606 Node *x = pop_math_arg(); 1607 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x)); 1608 1609 finish_pow_exp(result, x, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1610 1611 C->set_has_split_ifs(true); // Has chance for split-if optimization 1612 1613 return true; 1614} 1615 1616//------------------------------inline_pow------------------------------------- 1617// Inline power instructions, if possible. 1618bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) { 1619 assert(id == vmIntrinsics::_dpow, "Not pow"); 1620 1621 // Pseudocode for pow 1622 // if (x <= 0.0) { 1623 // long longy = (long)y; 1624 // if ((double)longy == y) { // if y is long 1625 // if (y + 1 == y) longy = 0; // huge number: even 1626 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1627 // } else { 1628 // result = NaN; 1629 // } 1630 // } else { 1631 // result = DPow(x,y); 1632 // } 1633 // if (result != result)? { 1634 // result = uncommon_trap() or runtime_call(); 1635 // } 1636 // return result; 1637 1638 _sp += arg_size(); // restore stack pointer 1639 Node* y = pop_math_arg(); 1640 Node* x = pop_math_arg(); 1641 1642 Node* result = NULL; 1643 1644 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1645 // Short form: skip the fancy tests and just check for NaN result. 1646 result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); 1647 } else { 1648 // If this inlining ever returned NaN in the past, include all 1649 // checks + call to the runtime. 1650 1651 // Set the merge point for If node with condition of (x <= 0.0) 1652 // There are four possible paths to region node and phi node 1653 RegionNode *r = new (C, 4) RegionNode(4); 1654 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE); 1655 1656 // Build the first if node: if (x <= 0.0) 1657 // Node for 0 constant 1658 Node *zeronode = makecon(TypeD::ZERO); 1659 // Check x:0 1660 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode)); 1661 // Check: If (x<=0) then go complex path 1662 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) ); 1663 // Branch either way 1664 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1665 // Fast path taken; set region slot 3 1666 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(if1) ); 1667 r->init_req(3,fast_taken); // Capture fast-control 1668 1669 // Fast path not-taken, i.e. slow path 1670 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(if1) ); 1671 1672 // Set fast path result 1673 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) ); 1674 phi->init_req(3, fast_result); 1675 1676 // Complex path 1677 // Build the second if node (if y is long) 1678 // Node for (long)y 1679 Node *longy = _gvn.transform( new (C, 2) ConvD2LNode(y)); 1680 // Node for (double)((long) y) 1681 Node *doublelongy= _gvn.transform( new (C, 2) ConvL2DNode(longy)); 1682 // Check (double)((long) y) : y 1683 Node *cmplongy= _gvn.transform(new (C, 3) CmpDNode(doublelongy, y)); 1684 // Check if (y isn't long) then go to slow path 1685 1686 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmplongy, BoolTest::ne ) ); 1687 // Branch either way 1688 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1689 Node* ylong_path = _gvn.transform( new (C, 1) IfFalseNode(if2)); 1690 1691 Node *slow_path = _gvn.transform( new (C, 1) IfTrueNode(if2) ); 1692 1693 // Calculate DPow(abs(x), y)*(1 & (long)y) 1694 // Node for constant 1 1695 Node *conone = longcon(1); 1696 // 1& (long)y 1697 Node *signnode= _gvn.transform( new (C, 3) AndLNode(conone, longy) ); 1698 1699 // A huge number is always even. Detect a huge number by checking 1700 // if y + 1 == y and set integer to be tested for parity to 0. 1701 // Required for corner case: 1702 // (long)9.223372036854776E18 = max_jlong 1703 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1704 // max_jlong is odd but 9.223372036854776E18 is even 1705 Node* yplus1 = _gvn.transform( new (C, 3) AddDNode(y, makecon(TypeD::make(1)))); 1706 Node *cmpyplus1= _gvn.transform(new (C, 3) CmpDNode(yplus1, y)); 1707 Node *bolyplus1 = _gvn.transform( new (C, 2) BoolNode( cmpyplus1, BoolTest::eq ) ); 1708 Node* correctedsign = NULL; 1709 if (ConditionalMoveLimit != 0) { 1710 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1711 } else { 1712 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1713 RegionNode *r = new (C, 3) RegionNode(3); 1714 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); 1715 r->init_req(1, _gvn.transform( new (C, 1) IfFalseNode(ifyplus1))); 1716 r->init_req(2, _gvn.transform( new (C, 1) IfTrueNode(ifyplus1))); 1717 phi->init_req(1, signnode); 1718 phi->init_req(2, longcon(0)); 1719 correctedsign = _gvn.transform(phi); 1720 ylong_path = _gvn.transform(r); 1721 record_for_igvn(r); 1722 } 1723 1724 // zero node 1725 Node *conzero = longcon(0); 1726 // Check (1&(long)y)==0? 1727 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpLNode(correctedsign, conzero)); 1728 // Check if (1&(long)y)!=0?, if so the result is negative 1729 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) ); 1730 // abs(x) 1731 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x)); 1732 // abs(x)^y 1733 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, absx, y) ); 1734 // -abs(x)^y 1735 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy)); 1736 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1737 Node *signresult = NULL; 1738 if (ConditionalMoveLimit != 0) { 1739 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1740 } else { 1741 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1742 RegionNode *r = new (C, 3) RegionNode(3); 1743 Node *phi = new (C, 3) PhiNode(r, Type::DOUBLE); 1744 r->init_req(1, _gvn.transform( new (C, 1) IfFalseNode(ifyeven))); 1745 r->init_req(2, _gvn.transform( new (C, 1) IfTrueNode(ifyeven))); 1746 phi->init_req(1, absxpowy); 1747 phi->init_req(2, negabsxpowy); 1748 signresult = _gvn.transform(phi); 1749 ylong_path = _gvn.transform(r); 1750 record_for_igvn(r); 1751 } 1752 // Set complex path fast result 1753 r->init_req(2, ylong_path); 1754 phi->init_req(2, signresult); 1755 1756 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1757 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1758 r->init_req(1,slow_path); 1759 phi->init_req(1,slow_result); 1760 1761 // Post merge 1762 set_control(_gvn.transform(r)); 1763 record_for_igvn(r); 1764 result=_gvn.transform(phi); 1765 } 1766 1767 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1768 1769 C->set_has_split_ifs(true); // Has chance for split-if optimization 1770 1771 return true; 1772} 1773 1774//------------------------------inline_trans------------------------------------- 1775// Inline transcendental instructions, if possible. The Intel hardware gets 1776// these right, no funny corner cases missed. 1777bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) { 1778 _sp += arg_size(); // restore stack pointer 1779 Node* arg = pop_math_arg(); 1780 Node* trans = NULL; 1781 1782 switch (id) { 1783 case vmIntrinsics::_dlog: 1784 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg)); 1785 break; 1786 case vmIntrinsics::_dlog10: 1787 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg)); 1788 break; 1789 default: 1790 assert(false, "bad intrinsic was passed in"); 1791 return false; 1792 } 1793 1794 // Push result back on JVM stack 1795 push_pair(trans); 1796 return true; 1797} 1798 1799//------------------------------runtime_math----------------------------- 1800bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1801 Node* a = NULL; 1802 Node* b = NULL; 1803 1804 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1805 "must be (DD)D or (D)D type"); 1806 1807 // Inputs 1808 _sp += arg_size(); // restore stack pointer 1809 if (call_type == OptoRuntime::Math_DD_D_Type()) { 1810 b = pop_math_arg(); 1811 } 1812 a = pop_math_arg(); 1813 1814 const TypePtr* no_memory_effects = NULL; 1815 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1816 no_memory_effects, 1817 a, top(), b, b ? top() : NULL); 1818 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0)); 1819#ifdef ASSERT 1820 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1)); 1821 assert(value_top == top(), "second value must be top"); 1822#endif 1823 1824 push_pair(value); 1825 return true; 1826} 1827 1828//------------------------------inline_math_native----------------------------- 1829bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1830 switch (id) { 1831 // These intrinsics are not properly supported on all hardware 1832 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1833 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS"); 1834 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1835 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN"); 1836 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1837 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN"); 1838 1839 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) : 1840 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG"); 1841 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) : 1842 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10"); 1843 1844 // These intrinsics are supported on all hardware 1845 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false; 1846 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false; 1847 1848 case vmIntrinsics::_dexp: return 1849 Matcher::has_match_rule(Op_ExpD) ? inline_exp(id) : 1850 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1851 case vmIntrinsics::_dpow: return 1852 Matcher::has_match_rule(Op_PowD) ? inline_pow(id) : 1853 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1854 1855 // These intrinsics are not yet correctly implemented 1856 case vmIntrinsics::_datan2: 1857 return false; 1858 1859 default: 1860 ShouldNotReachHere(); 1861 return false; 1862 } 1863} 1864 1865static bool is_simple_name(Node* n) { 1866 return (n->req() == 1 // constant 1867 || (n->is_Type() && n->as_Type()->type()->singleton()) 1868 || n->is_Proj() // parameter or return value 1869 || n->is_Phi() // local of some sort 1870 ); 1871} 1872 1873//----------------------------inline_min_max----------------------------------- 1874bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1875 push(generate_min_max(id, argument(0), argument(1))); 1876 1877 return true; 1878} 1879 1880Node* 1881LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 1882 // These are the candidate return value: 1883 Node* xvalue = x0; 1884 Node* yvalue = y0; 1885 1886 if (xvalue == yvalue) { 1887 return xvalue; 1888 } 1889 1890 bool want_max = (id == vmIntrinsics::_max); 1891 1892 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 1893 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 1894 if (txvalue == NULL || tyvalue == NULL) return top(); 1895 // This is not really necessary, but it is consistent with a 1896 // hypothetical MaxINode::Value method: 1897 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 1898 1899 // %%% This folding logic should (ideally) be in a different place. 1900 // Some should be inside IfNode, and there to be a more reliable 1901 // transformation of ?: style patterns into cmoves. We also want 1902 // more powerful optimizations around cmove and min/max. 1903 1904 // Try to find a dominating comparison of these guys. 1905 // It can simplify the index computation for Arrays.copyOf 1906 // and similar uses of System.arraycopy. 1907 // First, compute the normalized version of CmpI(x, y). 1908 int cmp_op = Op_CmpI; 1909 Node* xkey = xvalue; 1910 Node* ykey = yvalue; 1911 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) ); 1912 if (ideal_cmpxy->is_Cmp()) { 1913 // E.g., if we have CmpI(length - offset, count), 1914 // it might idealize to CmpI(length, count + offset) 1915 cmp_op = ideal_cmpxy->Opcode(); 1916 xkey = ideal_cmpxy->in(1); 1917 ykey = ideal_cmpxy->in(2); 1918 } 1919 1920 // Start by locating any relevant comparisons. 1921 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 1922 Node* cmpxy = NULL; 1923 Node* cmpyx = NULL; 1924 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 1925 Node* cmp = start_from->fast_out(k); 1926 if (cmp->outcnt() > 0 && // must have prior uses 1927 cmp->in(0) == NULL && // must be context-independent 1928 cmp->Opcode() == cmp_op) { // right kind of compare 1929 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 1930 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 1931 } 1932 } 1933 1934 const int NCMPS = 2; 1935 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 1936 int cmpn; 1937 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1938 if (cmps[cmpn] != NULL) break; // find a result 1939 } 1940 if (cmpn < NCMPS) { 1941 // Look for a dominating test that tells us the min and max. 1942 int depth = 0; // Limit search depth for speed 1943 Node* dom = control(); 1944 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 1945 if (++depth >= 100) break; 1946 Node* ifproj = dom; 1947 if (!ifproj->is_Proj()) continue; 1948 Node* iff = ifproj->in(0); 1949 if (!iff->is_If()) continue; 1950 Node* bol = iff->in(1); 1951 if (!bol->is_Bool()) continue; 1952 Node* cmp = bol->in(1); 1953 if (cmp == NULL) continue; 1954 for (cmpn = 0; cmpn < NCMPS; cmpn++) 1955 if (cmps[cmpn] == cmp) break; 1956 if (cmpn == NCMPS) continue; 1957 BoolTest::mask btest = bol->as_Bool()->_test._test; 1958 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 1959 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1960 // At this point, we know that 'x btest y' is true. 1961 switch (btest) { 1962 case BoolTest::eq: 1963 // They are proven equal, so we can collapse the min/max. 1964 // Either value is the answer. Choose the simpler. 1965 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 1966 return yvalue; 1967 return xvalue; 1968 case BoolTest::lt: // x < y 1969 case BoolTest::le: // x <= y 1970 return (want_max ? yvalue : xvalue); 1971 case BoolTest::gt: // x > y 1972 case BoolTest::ge: // x >= y 1973 return (want_max ? xvalue : yvalue); 1974 } 1975 } 1976 } 1977 1978 // We failed to find a dominating test. 1979 // Let's pick a test that might GVN with prior tests. 1980 Node* best_bol = NULL; 1981 BoolTest::mask best_btest = BoolTest::illegal; 1982 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 1983 Node* cmp = cmps[cmpn]; 1984 if (cmp == NULL) continue; 1985 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 1986 Node* bol = cmp->fast_out(j); 1987 if (!bol->is_Bool()) continue; 1988 BoolTest::mask btest = bol->as_Bool()->_test._test; 1989 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 1990 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 1991 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 1992 best_bol = bol->as_Bool(); 1993 best_btest = btest; 1994 } 1995 } 1996 } 1997 1998 Node* answer_if_true = NULL; 1999 Node* answer_if_false = NULL; 2000 switch (best_btest) { 2001 default: 2002 if (cmpxy == NULL) 2003 cmpxy = ideal_cmpxy; 2004 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) ); 2005 // and fall through: 2006 case BoolTest::lt: // x < y 2007 case BoolTest::le: // x <= y 2008 answer_if_true = (want_max ? yvalue : xvalue); 2009 answer_if_false = (want_max ? xvalue : yvalue); 2010 break; 2011 case BoolTest::gt: // x > y 2012 case BoolTest::ge: // x >= y 2013 answer_if_true = (want_max ? xvalue : yvalue); 2014 answer_if_false = (want_max ? yvalue : xvalue); 2015 break; 2016 } 2017 2018 jint hi, lo; 2019 if (want_max) { 2020 // We can sharpen the minimum. 2021 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2022 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2023 } else { 2024 // We can sharpen the maximum. 2025 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2026 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2027 } 2028 2029 // Use a flow-free graph structure, to avoid creating excess control edges 2030 // which could hinder other optimizations. 2031 // Since Math.min/max is often used with arraycopy, we want 2032 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2033 Node* cmov = CMoveNode::make(C, NULL, best_bol, 2034 answer_if_false, answer_if_true, 2035 TypeInt::make(lo, hi, widen)); 2036 2037 return _gvn.transform(cmov); 2038 2039 /* 2040 // This is not as desirable as it may seem, since Min and Max 2041 // nodes do not have a full set of optimizations. 2042 // And they would interfere, anyway, with 'if' optimizations 2043 // and with CMoveI canonical forms. 2044 switch (id) { 2045 case vmIntrinsics::_min: 2046 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2047 case vmIntrinsics::_max: 2048 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2049 default: 2050 ShouldNotReachHere(); 2051 } 2052 */ 2053} 2054 2055inline int 2056LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2057 const TypePtr* base_type = TypePtr::NULL_PTR; 2058 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2059 if (base_type == NULL) { 2060 // Unknown type. 2061 return Type::AnyPtr; 2062 } else if (base_type == TypePtr::NULL_PTR) { 2063 // Since this is a NULL+long form, we have to switch to a rawptr. 2064 base = _gvn.transform( new (C, 2) CastX2PNode(offset) ); 2065 offset = MakeConX(0); 2066 return Type::RawPtr; 2067 } else if (base_type->base() == Type::RawPtr) { 2068 return Type::RawPtr; 2069 } else if (base_type->isa_oopptr()) { 2070 // Base is never null => always a heap address. 2071 if (base_type->ptr() == TypePtr::NotNull) { 2072 return Type::OopPtr; 2073 } 2074 // Offset is small => always a heap address. 2075 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2076 if (offset_type != NULL && 2077 base_type->offset() == 0 && // (should always be?) 2078 offset_type->_lo >= 0 && 2079 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2080 return Type::OopPtr; 2081 } 2082 // Otherwise, it might either be oop+off or NULL+addr. 2083 return Type::AnyPtr; 2084 } else { 2085 // No information: 2086 return Type::AnyPtr; 2087 } 2088} 2089 2090inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2091 int kind = classify_unsafe_addr(base, offset); 2092 if (kind == Type::RawPtr) { 2093 return basic_plus_adr(top(), base, offset); 2094 } else { 2095 return basic_plus_adr(base, offset); 2096 } 2097} 2098 2099//-------------------inline_numberOfLeadingZeros_int/long----------------------- 2100// inline int Integer.numberOfLeadingZeros(int) 2101// inline int Long.numberOfLeadingZeros(long) 2102bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) { 2103 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros"); 2104 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false; 2105 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false; 2106 _sp += arg_size(); // restore stack pointer 2107 switch (id) { 2108 case vmIntrinsics::_numberOfLeadingZeros_i: 2109 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop()))); 2110 break; 2111 case vmIntrinsics::_numberOfLeadingZeros_l: 2112 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair()))); 2113 break; 2114 default: 2115 ShouldNotReachHere(); 2116 } 2117 return true; 2118} 2119 2120//-------------------inline_numberOfTrailingZeros_int/long---------------------- 2121// inline int Integer.numberOfTrailingZeros(int) 2122// inline int Long.numberOfTrailingZeros(long) 2123bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) { 2124 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros"); 2125 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false; 2126 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false; 2127 _sp += arg_size(); // restore stack pointer 2128 switch (id) { 2129 case vmIntrinsics::_numberOfTrailingZeros_i: 2130 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop()))); 2131 break; 2132 case vmIntrinsics::_numberOfTrailingZeros_l: 2133 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair()))); 2134 break; 2135 default: 2136 ShouldNotReachHere(); 2137 } 2138 return true; 2139} 2140 2141//----------------------------inline_bitCount_int/long----------------------- 2142// inline int Integer.bitCount(int) 2143// inline int Long.bitCount(long) 2144bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) { 2145 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount"); 2146 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false; 2147 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false; 2148 _sp += arg_size(); // restore stack pointer 2149 switch (id) { 2150 case vmIntrinsics::_bitCount_i: 2151 push(_gvn.transform(new (C, 2) PopCountINode(pop()))); 2152 break; 2153 case vmIntrinsics::_bitCount_l: 2154 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair()))); 2155 break; 2156 default: 2157 ShouldNotReachHere(); 2158 } 2159 return true; 2160} 2161 2162//----------------------------inline_reverseBytes_int/long/char/short------------------- 2163// inline Integer.reverseBytes(int) 2164// inline Long.reverseBytes(long) 2165// inline Character.reverseBytes(char) 2166// inline Short.reverseBytes(short) 2167bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) { 2168 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l || 2169 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s, 2170 "not reverse Bytes"); 2171 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false; 2172 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false; 2173 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false; 2174 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false; 2175 _sp += arg_size(); // restore stack pointer 2176 switch (id) { 2177 case vmIntrinsics::_reverseBytes_i: 2178 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop()))); 2179 break; 2180 case vmIntrinsics::_reverseBytes_l: 2181 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair()))); 2182 break; 2183 case vmIntrinsics::_reverseBytes_c: 2184 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop()))); 2185 break; 2186 case vmIntrinsics::_reverseBytes_s: 2187 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop()))); 2188 break; 2189 default: 2190 ; 2191 } 2192 return true; 2193} 2194 2195//----------------------------inline_unsafe_access---------------------------- 2196 2197const static BasicType T_ADDRESS_HOLDER = T_LONG; 2198 2199// Helper that guards and inserts a pre-barrier. 2200void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2201 Node* pre_val, int nargs, bool need_mem_bar) { 2202 // We could be accessing the referent field of a reference object. If so, when G1 2203 // is enabled, we need to log the value in the referent field in an SATB buffer. 2204 // This routine performs some compile time filters and generates suitable 2205 // runtime filters that guard the pre-barrier code. 2206 // Also add memory barrier for non volatile load from the referent field 2207 // to prevent commoning of loads across safepoint. 2208 if (!UseG1GC && !need_mem_bar) 2209 return; 2210 2211 // Some compile time checks. 2212 2213 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2214 const TypeX* otype = offset->find_intptr_t_type(); 2215 if (otype != NULL && otype->is_con() && 2216 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2217 // Constant offset but not the reference_offset so just return 2218 return; 2219 } 2220 2221 // We only need to generate the runtime guards for instances. 2222 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2223 if (btype != NULL) { 2224 if (btype->isa_aryptr()) { 2225 // Array type so nothing to do 2226 return; 2227 } 2228 2229 const TypeInstPtr* itype = btype->isa_instptr(); 2230 if (itype != NULL) { 2231 // Can the klass of base_oop be statically determined to be 2232 // _not_ a sub-class of Reference and _not_ Object? 2233 ciKlass* klass = itype->klass(); 2234 if ( klass->is_loaded() && 2235 !klass->is_subtype_of(env()->Reference_klass()) && 2236 !env()->Object_klass()->is_subtype_of(klass)) { 2237 return; 2238 } 2239 } 2240 } 2241 2242 // The compile time filters did not reject base_oop/offset so 2243 // we need to generate the following runtime filters 2244 // 2245 // if (offset == java_lang_ref_Reference::_reference_offset) { 2246 // if (instance_of(base, java.lang.ref.Reference)) { 2247 // pre_barrier(_, pre_val, ...); 2248 // } 2249 // } 2250 2251 float likely = PROB_LIKELY(0.999); 2252 float unlikely = PROB_UNLIKELY(0.999); 2253 2254 IdealKit ideal(this); 2255#define __ ideal. 2256 2257 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2258 2259 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2260 // Update graphKit memory and control from IdealKit. 2261 sync_kit(ideal); 2262 2263 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2264 _sp += nargs; // gen_instanceof might do an uncommon trap 2265 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2266 _sp -= nargs; 2267 2268 // Update IdealKit memory and control from graphKit. 2269 __ sync_kit(this); 2270 2271 Node* one = __ ConI(1); 2272 // is_instof == 0 if base_oop == NULL 2273 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2274 2275 // Update graphKit from IdeakKit. 2276 sync_kit(ideal); 2277 2278 // Use the pre-barrier to record the value in the referent field 2279 pre_barrier(false /* do_load */, 2280 __ ctrl(), 2281 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2282 pre_val /* pre_val */, 2283 T_OBJECT); 2284 if (need_mem_bar) { 2285 // Add memory barrier to prevent commoning reads from this field 2286 // across safepoint since GC can change its value. 2287 insert_mem_bar(Op_MemBarCPUOrder); 2288 } 2289 // Update IdealKit from graphKit. 2290 __ sync_kit(this); 2291 2292 } __ end_if(); // _ref_type != ref_none 2293 } __ end_if(); // offset == referent_offset 2294 2295 // Final sync IdealKit and GraphKit. 2296 final_sync(ideal); 2297#undef __ 2298} 2299 2300 2301// Interpret Unsafe.fieldOffset cookies correctly: 2302extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2303 2304bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2305 if (callee()->is_static()) return false; // caller must have the capability! 2306 2307#ifndef PRODUCT 2308 { 2309 ResourceMark rm; 2310 // Check the signatures. 2311 ciSignature* sig = signature(); 2312#ifdef ASSERT 2313 if (!is_store) { 2314 // Object getObject(Object base, int/long offset), etc. 2315 BasicType rtype = sig->return_type()->basic_type(); 2316 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2317 rtype = T_ADDRESS; // it is really a C void* 2318 assert(rtype == type, "getter must return the expected value"); 2319 if (!is_native_ptr) { 2320 assert(sig->count() == 2, "oop getter has 2 arguments"); 2321 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2322 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2323 } else { 2324 assert(sig->count() == 1, "native getter has 1 argument"); 2325 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2326 } 2327 } else { 2328 // void putObject(Object base, int/long offset, Object x), etc. 2329 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2330 if (!is_native_ptr) { 2331 assert(sig->count() == 3, "oop putter has 3 arguments"); 2332 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2333 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2334 } else { 2335 assert(sig->count() == 2, "native putter has 2 arguments"); 2336 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2337 } 2338 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2339 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2340 vtype = T_ADDRESS; // it is really a C void* 2341 assert(vtype == type, "putter must accept the expected value"); 2342 } 2343#endif // ASSERT 2344 } 2345#endif //PRODUCT 2346 2347 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2348 2349 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ]; 2350 2351 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words 2352 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0); 2353 assert(callee()->arg_size() == nargs, "must be"); 2354 2355 debug_only(int saved_sp = _sp); 2356 _sp += nargs; 2357 2358 Node* val; 2359 debug_only(val = (Node*)(uintptr_t)-1); 2360 2361 2362 if (is_store) { 2363 // Get the value being stored. (Pop it first; it was pushed last.) 2364 switch (type) { 2365 case T_DOUBLE: 2366 case T_LONG: 2367 case T_ADDRESS: 2368 val = pop_pair(); 2369 break; 2370 default: 2371 val = pop(); 2372 } 2373 } 2374 2375 // Build address expression. See the code in inline_unsafe_prefetch. 2376 Node *adr; 2377 Node *heap_base_oop = top(); 2378 Node* offset = top(); 2379 2380 if (!is_native_ptr) { 2381 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2382 offset = pop_pair(); 2383 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2384 Node* base = pop(); 2385 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2386 // to be plain byte offsets, which are also the same as those accepted 2387 // by oopDesc::field_base. 2388 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2389 "fieldOffset must be byte-scaled"); 2390 // 32-bit machines ignore the high half! 2391 offset = ConvL2X(offset); 2392 adr = make_unsafe_address(base, offset); 2393 heap_base_oop = base; 2394 } else { 2395 Node* ptr = pop_pair(); 2396 // Adjust Java long to machine word: 2397 ptr = ConvL2X(ptr); 2398 adr = make_unsafe_address(NULL, ptr); 2399 } 2400 2401 // Pop receiver last: it was pushed first. 2402 Node *receiver = pop(); 2403 2404 assert(saved_sp == _sp, "must have correct argument count"); 2405 2406 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2407 2408 // First guess at the value type. 2409 const Type *value_type = Type::get_const_basic_type(type); 2410 2411 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2412 // there was not enough information to nail it down. 2413 Compile::AliasType* alias_type = C->alias_type(adr_type); 2414 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2415 2416 // We will need memory barriers unless we can determine a unique 2417 // alias category for this reference. (Note: If for some reason 2418 // the barriers get omitted and the unsafe reference begins to "pollute" 2419 // the alias analysis of the rest of the graph, either Compile::can_alias 2420 // or Compile::must_alias will throw a diagnostic assert.) 2421 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2422 2423 // If we are reading the value of the referent field of a Reference 2424 // object (either by using Unsafe directly or through reflection) 2425 // then, if G1 is enabled, we need to record the referent in an 2426 // SATB log buffer using the pre-barrier mechanism. 2427 // Also we need to add memory barrier to prevent commoning reads 2428 // from this field across safepoint since GC can change its value. 2429 bool need_read_barrier = !is_native_ptr && !is_store && 2430 offset != top() && heap_base_oop != top(); 2431 2432 if (!is_store && type == T_OBJECT) { 2433 // Attempt to infer a sharper value type from the offset and base type. 2434 ciKlass* sharpened_klass = NULL; 2435 2436 // See if it is an instance field, with an object type. 2437 if (alias_type->field() != NULL) { 2438 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2439 if (alias_type->field()->type()->is_klass()) { 2440 sharpened_klass = alias_type->field()->type()->as_klass(); 2441 } 2442 } 2443 2444 // See if it is a narrow oop array. 2445 if (adr_type->isa_aryptr()) { 2446 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2447 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2448 if (elem_type != NULL) { 2449 sharpened_klass = elem_type->klass(); 2450 } 2451 } 2452 } 2453 2454 if (sharpened_klass != NULL) { 2455 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2456 2457 // Sharpen the value type. 2458 value_type = tjp; 2459 2460#ifndef PRODUCT 2461 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 2462 tty->print(" from base type: "); adr_type->dump(); 2463 tty->print(" sharpened value: "); value_type->dump(); 2464 } 2465#endif 2466 } 2467 } 2468 2469 // Null check on self without removing any arguments. The argument 2470 // null check technically happens in the wrong place, which can lead to 2471 // invalid stack traces when the primitive is inlined into a method 2472 // which handles NullPointerExceptions. 2473 _sp += nargs; 2474 do_null_check(receiver, T_OBJECT); 2475 _sp -= nargs; 2476 if (stopped()) { 2477 return true; 2478 } 2479 // Heap pointers get a null-check from the interpreter, 2480 // as a courtesy. However, this is not guaranteed by Unsafe, 2481 // and it is not possible to fully distinguish unintended nulls 2482 // from intended ones in this API. 2483 2484 if (is_volatile) { 2485 // We need to emit leading and trailing CPU membars (see below) in 2486 // addition to memory membars when is_volatile. This is a little 2487 // too strong, but avoids the need to insert per-alias-type 2488 // volatile membars (for stores; compare Parse::do_put_xxx), which 2489 // we cannot do effectively here because we probably only have a 2490 // rough approximation of type. 2491 need_mem_bar = true; 2492 // For Stores, place a memory ordering barrier now. 2493 if (is_store) 2494 insert_mem_bar(Op_MemBarRelease); 2495 } 2496 2497 // Memory barrier to prevent normal and 'unsafe' accesses from 2498 // bypassing each other. Happens after null checks, so the 2499 // exception paths do not take memory state from the memory barrier, 2500 // so there's no problems making a strong assert about mixing users 2501 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2502 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2503 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2504 2505 if (!is_store) { 2506 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile); 2507 // load value and push onto stack 2508 switch (type) { 2509 case T_BOOLEAN: 2510 case T_CHAR: 2511 case T_BYTE: 2512 case T_SHORT: 2513 case T_INT: 2514 case T_FLOAT: 2515 push(p); 2516 break; 2517 case T_OBJECT: 2518 if (need_read_barrier) { 2519 insert_pre_barrier(heap_base_oop, offset, p, nargs, !(is_volatile || need_mem_bar)); 2520 } 2521 push(p); 2522 break; 2523 case T_ADDRESS: 2524 // Cast to an int type. 2525 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) ); 2526 p = ConvX2L(p); 2527 push_pair(p); 2528 break; 2529 case T_DOUBLE: 2530 case T_LONG: 2531 push_pair( p ); 2532 break; 2533 default: ShouldNotReachHere(); 2534 } 2535 } else { 2536 // place effect of store into memory 2537 switch (type) { 2538 case T_DOUBLE: 2539 val = dstore_rounding(val); 2540 break; 2541 case T_ADDRESS: 2542 // Repackage the long as a pointer. 2543 val = ConvL2X(val); 2544 val = _gvn.transform( new (C, 2) CastX2PNode(val) ); 2545 break; 2546 } 2547 2548 if (type != T_OBJECT ) { 2549 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile); 2550 } else { 2551 // Possibly an oop being stored to Java heap or native memory 2552 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2553 // oop to Java heap. 2554 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2555 } else { 2556 // We can't tell at compile time if we are storing in the Java heap or outside 2557 // of it. So we need to emit code to conditionally do the proper type of 2558 // store. 2559 2560 IdealKit ideal(this); 2561#define __ ideal. 2562 // QQQ who knows what probability is here?? 2563 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2564 // Sync IdealKit and graphKit. 2565 sync_kit(ideal); 2566 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type); 2567 // Update IdealKit memory. 2568 __ sync_kit(this); 2569 } __ else_(); { 2570 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile); 2571 } __ end_if(); 2572 // Final sync IdealKit and GraphKit. 2573 final_sync(ideal); 2574#undef __ 2575 } 2576 } 2577 } 2578 2579 if (is_volatile) { 2580 if (!is_store) 2581 insert_mem_bar(Op_MemBarAcquire); 2582 else 2583 insert_mem_bar(Op_MemBarVolatile); 2584 } 2585 2586 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2587 2588 return true; 2589} 2590 2591//----------------------------inline_unsafe_prefetch---------------------------- 2592 2593bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2594#ifndef PRODUCT 2595 { 2596 ResourceMark rm; 2597 // Check the signatures. 2598 ciSignature* sig = signature(); 2599#ifdef ASSERT 2600 // Object getObject(Object base, int/long offset), etc. 2601 BasicType rtype = sig->return_type()->basic_type(); 2602 if (!is_native_ptr) { 2603 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2604 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2605 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2606 } else { 2607 assert(sig->count() == 1, "native prefetch has 1 argument"); 2608 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2609 } 2610#endif // ASSERT 2611 } 2612#endif // !PRODUCT 2613 2614 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2615 2616 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args 2617 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3); 2618 2619 debug_only(int saved_sp = _sp); 2620 _sp += nargs; 2621 2622 // Build address expression. See the code in inline_unsafe_access. 2623 Node *adr; 2624 if (!is_native_ptr) { 2625 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2626 Node* offset = pop_pair(); 2627 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2628 Node* base = pop(); 2629 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2630 // to be plain byte offsets, which are also the same as those accepted 2631 // by oopDesc::field_base. 2632 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2633 "fieldOffset must be byte-scaled"); 2634 // 32-bit machines ignore the high half! 2635 offset = ConvL2X(offset); 2636 adr = make_unsafe_address(base, offset); 2637 } else { 2638 Node* ptr = pop_pair(); 2639 // Adjust Java long to machine word: 2640 ptr = ConvL2X(ptr); 2641 adr = make_unsafe_address(NULL, ptr); 2642 } 2643 2644 if (is_static) { 2645 assert(saved_sp == _sp, "must have correct argument count"); 2646 } else { 2647 // Pop receiver last: it was pushed first. 2648 Node *receiver = pop(); 2649 assert(saved_sp == _sp, "must have correct argument count"); 2650 2651 // Null check on self without removing any arguments. The argument 2652 // null check technically happens in the wrong place, which can lead to 2653 // invalid stack traces when the primitive is inlined into a method 2654 // which handles NullPointerExceptions. 2655 _sp += nargs; 2656 do_null_check(receiver, T_OBJECT); 2657 _sp -= nargs; 2658 if (stopped()) { 2659 return true; 2660 } 2661 } 2662 2663 // Generate the read or write prefetch 2664 Node *prefetch; 2665 if (is_store) { 2666 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr); 2667 } else { 2668 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr); 2669 } 2670 prefetch->init_req(0, control()); 2671 set_i_o(_gvn.transform(prefetch)); 2672 2673 return true; 2674} 2675 2676//----------------------------inline_unsafe_CAS---------------------------- 2677 2678bool LibraryCallKit::inline_unsafe_CAS(BasicType type) { 2679 // This basic scheme here is the same as inline_unsafe_access, but 2680 // differs in enough details that combining them would make the code 2681 // overly confusing. (This is a true fact! I originally combined 2682 // them, but even I was confused by it!) As much code/comments as 2683 // possible are retained from inline_unsafe_access though to make 2684 // the correspondences clearer. - dl 2685 2686 if (callee()->is_static()) return false; // caller must have the capability! 2687 2688#ifndef PRODUCT 2689 { 2690 ResourceMark rm; 2691 // Check the signatures. 2692 ciSignature* sig = signature(); 2693#ifdef ASSERT 2694 BasicType rtype = sig->return_type()->basic_type(); 2695 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2696 assert(sig->count() == 4, "CAS has 4 arguments"); 2697 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2698 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2699#endif // ASSERT 2700 } 2701#endif //PRODUCT 2702 2703 // number of stack slots per value argument (1 or 2) 2704 int type_words = type2size[type]; 2705 2706 // Cannot inline wide CAS on machines that don't support it natively 2707 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8()) 2708 return false; 2709 2710 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2711 2712 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue; 2713 int nargs = 1 + 1 + 2 + type_words + type_words; 2714 2715 // pop arguments: newval, oldval, offset, base, and receiver 2716 debug_only(int saved_sp = _sp); 2717 _sp += nargs; 2718 Node* newval = (type_words == 1) ? pop() : pop_pair(); 2719 Node* oldval = (type_words == 1) ? pop() : pop_pair(); 2720 Node *offset = pop_pair(); 2721 Node *base = pop(); 2722 Node *receiver = pop(); 2723 assert(saved_sp == _sp, "must have correct argument count"); 2724 2725 // Null check receiver. 2726 _sp += nargs; 2727 do_null_check(receiver, T_OBJECT); 2728 _sp -= nargs; 2729 if (stopped()) { 2730 return true; 2731 } 2732 2733 // Build field offset expression. 2734 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2735 // to be plain byte offsets, which are also the same as those accepted 2736 // by oopDesc::field_base. 2737 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2738 // 32-bit machines ignore the high half of long offsets 2739 offset = ConvL2X(offset); 2740 Node* adr = make_unsafe_address(base, offset); 2741 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2742 2743 // (Unlike inline_unsafe_access, there seems no point in trying 2744 // to refine types. Just use the coarse types here. 2745 const Type *value_type = Type::get_const_basic_type(type); 2746 Compile::AliasType* alias_type = C->alias_type(adr_type); 2747 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2748 int alias_idx = C->get_alias_index(adr_type); 2749 2750 // Memory-model-wise, a CAS acts like a little synchronized block, 2751 // so needs barriers on each side. These don't translate into 2752 // actual barriers on most machines, but we still need rest of 2753 // compiler to respect ordering. 2754 2755 insert_mem_bar(Op_MemBarRelease); 2756 insert_mem_bar(Op_MemBarCPUOrder); 2757 2758 // 4984716: MemBars must be inserted before this 2759 // memory node in order to avoid a false 2760 // dependency which will confuse the scheduler. 2761 Node *mem = memory(alias_idx); 2762 2763 // For now, we handle only those cases that actually exist: ints, 2764 // longs, and Object. Adding others should be straightforward. 2765 Node* cas; 2766 switch(type) { 2767 case T_INT: 2768 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2769 break; 2770 case T_LONG: 2771 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2772 break; 2773 case T_OBJECT: 2774 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2775 // could be delayed during Parse (for example, in adjust_map_after_if()). 2776 // Execute transformation here to avoid barrier generation in such case. 2777 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2778 newval = _gvn.makecon(TypePtr::NULL_PTR); 2779 2780 // Reference stores need a store barrier. 2781 // (They don't if CAS fails, but it isn't worth checking.) 2782 pre_barrier(true /* do_load*/, 2783 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2784 NULL /* pre_val*/, 2785 T_OBJECT); 2786#ifdef _LP64 2787 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2788 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2789 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2790 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr, 2791 newval_enc, oldval_enc)); 2792 } else 2793#endif 2794 { 2795 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2796 } 2797 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true); 2798 break; 2799 default: 2800 ShouldNotReachHere(); 2801 break; 2802 } 2803 2804 // SCMemProjNodes represent the memory state of CAS. Their main 2805 // role is to prevent CAS nodes from being optimized away when their 2806 // results aren't used. 2807 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas)); 2808 set_memory(proj, alias_idx); 2809 2810 // Add the trailing membar surrounding the access 2811 insert_mem_bar(Op_MemBarCPUOrder); 2812 insert_mem_bar(Op_MemBarAcquire); 2813 2814 push(cas); 2815 return true; 2816} 2817 2818bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 2819 // This is another variant of inline_unsafe_access, differing in 2820 // that it always issues store-store ("release") barrier and ensures 2821 // store-atomicity (which only matters for "long"). 2822 2823 if (callee()->is_static()) return false; // caller must have the capability! 2824 2825#ifndef PRODUCT 2826 { 2827 ResourceMark rm; 2828 // Check the signatures. 2829 ciSignature* sig = signature(); 2830#ifdef ASSERT 2831 BasicType rtype = sig->return_type()->basic_type(); 2832 assert(rtype == T_VOID, "must return void"); 2833 assert(sig->count() == 3, "has 3 arguments"); 2834 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 2835 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 2836#endif // ASSERT 2837 } 2838#endif //PRODUCT 2839 2840 // number of stack slots per value argument (1 or 2) 2841 int type_words = type2size[type]; 2842 2843 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2844 2845 // Argument words: "this" plus oop plus offset plus value; 2846 int nargs = 1 + 1 + 2 + type_words; 2847 2848 // pop arguments: val, offset, base, and receiver 2849 debug_only(int saved_sp = _sp); 2850 _sp += nargs; 2851 Node* val = (type_words == 1) ? pop() : pop_pair(); 2852 Node *offset = pop_pair(); 2853 Node *base = pop(); 2854 Node *receiver = pop(); 2855 assert(saved_sp == _sp, "must have correct argument count"); 2856 2857 // Null check receiver. 2858 _sp += nargs; 2859 do_null_check(receiver, T_OBJECT); 2860 _sp -= nargs; 2861 if (stopped()) { 2862 return true; 2863 } 2864 2865 // Build field offset expression. 2866 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2867 // 32-bit machines ignore the high half of long offsets 2868 offset = ConvL2X(offset); 2869 Node* adr = make_unsafe_address(base, offset); 2870 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2871 const Type *value_type = Type::get_const_basic_type(type); 2872 Compile::AliasType* alias_type = C->alias_type(adr_type); 2873 2874 insert_mem_bar(Op_MemBarRelease); 2875 insert_mem_bar(Op_MemBarCPUOrder); 2876 // Ensure that the store is atomic for longs: 2877 bool require_atomic_access = true; 2878 Node* store; 2879 if (type == T_OBJECT) // reference stores need a store barrier. 2880 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type); 2881 else { 2882 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access); 2883 } 2884 insert_mem_bar(Op_MemBarCPUOrder); 2885 return true; 2886} 2887 2888bool LibraryCallKit::inline_unsafe_allocate() { 2889 if (callee()->is_static()) return false; // caller must have the capability! 2890 int nargs = 1 + 1; 2891 assert(signature()->size() == nargs-1, "alloc has 1 argument"); 2892 null_check_receiver(callee()); // check then ignore argument(0) 2893 _sp += nargs; // set original stack for use by uncommon_trap 2894 Node* cls = do_null_check(argument(1), T_OBJECT); 2895 _sp -= nargs; 2896 if (stopped()) return true; 2897 2898 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 2899 _sp += nargs; // set original stack for use by uncommon_trap 2900 kls = do_null_check(kls, T_OBJECT); 2901 _sp -= nargs; 2902 if (stopped()) return true; // argument was like int.class 2903 2904 // Note: The argument might still be an illegal value like 2905 // Serializable.class or Object[].class. The runtime will handle it. 2906 // But we must make an explicit check for initialization. 2907 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 2908 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 2909 // can generate code to load it as unsigned byte. 2910 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN); 2911 Node* bits = intcon(InstanceKlass::fully_initialized); 2912 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) ); 2913 // The 'test' is non-zero if we need to take a slow path. 2914 2915 Node* obj = new_instance(kls, test); 2916 push(obj); 2917 2918 return true; 2919} 2920 2921#ifdef TRACE_HAVE_INTRINSICS 2922/* 2923 * oop -> myklass 2924 * myklass->trace_id |= USED 2925 * return myklass->trace_id & ~0x3 2926 */ 2927bool LibraryCallKit::inline_native_classID() { 2928 int nargs = 1 + 1; 2929 null_check_receiver(callee()); // check then ignore argument(0) 2930 _sp += nargs; 2931 Node* cls = do_null_check(argument(1), T_OBJECT); 2932 _sp -= nargs; 2933 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0); 2934 _sp += nargs; 2935 kls = do_null_check(kls, T_OBJECT); 2936 _sp -= nargs; 2937 ByteSize offset = TRACE_ID_OFFSET; 2938 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 2939 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG); 2940 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 2941 Node* andl = _gvn.transform(new (C, 3) AndLNode(tvalue, bits)); 2942 Node* clsused = longcon(0x01l); // set the class bit 2943 Node* orl = _gvn.transform(new (C, 3) OrLNode(tvalue, clsused)); 2944 2945 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 2946 store_to_memory(control(), insp, orl, T_LONG, adr_type); 2947 push_pair(andl); 2948 return true; 2949} 2950 2951bool LibraryCallKit::inline_native_threadID() { 2952 Node* tls_ptr = NULL; 2953 Node* cur_thr = generate_current_thread(tls_ptr); 2954 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 2955 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 2956 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 2957 2958 Node* threadid = NULL; 2959 size_t thread_id_size = OSThread::thread_id_size(); 2960 if (thread_id_size == (size_t) BytesPerLong) { 2961 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG)); 2962 push(threadid); 2963 } else if (thread_id_size == (size_t) BytesPerInt) { 2964 threadid = make_load(control(), p, TypeInt::INT, T_INT); 2965 push(threadid); 2966 } else { 2967 ShouldNotReachHere(); 2968 } 2969 return true; 2970} 2971#endif 2972 2973//------------------------inline_native_time_funcs-------------- 2974// inline code for System.currentTimeMillis() and System.nanoTime() 2975// these have the same type and signature 2976bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 2977 const TypeFunc *tf = OptoRuntime::void_long_Type(); 2978 const TypePtr* no_memory_effects = NULL; 2979 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 2980 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0)); 2981#ifdef ASSERT 2982 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1)); 2983 assert(value_top == top(), "second value must be top"); 2984#endif 2985 push_pair(value); 2986 return true; 2987} 2988 2989//------------------------inline_native_currentThread------------------ 2990bool LibraryCallKit::inline_native_currentThread() { 2991 Node* junk = NULL; 2992 push(generate_current_thread(junk)); 2993 return true; 2994} 2995 2996//------------------------inline_native_isInterrupted------------------ 2997bool LibraryCallKit::inline_native_isInterrupted() { 2998 const int nargs = 1+1; // receiver + boolean 2999 assert(nargs == arg_size(), "sanity"); 3000 // Add a fast path to t.isInterrupted(clear_int): 3001 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int)) 3002 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3003 // So, in the common case that the interrupt bit is false, 3004 // we avoid making a call into the VM. Even if the interrupt bit 3005 // is true, if the clear_int argument is false, we avoid the VM call. 3006 // However, if the receiver is not currentThread, we must call the VM, 3007 // because there must be some locking done around the operation. 3008 3009 // We only go to the fast case code if we pass two guards. 3010 // Paths which do not pass are accumulated in the slow_region. 3011 RegionNode* slow_region = new (C, 1) RegionNode(1); 3012 record_for_igvn(slow_region); 3013 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow 3014 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL); 3015 enum { no_int_result_path = 1, 3016 no_clear_result_path = 2, 3017 slow_result_path = 3 3018 }; 3019 3020 // (a) Receiving thread must be the current thread. 3021 Node* rec_thr = argument(0); 3022 Node* tls_ptr = NULL; 3023 Node* cur_thr = generate_current_thread(tls_ptr); 3024 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) ); 3025 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) ); 3026 3027 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO); 3028 if (!known_current_thread) 3029 generate_slow_guard(bol_thr, slow_region); 3030 3031 // (b) Interrupt bit on TLS must be false. 3032 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3033 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS); 3034 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3035 // Set the control input on the field _interrupted read to prevent it floating up. 3036 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT); 3037 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) ); 3038 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) ); 3039 3040 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3041 3042 // First fast path: if (!TLS._interrupted) return false; 3043 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) ); 3044 result_rgn->init_req(no_int_result_path, false_bit); 3045 result_val->init_req(no_int_result_path, intcon(0)); 3046 3047 // drop through to next case 3048 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) ); 3049 3050 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3051 Node* clr_arg = argument(1); 3052 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) ); 3053 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) ); 3054 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3055 3056 // Second fast path: ... else if (!clear_int) return true; 3057 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) ); 3058 result_rgn->init_req(no_clear_result_path, false_arg); 3059 result_val->init_req(no_clear_result_path, intcon(1)); 3060 3061 // drop through to next case 3062 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) ); 3063 3064 // (d) Otherwise, go to the slow path. 3065 slow_region->add_req(control()); 3066 set_control( _gvn.transform(slow_region) ); 3067 3068 if (stopped()) { 3069 // There is no slow path. 3070 result_rgn->init_req(slow_result_path, top()); 3071 result_val->init_req(slow_result_path, top()); 3072 } else { 3073 // non-virtual because it is a private non-static 3074 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3075 3076 Node* slow_val = set_results_for_java_call(slow_call); 3077 // this->control() comes from set_results_for_java_call 3078 3079 // If we know that the result of the slow call will be true, tell the optimizer! 3080 if (known_current_thread) slow_val = intcon(1); 3081 3082 Node* fast_io = slow_call->in(TypeFunc::I_O); 3083 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3084 // These two phis are pre-filled with copies of of the fast IO and Memory 3085 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3086 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3087 3088 result_rgn->init_req(slow_result_path, control()); 3089 io_phi ->init_req(slow_result_path, i_o()); 3090 mem_phi ->init_req(slow_result_path, reset_memory()); 3091 result_val->init_req(slow_result_path, slow_val); 3092 3093 set_all_memory( _gvn.transform(mem_phi) ); 3094 set_i_o( _gvn.transform(io_phi) ); 3095 } 3096 3097 push_result(result_rgn, result_val); 3098 C->set_has_split_ifs(true); // Has chance for split-if optimization 3099 3100 return true; 3101} 3102 3103//---------------------------load_mirror_from_klass---------------------------- 3104// Given a klass oop, load its java mirror (a java.lang.Class oop). 3105Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3106 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3107 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT); 3108} 3109 3110//-----------------------load_klass_from_mirror_common------------------------- 3111// Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3112// Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3113// and branch to the given path on the region. 3114// If never_see_null, take an uncommon trap on null, so we can optimistically 3115// compile for the non-null case. 3116// If the region is NULL, force never_see_null = true. 3117Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3118 bool never_see_null, 3119 int nargs, 3120 RegionNode* region, 3121 int null_path, 3122 int offset) { 3123 if (region == NULL) never_see_null = true; 3124 Node* p = basic_plus_adr(mirror, offset); 3125 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3126 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) ); 3127 _sp += nargs; // any deopt will start just before call to enclosing method 3128 Node* null_ctl = top(); 3129 kls = null_check_oop(kls, &null_ctl, never_see_null); 3130 if (region != NULL) { 3131 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3132 region->init_req(null_path, null_ctl); 3133 } else { 3134 assert(null_ctl == top(), "no loose ends"); 3135 } 3136 _sp -= nargs; 3137 return kls; 3138} 3139 3140//--------------------(inline_native_Class_query helpers)--------------------- 3141// Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3142// Fall through if (mods & mask) == bits, take the guard otherwise. 3143Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3144 // Branch around if the given klass has the given modifier bit set. 3145 // Like generate_guard, adds a new path onto the region. 3146 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3147 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT); 3148 Node* mask = intcon(modifier_mask); 3149 Node* bits = intcon(modifier_bits); 3150 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) ); 3151 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) ); 3152 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) ); 3153 return generate_fair_guard(bol, region); 3154} 3155Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3156 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3157} 3158 3159//-------------------------inline_native_Class_query------------------- 3160bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3161 int nargs = 1+0; // just the Class mirror, in most cases 3162 const Type* return_type = TypeInt::BOOL; 3163 Node* prim_return_value = top(); // what happens if it's a primitive class? 3164 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3165 bool expect_prim = false; // most of these guys expect to work on refs 3166 3167 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3168 3169 switch (id) { 3170 case vmIntrinsics::_isInstance: 3171 nargs = 1+1; // the Class mirror, plus the object getting queried about 3172 // nothing is an instance of a primitive type 3173 prim_return_value = intcon(0); 3174 break; 3175 case vmIntrinsics::_getModifiers: 3176 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3177 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3178 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3179 break; 3180 case vmIntrinsics::_isInterface: 3181 prim_return_value = intcon(0); 3182 break; 3183 case vmIntrinsics::_isArray: 3184 prim_return_value = intcon(0); 3185 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3186 break; 3187 case vmIntrinsics::_isPrimitive: 3188 prim_return_value = intcon(1); 3189 expect_prim = true; // obviously 3190 break; 3191 case vmIntrinsics::_getSuperclass: 3192 prim_return_value = null(); 3193 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3194 break; 3195 case vmIntrinsics::_getComponentType: 3196 prim_return_value = null(); 3197 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3198 break; 3199 case vmIntrinsics::_getClassAccessFlags: 3200 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3201 return_type = TypeInt::INT; // not bool! 6297094 3202 break; 3203 default: 3204 ShouldNotReachHere(); 3205 } 3206 3207 Node* mirror = argument(0); 3208 Node* obj = (nargs <= 1)? top(): argument(1); 3209 3210 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3211 if (mirror_con == NULL) return false; // cannot happen? 3212 3213#ifndef PRODUCT 3214 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) { 3215 ciType* k = mirror_con->java_mirror_type(); 3216 if (k) { 3217 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3218 k->print_name(); 3219 tty->cr(); 3220 } 3221 } 3222#endif 3223 3224 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3225 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3226 record_for_igvn(region); 3227 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type); 3228 3229 // The mirror will never be null of Reflection.getClassAccessFlags, however 3230 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3231 // if it is. See bug 4774291. 3232 3233 // For Reflection.getClassAccessFlags(), the null check occurs in 3234 // the wrong place; see inline_unsafe_access(), above, for a similar 3235 // situation. 3236 _sp += nargs; // set original stack for use by uncommon_trap 3237 mirror = do_null_check(mirror, T_OBJECT); 3238 _sp -= nargs; 3239 // If mirror or obj is dead, only null-path is taken. 3240 if (stopped()) return true; 3241 3242 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3243 3244 // Now load the mirror's klass metaobject, and null-check it. 3245 // Side-effects region with the control path if the klass is null. 3246 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs, 3247 region, _prim_path); 3248 // If kls is null, we have a primitive mirror. 3249 phi->init_req(_prim_path, prim_return_value); 3250 if (stopped()) { push_result(region, phi); return true; } 3251 3252 Node* p; // handy temp 3253 Node* null_ctl; 3254 3255 // Now that we have the non-null klass, we can perform the real query. 3256 // For constant classes, the query will constant-fold in LoadNode::Value. 3257 Node* query_value = top(); 3258 switch (id) { 3259 case vmIntrinsics::_isInstance: 3260 // nothing is an instance of a primitive type 3261 _sp += nargs; // gen_instanceof might do an uncommon trap 3262 query_value = gen_instanceof(obj, kls); 3263 _sp -= nargs; 3264 break; 3265 3266 case vmIntrinsics::_getModifiers: 3267 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3268 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3269 break; 3270 3271 case vmIntrinsics::_isInterface: 3272 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3273 if (generate_interface_guard(kls, region) != NULL) 3274 // A guard was added. If the guard is taken, it was an interface. 3275 phi->add_req(intcon(1)); 3276 // If we fall through, it's a plain class. 3277 query_value = intcon(0); 3278 break; 3279 3280 case vmIntrinsics::_isArray: 3281 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3282 if (generate_array_guard(kls, region) != NULL) 3283 // A guard was added. If the guard is taken, it was an array. 3284 phi->add_req(intcon(1)); 3285 // If we fall through, it's a plain class. 3286 query_value = intcon(0); 3287 break; 3288 3289 case vmIntrinsics::_isPrimitive: 3290 query_value = intcon(0); // "normal" path produces false 3291 break; 3292 3293 case vmIntrinsics::_getSuperclass: 3294 // The rules here are somewhat unfortunate, but we can still do better 3295 // with random logic than with a JNI call. 3296 // Interfaces store null or Object as _super, but must report null. 3297 // Arrays store an intermediate super as _super, but must report Object. 3298 // Other types can report the actual _super. 3299 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3300 if (generate_interface_guard(kls, region) != NULL) 3301 // A guard was added. If the guard is taken, it was an interface. 3302 phi->add_req(null()); 3303 if (generate_array_guard(kls, region) != NULL) 3304 // A guard was added. If the guard is taken, it was an array. 3305 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3306 // If we fall through, it's a plain class. Get its _super. 3307 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3308 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) ); 3309 null_ctl = top(); 3310 kls = null_check_oop(kls, &null_ctl); 3311 if (null_ctl != top()) { 3312 // If the guard is taken, Object.superClass is null (both klass and mirror). 3313 region->add_req(null_ctl); 3314 phi ->add_req(null()); 3315 } 3316 if (!stopped()) { 3317 query_value = load_mirror_from_klass(kls); 3318 } 3319 break; 3320 3321 case vmIntrinsics::_getComponentType: 3322 if (generate_array_guard(kls, region) != NULL) { 3323 // Be sure to pin the oop load to the guard edge just created: 3324 Node* is_array_ctrl = region->in(region->req()-1); 3325 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset())); 3326 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT); 3327 phi->add_req(cmo); 3328 } 3329 query_value = null(); // non-array case is null 3330 break; 3331 3332 case vmIntrinsics::_getClassAccessFlags: 3333 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3334 query_value = make_load(NULL, p, TypeInt::INT, T_INT); 3335 break; 3336 3337 default: 3338 ShouldNotReachHere(); 3339 } 3340 3341 // Fall-through is the normal case of a query to a real class. 3342 phi->init_req(1, query_value); 3343 region->init_req(1, control()); 3344 3345 push_result(region, phi); 3346 C->set_has_split_ifs(true); // Has chance for split-if optimization 3347 3348 return true; 3349} 3350 3351//--------------------------inline_native_subtype_check------------------------ 3352// This intrinsic takes the JNI calls out of the heart of 3353// UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3354bool LibraryCallKit::inline_native_subtype_check() { 3355 int nargs = 1+1; // the Class mirror, plus the other class getting examined 3356 3357 // Pull both arguments off the stack. 3358 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3359 args[0] = argument(0); 3360 args[1] = argument(1); 3361 Node* klasses[2]; // corresponding Klasses: superk, subk 3362 klasses[0] = klasses[1] = top(); 3363 3364 enum { 3365 // A full decision tree on {superc is prim, subc is prim}: 3366 _prim_0_path = 1, // {P,N} => false 3367 // {P,P} & superc!=subc => false 3368 _prim_same_path, // {P,P} & superc==subc => true 3369 _prim_1_path, // {N,P} => false 3370 _ref_subtype_path, // {N,N} & subtype check wins => true 3371 _both_ref_path, // {N,N} & subtype check loses => false 3372 PATH_LIMIT 3373 }; 3374 3375 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3376 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 3377 record_for_igvn(region); 3378 3379 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3380 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3381 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3382 3383 // First null-check both mirrors and load each mirror's klass metaobject. 3384 int which_arg; 3385 for (which_arg = 0; which_arg <= 1; which_arg++) { 3386 Node* arg = args[which_arg]; 3387 _sp += nargs; // set original stack for use by uncommon_trap 3388 arg = do_null_check(arg, T_OBJECT); 3389 _sp -= nargs; 3390 if (stopped()) break; 3391 args[which_arg] = _gvn.transform(arg); 3392 3393 Node* p = basic_plus_adr(arg, class_klass_offset); 3394 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3395 klasses[which_arg] = _gvn.transform(kls); 3396 } 3397 3398 // Having loaded both klasses, test each for null. 3399 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3400 for (which_arg = 0; which_arg <= 1; which_arg++) { 3401 Node* kls = klasses[which_arg]; 3402 Node* null_ctl = top(); 3403 _sp += nargs; // set original stack for use by uncommon_trap 3404 kls = null_check_oop(kls, &null_ctl, never_see_null); 3405 _sp -= nargs; 3406 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3407 region->init_req(prim_path, null_ctl); 3408 if (stopped()) break; 3409 klasses[which_arg] = kls; 3410 } 3411 3412 if (!stopped()) { 3413 // now we have two reference types, in klasses[0..1] 3414 Node* subk = klasses[1]; // the argument to isAssignableFrom 3415 Node* superk = klasses[0]; // the receiver 3416 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3417 // now we have a successful reference subtype check 3418 region->set_req(_ref_subtype_path, control()); 3419 } 3420 3421 // If both operands are primitive (both klasses null), then 3422 // we must return true when they are identical primitives. 3423 // It is convenient to test this after the first null klass check. 3424 set_control(region->in(_prim_0_path)); // go back to first null check 3425 if (!stopped()) { 3426 // Since superc is primitive, make a guard for the superc==subc case. 3427 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) ); 3428 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) ); 3429 generate_guard(bol_eq, region, PROB_FAIR); 3430 if (region->req() == PATH_LIMIT+1) { 3431 // A guard was added. If the added guard is taken, superc==subc. 3432 region->swap_edges(PATH_LIMIT, _prim_same_path); 3433 region->del_req(PATH_LIMIT); 3434 } 3435 region->set_req(_prim_0_path, control()); // Not equal after all. 3436 } 3437 3438 // these are the only paths that produce 'true': 3439 phi->set_req(_prim_same_path, intcon(1)); 3440 phi->set_req(_ref_subtype_path, intcon(1)); 3441 3442 // pull together the cases: 3443 assert(region->req() == PATH_LIMIT, "sane region"); 3444 for (uint i = 1; i < region->req(); i++) { 3445 Node* ctl = region->in(i); 3446 if (ctl == NULL || ctl == top()) { 3447 region->set_req(i, top()); 3448 phi ->set_req(i, top()); 3449 } else if (phi->in(i) == NULL) { 3450 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3451 } 3452 } 3453 3454 set_control(_gvn.transform(region)); 3455 push(_gvn.transform(phi)); 3456 3457 return true; 3458} 3459 3460//---------------------generate_array_guard_common------------------------ 3461Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3462 bool obj_array, bool not_array) { 3463 // If obj_array/non_array==false/false: 3464 // Branch around if the given klass is in fact an array (either obj or prim). 3465 // If obj_array/non_array==false/true: 3466 // Branch around if the given klass is not an array klass of any kind. 3467 // If obj_array/non_array==true/true: 3468 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3469 // If obj_array/non_array==true/false: 3470 // Branch around if the kls is an oop array (Object[] or subtype) 3471 // 3472 // Like generate_guard, adds a new path onto the region. 3473 jint layout_con = 0; 3474 Node* layout_val = get_layout_helper(kls, layout_con); 3475 if (layout_val == NULL) { 3476 bool query = (obj_array 3477 ? Klass::layout_helper_is_objArray(layout_con) 3478 : Klass::layout_helper_is_array(layout_con)); 3479 if (query == not_array) { 3480 return NULL; // never a branch 3481 } else { // always a branch 3482 Node* always_branch = control(); 3483 if (region != NULL) 3484 region->add_req(always_branch); 3485 set_control(top()); 3486 return always_branch; 3487 } 3488 } 3489 // Now test the correct condition. 3490 jint nval = (obj_array 3491 ? ((jint)Klass::_lh_array_tag_type_value 3492 << Klass::_lh_array_tag_shift) 3493 : Klass::_lh_neutral_value); 3494 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) ); 3495 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3496 // invert the test if we are looking for a non-array 3497 if (not_array) btest = BoolTest(btest).negate(); 3498 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) ); 3499 return generate_fair_guard(bol, region); 3500} 3501 3502 3503//-----------------------inline_native_newArray-------------------------- 3504bool LibraryCallKit::inline_native_newArray() { 3505 int nargs = 2; 3506 Node* mirror = argument(0); 3507 Node* count_val = argument(1); 3508 3509 _sp += nargs; // set original stack for use by uncommon_trap 3510 mirror = do_null_check(mirror, T_OBJECT); 3511 _sp -= nargs; 3512 // If mirror or obj is dead, only null-path is taken. 3513 if (stopped()) return true; 3514 3515 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3516 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3517 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3518 TypeInstPtr::NOTNULL); 3519 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3520 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3521 TypePtr::BOTTOM); 3522 3523 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3524 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3525 nargs, 3526 result_reg, _slow_path); 3527 Node* normal_ctl = control(); 3528 Node* no_array_ctl = result_reg->in(_slow_path); 3529 3530 // Generate code for the slow case. We make a call to newArray(). 3531 set_control(no_array_ctl); 3532 if (!stopped()) { 3533 // Either the input type is void.class, or else the 3534 // array klass has not yet been cached. Either the 3535 // ensuing call will throw an exception, or else it 3536 // will cache the array klass for next time. 3537 PreserveJVMState pjvms(this); 3538 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3539 Node* slow_result = set_results_for_java_call(slow_call); 3540 // this->control() comes from set_results_for_java_call 3541 result_reg->set_req(_slow_path, control()); 3542 result_val->set_req(_slow_path, slow_result); 3543 result_io ->set_req(_slow_path, i_o()); 3544 result_mem->set_req(_slow_path, reset_memory()); 3545 } 3546 3547 set_control(normal_ctl); 3548 if (!stopped()) { 3549 // Normal case: The array type has been cached in the java.lang.Class. 3550 // The following call works fine even if the array type is polymorphic. 3551 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3552 Node* obj = new_array(klass_node, count_val, nargs); 3553 result_reg->init_req(_normal_path, control()); 3554 result_val->init_req(_normal_path, obj); 3555 result_io ->init_req(_normal_path, i_o()); 3556 result_mem->init_req(_normal_path, reset_memory()); 3557 } 3558 3559 // Return the combined state. 3560 set_i_o( _gvn.transform(result_io) ); 3561 set_all_memory( _gvn.transform(result_mem) ); 3562 push_result(result_reg, result_val); 3563 C->set_has_split_ifs(true); // Has chance for split-if optimization 3564 3565 return true; 3566} 3567 3568//----------------------inline_native_getLength-------------------------- 3569bool LibraryCallKit::inline_native_getLength() { 3570 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3571 3572 int nargs = 1; 3573 Node* array = argument(0); 3574 3575 _sp += nargs; // set original stack for use by uncommon_trap 3576 array = do_null_check(array, T_OBJECT); 3577 _sp -= nargs; 3578 3579 // If array is dead, only null-path is taken. 3580 if (stopped()) return true; 3581 3582 // Deoptimize if it is a non-array. 3583 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3584 3585 if (non_array != NULL) { 3586 PreserveJVMState pjvms(this); 3587 set_control(non_array); 3588 _sp += nargs; // push the arguments back on the stack 3589 uncommon_trap(Deoptimization::Reason_intrinsic, 3590 Deoptimization::Action_maybe_recompile); 3591 } 3592 3593 // If control is dead, only non-array-path is taken. 3594 if (stopped()) return true; 3595 3596 // The works fine even if the array type is polymorphic. 3597 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3598 push( load_array_length(array) ); 3599 3600 C->set_has_split_ifs(true); // Has chance for split-if optimization 3601 3602 return true; 3603} 3604 3605//------------------------inline_array_copyOf---------------------------- 3606bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3607 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3608 3609 // Restore the stack and pop off the arguments. 3610 int nargs = 3 + (is_copyOfRange? 1: 0); 3611 Node* original = argument(0); 3612 Node* start = is_copyOfRange? argument(1): intcon(0); 3613 Node* end = is_copyOfRange? argument(2): argument(1); 3614 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3615 3616 Node* newcopy; 3617 3618 //set the original stack and the reexecute bit for the interpreter to reexecute 3619 //the bytecode that invokes Arrays.copyOf if deoptimization happens 3620 { PreserveReexecuteState preexecs(this); 3621 _sp += nargs; 3622 jvms()->set_should_reexecute(true); 3623 3624 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT); 3625 original = do_null_check(original, T_OBJECT); 3626 3627 // Check if a null path was taken unconditionally. 3628 if (stopped()) return true; 3629 3630 Node* orig_length = load_array_length(original); 3631 3632 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0, 3633 NULL, 0); 3634 klass_node = do_null_check(klass_node, T_OBJECT); 3635 3636 RegionNode* bailout = new (C, 1) RegionNode(1); 3637 record_for_igvn(bailout); 3638 3639 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3640 // Bail out if that is so. 3641 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3642 if (not_objArray != NULL) { 3643 // Improve the klass node's type from the new optimistic assumption: 3644 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3645 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3646 Node* cast = new (C, 2) CastPPNode(klass_node, akls); 3647 cast->init_req(0, control()); 3648 klass_node = _gvn.transform(cast); 3649 } 3650 3651 // Bail out if either start or end is negative. 3652 generate_negative_guard(start, bailout, &start); 3653 generate_negative_guard(end, bailout, &end); 3654 3655 Node* length = end; 3656 if (_gvn.type(start) != TypeInt::ZERO) { 3657 length = _gvn.transform( new (C, 3) SubINode(end, start) ); 3658 } 3659 3660 // Bail out if length is negative. 3661 // Without this the new_array would throw 3662 // NegativeArraySizeException but IllegalArgumentException is what 3663 // should be thrown 3664 generate_negative_guard(length, bailout, &length); 3665 3666 if (bailout->req() > 1) { 3667 PreserveJVMState pjvms(this); 3668 set_control( _gvn.transform(bailout) ); 3669 uncommon_trap(Deoptimization::Reason_intrinsic, 3670 Deoptimization::Action_maybe_recompile); 3671 } 3672 3673 if (!stopped()) { 3674 3675 // How many elements will we copy from the original? 3676 // The answer is MinI(orig_length - start, length). 3677 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) ); 3678 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3679 3680 newcopy = new_array(klass_node, length, 0); 3681 3682 // Generate a direct call to the right arraycopy function(s). 3683 // We know the copy is disjoint but we might not know if the 3684 // oop stores need checking. 3685 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3686 // This will fail a store-check if x contains any non-nulls. 3687 bool disjoint_bases = true; 3688 // if start > orig_length then the length of the copy may be 3689 // negative. 3690 bool length_never_negative = !is_copyOfRange; 3691 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 3692 original, start, newcopy, intcon(0), moved, 3693 disjoint_bases, length_never_negative); 3694 } 3695 } //original reexecute and sp are set back here 3696 3697 if(!stopped()) { 3698 push(newcopy); 3699 } 3700 3701 C->set_has_split_ifs(true); // Has chance for split-if optimization 3702 3703 return true; 3704} 3705 3706 3707//----------------------generate_virtual_guard--------------------------- 3708// Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3709Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3710 RegionNode* slow_region) { 3711 ciMethod* method = callee(); 3712 int vtable_index = method->vtable_index(); 3713 // Get the Method* out of the appropriate vtable entry. 3714 int entry_offset = (InstanceKlass::vtable_start_offset() + 3715 vtable_index*vtableEntry::size()) * wordSize + 3716 vtableEntry::method_offset_in_bytes(); 3717 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3718 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT); 3719 3720 // Compare the target method with the expected method (e.g., Object.hashCode). 3721 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3722 3723 Node* native_call = makecon(native_call_addr); 3724 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) ); 3725 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) ); 3726 3727 return generate_slow_guard(test_native, slow_region); 3728} 3729 3730//-----------------------generate_method_call---------------------------- 3731// Use generate_method_call to make a slow-call to the real 3732// method if the fast path fails. An alternative would be to 3733// use a stub like OptoRuntime::slow_arraycopy_Java. 3734// This only works for expanding the current library call, 3735// not another intrinsic. (E.g., don't use this for making an 3736// arraycopy call inside of the copyOf intrinsic.) 3737CallJavaNode* 3738LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3739 // When compiling the intrinsic method itself, do not use this technique. 3740 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3741 3742 ciMethod* method = callee(); 3743 // ensure the JVMS we have will be correct for this call 3744 guarantee(method_id == method->intrinsic_id(), "must match"); 3745 3746 const TypeFunc* tf = TypeFunc::make(method); 3747 int tfdc = tf->domain()->cnt(); 3748 CallJavaNode* slow_call; 3749 if (is_static) { 3750 assert(!is_virtual, ""); 3751 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3752 SharedRuntime::get_resolve_static_call_stub(), 3753 method, bci()); 3754 } else if (is_virtual) { 3755 null_check_receiver(method); 3756 int vtable_index = Method::invalid_vtable_index; 3757 if (UseInlineCaches) { 3758 // Suppress the vtable call 3759 } else { 3760 // hashCode and clone are not a miranda methods, 3761 // so the vtable index is fixed. 3762 // No need to use the linkResolver to get it. 3763 vtable_index = method->vtable_index(); 3764 } 3765 slow_call = new(C, tfdc) CallDynamicJavaNode(tf, 3766 SharedRuntime::get_resolve_virtual_call_stub(), 3767 method, vtable_index, bci()); 3768 } else { // neither virtual nor static: opt_virtual 3769 null_check_receiver(method); 3770 slow_call = new(C, tfdc) CallStaticJavaNode(tf, 3771 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3772 method, bci()); 3773 slow_call->set_optimized_virtual(true); 3774 } 3775 set_arguments_for_java_call(slow_call); 3776 set_edges_for_java_call(slow_call); 3777 return slow_call; 3778} 3779 3780 3781//------------------------------inline_native_hashcode-------------------- 3782// Build special case code for calls to hashCode on an object. 3783bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3784 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3785 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3786 3787 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3788 3789 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 3790 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 3791 TypeInt::INT); 3792 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 3793 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 3794 TypePtr::BOTTOM); 3795 Node* obj = NULL; 3796 if (!is_static) { 3797 // Check for hashing null object 3798 obj = null_check_receiver(callee()); 3799 if (stopped()) return true; // unconditionally null 3800 result_reg->init_req(_null_path, top()); 3801 result_val->init_req(_null_path, top()); 3802 } else { 3803 // Do a null check, and return zero if null. 3804 // System.identityHashCode(null) == 0 3805 obj = argument(0); 3806 Node* null_ctl = top(); 3807 obj = null_check_oop(obj, &null_ctl); 3808 result_reg->init_req(_null_path, null_ctl); 3809 result_val->init_req(_null_path, _gvn.intcon(0)); 3810 } 3811 3812 // Unconditionally null? Then return right away. 3813 if (stopped()) { 3814 set_control( result_reg->in(_null_path) ); 3815 if (!stopped()) 3816 push( result_val ->in(_null_path) ); 3817 return true; 3818 } 3819 3820 // After null check, get the object's klass. 3821 Node* obj_klass = load_object_klass(obj); 3822 3823 // This call may be virtual (invokevirtual) or bound (invokespecial). 3824 // For each case we generate slightly different code. 3825 3826 // We only go to the fast case code if we pass a number of guards. The 3827 // paths which do not pass are accumulated in the slow_region. 3828 RegionNode* slow_region = new (C, 1) RegionNode(1); 3829 record_for_igvn(slow_region); 3830 3831 // If this is a virtual call, we generate a funny guard. We pull out 3832 // the vtable entry corresponding to hashCode() from the target object. 3833 // If the target method which we are calling happens to be the native 3834 // Object hashCode() method, we pass the guard. We do not need this 3835 // guard for non-virtual calls -- the caller is known to be the native 3836 // Object hashCode(). 3837 if (is_virtual) { 3838 generate_virtual_guard(obj_klass, slow_region); 3839 } 3840 3841 // Get the header out of the object, use LoadMarkNode when available 3842 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3843 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type()); 3844 3845 // Test the header to see if it is unlocked. 3846 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 3847 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) ); 3848 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 3849 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val)); 3850 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) ); 3851 3852 generate_slow_guard(test_unlocked, slow_region); 3853 3854 // Get the hash value and check to see that it has been properly assigned. 3855 // We depend on hash_mask being at most 32 bits and avoid the use of 3856 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 3857 // vm: see markOop.hpp. 3858 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 3859 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 3860 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) ); 3861 // This hack lets the hash bits live anywhere in the mark object now, as long 3862 // as the shift drops the relevant bits into the low 32 bits. Note that 3863 // Java spec says that HashCode is an int so there's no point in capturing 3864 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 3865 hshifted_header = ConvX2I(hshifted_header); 3866 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) ); 3867 3868 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 3869 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val)); 3870 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) ); 3871 3872 generate_slow_guard(test_assigned, slow_region); 3873 3874 Node* init_mem = reset_memory(); 3875 // fill in the rest of the null path: 3876 result_io ->init_req(_null_path, i_o()); 3877 result_mem->init_req(_null_path, init_mem); 3878 3879 result_val->init_req(_fast_path, hash_val); 3880 result_reg->init_req(_fast_path, control()); 3881 result_io ->init_req(_fast_path, i_o()); 3882 result_mem->init_req(_fast_path, init_mem); 3883 3884 // Generate code for the slow case. We make a call to hashCode(). 3885 set_control(_gvn.transform(slow_region)); 3886 if (!stopped()) { 3887 // No need for PreserveJVMState, because we're using up the present state. 3888 set_all_memory(init_mem); 3889 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode; 3890 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode; 3891 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 3892 Node* slow_result = set_results_for_java_call(slow_call); 3893 // this->control() comes from set_results_for_java_call 3894 result_reg->init_req(_slow_path, control()); 3895 result_val->init_req(_slow_path, slow_result); 3896 result_io ->set_req(_slow_path, i_o()); 3897 result_mem ->set_req(_slow_path, reset_memory()); 3898 } 3899 3900 // Return the combined state. 3901 set_i_o( _gvn.transform(result_io) ); 3902 set_all_memory( _gvn.transform(result_mem) ); 3903 push_result(result_reg, result_val); 3904 3905 return true; 3906} 3907 3908//---------------------------inline_native_getClass---------------------------- 3909// Build special case code for calls to getClass on an object. 3910bool LibraryCallKit::inline_native_getClass() { 3911 Node* obj = null_check_receiver(callee()); 3912 if (stopped()) return true; 3913 push( load_mirror_from_klass(load_object_klass(obj)) ); 3914 return true; 3915} 3916 3917//-----------------inline_native_Reflection_getCallerClass--------------------- 3918// In the presence of deep enough inlining, getCallerClass() becomes a no-op. 3919// 3920// NOTE that this code must perform the same logic as 3921// vframeStream::security_get_caller_frame in that it must skip 3922// Method.invoke() and auxiliary frames. 3923 3924 3925 3926 3927bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 3928 ciMethod* method = callee(); 3929 3930#ifndef PRODUCT 3931 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3932 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 3933 } 3934#endif 3935 3936 debug_only(int saved_sp = _sp); 3937 3938 // Argument words: (int depth) 3939 int nargs = 1; 3940 3941 _sp += nargs; 3942 Node* caller_depth_node = pop(); 3943 3944 assert(saved_sp == _sp, "must have correct argument count"); 3945 3946 // The depth value must be a constant in order for the runtime call 3947 // to be eliminated. 3948 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int(); 3949 if (caller_depth_type == NULL || !caller_depth_type->is_con()) { 3950#ifndef PRODUCT 3951 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3952 tty->print_cr(" Bailing out because caller depth was not a constant"); 3953 } 3954#endif 3955 return false; 3956 } 3957 // Note that the JVM state at this point does not include the 3958 // getCallerClass() frame which we are trying to inline. The 3959 // semantics of getCallerClass(), however, are that the "first" 3960 // frame is the getCallerClass() frame, so we subtract one from the 3961 // requested depth before continuing. We don't inline requests of 3962 // getCallerClass(0). 3963 int caller_depth = caller_depth_type->get_con() - 1; 3964 if (caller_depth < 0) { 3965#ifndef PRODUCT 3966 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3967 tty->print_cr(" Bailing out because caller depth was %d", caller_depth); 3968 } 3969#endif 3970 return false; 3971 } 3972 3973 if (!jvms()->has_method()) { 3974#ifndef PRODUCT 3975 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 3976 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 3977 } 3978#endif 3979 return false; 3980 } 3981 int _depth = jvms()->depth(); // cache call chain depth 3982 3983 // Walk back up the JVM state to find the caller at the required 3984 // depth. NOTE that this code must perform the same logic as 3985 // vframeStream::security_get_caller_frame in that it must skip 3986 // Method.invoke() and auxiliary frames. Note also that depth is 3987 // 1-based (1 is the bottom of the inlining). 3988 int inlining_depth = _depth; 3989 JVMState* caller_jvms = NULL; 3990 3991 if (inlining_depth > 0) { 3992 caller_jvms = jvms(); 3993 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth"); 3994 do { 3995 // The following if-tests should be performed in this order 3996 if (is_method_invoke_or_aux_frame(caller_jvms)) { 3997 // Skip a Method.invoke() or auxiliary frame 3998 } else if (caller_depth > 0) { 3999 // Skip real frame 4000 --caller_depth; 4001 } else { 4002 // We're done: reached desired caller after skipping. 4003 break; 4004 } 4005 caller_jvms = caller_jvms->caller(); 4006 --inlining_depth; 4007 } while (inlining_depth > 0); 4008 } 4009 4010 if (inlining_depth == 0) { 4011#ifndef PRODUCT 4012 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 4013 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth); 4014 tty->print_cr(" JVM state at this point:"); 4015 for (int i = _depth; i >= 1; i--) { 4016 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 4017 } 4018 } 4019#endif 4020 return false; // Reached end of inlining 4021 } 4022 4023 // Acquire method holder as java.lang.Class 4024 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4025 ciInstance* caller_mirror = caller_klass->java_mirror(); 4026 // Push this as a constant 4027 push(makecon(TypeInstPtr::make(caller_mirror))); 4028#ifndef PRODUCT 4029 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) { 4030 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); 4031 tty->print_cr(" JVM state at this point:"); 4032 for (int i = _depth; i >= 1; i--) { 4033 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8()); 4034 } 4035 } 4036#endif 4037 return true; 4038} 4039 4040// Helper routine for above 4041bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) { 4042 ciMethod* method = jvms->method(); 4043 4044 // Is this the Method.invoke method itself? 4045 if (method->intrinsic_id() == vmIntrinsics::_invoke) 4046 return true; 4047 4048 // Is this a helper, defined somewhere underneath MethodAccessorImpl. 4049 ciKlass* k = method->holder(); 4050 if (k->is_instance_klass()) { 4051 ciInstanceKlass* ik = k->as_instance_klass(); 4052 for (; ik != NULL; ik = ik->super()) { 4053 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() && 4054 ik == env()->find_system_klass(ik->name())) { 4055 return true; 4056 } 4057 } 4058 } 4059 else if (method->is_method_handle_intrinsic() || 4060 method->is_compiled_lambda_form()) { 4061 // This is an internal adapter frame from the MethodHandleCompiler -- skip it 4062 return true; 4063 } 4064 4065 return false; 4066} 4067 4068bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4069 // restore the arguments 4070 _sp += arg_size(); 4071 4072 switch (id) { 4073 case vmIntrinsics::_floatToRawIntBits: 4074 push(_gvn.transform( new (C, 2) MoveF2INode(pop()))); 4075 break; 4076 4077 case vmIntrinsics::_intBitsToFloat: 4078 push(_gvn.transform( new (C, 2) MoveI2FNode(pop()))); 4079 break; 4080 4081 case vmIntrinsics::_doubleToRawLongBits: 4082 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair()))); 4083 break; 4084 4085 case vmIntrinsics::_longBitsToDouble: 4086 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair()))); 4087 break; 4088 4089 case vmIntrinsics::_doubleToLongBits: { 4090 Node* value = pop_pair(); 4091 4092 // two paths (plus control) merge in a wood 4093 RegionNode *r = new (C, 3) RegionNode(3); 4094 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG); 4095 4096 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value)); 4097 // Build the boolean node 4098 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 4099 4100 // Branch either way. 4101 // NaN case is less traveled, which makes all the difference. 4102 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4103 Node *opt_isnan = _gvn.transform(ifisnan); 4104 assert( opt_isnan->is_If(), "Expect an IfNode"); 4105 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4106 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 4107 4108 set_control(iftrue); 4109 4110 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4111 Node *slow_result = longcon(nan_bits); // return NaN 4112 phi->init_req(1, _gvn.transform( slow_result )); 4113 r->init_req(1, iftrue); 4114 4115 // Else fall through 4116 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 4117 set_control(iffalse); 4118 4119 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value))); 4120 r->init_req(2, iffalse); 4121 4122 // Post merge 4123 set_control(_gvn.transform(r)); 4124 record_for_igvn(r); 4125 4126 Node* result = _gvn.transform(phi); 4127 assert(result->bottom_type()->isa_long(), "must be"); 4128 push_pair(result); 4129 4130 C->set_has_split_ifs(true); // Has chance for split-if optimization 4131 4132 break; 4133 } 4134 4135 case vmIntrinsics::_floatToIntBits: { 4136 Node* value = pop(); 4137 4138 // two paths (plus control) merge in a wood 4139 RegionNode *r = new (C, 3) RegionNode(3); 4140 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT); 4141 4142 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value)); 4143 // Build the boolean node 4144 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) ); 4145 4146 // Branch either way. 4147 // NaN case is less traveled, which makes all the difference. 4148 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4149 Node *opt_isnan = _gvn.transform(ifisnan); 4150 assert( opt_isnan->is_If(), "Expect an IfNode"); 4151 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4152 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) ); 4153 4154 set_control(iftrue); 4155 4156 static const jint nan_bits = 0x7fc00000; 4157 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4158 phi->init_req(1, _gvn.transform( slow_result )); 4159 r->init_req(1, iftrue); 4160 4161 // Else fall through 4162 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) ); 4163 set_control(iffalse); 4164 4165 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value))); 4166 r->init_req(2, iffalse); 4167 4168 // Post merge 4169 set_control(_gvn.transform(r)); 4170 record_for_igvn(r); 4171 4172 Node* result = _gvn.transform(phi); 4173 assert(result->bottom_type()->isa_int(), "must be"); 4174 push(result); 4175 4176 C->set_has_split_ifs(true); // Has chance for split-if optimization 4177 4178 break; 4179 } 4180 4181 default: 4182 ShouldNotReachHere(); 4183 } 4184 4185 return true; 4186} 4187 4188#ifdef _LP64 4189#define XTOP ,top() /*additional argument*/ 4190#else //_LP64 4191#define XTOP /*no additional argument*/ 4192#endif //_LP64 4193 4194//----------------------inline_unsafe_copyMemory------------------------- 4195bool LibraryCallKit::inline_unsafe_copyMemory() { 4196 if (callee()->is_static()) return false; // caller must have the capability! 4197 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size) 4198 assert(signature()->size() == nargs-1, "copy has 5 arguments"); 4199 null_check_receiver(callee()); // check then ignore argument(0) 4200 if (stopped()) return true; 4201 4202 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4203 4204 Node* src_ptr = argument(1); 4205 Node* src_off = ConvL2X(argument(2)); 4206 assert(argument(3)->is_top(), "2nd half of long"); 4207 Node* dst_ptr = argument(4); 4208 Node* dst_off = ConvL2X(argument(5)); 4209 assert(argument(6)->is_top(), "2nd half of long"); 4210 Node* size = ConvL2X(argument(7)); 4211 assert(argument(8)->is_top(), "2nd half of long"); 4212 4213 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4214 "fieldOffset must be byte-scaled"); 4215 4216 Node* src = make_unsafe_address(src_ptr, src_off); 4217 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4218 4219 // Conservatively insert a memory barrier on all memory slices. 4220 // Do not let writes of the copy source or destination float below the copy. 4221 insert_mem_bar(Op_MemBarCPUOrder); 4222 4223 // Call it. Note that the length argument is not scaled. 4224 make_runtime_call(RC_LEAF|RC_NO_FP, 4225 OptoRuntime::fast_arraycopy_Type(), 4226 StubRoutines::unsafe_arraycopy(), 4227 "unsafe_arraycopy", 4228 TypeRawPtr::BOTTOM, 4229 src, dst, size XTOP); 4230 4231 // Do not let reads of the copy destination float above the copy. 4232 insert_mem_bar(Op_MemBarCPUOrder); 4233 4234 return true; 4235} 4236 4237//------------------------clone_coping----------------------------------- 4238// Helper function for inline_native_clone. 4239void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4240 assert(obj_size != NULL, ""); 4241 Node* raw_obj = alloc_obj->in(1); 4242 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4243 4244 AllocateNode* alloc = NULL; 4245 if (ReduceBulkZeroing) { 4246 // We will be completely responsible for initializing this object - 4247 // mark Initialize node as complete. 4248 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4249 // The object was just allocated - there should be no any stores! 4250 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4251 // Mark as complete_with_arraycopy so that on AllocateNode 4252 // expansion, we know this AllocateNode is initialized by an array 4253 // copy and a StoreStore barrier exists after the array copy. 4254 alloc->initialization()->set_complete_with_arraycopy(); 4255 } 4256 4257 // Copy the fastest available way. 4258 // TODO: generate fields copies for small objects instead. 4259 Node* src = obj; 4260 Node* dest = alloc_obj; 4261 Node* size = _gvn.transform(obj_size); 4262 4263 // Exclude the header but include array length to copy by 8 bytes words. 4264 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4265 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4266 instanceOopDesc::base_offset_in_bytes(); 4267 // base_off: 4268 // 8 - 32-bit VM 4269 // 12 - 64-bit VM, compressed klass 4270 // 16 - 64-bit VM, normal klass 4271 if (base_off % BytesPerLong != 0) { 4272 assert(UseCompressedOops, ""); 4273 if (is_array) { 4274 // Exclude length to copy by 8 bytes words. 4275 base_off += sizeof(int); 4276 } else { 4277 // Include klass to copy by 8 bytes words. 4278 base_off = instanceOopDesc::klass_offset_in_bytes(); 4279 } 4280 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4281 } 4282 src = basic_plus_adr(src, base_off); 4283 dest = basic_plus_adr(dest, base_off); 4284 4285 // Compute the length also, if needed: 4286 Node* countx = size; 4287 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) ); 4288 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4289 4290 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4291 bool disjoint_bases = true; 4292 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4293 src, NULL, dest, NULL, countx, 4294 /*dest_uninitialized*/true); 4295 4296 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4297 if (card_mark) { 4298 assert(!is_array, ""); 4299 // Put in store barrier for any and all oops we are sticking 4300 // into this object. (We could avoid this if we could prove 4301 // that the object type contains no oop fields at all.) 4302 Node* no_particular_value = NULL; 4303 Node* no_particular_field = NULL; 4304 int raw_adr_idx = Compile::AliasIdxRaw; 4305 post_barrier(control(), 4306 memory(raw_adr_type), 4307 alloc_obj, 4308 no_particular_field, 4309 raw_adr_idx, 4310 no_particular_value, 4311 T_OBJECT, 4312 false); 4313 } 4314 4315 // Do not let reads from the cloned object float above the arraycopy. 4316 if (alloc != NULL) { 4317 // Do not let stores that initialize this object be reordered with 4318 // a subsequent store that would make this object accessible by 4319 // other threads. 4320 // Record what AllocateNode this StoreStore protects so that 4321 // escape analysis can go from the MemBarStoreStoreNode to the 4322 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4323 // based on the escape status of the AllocateNode. 4324 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4325 } else { 4326 insert_mem_bar(Op_MemBarCPUOrder); 4327 } 4328} 4329 4330//------------------------inline_native_clone---------------------------- 4331// Here are the simple edge cases: 4332// null receiver => normal trap 4333// virtual and clone was overridden => slow path to out-of-line clone 4334// not cloneable or finalizer => slow path to out-of-line Object.clone 4335// 4336// The general case has two steps, allocation and copying. 4337// Allocation has two cases, and uses GraphKit::new_instance or new_array. 4338// 4339// Copying also has two cases, oop arrays and everything else. 4340// Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4341// Everything else uses the tight inline loop supplied by CopyArrayNode. 4342// 4343// These steps fold up nicely if and when the cloned object's klass 4344// can be sharply typed as an object array, a type array, or an instance. 4345// 4346bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4347 int nargs = 1; 4348 PhiNode* result_val; 4349 4350 //set the original stack and the reexecute bit for the interpreter to reexecute 4351 //the bytecode that invokes Object.clone if deoptimization happens 4352 { PreserveReexecuteState preexecs(this); 4353 jvms()->set_should_reexecute(true); 4354 4355 //null_check_receiver will adjust _sp (push and pop) 4356 Node* obj = null_check_receiver(callee()); 4357 if (stopped()) return true; 4358 4359 _sp += nargs; 4360 4361 Node* obj_klass = load_object_klass(obj); 4362 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4363 const TypeOopPtr* toop = ((tklass != NULL) 4364 ? tklass->as_instance_type() 4365 : TypeInstPtr::NOTNULL); 4366 4367 // Conservatively insert a memory barrier on all memory slices. 4368 // Do not let writes into the original float below the clone. 4369 insert_mem_bar(Op_MemBarCPUOrder); 4370 4371 // paths into result_reg: 4372 enum { 4373 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4374 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4375 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4376 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4377 PATH_LIMIT 4378 }; 4379 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4380 result_val = new(C, PATH_LIMIT) PhiNode(result_reg, 4381 TypeInstPtr::NOTNULL); 4382 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO); 4383 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY, 4384 TypePtr::BOTTOM); 4385 record_for_igvn(result_reg); 4386 4387 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4388 int raw_adr_idx = Compile::AliasIdxRaw; 4389 4390 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4391 if (array_ctl != NULL) { 4392 // It's an array. 4393 PreserveJVMState pjvms(this); 4394 set_control(array_ctl); 4395 Node* obj_length = load_array_length(obj); 4396 Node* obj_size = NULL; 4397 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); 4398 4399 if (!use_ReduceInitialCardMarks()) { 4400 // If it is an oop array, it requires very special treatment, 4401 // because card marking is required on each card of the array. 4402 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4403 if (is_obja != NULL) { 4404 PreserveJVMState pjvms2(this); 4405 set_control(is_obja); 4406 // Generate a direct call to the right arraycopy function(s). 4407 bool disjoint_bases = true; 4408 bool length_never_negative = true; 4409 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4410 obj, intcon(0), alloc_obj, intcon(0), 4411 obj_length, 4412 disjoint_bases, length_never_negative); 4413 result_reg->init_req(_objArray_path, control()); 4414 result_val->init_req(_objArray_path, alloc_obj); 4415 result_i_o ->set_req(_objArray_path, i_o()); 4416 result_mem ->set_req(_objArray_path, reset_memory()); 4417 } 4418 } 4419 // Otherwise, there are no card marks to worry about. 4420 // (We can dispense with card marks if we know the allocation 4421 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4422 // causes the non-eden paths to take compensating steps to 4423 // simulate a fresh allocation, so that no further 4424 // card marks are required in compiled code to initialize 4425 // the object.) 4426 4427 if (!stopped()) { 4428 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4429 4430 // Present the results of the copy. 4431 result_reg->init_req(_array_path, control()); 4432 result_val->init_req(_array_path, alloc_obj); 4433 result_i_o ->set_req(_array_path, i_o()); 4434 result_mem ->set_req(_array_path, reset_memory()); 4435 } 4436 } 4437 4438 // We only go to the instance fast case code if we pass a number of guards. 4439 // The paths which do not pass are accumulated in the slow_region. 4440 RegionNode* slow_region = new (C, 1) RegionNode(1); 4441 record_for_igvn(slow_region); 4442 if (!stopped()) { 4443 // It's an instance (we did array above). Make the slow-path tests. 4444 // If this is a virtual call, we generate a funny guard. We grab 4445 // the vtable entry corresponding to clone() from the target object. 4446 // If the target method which we are calling happens to be the 4447 // Object clone() method, we pass the guard. We do not need this 4448 // guard for non-virtual calls; the caller is known to be the native 4449 // Object clone(). 4450 if (is_virtual) { 4451 generate_virtual_guard(obj_klass, slow_region); 4452 } 4453 4454 // The object must be cloneable and must not have a finalizer. 4455 // Both of these conditions may be checked in a single test. 4456 // We could optimize the cloneable test further, but we don't care. 4457 generate_access_flags_guard(obj_klass, 4458 // Test both conditions: 4459 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4460 // Must be cloneable but not finalizer: 4461 JVM_ACC_IS_CLONEABLE, 4462 slow_region); 4463 } 4464 4465 if (!stopped()) { 4466 // It's an instance, and it passed the slow-path tests. 4467 PreserveJVMState pjvms(this); 4468 Node* obj_size = NULL; 4469 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size); 4470 4471 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4472 4473 // Present the results of the slow call. 4474 result_reg->init_req(_instance_path, control()); 4475 result_val->init_req(_instance_path, alloc_obj); 4476 result_i_o ->set_req(_instance_path, i_o()); 4477 result_mem ->set_req(_instance_path, reset_memory()); 4478 } 4479 4480 // Generate code for the slow case. We make a call to clone(). 4481 set_control(_gvn.transform(slow_region)); 4482 if (!stopped()) { 4483 PreserveJVMState pjvms(this); 4484 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4485 Node* slow_result = set_results_for_java_call(slow_call); 4486 // this->control() comes from set_results_for_java_call 4487 result_reg->init_req(_slow_path, control()); 4488 result_val->init_req(_slow_path, slow_result); 4489 result_i_o ->set_req(_slow_path, i_o()); 4490 result_mem ->set_req(_slow_path, reset_memory()); 4491 } 4492 4493 // Return the combined state. 4494 set_control( _gvn.transform(result_reg) ); 4495 set_i_o( _gvn.transform(result_i_o) ); 4496 set_all_memory( _gvn.transform(result_mem) ); 4497 } //original reexecute and sp are set back here 4498 4499 push(_gvn.transform(result_val)); 4500 4501 return true; 4502} 4503 4504//------------------------------basictype2arraycopy---------------------------- 4505address LibraryCallKit::basictype2arraycopy(BasicType t, 4506 Node* src_offset, 4507 Node* dest_offset, 4508 bool disjoint_bases, 4509 const char* &name, 4510 bool dest_uninitialized) { 4511 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4512 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4513 4514 bool aligned = false; 4515 bool disjoint = disjoint_bases; 4516 4517 // if the offsets are the same, we can treat the memory regions as 4518 // disjoint, because either the memory regions are in different arrays, 4519 // or they are identical (which we can treat as disjoint.) We can also 4520 // treat a copy with a destination index less that the source index 4521 // as disjoint since a low->high copy will work correctly in this case. 4522 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4523 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4524 // both indices are constants 4525 int s_offs = src_offset_inttype->get_con(); 4526 int d_offs = dest_offset_inttype->get_con(); 4527 int element_size = type2aelembytes(t); 4528 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4529 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4530 if (s_offs >= d_offs) disjoint = true; 4531 } else if (src_offset == dest_offset && src_offset != NULL) { 4532 // This can occur if the offsets are identical non-constants. 4533 disjoint = true; 4534 } 4535 4536 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4537} 4538 4539 4540//------------------------------inline_arraycopy----------------------- 4541bool LibraryCallKit::inline_arraycopy() { 4542 // Restore the stack and pop off the arguments. 4543 int nargs = 5; // 2 oops, 3 ints, no size_t or long 4544 assert(callee()->signature()->size() == nargs, "copy has 5 arguments"); 4545 4546 Node *src = argument(0); 4547 Node *src_offset = argument(1); 4548 Node *dest = argument(2); 4549 Node *dest_offset = argument(3); 4550 Node *length = argument(4); 4551 4552 // Compile time checks. If any of these checks cannot be verified at compile time, 4553 // we do not make a fast path for this call. Instead, we let the call remain as it 4554 // is. The checks we choose to mandate at compile time are: 4555 // 4556 // (1) src and dest are arrays. 4557 const Type* src_type = src->Value(&_gvn); 4558 const Type* dest_type = dest->Value(&_gvn); 4559 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4560 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4561 if (top_src == NULL || top_src->klass() == NULL || 4562 top_dest == NULL || top_dest->klass() == NULL) { 4563 // Conservatively insert a memory barrier on all memory slices. 4564 // Do not let writes into the source float below the arraycopy. 4565 insert_mem_bar(Op_MemBarCPUOrder); 4566 4567 // Call StubRoutines::generic_arraycopy stub. 4568 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4569 src, src_offset, dest, dest_offset, length); 4570 4571 // Do not let reads from the destination float above the arraycopy. 4572 // Since we cannot type the arrays, we don't know which slices 4573 // might be affected. We could restrict this barrier only to those 4574 // memory slices which pertain to array elements--but don't bother. 4575 if (!InsertMemBarAfterArraycopy) 4576 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4577 insert_mem_bar(Op_MemBarCPUOrder); 4578 return true; 4579 } 4580 4581 // (2) src and dest arrays must have elements of the same BasicType 4582 // Figure out the size and type of the elements we will be copying. 4583 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4584 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4585 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4586 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4587 4588 if (src_elem != dest_elem || dest_elem == T_VOID) { 4589 // The component types are not the same or are not recognized. Punt. 4590 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4591 generate_slow_arraycopy(TypePtr::BOTTOM, 4592 src, src_offset, dest, dest_offset, length, 4593 /*dest_uninitialized*/false); 4594 return true; 4595 } 4596 4597 //--------------------------------------------------------------------------- 4598 // We will make a fast path for this call to arraycopy. 4599 4600 // We have the following tests left to perform: 4601 // 4602 // (3) src and dest must not be null. 4603 // (4) src_offset must not be negative. 4604 // (5) dest_offset must not be negative. 4605 // (6) length must not be negative. 4606 // (7) src_offset + length must not exceed length of src. 4607 // (8) dest_offset + length must not exceed length of dest. 4608 // (9) each element of an oop array must be assignable 4609 4610 RegionNode* slow_region = new (C, 1) RegionNode(1); 4611 record_for_igvn(slow_region); 4612 4613 // (3) operands must not be null 4614 // We currently perform our null checks with the do_null_check routine. 4615 // This means that the null exceptions will be reported in the caller 4616 // rather than (correctly) reported inside of the native arraycopy call. 4617 // This should be corrected, given time. We do our null check with the 4618 // stack pointer restored. 4619 _sp += nargs; 4620 src = do_null_check(src, T_ARRAY); 4621 dest = do_null_check(dest, T_ARRAY); 4622 _sp -= nargs; 4623 4624 // (4) src_offset must not be negative. 4625 generate_negative_guard(src_offset, slow_region); 4626 4627 // (5) dest_offset must not be negative. 4628 generate_negative_guard(dest_offset, slow_region); 4629 4630 // (6) length must not be negative (moved to generate_arraycopy()). 4631 // generate_negative_guard(length, slow_region); 4632 4633 // (7) src_offset + length must not exceed length of src. 4634 generate_limit_guard(src_offset, length, 4635 load_array_length(src), 4636 slow_region); 4637 4638 // (8) dest_offset + length must not exceed length of dest. 4639 generate_limit_guard(dest_offset, length, 4640 load_array_length(dest), 4641 slow_region); 4642 4643 // (9) each element of an oop array must be assignable 4644 // The generate_arraycopy subroutine checks this. 4645 4646 // This is where the memory effects are placed: 4647 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4648 generate_arraycopy(adr_type, dest_elem, 4649 src, src_offset, dest, dest_offset, length, 4650 false, false, slow_region); 4651 4652 return true; 4653} 4654 4655//-----------------------------generate_arraycopy---------------------- 4656// Generate an optimized call to arraycopy. 4657// Caller must guard against non-arrays. 4658// Caller must determine a common array basic-type for both arrays. 4659// Caller must validate offsets against array bounds. 4660// The slow_region has already collected guard failure paths 4661// (such as out of bounds length or non-conformable array types). 4662// The generated code has this shape, in general: 4663// 4664// if (length == 0) return // via zero_path 4665// slowval = -1 4666// if (types unknown) { 4667// slowval = call generic copy loop 4668// if (slowval == 0) return // via checked_path 4669// } else if (indexes in bounds) { 4670// if ((is object array) && !(array type check)) { 4671// slowval = call checked copy loop 4672// if (slowval == 0) return // via checked_path 4673// } else { 4674// call bulk copy loop 4675// return // via fast_path 4676// } 4677// } 4678// // adjust params for remaining work: 4679// if (slowval != -1) { 4680// n = -1^slowval; src_offset += n; dest_offset += n; length -= n 4681// } 4682// slow_region: 4683// call slow arraycopy(src, src_offset, dest, dest_offset, length) 4684// return // via slow_call_path 4685// 4686// This routine is used from several intrinsics: System.arraycopy, 4687// Object.clone (the array subcase), and Arrays.copyOf[Range]. 4688// 4689void 4690LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 4691 BasicType basic_elem_type, 4692 Node* src, Node* src_offset, 4693 Node* dest, Node* dest_offset, 4694 Node* copy_length, 4695 bool disjoint_bases, 4696 bool length_never_negative, 4697 RegionNode* slow_region) { 4698 4699 if (slow_region == NULL) { 4700 slow_region = new(C,1) RegionNode(1); 4701 record_for_igvn(slow_region); 4702 } 4703 4704 Node* original_dest = dest; 4705 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 4706 bool dest_uninitialized = false; 4707 4708 // See if this is the initialization of a newly-allocated array. 4709 // If so, we will take responsibility here for initializing it to zero. 4710 // (Note: Because tightly_coupled_allocation performs checks on the 4711 // out-edges of the dest, we need to avoid making derived pointers 4712 // from it until we have checked its uses.) 4713 if (ReduceBulkZeroing 4714 && !ZeroTLAB // pointless if already zeroed 4715 && basic_elem_type != T_CONFLICT // avoid corner case 4716 && !src->eqv_uncast(dest) 4717 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 4718 != NULL) 4719 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 4720 && alloc->maybe_set_complete(&_gvn)) { 4721 // "You break it, you buy it." 4722 InitializeNode* init = alloc->initialization(); 4723 assert(init->is_complete(), "we just did this"); 4724 init->set_complete_with_arraycopy(); 4725 assert(dest->is_CheckCastPP(), "sanity"); 4726 assert(dest->in(0)->in(0) == init, "dest pinned"); 4727 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 4728 // From this point on, every exit path is responsible for 4729 // initializing any non-copied parts of the object to zero. 4730 // Also, if this flag is set we make sure that arraycopy interacts properly 4731 // with G1, eliding pre-barriers. See CR 6627983. 4732 dest_uninitialized = true; 4733 } else { 4734 // No zeroing elimination here. 4735 alloc = NULL; 4736 //original_dest = dest; 4737 //dest_uninitialized = false; 4738 } 4739 4740 // Results are placed here: 4741 enum { fast_path = 1, // normal void-returning assembly stub 4742 checked_path = 2, // special assembly stub with cleanup 4743 slow_call_path = 3, // something went wrong; call the VM 4744 zero_path = 4, // bypass when length of copy is zero 4745 bcopy_path = 5, // copy primitive array by 64-bit blocks 4746 PATH_LIMIT = 6 4747 }; 4748 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 4749 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO); 4750 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type); 4751 record_for_igvn(result_region); 4752 _gvn.set_type_bottom(result_i_o); 4753 _gvn.set_type_bottom(result_memory); 4754 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 4755 4756 // The slow_control path: 4757 Node* slow_control; 4758 Node* slow_i_o = i_o(); 4759 Node* slow_mem = memory(adr_type); 4760 debug_only(slow_control = (Node*) badAddress); 4761 4762 // Checked control path: 4763 Node* checked_control = top(); 4764 Node* checked_mem = NULL; 4765 Node* checked_i_o = NULL; 4766 Node* checked_value = NULL; 4767 4768 if (basic_elem_type == T_CONFLICT) { 4769 assert(!dest_uninitialized, ""); 4770 Node* cv = generate_generic_arraycopy(adr_type, 4771 src, src_offset, dest, dest_offset, 4772 copy_length, dest_uninitialized); 4773 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4774 checked_control = control(); 4775 checked_i_o = i_o(); 4776 checked_mem = memory(adr_type); 4777 checked_value = cv; 4778 set_control(top()); // no fast path 4779 } 4780 4781 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 4782 if (not_pos != NULL) { 4783 PreserveJVMState pjvms(this); 4784 set_control(not_pos); 4785 4786 // (6) length must not be negative. 4787 if (!length_never_negative) { 4788 generate_negative_guard(copy_length, slow_region); 4789 } 4790 4791 // copy_length is 0. 4792 if (!stopped() && dest_uninitialized) { 4793 Node* dest_length = alloc->in(AllocateNode::ALength); 4794 if (copy_length->eqv_uncast(dest_length) 4795 || _gvn.find_int_con(dest_length, 1) <= 0) { 4796 // There is no zeroing to do. No need for a secondary raw memory barrier. 4797 } else { 4798 // Clear the whole thing since there are no source elements to copy. 4799 generate_clear_array(adr_type, dest, basic_elem_type, 4800 intcon(0), NULL, 4801 alloc->in(AllocateNode::AllocSize)); 4802 // Use a secondary InitializeNode as raw memory barrier. 4803 // Currently it is needed only on this path since other 4804 // paths have stub or runtime calls as raw memory barriers. 4805 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 4806 Compile::AliasIdxRaw, 4807 top())->as_Initialize(); 4808 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 4809 } 4810 } 4811 4812 // Present the results of the fast call. 4813 result_region->init_req(zero_path, control()); 4814 result_i_o ->init_req(zero_path, i_o()); 4815 result_memory->init_req(zero_path, memory(adr_type)); 4816 } 4817 4818 if (!stopped() && dest_uninitialized) { 4819 // We have to initialize the *uncopied* part of the array to zero. 4820 // The copy destination is the slice dest[off..off+len]. The other slices 4821 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 4822 Node* dest_size = alloc->in(AllocateNode::AllocSize); 4823 Node* dest_length = alloc->in(AllocateNode::ALength); 4824 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset, 4825 copy_length) ); 4826 4827 // If there is a head section that needs zeroing, do it now. 4828 if (find_int_con(dest_offset, -1) != 0) { 4829 generate_clear_array(adr_type, dest, basic_elem_type, 4830 intcon(0), dest_offset, 4831 NULL); 4832 } 4833 4834 // Next, perform a dynamic check on the tail length. 4835 // It is often zero, and we can win big if we prove this. 4836 // There are two wins: Avoid generating the ClearArray 4837 // with its attendant messy index arithmetic, and upgrade 4838 // the copy to a more hardware-friendly word size of 64 bits. 4839 Node* tail_ctl = NULL; 4840 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 4841 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) ); 4842 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) ); 4843 tail_ctl = generate_slow_guard(bol_lt, NULL); 4844 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 4845 } 4846 4847 // At this point, let's assume there is no tail. 4848 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 4849 // There is no tail. Try an upgrade to a 64-bit copy. 4850 bool didit = false; 4851 { PreserveJVMState pjvms(this); 4852 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 4853 src, src_offset, dest, dest_offset, 4854 dest_size, dest_uninitialized); 4855 if (didit) { 4856 // Present the results of the block-copying fast call. 4857 result_region->init_req(bcopy_path, control()); 4858 result_i_o ->init_req(bcopy_path, i_o()); 4859 result_memory->init_req(bcopy_path, memory(adr_type)); 4860 } 4861 } 4862 if (didit) 4863 set_control(top()); // no regular fast path 4864 } 4865 4866 // Clear the tail, if any. 4867 if (tail_ctl != NULL) { 4868 Node* notail_ctl = stopped() ? NULL : control(); 4869 set_control(tail_ctl); 4870 if (notail_ctl == NULL) { 4871 generate_clear_array(adr_type, dest, basic_elem_type, 4872 dest_tail, NULL, 4873 dest_size); 4874 } else { 4875 // Make a local merge. 4876 Node* done_ctl = new(C,3) RegionNode(3); 4877 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type); 4878 done_ctl->init_req(1, notail_ctl); 4879 done_mem->init_req(1, memory(adr_type)); 4880 generate_clear_array(adr_type, dest, basic_elem_type, 4881 dest_tail, NULL, 4882 dest_size); 4883 done_ctl->init_req(2, control()); 4884 done_mem->init_req(2, memory(adr_type)); 4885 set_control( _gvn.transform(done_ctl) ); 4886 set_memory( _gvn.transform(done_mem), adr_type ); 4887 } 4888 } 4889 } 4890 4891 BasicType copy_type = basic_elem_type; 4892 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 4893 if (!stopped() && copy_type == T_OBJECT) { 4894 // If src and dest have compatible element types, we can copy bits. 4895 // Types S[] and D[] are compatible if D is a supertype of S. 4896 // 4897 // If they are not, we will use checked_oop_disjoint_arraycopy, 4898 // which performs a fast optimistic per-oop check, and backs off 4899 // further to JVM_ArrayCopy on the first per-oop check that fails. 4900 // (Actually, we don't move raw bits only; the GC requires card marks.) 4901 4902 // Get the Klass* for both src and dest 4903 Node* src_klass = load_object_klass(src); 4904 Node* dest_klass = load_object_klass(dest); 4905 4906 // Generate the subtype check. 4907 // This might fold up statically, or then again it might not. 4908 // 4909 // Non-static example: Copying List<String>.elements to a new String[]. 4910 // The backing store for a List<String> is always an Object[], 4911 // but its elements are always type String, if the generic types 4912 // are correct at the source level. 4913 // 4914 // Test S[] against D[], not S against D, because (probably) 4915 // the secondary supertype cache is less busy for S[] than S. 4916 // This usually only matters when D is an interface. 4917 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4918 // Plug failing path into checked_oop_disjoint_arraycopy 4919 if (not_subtype_ctrl != top()) { 4920 PreserveJVMState pjvms(this); 4921 set_control(not_subtype_ctrl); 4922 // (At this point we can assume disjoint_bases, since types differ.) 4923 int ek_offset = in_bytes(objArrayKlass::element_klass_offset()); 4924 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 4925 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM); 4926 Node* dest_elem_klass = _gvn.transform(n1); 4927 Node* cv = generate_checkcast_arraycopy(adr_type, 4928 dest_elem_klass, 4929 src, src_offset, dest, dest_offset, 4930 ConvI2X(copy_length), dest_uninitialized); 4931 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 4932 checked_control = control(); 4933 checked_i_o = i_o(); 4934 checked_mem = memory(adr_type); 4935 checked_value = cv; 4936 } 4937 // At this point we know we do not need type checks on oop stores. 4938 4939 // Let's see if we need card marks: 4940 if (alloc != NULL && use_ReduceInitialCardMarks()) { 4941 // If we do not need card marks, copy using the jint or jlong stub. 4942 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 4943 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 4944 "sizes agree"); 4945 } 4946 } 4947 4948 if (!stopped()) { 4949 // Generate the fast path, if possible. 4950 PreserveJVMState pjvms(this); 4951 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 4952 src, src_offset, dest, dest_offset, 4953 ConvI2X(copy_length), dest_uninitialized); 4954 4955 // Present the results of the fast call. 4956 result_region->init_req(fast_path, control()); 4957 result_i_o ->init_req(fast_path, i_o()); 4958 result_memory->init_req(fast_path, memory(adr_type)); 4959 } 4960 4961 // Here are all the slow paths up to this point, in one bundle: 4962 slow_control = top(); 4963 if (slow_region != NULL) 4964 slow_control = _gvn.transform(slow_region); 4965 debug_only(slow_region = (RegionNode*)badAddress); 4966 4967 set_control(checked_control); 4968 if (!stopped()) { 4969 // Clean up after the checked call. 4970 // The returned value is either 0 or -1^K, 4971 // where K = number of partially transferred array elements. 4972 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) ); 4973 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 4974 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4975 4976 // If it is 0, we are done, so transfer to the end. 4977 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) ); 4978 result_region->init_req(checked_path, checks_done); 4979 result_i_o ->init_req(checked_path, checked_i_o); 4980 result_memory->init_req(checked_path, checked_mem); 4981 4982 // If it is not zero, merge into the slow call. 4983 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) )); 4984 RegionNode* slow_reg2 = new(C, 3) RegionNode(3); 4985 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO); 4986 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type); 4987 record_for_igvn(slow_reg2); 4988 slow_reg2 ->init_req(1, slow_control); 4989 slow_i_o2 ->init_req(1, slow_i_o); 4990 slow_mem2 ->init_req(1, slow_mem); 4991 slow_reg2 ->init_req(2, control()); 4992 slow_i_o2 ->init_req(2, checked_i_o); 4993 slow_mem2 ->init_req(2, checked_mem); 4994 4995 slow_control = _gvn.transform(slow_reg2); 4996 slow_i_o = _gvn.transform(slow_i_o2); 4997 slow_mem = _gvn.transform(slow_mem2); 4998 4999 if (alloc != NULL) { 5000 // We'll restart from the very beginning, after zeroing the whole thing. 5001 // This can cause double writes, but that's OK since dest is brand new. 5002 // So we ignore the low 31 bits of the value returned from the stub. 5003 } else { 5004 // We must continue the copy exactly where it failed, or else 5005 // another thread might see the wrong number of writes to dest. 5006 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) ); 5007 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT); 5008 slow_offset->init_req(1, intcon(0)); 5009 slow_offset->init_req(2, checked_offset); 5010 slow_offset = _gvn.transform(slow_offset); 5011 5012 // Adjust the arguments by the conditionally incoming offset. 5013 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) ); 5014 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) ); 5015 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) ); 5016 5017 // Tweak the node variables to adjust the code produced below: 5018 src_offset = src_off_plus; 5019 dest_offset = dest_off_plus; 5020 copy_length = length_minus; 5021 } 5022 } 5023 5024 set_control(slow_control); 5025 if (!stopped()) { 5026 // Generate the slow path, if needed. 5027 PreserveJVMState pjvms(this); // replace_in_map may trash the map 5028 5029 set_memory(slow_mem, adr_type); 5030 set_i_o(slow_i_o); 5031 5032 if (dest_uninitialized) { 5033 generate_clear_array(adr_type, dest, basic_elem_type, 5034 intcon(0), NULL, 5035 alloc->in(AllocateNode::AllocSize)); 5036 } 5037 5038 generate_slow_arraycopy(adr_type, 5039 src, src_offset, dest, dest_offset, 5040 copy_length, /*dest_uninitialized*/false); 5041 5042 result_region->init_req(slow_call_path, control()); 5043 result_i_o ->init_req(slow_call_path, i_o()); 5044 result_memory->init_req(slow_call_path, memory(adr_type)); 5045 } 5046 5047 // Remove unused edges. 5048 for (uint i = 1; i < result_region->req(); i++) { 5049 if (result_region->in(i) == NULL) 5050 result_region->init_req(i, top()); 5051 } 5052 5053 // Finished; return the combined state. 5054 set_control( _gvn.transform(result_region) ); 5055 set_i_o( _gvn.transform(result_i_o) ); 5056 set_memory( _gvn.transform(result_memory), adr_type ); 5057 5058 // The memory edges above are precise in order to model effects around 5059 // array copies accurately to allow value numbering of field loads around 5060 // arraycopy. Such field loads, both before and after, are common in Java 5061 // collections and similar classes involving header/array data structures. 5062 // 5063 // But with low number of register or when some registers are used or killed 5064 // by arraycopy calls it causes registers spilling on stack. See 6544710. 5065 // The next memory barrier is added to avoid it. If the arraycopy can be 5066 // optimized away (which it can, sometimes) then we can manually remove 5067 // the membar also. 5068 // 5069 // Do not let reads from the cloned object float above the arraycopy. 5070 if (alloc != NULL) { 5071 // Do not let stores that initialize this object be reordered with 5072 // a subsequent store that would make this object accessible by 5073 // other threads. 5074 // Record what AllocateNode this StoreStore protects so that 5075 // escape analysis can go from the MemBarStoreStoreNode to the 5076 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5077 // based on the escape status of the AllocateNode. 5078 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5079 } else if (InsertMemBarAfterArraycopy) 5080 insert_mem_bar(Op_MemBarCPUOrder); 5081} 5082 5083 5084// Helper function which determines if an arraycopy immediately follows 5085// an allocation, with no intervening tests or other escapes for the object. 5086AllocateArrayNode* 5087LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5088 RegionNode* slow_region) { 5089 if (stopped()) return NULL; // no fast path 5090 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5091 5092 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5093 if (alloc == NULL) return NULL; 5094 5095 Node* rawmem = memory(Compile::AliasIdxRaw); 5096 // Is the allocation's memory state untouched? 5097 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5098 // Bail out if there have been raw-memory effects since the allocation. 5099 // (Example: There might have been a call or safepoint.) 5100 return NULL; 5101 } 5102 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5103 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5104 return NULL; 5105 } 5106 5107 // There must be no unexpected observers of this allocation. 5108 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5109 Node* obs = ptr->fast_out(i); 5110 if (obs != this->map()) { 5111 return NULL; 5112 } 5113 } 5114 5115 // This arraycopy must unconditionally follow the allocation of the ptr. 5116 Node* alloc_ctl = ptr->in(0); 5117 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5118 5119 Node* ctl = control(); 5120 while (ctl != alloc_ctl) { 5121 // There may be guards which feed into the slow_region. 5122 // Any other control flow means that we might not get a chance 5123 // to finish initializing the allocated object. 5124 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5125 IfNode* iff = ctl->in(0)->as_If(); 5126 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5127 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5128 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5129 ctl = iff->in(0); // This test feeds the known slow_region. 5130 continue; 5131 } 5132 // One more try: Various low-level checks bottom out in 5133 // uncommon traps. If the debug-info of the trap omits 5134 // any reference to the allocation, as we've already 5135 // observed, then there can be no objection to the trap. 5136 bool found_trap = false; 5137 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5138 Node* obs = not_ctl->fast_out(j); 5139 if (obs->in(0) == not_ctl && obs->is_Call() && 5140 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5141 found_trap = true; break; 5142 } 5143 } 5144 if (found_trap) { 5145 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5146 continue; 5147 } 5148 } 5149 return NULL; 5150 } 5151 5152 // If we get this far, we have an allocation which immediately 5153 // precedes the arraycopy, and we can take over zeroing the new object. 5154 // The arraycopy will finish the initialization, and provide 5155 // a new control state to which we will anchor the destination pointer. 5156 5157 return alloc; 5158} 5159 5160// Helper for initialization of arrays, creating a ClearArray. 5161// It writes zero bits in [start..end), within the body of an array object. 5162// The memory effects are all chained onto the 'adr_type' alias category. 5163// 5164// Since the object is otherwise uninitialized, we are free 5165// to put a little "slop" around the edges of the cleared area, 5166// as long as it does not go back into the array's header, 5167// or beyond the array end within the heap. 5168// 5169// The lower edge can be rounded down to the nearest jint and the 5170// upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5171// 5172// Arguments: 5173// adr_type memory slice where writes are generated 5174// dest oop of the destination array 5175// basic_elem_type element type of the destination 5176// slice_idx array index of first element to store 5177// slice_len number of elements to store (or NULL) 5178// dest_size total size in bytes of the array object 5179// 5180// Exactly one of slice_len or dest_size must be non-NULL. 5181// If dest_size is non-NULL, zeroing extends to the end of the object. 5182// If slice_len is non-NULL, the slice_idx value must be a constant. 5183void 5184LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5185 Node* dest, 5186 BasicType basic_elem_type, 5187 Node* slice_idx, 5188 Node* slice_len, 5189 Node* dest_size) { 5190 // one or the other but not both of slice_len and dest_size: 5191 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5192 if (slice_len == NULL) slice_len = top(); 5193 if (dest_size == NULL) dest_size = top(); 5194 5195 // operate on this memory slice: 5196 Node* mem = memory(adr_type); // memory slice to operate on 5197 5198 // scaling and rounding of indexes: 5199 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5200 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5201 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5202 int bump_bit = (-1 << scale) & BytesPerInt; 5203 5204 // determine constant starts and ends 5205 const intptr_t BIG_NEG = -128; 5206 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5207 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5208 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5209 if (slice_len_con == 0) { 5210 return; // nothing to do here 5211 } 5212 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5213 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5214 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5215 assert(end_con < 0, "not two cons"); 5216 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5217 BytesPerLong); 5218 } 5219 5220 if (start_con >= 0 && end_con >= 0) { 5221 // Constant start and end. Simple. 5222 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5223 start_con, end_con, &_gvn); 5224 } else if (start_con >= 0 && dest_size != top()) { 5225 // Constant start, pre-rounded end after the tail of the array. 5226 Node* end = dest_size; 5227 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5228 start_con, end, &_gvn); 5229 } else if (start_con >= 0 && slice_len != top()) { 5230 // Constant start, non-constant end. End needs rounding up. 5231 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5232 intptr_t end_base = abase + (slice_idx_con << scale); 5233 int end_round = (-1 << scale) & (BytesPerLong - 1); 5234 Node* end = ConvI2X(slice_len); 5235 if (scale != 0) 5236 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) )); 5237 end_base += end_round; 5238 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) ); 5239 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) ); 5240 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5241 start_con, end, &_gvn); 5242 } else if (start_con < 0 && dest_size != top()) { 5243 // Non-constant start, pre-rounded end after the tail of the array. 5244 // This is almost certainly a "round-to-end" operation. 5245 Node* start = slice_idx; 5246 start = ConvI2X(start); 5247 if (scale != 0) 5248 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) )); 5249 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) ); 5250 if ((bump_bit | clear_low) != 0) { 5251 int to_clear = (bump_bit | clear_low); 5252 // Align up mod 8, then store a jint zero unconditionally 5253 // just before the mod-8 boundary. 5254 if (((abase + bump_bit) & ~to_clear) - bump_bit 5255 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5256 bump_bit = 0; 5257 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5258 } else { 5259 // Bump 'start' up to (or past) the next jint boundary: 5260 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) ); 5261 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5262 } 5263 // Round bumped 'start' down to jlong boundary in body of array. 5264 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) ); 5265 if (bump_bit != 0) { 5266 // Store a zero to the immediately preceding jint: 5267 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) ); 5268 Node* p1 = basic_plus_adr(dest, x1); 5269 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT); 5270 mem = _gvn.transform(mem); 5271 } 5272 } 5273 Node* end = dest_size; // pre-rounded 5274 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5275 start, end, &_gvn); 5276 } else { 5277 // Non-constant start, unrounded non-constant end. 5278 // (Nobody zeroes a random midsection of an array using this routine.) 5279 ShouldNotReachHere(); // fix caller 5280 } 5281 5282 // Done. 5283 set_memory(mem, adr_type); 5284} 5285 5286 5287bool 5288LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5289 BasicType basic_elem_type, 5290 AllocateNode* alloc, 5291 Node* src, Node* src_offset, 5292 Node* dest, Node* dest_offset, 5293 Node* dest_size, bool dest_uninitialized) { 5294 // See if there is an advantage from block transfer. 5295 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5296 if (scale >= LogBytesPerLong) 5297 return false; // it is already a block transfer 5298 5299 // Look at the alignment of the starting offsets. 5300 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5301 5302 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5303 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5304 if (src_off_con < 0 || dest_off_con < 0) 5305 // At present, we can only understand constants. 5306 return false; 5307 5308 intptr_t src_off = abase + (src_off_con << scale); 5309 intptr_t dest_off = abase + (dest_off_con << scale); 5310 5311 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5312 // Non-aligned; too bad. 5313 // One more chance: Pick off an initial 32-bit word. 5314 // This is a common case, since abase can be odd mod 8. 5315 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5316 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5317 Node* sptr = basic_plus_adr(src, src_off); 5318 Node* dptr = basic_plus_adr(dest, dest_off); 5319 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type); 5320 store_to_memory(control(), dptr, sval, T_INT, adr_type); 5321 src_off += BytesPerInt; 5322 dest_off += BytesPerInt; 5323 } else { 5324 return false; 5325 } 5326 } 5327 assert(src_off % BytesPerLong == 0, ""); 5328 assert(dest_off % BytesPerLong == 0, ""); 5329 5330 // Do this copy by giant steps. 5331 Node* sptr = basic_plus_adr(src, src_off); 5332 Node* dptr = basic_plus_adr(dest, dest_off); 5333 Node* countx = dest_size; 5334 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) ); 5335 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) ); 5336 5337 bool disjoint_bases = true; // since alloc != NULL 5338 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5339 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5340 5341 return true; 5342} 5343 5344 5345// Helper function; generates code for the slow case. 5346// We make a call to a runtime method which emulates the native method, 5347// but without the native wrapper overhead. 5348void 5349LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5350 Node* src, Node* src_offset, 5351 Node* dest, Node* dest_offset, 5352 Node* copy_length, bool dest_uninitialized) { 5353 assert(!dest_uninitialized, "Invariant"); 5354 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5355 OptoRuntime::slow_arraycopy_Type(), 5356 OptoRuntime::slow_arraycopy_Java(), 5357 "slow_arraycopy", adr_type, 5358 src, src_offset, dest, dest_offset, 5359 copy_length); 5360 5361 // Handle exceptions thrown by this fellow: 5362 make_slow_call_ex(call, env()->Throwable_klass(), false); 5363} 5364 5365// Helper function; generates code for cases requiring runtime checks. 5366Node* 5367LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5368 Node* dest_elem_klass, 5369 Node* src, Node* src_offset, 5370 Node* dest, Node* dest_offset, 5371 Node* copy_length, bool dest_uninitialized) { 5372 if (stopped()) return NULL; 5373 5374 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5375 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5376 return NULL; 5377 } 5378 5379 // Pick out the parameters required to perform a store-check 5380 // for the target array. This is an optimistic check. It will 5381 // look in each non-null element's class, at the desired klass's 5382 // super_check_offset, for the desired klass. 5383 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5384 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5385 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr()); 5386 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5387 Node* check_value = dest_elem_klass; 5388 5389 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5390 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5391 5392 // (We know the arrays are never conjoint, because their types differ.) 5393 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5394 OptoRuntime::checkcast_arraycopy_Type(), 5395 copyfunc_addr, "checkcast_arraycopy", adr_type, 5396 // five arguments, of which two are 5397 // intptr_t (jlong in LP64) 5398 src_start, dest_start, 5399 copy_length XTOP, 5400 check_offset XTOP, 5401 check_value); 5402 5403 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5404} 5405 5406 5407// Helper function; generates code for cases requiring runtime checks. 5408Node* 5409LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5410 Node* src, Node* src_offset, 5411 Node* dest, Node* dest_offset, 5412 Node* copy_length, bool dest_uninitialized) { 5413 assert(!dest_uninitialized, "Invariant"); 5414 if (stopped()) return NULL; 5415 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5416 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5417 return NULL; 5418 } 5419 5420 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5421 OptoRuntime::generic_arraycopy_Type(), 5422 copyfunc_addr, "generic_arraycopy", adr_type, 5423 src, src_offset, dest, dest_offset, copy_length); 5424 5425 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 5426} 5427 5428// Helper function; generates the fast out-of-line call to an arraycopy stub. 5429void 5430LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5431 BasicType basic_elem_type, 5432 bool disjoint_bases, 5433 Node* src, Node* src_offset, 5434 Node* dest, Node* dest_offset, 5435 Node* copy_length, bool dest_uninitialized) { 5436 if (stopped()) return; // nothing to do 5437 5438 Node* src_start = src; 5439 Node* dest_start = dest; 5440 if (src_offset != NULL || dest_offset != NULL) { 5441 assert(src_offset != NULL && dest_offset != NULL, ""); 5442 src_start = array_element_address(src, src_offset, basic_elem_type); 5443 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5444 } 5445 5446 // Figure out which arraycopy runtime method to call. 5447 const char* copyfunc_name = "arraycopy"; 5448 address copyfunc_addr = 5449 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5450 disjoint_bases, copyfunc_name, dest_uninitialized); 5451 5452 // Call it. Note that the count_ix value is not scaled to a byte-size. 5453 make_runtime_call(RC_LEAF|RC_NO_FP, 5454 OptoRuntime::fast_arraycopy_Type(), 5455 copyfunc_addr, copyfunc_name, adr_type, 5456 src_start, dest_start, copy_length XTOP); 5457} 5458 5459//----------------------------inline_reference_get---------------------------- 5460 5461bool LibraryCallKit::inline_reference_get() { 5462 const int nargs = 1; // self 5463 5464 guarantee(java_lang_ref_Reference::referent_offset > 0, 5465 "should have already been set"); 5466 5467 int referent_offset = java_lang_ref_Reference::referent_offset; 5468 5469 // Restore the stack and pop off the argument 5470 _sp += nargs; 5471 Node *reference_obj = pop(); 5472 5473 // Null check on self without removing any arguments. 5474 _sp += nargs; 5475 reference_obj = do_null_check(reference_obj, T_OBJECT); 5476 _sp -= nargs;; 5477 5478 if (stopped()) return true; 5479 5480 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5481 5482 ciInstanceKlass* klass = env()->Object_klass(); 5483 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5484 5485 Node* no_ctrl = NULL; 5486 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT); 5487 5488 // Use the pre-barrier to record the value in the referent field 5489 pre_barrier(false /* do_load */, 5490 control(), 5491 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5492 result /* pre_val */, 5493 T_OBJECT); 5494 5495 // Add memory barrier to prevent commoning reads from this field 5496 // across safepoint since GC can change its value. 5497 insert_mem_bar(Op_MemBarCPUOrder); 5498 5499 push(result); 5500 return true; 5501} 5502